TECHNICS

MAGAZINE OF THE UNION OF ENGINEERS AND TECHNICIANS OF SERBIA

ISSN 0040 - 2176 UDC: 62 (062. 2) (497.1)

Publisher: Union of Engineers and Technicians of Serbia, Belgrade

Kneza Miloa 7a, tel: +381/011/ 32 35 891, Fax: +381/11/32 30 067,
www.sits.org.rs; e-mail:tehnika@sits.rs; office@sits.rs

FOR PUBLISHER:
BRANISLAV VUJINOVI, M.Sc, Secretary General of the Union of Engineers and Technicians of
Serbia

HEAD EDITOR -IN -CHEF

Prof. SMILJAN VUKANOVI, D.Sc, Faculty of Transport and Traffic Engineering, Belgrade

EDITORIAL BOARD
Prof DRAGAN USKOKOVI, D.Sc (New Materials); Prof. BRANKO BOI, M.Sc (Our Civil
Engineering); Prof BOO KOLONJA, D.Sc (Mining, Geology and Metallurgy); Prof BOKO
RAUO, D.Sc (Mechanical Engineering); Prof. VUJO DRNDAREVI, D.Sc (Electrical Engineering);
Prof SMILJAN VUKANOVI, D.Sc (Traffic); Prof VUJADIN VEOVI, D.Sc (Management);
ZORAN PENDI, M.Sc (Quality - IMS, Standardization and Metrology)

Technical editors
Olja Jovii and Olivera osovi

MAGAZINE TECHNICS IS PARTLY FINANCED BY THE MINISTRY OF EDUCATION,

SCIENCE AND TECHNOLOGICAL DEVELOPMENT OF THE REPUBLIC OF SERBIA

CIP Published papers are being referred to in inter-

, nationally renown magazines, such as: Geotech-
62
TEHNIKA: asopis Saveza inenjera i
tehniara Srbije / glavni i odgovorni urednik
Smiljan Vukanovi. - God. 1, br. 1 (1946) -
- Beograd: Savez inenjera i tehniara
Srbije, 1946 - (Zemun: Grafiki atelje Dunav)
- 29 cm
Dvomeseno. - Ima podzbirke: Novi materijali =
ISSN 0354-2300, Nae graevinarstvo =
ISSN 0350-2619, Rudarstvo, geologija i
metalurgija = ISSN 0350-2627, Mainstvo =
ISSN 0461-2531, Elektrotehnika (Beograd) =
ISSN
0013-5836, Saobraaj = ISSN 0558-6208,
Menadment = ISSN 1450-9911, Kvalitet,
standardizacija i metrologija = ISSN 2334-7368
ISSN 0040-2176 = Tehnika (Beograd)
COBISS.SR-ID 2527490
nical Abstracts, Metal Abstracts, Chemical Abs-
tracts, Electrical and Electronics Abstracts, Scie-
nece Abstracts, Ergonomics Abstracts and reference
journals VINITI.
They also serve as background source of infor-
mation for international data base requirements:
SAIDC-el (ISKRA,SL), SAIDC-gr (CTK, SL),
INSPEC (IEEE, UK), METADEX (M. I., UK),
CASEARCH (CA, USA)
All manuscripts are not to be returned.
Computer typeface and paging:
Union of Engineers and Technicians of Serbia
Telefon: +381/11/32 35 891
Printed by: Grafiki atelje Dunav, Zemun
Circulation: 360 copies
 UNION OF ENGINEERS AND TECHNICIANS OF SERBIA
BACKGROUND
The roots of Serbian technical civilization date as
early as the time of the Nemanjics. Beginnings of
engineering activities were associated to the mining
and metallurgical undertakings (Novo Brdo) and to
building of magnificent medieval sacral structures of
the Serbian state.
After the First (1804) and second Serbian Up-
rising (1815) the technical tradition was renewed and
Serbian joined the then current European trends. First
educated engineers came in Serbia from Austro-Hu-
ngarian Empire in 1830s. At that time, the main pre-
occupations of engineers were railway construction,
town planning, construction of sewage disposal and
water supply systems, as well as creating of national
defense system. At that time 1834/35 from Austrian
Empire arrived first schooled engineers France Janke
and Franz Baron Kordon who served as so called
drzavni indzilirin or state engineers.
In Serbia in the 19th century there were a total
number of about 6000 engineers engaged in various
activities. In an eightyyear period from 1834-1914
the State Construction Administration (which from
1880 also included railways) employed one third of
these engineers. However other ministries were also
competent for some engineering affaires like, for ex-
ample the Ministry of Finance was responsible for
mining, or the Ministry of Education and Church Aff-
aires was responsible for education of technical stuff.
From 1838 this primarily referred to the Licej: acco-
rding to Establishment of public institutions of lear-
ning of 1844, the Department for Philosophy inclu-
ded also subjects such as Pure and Practical Geometry
and Higher Mathematics, and Architecture, while in
1853 a separate Natural Sciences and Technical De-
partment was introduced in the Licej and in 1863 the
Great School with Technical Faculty started ope-
rating. The first classes held at the Technical Faculty
of the Great School in 1863 marked turning point in
schooling of Serbian engineers.
Out of some 600 engineers, approximately one th-
ird were schooled in Serbia and one fifth of them stu-
died abroad as state grants students, while about
one fourth were foreigners and Serbs from across the
Danube.
In 1868 one of preconditions which might have
contributed to professional associating of engineers
was the numerosity of professionals and models from
abroad established half a century earlier (engineering
associations in Great Britain, Germany and America)
had influence on establishing professional associati-
ons in Serbia.
The Founding Assembly of the Technicians` So-
ciety was held on the 3rd February 1868 in the pre-
mises of Great School. Engineer Emilijan Josimovic
was elected for the first President of the Society. It is
important to mention that this happened only a year
after Turkish commander in Belgrade Ali -Riza pasha
gave the town and the fortress keys to duke Mihailo
Obrenovic. Shortly afterward in 1869 was established
Society for Agrarian Economy that is the Serbian
Agricultural Society. Association of Serbian Engine-
ers was established in 1890 while in 1896 was esta-
blished the Association of Serbian Engineers and Ar-
chitects.
The first scientific magazine published by this
Association in 1890 was Srpski tehnicki list The
Srpski tehnicki list besides professional articles also
published detailed information related to the work of
the Association. The members at that time, who nu-
mbered around one hundred of them, initiated a who-
le series of issues and demand the same to be solved
by the competent bodies. During the First World War,
two volumes of Srpski tehnicki list were published
in Thessaloniki. The magazine was initiated by the
engineers and architects who were in Thessloniki as
members of the Serbian Army. In Thessaloniki was
held the General Assembly of the Association in 1918
attended by 463 engineers.
During his short stay in Belgrade, in 1892, fa-
mous scientist Nikola Tesla was elected for the first
honorary member of the Association of Serbian En-
gineers.
Providing assets from its own incomes, bank lo-
ans, gifts and donations of its organizations-members
and its individual members Association built the
House of Engineers in Belgrade, Kneza Milosa 7 str
in 1932/35. The House of Engineers Nikola Tesla
in Belgrade Kneza Milosa 9-11 str was built between
1962 and 1969. In the premises of these two Houses
of Engineers besides the Union of Engineers and Te-
chnicians of Serbia today perform their activities 26
republic`s professional and multidisciplinary engine-
ering-technicians` associations out of 41 collective
members of UETS.
Besides Emilijan Josimovic who was first Pre-
sident of the Technicians` Society, prominent figure
of that time, Rector of Licej and Great School and
honorary member of the Serbian Royal Academy, to
work of our Union contributed as well: Kosta Al-
kovic, professor at the Great School, Minister of Co-
nstruction and member of Serbian Learned Society
and Serbian Royal Academy, Dimitrije Stojanovic
professor at the Technical Faculty, first Director of
Serbian State Railways, and member of Serbian
Learned Society and Serbian Royal Academy, Milos
Savcic, Minister of Construction and President of
Belgrade Municipality, famous businessman who
gave the greatest donation for the construction of Ho-
use of Engineers in 1932, as well as presidents of the
Serbian Academy of Sciences and Arts Josif Pancic,
Jovan Zujovic, Simo Lozanic, Kirilo Savic, Aleksa-
ndar Despic, Nikola Hajdin and other famous
scientists.
ACTIVITIES
The Union of Engineers and Technicians of Se-
rbia - Savez inenjera i tehnicara Srbije is a voluntary,
non-governmental, non-profit, scientific, interest,
professional, non-party organization of engineers and
technicians, and their organizations in the Republic of
Serbia, open for cooperation with other scientific, co-
mmercial and other organizations, on the basis of
mutual recognition, mutual respect and independence
in work.
Union of Engineers and Technicians of Serbia
and its collective member finance their own activities
from their own assets.
Purposes and tasks of UETS are:
Assembling and organizing of engineers and te-
chnicians of Serbia for the purpose of increase of
their expert knowledge, providing appropriate
status in the community, on the basis of their co-
ntribution to the, scientific-technological and eco-
nomic and development in general of Republic of
Serbia;
Joining, strengthening and massification of basic
engineering-technicians' organizations of Serbia,
development of mutual cooperation as well as the
cooperation with appropriate international organi-
zations of engineers and technicians;
Improvement of order-interest, reputation and pr-
otection of members of the engineering-techni-
cians' organization of Serbia;
Providing help to engineers and technicians in
scientific, expert improvement and organization
of appropriate forms of permanent education;
Monitoring contemporary development of engi-
neering and technology and pointing out the cur-
rents of events and changes in this area and pro-
viding opinions on optimality of engineering and
technological solutions in investment and other
enterprises;
Caring for and development of ethics of engine-
ering-technician profession, human rights and
liberties;
Stimulating, organization and publishing of scie-
ntific and expert papers, magazines and other pu-
blications of interest for engineering-technician
organization and technical intelligence;
Work on technical regulations (laws, regulations
and standards), providing its modernity, adequ-
acy, actuality and functionality;
Consideration and providing expert opinions on
plans, programs, analysis and other acts, which
are important for the development of engineering,
technology and production in the Republic of
Serbia;
Stimulating and helping the activities and ini-
tiatives, aiming to preserve the human environ-
ment and area organization, saving and rationali-
zation of spending of all sorts of energy;
Preparation and maintenance of the meetings wi-
th purpose of permanent education of engineers
and technicians;
Providing help in development of technology and
economy whose purposes are similar to the
purposes of engineering-technicians` organizati-
on;
Organization of multidisciplinary meetings and
meetings of wider social importance;
Cooperation with appropriate expert, commercial
organizations and other organizations and organs
at the realization of tasks of mutual interest;
 Management of Houses of Engineers and other
property of Union of Engineers and Technicians
of Serbia.
Union of Engineers and Technicians of Serbia has
developed cooperation with organs of local gove-
rnment, state ministries, Serbian Academy of Scie-
nces and Arts, Serbian, Chamber of Engineers, Engi-
neering Academy of Serbia, Chamber of Commerce
and Industry of Serbia, with numerous companies,
professional associations, faculties and universities
and other institutions. UETS also has developer inter-
national cooperation.
In accordance with the Law and Contract with
republic ministries in the framework of UETS are
organized and performed specialist` exams for several
engineering branches.
Union of Engineers and Technicians of Serbia has
several thousand individual members and 41
collective members in the Republic of Serbia: 19
republic`s professional associations (associations of
architects, town planners, mechanical engineers, ele-
ctrical engineers, mining and geological engineers,
surveyors, agricultural engineers, chemical engineers
etc) 7 republic`s multidisciplinary engineering-techni-
cians` associations (ecology, standardization and
quality, material protection and corrosion, informatics
etc) 1 provincial engineering-technicians` association,
14 municipal and regional engineering-technicians`
associations.
Union of Engineers and Technicians of Serbia is
founder of the Engineering Academy of Serbia, and
collective member of the Chamber of Commerce and
Industry of Serbia.
Union of Engineers and Technicians of Serbia, in
a cooperation with faculties, universities, enterprises,
economic and professional associations organizes
various scientific meetings, professional reunions, co-
ngresses, seminars, conferences. UETS members
publish their expert magazines; KGH; Procesna te-
hnika, Ecologica, Tekstilna industrija, Forum,
Sumarska industrija, Zastita materijala and main-
tain professional reunions, seminars, conferences and
congresses in branches of architecture, mechanical
engineering, chemistry, electrical engineering, agricu-
lture, forestry etc.
All activities of the Union are performed in acco-
rdance with the procedures and standards of QMS -
Quality Management System.
Union of Engineers and Technicians of Serbia is
National member of FEANI European Federation
of National Engineering Associations from Serbia.
FEANI is a federation of professional engineers that
unites national engineering associations from 32
European countries. Thus, FEANI represents the in-
terests of over 3,5 million professional engineers in
Europe. FEANI is striving for a single voice for the
engineering profession in Europe and wants to affirm
and develop the professional identity of engineers.
Through its activities and services, especially with the
attribution of the EUR ING professional title, FEANI
aims to facilitate the mutual recognition of engine-
ering qualifications in Europe and to strengthen the
position, role and responsibility of engineers in so-
ciety.
Union of Engineers and Technicians of Serbia is
member of COPISSE Permanent Conference of the
Engineers of Southeast Europe.
Collective members of UETS are members of
international professional associations and have deve-
loped international cooperation.
With all that has been done and with acco-
mplished results, objectively solid conditions have
been provided for further and more successful work,
business operation and development of the Union of
Engineers and Technicians of Serbia.
 CONTENTS TECHNICS 2014

NEW MATERIALS
Original scientific paper
Ivan Stoevski, Jelena Krsti, Nikola Voki, Zorica Kaarevi Popovi, epan Miljani,
Radiolityc Synthesis and Characterization of PVA and Chitosan Based Conductive
Polymer Membrans for Alkaline Fuel Cells ................................................................................9

OUR CIVIL ENGINEERING

Original scientific papers
Nikola Toi, Sneana Marinkovi, Calibrating and Validating a FE Model for Long-term
Bahavior of RC Beams.................................................................................................................17
ikica Teki, Experimental Determination of Load Bearing Capacity of Connections
Realized by Punched Plate Fastener...........................................................................................25

MINING, GEOLOGY AND METALLURGY

Original scientific papers
Sreten Beatovi, Dinko Kneevi, Lazar Stojanovi, Slavko Torbica, ivko Sekuli, Influence of
Ash Slurry Preparation Method on Characteristics of Ash Disposed from the
Thermal Power Plant Gacko ...................................................................................................35
Violeta Stefanova, Todor Serafimovski, Goran Tasev, Morphological and Chemical
Characteristics of the Placer Gold Deposits from Meckin Dol, Locality of Borov Dol,
Eastern Macedonia.......................................................................................................................44

MECHANICAL ENGINEERING
Original scientific papers
Najdan Vukovi, Zoran Miljkovi, Machine Learning of Radial Basis Function Neural
Networks with Gaussian Processing Units Using Kalman Filtering - Introduction...............53
Najdan Vukovi, Zoran Miljkovi, Machine Learning of Radial Basis Function Neural Networks with
Gaussian Processing Units Using Kalman Filtering Experimental Results .........................61
ELECTRICAL ENGINEERING
Original scientific paper
Miljana Mili, Vano Litovski, Statistical Delay Estimation in Digital Circuits Using VHDL ............71
Review paper
Mirjana Simi, Predrag Pejovi, Deterministic Approach for Mobile User Positioning in Radio
Systems ..........................................................................................................................................79

TRAFFIC
Original scientific paper
Vladimir ori, Methodology for Collecting Instanteneous Vehicle Emissions on Street
Network .........................................................................................................................................89
Previous announcement
Nikola elar, Analitical Delay Models for Signalized Intersection........................................................97

MANAGEMENT
Original scientific papers
Aleksandar Vukovi, Marko Mihi, Dejan Petrovi, Strategic Management of Energy
Efficiency Project Portfolio in Public Buildings ......................................................................105
Darko Boani, Dragan Pamuar, Boban orovi, Modification of Analytic Hierarchy Process
(AHP) Method and its Application in the Defense Decision-making.....................................111

QUALITY IMS, STANDARDIZATION AND METROLOGY

Previous announcement
Radoje Jevti, Simulation of the Shopping Centre Zona I Evacuation............................................121
Professional paper
Goran Oraanin, Dragana Risti, Dalibor Vlaki, Branko Vuijak, Development of Methodology
for Evaluation and Real Water Loss Reduction in Water Supply Systems in
Transition Countries ..................................................................................................................126
 ISSN 0354-2300

TECHNICS

NEW
MATERIALS
Novi materijali Nouveaux matriaux
Neue Materialien
YEAR 23 2014.

EDITOR-IN-CHEF
Prof. Dragan U s k o k o v i , D.Sc, Institute of Technical
Sciences of SASA, Belgrade

EDITORIAL COUNCIL:
Prof. Vera D o n d u r, D.Sc, University of Belgrade,
Faculty of Physical Chemistry, Belgrade
Nenad I g nj a t o v i , D.Sc, Institute of Technical Sciences
of SASA, Belgrade
Slobodan M i l o nj i , D.Sc, Vina Institute of Nuclear
Sciences, Belgrade
Prof. Neboja M i t r o v i , D.Sc, University of
Kragujevac, Technical Faculty aak
Prof. Milenko P l a v i , D.Sc, University of Belgrade,
Faculty of Technology and Metallurgy, Belgrade
Prof. Zoran Po p o v i , D.Sc, Institute of Physics, Zemun
Prof. Dejan R a k o v i , D.Sc, University of Belgrade,
School of Electrical Engineering, Belgrade
Prof. Mira V u k e v i , D.Sc, University of Montenegro,
Faculty of Metallurgy of Technology, Podgorica

EDITORIAL OFFICE: Union of Engineers and Technicians

of Serbia, 11000 Belgrade, Kneza Miloa 7a/I, Tel. +381/11/
32 35 891, Fax +381/11/ 32 30 067
I. STOEVSKI at al. RADIOLYTIC SYNTHESIS AND CHARACTERIZATION OF PVA...

Radiolytic Synthesis and Characterization of PVA and Chitosan Based

Conductive Polymer Membranes for Alkaline Fuel Cells

IVAN. D. STOEVSKI, University of Belgrade, Original scientific paper

Faculty of Physical Chemistry, Belgrade UDC: 621.355.8:577.352
JELENA I. KRSTI, University of Belgrade,
Vina Institute of Nuclear Sciences,
Laboratory for Radiation Chemistry and Physics, Belgrade
NIKOLA . VOKI, University of Belgrade,
Faculty of Physical Chemistry, Belgrade
ZORICA M. KAAREVI POPOVI, University of Belgrade,
Vina Institute of Nuclear Sciences,
Laboratory for Radiation Chemistry and Physics, Belgrade
EPAN S. MILJANI, University of Belgrade,
Faculty of Physical Chemistry, Belgrade

Poly(vinyl alcohol) (PVA) and chitosan (CS) based polymer membranes for alkaline fuel cells were
synthesized by gamma irradiation method. They were swollen with 6 M KOH solution and their ionic
conductivity and gas permeance were investigated as a function of temperature. They show high ionic
conductivities at room temperature, which wasnt reduced over a period of few months. No gas flow
through membranes was detected at any temperature and pressure. These properties show that the
membranes could be potentially applied in alkaline fuel cells.
Key words: alkaline fuel cell, membrane, gamma radiation, poly(vinyl alcohol) (PVA), chitosan

1. INTRODUCTION line fuel cells have some disadvantages if a liquid

Hydrogen fuel cells are environmental friendly
electrochemical devices which convert chemical ener-
gy of hydrogen into electrical energy. Therefore, they
may have a significant role in development of sustai-
nable and ecological society. The most investigated
and also most efficient type of fuel cells today is a pro-
ton-exchange membrane fuel cell (PEMFC). However,
their widespread commercial application is limited by
high price [1], since they require use of noble metals
as catalysts, mostly platinum, and expensive polymer
membranes, like NafionTM. Alkaline fuel cells, on the
other hand, can operate with non-platinum catalysts,
because in alkaline environment, kinetics of oxygen
reduction reaction is enhanced, while the catalysts are
more stabile than in acid medium [2]. However, alka
Authors address: Ivan Stoevski, University of
Belgrade, Faculty of Physical Chemistry, Belgrade,
Studenstki trg 12-16
Paper received: 02.03.2014.
Paper accepted: 03.03.2014.
electrolyte is used.
The very first fuel cells used the liquid electrolyte,
in most cases aqueous solutions of potassium-hydro-
xide. Highly concentrated KOH solutions, besides ha-
ving corrosive effect, are sensitive to present of CO2
[3], which both result in degradation processes in
electrodes [4]. When air is fed to the cathode, hy-
droxide ions from the electrolyte can react with CO2
from air forming CO32- ions which can precipitate in
form of K2CO3. This decreasing of hydroxide ions
concentration leads to reducing of ionic conductivity
and increasing of activation loses on the anode, while
the precipitation itself can block pores within the gas-
diffusion layer. There are few approaches that can be
used to overcome CO2 contamination of the electro-
lyte, such as: absorption of CO2 from air, using liquid
hydrogen etc. [5, 6]. However, use of solid anion-
conducting polymer membranes instead of the liquid
electrolyte solves problems regarding CO2, electrolyte
leakage and corrosivity. At the moment, there are ma-
ny researchers in the world working on development
of these types of membranes [7-11].

TECHNICS NEW MATERIALS (2014) 9

I. STOEVSKI at al. RADIOLYTIC SYNTHESIS AND CHARACTERIZATION OF PVA...
Poly(vinyl alcohol) (PVA), because of its good
chemical stability and high hydrophilicity, and chi-
tosan (CS) due to its low cost, high abundance and also
high hydrophilicity, are materials which are commonly
used for polymer membranes for alkaline fuel cells. In
the most cases these polymers are chemically cross-
linked. Instead of chemical crosslinking there are other
crosslinking techniques that can be used, like gamma
irradiation method [12]. Gamma irradiation method
has few advantages over chemical crosslinking like:
high speed, no need for catalysts, very low byproducts
formation etc. [13].
The main goal of this paper is to obtain chemically
stable PVA and CS based polymer membrane, which
can be efficiently used in alkaline fuel cells and enable
use of inexpensive non-platinum catalysts. Their ionic
conductivity and gas separation capabilities were tes-
ted in conditions similar to those in operating the al-
kaline fuel cell. Also, influence of chitosan addition on
ionic conductivity is analyzed. Membranes were syn-
thesized using irradiation method, in which aqueous
solutions of PVA and CS were exposed to gamma rays
in order to form hydrogels. Obtained hydrogels were
immersed in potassium-hydroxide solutions.
2. EXPERIMENTAL PART
2.1. Materials
PVA with average molecular weight of 72000
g/mol and 99% of minimal degree of hydrolysis, was a
product of Merck. Chitosan with low molecular weight
and 75-85% degree of deacetylation, and potassium
hydroxide (KOH) were obtained from Sigma-Aldrich,
while acetic acid (CH3COOH) was a product of Zorka
Pharma. All chemicals were commercial products of
the analytical grade and were used without additional
purification. Water from Millipore Milli-Q system was
used in all experiments, while the high purity argon gas
(99.5%) from Messer Tehnogas was used for removing
the oxygen from solutions. The irradiations were per-
formed at room temperature using the 60Co gamma
source.
2.2 Synthesis of chitosan/PVA blended hydrogel
membranes
The synthesis of hydrogel membranes was perfor-
med by specific procedure that consists of three succe-
ssive steps: preparation of the polymer solutions, -
irradiation of the solutions and swelling of obtained
hydrogels in KOH solutions. PVA solution (5% w/w)
was obtained by the dissolution of PVA in distilled
water (90 C) for 6 h, with constant stirring. The
solution of chitosan (5% (w/w)) was prepared by
dissolving the chitosan in aqueous solution of
CH3COOH (6% (v/v)), at room temperature under the
10 TECHNICS NEW MATERIALS (2014)
constant stirring for 3 h. The mixtures of PVA and
chitosan polymers were stirred with different mass
ratios (90/10 and 95/5). All solutions were bubbled
with Ar for 30 min to remove oxygen before exposing
to -rays. Polymer solutions were poured into specially
designed molds consisted of two glass plates, which
were mutually separated by rubber spacers. The pure
PVA solution (5% (w/w)) was irradiated to the di-
fferent absorbed doses of 14 kGy and 25 kGy (at a dose
rate of 0.52 kGy/h) to induce polymer crosslinking
with a different mass yield of the hydrogel. After
synthesis, one part of the obtained PVA hydrogels was
left to swell in 6M KOH solution, while the rest were
dried to the constant weight to form xerogels.
2.3. Characterization of hydrogel membranes
The mass yield of the hydrogel gel fraction
(Gel%) was obtained by the following expression:
%= 100 (1)
where Wg is the weight of the xerogel after extraction
and W0 is the initial weight of the polymer. The
extraction of the uncrosslinked polymer was perfor-
med in duration of 7 days. The obtained xerogels (Wo)
were immersed into excess of deionized water, which
was changed periodically, after each 24 h. Than the
hydrogels were left to dry at room temperature, until
their masses have become constant (Wg).
Swelling degree of hydrogels in the equilibrium
state (EDS) was obtained by the following equation:
= (2)
where We is the weight of the swollen hydrogel in the
equilibrium state and Wg is the initial weight of the
xerogel.
Chemical stability of the hydrogel membranes was
investigated by monitoring the influence of the alka-
line environment on the color change, morphological
structure and ionic conductivity of the hydrogel mem-
branes. Surface morphology of the hydrogel membra-
nes was examined by scanning electron microscopy.
Before measurement on a JEOL JSM-6610LV in-
strument, the samples were dried for 72 h at the tem-
perature of 60 C, and then coated with a thin gold
layer.
Ionic conductivity measurements were carried out
by impedance spectroscopy performed on the pote-
ntiostat/galvanostat instrument (Model 273A) coupled
with a phase sensitive lock-in amplifier, Model 5210
(Princeton Applied Research) Princeton Applied Rese-
arch. Impedance spectra were recorded in the frequ-
ency range from 100-1 kHz with an amplitude po-
tential of 10 mV.
I. STOEVSKI at al. RADIOLYTIC SYNTHESIS AND CHARACTERIZATION OF PVA...

differential pressure between the two sides of the me-

Figure 1 - Cross-section of the cell for gas perme-
ability measurements
Resistance of the hydrogel membranes (R) which
represents its ionic resistance, was determined from the
intersection of the curve -Zim = f (Zre) to the axis of real
impedance, where the frequencies are approximately
100 kHz. Ionic conductivities of the hydrogel me-
mbrane, , are obtained from the equation 3:
(3)
where l is membrane thickness, and A is the electrode
surface. The electrodes made of stainless steel (with
surface area 0.785 cm2) and polyamide cell casing are
not commercial, but rather are specially made for this
purpose. Measurements were carried out in the tem-
perature range of 20-70 C in conditions of air relative
humidity of 100%. The cell prior to the series of
measurements is inserted with a sufficient amount of
water, so that at each temperature, after a certain time,
equilibrium was achieved between the two phases th-
rough establishing the saturation of the air with water
vapor. Prior to the measurements, membranes were
immersed into 6M KOH solution for 24 h.
The permeability of gases through the hydrogel
membrane was studied using a self-made stainless ste-
el cells (figure 1). The gas flow at the exit of the cell
was measured using a bubble flowmeter, at a given
Table 1. Gel fraction, equilibrium degree of swelling and ionic conductivity of the membranes (xerogels) at room
temperature
mbrane (input and output side of the cell). The gas
pressure at the input side (H2, N2, O2) is maintained
with the regulation valve to 4 bar, relative to output
(atmospheric) pressure. The relative humidity of the
gas was 100%, and the temperature of the membrane
was changed in the range of 20-80 C
3. RESULTS AND DISCUSSION
Irradiation of polymer solutions produces trans-
parent hydrogels that have ability to absorb large amo-
unts of water. After immersion of the hydrogels in
KOH solution for 24 hours, they became transparent-
white membranes with high ionic conductivity and
enhanced mechanical characteristics.
If the hydrogels were dried after preparation
(xerogels) and then immersed in KOH solution, their
thicknesses are 100 m. However, if water swollen
hydrogels were immersed in KOH solution, their
thicknesses were around 500 m. By applying pressure
on them, their thickness will be reduced permanently
from 500 m down to 100 m, without reducing their
ionic conductivity.
PVA crosslinking in aqueous solution by gamma
irradiation take place indirectly, via OH and H radi-
cals, which are formed by water radiolysis [14,15]. On
the other hand, gamma irradiation of Chitosan aqueous
solution causes Chitosan degradation and reduction of
its molar mass [16-18].
If PVA/Chitosan blend is irradiated, simultaneous
crosslinking of PVA, and molar mass reduction of CS
occurs, so obtained PVA/CS hydrogel will have less
gel fraction than pure PVA hydrogel of the same
concentration and absorbed dose [19].
We have also confirmed that addition of chitosan
to PVA reduces gel fraction of the obtained hydrogel
(table 1). From table 1 it can be seen that the gel
fraction of the PVA membranes are almost the same,
while it is reduced in PVA/CS membranes, especially
in 90/10 PVA/CS25kGy, which has larger fraction of CS.

Equilibrium degree of Ionic conductivity

Membrane Gel fraction (%)
swelling (S cm-1)

5%PVA14kGy 87 6.5 0.021

5%PVA25kGy 89 6.2 0.043

95/5 PVA/CS25kGy 79 7.9 0.17

90/10 PVA/CS25kGy 69 6 0.23

TECHNICS NEW MATERIALS (2014) 11

I. STOEVSKI at al. RADIOLYTIC SYNTHESIS AND CHARACTERIZATION OF PVA...
By comparison of equilibrium degree of swelling
(EDS) of membranes which have been dried, xerogels
(table 1), and those which havent (table 2), it can be
seen that drying considerably decreases membranes
ability to absorb liquids.
Peppas and Merrill [20, 21] have shown that dehy-
dration and heating of PVA hydrogel increase fraction
of crystalline phase in hydrogel, which further reduces
its absorption ability. This structural modification of
membranes, drastically affects their ionic conductivity.
For example, ionic conductivity of 5%PVA25kGy mem-
brane, which was dried before it was immersed in
KOH solution (table 1), is for one order of magnitude
lower than the non-dried one (table 2).
Table 2. Equilibrium degree of swelling and ionic con-
ductivity of the membranes (non-dried) at room
temperature
Equilibrium Ionic conductivity
Membrane
degree of swelling (S cm-1)
5%PVA25kGy 19.2 0.17
90/10
31.0 0.33
PVA/CS25kGy
The reasons for such a significant increase of the
dried membrane resistance are just mentioned increase
of the crystalline phase and reduced absorption ability.
Higher content of the crystalline phase in a membrane
is related with lower mobility of polymer chains [22],
while the poorer absorption of KOH solution decreases
the number of OH- ions, responsible for charge tra-
nsfer. The membrane with addition of CS, 90/10
PVA/CS25kGy, has the highest ionic conductivity, no
Figure 2 - Temperature dependence of ionic conductivity of the membranes.
12 TECHNICS NEW MATERIALS (2014)
matter it was dried or no. The reason for this can be a
lower gel fraction in comparison to other membranes,
which contributes to higher mobility of polymer chains
[23] and formation of larger pores and voids that
enhances electrolyte mobility.
Electrical ionic resistance through the membranes
5%PVA25kGy and 90/10 PVA/CS25kGy was measured in
function of the temperature and by means of equation
(3) transformed in ionic conductivity. Figure 2 shows
that ionic conductivity linearly raises with the tem-
perature. The mechanism of the ionic transport in poly-
meric electrolytes isnt completely clear, but generally,
the increase of ionic conductivity with temperature is
due to local structural relaxation, and segmental mo-
tions of the polymer [24]. The ionic conductivity of
non-dried membrane at room temperature is 0,33 S cm-
1, while at the temperature of 75 oC is 0,44 S cm-1. This
has shown that membranes have high ionic condu-
ctivity that is stable at elevated temperatures, which is
desirable for application in fuel cells.
Chemical stability of the membranes was moni-
tored visually through color changes and through ionic
conductivity changes over a period of time. For this
purpose, the non-dried 5%PVA25kGy membrane was
immersed in 6 M KOH solution. First 3 months it was
transparent-white, after which it changed color to
yellow. Ionic conductivity of the membrane after 3
months remained same as it was after 24 hours. This
indicates that the membranes have stabile ionic con-
ductivity in alkaline medium, which is very important,
since many anion-exchange membranes have proble-
ms with stability in highly alkaline environment.
I. STOEVSKI at al. RADIOLYTIC SYNTHESIS AND CHARACTERIZATION OF PVA...
However, it was noticed that membranes that were
swollen with KOH solution and left few days on
ambient conditions (25 oC and relative humidity of air,
60-70 %) also slightly changed their color to yellow. It
is supposed that drying causes color changes, and to
confirm that, 5%PVA25kGy membrane was immersed in
KOH solution for 24 hours and then dried for 48 hours
at the temperature of 60oC. After drying its color be-
came dark-brown, while the ionic conductivity was
reduced from 0.17 S cm-1 to 0.0035 S cm-1. By SEM
analysis of membranes surface it was noticed damage
of the gel structure (figure 3.b), as well as formation of
KOH crystals (figure 3.c).
On the other hand, surface of the 5%PVA25kGy
membrane, which wasnt immersed in KOH solution
at all, looks relatively smooth, with the presence of
blisters, but without any damage (figure 3a). Water
evaporation from the membrane that was previously
immersed in the KOH solution increases concentration
of OH- ions, which causes damage on polymer chains
and change of color. Further evaporation of water leads
to crystal formation and complete drying of membrane.
Mobility of OH- ions in crystals is much lower than in
Figure 3 - SEM images of the surface of: a) dried, non-doped 5%PVA25kGy; b) and c) dried, KOH doped
5%PVA25kGy membrane.
Gas separation capabilities of the membranes was
investigated under conditions similar to those in fuel
cells, i.e. temperature was raised up to 80 oC and
relative humidity of the gaseous were 100%. Gas flow
through the membranes wasnt detected even at the
pressure of 4 bar. A minimal flow rate that can be
accurately detected by the measuring system is 0.01
TECHNICS NEW MATERIALS (2014)
the liquid ele and this is the main reason for the rise of
membranes resistance.
It was noticed that the electrolyte leaks out of the
membrane when it is immersed in water. Namely, 5%-
PVA25kGy membrane was immersed in water for 24
hours, which has increased pH of water up to 12 and
reduced ionic conductivity of the membrane from 0.17
S cm-1 to 0.018 S cm-1. This indicates that a part of the
electrolyte diffuses from the membrane to water,
because of weaker interaction with the polymer matrix
than with the other part, which remains inside the
matrix even if the membrane was soaked with water.
Presence of the free electrolyte inside the mem-
branes is responsible for their high ionic conductivity,
as well as for easy evaporation of water, which
degrades their structure.
Anion-exchange membranes dont have problems
with electrolyte leakage, since they have functional
groups for OH- ion exchange that are strongly bonded
to the matrix [25]. Our future research will address the
problem of electrolyte leakage, by incorporating
different functional groups inside the PVA matrix.
cm3 min-1, which means that gas permeability of the
membranes is lower than 0.12 cm3 m m-2 h-1 bar-1, in
conditions similar to those in fuel cells.
4. CONCLUSIONS
Hydrogel membranes for the alkaline fuel cell ba-
sed on poly(vinyl alcohol), chitosan and KOH have
13
I. STOEVSKI at al. RADIOLYTIC SYNTHESIS AND CHARACTERIZATION OF PVA...

been synthesized via gamma crosslinking method. It
was shown that they have high ionic conductivities
thanks to ability to absorb a great amount of KOH
solution. Also, they have shown good tolerance in al-
kaline medium and their conductivity remained un-
changed for several months. Gas separation capabi-
lities of the membranes were shown to be very good
even at the temperature of 80 oC and differential pre-
ssure of 4 bar. Main disadvantages of these membranes
are electrolyte leakage and easy dehydration. In our
future research we will address these issues by inco-
rporating functional groups in the PVA matrix. Testing
the membranes in real conditions, in fuel cells will also
be one of our future goals.
5. ACKNOWLEDGEMENTS
This work was financially supported by the Mini-
stry of Education, Science and Technological Deve-
lopment of the Republic of Serbia through projects III-
45014 and III-45005.
REFERENCES
[1] Wu J., Yuan X. Z., Martin J. J., Wang H., Zhang J.,
Shen J., Wu S., Merida W., J. Power Sources 184,
104-119, 2008.
[2] Li, X., Liu G., Popov N. P., J. Power Sources 195,
6373-6378, 2010.
[3] Tewari, A., Sambhy V., Urquidi Macdonald M., Sen
A., J. Power Sources 153, 1-10, 2006.
[4] Schulze M., Glzow E., J. Power Sources 127, 252-
263, 2004.
[5] McLean G. F., Niet T., Prince-Richard S., Djilali N.,
Int. J. Hydrogen Energy 27, 507-526, 2002.
[6] Ni M., Leung K. H. M., Leung Y. C. D., WHEC 16,
Lyon, 2006.
[7] Merle G., Wessling M., Nijmeijer K., J. Membr. Sci.,
377, 1-35, 2011.
[8] Wang Y. J., Qiao J., Baker R., Zhang J., Chem. Soc.
Rev. 42, 5768-5787, 2013.
[9] Couture G., Alaaeddine A., Boschet F., Ameduri B.,
Prog. Polym. Sci., 36, 1521-1557, 2011.
[10]Zhang H., Shen K. P., Chem. rev. 112, 2780-2832,
2012.
[11]Varcoe, R. J., Slade C. T. R., Fuel cells 5, 187-200,
2005.
[12]Nikolic M. V. Krkljes A., Kacarevic Popovic Z.,
Lausevic V. Z., Miljanic S. S., Electrochem.
Commun., 9, 2661-2665, 2007.
[13]Bhattacharya A., Prog. Polym. Sci., 25, 371-401,
2000.
[14]Rosiak J. M., Ulaski P., Radiat. Phys. Chem., 55,
139-151, 1999.
[15]Wang B., Mukataka S., Kokufuta E., Kodama M.,
Radiat. Phys. Chem., 59, 91-95, 2000.
[16]Gryczka U., Dondi D., Chmielewski A., Migdal W.,
Butaffava A., Faucitano A., Radiat. Phys. Chem., 78,
543-548, 2009.
[17]Ulanski P., Sonntag C., J. Chem. Soc., Perkin Trans.
2 10, 2022-2028, 2000.
[18]Kang B., Dai Y. D., Zhang H. Q., Chen D., Polym.
Degrad. Stab. 92, 359-362, 2007.
[19]Tahtat D., Mahlous M., Benamer S., Khodja N. A.,
Youcef L. S., Hadjarab N., Mezaache W., J Mater
Sci: Mater Med, 22, 2505-2512, 2011.
[20]Peppas N. A., Merrill W. E., J. Polym. Sci Part A:
Polym. Chem. 14, 441-457, 1976.
[21]Peppas N. A., Merrill W. E., J. Appl. Polym. Sci., 20,
1457-1465, 1976.
[22]Yang, C. C., Lin S. J., Hsu S. T., J. Power Sources
122, 210-218, 2003.
[23]Zhou J., Unlu M., Anestis-Richard I., Kohl A. P., J.
Membr. Sci., 350, 286-292, 2010.
[24]Yang C. C., Lin S. J., Mater. Lett.57, 873-881, 2002.
[25]Merle G., Hosseiny S. S., Wessling M., Nijmeijer K.,
J. Membr. Sci., 409, 191-199, 2012

REZIME
RADIOLITIKA SINTEZA I KARAKTERIZACIJA PROVODNIH POLIMERNIH
MEMBRANA ZA ALKALNE GORIVNE ELIJE, NA BAZI PVA I HITOZANA

Sintetisane su polimerne membrane za alkalne gorivne elije na bazi poli(vinil alkohola) i hitozana, uz
korienje gama-zraenja za njihovo umreavanje. Nakon bubrenja u 6 M rastvoru kalijum-hidroksida,
ispitivani su jonska provodljivost i propustljivost na gasove dobijenih membrana, u funkciji temperature.
Utvreno je da poseduju visoku jonsku provodljivost u irokom opsegu temperatura, koja se nije menjala
tokom nekoliko meseci. Na svim temperaturama i pritiscima na kojima su vreni eksperimenti, nije
zapaeno da su membrane propustljive za gasove. Ova svojstva ukazuju da membrane mogu imati
potencijalnu primenu u alkalnim gorivnim elijama.
Kljune rei: alkalna gorivna elija, membrana, gama-zraenje, poli(vinilalkohol) (PVA), hitozan

14 TECHNICS NEW MATERIALS (2014)

 ISSN 0350-2619
TECHNICS

OUR CIVIL
ENGINEERING
Our Civil Engineering - Notre gnie civil
- Unser Bauwesen

YEAR 68 - 2014

EDITOR-IN-CHEEF:
Prof. Branko B o i , D.Sc University of Belgrade, Faculty
of Civil Engineering, Belgrade

EDITORIAL COUNCIL:

Prof. Radomir F o l i , D.Sc University of Novi Sad,

Faculty of Technical Sciences, Novi Sad
Prof. Nenad I v a n i e v i , D.Sc, University of
Belgrade, Faculty of Civil Engineering, Belgrade
Prof. Radomir K a p o r, D.Sc, University of Belgrade,
Faculty of Civil Engineering, Belgrade
Prof. Milo K n e e v i , D.Sc, University of Montenegro,
Faculty of Civil Engineering, Podgorica
Docent Goran M l a d e n o v i , D.Sc, University of
Belgrade, Faculty of Civil Engineering, Belgrade
Docent Ivan I g n j a t o v i , D.Sc, University of Belgrade,
Faculty of Civil Engineering, Belgrade
Docent Budimir S u d i m a c, M.Sc, University of
Belgrade, Faculty of Architecture, Belgrade
Docent Stevan R a d o j i , D.Sc, Military Geographical
Institute, Belgrade
Marijan e h, D.Sc University of Ljubljana, Faculty of
Civil and Geodetic Engineering, Ljubljana
Assistant Professor Tomas H a n a k, Brno University of
Technology, Faculty of Civil Engineering, Institute
of Structural Economics and Management, Brno

REDAKCIJA I ADMINISTRACIJA: Savez inenjera i

tehniara Srbije, 11000 Beograd, Kneza Miloa 7a/I, Telefon
(011) 32 35 891, Fax (011) 32 30 067
N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

Calibrating and Validating a FE Model for Long-term Behavior of RC Beams

NIKOLA D. TOI, University of Belgrade, Original scientific paper

Faculty of Civil Engineering, Belgrade UDC: 624.012.45.042
SNEANA B. MARINKOVI, University of Belgrade,
Faculty of Civil Engineering, Belgrade

This study presents the research carried out in finding an optimal finite element (FE) model for
calculating the long-term behavior of reinforced concrete (RC) beams. A multi-purpose finite element
software DIANA was used. A benchmark test in the form of a simply supported beam loaded in four-
point bending was selected for model calibration. The result was the choice of 3-node beam elements, a
multi-directional fixed crack model with constant stress cut-off, nonlinear tension softening and constant
shear retention and a creep and shrinkage model according to CEB-FIP Model Code 1990. The model
was then validated on 14 simply supported beams and 6 continuous beams. Good agreement was found
with experimental results (within 15%).
Key words: creep, shrinkage, reinforced concrete, beams, FE

1. INTRODUCTION 2. FE MODEL SELECTION AND CALIBRATION

Non-linear computer analysis methods have seen
remarkable advancement in the last half-century with
a lot of research activity in the field of constitutive
modeling of reinforced concrete behavior and in the
development of sophisticated analysis algorithms [1].
The state-of-the-art in non-linear finite element
analysis of reinforced concrete has progressed to the
point where such procedures are close to being pra-
ctical, every-day tools for design office engineers.
At the same time the long-term behavior of con-
crete i.e. creep and shrinkage, remains one of the most
hotly debated areas in the field of constitutive mo-
deling of concrete. There is no uniform view on this
phenomenon and a vast variety of models exist for
describing the rheological properties of concrete.
The aim of the present study is to find a relatively
simple and robust material and numerical model that
can be understandable to the practicing engineer. For
this purpose a benchmark test was selected and a FE
model calibrated. The model was then validated on
other experiments with good agreement while drawing
significant conclusions.
Authors address: Nikola Toi, University of Bel-
grade, Faculty of Civil Engineering, Belgrade, Bulevar
kralja Aleksandra 73
Paper received: 24.03.2014.
Paper accepted: 25.04.2014.
2.1 TNO DIANA
DIANA is a general purpose finite element code,
based on the Displacement Method (DIANA=DIs-
placement ANAlyser) [2]. It has been under deve-
lopment at TNO DIANA BV since 1972. Since then it
was proven and tested to be an extensive multi-purpose
finite element software package dedicated, but not
exclusive to a wide range of problems arising in civil
engineering including structural, geotechnical,
tunneling, earthquake disciplines and oil and gas en-
gineering. The program includes extensive material
models, element libraries and analysis procedures,
which are based on the latest and the most advanced
finite element analysis techniques. DIANA has been
equipped with powerful solvers in order to optimize
the solution procedures for all types of linear and no-
nlinear complex models with accurate results and fast
computations. Its more than abundant offer of concrete
material models and model code libraries, finite
elements and nonlinear static and transient analysis
procedures was the reason for choosing it in this study.
2.2 Experiment used as creep and shrinkage
benchmark test
DIANA Users Manual Concrete and Masonry
Analysis (ConcMas) [3] elaborates numerous exam-
ples of analysis of concrete and masonry structures.
Part IV Chapter 13 presents an example of long-term
behavior of a reinforced concrete beam. The example
actually models the experiment carried out by Jaccoud

TECHNICS OUR CIVIL ENGINEERING (2014) 17

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

and Favre from 1979 through 1982. [4]. In this well- was analyzed since it was cracked and is a more
known experiment various reinforced simply su- realistic representation of real-life structures.
pported concrete one-way slabs were loaded with 2.3. Initial FE model description
different load intensities and their long-term behavior
In the adopted initial FE model only one half of the
was monitored. The quality of this research is its
slab was modeled, because of symmetry. This half was
scope and thoroughness. One-way slabs were executed
modeled with seven three-node CL9BE beam
well in terms of achieved material properties,
elements, 200 mm length each [3]. CL9BE is a nume-
reinforcement ratio and span-to-depth ratio. All rele-
rically integrated two-dimensional element in which
vant material properties (compressive and tensile stre-
shear deformation is included according to the Min-
ngth, modulus of elasticity) were measured.
dlin-Reissner theory.
The most numerous experimental series series C,
In the axial direction, a default 2-point Gauss
consisted of 11 slabs with the same reinforcement ratio
integration is used and in the thickness direction a 9-
loaded in four-point bending with different loads,
point Simpsonian integration. Due to the two-
ranging from 0.2 to 0.6 of the slabs ultimate load
dimensional idealization, all reinforcements at one side
(UL). There were slabs that remained uncracked as
were modeled by one embedded bar.
well as those which cracked under the loading. The
slabs were loaded at the age of 28 days, sufficiently Two load sets were modeled load set 1 was a
quickly and most importantly, the initial deflection was nodal force of 15.725 kN applied at the third of the span
measured quickly (5 minutes after loading). Deflection and load set 2 was dead weight loading.
was measured for one year. Because of these qualities The necessary material properties were standard
in 1993 the RILEM committee TC 114 adopted the C- elastic properties, creep, shrinkage and cracking pro-
series as benchmark tests for computer software for perties.
creep and shrinkage [5], [6]. Table 1 lists the most The elastic properties are the concrete modulus of
important data for slab_C15. elasticity, Poissons ratio and the reinforcement steel
ConcMas [3] presents the numerical analysis for modulus of elasticity. A divergence was made here
slabs C11 and C15 (slab C11 is loaded with 0.2UL from the example in ConcMas [3]. The material pro-
and slab C15 with 0.6UL). In this manual a material perties used in the Users Manual were average pro-
and FE model and analysis procedure, deemed most fit perties for slabs C11 and C15 whereas in this study the
by DIANA developers were selected and results exact properties for slab C15 were used, Table 1.
presented. However, in this paper only the slab C15
Table 1. Experimental data from [4]
Span Width Height Aa1b Aa2c RH fcm,28d
Experiment Type
(mm) (mm) (mm) (mm2) (mm2) (%) (MPa)
J&F C15 4PBa 3100 750 160 565 57 60 29.29
a
four-point bending b
tension reinforcement
c
compression reinforcement d
compressive strength after 28 days
Ecm,28f tloading tfinal Mcrg Mswh M Fi w(i)
fct,28e (MPa) w(t) (mm)
(GPa) (days) (days) (kNm) (kNm) (kNm) (mm)
2.90 28.70 28 393 9.280 3.54 15.725 10.0 19.50
e
tensile strength after 28 days f elasticity modulus after 28 days
g
cracking moment h
moment due to self-weight i
moment due to applied load

For cracking, a Multi-Directional Fixed Crack
model was used. Besides this model, DIANA also
offers a Total Strain Crack model. The advantages of
the chosen model in similar applications is thoroughly
elaborated in literature [7], [8], [9]. This model re-
quires the definition of stress cut-off, tension softening
and shear retention.
For the initial model constant stress cut-off, nonli-
near tension-softening according to Moelands and Re-
inhardt [10] and constant shear retention were chosen.
The additional input parameters are the tensile strength
ft, fracture energy Gf, crack bandwidth hcr and shear
retention factor . The fracture energy is considered a
material property and was calculated according to the
CEB-FIP Model Code 1990 [11]. The crack bandwidth
is dependent on the element size (equals the length of
the beam elements; in reality can be estimated from
stirrup spacing). These three parameters - ft, Gf and hcr
determine the release of Mode-I fracture energy if the
tensile strength is violated. It is possible that the
elements of the discretization are so large that the
equivalent length of an element results in a snap-back

18 TECHNICS OUR CIVIL ENGINEERING (2014)

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

in the constitutive model and the concept of objective of 365 days. Since there is no cracking under the dead
fracture energy which has been assumed is no longer weight load it was applied in one step with no step size
satisfied. A snap-back in the constitutive model is control. For the point load automatic load step control
possible if the absolute value of the initial slope of the is used. This is done using an Iteration Based Adaptive
softening diagram is greater than the Youngs modulus Loading method combined with a Spherical Path Arc-
of the material, if it is assumed that the initial tangent length method to control for snap-back behavior [13].
of the tension softening diagram results in the greatest Several parameters such as maximum, minimum and
value of the tangent stiffness. This criterion is always initial step size, number of load steps, number of
violated in the case of tension softening according to maximum iterations and convergence criteria are
Moelands and Reinhardt [12]. The shear retention defined. Creep and shrinkage is analyzed using time
factor was in this example chosen rather arbitrarily and steps in days. The time stepping scheme is composed
there is no discussion on this choice. A measure of by incrementing the time steps logarithmically [3]. The
aggregate interlock, shear retention across a crack is selected output data are internal forces and moments,
best described by a descending linear function of the global displacements, concrete and reinforcement
crack strain. Since DIANA doesnt offer such an stresses and crack strains. Results are presented in
option, a constant value of 0.2 chosen in this example Table 3.
is accepted as appropriate. Table 3. Results of FE analysis compared to measured
Creep is modeled by rheological Maxwell chains, values
consisting of several spring and dashpot chains [3]. Deflection (mm)
Time FE/Exp.
DIANA offers the option of deriving the chain pro- Exp.
(days) (%)
FE Experiment
perties from standard creep models like the CEB-FIP
Model Code 1990 [11]. The example in ConcMas [3] J&F 0 13.2 10.0 132
C15 365 21.4 19.5 109.7
uses a built-in CEB-FIP Model Code creep model
which requires only parameters such as the loading age As can be seen the calculated deflections at slabs
t0, modulus of elasticity Et0, mean compressive midspan fit fairly good with the measured values.
strength fcm, notational size h0 (= 2A/O), cement type, There is a larger discrepancy in the initial deflection,
relative humidity RH and ambient temperature Tenv. which can only partially be explained by uncertainties
From these parameters the creep function J(t,t0) is tied to the tensile strength and fracture energy. Still the
calculated. This model is only valid within the range of results are within 30% which can be considered a limit
linear creep i.e. when c 0.4fck. for satisfactory models [14]. The results are more
Shrinkage is modeled as age-dependent initial st- accurate for the long-term deflection which follows the
rain. The built-in CEB-FIP model was used [11]. The experimental curve rather well and ends within 10% of
required parameters are the same as for the creep mo- the measured value, enough to deem the model good
del. Shrinkage strain starts at the end of the curing [14]. However, the fact that the difference is reduced
period, in this case 7 days. from 30% to 10% during the time analysis can indicate
that the creep and shrinkage models underestimate
The input parameters are shown in Table 2.
deflection.
Table 2. Input parameters for the initial FE model
2.4. Model calibration
Reinforcement Ea 210 kN/mm2
modulus The next step in this study involved finding
Concrete modulus Ec 28.7 kN/mm2
whether it was possible to find a more accurate model
for this experiment. The varied model parameters
Tensile strength ft 2.9 N/mm2
were:
Loading age t0 28 days
Finite element type
Relative humidity RH 60 %
Tension softening relation
Ambient temperature Tenv 20 C
Creep and shrinkage model
Notational member h0 132 mm
size For finite element types, two different beam ele-
Crack bandwidth hcr 200 mm ments were used CL12B and CL15B with 4 and 5
nodes, respectively. Element length was kept constant
Fracture energy Gf 0.064 Nmm/mm2
at 200 mm. In the axial direction a default 3-point and
Shear retention factor 0.2
4-point Gauss integration is used respectively while in
For the analysis procedure the following was the thickness direction a 9-point Simpsonian in-
adopted: first the dead weight load is applied, then the tegration is kept same as before. The use of plane stress
point load and finally a transient analysis is carried out, and solid elements was also investigated. A qu-
calculating the creep and shrinkage effects for a period adrilateral, 8 node plane stress element, CQ16M was

TECHNICS OUR CIVIL ENGINEERING (2014) 19

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

used with a default 2x2 Gauss integration scheme. constant stress cut-off, constant shear retention and
Element dimensions were 50x40 mm. A brick, 8 node creep and shrinkage according to CEB-FIP Model
solid element, HX24L was also used with a default Code 1990 is the most accurate model.
2x2x2 Gauss integration scheme. Element dimensions
Table 4. Results of the FE parametric study
were 75x80x75 mm.
For tension softening relations, DIANA offers 5 Varied w(i) FE/Exp. w(t) FE/Ex
models brittle, linear, multilinear, nonlinear acco- parameter (mm) (%) (mm) p. (%)
rding to Moelands and Reinhardt [10] and nonlinear Initial model 13.2 132 21.4 109.7
according to Hordijk et al. [15]. Multilinear tension CL12B element 13.2 132 21.5 110.3
softening was not considered as it is simply a crude CL15B element 13.2 132 21.4 109.7
transition between linear and nonlinear relations. In CQ16M 112.8
linear and nonlinear (Hordijk et al.) tension softening 13.0 130 22
element
it was checked whether the absolute value of the initial HX24L element 13.9 139 22.4 114.9
slope of the softening diagram is greater than the
Brittle cracking 13.3 133 21.4 109.8
Youngs modulus of the material.
Linear tension 109.8
This criterion can be written as: soft.
13.3 133 21.4

cr f t dy
nn ,ult |x=0 (1)
E dx
where nc is the ultimate crack strain.
For the selected input parameters it was found that
the criterion was violated in the case of nonlinear
tension softening according to Hordijk et al. The re-
commended action in that case is the reduction of the
tensile strength [12] in accordance with:
1 subject of FE model validation [1]. In this process the
l 2
Gf
f t 0.739 (2)
h
The choice of reducing the tensile strength has
physical meaning since the variations in tensile stre-
ngth can be relatively large and can increase with in-
creasing sampling size (i.e. in larger finite elements).
In this way the tensile strength was reduced from 2.90
MPa to 2.61 MPa.
For modeling creep and shrinkage by using model
codes, besides the CEB-FIP Model Code 1990, DIA-
NA offers the possibility of using the American ACI
209R-92 model code [16] and the Dutch NEN 6720
model code [17]. DIANAs built-in module was used,
where the user only needs to specify the required input
parameters without further calculation.
The analysis procedure was kept constant. In total
this gave 9 parametric variations. The results are pre-
sented in Table 4. The results show that all the para-
meter variations offer little improvement. It has to be
noted that since only one parameter was varied in each
model there is no information on possible parameter
interactions and possible synergetic effects. However
it was found that, within this kind of parametric study,
a choice of nonlinear tension softening according to
Hordijk et al. together with CL9BE beam elements,
Nonlinear tens.
soft. Hordijk et 13.0 130 20.1 103.1
al.
ACI 209R-92 13.2 132 17.6 90.3
NEN 6720 13.2 132 21.3 109.2
3. FE MODEL VALIDATION
3.1. Model validation on simply supported beams
In various literature guidance can be found on the
most important step is the validation of the proposed
model on a set of carefully chosen experiments of the
same type. In this case that includes simply supported
beams loaded in four-point bending or by a uniformly
distributed load (UDL). The number of such experi-
ments properly and carefully executed isnt large -in
1988. Espion found only 29 research programs of
sufficient quality [18]. In this work 13 different models
from 3 research programs were validated. The chosen
experiments were Washa and Fluck, 1952 [19],
Glanville and Thomas, 1939 [20] and Portland Cement
Association, 1950 [21].
The research performed by Washa and Fluck [19]
is practically the first systematical and methodical
research of long-term behavior of RC beams [14]. Be-
ams of different cross sections and different amounts
of compressive reinforcement were loaded by a uni-
formly distributed load after 14 days and deflections
were measured for 21/2 years (913days). Among the
necessary parameters, relative humidity was not con-
trolled but in [18] a value of 50% is recommended. The
tensile strength of concrete is not reported. The 14-day
tensile strength was calculated according to CEB-FIP
Model Code 1990. In the analysis procedure, step size
control was used for dead weight load as it is
practically identical to the cracking load as well as for
the uniformly distributed additional load.

20 TECHNICS OUR CIVIL ENGINEERING (2014)

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

Glanville and Thomas [20] carried out the first re- Skokie, Illinois were never published but are me-
search of long-term behavior of RC beams under co- ntioned by various researchers [18]. Two beams with
ntrolled environmental conditions (RH=65%). Two different reinforcement ratios are considered.
rather heavily reinforced beams are considered here, The beams are loaded at 28 days by different loads
loaded at 24 and 25 days in four-point bending for 200 and deflections are monitored over a period of 270
days. Unfortunately data had to be taken from [18] as days. The values of the modulus of elasticity and con-
the original was not available to the authors. The crete tensile strength arent presented and had to be ca-
values of the modulus of elasticity and concrete tensile lculated according to CEB-FIP Model Code 1990. The
strength arent presented and had to be calculated analysis procedure is identical to the one in section 2.3
according to CEB-FIP Model Code 1990. i.e. dead weight load is applied in one step while step
The analysis procedure is identical to the one in se- size control is used for the point load.
ction 2.3 i.e. dead weight load is applied in one step The experimental data is presented it Table 5 and
while step size control is used for the point load. The the FE model data in Table 6. The results of the nu-
results of the research by the Portland Cement merical analyses can be seen in Table 7.
Association [21] performed in the laboratories in
Table 5. Experimental data from [19], [20] and [21]
Aa1 Aa2 fcm,28 w(i) w(t)
Experiment L (mm) b (mm) d (mm)
(mm2) (mm2) (MPa) (mm) (mm)
W&F A1/A4 6096 203.2 304.8 852 852 28.1 13.46 23.62

W&F A2/A5 6096 203.2 304.8 852 400 28.1 15.75 32.26

W&F A3/A6 6096 203.2 304.8 852 0 28.1 17.02 44.70

W&F B1/B4 6096 152.4 203.2 400 400 23.58 23.37 51.05

W&F B2/B5 6096 152.4 203.2 400 200 23.58 24.89 65.02

W&F B3/B6 6096 152.4 203.2 400 0 23.58 26.42 86.36

W&F D1/D4 3810 304.8 127 516 516 25.11 11.94 27.69

W&F D2/D5 3810 304.8 127 516 258 25.11 14.22 33.78

W&F D3/D6 3810 304.8 127 516 0 25.11 17.78 48.51

G&T X49 1830 101.6 209.6 400 0 24.0 1.42 3.48

G&T 88D 1830 101.6 209.6 400 0 28.1 1.57 3.63

PCA 40NA 3048 152 305 849 0 26.9 4.25 10

PCA 60NA 3048 152 305 1019 0 37.4 4.9 9.9

Table 6. FE model data for [19], [20] and [21]

Exp. Elem. length Ec (kN/mm2) ft (N/mm2) RH (%) Gf (Nmm/mm2)

W&F A-ser. 381 20.37 2.01 50 0.058

W&F B-ser. 254 18.75 1.70 50 0.051

W&F D-ser. 210 18.40 1.80 50 0.053

G&T X49 230 28.62 1.88 65 0.055

G&T 88D 230 30.21 2.20 65 0.061

PCA 40NA 250 29.90 2.14 50 0.06

PCA 60NA 250 33.37 2.87 50 0.076

TECHNICS OUR CIVIL ENGINEERING (2014) 21

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

Table 7. Results of FE analyses vs. measured values, shrinkage model underestimates the long-term defle-
[19], [20] and [21] ction development (although in Washa and Flucks
w(i) FE/Exp. w(t) FE/Ex study relative humidity wasnt controlled; the adopted
Exp.
(mm) (%) (mm) p. (%) value of 50% is assumed to be an average).
W&F A1/A4 15.5 115.2 24.3 102.9 3.2. Model validation on continuous beams
W&F A2/A5 16.4 104.1 30.5 94.5
In [1], among the guidelines given for proper FE
W&F A3/A6 17.3 101.7 43.7 97.8
model validation, it is stated that a model should be
W&F B1/B4 27.1 116.0 48.8 95.6
tested in several different applications, not only in the
W&F B2/B5 27.7 111.3 58.1 89.4
application for which it was calibrated. In this study a
W&F B3/B6 28.5 107.9 80.4 93.1
choice was made to further validate the model on
W&F D1/D4 15.8 132.3 26.9 97.1
continuous beams. There exists a small number of
W&F D2/D5 16.4 115.3 32.8 97.1
good research programs in this area [18]. Continuous
W&F D3/D6 17.1 96.2 47.9 98.7
beams are also significant because of a new factor that
G&T X49 1.66 116.9 3.65 104.9
influences their long-term behavior - moment redi-
G&T 88D 1.63 103.8 3.42 94.2
stribution due to cracking. The aim of the final step in
PCA 40NA 4.09 96.2 8.63 86.3
this study was to test whether the proposed model
PCA 60NA 4.74 96.7 8.94 90.3
captures these influences.
= 108.7 = 95.5
One research was selected for this phase an ex-
= 10.5 = 5.2
periment carried out by Washa and Fluck in 1956 [22]
Very good agreement with measured values is after the one in 1952 [19]. Together these two
achieved with all results, except one, falling within experiments were conceived in such a way that they
15% of the experimental values. There is one dra- can be interpreted together. The two-span beams were
wback that significantly diminishes the quality of any of the same cross-sections and spans and long-term
FE analysis prediction. In Washa and Flucks study no behavior was monitored for 21/2 years (910 days). In
tensile strength is reported while in the other two this study series X and Y were selected which co-
studies neither the tensile strength nor the elasticity rrespond to series B and D in [19].
modulus are given. Having in mind this drawback, so- Compressive reinforcement was again varied from
me trends can be outlined. As in the benchmark test, being equal to the tension reinforcement, half of the
accuracy is lower in the case of initial deflection. This tension reinforcement and zero. The load applied at the
can in part be explained by the unreported material age of 28 days was selected in such a way that the
properties and in the other part by the inadequacy of maximum bending moment in the span is equal to the
the FE model itself. In case of the initial deflection corresponding moment of the simply supported beams
there is a clear trend of increasing accuracy with de- in [19]. Again the tensile strength of concrete isnt
creasing compressive reinforcement. reported and had to be calculated according to CEB-
It would seem that the adopted Multi-Directional FIP. The FE model data is presented in Table 8 and the
Fixed Crack model is intrinsically less stiff than real- experimental data in Table 9. Numerical analysis re-
life structures and the presence of compressive rein- sults are presented in Table 10.
forcement only further softens it i.e. there is a re- Table 8. FE model data for experiment [22]
duction of concrete stiffness due to compressive
El.
reinforcement presence, but it isnt adequately acco- Exp. length
Ec,28 ft,28 RH Gf
unted for in the overall stiffness. As for the long-term (kN/mm2) (N/mm2) (%) Nmm/mm2
(mm)
deflection, accuracy seems to be better. However X-ser. 254 23.24 2.02 0 0.058
taking into account the overestimation of initial Y-ser. 210 23.27 2.19 0 0.061
deflection, it seems that the CEB-FIP creep and
Table 9. Experimental data from [22]
Span Mid-support
fcm,28 w(i)
Exp. L (mm) b (mm) d (mm) w(t) (mm)
Aa1 (mm )
2
Aa2 (mm )
2
Aa1 (mm2) Aa2 (mm2) (MPa) (mm)

X1/X4 6096 203.2 304.8 400 400 600 685 25.34 14.22 28.96
X2/X5 6096 203.2 304.8 400 200 600 685 25.34 14.48 32.26
X3/X6 6096 203.2 304.8 400 0 600 685 25.34 15.75 37.85
Y1/Y4 6340 152.4 203.2 516 516 1000 1000 27.51 22.61 45.97
Y2/Y5 6340 152.4 203.2 516 258 1000 1000 27.51 23.62 49.78
Y3/Y6 6340 152.4 203.2 516 0 1000 1000 27.51 25.40 59.94

22 TECHNICS OUR CIVIL ENGINEERING (2014)

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

Table 10. FE analyses vs. measured values, [22] 5. ACKNOWLEDGMENTS

Exp.
w(i) FE/Exp. w(t) FE/Exp. The work reported in this study is a part of the
(mm) (%) (mm) (%)
investigation within the Research Project TR36017:
X1/X4 16.2 113.9 29.5 101.9 Utilization of by-products and recycled waste materi-
X2/X5 16.3 112.6 32 99.2 als in concrete composites in the scope of sustainable
X3/X6 16.5 104.8 36.7 97.0 construction development in Serbia: investigation and
Y1/Y4 24.6 108.9 42.3 92.0 environmental assessment of possible applications,
Y2/Y5 25 105.8 46.5 93.4 supported by the Ministry for Education, Science and
Y3/Y6 25.5 100.4 54.5 90.9 Technology, Republic of Serbia This support is
= 107.7 = 95.7 gratefully acknowledged.
= 5.1 = 4.3
REFERENCES
Again very good agreement with measured values
is achieved, all results are within 15% of the experi- [1] Maekawa K. ed., Vecchio F. ed., Foster S.ed., Pra-
mental values. ctitioners guide to finite element modeling of rei-
nforced concrete structures Lausanne: International
The same FE model behavior and trends outlined
Federation for Structural Concrete, 2008.
in section 3.2 are also present here i.e. overestimation
of initial deflection (along with increasing accuracy [2] Jonna M. ed., Schreppers G.J. ed., DIANA Finite
with decreasing compressive reinforcement) and und- Element Analysis Users Manual Getting Started
erestimation of long-term deflection development Release 9.4.4. Delft: TNO DIANA BV, 2012.
(again the uncontrolled relative humidity to an extent
[3] Jonna M. ed., DIANA Finite Element Analysis
diminishes the value of this conclusion). Users Manual Concrete and Masonry Analysis
Release 9.4.4. Delft: TNO DIANA BV, 2011.
4. CONCLUSIONS
This work presented a calibration and validation [4] Jaccoud J. P., Favre R., Flche des structures en b-
process of a numerical model for modeling the long- ton arm, Annales d linstitut technique de btiment
et des Travaux Publics, 406 1982.
term behavior or reinforced concrete beams. From the
previous sections the following conclusions are drawn: [5] Espion, B. Benchmark examples for creep and shri-
DIANA is a versatile software with a wide range nkage analysis computer programs, Proceedings of
of possible material and element models and ana- the 5th Int. RILEM Symposium on Creep and Shri-
lysis procedures. It is well suited for modeling the nkage of Concrete Barcelona, 877888, 1993.
long-term behavior of reinforced concrete beams. [6] Vonk, R. A., Rots, J. G., Kanstad, T., Ulm, F. J., and
The initial FE model presented in DIANA Users Navratil, J. Examples of evaluation of computer
Manual ConcMas [3] as a benchmark test for long- codes for creep and shrinkage analysis of concrete
term behavior of RC beams, can be improved upon structures computer programs, Proceedings of the
through a parametric study. Within the boundaries 5th Int. RILEM Symposium on Creep and Shrinkage
of parametric study it was shown that a Multi- of Concrete Barcelona, 889924, 1993.
Directional Fixed Crack model with constant
[7] Borst, de, R., and Nauta, P. Non-orthogonal cracks in
stress cut-off, nonlinear tension softening a smeared finite element model. Engineering Co-
according to Hordijk et al. [15] and constant shear mputations 2, 3546, 1985.
retention is best suited for modeling the cracking
behavior of concrete in this application. For [8] Riggs, H. R., and Powell, G. H. Rough crack model
modeling creep and shrinkage the CEB-FIP Model for analysis of concrete. J. Eng. Mech. Div., ASCE
Code 1990 was shown to be the most precise 112, 5, 448464, 1986.
model in the range of linear creep. [9] Rots, J. G. Computational Modeling of Concrete
The model was successfully validated on 14 sim- Fracture. PhD thesis, Delft University of Techno-
ply supported and 6 continuous beams with very logy, 1988.
good agreement (15%). The model systemati-
cally overestimates the initial deflection. Accuracy [10] Reinhardt, H. W. Fracture mechanics of an elastic
is better when there is no compressive reinfor- softening material like concrete. Heron 29, 2, 1984.
cement. The CEB-FIP creep and shrinkage model [11] CEB-FIP. CEB-FIP Model Code 1990. Comit
tends to undersestimate the long-term deflection Euro-International du Bton, 1993.
development.

TECHNICS OUR CIVIL ENGINEERING (2014) 23

N. TOI at al. CALIBRATING AND VALIDATING A FE MODEL FOR LONG-TERM...

[12] Jonna M. ed., Kikstra W. P. ed., DIANA Finite Ele-
ment Analysis Users Manual Material Library Re-
lease 9.4.4. Delft: TNO DIANA BV, 2012.
[13] Jonna M. ed., Kikstra W. P. ed., DIANA Finite Ele-
ment Analysis Users Manual Analysis Procedures
Release 9.4.4. Delft: TNO DIANA BV, 2012.
[14] Pecic, N. Improved method for deflection control of
reinforced concrete structures. PhD thesis, Faculty of
Civil Engineering University of Belgrade, 2012.
[15] Hordijk, D. A., Cornelissen, H. A. W. and Reinhardt,
H. W. Experimental determination of crack softening
characteristics of normal weight and lightweight
concrete. Heron 31, 2, 1986.
[16] ACI. Prediction of Creep, Shrinkage, and Tempe-
rature Effects in Concrete Structures. Tech. Rep. ACI
209R-92, American Concrete Institute, 1997.
[17] NEN. TGB 1990 Regulations for concrete
Stuctural requirements and calculation methods.
Tech. Rep. NEN 6720, Nederlands Normalisatiein-
stituut, 1995.
[18] Espion, B. Long term sustained loading tests on
reinforced concrete beams A selected data base,
Universit Libre de Bruxelles Service Gnie Civil,
Bulletin 88-1, 1988.
[19] Washa G. W., Fluck P. G., Effect of compressive
reinforcement on the plastic flow of reinforced
concrete beams, Proc. ACI, 49, 1952, 89-108
[20] Glanville W. H., Thomas F.G., Studies in reinforced
concrete: IV, Further studies on the creep or flow of
concrete under load, Building research technical pa-
per, 21, Department of scientific and industrial rese-
arch (London, H.M.S.O.), 43 pp, 1939.
[21] Hognestad E. (principal consultant at Portland Ce-
ment Association), private correspondence with B.
Espion, August 1987.
[22] Washa G.W., Fluck P.G., Plastic flow (creep) of
reinforced concrete continuous beams, Proc. ACI,
52, 549-561, 1956.

REZIME
KALIBRACIJA I VERIFIKACIJA MODELA KONANIH ELEMENATA ZA DUGOTRAJNO
PONAANJE AB GREDA

U ovom radu je predstavljeno istraivanje sprovedeno u svrhu pronalaenja optimalnog modela

konanih elemenata za predvianje ponaanja armirano-betonskih greda pod dugotrajnim optere-
enjem. U istraivanju je korien vienamenski komercijalni softver DIANA. Za kalibraciju modela je
odabran benmark test u vidu proste grede optereene savijanjem u etiri take. Kao rezultat kalibracije,
odabrani model se sastojao od linijskih konanih elemenata sa tri vora. Odabran je model fiksne prsline
u vie pravaca. Sadejstvo zategnutog betona izmeu prslina opisano je nelinearnom funkcijom.
Odabrana je konstantna rezidualna nosivost pri smicanju. Skupljanje i teenje je modelirano prema
CEB-FIP Model Code 1990. Model je dodatno verifikovan na 14 prostih greda i 6 kontinualnih greda
sa dva polja. Rezultati su pokazali dobro slaganje sa eksperimentalnim vrednostima (15%).
Kljune rei: teenje, skupljanje, armirani beton, grede, konani elementi

24 TECHNICS OUR CIVIL ENGINEERING (2014)

. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...

Experimental Determination of Load Bearing Capacity of Connections

Realized by Punched Metal Plate Fastener

IKICA M. TEKI, University of Belgrade, Original scientific paper

Faculty of Architecture, Belgrade UDC: 624.014.2.072.9
SAA M. OREVI, University of Belgrade,
Faculty of Architecture, Belgrade

This paper demonstrates the results of experimental determination of load bearing capacity of
structural timber member connections realized by WOLF and LKVC metal connector plates.
Considering the complexity of the connections realized by these modern mechanical fasteners, this
paper deals only with plate anchorage capacity (stress in the metal-wood contact). The aim of the
conducted experimental study was to determine the metal connector plate anchorage capacity in
accordance with the provisions of Eurocode 5 and also to analyze the ratio of the load bearing
capacities of these two types of connectors in terms of their geometry. Experimental testing was
conducted by loading of multiple samples up to the limit plate anchorage capacity. Discussion of the
test results included the analysis of the connection deformation for different levels of load, as well as
the mode of reaching the limit plate anchorage capacity. Review of the determined limit plate
anchorage capacities, for the determined displacements of connection, was given in the conclusion,
together with the comment on test results.
Key words: metal connector plate, tooth, joint, limit bearing capacity, displacement

1. INTRODUCTION number of teeth per unit area.

Metal connector plates are modern mechanical
fasteners, which are used in contemporary timber
structures. The emergence of these fasteners was
aimed to improve the level of production, in terms of
simplicity of the connections of structural elements,
namely, construction of the connections which will be
of high-quality and safe, in terms of transferring the
load [1].
In the world and in our country there are several
standard types of metal connectors which are
produced by perforating metal sheet and positioning
the teeth perpendicular to a sheet plane. These type of
connectors have been successfully used for the
realization of connections between members of the
timber trusses. The geometry of the standard types of
metal connectors differ from manufacturer to
manufacturer, in relation to the sheet thickness, the
shape and the length and position of the teeth and the
These parameters, as well as the production
technology of standard types of metal connectors,
affect the load bearing capacity of connections in
timber structures.
Load bearing capacity of metal connectors can be
experimentally determined in accordance with
Eurocode 5 and it is used as a parameter for
dimensioning of connections formed by metal
connector plates.
For experimental determination of the load
bearing capacity of connections formed by punced
metal plate, the LKVC plate type is chosen, whose
manufacturer is LKV Center from Serbia [2], so as
the WOLF 15N plate type, whose manufacturer is
WOLF from Austria [3] (Figure 1).

Authors address: ikica Teki, University of

Belgrade, Faculty of Architecture, Belgrade, Bulevar
kralja Aleksandra 73
Paper received: 04.09.2014. Figure 1 - Punched metal plate fastener LKVC (on
Paper accepted: 22.09.2014. the left) and WOLF (on the right)

TECHNICS OUR CIVIL ENGINEERING (2014) 25

. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...

Metal connectors are manufactured in acco- where:

rdance with the provisions of EN 10147 [4]. Both
have same thickness of the sheet of 1.5 mm, which is
important in order to make certain comparison in
terms of load bearing capacity of realized
connections. Presented connectors have different
geometry of teeth, their arrangement and number per
unit area. Distance between the teeth in the two
orthogonal directions is shown in Table 1, from
which we can find the corresponding tooth surface of
169 mm2 for LKVC plate and 215 mm2 for WOLF
plate.
Table 1. LKVC and WOLF plate geometry
LKVC
(mm)
Distance between teeth in
X direction 22.50
Distance between teeth in
Y direction 7.50
Length of teeth 13.50
fa,, - limit plate anchorage capacity for given
angles and (for one connector in connection)
Fa,,,max - maximum (limit) load,
Aef - effective connector area,
k - characteristic density of wood, for certain
class of wood,
- density of wood, for the test sample,
c - dimensionless coefficient.
3. SAMPLES FOR EXPERIMENTAL TESTING
Experimental testing of plate anchorage capacity
was conducted for the angle values of =90 and
=0. For determination of the limit metal connector
WOLF
plate anchorage capacity, in accordance with the
(mm) provisions of EN 14545 [8], standard samples were
tested, with 5 samples in one series, for the connector
25.40 LKVC and the same for the connector WOLF (Figure
2 and Figure 3).
8.46
15.00

2. DIMENSIONING ANCHORAGE CAPACITY OF

THE CONNECTIONS FORMED BY METAL
CONNECTORS
Plate anchorage capacity (which includes analysis
of stresses in the metal-wood contact) is determined
on the basis of laboratory tests, loading the test
Figure 2 - Types of samples of series B (LKVC)
samples to the limit bearing capacity, for different
positions of the longitudinal connector axis, in
relation to the direction of the force and direction of
the wood fibres. Typical values of stresses in the
metal-wood contact, are expressed per unit area of the
metal connector plate, for specific angles and [5],
where is angle between the direction of force and
the direction of the longitudinal connector axis, and is
angle between the direction of force and the
direction of the longitudinal axis of the timber
member. Longitudinal connector axis is direction of
the sheet rolling, namely, direction of the teeth Figure 3 - Types of samples of series D (WOLF)
pulling out of the sheet plane. The longitudinal axis of
the timber member is defined by the direction of the The geometry of the samples was derived
fibres. according to EN 1075. Connector dimensions were
Limit metal connector plate anchorage capacity is chosen in way that approximately the same effective
defined by the provisions of EN 1075 [6] and EN area for each connector type could be obtained, and at
28970 [7], in function of the limit load, effective the same time that there is no fracture in the cross-
connector area and wood density, and it is expressed section of connector, but only in the metal-wood
in N/mm2: contact zone. This is achieved by the ratio of
c connector length L and connector width B, where
F , ,max k L>B/2. Wood that is used for samples is from class II
f a, , (1)
2 Aef conifers, with humidity of 15%. The geometry of all
the samples that were subjected to the test is shown in

26 TECHNICS OUR CIVIL ENGINEERING (2014)

. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...
Figure 4 and Figure 5, noting that the width of the
cross section of the wood element was b=44 mm.
Figure 4 - Types of samples of series B (LKVC)
Figure 5 - Types of samples of series D (WOLF)
4. EXPERIMENTAL TESTING PROCEDURE
The experimental determination of load bearing
capacity of connection of timber elements requires
implementation of specific tests in laboratories, which
are equipped for this purpose, in order to obtain
specific information on the strength of these
connections. Preparation of test samples included
conditioning of the wooden elements, prior to the
construction of the joints, and then the conditioning
of the finished samples, after the connection was
formed. According to ISO 554 standard, atmosphere
for the preparation of samples is 20/65, which means
the air temperature of 20 and humidity of 65%. By
respecting these rules, load bearing capacity of the
tested connections can be compared, including the
comparison of the test results from different
laboratories.
Testing of the load bearing capacity of realized
connections was performed in the Institute of
Materials and Structures, at Faculty of Civil
Engineering in Belgrade, on the hydraulic tensile
testing machine, made by Amsler. During the
application of particular force, deformations were
registered with mechanical deformation indicators,
with 0.01 mm accuracy, which were positioned in
joint area in order to obtain relevant values of
deformations in the connection (Figure 6). Two of
TECHNICS OUR CIVIL ENGINEERING (2014)
these mechanical deformation gauges were positioned
on each sample, in order to record the deformation of
connection on the both sides of the sample.
Figure 6 - Samples of series B (LKVC) and series D
(WOLF)
Testing procedure was conducted according to
EN 26891 [9]. Load was applied in value of 40% of
the estimated maximum load (0.4Fest) and retained
for 30 seconds. After that, the deformation was read,
as well as load that caused that deformation. Then the
load is reduced to a value of 0.1Fest and retained for
30 seconds, whereupon reading was carried out. The
load was then raised until limit load or sliding was
attained. Loads up to the value of 0.7Fest were
approximately 0.2Fest per minute, while for the
values above 0.7Fest limit load or sliding was attained
for about 3 to 5 minutes. Total test time of one sample
was between 10 and 15 minutes. Application of
tensile force was carried out through certain
mechanical additions (Figure 7), with an increment of
2.0 kN, and it was in line with loading procedure, in
function of estimated maximal load. Destruction of
connections came in two ways: due to the extraction
of teeth from the timber, before reaching allowable
displacement and due to exceeding the allowable
displacement in connection. Testing of each sample
was discontinued after the destruction of connections,
when it was visually ascertained that teeth started
pulling out of wood, or the strain in the connection
reached the value of about 2.0 mm.
Figure 7 - Samples of series B and series D (axial
tension - a schematic representation)
27
. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...

5. RESULTS OF THE EXPERIMENTAL TESTING displacement for the value of 10% of estimated
Results of the experimental testing, for both plate maximal load, v04 - displacement for the value of 40%
types, are shown in Tables 2, 3 and 4 and in force- of estimated maximum load i vmax - maximum
strain diagrams (Figure 8), Tables and diagrams are displacement of connection during testing). Load Fest
showing the individual values of connection was estimated for the allowable displacements of 1.5
displacement and the mean average value. mm [10]. In the discussion of the test results, ratio of
Characteristic parameters in accordance to EN 26891 the value of applied load and the measured
are shown in Table 3 and Table 4 (Fest - estimated displacement of connection was analyzed,
maximum load, Fmax - maximum load for the adopted respectively, the limit bearing capacity of connections
allowable displacements of connection, v01 - in function of defined allowable displacement are
given.
Table 2. Test results (individual and mean value displacement)
Load (kN)
Sample 0.0 10.0 2.0 6.0 10.0 14.0 18.0 22.0 26.0 30.0
Displacement (mm)
B1 0.0 0.055 0.020 0.035 0.065 0.115 0.200 0.325 0.535 0.985
B2 0.0 0.060 0.010 0.040 0.070 0.110 0.180 0.280 0.405 0.640
B3 0.0 0.085 0.030 0.065 0.090 0.135 0.195 0.285 0.415 0.715
B4 0.0 0.075 0.020 0.060 0.085 0.130 0.205 0.310 0.515 0.835
B5 0.0 0.085 0.035 0.065 0.095 0.150 0.225 0.355 0.535 0.940
Mean value
displacement (mm) 0.0 0.072 0.023 0.053 0.081 0.128 0.201 0.311 0.481 0.823

D1 0.0 0.120 0.060 0.105 0.135 0.245 0.460 0.850 2.140

D2 0.0 0.115 0.045 0.090 0.120 0.215 0.385 0.770 1.765
D3 0.0 0.110 0.045 0.085 0.120 0.210 0.385 0.775 1.795
D4 0.0 0.120 0.055 0.100 0.125 0.225 0.435 0.920 2.320
D5 0.0 0.105 0.060 0.085 0.115 0.220 0.405 0.790 2.035
Mean value
displacement (mm) 0.0 0.114 0.053 0.093 0.123 0.223 0.414 0.821 2.011

L K V C - sa m ple s B 1 to B 5 WOL F - sa m ple s D1 to D5

30.0 30.0
28.0 28.0
26.0 26.0
24.0 24.0
22.0 22.0
20.0 20.0
L oad (kN)

18.0
L oad (kN)

18.0
16.0 16.0
14.0 14.0
12.0 12.0
10.0 10.0
8.0 8.0
6.0 6.0
4.0 4.0
2.0 2.0
0.0 0.0
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0

Dis plac e m e nt (m m ) Dis plac e m e nt (m m )

Figure 8 - Force-displacement diagram (mean value displacement)

28 TECHNICS OUR CIVIL ENGINEERING (2014)

. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...

Table 3. Test results (individual value displacement)

Load Displacement
Series Sample Fest Fmax v01 v04 vmax
(kN) (kN) (mm) (mm) (mm)
B1 26.0 30.00 0.020 0.065 0.985
B2 26.0 30.00 0.010 0.070 0.640
B B3 26.0 30.00 0.030 0.090 0.715
B4 26.0 30.00 0.020 0.085 0.835
B5 26.0 30.00 0.035 0.095 0.940
D1 26.0 24.51 0.060 0.135 2.140
D2 26.0 25.20 0.045 0.120 1.765
D D3 26.0 25.11 0.045 0.120 1.795
D4 26.0 23.93 0.055 0.125 2.320
D5 26.0 24.69 0.060 0.115 2.035

Table 4. Test results (mean value displacement)

Load Displacement
Series Fest Fmax v01 v04 vmax
(kN) (kN) (mm) (mm) (mm)
B LKVC 26.00 30.00 0.023 0.081 0.823
D WOLF 26.00 24.69 0.053 0.123 2.011

6. DISCUSSION OF THE TEST RESULTS
Based on the test results, there are evident diffe-
rences in the strain of connections realized by plates
made by diferent manufacturers, for the same value of
applied load. For all samples, and both plate manu-
facturers (LKVC and WOLF), it can be distinguished
only a slight difference in the sample strain to the ex-
tent of 40% of the maximum estimated load (about
10.0 kN), after which the increment in deformation is
larger for WOLF plate, comparing to LKVC plate and
for the same value of applied load. Since the effective
area of LKVC plate is only 3.57% higher than the
WOLF plate, for the same value of displacement, it
can be concluded that LKVC plate has a higher load
bearing capacity per unit area, compared to WOLF
plate, for the applicable load. If we analyze the
behavior of connections in relation to the boundary
displacement of 1.5 mm, certain differences in the
achievement of ultimate bearing capacity of conne-
ctions for both plate types can be identified. The
LKVC plate never reaches the boundary displacement
of 1.5 mm, since at a significantly lower amount of
displacement the extraction of teeth from the timber
began and in that point further testing was stopped.
The point in wich teeth start pulling out of the timber
represents a physical destruction of the connection,
which practically defines the breaking force, which in
this case corresponds to the shift of only 0.823 mm,
taking into account the average value of five samples
tested. In the other hand, the WOLF plate reached the
border displacement of 1.5 mm, but at lower load
applied compared to LKVC plate, which eventually
resulted in a smaller load bearing capacity per a unit
area for a WOLF plate. Testing of load bearing
capacity of connections realized by WOLF metal
plates was stopped at an average shift of about 2.0
mm, since it was visually ascertained extraction of
teeth from the timber.
Having in mind that testing was conducted in
accordance with standard ISO 554, and the fact that a
same class of wood, with a same density, was used
for all samples, value of c=0 is taken for
dimensionless coefficient c (1), from the provisions of
EN 1075 and EN 28970. In this way the effect of ratio
of characteristic and real wood density on the load
bearing capacity of connection is excluded.
Accordingly, limit plate anchorage capacities are
determined for both plate types, as well as the ratio

TECHNICS OUR CIVIL ENGINEERING (2014) 29

. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...
between limit plate anchorage capacities of LKVC
and WOLF plates (Table 5).
Table 5. Limit plate anchorage capacity
Plate type Plate anchorage capacity
fa,90,0
LKVC 2.67 N/mm2
WOLF 2.27 N/mm2
LKVC / WOLF 17.62 %
If we analyze the spacing between teeth in two
orthogonal directions, and i another words we take
into account the number of teeth per unit area, the lo-
adbering capacity of one teeth can be identified. The
corresponding surface for one tooth of LKVC plate is
169 mm2, and 215 mm2 for WOLF plate, and so we
can conclude that the capacity of a single tooth of
WOLF plate by about 8.20% higher than the capacity
of a single tooth of LKVC plate (Table 6). This data
is not relevant for load bearing capacity of
connections, in terms of dimensioning of the
Figure 9 - Samples of series B (LKVC) and series D (WOLF)
7. CONCLUSION
Experimental testing of punced metal plate faste-
ners anchorage capacity were carried out on samples
made of timber connected by plates of two diferent
types: LKVC and WOLF 15N. Samples were sub-
jected to axial tension, in accordance with EN 1075,
in order to determine the behavior of these two types
of plates in connections, where the behavior of conne-
ction is analyzed in a linear and the non-linear part of
the strain diagram. The results gave a picture about
the relation between carrying capacity of these two
types of plates, in terms of their geometry, for various
levels of load of timber connections, and the values of
angles =90 and =0. plate.
30 TECHNICS OUR CIVIL ENGINEERING (2014)
connection, because bearing capacity of connector is
expressed per unit of surface area and not by bearing
capacity of one tooth. This data can be useful during
the designing of connectors geometry and geometry
of tooth, in order to increase the connector load
bearing capacity.
Table 6. Limit bearing capacity of one tooth
Connector Capacity of one tooth
LKVC 451.0 N
WOLF 488.0 N
WOLF / LKVC 8.20 %
Figure 9 is showing samples of B and D series in
the phase of testing, as well as the plate position
relative to the timber, in a moment when testing was
stopped.
Ultimate anchorage capacity of connections is
determined by the adopted permissible displacement
of connection of 1.5 mm, which proved to be a key
criterion for conections executed by WOLF plates,
unlike the connections executed by LKVC plates,
where the failure occurred before reaching
permissible displacement of 1.5 mm. For both plate
types only small differences in capacity were evident
in the area of of elasticity, where the amount of strain
was approximately 0.1 mm, while in the area of
plastic deformation LKVC plate performed better, as
it was noted in ratio of ancorage capacity of these two
type of plates which is about 18% in favor of LKVC
. TEKI at al. EXPERIMENTAL DETERMINATION OF LOAD BEARNIG CAPACITY...

Higher capacity of one tooth for WOLF plate, by
about 8% compared to the one tooth of LKVC plate
may be partially explained by greater length of
WOLF plate teeth by 2.0 mm compared to the LKVC
plate teeth, but it is not the only parameter which
determines the anchorage capacity of plates. Number
of teeth per unit area of plate proved to be an
important parameter in the function of the anchorage
capacity. The corresponding surface of a single tooth
of WOLF plate is 27% higher compared to that of the
LKVC plate, which is evident in determining the
ratios of limit load carying capacity per unit area, in
favor of LKVC plate, despite a greater capacity of
single teeth of WOLF plate.
For the final comparative assessment of the
quality of these two types of plates, in terms of
anchorage capacity, it is necessary to acquire data on
the anchorage capacity for other combinations of
angles and . Also, the distance of the farthest row
of teeth to the line of splice may be of importance to
limit anchorage capacity, or to teeth beginning
extracting out of wood. Therefore, new experimental
tests of anchorage capacity of metal plates, should
analyse the impact of plate size to a limit anchorage
capacity.
REFERENCES
[1] Teki, ., Savremeni koncepti primene metalnih ko-
nektera u sistemima drvenih struktura, Doktorska
disertacija, Arhitektonski fakultet Univerziteta u Be-
ogradu, Beograd, 2005.
[2] Karakteristike konektera LKVC, www.lkvcen-
tar.com
[3] Karakteristike konektera WOLF,
www.wolfsystem.at
[4] EN 10147:2000, Specification for continuously hot-
dip zinc coated structural steel sheet and strip -
Technical delivery conditions.
[5] Eurocode 5, Design of timber structures. EN 1995-
1-1, Part 1-1: General - Common rules and rules for
buildings, 2004, Brussels: European committee for
standardization.
[6] EN 1075:1999, Timber structures - Test methods.
Testing of joints made with punched metal plate
fasteners.
[7] EN 28970:1991, Timber structures. Testing of joints
made with mechanical fasteners; requirements for
wood density (ISO 8970:1989).
[8] EN 14545, Timber structures - Connectors -
Requirements.
[9] EN 26891:1991, Timber structures. Joints made
with mechanical fasteners. General principles for
the determination of strength and deformation
characteristics.
[10]SRPS U.C9.200, Projektovanje i izvodenje drvenih
konstrukcija - Konstrukcije od monolitnog drveta i
ploa.

REZIME
EKSPERIMENTALNO ODREIVANJE NOSIVOSTI VEZA OSTVARENIH METALNIM
KONEKTERIMA

U radu je prikazano eksperimentalno odreivanje nosivosti veza ostvarenih metalnim konekterima tipa
WOLF i LKVC. S obzirom na kompleksnost veza ostvarenih ovim savremenim mehanikim spojnim
sredstvima, predmet rada obuhvata samo nosivost bone veze (naprezanje u spoju metal-drvo). Cilj
sprovedenih eksperimentalnih ispitivanja je bio da se utvrdi nosivost bone veze ostvarene metalnim
konekterima, u skladu sa odredbama Evrokoda 5 i da se paralelno sa tim analizira odnos nosivosti ova
dva tipa konektera, sa aspekta njihove geometrije. Eksperimentalno ispitivanje je sprovedeno na vie
uzoraka, optereivanjem uzoraka do dostizanja granine nosivosti bone veze. Diskusija rezultata
ispitivanja je obuhvatila analizu pomerljivosti veza za razliite nivoe optereenja, kao i nain
dostizanja granine nosivosti bone veze. U zakljuku je dat komentar o utvrenim graninim
nosivostima bone veze, za odreena pomeranja veze, i zauzet je stav po pitanju dobijenih rezultata.
Kljune rei: metalni konekter, zubac, vorna veza, granina nosivost, pomeranje

TECHNICS OUR CIVIL ENGINEERING (2014) 31

 TEHNIKA
INVITATION TO COOPERATION AND SUBSCRIPTION

TEHNIKA is the leading Serbian scientific and technical magazine covering areas from science
and professional topics. It has been published for 69 years in a row by the Union of Engineers
and Technicians of Serbia thus creating a unique scientific and technical basis for transfer of
scientific, technologic, commercial and business information originating both locally and
internationally. According to the Ministry of Education, Science and Technological
Development of Republic of Serbia this magazine has been thought of as a first class magazine
in its category.
All published contributions and papers are subjected to professional review in order to
guarantee highest scientific and professional level. Published papers are being referred to in
internationally renown magazines, such as: Geotechnical Abstracts, Metals Abstracts, Chemical
Abstracts, Electrical and Electronics Abstracts, Science Abstracts, Ergonomics Abstracts and
reference journals VINITI. They also serve as background source of information for
international data base requirements: INSPEC (IEEE, London), METADEX (M. I., England),
CASEARCH (CA, USA).
The magazine is being disseminated both throughout the country and abroad (in about 30
countries) through subscriptions and scientific magazine exchange.
Readership in Serbia includes businesses and companies, professional libraries, military
readership, faculties, institutes, colledge level and high schools, social and political entities and
organizations, embassies, representation offices, trade and public relations agencies, chambers
of commerce, individuals and other.
The magazine offers its space to papers classified under areas, such as: New Materials, Our
Civil Engineering, Mining, Geology and Metallurgy, Mechanical Engineering, Electrical
Engineering, Transport and Traffic Engineering, Management, and Standardization, Quality and
Metrology.
In addition, there are columns permanently covering a variety of topics, such as overviews of
books and magazines, review of scientific and professional gatherings both local and
international, technical news information.
This 180 page magazine is being published bi-monthly in A-4 format.
Cooperation with the magazine offers you the possibility to present not only the results of your
efforts and expertise but also the knowledge and developments in the company you work for in
a way that makes it known and available to relevant scientific and professional circles.
UNION OF ENGINEERS AND TECHNICIANS OF SERBIA
Kneza Milosa 7a/I, 11000 Belgrade, Serbia
Tel.: Editorship: (011) 32-35-891, 3237 363
e-mail: tehnika@sits.rs; office@sits.rs
_____________________________________________________________________________
 ISSN 0350-2627

TECHNICS

MINING
GEOLOGY AND
METALLURGY
Rudarstvo, geologija i metalurgija - Exploitation des mines,
gologie et metallurgie - Bergbau, Geologie und Metallurgie -

YEAR 65 2014.

EDITOR-IN-CHEF
and Editor for Mining:
Prof. Boo K o l o n j a, D.Sc, University of Belgrade,
Faculty of Mining and Geology, Belgrade

Editor for Geology:

Prof. Aleksandar Kosti, D.Sc, University of Belgrade,
Faculty of Mining and Geology, Belgrade

Editor for Metallurgy:

Prof. Zvonko Guliija, D.Sc, Institute for Technology of
Nuclear and other Mineral Raw Materials, Belgrade

EDITORIAL COUNCIL:
Prof. Radule T o o v i , D.Sc, University of Belgrade,
Faculty of Mining and Geology, Belgrade
Prof. Todor S e r a f i m o v s k i, D.Sc, Faculty of Mining
and Geology, tip, Macedonia
Mirko M a k s i m o v i , B.Sc, Association of Mining and
Geological Engineers of Serbia, Belgrade
Prof. Nada trbac, D.Sc, University of Belgrade, Technical
Faculty in Bor, Bor
Miroslav Soki, D.Sc Institute for Technology of Nuclear
and other Mineral Raw Materials, Belgrade

EDITORIAL OFFICE: Union of Engineers and

Technicians of Serbia, 11000 Belgrade, Kneza Miloa 7a/I,
Tel. +381/11/ 32 35 891, Fax +381/11/ 32 30 067
S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...

Influence of Ash Slurry Preparation Method on Characteristics of Ash

Disposed from the Thermal Power Plant "Gacko"

SRETEN M. BEATOVI, Mine and Thermal Power Plant Original scientific paper
"Gacko", Gacko, Republic of Srpska UDC: 662.613.11:628.472.3(497.6)
DINKO N. KNEEVI, University in Belgrade,
Faculty of Mining and Geology, Belgrade
LAZAR B. STOJANOVI, University in Banja Luka,
Faculty of Mining, Prijedor, Republic of Srpska
SLAVKO M. TORBICA, University in Belgrade,
Faculty of Mining and Geology, Belgrade
IVKO T. SEKULI, Institute for Technology of Nuclear and
other Mineral Raw Materials, Belgrade

Over the previous period ash with high calcium content from the Power Plant "Gacko" (Republic of
Srpska) has been disposed into 2 cassettes. The slurry disposed into cassette no. 1 was prepared in a
non-industrial horizontal drum with nozzles, while ash slurry disposed into cassette no. 2 was mixed in
an agitator. This paper presents the investigation results of samples taken from both cassettes in order
to determine the influence of slurry preparation method on the characteristics of the ash that has been
disposed of. The results obtained show that the slurry preparation method may have a significant
influence on the characteristics of this ash.
Key words: ash disposal site of "Gacko" Thermal Power Plant, ash with high calcium content,
spontaneous solidification, disposed ash characteristics

1. INTRODUCTION Bad, but somehow better average situation is in

In the process of coal combustion for electrical
power generation ash slurry transport and disposal
presents a continuous problem. More than 700 million
tons of ash is produced from coal worldwide [1]. It
gives a ratio of 100 kg of ash per 1 inhabitant per year
and 47 kg/km2 of Earth surface.
China, USA and India are the countries with the
largest ash production. In relation to the surface the
area of country most unfavorable situation is in Gre-
ece, 91.3 t/km2. Large production is also present in
Slovenia, Serbia, Poland, Slovakia, Bulgaria, Ger-
many, China, etc.
Considering the number of inhabitants, Greece has
again the most unfavorable ration, since it generates.
More than 1 ton of ash per year per person each year.
Author`s address: Sreten Beatovi, Mine and The-
rmal Power Plant Gacko, Gacko, Graanica bb,
Republic of Srpska
Paper received: 13.05.2014.
Paper accepted: 19.05.2014.
Serbia, Bulgaria and Slovenia. China as the largest
"producer" of ash and the country with the greatest
number of inhabitants has a load of 320 kg per
inhabitant, while in India this ration is much smaller,
80 kg per inhabitant. The analyzed results just show
average values. If the actual conditions were analyzed
on micro locations where coal, energy and ash are
"produced" the results would be far worse.
Lignite, with caloric value ranging from 8000 to
8500 kJ/kg, is combusted in the 300 MW Thermal
Power Plant (TPP) Gacko (Republic of Srpska, Bos-
nia and Herzegovina). The annual production of ash is
about 400,000 tons. According to American standard
ash from TPP Gacko is class C [2], by Russian it
is "active" [3, 4] and by Serbian classification it is "ca-
lcium" [5]. The basic characteristic of these type of
ashes is spontaneous solidification on disposal sites
where the amount of added water is adjusted to the
content of the active substance in ash.
The disposal of calcium ash in the form of thick
slurry generates spontaneous self-solidification of ash

TECHNICS MINING, GEOLOGY AND METALLURGY (2014) 35

S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...
and brings various technological and geotechnical be-
nefits.
Solidification is a known method enlarges small
particles transferring them into stabile form, techno-
logically and ecologically. Another classification di-
vided these methods into six groups by the way of
stabilization: self-solidification, by lime or cement ad-
ding, by silicate adding (silica gel or cement), thermo
plastic impregnation (asphalt, bitumen, paraffin,
polyethylene etc), impregnation of organic polymers
(urea-formaldehyde for polymerization) and inorganic
precipitation (ferric ions) Worldwide practice shows
that the most common use have the methods from first
two groups stabilization on the basis of lime or ce-
ment adding and self-solidification.
They have the most common use in solidification
of nuclear and toxic waste [6]. Solidification of nuclear
waste is broadly accepted because of its harness, low
leachability of toxic elements, retention of stabile form
in long period of time, incombustible feature, etc. Ash
mixing with high concentration of calcium, cement and
nuclear waste is often used in order to transfer nuclear
waste into stabile form, i.e. to prevent any leachability
out of nuclear waste disposal site. Name of this
technology is CFS Chemical fixation and
solidification [7-13].
In Malaysia, agricultural waste and waste from ag-
ro-processing industries were mixed with other toxic
waste and combusted. Ash as a solid residue was
solidified and stabilized by adding Portland cement
[14]. Research has shown that cement cant succe-
ssfully stabilize hard metals, so the process was im-
proved by adding active coal and rice husk.
Solidification of ash which was made by combu-
sting medical waste is done by cement adding (up to
60% in ration of ash mass). A stable mass is formed
thus preventing leaching of hard metals and the
material disposed become geotechnical stable [15].
Ash obtained from communal waste combustion is
solidified by extrusion with polyethylene of low de-
nsity, polystyrene, polyvinyl chloride and their mi-
xture. The results obtained open new perspectives in
usage of plastic waste for ash stabilization [16].
Kmiecik patented the ash solidification method by
adding: CaO, CaCl2 .H2O, KNO3, NaHSO4 or KHSO4.
The disposing system has experienced semi-industrial
usage on TPP Lagisa [17].
The research of silica ash solidification conditions
in TP "Kostolac-B" with lime and bentonite usage have
given positive results when additive usage was 5-10%
[18].
Ash self-solidification has been noted in Vis-
koza Heat Power Plant from Loznica and on the Hy-
draulic Ash Disposal Site TPP "Kosovo-A", in
36 TECHNICS MINING, GEOLOGY AND METALLURGY (2014)
Estonian TPP Estonskaja GRES and on the Ash
Disposal Site TPP Kakanj in Bosnia [19].
Spontaneous solidification in controlled conditio-
ns of preparation and ash disposal is noted on TPP
Kosovo-B ash disposal site [20].
Solidification of ash from coal in an autoclave was
researched in Japan [21]. Fly ash was first of all co-
mpacted under the pressure of 20 to 50 MPa and then
hydrothermally treated in an autoclave at temperatures
ranging from 150 to 250oC during 15-60 hours. Mi-
neral tobermorite or calcium silicate hydrate similar to
tobermorite is formed by solidification. The hardest
product was obtained when the ratio Cao/SiO2 was
close to 0.83. Compressive strength achieved under
hydrothermal conditions was above 10 MPa.
Russian authors [3, 4] who were involved in the
research process of coal ash solidification came to the
conclusion that, because of the complexity of the
composition of the system it is very hard to research
the process, so they gave the precedence to the less
complicated research of individual constituent phases.
Up to 40 minerals were noted in the ash composition.
Those with various speed, bring in the reaction with
water thus making hydration and hydrolyses products
of various composition.
2. RESEARCH TOPIC
TPP "Gacko" disposing of ash in the excavated
area of the open pit mine. Backfilling of the disposal
site started in 1996 and it is still active. It consist of
two cassettes.
Inside cassette no. 1, which was active from 1996
until 2007, ash was disposed as low concentration slu-
rry by using a drum device for moistening and slurry
preparation. Since 2007, disposal has been shifted to
cassette no. 2 and slurry preparation is due in specially
made device.
The disposal into cassette no. 1 was basically cha-
racterized by a technological disorder in the slurry pre-
paration process and great alteration of slurry density.
On the other hand, the disposal in cassette no. 2 is
carried out by implementing technologically comple-
tely regulated system of slurry preparation which auto-
matically maintains process parameters in set limits.
So, there are two different methods for slurry
preparation. As a result from the previous some basic
objectives should be noted. It has been considered in
the research, as follows: the comparison of the results
obtained on the mass disposed in order to see the
benefits of regulated disposal technology usage.
All research, monitoring and overview of cha-
racteristics of mass disposed are published "in situ"
with conscious neglecting of side effects and actions
S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...

on and around disposal site itself (unequal geotec-
hnical cassette arrangement, water decantation out of
cassette no. 1 into cassette no. 2, different height of di-
sposed mass in cassettes, different age of disposed
mass, etc).
Ash disposal site is formed on excavated area of the
active open pit "Gracanica" in the part that is called
"Field A", figures 1 and 2.
Figure 1 - Location of cassette no. 1 and 2 inside of the
open pit mine "Graanica, Field A"
In order to completely backfill the excavated area
three cassettes are planned in total. So far, two of
similar size are active (averagely 500 x 250 m). Both
cassettes are completely hydrotechnically arranged
which includes lining of bottom and sides with HDPE
geomembrane.
3. PRIOR RESEARCH
When water comes into contact with ash from TPP
"Gacko" a turbulent exothermic reaction. When the
mass concentration of ash is below 40% there is not
complete binding of technological water and the ve-
locity of the reaction is slow. When the concentration
is between 40% and 60% a complete binding of
technological water occurs and the velocity of water
binding goes from several hours (when it is 40%) to
less than 30 minutes (when it is 60%).
When mass concentration exceeds 60% the rea-
ction is extremely fast and it binds technological water
in the time of 10-13 minutes. For industrial purpose the
mass concentration of 50% of ash was chosen. It
assures complete binding and evaporation of added
water with the possibility of binding all atmospheric
waters contained in the disposal site [22].

Figure 2 - Cross section through cassette no. 1 and 2 of ash disposal site
4. SLURRY PROCESSING AND DISPOSAL IN One or two (out of three) installed lines for ash
CASSETTE NO. 1 transport are active in every moment.
Ash pneumatically transported, by pressure ves- Ash moistening is done in the steel drum, length
sels, in a moistening drum which is located at the edge 6 m, width 1.5 m, figure 3. The nozzles are placed
of the disposal site. Transport is performed dis- around the circumference of the cylinder in several
continuously. lines in order to form water curtain in the cylinder [23].

Figure 3 Scheme and external appearance of the ash moistening drum

The supply of ash into the drum is discontinuous Monitoring results of moistening device are shown
while water inflow is continuous. Slurry flows out of in Table 1 [24].
drum into a previously prepared trench that it uses and In every hour, moistening system is working be-
it is directed into the active cassette. tween 25-32% without ash, just with water. Naturally,

TECHNICS MINING, GEOLOGY AND METALLURGY (2014) 37

S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...

a disorder in water balance occurs, because the water control and regulation. Movement of ash concentration
level in the disposal site is higher than required. In in slurry is shown in figure 5.
periods when device is "feed" with both ash and water,
slurry density range in intervals between 1.100-1.250
kg/m3 (Cm = 14 26%)
Table 1. Overview of the operation of devices for moi-
stening with slurry or water only
Time of Operation with..., %
Day
Slurry Water only
1. 68.8 31.2
2. 68.4 31.6
3. 75.2 24.8 Figure 5 Slurry density movement
4. 72.3 27.7 Low ash concentration in slurry in the beginning is
5. 74.5 25.5 a consequence of the slow reaching of the previously
Average 71.9 28.1 set concentration, and at the end flushing of pipeline.

5. SLURRY PROCESSING AND DISPOSAL ON 6. RESULTS AND DISCUSSION

CASSETTE NO. 2 Samples. Dry ash samples were taken below ele-
Dry ash is transported pneumatically, by pressure ctrostatic precipitator. Slurry samples were taken at the
vessels, into one of two concrete silos. Silos provide end of pipeline from the disposal site. Samples of ash
smooth operation of power plant in three shifts. Slurry disposed in both cassettes were taken in research drills,
preparation system has two independent, identical core sampling. Research drillings was done mechani-
lines. cally, rotationally with continuous coring and with mi-
nimal water usage. One research drill was done on
every cassette.
Research method. Chemical composition is deter-
mined by using standard methods of chemical analysis.
In order to determine mineralogical alterations samples
were researched by the method of X-ray diffraction on
the poly-crystal sample (dust) by using diffracto-meter
Philips PW1710. Granulometric composition is deter-
mined by sieving on series of Tyler laboratory screens.
Permeability coefficient is set in permeameter by water
flow measuring via sample with declining water pre-
ssure. Slurry density, temperature and temperature of
air and water were determined on the very location.
Geotechnical parameters were determined in accor-
dance with Serbian standards.
Figure 4 - Technological scheme of slurry preparation Dry ash characteristic. Chemical composition of
ash is given in table 2.
Dry ash is dozed into a covered agitator via sector
feeder. Slurry preparation and process parameters are Table 2. Chemical composition of ash
regulated in the agitator. Hydraulic transportation is Compo- Cassette no. 1 Cassette no. 2
done through two pipelines up to cassette no. 2. nents, % Range Average range average
Technological scheme of slurry preparation is shown SiO2 3.5-8.8 4.5 8.5-12.4 9.9
in figure 4. Al2O3 1.6-6.7 4.9 7.7-11.7 9.6
Disposal is done by free discharge of slurry in the Fe2O3 1.6- 2.7 2.0 2.2-4.5 2.5
accumulation area. CaO 63-69 66.3 53-63 59.9
Given mass ratio between ash and water is 1:1, i.e. MgO 1.1-3.5 1.9 4.4-5.9 5.7
slurry density is 1482 kg/m3. TiO2 0.1-0.5 0.2 0.3-0.4 0.4
K2O 0.3-0.7 0.4 0.6-0.9 0.8
Usually parameters of hydraulic transport are not
kept constant, but in optimal or acceptable boundaries Na2O 0.5-0.9 0.6 0.5-0.9 0.5
with the assistance of the measures that were taken for SO3 7.9-14.5 12.4 10.1-13.5 12.0

38 TECHNICS MINING, GEOLOGY AND METALLURGY (2014)

S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...

CaO is the prevailing component in all samples. Geomechanical Geomechanical

Participation of SO3 group is significant which Eleva- material Eleva- material
indicates that spontaneous desulphurization occurs in tion description in tion description in
the thermal plant boiler. The presence of other layer layer
components is mostly minor. CASSETTE NO. 1 CASSETTE NO. 2
According to its mineral composition, the active 933.8
matter is made of calcium oxide with the presence of
minerals of such as cement, calcium silicate, alumino
ferrite and increased quantities of anhydrite.
According to its coarseness it is finely pulverized
material with over 92% of fraction -0.074+0 mm, and
all particles are finer than 0.147 mm. 1 Moderate
The average density of dry ash which is disposed solidified ash 918.5
in both cassettes was 2,800 kg/m3.
It may be concluded that the dry ash disposed in
both cassettes had similar characteristics and that it
could not induce altered behavior of slurry during 1- Moderate
solidified ash
disposal nor could it affect the different results
915.5
obtained for the ash disposed.
910.3 4- Solid,
Slurry density. Slurry density measuring and
904 compact
concentration of solid phase certainly show the 2 moist 904.0
difference and difference in solid phase concentration. 899.4
3 poorly
solidified 1 - Moderate
Low concentration slurry density is in a range of Bottom of cassette no. 1 solidified ash
1.050-1.250 kg/m3, and concentration of solid phase is 895.5
in range of 7-31%. Density of high concentration 4- Solid,
slurry was in a range of 1.330-1.360 kg/m3 and 892.3 compact
concentration was in a range of 39-41%. Bottom of cassette no. 2
There is an obvious difference in slurry density and Figure 6 Geomechanical description of lithological
also difference in relation to designed density of 1.482 layers from research bore holes
kg/m3. Data from the period when disposing was done Results of research work on cassette no. 2 indicate
in cassette no. 1 shows that in all circumstances in that on relative equalized parameters, i.e. that heteroge-
cassette there was a water surplus. neous degree of dispose material is significantly less
During the disposal period in cassette no. 2 it can than on cassette no. 1. Disposed ash has expressed ho-
be noted that the system operates at a boundary of rizontal layering and anisotropy which leads to several
about 40%, i.e. at the limit when water surplus may times more water permeability in horizontal than ve-
appear. rtical direction.
In both cases the slurry that is being disposed Moisture, figure 7, in cassette no. 1 significantly
allows appearance of water surplus. The situation was varied when depth is concerned without particular or-
worse in cassette no. 1 than in cassette no. 2, but due der and regularity, while in cassette no. 2 is stable and
to cassette configuration (figure 2) there is a possibility without significant effect of depth.
that the water from cassette no. 1 could easily spills
outs into cassette no. 2.
Lithological composition of ash from cassette no.
1 and cassette no. 2. In both cassettes, in terms of
compactness, water sufficiency and connection degree
of solidified mass, the material disposed is extremely
heterogeneous and anisotropic, figure 6.
Based on the measuring results obtained from
cassette 1 it has been established that up to the ele-
vation 903 the ash disposed is under water. Above the
903 elevation the ash layers differ in their physical Figure 7 Moisture depending on the depth of the bore
mechanical parameters. hole: 1 Cassette no. 1, 2 Cassette no. 2

TECHNICS MINING, GEOLOGY AND METALLURGY (2014) 39

S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...
Angle of inner friction and cohesion. Comparative
overview of inner friction angle movement, in function
of depth of disposed material, it was shown on figure
8 for both cassettes.
Figure 8 Angle of inner friction depends of bore hole
depth: 1 cassette no. 1, 2 cassette no. 2
The angle of inner friction in cassette no. 1 varies
significantly and certain regularity cannot be drawn,
while on cassette no. 2 it is constant and it increases
slightly along with sampling depth. Considering that
the angle of inner friction is part of stability calculation
patterns it may be concluded that the situation in
cassette no. 2 is more favorable.
The comparative overview of movement cohesion
in both cassettes considering the depth of the material
disposed is shown in figure 9.
Comparative overview shows great cohesion
variations in cassette no. 1. The cohesion is very high
except in cassette surface. It is even high in the layers
that are below the level of free water. In cassette no. 2
there is no significant difference in relation to the
disposal site depth, except that area 2 is characterized
as a "solid, compact ash" and it jumps out of fairly
equalized values.
Lower values of cohesion are generally noted in
second cassette, despite the description of lithological
members which determine the material disposed as
more favorable than in cassette no. 1.
Figure 9 Cohesion depends on bore hole depth: 1
cassette no. 1, 2 cassette no. 2
Compressibility module. Comparative overview of
compressibility module is given on figure 10.
40 TECHNICS MINING, GEOLOGY AND METALLURGY (2014)
Figure 10 Values of module compressibility under
load of 200 400 kN/m2, depending on bore
hole depth: 1 cassette no. 1, 2 cassette no.
2
Comparative overview of compressibility module
movement in function of the depth of ash disposed on
cassette no. 1 shows that the biggest module is at
elevation 907, about 4 m above water level and after
that it continually decreases. In cassette no. 2 there are
no extreme variations and the highest value was
recorded around elevation 907. Declining trend is
noted in cassette no. 2, although very slow in direction
to disposal site surface.
Relative compression. Comparative overview of
the results obtained from relative compression expe-
riment with load increase of 100 200 kN/m2 and from
200 to 400 kN/m2 is shown in figure 11.
It may be concluded that in the cassette no. 1
relative compression increases with disposal depth
while in the cassette no. 2 a reverse process occurs due
to that fact that the compressibility is higher in surface
layers than at the bottom of the cassette. Results from
the cassette no. 1 are given for the ash located above
the level of free water, while with cassette no. 2 it is
also shown compression on samples taken below level
of free water. High relative compression in lower
layers of cassette no. 1 is possibly consequence of
water presence.
Figure 11 Relative compression at different load
depending of bore hole depth
The permeability that depend of the depth of the
ash is higher on cassette no. 1 because the coefficient
at all depth is 10-5 m/s, while in cassette no. 2 it ranges
from 10-6 up to 10-7 m/s. The lowest permeability is
S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...
recorded in area 2 with ash which was classified as
solid, compact. The values obtained indicate that the
state of the ash disposed in cassette no. 2 is more
favorable due to the fact that the material is better
compacted so it leads to a slightly slower water flow
through the ash disposed.
Mineral composition. The presence of ettringite,
portlandite, magnetite, calcite, olivine has been
determined in analyzed samples. The most common
phases are ettringite and portlandite.
7. CONCLUSION
Hydraulic disposal process in "Gacko" TPP ash
disposal site has been progressing in two phases and in
two cassettes. Hydraulic disposal in cassette no. 1 and
cassette no. 2 differed in slurry preparation method, in
the amount of added water and in the way slurry was
disposed.
Studies have been conducted about disposal ap-
plied in both cassettes based on the knowledge of the
characteristics of dry ash from both periods based on
the studies of chemical, mineralogical, geomechanical
and hydrogeological characteristics of disposed mass.
Dry ash that was disposed in both cassettes had
almost identical characteristics, i.e. dry ash could not
be reason for different behavior of slurry during dis-
posing or for getting various results on the ash dis-
posed.
Data from the period when ash was disposed into
cassette no. 1 showed that surplus of free water appe-
ared in the cassette. Slurry density ranged from 1.000
kg/m3 (just water) up to 1.250 kg/m3, and in solid phase
concentration 0-31%.
By geological mapping of the research bore hole
core in cassette no. 1 three areas have been identified:
Area 1 from surface up to elevation 910 hete-
rogeneous composition, solidified to poorly soli-
dified,
Area 2 from elevation 910 to 903 ash is wet
Area 3 - from elevation 903 up to bottom non
solidified to poorly solidified, wet.
The moisture in cassette no. 1 significantly varies
without special order and rules. The angle of inner
friction is very variable and certain rule cannot be
determined. Comparative overview shows great vari-
ations in cohesion. Although cohesion is not high in
the surface area of the cassette, cohesion is extremely
high even in layers that are visually characterized as
insufficiently solidified, i.e. in layers that are below the
level of free water.
It can be noted that the system has been working at
the limit of 40% of solid during disposal in cassette no.
2. In terms of the influence of dense slurry on the
TECHNICS MINING, GEOLOGY AND METALLURGY (2014)
characteristic of the mass disposed it may be concluded
that in cassette no. 2 slurry enables the appearance of
free water.
The moisture in cassette no. 2 is stable and without
significant influence of depth. Solidified ash is not
compressible, wet shows little bit increased com-
pressible while non-solidified ash is compressible. Re-
search area of cassette no. 2 comprises homogenous or
quasi homogenous layers of similar physical-mecha-
nical characteristics. The results shown indicate rela-
tively uniform parameters, i.e. that the degree of he-
terogeneous material disposed is significantly smaller
than in cassette no. 1. The ash disposed has great hori-
zontal stratification and anisotropy as a result of the
disposal method. Permeability coefficient of the mass
disposed in cassette no. 2 is lower than in cassette no.
1.
The angle of inner friction in cassette no. 2 is sig-
nificantly constant and slightly increases with sam-
pling depth. Cohesion in cassette no. 2 does not differ
significantly with the depth of disposal site, except for
the area 2, which is marked as a "solid compact ash",
and it differs from uniform values. Lower values of
cohesion are noted in the cassette no. 2. In ash disposed
in cassette no. 2 the presence of ettringite, portlandite,
magnetite, calcite and olivine has been found. The
most common phases are ettringite and portlandite.
It is obvious that the ash disposal system as a high
concentration slurry operates within narrow margins of
about 40% of solid phase. With those densities the
possibility of slurry solidification during transport has
been prevented. According to that, the higher the water
content is, the transport is easier, but keeping the sta-
bility on disposal site and its build up is more com-
plicated, because solidification is not occurring or it is
significantly slowed down.
By adding precisely determined water quantity
transport becomes sensitive and complicated but the
actual disposal and keeping the disposal site in stabile
state becomes technically easier and cheaper. The
benefits from using small amounts of water reflects in
inducing the process of spontaneous solidification by
using completely the active component which natu-
rally occurs in ash.
An unambiguous benefit is determined from the
application of the new method of ash preparation and
disposal in the form of high concentration slurry in
order to form the geotechnical and ecological stable
ash disposal site of TPP "Gacko".
REFERENCES
[1] Kulkarni D, Lokeshappa B, Kandarp S, Anil K D,
Mobilization of Metals from Indian Coal Fly Ash
under Dynamic Conditions, World Environment
2(1):1-6, 2012
41
S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...
[2] ASTM Standard C33, "Specification for Concrete
Aggregates," ASTM Inter., West Conshohocken, PA,
www.astm.org. 2003Melentjev V. A. (redaktor),
Zololakovnije materijali i zolootvali, Energija,
Moskva, 1978.
[4] Panteleev V. G., Sostav i svoystva zoli i laka TES,
Energoatomizdat, Leningrad, 1985
[5] Kneevi D., Ash disposal as high concentration slu-
rry, Elektroprivreda 3, Belgrade, pp. 107-111 (in
Serbian), 1988.
[6] Santhanam, C. J.; Ullrich, C. R. (eds.) Flue gas clea-
ning waste disposal and utilization, Noyes data cor-
porration, N. J., 1981
[7] Cotte P., Gilliam M., (editors), Environmental aspe-
cts of stabilization and solidification of hazardous
and radioactive waste, ASTM publication 1033, 1989
[8] Conner J. R., Chemical Fixation and solidification of
hazardous wastes, Van Nostrand Reinhold, New
York, 1990
[9] Jiang, W. X. Wu and D.M. Roy, Alkali-Activated Fly
Ash-Slag Cement Based Nuclear Waste Forms. MRS
Proceedings, 294 , 255, 1992
[10]Fan, L. T., Experimental Study of Stabilization/ Soli-
dification of Hazardous Wastes, Kansas State Uni-
versity, Great Plains/Rocky Mountain HSRC, EPA,
1992
[11]Parsa J, Munson-McGee S. H., Steiner R., Stabi-
lizationSolidification of Hazardous Wastes Using
Fly Ash, Journal of Environmental Engineering,
Volume 122, Issue 10, 935, 1996
[12]Shi C, Fernandez-Jimenez A., Stabilization/ soli-
dification of hazardous and radioactive wastes with
alkali-activated cements, Journal of Hazardous
Materials B137, Elsevier, 16561663, 2006
[13]xxx, Cementation of radioactive waste, NUKEM
Technologies GmbH, 2007
[14]Agamuthu P, Chitra S, Satoshi M, Hiroshi T, Soli-
dification and stabilization of Fly ash from mixed
hazardous waste incineration using ordinary Portland
cement, Environmental science, MY Tokio, 13, 5,
289-296, 2005.
42 TECHNICS MINING, GEOLOGY AND METALLURGY (2014)
[15]Anastasiadou K, Christopoulos K, Mousios E,
Gidarakos E., Solidification/stabilization of fly and
bottom ash from medical waste incineration facility,
J Hazard Mater., Elsevier, 15; 207-208:165-70, 2012
[16]Massardier V.; Moszkowicz P.; Taha M., Fly ash
stabilization-solidification using polymer-concrete
double matrices, European Polymer Journal, Else-
vier, Volume 33, Number 7, 1081-1086(6), 1997
[17]Hycnar J., Skladowanie otpadow z elektrowni, Gaz,
woda i technika sanitarna, No. 4, Katowice, 100-106,
1985
[18]Brankovi B., Kneevi D., Miloevi M., Joksi-
movi V., Monevski B., Opportunities and condi-
tions of ash disposal in Kostolac-B thermal power
plant, Rudarski glasnik, no. 1, Belgrade, pp. 53-59 (in
Serbian), 1991.
[19]Kneevi D., Practice and problems of ash and
bottom ash transportation and disposal in Serbia and
Montenegro, Proceedings of 1st congress of ash and
bottom ash disposal site, Obrenovac 19-27. (in
Serbian), 2005
[20]Kneevi D.: Possibilities and conditions for ash self-
solidification ash generated from Kosovo-B Thermal
Power Plant, Mining Science and Technology, No.1-
2, Vol.1, Belgrade, 19-36, 1996
[21]Jing Z, Matsouka N, Jin F, Yamasaki N, Suzuki K,
Hashida T, A Novel Method Of Advanced Materials
Processing, Solidification of coal fly ash using
hydrothermal processing method, J Mater Sci 41,
Springer, 15791584, 2006
[22]Kneevi D., Simovi I., Vujii Lj., Radoii M.,
Disposal problems of high calcium ashes, TPP Gacko
case, Proceedings of XII Yugoslav mineral proce-
ssing symp., Tuzla, pp.347-352 (in Serbian), 1989.
[23]Beatovi S., Analyses of ash hydraulic disposal on
"Gacko" TPP, BSc thesis, Faculty of mining and
geology, Belgrade, (in Serbian), 1998.
[24]Beatovi S. M., Influence of slurry preparation
methodology on deposited ash characteristics on the
example of thermal power plant "Gacko", Gacko,
Mgr`s thesis, Faculty of mining and geology, Bel-
grade, 2013 (in Serbian), 2013.
S. BEATOVI at al. INFLUENCE OF ASH SLURRY PREPARATION METHOD ON...

REZIME
UTICAJ NAINA PRIPREME HIDROMEAVINE NA KARAKTERISTIKE
DEPONOVANOG PEPELA IZ TERMOELEKTRANE GACKO

U proteklom periodu pepeo, sa visokim ueem kalcijuma, iz termoelektrane "Gacko" (Republika

Srpska) deponovan je u 2 kasete. Na 1. kaseti priprema hidromeavine obavljana je u priruno uraenom
horizontalnom bubnju sa mlaznicama, a na 2. kaseti meanjem suvog pepela i vode u kondicioneru. U
ovom radu su prikazani rezultati ispitivanja uzoraka iz obe kasete radi utvrivanja uticaja naina
pripreme hidromeavine na karakteristike deponovanog pepela. Dobijeni rezultati su pokazali da nain
pripreme hidromeavine bitno utie na karakteristike deponovanog pepela.
Kljune rei: deponija pepela TE "Gacko", pepeo sa visokim ueem kalcijuma, spontana solidifika-
cija, karakteristike deponovanog pepela

TECHNICS MINING, GEOLOGY AND METALLURGY (2014) 43

V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...

Morphological and Chemical Characteristics of the Placer Gold Deposits from

Meckin Dol, Locality of Borov Dol, Eastern Macedonia

VIOLETA K. STEFANOVA, University Goce Delev, tip Original scientific paper

Faculty of Natural and Technical Sciences, R. Macedonia UDC: 553.411(497.7)
TODOR S. SERAFIMOVSKI, University Goce Delev, tip,
Faculty of Natural and Sciences, R. Macedonia
GORAN K. TASEV, University Goce Delev, tip,
Faculty of Natural and Technical Sciences, R. Macedonia

This paper presents the latest results from gold panning which was performed along the creek Meckin
Dol. Research showed that the size of the tested gold aggregates ranges from 30 microns to about 600
microns, and the most frequently present form of gold aggregates was isometric - irregularform. The
chemical composition of gold aggregates shows heterogeneity with an average gold content ranging
from 74.15% - 99.32%, followed by silver with an average content of 0.3% - 24.68%,and an average
copper content of 0.04 to 11.06%. Other impurities such as Fe, Bi, Se, and Te are represented with
content less than 1%. In general it can be said that the examined gold aggregates, by their chemical
composition, are characterized by high purity ranging from 750 to 996.
Key words: gold morphology, flatness index, chemical composition

1. INTRODUCTION ailed schlich prospection of certain parts of the primary

Copper and gold occurrences in the Borov Dol lo-
cality and its vicinity were known as early as the 19th
century, which is reflected in the remains of old mining
and archaeological artefacts found. Old gutters along
the River Kriva Lakavica show the presence of gold in
the recent alluvium of the River Lakavica.
In the period from 1952 until today, the Borov Dol
site, its immediate vicinity and the connection of the
Borov Dol with the River Kriva Lakavica, have been
the subject of regional and detailed geological surveys
and studies that resulted in the definition of the copper
and gold porphyry type of deposit in the locality of
Borov Dol and the presence of placer gold in the
drainage systems of Borov Dol and along the River
Kriva Lakavica. Contributions to these numerous
studies are presented in the works of [6], [4], [11], [17],
[2], [13], [14], [15], [19], [12]. The results of these tests
instigated prospection research with det-
Authors address: Violeta Stefanova, University
Goce Delev, Faculty of Natural and Technical Sciences
tip, Goce Delev 89
Paper received: 28.01.2014.
Paper accepted: 28.03.2014.
and disintegrated mineralization material of the
narrower and wider environment of the copper and
gold deposits of the Borov Dol. The Meckin Dol creek
was selected as the best subject for our tests and study
after recognizing that this mineralized space, whose
spring area is in direct contact with the primary copper
and gold mineralization in the Borov Dol site, makes
the disintegrated material a potential carrier of eluvial-
proluvial-alluvial gold. Along the Meckin Dol creek,
certain sampling points were selected for gold panning
tests, where, during field work, a total number of 13
schlichs were sampled. The procedure for their further
treatment is described in detail in the section on the
methodology of the work. In the context of the above
mentioned, and taking these facts into account in the
framework of this paper, studies examining the rank of
the physical characteristics, such as size and shape of
the gold aggregates and their chemical composition,
were undertaken.
2. GELOGICAL CHARACTERISTICS OF THE
TERRAIN
The Borov Dol deposit is part of the mining region
Bucim-Damjan-BorovDol, that is, it occupies the

44 TECHNICS MININIG, GEOLOGY AND METALLURGY

V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...
southern parts. The metallogeny of this ore region is
closely related to the evolution of the Tertiary mag-
matism represented by sub-volcanicvolcanic facies of
andesite, latites, quartz latites, trachyrhyolites and so
forth, which are the product of intermediate to acid
calc-alkaline magmatism. Numerous interesting mi-
neralizations of Cu, Fe, Pb-Zn as well as Au mine-
ralizations are related to this magmatism. The geo-
logical structure (Figure 1) of the Borov Dol deposit,
which is quite complex, contains Paleogene, Neogene
and volcanogenic-sedimentary and volcanic rocks
[11], [17], [12].
Figure 1 - Geological setting from Bucim-Damjan-Bo-
rovDol ore district (Denkovski, 1995, suppl-
ment, Stefanova, 2005): 1. Paleogene, neoge-
ne quarter sediments, 2. Pyroclasts, 3. Ande-
site and latite, 4. Cretaceus flusch, 5. Carbo-
ntites, 6. Carbonate-schists series, 7. Granites,
8. Serpentinites, 9. Micasists, 10. Gneiss, 11.
Pb-Zn vein type of mineralization, 12. Fe
skarns type of mineralization, 13. Porphyry
type of mineralization
Paleogene sediments occupy a large part of the
terrain and are represented by conglomerates, Pale-
ogene flysch and series of tuffites and sands. Con-
glomerates are heavily modified and built mostly of
fragments of gneiss and quartz. Flysch facies are built
of thin layers of fine-grained and large-grained sand-
stone and conglomerates. The volcanogenic sediment
TECHNICS MININIG, GEOLOGY AND METALLURGY
series is built of sandstones, marly sandstone and li-
mestone, and pelitic tuffs and tuffites. Tuffs are dete-
rmined as andesite and their presence indicates that the
volcanic activity is synchronous with sedimentation of
Paleogenic sediments. Neogenic sediments, which are
represented by conglomerates, are built of hetero-
geneous material dominated by pieces of quartz, cry-
stalline shale and Paleogenic sediments.
Magmatic rocks are represented by volcanic and
sub-volcanic facies of latites, quartz latites and
andesites. Andesites in turn, depending on the degree
of change and ore amount are divided into: propilytized
andesites, hydrothermally altered andesites, hydrother-
mally altered and mineralized andesites and andesite
lava. Effusive rocks occur in the form of outcrops and
lava.
Large porphyritic propilytized andesites, that is,
latites and quartz latites, that occupy most of the te-
rrain, are hydrothermally altered and mineralized. La-
tites and quartz latites are products of the older phase,
hacked with dark grey fine-grained biotite amphibole-
andesite which, in the form of a neck, are imprinted on
the central part of the Borov Dol circular structure.
3. METHODS AND MATERIALS
Gold prospection was applied, from which mate-
rial of about 1015 kg from each sampling location
was taken. After panning, magnetic separation of the
magnetic fraction was undertaken manually. Both fra-
ctions obtained were observed under binoculars. The
gold aggregates found were separated manually and
subjected to further study.
To determine the morphological features of the
gold primarily the roundness, flatness, and serration
of the grains scanning electron microscopy (SEM)
was used. This type of analysis allowed the length of
the transport and nature of the environment in which
the pan material was deposited to be determined. The
SEM study was undertaken in the electron microscopy
laboratory within the Faculty of Stip University,
Macedonia.
Analyses were performed on a VEGA3 LMU.
The etalons were from TESCAN. The specific
operating conditions were
Tension: 20 keV
Test Method: EDS
Type of analysis: Quantitative
X-act: 10 mm2 (Slicon Drift detector)
Max resolution: 125 EV
Resolution of MnK, FK,CK: according to ISO
15632:2012
45
V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...

4. RESULTS AND DISCUSSION in some galenite was significantly represented. Less

Studies were performed along the Meckin Dol prevalent are sphalerite, malachite, azurite and
over a length of about 0.6 km (Figure 2). specularite. The non-metal minerals, zircon, mica, and
epidote are present.
Table 1. Results from pan material investigations

Sample No. Number of gold agregates

MD-1 5
MD-3 37
MD-4 3
MD-5 4
MD-7 4

The most important mineral from this prospection

Figure 2 - Locationof sampling sites on shlich samples
from locality MeckinDol
Out of the total of 13 samples, 5 samples were
taken for more detailed processing. We found a total of
56 gold grains of different shapes and sizes (Table 1).
The second mineral represented is chalcopyrite which
is the copper bearing mineralization in this locality.
Another metallic mineral in the material is pyrite, and
is gold. It is characterized by a beautiful yellow colour
and it has various sizes. Gold is found in elongated,
dendritic irregular shapes and less frequently in
spherical shapes. In order to precisely define and
determine the size and shape of the gold aggregates,
detailed SEM investigations were undertaken (Fig. 3).
From these examinations, it can be said that an
isometric form prevails (Fig. 3f) while elongated gold
aggregates (Figure 3a and 3b), spherical shapes (Figure
3b and 3c) and plate forms were also found (Figure 3e).

-elongated irregular gold grain b-elongated irregular gold grain c-Aggregate with spherical form

d-rounded form e- Aggregate with platy form fisometric gold grain

Figure 3 - Morphological forms of matching gold aggregates of Meckin Dol, Borov Dol locality
The physical characteristics of the gold, such as the function of the length of transport can serve as an
size of the grains, depend on the type of primary effective tool in determining the distance from the
mineralization, type and length of transport and the main source, that is, the identification of the proximity
erosion processes to which the terrain was subjected of the primary gold source. Apart from the form and
[8]. The morphology of the gold aggregates as a shape of the gold aggregates as a real indicator of

46 TECHNICS MININIG, GEOLOGY AND METALLURGY

V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...
length of transportation, the flatness index can be used,
which is calculated by the equation:(L+b) / 2t, where L
is the grain length, b is the width, and t is the breadth,
although the degree of roundness and the appearance
of edges are also a good indicator ([16]).
Studies of the forms of gold aggregates from the
tested locality show that irregular isometric shapes pre-
vail in the vicinity of the primary source, while down-
stream from the creek, as a result of transport, these
forms are lost, that is, the grains are elongated and
rounded.
At a distance of up to 50 m from the primary
mineralization, almost identical shapes as well as
Table 2. Morphological characteristics of Au grains associated with distance from source
Distance
General shape, outline-surface of gold aggregates
to source
Flatness index 1-2
to 50
Flatness index 2.1-3.3
50-300
The first group of grains (Table 2) is characterized
by a well preserved original form: square to re-
ctangular shapes, angular edges. The general shape of
the grains is irregular. They can often be found in the
elongated shapes. The flatness index ranges from 1 to
2. The second group is represented by aggregates with
a flatness index of 2.13.3 and are characterized by
slightly elongated shapes. The general appearance is
represented by a relatively regular shape.
It could be said that the external shape of the stu-
died grains of the locality varies widely, which, acco-
rding to [18], is due to the fact that primary grains show
a wide variety of initial shapes. Such characteristics
suggests that this gold probably has similar
characteristics as the primary gold, [10] and that it has
not suffered long transportation. Square or flat shapes
as well as elongated shapes found in the test locality
TECHNICS MININIG, GEOLOGY AND METALLURGY
primary gold prevail. At a distance of 50 to 300 m se-
condary grains become elongated, they even become
triangular in shape with the appearance of small voids
on the grain surface. At a distance greater than 300 m
the primary shapes of the grains are lost ([16] and [9]).
The flatness index of the examined gold aggregates
from the Meckin Dol creek ranges from 1.05 to 3.27.
Based on morphological characteristics, form, shape,
external appearance, and flatness index, two groups of
gold aggregates can be distinguished (Table 2). The
aggregates up to 50 m belong in the first group, and the
gold aggregates from 50 to 300m belong in the second
group.
are a feature of primary gold that typically occurs in
such shapes, although it sometimes may occur in
square to rectangular but very irregular shapes with
expressed sharp contours and edges on the grains [8],
[16], [7], [9].
Beside the shape, analyses of the chemical com-
position of the gold aggregates were undertaken, whi-
ch showed that it was native gold characterized by high
purity. According to [20] the gold with a purity of 900
950 is called high grade gold, and the gold whose
purity ranges from 800899 belongs to the group of
moderate grade gold. Given this grouping, it can be
said that (Table 3) most of the tested gold aggregates
belong to the group of high grade gold where the gold
content ranges from 92% to a maximum of 99%, and a
smaller amount with lower purity where the gold
content ranges from 71% to 84%
47
V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...

Table 3. Chemical composition of gold aggregates Meckin Dol, Borov Dol locality (%)
sample position of Fe Cu Ag Au Bi Se Te
analysis
periphery 0.25 0.09 0.23 99.42
MD-1
centre 0.72 99.21 0.07
grain-1
periphery 0.22 0.03 0.4 99.35
periphery 0.28 7.51 91.40 0.82
MD-1
centre 0.05 0.16 10.46 88.13 0.74 0.16 0.31
grain-2
periphery 0.03 0.18 4.83 93.75 0.74 0.47
periphery 0.3 99.12 0.5 0.07
MD-1
centre 0.01 0.74 99.25
grain-3
periphery 0.09 0.48 99.43
periphery 0.15 0.64 6.21 92.68 0.30
MD-3
centre 0.02 0.65 7.24 92.09
grain-1
periphery 0.29 0.14 9.55 89.52 0.2 0.30
periphery 1.98 0.46 3.07 94.48
MD-3
centre 1.61 5.29 92.80 0.29
grain-2
periphery 0.11 0.59 3.06 96.24
periphery 0.23 11.04 87.98 0.74 0.01
MD-3
centre 0.11 0.19 13.49 85.73 0.08 0.01 0.39
grain-3
periphery 0.30 0.25 13.23 86.20
periphery 0.11 26.91 71.56 0.53
MD-4
centre 25.91 73.24 0.42
grain-1
periphery 0.14 21.23 77.64 0.91
periphery 0.6 4.41 94.98
MD-4
centre 0.54 2.21 97.26
grain-2
periphery 21.52 78.48
periphery 0.14 2.28 97.58
MD-4
centre 0.59 2.39 97.02
grain-3
periphery 0.42 99.58
periphery 0.02 0.04 20.40 78.61 0.63 0.3
MD-5
centre 0.28 14.16 85.43 0.14
grain-1
periphery 0.03 0.05 9.92 89.97 0.04
periphery 0.32 10.18 89.10 0.4
MD-5
centre 0.01 0.34 7.59 91.83 0.13 0.1
grain-2
periphery 0.75 0.29 9.21 89.28 0.13 0.33
periphery 0.02 20.52 79.29 0.17
MD-5
centre 0.03 0.84 12.64 86.41
grain-3
periphery 1.01 0.31 21.89 76.02 0.4 0.36
periphery 0.1 21.65 77.53 0.41 0.32
MD-7
centre 0.3 19.30 80.19 0.21 0.08
grain-1
periphery 0.16 19.46 79.89 0.44 0.06
MD-7 periphery 0.21 0.19 22.84 76.43 0.06 0.27
grain-2 centre 1.36 0.26 25.78 72.26 0.34
periphery 0.18 0.49 11.13 88.20
MD-7
centre 0.24 0.03 12.91 86.73 0.09
grain-3
periphery 0.11 0.06 13 86.47 0.35

48 TECHNICS MININIG, GEOLOGY AND METALLURGY

V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...
Examination of the gold aggregates was under-
taken in peripheral areas and one analysis in the central
area in order to illustrate whether there was zonality in
the chemical composition. As the dominant admixture
in gold aggregates, silver is present in the rank ranging
Figure 4 - Histogram display the contents of the tested silver gold aggregates of Meckin Dol
In addition, regularity in the silver content was not
noticed. Some of the central gold aggregates are richer
in silver than those from the periphery, which can be
seen in Table 3 (MD-1 grain 2, MD-7 grain 2) while in
others reverse zonality is observed the central gold
aggregates are poorer in silver than those from the
periphery. In most of the tested gold aggregates there
is no regularity in terms of silver content.
Studies of the relationship between the composi-
tion of gold and the length of transportation showed
that there is no change in the microchemical record of
the alluvial gold during transportation ([3]).
As for other ingredients, the presence of Fe, Cu,
Bi, Se, and Te is determined. Iron content ranges from
0.01 to 1.98%; copper content from 0.03 to 0.84% and
other impurities, such as Bi, Se, and Te, are under-
represented and in some aggregates they are even
absent. Based on these data, it can be said that gold
aggregates in general have a homogeneous co-
mposition which leads to the assumption that gold
aggregates derive from one source ([5]).
5. CONCLUSION
The results of the investigations into the mo-
rphology of gold aggregates and the flatness index 1
2, from the locality of Meckin Dol suggests that there
is a primary deposit in the immediate vicinity which
has disintegrated to produce the material. The size of
the gold aggregates found ranges from 30 to about 600
microns. Based on tests it was determined that gold
aggregates commonly occur in irregular isometric
form, then in a round-spherical shape, and in the shape
of plates. The irregular shape is most common.
TECHNICS MININIG, GEOLOGY AND METALLURGY
from <1% to 26.91%. Fig. 4 shows a histogram
displaying the silver content in the tested gold
aggregates. In alluvial gold, the silver content can vary
from 32 to 50% (with an average fineness of 630), with
other elements not exceeding 1% [1].
Studies of the chemical composition showed that
the gold is characterized by high purity, which ranged
from 750 to 996. Of all the impurities, the silver con-
tent is the greatest from <1% to 26.91%. As for the
other ingredients, the Fe, Cu, Bi, Se, and Te content
was found to be low.
REFERENCES
[1] Bahna B, Smirnov A, Chovan M. & Bakos F River
transport induced hanges in chemical composition
of alluvial gold (documented on localities of Western
Carpathians).In:Geol.Carpath. (Bratislava), 53 (spec.
iss.), 105 107. 2002
[2] Bogoevski, K., Gold in Macedonia-geology, meta-
logenic, features, ore occurrences, gold ore deposits
and their evaluation. Special Issue No. 7 pp. 249,
1998
[3] Chapman R J, Leake R C, Moles NR, Earls G,
Cooper C, Harington K, Berzins R, The aplication of
Microchemical Analysis od Alluvial Gold Grains to
the Understanding of Complex Local and Regional
Gold Mineralizaction: A case Study in the Irish and
Scottish Caledonides, Economic Geology, Vol 95.
pp. 1753-1773, 2000
[4] Denkovski, G., Stefanova, G., Bandilov, L., Izvestaj
za slihovskata prospekcija I metalogenetskata pro-
spekcija na lakavickiot region. Strucen fond Bucim,
1995
[5] Florencia M, Southam G, Graig R James, Galliski M
A Morphological and Chemical study of placer gold
from the San Luis range, Argentina.The Canadian
Mineralogist, Vol. 42 pp. 169-182, 2004
49
V. STEFANOVA at al. MORPHOLOGICAL AND CHEMICAL CHARACTERISTICS OF THE...

[6] Ivanov, T., Metalogenetska studija na rudniot reon
Bucim-Damjan-Borov Dol. Strucen fond Bucim,
1982
[7] McClenaghan MB Indicator mineral methods in
mineral exploration; Geological Society ofLondon,
v. 5, p. 233245, 2005
[8] Mudaliar1 GG, Richards1 JP, Eccles DR Gold, Pla-
tinum and DiamondPlacer Deposits in AlluvialGra-
vels, Whitecourt, Alberta, EUB/AGS Special Report
089 (May 2007) pp 35, 2007
[9] Nakagawa M, Santosh M, Nambiar CG, Matsubara C
Morphology and chemistry ofplacer gold from
Attapadi Valley, southern India; Gondwana Rese-
arch, v. 8, no. 2, p. 213222, 2005
[10] . . , covered areas; Geological Society of London, v. 3, p.
, . 347, 1973
[11]Serafimovski, T., Metalogenija na zonata Lece-Ha-
lkidik, PhD, Stip, pp 380, 1990
[12]Serafimovski. V., Stefanova, V., Volkov. A. V.,
2010: Dwarf Copper-Gold Porphyry Deposit of the
Buchim-Damjan-Borov Dol Ore District, Republic
of Macedonia (FYROM). Geologiya Rudnykh
Mestorozhdenii, Vol. 52, No3, pp 203-220.
[13]Stefanova, V., Eluvijalno-deluvijalni pojavi na zlato
povrzani za tercieren magmatizam vo R. Makedonija.
Doktoska disertacija, pp.198, 2005
[14]Stefanova, V., Kovacev, V., Mladenov, V., Stani-
mirova, C., Eluvial-alluvial gold from gold-copper
occurrence Borov Dol (R. Macedonia). Part II: Mi-
neralogy of gold and stream sediments. Review of the
Bulgarian Geological Society, vo. 68, part 1-3, pp.
77-91, 2007
[15]Stefanova, V., Serafimovski, T., Tasev, G. and von
Quadt, A., Placer gold prospecting around the Te-
rtiary occurrences in the Republic of Macedonia.
(Eds. A. von Quadt & T. Serafimovski) Diversity of
copper and gold deposits in the Eastern Europe Ba-
lkan, Carpathian and Rhodopean belts: tectonic, ma-
gmatic and geochronological investigations; SCO-
PES Project-International Conference, May 29 - June
02, 2012 - Izgrev Hotel, Stip, Macedonia, pp. 33,
2012
[16]Townley BK, Herail G, Maksaev V, Palacios C, de
Parseval P, Sepulveda F, Orellana R, RivasP, Ulloa
C Gold grain morphology and composition as an
exploration tool: applicationto gold exploration in
2938, 2003
[17]Tudzarov, N., Metalogenija na naogalisteto na bakar
Borov Dol, doktorska disertacija, Stip pp. 195, 1993
[18]Utter T Rounding of Ore Particles from the Witwa-
tersrand Gold and Uranium Deposit (South Africa) as
an Indicator of their Detrial Origin, Journal of
Sedimenatry Petrology, V., 50, No 1, pp 71-76 (
Knight,et. all., 1994), 1980
[19]Volkov, A.V., Stefanova, V., Serafimovski, T., Si-
dorov, A.A. R., Native Gold of the Porphyry Copper
Mineralization in the Borov Dol Deposit (Republic
of Macedonia). Doklady Earth Sciences, Vol. 422,
No. 7, pp. 1013-1017, 2008
[20]Zaharova, E. M., Minerologia Rosipei, Nedra, Mo-
skva pp.191, 1994.

REZIME
MORFOLOKE I HEMIJSKE KARAKTERISTIKE NANOSNOG ZLATA SA MEKIN
DOLA, LOKALITET BOROV DOL, ISTONA MAKEDONIJA

U radu su prikazani najnoviji rezultati lihovske prospekcije zlata na lokalitetu Mekin Dol. Istraivanja
su pokazala da je veliina ispitivanih agregata zlata od 30 do oko 600 mikrona, kao i da je najei oblik
njihovog pojavljivanja izometrino-nepravilni. Hemijski sastav zrna zlata ukazuje na njihovu
heterogenost. Proseni sadraj je 74.15 99.32%, srebra od 0.3-24.68% i bakra od 0.04 do 11.06%.
Ostale primese (Fe, Bi, Se, Te) su zasupljene u koliini manjoj od 1%. Generalno, moe se zakljuiti da
su ispitivani agregati zlata hemijski cisti, na nivou od 750 do 996.
Kljune rei: morfologija zlata, indeks spljeskanosti, hemiski sastav

50 TECHNICS MININIG, GEOLOGY AND METALLURGY

 ISSN 0461-2531
TECHNICS

MECHANICAL
ENGINEERING
Mainstvo - Constructions mecaniques - Maschinenbau -

YEAR 63 - 2014.

EDITOR-IN-CHEF:
Prof. Boko Rauo, D.Sc, University of Belgrade, Faculty
of Mechanical Engineering, Belgrade

EDITORIAL COUNCIL:

Prof. Ljubodrag Tanovi, D.Sc, University of Belgrade,

Faculty of Mechanical Engineering, Belgrade
Prof. Jovan Vladi, D.Sc, University of Novi Sad, Faculty
of Technical Sciences, Novi Sad
Prof. Duan G r u d e n, D.Sc, F. Porsche
Aktiengesellschaft Stuttgart, Germany
Prof. Miroslav D e m i , D.Sc, University of Kragujevac,
Faculty of Technical Engineering, Kragujevac
Prof. Dragoljub i v k o v i , D.Sc, University of Ni,
Faculty of Mechanical Engineering, Ni
Prof. Radan D u r k o v i , D.Sc, University of
Montenegro, Faculty of Mechanical Engineering,
Podgorica
Prof. Zdravko Milovanovi, D.Sc University of Banja
Luka, Faculty of Mechanical Engineering, Banja
Luka

EDITORIAL OFFICE: Union of Engineers and

Technicians of Serbia, 11000 Belgrade, Kneza Miloa 7a/I,
Tel. +381/11/ 32 35 891, Fax +381/11/ 32 30 067
N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

Machine Learning of Radial Basis Function Neural Networks with Gaussian

Processing Units Using Kalman Filtering - Introduction

NAJDAN L. VUKOVI, University of Belgrade, Original scientific paper

Innovation Center, UDC: 004.85:004.032.26]:681.5.015.44
Faculty of Mechanical Engineering, Belgrade
ZORAN . MILJKOVI, University of Belgrade
Faculty of Mechanical Engineering, Belgrade
In this paper we analyze performance of Radial Basis Function network with Gaussian processing units
trained with algorithms based on Kalman filtering. Three distinct algorithms are developed and tested:
the first algorithm is extended Kalman filter, the second algorithm applies dual of Kalman filtering - the
Information filter, while the third algorithm uses unscented transformation to propagate first and second
moment of Gaussian random variable-the unscented Kalman filter. In paper we analyze advantages and
disadvantages of these learning algorithms when compared to conventional learning algorithms. We
analyze their basics so as to apply these algorithm for leaning of Radial Basis Function neural networks
with Gaussian processing units for real world engineering problems.
Key words: machine learning, neural networks, Kalman filter

1. UVOD Optimization of network parameters-machine

Neural networks are versatile mathematical and
software tool for modeling of problems for which we
are not able to explicitly build mathematical model on
first principles. During years of their development and
implementation, neural networks have been applied
for various problems in engineering practice [1], [2],
[3], [4]. However, for some problems in engineering
practice it is hard, if not impossible, to develop reliable
mathematical models using first principles. Neural
networks are especially suitable for modeling of these
real world problems.
To develop model based on neural networks one
typically needs to perform following modeling steps:
Define input and output variables of the system;
Choose of system for data acquisition;
Data acquisition;
Preprocessing of data;
Choose neural network (type of network, archi-
tecture, activation function, number of processing
units etc.);
Type of machine learning (supervised, unsu-
pervised or combined) and learning algorithm;
______________________
Authors address: Najdan Vukovi, University of Be-
lgrade, Innovation Center, Faculty of Mechanical
Engineering, Belgrade, Kraljice Marije 16
Papaer received: 07.05.2014
Papar accepted: 12.06.2014.
learning;
Validation of network performance;
Testing of network performance.
In the first step, it is assumed that we as engineers
are well acquainted with the problem; we ought to
know basic physical characteristics and dependences
between physical properties and parameters of
problem being analyzed. In the light of these facts,
steps #2 and #3 are uniquely defined. In the fourth step
we perform preprocessing of data so as to remove
possible noise, outliers and to perform scaling of input
and output values. Scaling is important because it
enables us to take all input and output dimensions of
problem and put them on the same scale; regardless of
their real numerical values, with scaling we take all
dimensions into consideration.
Having done this, we should choose type of neural
network (for example, feedforward or recurrent), nu-
mber of hidden layers (if we use multilayered net-
works), activation function etc. [1], [4]. In #6 and #7
we ought to define type of machine learning suitable
for our problem (supervised, unsupervised or com-
bined). From seventh step, we start iterative process in
which we perform learning of neural network using
training data and test its performance using test data
(steps #8 and #9). In this paper we analyze learning
algorithms for optimization of Radial Basis Function
networks with Gaussian processing units.

TECHNICS MECHANICAL ENGINEERING (2014) 53

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

2. RADIAL BASIS FUNCTION NETWORK WITH radial basis activation functions f xi . Based on their
GAUSSIAN PROCESSING UNITS
mathematical formulation we may conclude that
Radial Basis Function network with Gaussian activation of neuron given input vector xi nx is
processing units (GRBF) is given as [1], [2], [5]:
defined with vector ij , which defines difference
N
y i w ij f xi , j , j between input vector xi nx and output vector
j 1
(1) j nx .
The main parameters of GRBF networks are:
Weights of GRBF, also known as weight coeffi-
w ij
cients ;
j nx
Mean vector (prototype vector) , where
nx denotes number of dimensions of input vector,
nx dim x
i.e. ;
j ; j 1
Radial distance (spread) ;
f , ,
processing units (neurons); each neuron
has same activation function;
n Figure 1 Types of RBF activation functions: a)
xi nx and y i y represent input and output
Gaussian; b) Inverse quadratic activation
vetcors and they form training data set. function; c) Quadratic activation function; d)
If one decides to use Radial Basis Function Logarithmic activation function.
network (RBF) one can choose between following If we observe Gaussian and inverse quadratic
activation functions: activation function, we may conclude that activation of
Gaussian activation function: the neuron increases if difference ij between input

1
vector and prototype vector decreases; that is, if
f xi exp 2 Tij ij similarity between input and output vector decreases
2 j
(2) that means that new input sample x i is similar to
prototype vector j n , hence the neuron activation
x
Inverse quadratic activation function:
is increased.
f xi Tij ij 2j
1 2

(3) On the other hand, in case of quadratic and

logarithmic activation functions, one may conclude
Quadratic activation function: that neuron activation increases with increase of
difference between input vector and prototype vector.
f x i Tij ij 2j
12

(4) Therefore, in contrast to Gaussian and inverse

quadratic activation functions, quadratic and loga-
Logarithmic activation function: rithmic activation functions are active only if input ve-
f x i Tij ij ln ij ctor is sufficiently far away from center. Finally, we
(5) may observe that activation of neuron is always po-
sitive f xi 0, xi n .

x

where ij x i j ; ij nx stands for difference

between i -th input vector xi and j -th mean vector
j . Radial distance j defines width of activation
function, i.e. it defines boundary (spread) of one
prototype vector; everything within the boundary is the
part of the set represented with prototype vector: i -th
input vector xi is member of set represented with j -
th prototype vector j . Figure 1 shows four basic
In this paper we used Radial Basis Function
network with Gaussian processing units (GRBF). The
main reason for the choice of this network with
aforementioned activation function is in the ability of
this network to generate one prototype vector who can
represent a great number of input vectors; furthermore,
as stated earlier, activation of Gaussian unit increases
with decrease of difference between input vector and
prototype vector (and vice versa, activation decreases

54 TECHNICS MECHANICAL ENGINEERING (2014)

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...
with increase of difference between input vector and
prototype vector). It is important to emphasize the
main difference between the mechanisms how the
Multi-Layer Perceptron (MLP) neural network and
GRBF network model data; namely, MLP tries to
establish decision boundary between elements of a set
while GRBF tries to establish similarity between input
data and to find one vector which will represent a great
number of input vectors [1]. Having said that, in Figure
2 one may see GRBF for different values of radial
distance (spread) i .
Figure 2 Gaussian activation function for different
values of radial distance i : a) i 0,1 ; b)
i 1 ; c) i 5 ; d) i 10 . With increase
of i the support of activation function incre-
ases as well.
3. OPTIMIZATION OF PARAMETERS OF
RADIAL BASIS FUNCTION NEURAL
NETWORK WITH GAUSSIAN PROCESSING
UNITS-MACHINE LEARNING
The main advantage of GRBF network is in the
possibility to decouple optimization of activation
functions and weights. For example, one can apply
unsupervised learning [6], [7] to optimize prototype
vectors and radial distances using only input vectors
xi nx . Then, in the second step, we apply
supervised learning in which we use both input and
output values in learning process; both activation
functions and weights are optimized simultaneously
[7], [8]. Gradient descent is frequently applied [9], [10]
or recursive least squares [4]. One can perform
learning of unknown parameters and simultaneously
estimate needed number of neurons (processing units)
with respect to learning error [2], [11], [12], [13].
These methods typically generate network topologies
with optimal number of neurons (optimal in some
TECHNICS MECHANICAL ENGINEERING (2014)
previously defined sense) with increased precision and
prediction ability. In this paper we show and per-
formance of three learning algorithms for optimization
of GRBF network:
Extended Kalman filter (EKF);
Extended Information filter (EIF);
Unscented Kalman filter (UKF).
EKF is applied in [10] as learning algorithm. Ho-
wever, in this paper we perform sequential data
processing, in which learner has access to one and only
one training example. Furthermore, EKF serves as
standard against which you may assess performance
of other algorithms. To best of our knowledge, EIF has
not been applied for machine learning of neural
networks, which is the reason why we decided to test
its performance. During first decade of 21st century
UKF has gained much attention in community due to
its (relatively) simple implementation and increased
accuracy when compared to EKF. In this paper we test
performance of these three algorithms so as to see how
different theoretical concepts influence learning
process and network performance.
In this section of the paper we will introduce basic
mathematical notation and main assumptions. In the
first step, let us rewrite equation (1) in the following
form:
w10 w11 w1J 1
w
w21 w2 J f1 (xi , 1 , 1 )
yi (6)
20
wi 0 wi1 wij f (xi , J , J )
I.e. written in matrix form:
y i Wi j G J 1 (7)
where Wi j denotes matrix of all coefficients
(weight matrix), and G J 1 is activation vector of all
activation functions.
State vector is formed of all parameters of neural
network which need to be optimized in accordance
with previously defined optimization criterion.
Knowing that each row of weight matrix Wi j is a
vector, we may rewrite it in the following manner:
w10 w11 w1J w1T
w w21 w2 J w T2
Wij
20
wi 0 wi1 wij w Ti
(8)
To form the state vector, let us rewrite weight
matrix as:
55
N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

w w 1T w T2 w Ti
T
Algorithm 1 - Pseudocode of extended Kalman
(9) filter (EKF) for GRBF neural network opti-
mization.
Prototype vectors J and spreads J are the other
1. Algorithm EKF_GRBF ( k 1| k 1 , Pk 1| k 1 , x k , y k )
elements of the state vector, i.e.

E 0|0 0|0 0|0 0|0 P
T
T 2.E 0|0 0|0 ;

0|0
t 1T T
2 T
J T
1 T
2 T
J
(10) 3.Q diag dim ; R diag dim y k
Therefore, the state vector is formed of all weights 4. while Er Ermin
Wi j and parameters of processing units t : 5.
k | k 1 k 1| k 1

T 6. Pk | k 1 Pk 1| k 1 Q
wT tT
T
7.
y k g k | k 1 , x k
w w ... w ... ...
T
1
T
2
T
i
T
1
T
2
T
J
T
1
T
2
T
J 8. H k y w g(, ) g(, ) g(, )
i
(11) k

Dimension of the state vector is: 9. K k Pk | k 1HTk (H k Pk | k 1HTk R ) 1

10. k | k k | k 1 K k (y k y k )
N p dim
11. Pk | k (I K k H k )Pk | k 1
nneurons (dim(x) dim(y) 1)
(12) 1
y g
T
12. Er y g k | k , x k |k ,x
2
where function counts number of dim() 13. endwhile
dimension of its argument (in this particular case, input
14. return , P
vector x and output vector y ) and nneurons denotes
k |k k |k

Based on characteristics of the problem/system we

total number of neurons of the GRBF network.
need to set initial values of the state vector 0|0 , state
Having defined the state vector of GRBF neural
network, we may proceed and introduce EKF, EIF and covariance P0|0 , system noise covariance matrix Q,
UKF as learning algorithms. and measurement noise covariance R. Algorithm takes
Linear dynamical system in discrete form is current estimate of the state vector k 1| k 1 , state
defined with:
covariance Pk 1| k 1 and desired output y k . The-
k k 1| k 1
refore, desired output y k represents the measure-
y k g k | k 1 , x k
(13) ment. In line #5 and line #6 algorithm performs
prediction of the state vector and covariance. In line #7
g k | k 1 , xk denotes neural network. and are algorithm predicts measurement, which is the output
white noises ( ~ N , Q ; ~ N , R ); yk is of the neural network; in line #8, we need to calculate
desired output. Parameter optimization implies Jacobian of neural network y k g k | k 1 , x k with
identical mapping k 1| k 1 k . Let us make respect to the state vector . The Jacobian is cal-
following assumptions: culated because in EKF we linearize nonlinear equ-
ations (in this case measurement equation) with Taylor

E 0|0 0|0 0|0 0|0 P
T
E 0|0 0|0 ; series expansion. The filter gain is calculated in line
0|0
#9, while the correction step of the algorithm is
E k 0 ; E k Tk 1 Q k ,k 1 performed in line #10 for state vector and in line #11
E k 0 ; E k Tk 1 R k ,k 1
for state covariance. Innovation vector is defined as
difference between desired output y k and current
where k , k 1 is Kroneckers delta. In the reminder of
this section we introduce all three learning algorithms neural network output y k g k | k 1 , x k which is
based on Kalman filtering theory. the function of state vector k |k 1 and current input
3.1 Machine learning of GRBF neural network
using extended Kalman filter vector x k . Time index k denotes that in one itera-
tion algorithm uses one and only one example from
EKF for GRBF network optimization is shown
in Algorithm 1. training set x k , y k . Finally, having performed

56 TECHNICS MECHANICAL ENGINEERING (2014)

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

update step of the filter (correction), algorithm pro- inversion. Jacobian of neural network is calculated in
ceeds with calculation of current best estimate of the line #8. The correction step of the algorithm starts with

desired output g k | k , x . In the final step, we check calculation of information matrix k | k at k | k time
the convergence of the algorithm by calculating lea- instant and filter gain K k . In lines #9 and #11
rning error Er , which is the quadratic form of algorithm modifies state vector and information matrix
differences between current output of network for all using innovation y k y k ; learning error is
seen training examples and desired outputs.
calculated in line #12.
3.2 Machine learning of GRBF neural network Algorithm 2 Pseudocode of extended information
using extended information filter (EIF) filter (EIF) for GRBF neural network
Unlike EKF, the EIF uses different mathematical optimization.
form to parameterize Gaussian distribution. Instead of
1. Algorithm EIF_GRBF ( k 1| k 1 , Pk 1| k 1 , x k , y k )
mean and covariance, in information filtering one uses

1
information vector and information matrix, defined as

E 0|0 0|0 0|0 0|0
T
2.E 0|0 0|0 ; 0|0
follows:
P 1 (14) 3.Q diag dim ; R diag dim y k
where P represents the covariance matrix, which 4. while Er Ermin
tells us how much we do not know about our system; 5.
k | k 1 k 1| k 1
bigger P means more uncertainty, i.e. less knowledge

1
k | k 1 k 1| k 1 Q
1
of system/problem. If we let P than it means that 6.
we have no knowledge of our systems initial state.
Similarly, when P 0 than it means that we have all 7.
y k g k | k 1 , x k
available information about initial state of our system 8. H k y w g () g () g ()
i
k
(needless to say that we as engineers are aware that this
9. k | k k | k 1 HTk R k1H k
ideal situation is not possible). Now, from
K k k | k HTk R k1
1
computational perspective the problem arises when we 10.
have no initial knowledge of systems state, i.e.
11. k | k k | k 1 K k y k y k
P ; the problem is how to tell the computer this
1
y g
T
information. This is why estimation is performed in 12. Er y g k | k , x k |k ,x
information space; when we invert covariance matrix 2
13. endwhile
it is possible to tell the computer that we have little or
no knowledge of systems initial state, i.e. lack of 14. return , k |k k |k

knowledge is represented as:

P0
(15)
P 0
Therefore, when our knowledge is scarce ( P
) we may change our estimation space and move to
information space in which this lack of knowledge is
easily defined with 0 . This is the main reason of
EIF deployment in this paper; we would like to model
and test those problems in which designers have little
initial knowledge of the system.
Liner dynamical system is given by (13), while the
main assumptions hold as well. Pseudocode is given in
Algorithm 2. Algorithm is initialized in lines #2 and #3
where we define initial state vector 0|0 , initial
covariance P0|0 , as well as system noise Q and
measurement covariance R . In lines #5 and #6
algorithm predicts state, similar to EKF. The first
difference arises when algorithm predicts information
matrix (line #6); this operation implies double matrix
3.3 Machine learning of GRBF neural network
using unscented Kalman filter (UKF)
Linearization of nonlinear systems inevitable
results in errors because Taylor series expansion
cannot accurately model state transition. Furthermore,
Jacobians are calculated with our best current estimate
of the state vector (systems state) instead of using real
state vector. As a result, the filter optimality is
jeopardized and it may cause divergence.
Unscented Kalman filter (UKF) does not linearize
nonlinear systems; it is based on intuition that it is
rather easier to approximate known probability dis-
tribution than nonlinear state and observation models
[1]. UKF is based on a special kind of transformation
of random variable which we will introduce in the
reminder of the paper.
Suppose we are given an arbitrary nonlinear
function y f x ; let function argument x be a Gau-
ssian random variable (GRV) with known statistics x

TECHNICS MECHANICAL ENGINEERING (2014) 57

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

and P . Now, we need to calculate unknown statistics Having introduced basics of UKF, we may pro-
of function y after nonlinear transformation is applied. ceed with introduction of UKF as learning algorithm
In the first step, let us form new matrix , whose for GRBF neural network parameter optimization.
columns are called sigma points, in the following UKF is similar to basic form of EKF, but the main
manner: difference is in linearization of nonlinear equations,
which is not performed in UKF. The UKF learning
0 x algorithm is given in Algorithm 3.
i x dim x P , UKF algorithm is initialized in lines #2 and #3
i
where we need to define initial values of state vector
i 1,...,dim x
0|0 , state covariance P0|0 , process noise covariance
i x dim x P i dim x
, Q and measurement covariance R; all these are
i dim x 1,.., 2 dim x
determined in accordance with our knowledge of
(16) system. The first difference between UKF and EKF
is scalar determining the spread of sigma points i may be seen in line #7 where matrix of sigma points
and it is calculated as: is calculated (as stated, columns of this matrix are
called sigma points). In line #8 algorithm calculates the
2 dim x dim x network output using sigma points i . In line #9
(17)
algorithm calculates conditional expectation of y gi-
Other scalar parameters are 104 ,1 ;
ven current input vector x k and sigma points i .
0, dim x . Therefore, matrix is composed of
Learning proceeds with calculation of covariance
2dim x 1 sigma points. Having formed the sigma Py k y k in line #10, cross-covariance P y in line #11
k k

point matrix i , we apply nonlinear function f and filter gain K k in line #12. The update step of the
onto sigma points: filter is given in lines #13 and #14. Learning error is
calculated in last line of the algorithm.
i f i , i 0,..., 2dim x
(18) Algorithm 3 - Pseudocode of unscented Kalman filter
i is a new matrix formed of all numerical values (UKF) for GRBF neural network optimization.
1.Algorithm UKF_GRBF ( k 1| k 1 , Pk 1| k 1 , x k , y k )
function y f x can take knowing that function

E 0|0 0|0 0|0 0|0 P
T
argument x is GRV with mean x and covariance P . 2.E 0|0 0|0 ;
0|0

Mean (first moment) and covariance (second 3.Q diag dim ; R diag dim y k
moment) of y are: 4. while Er Ermin
2dim x
5.
y
i 0
Wi m i
6.
k | k 1 k 1| k 1

Pk | k 1 P k 1| k 1 Q
2dim x
7. i k | k 1 k | k 1 P k | k 1 P
Wi k i y i y
T
P
i 0 (19) 8. i , k |k 1 g i , x k
Constant weight coefficients are given as: 2dim x
9. y k
Wi m i , k |k 1 E g k |k 1 , xk

W
0
m
dim x i 1
2dim x
Wi k i , k |k 1 y k i , k |k 1 y k R
T

W
k
dim x 1 2 10. Py k y k
0 i 1
2dim x

W 1 2 dim x , y k
T
Wi m
i
k
11. P k
Wi k i k i , k |k 1
ky
i 1

i 1,..., 2dim x 12. K k P P

1
(20) k
ky
y k y k

The optimal value of parameter for Gaussian 13. k | k k | k 1 K k (y k y k )

distribution is 2 [1]. Unscented Kalman filter is 14. Pk | k Pk | k 1 K k Py k y k K Tk

1
y g
T
of higher accuracy than extended Kalman filter 15. Er y g k | k , x k |k ,x
because it performs nonlinear transformation of 2
16. endwhile
function y f x instead of linearization with Taylor
17. return , P
series expansion. k |k k |k

58 TECHNICS MECHANICAL ENGINEERING (2014)

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...
4. DISCUSSION
At the end of the paper we compare these three
learning algorithms built on basis of Kalman filtering,
with conventional approaches [4]. In conventional lea-
rning algorithms one can find constant parameters
known as learning rates, which purpose is to control
the speed of learning process. Analogously with this
parameter, in Kalman filtering based learning appro-
ach, process noise covariance Q has a direct influence
on covariance estimation; if we increase Q we tell filter
to take larger steps during learning. Similarly, when
we decrease Q it will result in smaller learning steps.
On the other hand, when we increase measurement
noise covariance R we tell the filter to put less em-
phasis on measurement y k and more on state pre-
diction. Therefore, with Q and R we can control the
increment algorithm takes during learning process and
filters confidence in measurements. These findings
are important because if measurements are erroneous
we are able to inform our learning algorithm of this.
Kalman filter can converge to local minimum
instead of global one, which is the familiar problem
with other iterative procedures as well. However, Kal-
man filter takes the second moments of state (Hessian
of objective function) and by doing this it takes local
curvature of error surface into consideration. The-
refore, Kalman filter is the algorithm of second order.
This is important fact because it enables our algorithm
to bypass local minima and search the global ones;
however, it should be stressed that it is not guaranteed
that algorithm will find the global minima.
In the second part of this paper, we show
experimental results of Kalman filter deployment for
optimization (machine learning) of Radial Basis
Function neural network with Gaussian processing
units. We test performance of three sequential algor-
ithms on four real world problems from engineering
practice.
5. ACKNOWLEDGMENT
This work is supported by the Serbian Government
- the Ministry of Education, Science and Techno-
logical Development - Project title: An innovative,
ecologically based approach to the implementation of
intelligent manufacturing systems for the production
of sheet metal parts (2011-2015) under grant
TR35004.
REFERENCES
[1] Vukovi, N., Machine Learning of Intelligent Mobile
Robot Based on Artificial Neural Networks. Doctoral
dissertation (in Serbian). University of Belgrade
Faculty of Mechanical Engineering, 2012. DOI:
10.2298/BG20120928VUKOVIC
TECHNICS MECHANICAL ENGINEERING (2014)
[2] Vukovi, N., Miljkovi, Z., A Growing and Pruning
Sequential Learning Algorithm of Hyper Basis Fun-
ction Neural Network for Function Approximation,
Neural Networks, 46C, pp. 210-226, (DOI:
10.1016/j.neunet.2013.06.004), 2013.
[3] Miljkovi, Z., Systems of artifical neural networks in
production technologies, University of Belgrade-
Faculty of Mechanical Engineering, 2003.
[4] Miljkovi, Z., Aleksendri, D., Artificial neural net-
workssolved examples with theoretical back-
ground (In Serbian). University of Belgrade-Faculty
of Mechanical Engineering, Belgrade, 2009.
[5] Vukovi, N., Miljkovi, Z., Babi, B., Bojovi,B.
Training of Radial Basis Function Networks with H
Filter Initial Simulation Results, Proceedings of the
6th International Working Conference Total Quality
Management-Advanced and Intelligent Approa-
ches, pp. 163-168, Belgrade, Serbia, 2011.
[6] Lazaro, M., Santamari I., Pantaleo C., A new EM-
based training algorithm for RBF networks, Neural
Networks, 16, pp. 69-77, 2003.
[7] Schwenker, F., Kestler, H. A., Palm, G., Three
learning phases for radial-basis-function networks.
Neural Networks, 1(4-5), pp. 439-58, 2001.
[8] Vukovi, N., Miljkovi, Z., Babi, B., ovi, N.,
Towards implementation of intelligent mobile robots
in a manufacturing environment, Proceedings of the
4th International Conference on Manufacturing
Engineering (ICMEN 2011), pp 367-376,
Thessaloniki, Greece, 2011.
[9] Karayiannis, N. B., Reformulated Radial Basis Neu-
ral Networks Trained by Gradient Descent, IEEE
Transactions on Neural Networks, 10(3), pp. 657
671, 1999.
[10]Simon, D, Training Radial Basis Function Neural
Networks with the Extended Kalman Filter, Neuro-
computing, 48, pp. 455-475, 2001.
[11]Huang, G. B., Saratchandran, P., Sundararajan, N., A
generalized growing and pruning RBF (GGAP-RBF)
neural network for function approximation. IEEE
Transactions on Neural Networks, 16(1), pp. 5767,
2005.
[12]Bortman, M., Aladjem M., A Growing and Pruning
Method for Radial Basis Function Networks, IEEE
Transactions on Neural Networks, 20(6), pp. 1039-
1045, 2009.
[13]Han,H. G., Chen,Q.L., Qiao, J. F., An efficient self-
organizing RBF neural network for water quality
prediction. Neural Networks, 24(7), pp. 717-725,
2011.
59
N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

REZIME
MAINSKO UENJE VETAKE NEURONSKE MREE SA RADIJALNIM
AKTIVACIONIM FUNKCIJAMA GAUSOVOG TIPA NA BAZI KALMANOVOG FILTRA-
OSNOVE TEORIJE

U ovom radu se analizira problem mainskog uenja vetake neuronske mree sa radijalnim akti-
vacionim funkcijama Gausovog tipa na bazi Kalmanovog filtra. Prikazana su tri nova sekvencijalna
algoritma mainskog uenja: prvi algoritam direktno primenjuje linearizovani Kalmanov filtar kao al-
goritam mainskog uenja, drugi algoritam primenjuje dual Kalmanovom filtru pod nazivom li-
nearizovani informacioni filtar, dok trei algoritam na poseban nain aproksimira prvi i drugi moment
Gausove raspodele. U radu se naglaavaju osnovne prednosti koje pomenuti algoritmi imaju u po-
reenju sa konvencionalnim vidovima mainskog uenja. Za sva tri algoritma razvijen je odgovarajui
matematiki model vetake neuronske mree sa radijalnim aktivacionim funkcijama Gausovog tipa.
Analizirane su osnovne postavke izvedenih algoritama u cilju njihove primene za sloene probleme u
inenjerskoj praksi.
Kljune rei: mainsko uenje, vetake neuronske mree, Kalmanov filtar

60 TECHNICS MECHANICAL ENGINEERING (2014)

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

Machine Learning of Radial Basis Function Neural Networks with Gaussian

Processing Units Using Kalman Filtering Experimental Results

NAJDAN L. VUKOVI, University of Belgrade Original scientific paper

Innovation Center, UDC: 004.85:004.032.26]:681.5.015.44
Faculty of Mechanical Engineering, Belgrade
ZORAN . MILJKOVI, University of Belgrade
Faculty of Mechanical Engineering, Belgrade
In this paper we test three new sequential machine learning algorithms for radial basis function (RBF)
neural network based on Kalman filter theory. Three new algorithms are derived: linearized Kalman
filter, linearized information filter and unscented Kalman filter. Having introduced and derived
mathematical model of each algorithm in the previous part of the paper, in this part we test and assess
their performance using standard test sets from machine learning community. RBF neural network and
three developed algorithms are implemented in MATLAB programming environment. Experimental
results obtained on real data sets as well as on real engineering problem show that developed algorithms
result in more accurate models of the problem being investigated based on radial basis function neural
network.
Key words: Machine Learning, Neural Networks, Kalman filter.

1. INTRODUCTION chine learning algorithms it is essential to test their

In the first part of this paper we developed and ana-
lyzed three distinct sequential machine learning algo-
rithms: extended Kalman filter (EKF), extended infor-
mation filter (EIF) and unscented Kalman filter (UKF).
In this part we show experimental results of their
performance for optimization of Radial Basis Function
network with Gaussian processing units (GRBF). All
codes are written and developed in MATLAB progra-
mming environment running on laptop (2.0 GHz CPU
with 2 GB of RAM). To fully assess accuracy and
precision of three Kalman filter based machine lear-
ning algorithms we tested their performance using the
following real world datasets:
Four real world data for testing machine learning
algorithms;
Modeling of 2D surface.
2. REAL DATA SETS FOR TESTING OF
MACHINE LEARNING ALGORITHMS
When one develops new or modifies existing ma
Authors address: Najdan Vukovi, University of Be-
lgrade, Innovation Center, Faculty of Mechanical Engi-
neering, Belgrade, Kraljice Marije 16
Paper received: 07.05.2014.
Paper accepted: 12.06.2014.
performance in terms of convergence, accuracy and
precision on datasets recognized widely in a machine
learning community. By doing this, the research co-
mmunity can perceive the advantages and disadva-
ntages of newly developed or significantly impro-
ved/modified algorithms. In this research, we decided
to use real world datasets given in Table 1.
50% of available data points are used for training
and the other 50% for testing of the network per-
formance. For all three algorithms we test GRBF arc-
hitectures with ten, 15 and 20 processing units (neu-
rons) in the hidden layer, i.e. network architectures are
given as:
Dataset #1: 8-[5]1-1; 8-[10]1-1; 8-[15]1-1; 8-[20]1-
1;
Dataset #2: 10-[5]1-1; 10-[10]1-1; 10-[15]1-1; 10-
[20]1-1:
Dataset #3: 8-[5]1-1; 8-[10]1-1; 8-[15]1-1; 8-[20]1-
1.
When referring to network architecture, the first
number denotes the number of dimensions of the input
vector, the third number shows the output number of
dimensions; the number in square brackets tells us how
many neurons are there in the hidden layer of the
GRBF network. This notation is developed in [4].

TECHNICS MECHANICAL ENGINEERING (2014) 61

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

Table 1. Characteristics of real world datasets

Num. of Num. of Num. of

Num. of Num. of
Num. Test set Training Dimensions Dimensions of
Examples Testing Pairs
Pairs of Input Output
1. Abalone 4177 2089 2088 8 1
Weather
2. 1609 805 804 10 1
Ankara
3. Auto-mpg 398 199 199 8 1
Root Mean Square Error (RMSE) statistics is used process is repeated 20 times, and at each time step the
to assess accuracy of network: new prototype vectors and network weights are
randomly initialized. Table 2 shows RMSE statistics
N
for Abalone dataset, Table 3 for Weather Ankara
y
2
i g w , , , x i 2 and Table 4 for Auto-MPG; as stated, RMSE sta-
s i 1

N (1) tistics is calculated over 20 independent trials.

RMSE is calculated using three datasets:
s test - RMSE for testing set (consists of 50% of
initial set but these data points are not used for
training);
s opt - RMSE for training set (consists of
remaining 50% of initial set but these examples
have been shown to the network during training
process);
s all - RMSE for all available data (RMSE is
calculated for all available data).
Prior to training, the input data are scaled into
1,1 while the output data to 0,1 . At the beginning
of the learning process, prototype vectors and GRBF
network weights are randomly initialized; having
initialized the prototype vectors and network weights,
these are sent to learning algorithm to modify them in
learning process using current data point. Learning
Table 2. Experimental results for Abalone test set.
The first dataset is highly nonlinear. According to
calculated RMSE statistics given in Table 2, one may
observe that expected value of RMSE statistics is
stest 0.1 . UKFs RMSE statistics is 0.0993 for GRBF
network with 15 processing units in the hidden layer.
Experimental results indicate that standard devi-
ation of RMSE statistics is small enough, which means
that learning algorithm found GRBF parameters with
high probability for repetition of results shown in
Tables 2-4. In literature, the common value of RMSE
test statistics is stest 0.08 , 0.09 .
In our research, we obtained slightly higher results
(~0.01), but as reader may see, these results are of the
same order of magnitude.
This means that GRBF network trained with
sequential EKF, EIF or UKF generates similar results
as the state of the art machine learning algorithms [5].

Num. of Processing units/neurons in the hidden layer

Alg. s
5 10 15 20
s test 0.09960.0039 0.10150.0061 0.10150.0034 0.10250.0030

EKF s opt 0.10090.0009 0.09910.0029 0.10040.0018 0.10130.0017

s all 0.10020.0019 0.10030.0037 0.10100.0021 0.10190.0018

s test 0.10140.0046 0.10010.0038 0.09990.0068 0.10110.0052

EIF s opt 0.10040.0024 0.09900.0035 0.09920.0031 0.09910.002

s all 0.10090.0034 0.09960.0035 0.09960.004 0.10010.0036

s test 0.09970.0040 0.09940.0048 0.09930.0045 0.10060.0034

UKF
s opt 0.09900.0016 0.09810.0020 0.09970.0026 0.10040.0014

62 TECHNICS MECHANICAL ENGINEERING (2014)

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

s all 0.09940.0021 0.09870.0031 0.09950.0031 0.10050.0020

Table 3. Experimental results for Weather Ankara test set.

Num. of Processing units/neurons in the hidden layer
Alg. s
5 10 15 20
s test 0.04150.0027 0.04160.0042 0.04190.0034 0.04020.0032

EKF s opt 0.03980.0010 0.04030.0006 0.03950.0009 0.03970.0009

s all 0.04070.0015 0.04100.0022 0.04070.0020 0.04000.0017

s test 0.04090.0029 0.04110.0030 0.04030.0026 0.04050.0032

EIF s opt 0.03980.0008 0.03970.0012 0.03950.0011 0.03980.0008

s all 0.04040.0016 0.04040.0017 0.03990.0017 0.04020.0019

s test 0.04050.0028 0.03930.0031 0.04040.0027 0.03950.0028

UKF s opt 0.03980.0009 0.03870.0018 0.03930.0013 0.03880.0017

s all 0.04020.0016 0.03900.0022 0.03990.0017 0.03920.0019

Table 4. Experimental results for Auto-mpg test set.

Num. of Processing units/neurons in the hidden layer
Alg. s
5 10 15 20
s test 0.09230.0074 0.09450.0105 0.09770.0095 0.08560.0095

EKF s opt 0.08720.0012 0.08710.0022 0.08720.0021 0.07560.0010

s all 0.08980.0043 0.09100.0057 0.09270.0053 0.08090.0049

s test 0.09320.0074 0.09460.0096 0.09330.0092 0.08600.0099

EIF s opt 0.08690.0032 0.08720.0022 0.08700.0024 0.07520.0010

s all 0.0902.0.0041 0.09110.0053 0.09040.0048 0.08090.0055

s test 0.08870.0080 0.09680.0073 0.09170.0060 0.08460.0065

UKF s opt 0.08710.0015 0.08710.0019 0.08680.0024 0.07590.0011

s all 0.09210.0039 0.09210.0039 0.08930.0031 0.08040.0033

In the second example (Weather Ankara test set-
Table 3) RMSE test statistics is stest 0.04 , while
standard deviation of RMSE calculated over 20
independent trials is quite small. Similarly to the
previous case, UKF outperforms both EKF and EIF in
terms of accuracy; UKF generates stest = 0.0387, while
EKF and EIF generate stest = 0.0397 and stest = 0.0395,
respectively. In [5] one may find values close to
stest 0.06 , which is higher than RMSE values obta-
ined in this research. The final experimental results are
given in Table 4 for Auto-mpg test set. RMSE values
are approximately stest 0.09 . In [5] one may find va-
lues of stest 0.1165 , which indicates that GRBF net-
work trained with UKF generates results with higher
accuracy and precision. Experimental results shown in
Table 2 for Abalone, Table 3 for Weather Ankara
and Table 4 for Auto-mpg demonstrate that RMSE
test statistics (RMSE for all available data s all , testing
set -50% of available data s test and training RMSE-
remaining 50% of available data s opt ) has the same

TECHNICS MECHANICAL ENGINEERING (2014) 63

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...
order of magnitude (or even smaller) of RMSE sta-
tistics when compared to other state of the art
algorithms.
3. IRIS TEST SET
Iris test set has been in focus of machine and
statistical learning community since 1930, frequently
used for assessment of accuracy and precision of
estimator for classification problem. Iris test set
consists of 150 examples classified into three classes,
where each class is represented with 50 distinct
examples. The input vector is four dimensional vector,
i.e. x 4 and the test set is shown in Figure 1.
Figure 1 Input data for Iris test set. On the main
diagonal one may observe the elements (com-
ponents) of the input vector. Each diagram
represents relation between two elements of
input vector (element xi with respect to the
element x j ).
Table 5. Experimental results for Iris test set.
Num. of Processing units/neurons in the hidden layer
Alg. s
5 10
Er 0.36350.0105 0.36110.0120
EKF
% 94.85190.0105 95.39130.0120
Er 0.36620.0168 0.36830.0132
EIF
% 94.66670.0168 95.10140.0132
Er 0.35820.0138 0.37130.0192
UKF
% 94.33330.0138 94.72460.0192
Iris test set is one of the most used real world test
sets for assessing classification accuracy. Efficient
predictor should reach accuracy of at least 95%.
Experimental results in Table 5 show that all three
algorithms achieve this level of accuracy. As in the
previous analyzed cases, UKF outperforms EKF and
EIF: maximum UKF is 96.1053%, maximum EKF is
95.3913%, and maximum EIF is 95.6491%. It may be
observed that classification accuracy increases when
the number of processing units increases.
64 TECHNICS MECHANICAL ENGINEERING (2014)
The input is normalized using expected value
(mean) x and standard deviation x , with the fo-
llowing equation: xi xi x x . The output data
is transformed into three dimensional vector y, so that
each element of the output vector y is binary variable
determining whether k-th input vector is member of j-
th class:
1 i j
y ik
0 i j (2)
By doing this the output vector has already been
normalized.
RMSE is not a good measure of accuracy of esti-
mator for classification problems (although one may
find a number of highly influencing research papers
showing classification accuracy of the tested algo-
rithms with RMSE criterion). However, in this paper
we used more intuitive accuracy measure. The pro-
blem of classification allows direct measure of accu-
racy: if GRBF network correctly classifies k-th exa-
mple into j-th class, than all we have to do is to
memorize whether that is valid; similarly, if GRBF
network classifies k-the example into j-th class altho-
ugh it does not belong there, we need to memorize this
information.
When this is done for all testing data, we may
easily calculate the percentage of data being co-
rrectly/incorrectly classified. This is more intuitive and
natural way to measure success of model for classi-
fication problems.
15 20
0.36190.0127 0.36450.0178
94.59650.0127 95.15790.0178
0.36350.0186 0.36030.0138
94.70180.0186 95.64910.0138
0.36270.0175 0.35710.0132
95.33330.0175 96.10530.0132
When directly compared to other research results
devoted to modeling of Iris test set [7], one may
conclude that EKF, EIF and UKF achieve the same
level (or little bit higher) of accuracy and precision.
4. MODELING OF 2D SURFACE WITH RADIAL
BASIS FUNCTION NEURAL NETWORKS
WITH GAUSSIAN PROCESSING UNITS
In this part of the paper we present experimental
results obtained using Radial Basis Function network
with Gaussian processing units for modeling of
N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

topology of surface [3]. The main idea behind this Developing of surface roughness models is
experiment is not to demonstrate inefficiency of current entirely different field of research. In our research we
advanced research methods in the field of surface treated this problem on simple basis: input values are
topology modeling, but rather to show the applicability related to output values through some unknown
of developed algorithms for modeling of topology nonlinear function. To find this unknown complex
using GRBF network. It should be stressed that in this function we employed GRBF network and trained it
setup we observe this problem as input-output problem: with three algorithms: extended Kalman filter,
there are some input values, and there are some output extended information filter and unscented Kalman
values; find function that achieves mapping from input filter.
to output space.
Experimental data are acquired using atomic force
microscopy given by 512 x 512 matrix; therefore, there
are 262144 training examples. However, this large
number of training examples means extensive time
needed for modeling. To solve this problem, we choose
mini-batch of 30x30 dimensions. All input data are
scaled into [-1,1] and output into [0,1]. As in previous
case, 50% of initial data (30x30=>900 training
examples) is used for training, while the remaining
50% is used for testing. We tested architectures with 30
and 50 processing units in the hidden layer.
Experimental results are shown in Figure 2. Table 6
shows Root Mean Square Error for all data calculated
over 10 independent trials.
Table 6. Experimental results for GRBF modeling of 2D a)
surface
Num. of Processing units/neurons in the
Alg. s hidden layer
30 50

EKF
s test 0.05600.0017 0.05940.0022

s opt
EIF 0.05560.0003 0.05790.0030

UKF
s all 0.05500.0004 0.05940.0061

As it may be seen in Figure 2, tested GRBF has

problem to recognize minor imperfections on the
surface. GRBF network trained with either EKF, EIF or b)
UKF generates smooth estimate of surface roughness.
Figure 2 Chosen mini-batch of 30x30 dimension: a)
It is interesting to observe that GRBF network with GRBF network with 30 processing units in the
smaller number of neurons in the hidden layer gene- hidden layer; b) GRBF network with 50
rates lower values of RMSE statistics. However, we processing units in the hidden layer.
point out that this could not be generalized because that
claim implies thorough theoretical investigation and 5. DISCUSSION
experimental validation. One may observe that EKF
Radial Basis Function Networks with Gaussian pr-
generates values of RMSE statistics which are higher
ocessing units can be applied for function approxi-
than EIF and UKF; we believe this is due to inability of
mation, classification, signal processing, process mo-
generic EKF to model highly unknown nonlinear
deling etc. [1], [2], [3], [4]. Having this in mind, it is
functions. Similarly, it is interesting to note that EKF
important to stress advantages and disadvantages of
generates lower values of RMSE statistics for GRBF
developed sequential learning algorithms built on
network with 50 processing units.
Kalman filter.

TECHNICS MECHANICAL ENGINEERING (2014) 65

N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...
Obtained experimental results demonstrate suffi-
cient level of accuracy needed to model real world pro-
blems in engineering. To fully assess applicability of
developed sequential machine learning algorithms, we
tested their performance in terms of optimization of
parameters of Radial Basis Function Neural Network
with Gaussian processing units (GRBF) using standard
datasets. In machine learning community, these data-
sets are used so as to test performance of newly deve-
loped algorithms and to compare their results with state
of the art algorithms in the field.
Experimental results confirmed that sequential
learning algorithms have the same level of accuracy
and precision as similar algorithms in the field (in some
of studied cases developed sequential algorithms even
slightly outperformed some of the well known
algorithms in the field). These results are of extreme
importance to assess performance of GRBF network
trained with tested learning algorithms.
Parameters of extended Kalman filter (EKF), exte-
nded information filter (EIF) and unscented Kalman
filter (UKF) such as initial state vector 0|0 , initial
covariance matrix P0|0 , process noise matrix Q and
noise covariance matrix R , define/determine initial
point of estimation, initial knowledge of systems state,
rate of change of systems state and the confidence our
system should have in measurements and acquired
information. When compared to parameters of deter-
ministic machine learning algorithms, parameters of
EKF, EIF and UKF have deeper meaning.
To be more precise, in this setup we estimate the
probability distribution of parameters rather than solely
treating them as deterministic. By doing things in this
manner, we try to deal with uncertainty in data. This
insight is important especially from the standpoint of
implementation of GRBF networks for real world
engineering problems.
Developed algorithms are sequential and enable on
line learning, in which learner (i.e. GRBF network in
this particular case) processes data one by one. This
ability has been exploited for simultaneous localization
and mapping of mobile robot [1].
Conventional approach in this field is based on
Kalman filter. To compensate unknown stochastic and
modeling influences, we integrated extended Kalman
filter and feedforward neural networks. Experimental
results and conclusions drawn from extensive
experimental study inspired us to enhance generic EKF
with feedforward neural network so as to enable
modeling of unknown influences and disturbances.
However, although these tested sequential learning
algorithms have proven their applicability for complex
66 TECHNICS MECHANICAL ENGINEERING (2014)
systems modeling, we ought to emphasize some of their
shortcomings as well.
One of the very first problems is related to
numerical instability during learning. Namely, due to
numerical imprecision inversion of matrices may
become impossible. One of the causes may be in direct
implementation of equations without taking care of
possible instability. This problem may be overcome
with artificial increase or decrease of error covariance
matrix Pk |k .
The other problem is related to convergence of
learning error. If problem is nonlinear (all real world
problems are nonlinear-Abalone, Weather Ankara,
Auto-mpg, Iris, 2D surface modeling) convergence of
error becomes an important issue. We noticed that
learning algorithm converges to one value of learning
error, which results in either ending of the learning
process or with increased number of iterations. Conve-
rgence of EKF, EIF, and UKF can be improved with
artificial increase/decrease of error covariance matrix
during learning process. In this manner, we may even
avoid to have algorithm being trapped in local minima.
The future development of EKF, EIF and UKF
sequential learning algorithms is devoted to ability to
change architecture of GRBF network during learning
process. Namely, to prevent rather tedious experime-
ntal procedure in which engineer/designer tests per-
formance of network with previously defined number
of processing units, we may develop a network with the
ability to add neurons and delete them if needed. It goes
without saying that this process of growing and pruning
of the network is to be performed during learning
process [2], [5], [8].
This kind of adaption would result in a decrease in
number of parameters needed to code piece of
information and consequently in faster learning.
Having this in mind, authors have already developed an
algorithm that enables growing and pruning of neurons
while actually performing learning of feedforward
neural network [2].
In [2], we analyzed one way of Gaussian
processing unit generalization; namely, radial distance
(refer to the first part of this paper) is set to be vector
nx rather than scalar 1 , which enables local
scaling of data along all dimensions during learning
process. For details and additional information in terms
of approximation of function, chaotic time series
modeling and identification of nonlinear systems, the
reader is kindly referred to [2].
From theoretical standpoint, it would be interesting
to see whether integration of unsupervised machine
learning [9], [10], [11] with already developed
sequential learning algorithms would result in higher
N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...
accuracy or/and faster convergence (for example, un-
supervised learning may be used to estimate the optimal
initial value of state vector 0|0 and error covariance
matrix P0|0 ). In this manner, we would extract
maximum amount of information from available data.
These improvements will be in focus in forthcoming
research.
Convergence analysis of Radial Basis Function
neural networks with Gaussian processing units trained
with EKF, EIF and UKF is given in [1]. The main
conclusion of convergence analysis is that given same
initial conditions, UKF converges faster than EKF and
EIF, i.e. it takes less number of iterations to reach
predefined accuracy defined by learning error Ermin .
The other conclusion is drawn for comparison of EKF
and EIF: EIF converges faster to previously defined
learning error than EKF.
In addition, we ought to note the influence that
parameters of sequential learning algorithms have on
network learning. Decrease of initial covariance matrix
P0|0 results in increase of number of iterations needed
for convergence. Experimental results given in [1]
indicate that increase in initial covariance matrix P0|0
decreases number of iterations needed for convergence
of algorithm. Based on thorough experimental study
given in [1] it may be concluded that increase in initial
covariance matrix may result in decrease of number of
iterations (faster learning) by order of magnitude. This
insight has tremendous importance for modeling of real
world engineering problems. For detailed information
in terms of convergence analysis the reader is referred
to [1] and [2].
Larger process noise covariance Q decreases
number of iterations needed for convergence. This in-
sight is logical ramification of basic Kalman filter se-
tup: prediction of covariance is addition of current best
estimate of covariance and process noise (the first part
of the paper, Algorithms 1, 2 and 3).
Similarly, larger measurement noise covariance R
increases the number of iterations needed for con-
vergence. Covariance of innovation vector is directly
influenced by R, and its increase decreases Kalman
gain K, which directly influences correction/update
step of the filter.
6. CONCLUSION
In this paper we tested and analyzed three sequ-
ential machine learning algorithms based on Kalman
filter for optimization of parameters of Radial Basis
Function Neural Networks with Gaussian processing
units (GRBF). The first step towards testing of algo-
rithms and building of GRBF network is to define
TECHNICS MECHANICAL ENGINEERING (2014)
learning problem as linear dynamical system in discrete
form. The first algorithm is based on linearization of
nonlinear observation model by expanding it into
Taylor series and it is known as extended Kalman filter
(EKF); the second algorithm exploits different para-
meterization in terms of canonical form of Gaussian
probability distribution and it is called extended
information filter (EIF). The third algorithm is based on
a particular nonlinear transformation of Gaussian
probability distribution known as unscented transfor-
mation and hence the name unscented Kalman filter
(UKF). In the first part of the paper, we showed deve-
loped mathematical models for all three algorithms
while in the second part we demonstrated their per-
formance when training GRBF network using well
known real world engineering problems.
All three sequential learning algorithms may be
applied for optimization of GRBF networks employed
for regression or classification. Development and tes-
ting of these sequential machine learning algorithms is
conducted while taking into account possibility to em-
ploy GRBF network for a wider set of engineering
problems. In other words, almost any type of real world
problem from engineering practice can be modeled
with GRBF network trained with one of these sequ-
ential learning algorithms. Experimental results sho-
wed that GRBF neural network trained with these three
algorithms generates better or at least as good results as
other state of the art algorithms.
7. ACKNOWLEDGMENT
This work is supported by the Serbian Government
- the Ministry of Education, Science and Technological
Development - Project title: An innovative,
ecologically based approach to the implementation of
intelligent manufacturing systems for the production of
sheet metal parts (2011-2015) under grant TR35004.
REFERENCES
[1] Vukovi, N., Machine Learning of Intelligent Mobile
Robot Based on Artificial Neural Networks. Doctoral
dissertation (in Serbian). University of Belgrade
Faculty of Mechanical Engineering, DOI:
10.2298/BG20120928VUKOVIC, 2012.
[2] Vukovi, N., Miljkovi, Z., A Growing and Pruning
Sequential Learning Algorithm of Hyper Basis Fun-
ction Neural Network for Function Approximation,
Neural Networks, 46C, pp. 210-226, (DOI:
10.1016/j.neunet.2013.06.004), 2013.
[3] Vukovi, N., Miljkovi, Z., Babi, B., Bojovi,B.
Training of Radial Basis Function Networks with H
Filter Initial Simulation Results, Proceedings of the
6th International Working Conference Total Quality
Management-Advanced and Intelligent Approaches,
pp. 163-168, Belgrade, Serbia, 2011.
67
N. VUKOVI at al. MACHINE LEARNING OF RADIAL BASIS FUNCTION NEURAL...

[4] Miljkovi, Z., Aleksendri, D., Artificial neural net- neural network for function approximation. IEEE
workssolved examples with theoretical background Transactions on Neural Networks, 16(1), pp. 5767,
(In Serbian). University of Belgrade-Faculty of 2005.
Mechanical Engineering, Belgrade, 2009.
[9] Lazaro, M., Santamari I., Pantaleo C., A new EM-
[5] Bortman, M., Aladjem M., A Growing and Pruning based training algorithm for RBF networks, Neural
Method for Radial Basis Function Networks, IEEE Networks, 16, pp. 69-77, 2003.
Transactions on Neural Networks, 20(6), pp. 1039-
[10]Schwenker, F., Kestler, H. A., Palm, G., Three
1045, 2009. learning phases for radial-basis-function networks.
[6] Fisher, R., The Use of Multiple Measurements in Ta- Neural Networks, 1(4-5), pp. 439-58, 2001.
xonomic Problems, Annals of Eugenics, 7, pp.179
[11]Vukovi, N., Miljkovi, Z., Babi, B., ovi, N.,
188, 1936. Towards implementation of intelligent mobile robots
[7] Simon, D, Training Radial Basis Function Neural in a manufacturing environment, Proceedings of the
Networks with the Extended Kalman Filter, Neuro- 4th International Conference on Manufacturing Engi-
computing, 48, pp. 455-475, 2001. neering (ICMEN 2011), pp 367-376, Thessaloniki,
Greece, 2011.
[8] Huang, G. B., Saratchandran, P., Sundararajan, N., A
geeralized growing and pruning RBF (GGAP-RBF)

REZIME
MAINSKO UENJE VETAKE NEURONSKE MREE SA RADIJALNIM
AKTIVACIONIM FUNKCIJAMA GAUSOVOG TIPA NA BAZI KALMANOVOG FILTRA
REZULTATI PRIMENE

U ovom radu su prikazani eksperimentalni rezultati primene tri nova sekvencijalna algoritma mainskog
uenja u cilju optimizacije parametara vetake neuronske mree sa radijalnim aktivacionim funkcijama
Gausovog tipa na bazi Kalmanovog filtra. Uvedena su tri nova sekvencijalna algoritma mainskog
uenja: linearizovani Kalmanov filtar, linearizovani informacioni filtar, algoritam specifine
aproksimacije momenata Gausove raspodele. Nakon prikaza odgovarajuih matematikih modela datih
u prvom delu ovog rada, u ovom delu razvijeni algoritmi su testirani u MATLAB programskom
okruenju razvojem odgovarajueg softverskog koda i primenom test skupova podataka. Iako svi
izabrani test skupovi podataka predstavljaju realne probleme, razvijeni algoritmi su testirani i na
realnom inenjerskom problemu modeliranja izgleda segmenta obraene povrine. Sva tri algoritma su
prilikom modeliranja ovih problema pokazala visok stepen tanosti.
Kljune rei: mainsko uenje, vetake neuronske mree, Kalmanov filtar

68 TECHNICS MECHANICAL ENGINEERING (2014)

 ISSN 0013-5836

TECHNICS

ELECTRICAL
ENGINEERING
Elektrotehnika Constructions lectrotechniques
Elektrotechnik

YEAR 63 2014.

EDITOR-IN-CHEF

Prof. Vujo D r n d a r e v i , D.Sc, Univesity of Belgrade,

School of Electrical Engeneering, Belgrade

EDITORIAL COUNCIL
Ljiljana H a d i b a b i , Energy Agency of the Republic
of Serbia, Belgrade
Rade F i l i p o v i , D.Sc, Thermal Power Plants Nikola
Tesla, Obrenovac
Prof. Ninoslav S t o j a d i n o v i , D.Sc, University of Ni,
Faculty of Electronic Engineering, Ni
Prof. Vladimir K a t i , D.Sc, University of Novi Sad,
Faculty of Technical Sciences, Novi Sad
Gojko D o t l i , M.Sc, Public Enterprise for Electric
Energy Transmission and Transmission System
Control ''Elektromrea Srbije'', Belgrade
Prof. Mili S t o j i , D.Sc, University of Belgrade, School
of Electrical Engineering, Belgrade
Prof. Ilija V u j o e v i , D.Sc, University of Montenegro,
Faculty of Electrical Engineering, Podgorica
Prof. Boris L o n a r, D.Sc, University of Belgrade,
Faculty of Technology and Metallurgy, Belgrade

EDITORIAL OFFICE: Union of Engineers and

Technicians of Serbia, 11000 Belgrade, Kneza Miloa 7a/I,
Tel, +381/0111 32 35 891, Fax +381/011/ 32 30 067
M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL

Statistical Delay Estimation in Digital Circuits Using VHDL

MILJANA Lj. MILI, University of Ni, Original scientific paper

Faculty of Electronic Engineering, Ni UDC:621.3.049.77:004.383.4 ; 004.438VHDL
VANO B. LITOVSKI, University of Ni,
Faculty of Electronic Engineering, Ni

The most important feature of modern integrated circuit is the speed. It depends on circuits delay. For
the design of high-speed digital circuits, it is necessary to evaluate delays in the earliest stages of design,
thus making it easy to modify and redesign a circuit if its too slow. This paper gives an approach for
efficient delay estimation in the describing phase of the circuit design. The method can statistically esti-
mate the minimum and maximum delay of all possible paths and signal transitions in the circuit, consid-
ering the practical implementation of circuits, and information about the parameters tolerances. The
method uses a VHDL description and is verified on ISCAS85 benchmark circuits. Matlab was used for
data processing.
Key words: VHDL, Monte-Carlo analysis, fanout, transition, timing analysis

1. INTRODUCTION the input vectors. To determine the maximum delay of

During the design of digital integrated circuits
(ICs) with limited time constraints it is extremely im-
portant to determine path delays of the signal in the cir-
cuit in time, since they determine the circuit operating
frequency. With modern IC design techniques, a relia-
ble determination of the circuit delays is possible after
in the final design stages. The delay that is extracted
using special computer programs using the data from
the final layout of the circuit. If the circuit is not fast
enough, it is necessary to redesign it, which increases
the design costs. In that way, we conclude that the spe-
ed of some circuit is its unknown property during its
design. Because of that, it is convenient to estimate the
circuit delays in some earlier stage of the design, when
the designer is able to recognize bad solutions and to
re-synthesize the circuit, saving time and money.
The simplest method for determining delays is
simulation [1]. SPICE simulation of complex circuits
on transistor level lasts too long [2]. To simulate com-
plex circuits, timing simulators have been developed,
that use simple models of the circuits time behavior.
By simplifying models of logic circuits we come to the
logical simulations, where the delay is dependent on
Authors address: Miljana Mili, University of Ni,
Faculty of Electronic Engineering, Ni, Aleksandra Med-
vedeva 14
Paper received: 23.01.2014.
Paper accepted: 10.03.2014.
a logic circuit is necessary to apply 2n possible combi-
nations of input vector, where n is the number of pri
mary inputs. This approach is not efficient for circuits
with too many inputs.
One approach to delays analysis that avoids the de-
pendence of input combination is the timing verifica-
tion [3]. Here, signal graphs are used to determine
maximal delay between inputs and internal nodes of
the circuit. This delay is referred to as topological de-
lay, because the topology of the circuit is analyzed, in-
stead of logic function of circuits elements. On the
other hand, some timing analyzers [4], [5], determine
circuit delays depending on a limited set of input vec-
tors. Synthesis of such test signals can be a complex
task, the increase of the number of inputs, exponen-
tially increases the number of possible combinations of
test vectors.
It was shown that such an analysis can give pessi-
mistic prediction of delays since some paths cannot be
activated by any of test vector. These are false paths
and are the result of logical redundancy in the circuit.
To enable the path senzitation, it is necessary that test
vector satisfies the following: 1) input signal should
bring rising or falling edge at the beginning of the path
and, 2) to enable the propagation of signals along the
path that is, to provide certain logical conditions at side
inputs of logic elements along the path. With false
paths these conditions are contradictory, and the path
cannot be activated. The precise delays of the circuits

TECHNICS ELECTRICAL ENGINEERING (2014) 71

M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL
become known only after the layout realization, when
dimensions of connections between elements are avail-
able. This requires a good understanding of the IC te-
chnology, and mastering the tools that extract delays
from the layout. It is clear that an important decision
should be made in the early design stages: whether to
wait the layout design of the circuit and then to deter-
mine its delays and speed using expensive tools, or to
estimate the delays, right after the logic synthesis of
circuits, using standard logic simulators, allowing the
fast correction of the description if necessary.
The hierarchical decomposition of the system [6],
can speed up and facilitate the verification of timing
requirements of the system. During this procedure the
largest part of the system is observed as a purely logi-
cal behavior. Reduced complexity obtained in this way
facilitates the identification of critical paths, and find-
ing tests that activate these paths. The disadvantage of
this method is its impossibility to consider the impact
of the tolerance parameters of system components.
There is also a restriction considering known set of in-
put vectors.
In the mass production values of the gates param-
eters are random in nature. Variations are the result of
the manufacturing process and the environment in
which the circuit operates. The influence of parame-
ters tolerances to the circuit response can be deter-
mined using statistical formulas. This implies data
from few hundred samples. The only delay analysis al-
gorithm that has the ability to simulate the actual pro-
duction process is the Monte-Carlo analysis - MCA
[1]. Beside the sampling methods used for the analysis
of the worst case delays and statistical analysis of the
delays, there are also direct methods, where one uses
formulas for mapping tolerance parameters in toler-
ance response. These methods work with small varia-
tions of the parameters, and one such method considers
time as a function of variation of parameters [7]. Alt-
hough this approach can be up to 200 times faster and
20% more accurate than the MCA, it may last unac-
ceptably long, since it requires the determination of the
sensitivity of logic gate to all possible variations in the
circuit
In this paper, new possibilities of the standard VH-
DL logic simulator to estimate the topological delay of
rising and falling edges of each signal in a digital cir-
cuit will be explored [8]. Using the information about
the fan-out of each gate, delay calculation becomes
more realistic. For statistical delay estimation a ran-
dom number generator is built into gates models. Dis-
tributions of worst-case delay values of logic circuits
are defined by their means and deviations. Determina-
tion of all circuit delays will be performed at the most
efficient way with only one MCA sampling which is
72 TECHNICS ELECTRICAL ENGINEERING (2014)
described in [9]. The entire procedure was imple-
mented in VHDL, using ISCAS benchmark circuits
[10]. Results confirm excellent performances of the
proposed method.
2. WORST-CASE DELAY ESTIMATION
The proposed method estimates delays of structur-
ally shortest and longest signal paths in the circuit. To
achieve it with only one run of the logic simulator, a
simultaneous i.e. parallel propagation of all possible
combinations of input vectors is applied, and the infor-
mation carrying delay values is accumulated while
passing through each logic gate, until it reaches a cer-
tain node or output [11]. Results of this process are the
maximal and minimal delays of rising and falling sig-
nal edges to all outputs in the circuit. The speed and
simplicity of the method, as well as the possibility of
implementation in VHDL, makes it very applicable in
first stages of the circuit design. Similar mechanisms
may also be used for the evaluation of other circuit
characteristics such as power consumption. The unique
circuit modeling can be applied in logical and timing
simulators.
Figure 1 shows the principle of delay estimation.
Both rising and falling signal transitions are simultane-
ously applied to circuit primary inputs, while their ini-
tial delays are set to zero. Each gate is described with
delays of rising and falling edges. The maximal delays
of rising and falling signal edges to the primary output
are calculated. The information about delays of sig-
nals transition is simultaneously updated while pass-
ing through each gate until the primary output are
reached. For example, the delay of the rising edge of
the last logic gate in Figure 1 is obtained from received
information about the delays from two input signals of
the gate (the delay on the rising edge of the input d is
2, while the delay of the rising edge at the input c is 0).
First, process selects the greater one, and adds it to the
rising edge delay of the last logic gate which is 2. The
resulting value of maximum delay of the rising edge at
the output will be 4.
2 = max {0, 0} + 2
3 = max {0, 0} + 3
2/3 4 = max {2, 0} + 2
(0) (0)
a (2) (3)
6 = max {3, 0} + 3
(0) (0)
2/3
b d
(4) (6)
(0) (0) e
c
Figure 1 Illustration of maximal delay estimation
It should be emphasized that the suggested method
can also estimate minimal delays of all transitions in
the circuit. This information is useful for determining
M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL
the minimal width of the clock impulse, and was al-
ways neglected in most of other delay estimation meth-
ods. Estimation of minimal delay for rising and falling
transition is similar to the maximal delay estimation
procedure. Using record type of a signal, enables them
to carry both logical state and other additional data of
the signal, which is the key issue that enables imple-
mentation of our method [12], [13].
3. MODIFICATIONS OF THE ALGORITHM
Including the information about fan-out of each
gate in the circuits increases the accuracy of the delay
estimation. This requires some modifications of the
basic algorithm. Descriptions of each gate, and a netlist
of digital system, need to be modified.
It is known that delay of a logic gate depends on
the number of gates that are connected to its output. If
its output is connected to two circuits, signal delay, ac-
cording to rough approximations will be doubled than
in the case of a single load.
The delay model of a gate has two parts. The first
calculates the maximal and minimal delays of rising
and falling transitions as described in Figure 1. The
second assigns delays depending on the fan-out of the
particular gate, which depends of the circuit netlist.
When a logic gate is instantiated in the netlist in, we
have to search for those signals that are connected to
the observed output. Then, we search through the netli-
st to determine the number of occurrences of this signal
in the circuit. Gates and signals have to be unilateral,
and one output signal can only be attached to inputs of
other gates.
After having the name of the output signal for
some gate, we count the number of appearances of that
name in the netlist, N0. Fan-out is calculated as f0=N0-
1, to exclude the output signal of the observed gate.
This number is then assigned to the instance of each
gate individually. In the case of VHDL this infor-
mation is described as a generic. The gate delay is then
multiplied with this number, and that value is used for
calculating the path delay. Implementation of this will
be described later.
The next section will show how the proposed
method with its modifications can be implemented in
the MCA loop to determine statistical delay.
4. STATISTICAL DELAY ANALYSIS
Statistical evaluation of delays using logic simula-
tor, requires adoption of statistical worst-case delay
models for all gates in the system. Such gates will not
have a fixed range of delays for rising and falling tran-
sitions, but will be generated randomly using random
number generator (RNG). If we neglect the variations
TECHNICS ELECTRICAL ENGINEERING (2014)
of the manufacturing process during the delay estima-
tion, it can happen that the maximal estimated delay is
not a true maximal delay. This reflects to significant
yield drop.
It should be noted that these are not real logic sim-
ulations, and it is not necessary to consider the logical
function of the gate, but only the information about the
delay. This makes the estimation method very fast.
These simulations do not require the stimuli, because
each simulation simultaneously considers all possible
combinations of input signals. In our case gate gener-
ates its maximal and minimal delays for rising and fall-
ing edges according to the Gaussian distribution, given
by:
exp[( p p )2 /(22p )]
( p)
[ p 2 ]
(1)
where p represents mean i.e. mathematical expecta-
tion, and 2 represents variance of the random variable
p [1]. The shape of this function can be modified.
Distribution functions are implemented in the
RNG for each gate, avoiding arithmetical operations
over some distribution. This accelerates the accumula-
tion of delays.
Considering the number of required simulations, it
should be sufficient to allow correct statistical analysis
of simulation results (e.g. 600). From the obtained val-
ues of the delay, the mean value of worst-case delays
for rising and falling edges can be determined from ex-
pression (2), while deviations can be calculated using
the equation (3) [1].
N
p [ pi ] / N
i 1 (2)
N
2 [ pi p 2 ] / N
i 1 (3)
Each logic gate is modeled with following param-
eters:
d1mn - delay of the output signal rising edge for
the shortest path,
d0mn - delay of the output signal falling edge for
the shortest path,
d1mx delay of the output signal rising edge for
the longest path,
d0mx delay of the output signal falling edge for
the longest path.
These parameters represent means of the distribu-
tion, while the deviation is set to 3%.
73
M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL
5. VHDL IMPLEMENTATION
The main idea of the method is to enable statistical
analysis of the delay in the earliest design stages. For
this, we use logic simulator. The most often logic sim-
ulator, and a hardware description language, is VHDL.
The availability and capabilities of this simulator are
the main reasons for such choice for the method imple-
mentation. For processing the obtained statistical data
(calculations of the mean, deviations delays and histo-
gram display) we have used Matlab.
entity nandg is
generic (ifo_outp: integer:= 1;
tpd_hlmn : time := 0.9 ns;
tpd_lhmn : time := 1 ns;
tpd_hlmx : time := 0.95 ns;
tpd_lhmx : time := 1.05 ns);
port (out1: out DCSM_std_logic);
in1, in2: in DCSM_std_logic);
end nandg;
architecture only of nandg is
begin
...
logic function process
p3:process(in1.d0mn,in1.d1mn,in1.arr0mn,
in1.arr1mn,in2.d0mn,in2.d1mn,in2.arr0mn,
in2.arr1mn)
variable multipl : real;
begin
multipl := real(ifo_outp);
r:= ((mulitpl*1.00) + (0.03*(gauss_rng)));
p:= ((mulitpl*0.90) + (0.03*(gauss_rng)));
if (in1.arr0mn or in2.arr0mn) then
out1.d1mn<=(min(in1.d0mn,in2.d0mn))
+(multipl*r);
out1.arr1mn <= true;
end if;
if (in1.arr1mn and in2.arr1mn) then
out1.d0mn<=(min(in1.d1mn,in2.d1mn))
+(multipl*p);
out1.arr0mn <= true;
end if;
end process p3;
p4:process(in1.d0mx,in1.d1mx,in1.arr0mx,
in1.arr1mx,in2.d0mx,in2.d1mx,in2.arr0mx,
in2.arr1mx)
variable multipl : real;
begin
multipl := real(ifo_outp);
r:= ((mulitpl*1.05) + (0.03*(gauss_rng)));
p:= ((mulitpl*0.95) + (0.03*(gauss_rng)));
if (in1.arr0mx or in2.arr0mx) then
out1.d1mx<=(max(in1.d0mx,in2.d0mx))
+(multipl*r);
out1.arr1mx <= true;
end if;
if (in1.arr1mx and in2.arr1mx) then
out1.d0mx <=(max(in1.d1mx,in2.d1mx))
+(multipl*p);
out1.arr0mx <= true;
end if;
end process p4;
Figure 2 Part of the VHDL code modeling two-input
NAND gate
At the beginning of the algorithm implementation,
we need to choose signal types. Signals that occur in
digital circuits should not only carry information about
the logical states of individual nodes, but also infor-
mation about the accumulated delays for each signal.
The type of such a signal S is record, and beside the
information about the accumulated delays (they are
74 TECHNICS ELECTRICAL ENGINEERING (2014)
similar to the parameters that define the delay of indi-
vidual logic gates) d0mn (S), d0mx (S), d1mn (S),
d1mx (S), it contains the following flags:
arr1mn denotes the arrival of the rising edge de-
lay to determine the shortest path,
arr0mn denotes the arrival of the falling edge de-
lay to determine the shortest path,
arr1mx denotes the arrival of the rising edge to
determine the longest path,
arr0mx denotes the arrival of the falling edge to
determine the longest path.
It is assumed that the circuit is described at the
structural level, and that ranges of each transition de-
lays are known for each gate.
VHDL description of the two input NAND gate
that describes the fan-out dependence of the delay is
shown in Fig. 2. It contains generics which define
worst-case delays. The generic ifo_outp represents
fanout of the gate, and has the unit initial value, while
it real value is assigned during the instantiation in the
circuit.
-- Circuit structure:
g10gat: nandg
generic map (0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s10, s1, s3);
g11gat: nandg
generic map (0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s11, s3, s6);
g16gat: nandg
generic map (0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s16, s2, s11);
g19gat: nandg
generic map (0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s19, s11, s7);
g22gat: nandg
generic map (0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s22, s10, s16); -- global output
g23gat: nandg
generic map (0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s23, s16, s19); -- global output
a)
-- Circuit structure:
g10gat: nandg
generic map (1, 0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s10, s1, s3);
g11gat: nandg
generic map (2, 0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s11, s3, s6);
g16gat: nandg
generic map (3, 0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s16, s2, s11);
g19gat: nandg
generic map (2, 0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s19, s11, s7);
g22gat: nandg
generic map (1, 0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s22, s10, s16); -- global output
g23gat: nandg
generic map (1, 0.9ns, 1.0ns, 0.95ns, 1.05ns)
port map (s23, s16, s19); -- global output
b)
Figure 3 Netlist a) before b) after modification
Process p3 calculates minimal delay of the rising
and falling edges of the gate, while p4 simultaneously
calculates their maximal delays. For the particular gate
type, each falling transition of input signal generates a
rising edge at the output of the gate, while the rising
M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL

edge at one of gates inputs causes the falling transi- apply following replacements to the selected pares
tion, only if the rising edge appeared at the second in- (v1, v2)
put of the circuit [14], [15].
2 lnR 2 ln R
Figure 3, shows netlists of the benchmark circuit XS v1 X P v2
ISCAS c17, before a), and after b) the described mod- R i R (5)
ifications. After the modification primary outputs of VHDL implementation of the delays RNG is sho-
the system have zero fanout values. They need to be wn in Figure 4. Function rand, generates normally
set to the value of one, which corresponds to the unit
distributed random numbers in the interval [0, 1] [17].
load of the primary outputs. Now, a standard logic sim-
ulation or statistical delay estimation can be per-
formed.
The fanout information is now present in the mod-
ified circuit netlist. To modify it, a simple C program
is written. It analyzes the netlist as a text file and per-
forms text searches. In our case, this modification is
facilitated by the fact that the verification algorithm
uses ISCAS benchmark circuits, where port and ge-
neric maps are very regular. In each instance of the
gate, first, its output signal is given. Then the program
searches for the number of occurrences of that signal
in the rest of the netlist. The resulting number, when
decremented, gives the real fanout value of the ob-
served gate, and is overwritten over its initial value. Figure 5 Histogram of delays for the NAND gate
function gauss_rng return real is Verification of the model for the NAND in Fig. 2
variable u1, u2, v1, v2, r, q, p: real;
begin implies 600 simulations of the circuit. The obtained re-
loop u1:=rand;
u2:=rand; sults represent the random generation of worst-case de-
v1:=u1*2.0 - 1.0;
v2:=u2*2.0 - 1.0;
lays for rising and falling edges. These values are
r:=v1*v1 + v2*v2; memorized as text, and then processed in Matlab. For
exit when r<1.0;
end loop; circuits with a small number of outputs, results are pre-
q:=log2(r);
p:=(sqrt((0.0-2.0)*q/r))* v1; sented as a histogram, and are shown in Fig. 5. The
return p;
end function gauss_rng; mean of the delay is 1ns, and the corresponding devia-
tion is 3%. The horizontal axis shows the value of de-
lay, while the vertical axis shows numbers of delay oc-
Figure 4 VHDL implementation of RNG with a nor- currences in a particular delay range.
mal distribution
Composite types of signal are used in larger sys-
A statistical delay evaluation is enabled by gener- tems. They carry information about the accumulated
ation of random worst-case delays for each transition delays until the particular moment, i.e. up to the ob-
in the logic gate. For VHDL implementation of this
served logic gate (real data type), and the infor-
idea, a RNG with a normal distribution was developed.
mation about the arrival of the particular transition at
VHDL simulator can generate random numbers with
the output of the gate (Boolien data type). If the in-
uniform distribution on the interval [0, 1]. In order to
obtain normal distribution, the following procedure formation about the arrival of the transition is true, de-
should be applied [16]: lay calculation begins, resulting with ranges of worst-
case delays (they depends on the previously arrived de-
select two numbers u1 and u2 which are uniformly lay information about different transitions, and of
distributed on the interval [0, 1] newly generated delay of the logic gate).
map the interval [0, 1] into the interval [-1, 1] us-
ing the following replacements: v1=2u1-1 i To simulate and analyze logic circuits, it is neces-
v2=2u2-1 sary to write a VHDL testbench program where the
system netlist is instantiated several hundred times us-
calculate R using the following formula:
ing a loop. Input signals and their responses are mem-
R v12 v22 orized as matrixes.
(4)
among all pares (v1, v2) obtained in this way, select 6. RESULTS
those that satisfy R<1 The described method was verified on ISCAS'85
benchmark circuits. Table 1 shows the results obtained

TECHNICS ELECTRICAL ENGINEERING (2014) 75

M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL

after estimates worst-case delay ranges without consid- After the selection of the IC manufacturing tech-
ering fanouts of logic gates. Table shows: the number nology, when technical information on logic gates is
of signals, the number of outputs, number of logic available, these values can be prcised and corrected.
gates, the maximal topological level of the signal, and, Results in table represent estimation of delays before
the minimal and maximal delays of falling and rising and after (marked with ') including fanouts. If we ana-
edges of the signal. lyze these results, we can conclude that the load of a
Table 2 gives the results of an analysis of the same logic gate is not purely capacitive, since the delay is
circuit, as well as their change (in percentage) when not a linear function of fanout value. Loads of gates
taking into account fanouts. In Table 3, shows the com- due to metal connections have resistive components.
parison of delays for ISCAS benchmark C432. The in- With this information one can develop a delay model
itial worst-cases delays for all gates are the same and of a gate that nonlinearly depends of the fanout value.
equal to tfmn = 0.9ns - minimal delay of falling edge, It is important to emphasize that each pass of the sim-
trmn=1ns - minimal delay of rising edges, ulator through the circuit takes less than a second, even
tfmx=0.95ns-maximal delay of falling edge, and trmx for the most complex ISCAS'85 benchmark circuit.
= 1.05ns - maximal delay of rising edge. Their values This proves high efficiency of the present method, and
are selected to facilitate the verification of delays esti- thus guarantees and encourages its use as a tool for ac-
mation method (unit value of the delay in nanoseconds, celerating the MNA, where accuracy is ensured by re-
with a tolerance of 5%). peating the simulation several hundred times.
Table 1. Worst-case delay estimation without including fanout information
Circuit No. of signals No. ins/outs No. of gates Topol.level Dfmn [ns] Dfmx [ns] Drmn [ns] Drmx [ns]
c17 11 5/2 6 3 1.9 2.9 1.9 3
c432 196 36/7 160 17 1.9 16.9 1.9 17
c499 243 41/32 202 11 0.9 11.3 1 11.4
c880 443 60/26 383 24 1.8 24 2 24.2
c1355 587 41/32 546 24 2.8 24.1 2.9 23.9
c2670 1426 233/140 1193 32 0 32.3 0 32.4
c3540 1719 50/22 1669 47 1.8 47.5 1.9 47.7
c5315 2485 178/123 2307 49 0.9 49.3 1 48.6
c6288 2448 32/32 2416 124 0.9 124 1 124
c7552 3719 207/108 3512 43 0 42.9 0 43

Table 2. Estimated delays and their increments, when fanouts are calculated for ISCAS85 circuits
Circuit Dfmn [ns] fmx [%] Dfmx [ns] fmn [%] Drmn [ns] rmx [%] Drmx [ns] rmn [%]
c17 1.9 0 6 500 1.9 0 6 200
c432 4.6 142 53 214 2.8 47 53 212
c499 0.9 0 29.3 159 1 0 29.4 158
c880 1.8 0 50.1 109 2 0 50.3 108
c1355 2.8 0 53.9 120 2.9 0 54.1 126
c2670 0 0 80.5 149 0 0 80.6 149
c3540 1.8 0 93.8 97 1.9 0 94.8 99
c5315 1.8 100 89.9 82.3 1 0 88.2 81.5
c6288 5.5 511 262 261 6.8 580 262.7 112
c7552 0 0 101.1 136 0 0 102 137

Table 3 Estimated ranges of worst-case delays before and after including fanouts for C432 circuit
Output Dfmn [ns] Dfmn ' [ns] Dfmx [ns] Dfmx' [ns] Drmn [ns] Drmn ' [ns] Drmx [ns] Drmx' [ns]
1 2.9 5.9 4 7.15 2.8 5.5 4 4.95
2 3.8 6.8 8 22.8 3.8 6.5 8 20.7
3 5.7 6.7 12 35.5 5.7 6.6 12 34.4
4 2.8 2.8 16 52.1 2.8 2.8 16 51.8
5 1.9 4.9 16.95 53 1.9 4.6 17.05 53
6 1.9 4.9 16.95 52.9 1.9 4.6 17.05 53
7 1.9 4.6 16.95 53 1.9 4.6 17.05 53

76 TECHNICS ELECTRICAL ENGINEERING (2014)

M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL

The entire MCA delays estimation and for the most Table 5. Standard deviations of delays for C3540 depend-
complex benchmark circuit lasts no longer than a few ing on the signal topological level
minutes. A personal computer with modest performa- Output t.l.mn mn_f mn_r t.l.mx mx_f mx_r
nces is used, having the AMD Athlon processor run-
ning at 1.5 GHz, and 1 GB of RAM. 1 3 0.0597 0.0625 4 0.0685 0.0655

The use of VHDL simulator Aldec Active-HDL, 2 2 0.0506 0.0507 4 0.0694 0.0703
eliminates the need for expensive computers and tools. 3 4 0.0814 0.0705 8 0.0930 0.0860
Table 4 illustrates the effectiveness of delay estima- 4 10 0.1009 0.1081 13 0.0941 0.0998
tion. It shows the duration of the simulation and the 5 10 0.1098 0.1150 15 0.1181 0.1124
corresponding memory usage for ISCAS benchmark 6 5 0.0592 0.0588 15 0.1141 0.1139
circuits. Values are given for 100 simulation/circuits 7 7 0.0690 0.0739 18 0.1249 0.1210
samples, where one simulating simultaneously estima- 8 2 0.0511 0.0533 22 0.1595 0.1613
tes minimal and maximal delays of rising and falling 9 5 0.0728 0.0692 24 0.1733 0.1749
edges for each signal in the circuit, while considering 10 6 0.0821 0.0835 25 0.1722 0.1693
their fanouts. 11 6 0.0858 0.0826 28 0.1901 0.1937
Table 4. Durations of simulations and a memory oc- 12 5 0.0773 0.0831 32 0.1806 0.1874
cupation 13 6 0.0842 0.0875 34 0.2131 0.2052
14 6 0.0860 0.0836 33 0.1972 0.2080
ISCAS circuit CPU time [s] memory requirements [MB]
15 6 0.0870 0.0861 33 0.1800 0.1852
c17 2 3.9
16 6 0.0871 0.0809 33 0.2054 0.1933
c432 62 68.7
17 6 0.0863 0.0784 35 0.1977 0.2100
c880 160 152.6 18 6 0.0883 0.0848 35 0.2149 0.2086
c1355 283 178.8 19 8 0.0808 0.0834 37 0.2077 0.2133
c3540 817 382.6 20 2 0.0521 0.0504 39 0.2207 0.2054
21 7 0.0872 0.0878 47 0.2114 0.2323
When a complex digital circuit is simulated 600
22 14 0.1142 0.1348 47 0.2460 0.2393
times, a huge amount of data is obtained. One simula-
tion consists of 4 parallel processes - for each transition 0.24
of the signal, having 4 * 600 = 2400 simultaneous sim- 0.22
d
ulations executed. e
0.20
v
On the other hand, each gate generates random va- i 0.18
lues for all types of worst-case delays, which requires a
t 0.16
additional time. The obtained data should be presented i
o 0.14
in an appropriate manner. n
0.12
For a circuit with a small number of outputs, the
0.10
data can be presented as a histogram. ISCAS C17 has
0.08
only two outputs and it is possible to generate histo-
grams of maximal and minimal delays for rising and 0.06
0 5 10 15 20 25 30 35 40 45 50
falling edges for each output (8 histograms). topological level

However, the circuit C3540, has 22 outputs, and a
histogram is not a good solution. For such cases, a spe-
cial Matlab program is created. Using the observed
data, program calculates deviation of the worst-case
delays of each signal transition for all outputs of the
circuit.
These results are shown in Table 5. It shows devi-
ation values of the minimal and maximal delays for
falling edge of the output signal - mn_f mx_f, minimal
and maximal delays for rising edge of the output signal
- mn_r mx_r, and the minimal and maximal topologi-
cal levels of the output signals. Using these results, we
can analyze the influence of the path length on the de-
viation of delay for one output signal. This dependence
is shown in Figure 6.
Figure 6 Dependence of delay deviation from the
topological level of the signal for C3540
7. CONCLUSION
In this paper, we propose a new approach to digital
circuits delay analysis that has several novelties. First,
this kind of delay analysis is performed in the early
stages of digital circuit design, which allows the eval-
uation of the speed even during its circuit description.
Then, within the delay estimation algorithm the actual
realization of the circuit is considered. Introducing pa-
rameters tolerances of logic gates into delay analysis
is also a novelty. The estimation method is imple-
mented in analytical phase of MCA to obtain statisti-
cally acceptable estimates of the worst-case delay. The

TECHNICS ELECTRICAL ENGINEERING (2014) 77

M. MILI at al. STATISTICAL DELAY ESTIMATION IN DIGITAL CIRCUITS USING VHDL

method is implemented in VHDL, while its efficiency
is demonstrated on the ISCAS85 benchmark circuits.
8. ACKNOWLEDGEMENT
This research was partly funded by The Ministry
of Education and Science of Republic of Serbia under
contract No. TR32004.
REFERENCES
[1] V. Litovski, Electronic circuit design (in Serbian),
DGIP Nova Jugoslavija - Vranje, Ni, Serbia,
2000.
[2] L. Nagel, SPICE-2: A computer program to
simulate semiconductor circuits, ERL Memo ERL-
M520, Univ. of California, Berkley, CA, May 1975.
[3] J. Ousterhout, Crystal: A timing analyzer for nMOS
VLSI circuits, Proc. 3rd Caltech Conf. on VLSI,
March 1983.
[4] N. Jouppi, TV: An NMOS Timing Analyzer, Proc.
3rd Caltech Conf. on VLSI, Caltech, pp. 71-85,
March 1983.
[5] C. Oh, M. Mercer, Efficient Logic-Level Timing
Analysis Using Constraint-Guided Critical Path Sea-
rch, IEEE Trans. on VLSI Systems, Vol. 4, No. 3,
pp. 346-355, September 1996.
[6] S. Abbaspour, H. Fatemi, M. Pedram, VGTA:
Variation Aware Gate Timing Analysis, Proc. of
the IEEE Int. Conf. on Computer Design, San Jose,
California, pp. 351-356, 2005
[7] A. Krishnamachary, A.Abraham, R.Tupuri, Timing
verification and Delay Test Generation for Hiera-
rchical Designs, Proc. of the 14th International
Conference on VLSI Design, Bangalore, India, pp.
157-162, 2001.
[8] V. Litovski and M. Zwolinski, VLSI Circuit Simu-
lation and Optimization, UK: Chapman and Hall,
1997.
[9] M. Sokolovi and V. Litovski, Using VHDL Simu-
lator to Estimate Logic Path Delays in Combinational
and Embedded Sequential Circuits, Proc. of the
IEEE Region 8 EUROCON 2005, Conference, pp.
547-550, 2005.
[10]http://courses.ece.uiuc.edu/ece543/iscas89.html
[11]D. Maksimovi, V. Litovski, Tuning logic simula-
tors for timing analysis, Electronics Letters, Vol. 35,
No. 10, pp. 800-802, Stevenage, UK, May 1999.
[12]D. Maksimovi, Logic simulation Estimation of
worst-case properties in digital circuit design, PHD
thesis, Faculty of Electronic engineering, Ni, Serbia,
June 2000.
[13]M. Sokolovi, V. Litovski, Efficient Calculation of
the Statistical Worst Case Delay in Complex
Digital Circuits, XLX Conf. of ETRAN, Belgrade,
Vol. 1, pp. 23-26, June 2006.
[14]M. Sokolovi, D. Maksimovi, Estimation of path
delay using VHDL logic simulator, XLIX Conf. of
ETRAN, Budva, vol. 1. pp 99-102, June 2005.
[15]D. Maksimovi V. Litovski, "Logic Simulation Met-
hods for Longest Path Delay Estimation", IEE Proc.-
Comput. Digit. Tech., Vol. 149, No. 2, March 2002.
[16]K. Cheng and V. Agrawal, Unified Methods for VL-
SI Simulation and Test Generation, Boston: Kluwer
Academic Publishers, 1989.
[17]M. Zwolinski, Digital System Design with VHDL,
Pearson: Prentice Hall, 2004.

REZIME
STATISTIKA PROCENA KANJENJA DIGITALNIH KOLA PRIMENOM VHDL-A

Najvanija karakteristika savremenih integrisanih kola je brzina. Ona direktno zavisi od kanjenja kola.
Za projektovanje brzih digitalnih kola, neophodno je procenjivati kanjenja u najranijim fazama
projektovanja. Time se olakava modifikacija i resinteza kola ukoliko ono nije dovoljno brzo. U ovom
radu predloen je postupak kojim se pouzdano moe proceniti kanjenje digitalnih kola jo u fazi
njihovog opisivanja. Metod moe da statistiki proceni minimalna i maksimalna kanjenja svih moguih
puteva i tranzicija signala u kolu, uzimajui u obzir fiziku realizacija kola, i informacije o tolerancijama
parametara. Metod koristi VHDL opis, a verifikovan je na ISCAS85 benchmark kolima. Za procesiranje
dobijenih podataka korien je Matlab.
Kljune rei: kanjenje, Monte-Carlo analiza, fanout, vremenska analiza, VHDL

78 TECHNICS ELECTRICAL ENGINEERING (2014)

M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...

Deterministic Approach for Mobile User Positioning in Radio Systems

MIRJANA I. SIMI, University of Belgrade,

School of Electrical Engineering, Belgrade Review paper
PREDRAG V. PEJOVI, University of Belgrade UDC: 621.396.74
School of Electrical Engineering, Belgrade

This paper analyzes deterministic methods for user positioning in radio systems. Radio signal
parameters that can be used for positioning are considered, as well as the methods of their processing
in order to obtain the unknown location. Paper contains a brief comparison of probabilistic and
deterministic mathematical approaches in positioning, while all deterministic methods are analyzed in
detail. Signal parameters used in angulation, circular and hyperbolic lateration are presented, as well
as their algorithms. The paper discusses the minimum requirements for each deterministic method,
indicates problems that can occur, and the ways to overcome them.
Key words: positioning, angulation, lateration.

1. INTRODUCTION services), emergency services, enhanced call routing,

The problem of user positioning in radio systems
did not lose its actuality in the world of teleco-
mmunications for a long time. The idea about user
positioning in radio networks firstly appeared in ce-
llular networks for safety purposes, in 1996 in the
USA. The development of cellular radio systems in-
creased the number of users, resulting in an increase in
the number of emergency calls, 911. The emergency
calls are sent from mobile stations (natural disasters,
traffic accidents, terrorist attacks, etc.). The problem to
provide a timely response to these calls and appropriate
assistance, is that the mobile users were frequently
unable to provide precise information about their loca-
tion. Therefore, the United States Federal Commu-
nications Commission (FCC) imposed a request for
operators of mobile communication networks to be
able to determine locations of mobile callers that re-
quest emergency assistance via 911. Later, the infor-
mation about mobile user location opened new com-
mercial possibilities for mobile network operators. Lo-
cation Based Services (LBS) provide users with per-
sonalized services at their current location. Today,
there is a large number of LBS and many of them are
standardizedlocation based charging, tracking servi
Authors adddress: Mirjana Simi, University of Bel-
grade, School of Electrical Engineering, Belgrade, Bu-
levar kralja Aleksandra 73
Paper received: 25.07.2014.
Papaer accepted: 04.09.2014.
location based information services (navigation,
location dependent content broadcast, mobile yellow
pages, location sensitive Internet, etc) [1]. In addition,
the positioning can be used for planning and opti-
mization of cellular networks. A large number of appli-
cations in radio networks based on the user location
information in outdoor environment initiated deve-
lopment of similar applications for indoor enviro-
nment. Some of the indoor LBS are inventory tracking,
location detection of products stored in a warehouse,
location detection of medical personnel or equipment
in a hospital, location detection of firemen in a building
on fire, virtual tourist guides in museums and galleries,
etc. [2].
Starting from 1996 until today, many positioning
methods are developed, and they are classified in di-
fferent ways [3]. One of the mostly used classifications
is based on the type of the network in which po-
sitioning is implemented and operated. According to
this classification, there are: satellite, cellular, and
indoor positioning. Satellite positioning relies on sa-
tellite infrastructure, usually on Global Positioning Sy-
stem (GPS). Cellular positioning refers to mechanisms
in cellular networks like Global System for Mobile
Communications (GSM), Universal Mobile Teleco-
mmunications System (UMTS) and Long-Term Evo-
lution (LTE) for obtaining position of a subscriber,
while indoor positioning relies on methods specially
adapted for positioning in conditions characterized by
complex propagation inside buildings.

TECHNICS ELECTRICAL ENGINEERING (2014) 79

M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...
Another popular classification of positioning me-
thods is based on the part of the radio system which
performs the measurements required by the positio-
ning, as well as the calculation of the mobile station
location. This results in a distinction between network-
based and mobile-based positioning. In network-based
positioning, this is done by the network, while for
mobile-based positioning the mobile station (terminal)
carries out the measurements and position calculations.
In a hybrid approach the mobile station performs
measurements only, and transmits the results to the
network where the position is calculated. This is called
mobile-assisted network-based positioning, or simply
mobile-assisted. The reverse configuration is referred
to as network-assisted mobile-based positioning, but
this is applied quite rarely.
Finally, an interesting classification of the posi-
tioning methods is based according to a mathematical
approach for mobile user location determination, i.e.
characteristics of the mathematical model used in the
positioning. According to this classification, there are
deterministic and probabilistic positioning methods. In
deterministic methods, the positioning problem is
usually modeled as a problem of solving intersection
of circles or hyperbolas, where the distance estimation
is obtained by measuring various parameters of the
signal propagation (propagation time, propagation
time differences, received signal power). The biggest
challenge in application of these methods is to find a
good algorithm for solving the system of nonlinear
equations. Contrary to the deterministic methods, pro-
babilistic positioning methods are based on proba-
bilistic models that describe dependence of the signal
characteristics received by the mobile station from the
location of the mobile station. Probabilistic algorithms
treat the input information about the mobile station
location as spatial probability density functions and
combine the probability density functions to improve
the information about the mobile station possible
location.
This review paper analyzes deterministic methods
for user positioning in radio systems, their geometric
interpretation, the model solving algorithms, including
problems typical for the application of this kind of
mathematical models in positioning. Chapter 2 prese-
nts the radio signal parameters that can be used for
positioning. Chapter 3 provides basic information abo-
ut the deterministic approach and positioning scenario
that will be considered in this paper. Chapters 4 and 5
analyze the deterministic methods: angulation, circular
and hyperbolic lateration. For each of these methods
the algorithms to determine the mobile station position
are provided, as well as the problems that may occur
by the use of these methods are discussed. Chapter 6
contains the concluding remarks.
80 TECHNICS ELECTRICAL ENGINEERING (2014)
2. POSITIONING PARAMETERS
Regardless of the positioning type, location of any
object in space can be determined by measuring angles
or by measuring distances. While the angle is
determined by direct measurement, the distance is
obtained indirectly, by measuring level of the received
signal and using an appropriate propagation model, or
by measuring the signal propagation time (or time
difference), using the fact that signals propagate at the
speed of light. Finally, there are four signal parameters
that can be used for positioning in radio networks [4]:
AOA (Angle Of Arrival), TOA (Time Of Arrival),
TDOA (Time Difference Of Arrival), and RSS
(Received Signal Strength).
The received signal strength, RSS, is the signal
parameter that can be used for distance estimation
between the reference point and the mobile station.
Mobile station at unknown location measures the
signal strength from the transmitter with known co-
ordinates (reference point), and implements some of
the propagation models in order to estimate the dis-
tance. The problem of RSS parameter is a high sen-
sitivity to propagation conditions, especially NLOS
(Non-Line-Of-Sight) and multipath propagation, whi-
ch results in large errors in the distance estimation and
consequently the mobile station location. Therefore,
RSS parameter is not used in any of the standardized
positioning methods in cellular radio systems. This
parameter has the same drawback in indoor positi-
oning, but its application is significantly higher there,
primarily due to the lack of alternative.
The time of arrival, TOA, is the signal parameter
that can be used for distance estimation between the
reference point and the mobile station. Mobile station
with unknown location measures the signal propa-
gation time from the transmitter with known coor-
dinates (reference point), and the distance estimation is
easy since the signal propagates with the speed of light.
It is significantly more reliable parameter for posi-
tioning than the RSS parameter. However, TOA para-
meter is also sensitive to NLOS propagation, when the
distance between the mobile station and the reference
point is estimated greater than what it really is. In fact,
TOA provides a signal propagation time between the
mobile station and the reference point that corresponds
to the distance between them only in the case of the
Line-Of-Sight (LOS) propagation.
Parameter TOA has another serious drawback: the
requirement for synchronization between the reference
point and the mobile station. Unfortunately, none of
the existing public mobile systems (GSM, UMTS,
LTE) meet this condition, as well as the most of indoor
systems. Therefore, TOA parameter is practically ina-
pplicable in these systems if the mobile operators do
M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...

not want to buy extra hardware to provide sync- x BSk , y BSk , k 1, 2. The only information the base
hronization. The request for synchronization can be stations provide are the angles k , k 1, 2 , the rays
avoided if the time difference of arrival (TDOA)
parameter is observed instead of the time of arrival (half-lines) that start from the base station BSk and
(TOA) parameter. This is the main reason for intro- point towards the mobile station form with the positive
ducing the parameter TDOA in positioning. In the case ray of the x -axis, y 0, x 0 . The angles are
of cellular systems, this means that time difference in essentially azimuth angles, except the azimuth angles
signal propagation between the two base stations and are referred to the north, and the positive ray of the x
the mobile stations is measured. All standardized -axis points to the east. The choice is made to comply
positioning methods in the existing cellular radio sy- with common notation of analytical geometry. The
stems which are based on the time measurement apply angle measurement is illustrated in Fig. 1, where the
TDOA parameter. In TDOA, sensitivity to NLOS mobile station located at x MS , y MS 5, 5 is
propagation conditions still remains present. observed from three base stations, x BS 1 , y BS 1 3, 5
Historically, angle of arrival parameter, AOA. was
the first parameter exploited to determine position of with 1 0 , x BS 2 , y BS 2 5, 2 with 2 90 , and
radio transmitters, as utilized in goniometric methods. x BS 3 , y BS 3 9, 8 with 3 143.13 .
The angle of signal arrival might be determined ap-
plying direction sensitive antenna systems. Applica-
tion of specific antenna systems is the main drawback
for application in cellular networks, since specific
additional hardware is required, especially in the GSM
and UMTS systems, while in LTE parameter AOA is
used only as an auxiliary positioning parameter. Be-
sides, to extract useful information from the angle of
arrival, LOS propagation is required again, which fre-
quently is not the case, especially in indoor and urban
environments.

3. DETERMINISTIC POSITIONING METHODS

Deterministic methods apply geometric relations
to determine position of the mobile station according
to known coordinates of base stations and distances
and/or angles extracted from the signal radio para-
meters RSS, TOA, TDOA or AOA. The positioning
scenario assumes that there are points of known co-
ordinates available, so-called reference points (e.g., Figure 1 - Angulation
base station, access point, etc.), and that the location of
Coordinates of the base stations and the mobile
mobile station is determined applying deterministic
station observation angles locate the mobile station on
algorithms. It is also assumed that the reference points,
a line
as well as the mobile station, are located in the same
plane, i.e. that the problem is two-dimensional. In this y MS y BSk
paper, the considered radio system is cellular, i.e. the tan k
x MS x BSk (1)
reference points are the base stations, but all analysis
and conclusions can be applied for positioning in any which can be transformed to
other radio system.
y MS x MS tan k y BSk x BSk tan k (2)
4. ANGULATION
if k 2 n , n Z , i.e. x BSk x MS . In the case
Angulation is a deterministic metod to determine
the mobile station location based on known angles of k 2 n , the equation degenerates to
arriving signals from the mobile station to at least two
x MS x BSk . (3)
reference points, i.e. the base stations (BS).
To determine coordinates xMS , y MS of a mobile The observation angle provides more information
station (MS) applying angulation method, at least two than contained in (1), locating the mobile station on the
base stations are needed with known coordinates ray given by (2) and x MS x BSk , for

TECHNICS ELECTRICAL ENGINEERING (2014) 81

M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...
2 k 2 , or on the ray x MS x BSk , for
k 2 or 2 k . This might be
used as a rough test for the solution consistency in the
case of ill-conditioned equation systems.
To determine the mobile station coordinates, at
least two base stations are needed. In general, two base
stations k 1, 2 form the equation system
tan 1 1 x MS x BS1 tan 1 y BS1
tan
1 y MS x BS 2 tan 2 y BS 2
2
assuming finite values for tan k . In the opposite case,
corresponding equation should be replaced by an equ-
ation of the form (3).
In the case tan 1 tan 2 , the base stations and
the mobile station are located on the same line, and the
equation system (4) is singular. An additional base
station is needed to determine the mobile station
coordinates, but it should not be located on the same
line as the two base stations initially used. Further-
more, mobile station positions close to the line defined
by the two base stations result in ill-conditioned
equation system (4). This motivates introduction of
additional base stations, and positions of three or more
base stations on the same line, or close to a line, should
be avoided. In the example of Fig. 1, three base stations
are available, and taking any two of the base stations
to form (4) correct coordinates of the mobile station are
obtained, since the systems are well-conditioned and
the data are free from measurement error. If BS2 is
involved, equation of the form (3) should be used.
In practice, more than two base stations might be
available, and an overdetermined equation system
might be formed
tan 1 1 x BS 1 tan 1 y BS 1
x MS
y
tan n 1 MS x BSn tan n y BSn
where n N and n 2 . The system (5) may be
written in a matrix form
x the intersection points corresponds to the mobile
A MS b
y MS
where A is the system matrix (coefficient matrix),
and b is the right-hand-side vector. The system (6) can
be solved in a least-squares sense [5] forming the
system
x line, resulting in tangent circles. In this case, the third
A T A MS A T b
y MS . (7)
82 TECHNICS ELECTRICAL ENGINEERING (2014)
A similar technique is applied in surveying,
frequently referred to as triangulation, since the object
position is located in a triangle vertex, while the
remaining two vertexes of the triangle are the base
stations.
Two angles are measured in order to determine the
object position. The angles are frequently measured
relative to the position of the other base station. Having
the coordinates of the base stations known, the angles
can be recalculated and expressed in terms used here.
(4)
5. LATERATIONS
Unlike angulation, within lateration deterministic
methods unknown mobile station location is dete-
rmined based upon the distance or distance difference
between the mobile station and base stations.
5.1. Circular lateration
Circular lateration is a method based on infor-
mation about the distance rk of the mobile station from
at least three base stations BSk , k 1, ... n, n 3 .
Coordinates x BSk , y BSk of the base stations are
known. An example for circular lateration using the
same coordinates of the base stations and the mobile
station as in the angulation example is presented in
Fig. 2, where information about the mobile station
position is contained in distances rk instead of the
angles k .
Let us consider a minimal system of equations for
circular lateration
xMS x BSk 2 y MS y BSk 2 rk2 (8)
for k 1, 2, 3 . The equation system is nonlinear. Ac-
cording to the geometrical interpretation depicted in
Fig. 2, each of the equations (8) represents a circle,
centered at the corresponding base station, hence the
name of the methodcircular lateration.
(5)
If the system is consistent, each pair of the circles
provides two intersection points, and location of the
mobile station is determined using the information
provided by the third base station, indicating which of
(6)
station location.
The problem becomes more complicated in the
presence of measurement uncertainties, making exact
intersection of three circles virtually impossible. An
exception from this situation is the case when the two
base stations and the mobile station are located on a
base station wont be needed to determine the mobile
station position. Due to the measurement uncertainty,
M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...

it is also possible that measured distances result in for k 2, 3 . Each of the equations of the form (12)
circles that do not intersect. determines a line in the x, y plane. The line equation
The nonlinear system of equations (8) could be is formed manipulating equations of the circles
transformed to a linear system of equations [6] ap- centered at BS1 and BSk , and it is satisfied at both
plying algebraic transformations. intersections of the circles, if the intersections exist.
Thus, the line obtained from the two circle equa-
tions passes through the circle intersections. This geo-
metrical interpretation is illustrated in Fig. 2 for both
of the line equations (12).
The system of equations (12) could be expressed
in a matrix form
x BS1 x BS 2 y BS1 y BS 2 x MS
x x
BS1 BS 3 y BS1 y BS 3 y MS
(13)
1 r2 r1 x BS1 x BS 2 y BS1 y BS 2
2 2 2 2 2 2

2
.
2 r32 r12 x BS
2
1 x BS 3 y BS 1 y BS 3
2 2

In the case all three of the base stations are located

on the same line, their coordinates satisfy
y BS1 y BS 2 y BS1 y BS 3

x BS1 x BS 2 x BS1 x BS 3 (14)
which results in a singular system (13) since the
Figure 2 - Circular lateration
determinant of the system matrix is zero. In that case,
This removes problems associated with the solu- measurements from additional base stations should be
tion methods for nonlinear equations and ambiguities used to determine the mobile station location, similar
about the mobile station location in the case the circle to the case the angulation method is applied and a
intersections that do not match. The first step in singular system is reached. Also, base stations located
algebraic transformations is to expand the squared in a close to line arrangement result in ill-conditioned
binomial terms from (8) system (13) and huge sensitivity on the distance me-
2
xMS 2
2 xMS xBSk xBSk asurement error. Again, to avoid such situation, addi-
. (9) tional base stations are needed, and their close to line
2 2
yMS 2 yMS yBSk yBSk rk2 arrangements should be avoided as much as possible.
Next, all squared terms are moved to the right- In the case measurements from n base stations,
hand side n 3 , are available, the system of equations is
overdetermined and it takes a matrix form
2 xMS xBSk 2 yMS yBSk
(10) x BS 1 x BS 2 y BS 1 y BS 2
rk2 xBSk
2 2
yBSk 2
xMS 2
yMS x MS
y
Up to this point, equations of the form (8) were x BS 1 x BSn y BS 1 y BSn MS
subjected to transformation separately. Now, let us add r22 r12 x BS
2
1 x BS 2 y BS 1 y BS 2
2 2 2

the equation for k 1 multiplied by 1 1

2 2
.
2 xMS xBS1 2 yMS yBS1 rn r12 x BS

2 2 2 2
1 x BSn y BS 1 y BSn
(11) . (15)
r12 xBS
2 2 2 2
1 y BS1 xMS yMS This system takes the same compact form (6) as in
the case of angulation, and should be solved as (7).
to the remaining two equations. Terms x and y 2
MS
2
MS
Presented analysis is performed under the
on the right hand side are cancelled out, resulting in a
assumption of consistent data, resulting in intersection
linear system of two equations in the form
of all circles in a single point. However, even in the
2x BS1 x BSk x MS 2 y BS1 y BSk y MS case when only three base stations are considered (13)
it is likely that the three circles do not intersect in the
rk2 r12 x BS
2
1 x BSk y BS 1 y BSk
2 2 2
(12) same point. The mobile station location is determined

TECHNICS ELECTRICAL ENGINEERING (2014) 83

M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...

as an intersection of lines defined by intersections of problem to a single quadratic equation or even to a

pairs of circles. Obtained solution may be checked for system of linear equations applying appropriate alge-
compliance with the starting circle equations (8) to braic transformations.
verify the solution and to estimate the error margins. Let us start with the equation (12) derived for the
This also applies in the case the linear least-squares method of circular lateration. Instead of having
method of (7) is applied to solve (15).
distances rk , distance differences
5.2. Hyperbolic lateration
Hyperbolic lateration is a method to determine the d k d k ,1 rk r1
. (16)
mobile station location applying information about
differences in distance of the mobile station to a are available. To eliminate the terms that involve rk
number of pairs of base stations with known consider
coordinates x BSk , y BSk . To determine the mobile
rk2 r12 r1 d k r12 2d k r1 d k2 .(17)
2

station position, this would require at least four base

stations BSk , k 1, ... n , n 4 , and information abo- In this manner, rk is eliminated, while r1 remains
ut at least three differences in distance. To illustrate present in a linear term. After the transformation, the
hyperbolic lateration, the same set of base stations as set of equations (12) becomes
in the case of angulation (Fig. 1) and circular lateration
(Figure 2) are used, as shown in Fig. 3. Differences in 2xBS1 xBSk xMS 2 y BS1 y BSk y MS
distances between the base stations and the mobile 2d k r1 d k2 xBS
2 2 2 2
1 x BSk y BS1 y BSk (18)
station are d k d k ,1 rk r1 , where rk is distance of
which for k 2, 3 results in the equation system
the mobile station from the base station BSk , as in the
case of circular lateration. expressed in a matrix form
The locus of points with constant distance diffe- x
rence between the mobile station and two base stations A MS r1b1 b 0
yMS (19)
is a hyperbola (Fig. 3), hence the name hyperbolic
lateration. These hyperbolas pass through the interse- where
ctions of circles defined by fixed distance between the
base stations and the mobile station, similarly to the x x y BS1 y BS 2
A BS1 BS 2
lines in the case of linearized system of equations for xBS1 xBS 3 y BS1 y BS 3
circular lateration (Fig. 2). (20)
d
b1 2
d3 (21)
and
1 d 2 xBS1 xBS 2 y BS1 y BS 2
2 2 2 2 2

b0 2 2
. (22)
2 d 3 xBS
2
1 x BS 3 y BS 1 y BS 3
2 2

Solution of the linear system is

xMS
y A b1 r1 A b 0
1 1

MS (23)
where r1 is not known yet, while coordinates of the
mobile station are provided as linear functions of r1
x MS k x r1 n x (24)
Figure 3 - Hyperbolic lateration
It is possible to determine the mobile station loca- and
tion as an intersection of hyperbolas directly solving y MS k y r1 n y
the nonlinear system of equations. This approach . (25)
requires iterative solution and raises convergence The value of r1 is computed from
issues related to the numerical methods for nonlinear
equation systems. However, it is possible to reduce the xMS x BS1 2 y MS y BS1 2 r12 (26)

84 TECHNICS ELECTRICAL ENGINEERING (2014)

M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...

which after substitution of (24) and (25) results in a and it can be solved in the least-squares sense (7). The
quadratic equation method that involves the quadratic equation over r1
k 2
x k y2 1r12 2k x nx xBS1 k y n y y BS1 r1
(27) (27) should be applied in cases where information from
nx xBS1 2 n y y BS1 2 0. only three base stations are available, which results in
uniquely determined position of the mobile station in
In general, the quadratic equation provides two some cases, and in two possibilities for the mobile
solutions. If the equation corresponds to the physical station location in the remaining cases.
model, at least one of the solutions should be positive.
Possible negative solution for r1 should be rejected for 6. CONCLUSION
the lack of physical meaning. This paper analyzes angulation, circular and hy-
However, it is possible to obtain two positive perbolic lateration within a deterministic approach for
solutions for r1 , which faces us with a dilemma where determining the location of mobile users in radio
the mobile station is located. In this case, an additional systems. The minimum requirements in the number of
source of information is needed, i.e. another base reference points are two in the case of angulation, three
station that provides information about the difference in the case of circular lateration, and four in the case of
in distances. hyperbolic lateration.
However, in the case an additional source of infor- Each of these methods can be solved by tran-
mation is available, solving of the quadratic equation sformation to a system of linear equations. To some
is not required, and the problem could be transformed extend, the exception is hyperbolic lateration which
to linear [7]. can increase availability requiring only three reference
points, instead of the aforementioned four reference
Let us consider (18). Variable r1 is unknown, and
points, when positioning problem is reduced to solving
it is represented by a linear term on the right hand side. quadratic equation instead the system of linear
Simple transfer of the term that involves r1 to the left equations.
hand side of (18) results in However, in this case, there is a risk of mobile
2xBS 1 xBSk xMS 2 y BS 1 xBSk y MS 2d k r1 station location ambiguity. Therefore, this way of
(28) solving hyperbolic lateration should be applied only in
d k2 xBS
2
1 x BSk y BS 1 y BSk .
2 2 2
cases where only three reference points are available.
Adding the information that originates from the
fourth base station, BS4 , the system of linear equa- REFERENCES
tions over x MS , y MS and r1 is obtained as [1] 3GPP TS 22.071, release 8 (v8.0.0), Location Ser-
vices (LCS) Service description; Stage 1, 2007.
x BS 1 x BS 2 y BS 1 y BS 2 d 2 x MS Available:http://www.3gpp.org/ftp/Specs/archive/22
x x y BS 1 y BS 3 d 3 y MS
BS 1 BS 3 _series/22.071.
x BS 1 x BS 4 y BS 1 y BS 4 d 4 r1 [2] H. Liu, H. Darabi, P. Banerjee, and J. Liu, Survey of
. (29) wireless indoor positioning techniques and systems,
d 22 x BS2
1 x BS 2 y BS 1 y BS 2
2 2 2
IEEE Transactions on Systems, Man and Cyber-
1 2 2
d 3 x BS2
1 x BS 3 y BS 1 y BS 3
2 2
netics, vol. 37, no. 6, pp. 10671080, 2007.
2 2
d 4 x BS 2 2 2 2
1 x BS 4 y BS 1 y BS 4 [3] A. Kpper, Fundamentals of Positioning in Loca-

tion-based Services: fundamentals and operation,
Solution of the system (29) provides unique infor- Wiley , Chichester, England, 2005.
mation about the mobile station coordinates expressed
[4] M. Simi and P. Pejovi, Positioning in Cellular
in a closed-form.
Networks in Cellular Networks - Positioning, Perfo-
In the case more than four base stations provided rmance Analysis, Reliability. A. Melikov, InTech,
information about the distance difference ( n 4 ), an pp. 51-76, 2011.
overdetermined system of equations is obtained as [5] I. N. Bronshtein, K. A. Semendyayev, G. Musiol and
x BS 1 x BS 2 y BS 1 y BS 2 d 2 x MS H. Muehlig, Handbook of Mathematics, Springer,
y MS Berlin, Heidelberg, New York, 2007.
[6] A. Bensky, Wireless Positioning Technologies and
x BS 1 x BSn y BS 1 y BSn d n r1 Applications, Artech House, Boston, London, 2008.
(30)
d 22 x BS
2
1 x BS 2 y BS 1 y BS 2
2 2 2
[7] M. D. Gillette and H. F. Silverman, A linear closed-
1 form algorithm for source localization from time-

2 2 differences of arrival, IEEE Signal Processing
d n x BS
2
x 2
y 2
y 2
Letters, vol. 15, no. 1, pp. 14, 2008.
1 BSn BS 1 BSn

TECHNICS ELECTRICAL ENGINEERING (2014) 85

M. SIMI at al. DETERMINISTIC APPROACH FOR MOBILE USER POSITIONING...

REZIME
DETERMINISTIKI PRISTUP ODREIVANJU LOKACIJE MOBILNOG KORISNIKA U
RADIO SISTEMIMA

U ovom radu analizirane su metode u okviru deterministikog pristupa odreivanju lokacije mobilnog
korisnika u radio sistemima. Razmatrani su parametri radio signala koji se mogu koristiti za pozi-
cioniranje, kao i naini njihove obrade u cilju dobijanja nepoznate lokacije korisnika. Dato je kratko
poreenje dva razliita matematika pristupa u raunanju nepoznate lokacije mobilnog korisnika,
probabilistiki i deterministiki, dok su sve metode deterministikog pristupa detaljno analizirane. Pre-
dstavljeni su parametri signala koji se koriste u metodama angulacije, cirkularne i hiperbolike
lateracije, kao i sami algoritmi ovih deterministikih metoda. Definisani su minimalni zahtevi svake od
deterministikih metoda u cilju prorauna lokacije korisnika, prikazani problemi koji se mogu javiti
primenom svake od ovih metoda, kao i naini njihovog prevazilaenja.
Kljune rei: pozicioniranje, angulacija, lateracije

86 TECHNICS ELECTRICAL ENGINEERING (2014)

 ISSN 0558-6208
TECHNICS

TRAFFIC
Saobraaj Traffic Verkehr

YEAR 61 2014.

EDITOR-IN-CHEF

Prof. Smiljan V u k a n o v i , D.Sc, University of

Belgrade, Faculty of Transport and Traffic
Engineering, Belgrade

EDITORIAL COUNCIL

Prof. Vladeta o l i , D.Sc, University of Belgrade,

Faculty of Transport and Traffic Engineering,
Belgrade
Vladimir D e p o l o, D.Sc, Belgrade Land Development
Public Agency, Belgrade
Imre K e r e k e s, M.Sc, City Planning Institute of
Subotica
Prof. Milan M a r k o v i , D.Sc, University of Belgrade,
Faculty of Transport and Traffic Engineering,
Belgrade
Prof. Duan T e o d o r o v i , D.Sc, University of
Belgrade, Faculty of Transport and Traffic
Engineering, Belgrade
Prof. Jadranka J o v i , D.Sc, University of Belgrade,
Faculty of Transport and Traffic Engineering,
Belgrade
Prof. Pavle G l a d o v i , D.Sc, Univesity of Novi Sad,
Faculty of Technical Sciences, Novi Sad

EDITORIAL OFFICE: Union of Engineers and

Technicians of Serbia, 11000 Belgrade, Kneza Miloa 7a/I,
Tel. +381/11/ 32 35 891, Fax +381/11/32 30 607
V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

Methodology for Collecting Instantaneous Vehicle Emissions on Street

Network

VLADIMIR D. ORI, University of Belgrade, Original scientific paper

Faculty of Transport nad Traffic Engineering, Belgrade UDC: 621.43.068 ; 504.06

In this paper, the methodology for collecting second-by-second instantaneous emission data of basic
exhaust gases is presented. The data serves as a base for vehicle emission estimation of traffic flow.
Measurement system is based on GPS and gas analyser integration. Survey was conducted in real traffic
conditions in order to capture locally specific traffic and emission characteristics. This approach
overcomes the shortcomings of laboratory emission tests. The solutions for data collection problems are
presented. Focus was given to data time synchronization and gas analyser response delay reduction.
Key words: instantaneous vehicle emissions, instantaneous speed, emission modelling

1. INTRODUCTION of vehicles in the traffic flow is the basis for emission

Emission of a single vehicle depends on a wide
scale of parameters which can be grouped into [1]:
Operating conditions of the vehicle (cold or hot
state of the vehicle),
Traffic flow characteristics,
Technical characteristics of the vehicle and
Ambient parameters.
The same vehicle has different emission characte-
ristics under different traffic characteristics (speed and
acceleration) [2]. And traffic flow characteristics re-
presented by the relation between volume, speed and
density and are influenced by traffic management me-
asures. Management measures based on traffic signals
have significant influence on traffic flow emissions
[3]. On the other hand, traffic management improve
traffic flow conditions and decrease travel times, whi-
ch can generate new transport demands [4].
Traffic congestions together with quantity of vehi-
cle kilometres travelled, are frequently highlighted as
the main factor of increased traffic flow emission and
air pollution [5]. Traffic congestions are mostly con-
nected to increase in CO and HC emissions [6]. Emi-
ssion tests on modern cars have demonstrated that their
emissions are particularly sensitive to the occurrence
of congestion [7]. Emission characteristics estimation
Authors address: Vladimir ori, University of
Belgrade, Faculty of Transport and Traffic Engineering,
Belgrade, Bulevar vojvode Stepe 305
Paper received: 04.11.2013.
Paper accepted: 08.01.2014
models, integration with transport models and qu-
antification of environmental influence of traffic [8,
9].
2. THE PROBLEM OF DATA COLLECTION AND
VEHICLE EMISSION MODELING
The choices of data collection method and emi-
ssion model type are closely related. The key element
is the potential for precise definition of locally specific
emissions. Data base on vehicle emissions should
enable absolute quantification of emissions with the
highest possible level of precision [9].
Data collection problem
There are several methods for emission measu-
rement [10]:
laboratory engine bench testing or laboratory
chassis dynamometer testing,
road measurements and
on-board measurements.
The most common methods for measuring vehicle
emissions have been the use of dynamometer tests in a
laboratory test facility or roadside, remote sensing at
specific on-road locations, and tunnel studies [11].
Laboratory tests are often used to validate the data
collected using other methods [12]. In a laboratory,
vehicles are driven in a predetermined driving pattern
(driving cycle1) and emissions are measured using gas
1
A driving cycle is a predefined driving pattern in
which operating conditions in terms of instantaneous
speed, and in some cases gearshift points, are specifically
defined [6].

TECHNICS TRAFFIC (2014) 89

V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...
analysers [13]. The use of driving cycles provides the
simulation of characteristic traffic flow conditions [13]
Remote sensing is used as a method for tracking
CO, HC and NOx of traffic flow on a traffic network
section using infrared technology. The ratio of po-
llutants is measured in an exhaust plume by passing an
infrared light beam through the plume and measuring
the radiation absorbed by pollutants at given fre-
quencies [14]. Remote sensing approach enables mea-
suring total emissions of a traffic flow consisting of
different vehicles.
On-board emission measurement is based on in-
vehicle systems for collection of emission and driving
pattern data. That way, real emissions data are reco-
rded using measurement equipment [15, 16, 17] and
data can be analysed after, for average emission cal-
culation. Otherwise, data can be assessed and stored in
real time for production of instantaneous emissions.
Emission modelling problem
There are several types of emission models in re-
lation to methods of data collection:
Agregated emission factor models
Average speed models
Modal models
Second-by-second emision models
Aggregated emission factor models operate on the
simplest level, with a single emission factor being used
to represent a particular type of vehicle and a general
type of driving (or road category. The emission factors
are calculated as mean values of measurements on a
number of vehicles over given driving cycles, and are
usually stated in terms of the mass of pollutant emitted
per vehicle and per unit distance (g /vehicle/km) or per
unit of fuel consumed (g /litre). Given their simplicity,
these factors are of most use in applications on a large
spatial scale, such as national and regional emissions
inventories, where little detailed information on vehi-
cle operation is required [18].
Average-speed models are based upon the prin-
ciple that the average emission factor for a certain
pollutant and a given type of vehicle varies according
to the average speed during a trip. The emission factor
is usually stated in grams per vehicle-kilometre (g
/vehiclekm). However, there are several limitations as-
sociated with average-speed models, including the fo-
llowing: trips having very different vehicle operation
characteristics, and therefore different emission levels,
can have the same average speed; and average speed
models do not allow for detailed spatial resolution in
emission predictions [18]. Most often used average
emission model in Europe is Copert data base.
In modal models emission factors are allocated to
the specific modes of vehicle operation encountered
90 TECHNICS TRAFFIC (2014)
during a trip. Aim is to provide a more precise des-
cription of vehicle emission behaviour by relating emi-
ssion rates to vehicle operation during a series of short
time steps (often one second). In order to estimate the
effects of driving dynamics the vehicle operating co-
nditions and emissions in different traffic flow con-
ditions are analysed. The current state of the vehicle,
described by current speed and acceleration is simu-
ltaneously recorded with instantaneous emissions. Dri-
ving is typically divided into four modes: idle, acce-
leration, deceleration and cruise [19]. Several different
terms (as well as modal) have been used to describe the
more detailed type of model, including microscale,
continuous, on-line and instantaneous, which is
considered to be the most adequate [20].
Models based on second-by-second emission data
are in fact modal models but from a different perspe-
ctive [21]. Vehicle Specific Power (VSP) is taken into
account, as it is been identified as a useful parameter,
for estimation of passenger car emissions. Models are
based on relations between emissions and road grade,
vehicle speed and acceleration. Relatively high nu-
mber of repeated measurements Models based on
second-by-second on-board emission measurements
use relatively large sample sizes of repeated runs as a
basis for developing average emission rates and
confidence intervals for the average emissions [10].
After data collection from on-board measurements and
laboratory dynamometer, this information is then
compiled for use in developing a conceptual modelling
approach [10].
In the paper, it is chosen to develop a method for
emission measurement that should provide on-board,
high resolution data (in each second) suitable for
developing any of the previously described models.
3. SURVEY METHODOLOGY
General methodology for transport survey invo-
lves:
Survey calendar (valid days)
Time resolution
Spatial resolution
Survey Sample
Survey method
Measurement instalation
Training for equipment handling
Testing the data collection procedure
Data processing
Survey calendar (valid days). Survey is conducted
in 2012, July and August. Representativenes of chosen
months, considering traffic volumes and traffic flow
structure is considered to be of no significance. The
only criteria was to collect data for all combinations of
V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

vehicle speed and acceleration. Chosing this period for concentrations. The collection of data on concentration
survey enabled colecting data with minimal influence of O2, CO, NO, NO2, SO2, H2S, H2, CH4 is provided in
of traffic flow volumes. ppm (particles per million) with automatic conversion
Time frame. Survey covered 15 days of research to mg/m3. Additional data are provided through cal-
activity with a single run in each day. Single runs culation of CO2, (oxygen in exhaust gas), combustion
include 90 minute uninterrupted measurement inter- efficiency, temperature losses and efficiency, dew
vals during peak and off peak traffic flow periods. Ve- point, NOx, SO2, CO/CO2 relation. Temperatures of
hicle runs were mostly conducted in the period bet- exhaust gas and ambient temperature are also ava-
ween 7 and 9 AM. ilable.
Spatial resolution. Considering the survey com- The data on driving pattern are provided using
plexity it wasnt possible to collect the valid set of data GPS unit (1Hz data collection) powered by a laptop
for the whole city territory. The focus of the survey was computer. The accuracy of the instantaneous speed is
on streets without significant grade. The vehicle routs 0.2 km/h in steady state (which increases in motion).
were defined so that all of the main street network In order to use measurement installation on-board,
categories are represented (high-way, arterial, primary three problems needed to be treated [9]:
and secondary streets). Stable power supply in the survey time frame,
Sample size. During the time frame of the survey Data conversion to absolute emissions and
vehicle operational characteristics and emissions were Data storage.
collected with one second resolution. All vehicle runs
Stable power supply was provided by using SLA
are conducted using a single vehicle (1.6 fuel based)
(SLA Sealed Lead Acid) 12V battery (Figure 1- mark
with before ECE emission standard and no catalytic
1) which has the ability for a large number of full
converter. The test vehicle is in the category of large
charges and discharges. The battery capacity was
polluters, manufactured in 1992 and went over 230000
chosen to provide for full operational capacity for at
km. The survey didnt involve the differences in dri-
least 2 hours.
ving style.
The power supply provides power for heated
Survey method. On-board measurements showed
sample line which is heated to over 100 C (Figure 1
the most potential among all mentioned survey
mark 2) and the measurement unit itself (Figure 1
methods (in Paragraph 2). On the other hand, the
mark 3). The consumption was reduced by keeping the
methodology of the survey doesnt ensure controlled
installation connected to main power source (220V)
environment like in laboratory tests.
during preparatory phase which involves unit zeroing
Measurement installation. Special attention was (cleaning the sensors with ambient air) and sample line
be given to measurement installation for second-by- initial heating. DC/AC converter (Figure 1- mark 4) is
second on-board emission collection. Detailed descri- used to provide stable 220V power supply during on-
ption will be given in the next paragraph. board measurement.
Data collection procedure and user training were
conducted simultaneously in the course of preparatory
test runs. Errors in data collection were determined in
this phase, and find to be mostly related to mistakes in
measurement procedure and errors in some part of the
measurement equipment. Preparatory procedures for
measurement installation were improved in order to
minimize the quantity of false data.
Data processing. Data storing was provided in
tabular form together with graphical presentation. This
enabled data filtering and graphical control of data
consistency.

4. MEASUREMENT INSTALATION
Measurement system is based on integration of gas
analyser unit for collecting vehicle emissions (Vario+
measuring unit) and driving pattern data (Garmin GPS Figure 1 - Experimental design - in vehicle setup
18x). The measurement of toxic gases is based on the The vehicle was adjusted for measurement of
electrochemical principle (for lower emission ranges) emissions and gas flow velocity. The velocity of the
together with the infrared bench (NDIR) for higher gas flow was necessary in order to convert emissions

TECHNICS TRAFFIC (2014) 91

V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

from ppm and mg/m3 to g/s. Gas velocity is expressed were used for data assessment. Acceleration values are
in m/s and measured using Pitot tube2. in the range from -16 to16 m/s2, and speeds between 0
The absolute quantity of gas (Qgas) was calculated and 90 km/h. Reliability of the installation was proven
as a product of gas velocity (Vgas) and tailpipe cross with only 5% of the false collected data.
section area (P) (Equation 1). Tailpipe cross section The operational conditions of vehicle (e.g. cold
was calculated to be 0.002 m2, according to tailpipe start) can influence the overestimation of data.
diameter. Collected data were controlled considering emissions
in cold start conditions. Collected data and their
m3 m

Qgas Vgas P m 2 exhaust gas temperature are shown in Figure 3.
s s
(1)
Considering the resolution of all the data to be one
second, the conversion of emissions to mg/s is enabled
using Equation 2.
mg mg m3
Emis Emis 3 Qgas
s m s (2)
The position of Pitot tube (Figure 2 mark 1) and
gas sample probe (Figure 2 mark 2) are presented.
Sample probe is placed orthogonally with the pipeline
and in the middle of gas flow. Pitot tube is placed after
the sample probe, further away so that the position of
probe does not disturb the flow.
Figure 3 - Exhaust gas temperature for data base
There is a clear distinction between the data with
exhaust gas of over 80 C and those under that value.
When compared, the idling emissions between the
under 80 C data and the whole set, there arent any
significant differences. It is concluded that all the data
on emissions are collected in heated state of the
vehicle.
The results are stored in tabular form and also
graphically, in the file that can be read in Google Earth
application. Visualization of data is enabled and easy
connection to location (location criteria search). Each
data sample is connected with the location of recording
Figure 2 - Positions of Pitot tube and gas sample probe and the data table holding the complete data (Figure 4)

The internal memory of measurement unit enables

storing up to 8500 data samples. On line presentation
of data (using RS232 protocol) is slow and produces
the loss of data, enabling 1 out of 6 data samples to be
collected.

5. DATA BASE
During survey period, 15 runs using test vehicle
were made on a main street network of Belgrade,
Serbia. Close to 55000 data entries (seconds) were
collected and after control and validation about 49000 Figure 4 - Data in kml file
Connecting geo-referenced spatial data with emi-
2
A Pitot tube is a pressure measurement instrument ssion and travel pattern data enables generation of new
used to measure fluid flow velocity and quantity. It is data types. Consequently, street category can be linked
widely used to determine the speed of air, gas or water in to data base. There is also a potential for banding
the industry or the speed of a boat of an aircraft. together with other data bases, like Transport Model of

92 TECHNICS TRAFFIC (2014)

V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

Belgrade, to provide for additional street characte-
ristics (e.g. capacity, speed restrictions, free flow spe-
eds etc.).
Calculation of instantaneous accelerations is based
on instantaneous vehicle speed as a difference between
velocities in time step n and n-1. Acceleration is ex-
pressed in km/h/s. Vehicle emissions are highly
Table 1. Structure of data base for model generation
Data category Parameters
dependent on vehicle power demand [22, 23]. Vehicle
power demand is calculated as a product of insta-
ntaneous speed and acceleration and expressed in
km2/h2/s. The absolute quantity of emitted gas (g/s) is
calculated using Equations 1 and 2 (given in Section
4). Elements of the final data base are presented in
(Table 1).

Vehicle characteristics Engine size; Odometer readings ; ECE category

Ambient conditions and emission Ambient temperature (C); Gas temperature (C); Gas velocity
characteristics (m/s); Temperature losses; Combustion efficiency

Position and street category Latitude; Longitude; Altitude; Street category;

Vehicle activity Instantaneous speed (km/h); Instantaneous acceleration (km/h/s);

Power demand (km2/h2/s); Date; Time; ;

Vehicle emissions HC, CO, NO, CO2 i O2 emissions (ppm, mg/m3, g/sec, %)

Corrected data base emission correction
Several sources of error in data base collection are
identified. The most important are related to time
synchronization of emissions and travel pattern data.
After the emission, in the course of exhaust gas
transport to measurement instrument, time losses are
present. There is an offset between emission signal
reading on the gas analyser and the time of its emi-
ssion.
Consequently, there is an offset between traffic
characteristics that influence the quantity of emission
and the readings on the measurement instrument. Even
thou, this effect is well known, the methods for
overcoming this problem are not widely spread [24,
25]. Fixed time offset is often used to compensate for
this time delay.
However, the offset is not in constant delay, and it
varies more than one second according to the gas flow
in the analyser sample line and tailpipe. Considerable
variations in time delay occur because of the different
gas flow velocities. Generally, slow gas flow at low
engine speeds (rpm) and low engine load results in
large delay [24].
Besides, time delay in the tailpipe, there is an on-
going process of gas diffusion. Consequently, the emi-
ssion peaks are flattened, and real emission values are
reached with the delay. Diffusion dynamics and peak
flattening depend on gas flow volume and velocity
[26], so the emission peaks often dont represent spe-
cific traffic conditions.
Certain compensation must be applied to the data
to cancel these effects and return missing extreme
values. Vehicle and measurement installation system
were observed as a system with constant and dynamic
part of delay. The constant part of the delay was de-
termined visually for every vehicle run.
The moment in the speed time trace when the
vehicle starts to accelerate (from zero speed) was used
to match with the emission increase. The dynamic part
of the response delay was determined using theory of
the system control [27]. The response of the system
(vehicle and gas analyser) was recorded in the stabile
conditions (vehicle in the steady state).
The emission signal of the gas analyser was re-
corded as a response to constant input. This kind of
generalization is considered to be appropriate for the
purpose of the research. The emissions were recorded
until they reach stable value. Constant input is an
approximation of step function signal (Heaviside step
function) where input value instantly changes from mi-
nimal to maximal. Results of the response to step
function for CO are presented in the Figure 5. The
results of two separate tests are presented, where dark
line presents data used for system response identi-
fication.

TECHNICS TRAFFIC (2014) 93

V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

Figure 5 - Gas analyser static response to Heaviside step function CO emission change
Static response data were used to define the tra- up in Figure 6). However, lost information is not
nsfer function of the system (gas analyser). Transfer retrievable. It is only possible, to get closer to reality
function determines the dynamics of the recorded with faster response times (rapid accelerations) and
value on the gas analyser when the input signal is Hea- controlled signal boost (cancelling the signal fla-
viside step function. The identification of the signal is ttening).
performed in MATLAB (System identification modu-
le). Transfer functions are identified for CO, CH4 and
NOx for both, increasing (min to max) and decreasing
(max to min) mass emissions. The elements of the
transfer function are poles and zeroes. Systems are
identified as a third order dynamic system with two
imaginary and one real pole. Positions of imaginary
poles are influenced by damping ratio () and natural
undamped frequency (n). The general form of the
transfer function in s domain is given in Equation 3:
Kp
G(s) =
1 2 n
s n s 2 1 Tp3 s
(1)
where: Kp is the proportional gain of the system; is Figure 6 - Output signal adjustment (Simulink)
damping ratio; n presents natural undamped frequ- The total modification of the signal is near 2%
ency and Tp3 is the real pole. which is considered not to be significant. Time line
Based on a control system theory, the estimation data coordination provided increased instantaneous
of the system input (emission concentration at the emission accuracy for speed/acceleration data.
tailpipe) that caused a certain system output (analyser
indication) can be given by the inverse transfer fun- 6. CONCLUSIONS
ction [28]. Since the response of the gas analyser is not In this paper, a measurement installation for co-
instantaneous, inverse function is used to reconstruct llection of instantaneous emissions and speeds is pre-
the emission signal at the end of the tailpipe, according sented. Problems in the process of data collection are
to signal recorded by gas analyser. Inverse transfer identified and solutions to solving these problems are
function is simply used to speed up the recorded proposed.
signal and cancel the analyser influence on signal
Developed measurement installation is an integral
distortion.
part of the methodology for data collection. The result
Signal adjustment was performed in MATLAB is a data base on vehicle instantaneous emissions,
(Simulink). A block diagram was used to simulate the speed, acceleration and power demand in real traffic
effects of the inverse transfer function. The result is flow. Additional data and potentials for data base
corrected signal with faster reaction (emission change) expansion are also presented.
to instantaneous speed change (Figure 6). Overshot is The focus is given to modification of measured
the consequence of the transfer function type (con- emission data that is distorted as a consequence of
sequently, inverse transfers function to). recording delay of the gas analyser. Control system
Adjusted gas analyser response is slightly faster in theory was used to reduce the effects of gas analyser
both increasing and decreasing direction (6 sec speed- on measurement precision.

94 TECHNICS TRAFFIC (2014)

V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

Emission data are a base for emission models
development. High frequency data (in each second)
enable development of all types of emission models,
from general aggregated to instantaneous models. In
order to produce locally specific emission models it is
necessary to expand the data base, using instantaneous
emissions for different vehicle types according to
engine size, fuel type and emission standard.
5. ACKNOWLEDGEMENT
This research was conducted as a part of Serbian
Ministry of Education, Science and Technological
Development scientific project 36021 - The Effects of
Global Challenges on Transportation Planning and
Traffic Management in Cities
REFERENCES
[1] Den Braven, K., Abdel-rahim, A., & Battles, A..
Modeling Vehicle Fuel Consumption and Emissions
at Signalized Intersection Approaches: Integrating
Field-Collected Data Into Microscopic Simulation
(p. 59), 2012.
[2] NRC. Modeling Mobile-Source Emissions (p. 239).
Washington, D.C: National Academy Press, 2000.
[3] De Coensel, B., Can, A., Degraeuwe, B., De Vlieger,
I., & Botteldooren, D. Effects of traffic signal
coordination on noise and air pollutant emissions.
Environmental Modelling & Software, 35, 7483.
doi:10.1016/j.envsoft.2012.02.009, 2012.
[4] Vukanovi, S. ITS and traffic management and co-
ntrol: Part II. Tehnika - Saobraaj, 57(2), 1926,
2010.
[5] Jovic, J., & Djoric, V. Traffic and Environmental
Street Network Modelling: Belgrade Case Study.
Transport, 25(2), 155162, 2010.
[6] Smit, R. An Examination of Congestion in Road
Traffic Emission Models and their Application to
Urban Road Networks. Griffith University, 2006.
[7] Smit, R., Brown, A. L., & Chan, Y. C. Do air po-
llution emissions and fuel consumption models for
roadways include the effects of congestion in the
roadway traffic flow? Environmental Modelling &
Software, 23(10-11), 12621270, 2008.
[8] ori, V. Modeliranje transportnih zahteva sa
aspekta odrivog transporta i energetske efikasnosti.
Tehnika - Saobraaj, 55(6), 1724, 2008.
[9] ori, V.. Istraivanje i modeliranje emisija vozila u
funkciji modeliranja saobraajnih tokova na ulinoj
mrei. Doktorska disertacija, Univerzitet u Beogradu
- Saobraajni Fakultet, 2013
[10]Pimentel, M. Transportation Modelling for Enviro-
nmental Impact Assessment Porto Metropolitan Area
Case Study. University of Porto, 2008.
[11]Frey, C., Unal, A., & Chen, J. Recommended Stra-
tegy for On-Board Emission Data Analysis and
Collection for the New Generation Model (p. 272).
US, 2002.
[12]Kousoulidou, M., Ntziachristos, L., Hausberger, S.,
& Rexeis, M. Validation and Improvement of CO-
RINAIRs Emission Factors for Road Transport
using Real-World Emissions Measurements (p. 144),
2010.
[13]Ntziachristos, L., & Samaras, Z.. EMEP/EEA
emission inventory guidebook 2009 (pp. 1129). EU,
2010.
[14]Grant, D. Representative Vehicle Operating Mode
Frequencies: Measurement and Prediction of Vehicle
Specific Freeway Modal Activity. Georgia Institute
of Technology, 1998.
[15]Frey, C., Rouphail, N., Unal, A., & Colyar, J. Emi-
ssions Reduction Through Better Traffic Manage-
ment: An Empirical Evaluation Based Upon On-Ro-
ad Measurements (p. 369). US, 2001a.
[16]Frey, H. C., Unal, A., Rouphail, N. M., & Colyar, J.
D. On-road measurement of vehicle tailpipe emi-
ssions using a portable instrument. Journal of the Air
& Waste Management Association (1995), 53(8),
9921002, 2003.
[17]Unal, A. On-Board Measurement and Analysis of
On-Road Vehicle Emissions. North Carolina State
University, 2002a.
[18]Barlow, T., Boulter, P., & Mccrae, I. A review of
instantaneous emission models for road vehicles (p.
64). UK, 2007.
[19]Frey, C., Rouphail, N., Unal, A., & Colyar, J. Mea-
surement of On-Road Tailpipe CO, NO, and Hydro-
carbon Emissions Using a Portable Instrument. In
Annual Meeting of the Air & Waste Management
Association (pp. 120), 2001b.
[20]De Haan, P., & Keller, M. Emission factors for
passenger cars: application of instantaneous emission
modeling. Atmospheric Environment, 34(27), 4629
4638. doi:10.1016/S1352-2310(00)00233-8, 2000.
[21]Unal, A. On-board Measurement and Analysis of On-
Road Vehicle Emissions, 2002b.
[22]Frey, C., Unal, A., Chen, J., & Li, S. Modeling
Mobile Source Emissions Based Upon In-Use and
Second-by-Second Data: Development of Con-

TECHNICS TRAFFIC (2014) 95

V. ORI METHODOLOGY FOR COLLECTING INSTANTANEOUS VEHICLE...

ceptual Approaches for EPA s New MOVES Mo- rate instantaneous emission measurement. Atmo-
del. In Annual Meeting of the Air & Waste Mana- spheric Environment, 37(37), 51375145. doi:
gement Association (pp. 120), 2003. 10.1016/j.atmosenv.2003.05.004, 2003.

[23]Song, G., Yu, L., & Tu, Z. Distribution Chara-
cteristics of Vehicle-Specific Power on Urban Restri-
cted-Access Roadways. JOURNAL OF TRANS-
PORTATION ENGINEERING, (February), 202
209. doi:10.1061/(ASCE)TE, 2012.
[24]Ajtay, D., & Weilenmann, M. Compensation of the
exhaust gas transport dynamics for accurate instanta-
neous emission measurements. Environmental Scie-
nce & Technology, 38(19), 51418, 2004.
[25]Weilenmann, M., Soltic, P., & Ajtay, D. Describing
and compensating gas transport dynamics for accu-
[26]Weilenmann, M., Bach, C., & Rdy, C. Aspects of
instantaneous emission measurement. International
Journal of Vehicle Design, 27(1-4), 94104, 2001.
[27]Ogata, K. Modern Control Engineering (Fifth edit.,
p. 905). Pearson Education, Inc, 2010.
[28]Geivanidis, S., & Samaras, Z. Development of a dy-
namic model for the reconstruction of tailpipe emi-
ssions from measurements on a constant volume sa-
mpling dilution system. Measurement Science and
Technology, 19(1), 12, 2008.
[29]doi:10.1088/0957-0233/19/1/015404

REZIME
ISTRAIVANJE TRENUTNIH EMISIJA IZDUVNIH GASOVA VOZILA NA ULINOJ
MREI

U radu je predstavljena metodologija prikupljanja podataka o trenutnim emisijama (na nivou svake
sekunde) osnovnih zagaujuih materija koje nastaju kao posledica uslova u saobraajnom toku na
ulinoj mrei. Sistem za merenje je baziran na integrisanju GPS ureaja i ureaja za merenje emisija.
Prikupljanje podataka je realizovano u realnim uslovima vonje ime se opisuju lokalne karakteristike
emisije. Na taj nain su prevazieni nedostaci prikupljanja podataka u laboratorijskim uslovima. U radu
su predstavljena reenja za prevazilaenje problema ovog naina prikupljanja podataka sa akcentom na
vremensko usaglaa-vanje podataka i greke u odzivu merne instalacije.
Kljune rei: trenutna emisija izduvnog gasa vozila, trenutna brzina, modeliranje emisija

96 TECHNICS TRAFFIC (2014)

N. ELAR ANALITICAL DELAY MODELS FOR SIGNALIZED INTERSECTION

Analitical Delay Models for Signalized Intersection

NIKOLA . ELAR, University of Belgrade, Previous announcement

Faculty of Transport and Traffic Engineering, Belgrade UDC: 656.1.05

Delay, defined as additionally spent time in travel realization, represent the main indicator of the traffic
process realization quality on intersection and input parameter in procedure of traffic signal settings
optimization. Their value can be determined using no experimental (analysis and simulation) and
experimental research methods. Analytical models provide fast, simple evaluating procedure of delay.
In research is implemented procedure of value comparison of delay obtained by using relevant analytical
models and the real value determined during research. On the basis of the obtained results it is
concluded that either of analyzed models doesnt interpret real values of delay, and also have shown
that the minimum deviation in evaluation of delay value is carried out by using analytical model from
Highway Capacity Manual 2010 (HCM 2010).
Key words: signalized intersection, delay, delay models

1. INTRODUCTION flow values and passenger car equivalent) must re-

Delay value on signalized intersection most often
is determined by using analytical models. These mo-
dels are based on simplified theoretical interpretation
of traffic process and presumption that it is actualized
in typical, objective required conditions and behaviors
of system elements. Delay value is determined on
macroscopic level with simple relations between basic
parameters of the traffic flow and traffic signals timing
parameters.
The use of formed analytical models, from engi-
neer aspect, represents the fastest and the simplest way
of delay evaluation. On the other side by application of
original models in different traffic systems raises the
question of objectivity and precision of given delay
value.
Problem of real delay interpretation by using ana-
lytical models can appear as consequence:
Use of recommended, base values of input model
parameters;
Inadequate structures and formulations of
analytical model.
Therefore the procedure of real values deter-
minations of input parameters in research (saturated
Authors address: Nikola elar, University of Bel-
grade, Faculty of Transport and Traffic Engineering, Bel-
grade, Vojvode Stepe 305
Paper received: 06.12.2013.
Paper accepted: 13.03.2014.
present indispensable step in practical model use.
Problem of inadequate model structure (analytical
expression for delay) can be solved in one of these
ways:
transforming the existing, or creating new model;
calibration of existing model (most often by
introducing correcting factors in original
expression for delay determination). Similar
transformation of Websters model and its
adjustment to domestic conditions have been
already done in Yugoslavia (Vukanovic, 1980)
[9].
Both mentioned procedures represent long-term
and demanding researching process
Research problem, considering work, is directed to
determine level of real delay values interpretations, in
domestic conditions, which are obtained by using ori-
ginal analytical source of relevant model formulation..
2. OVERVIEW OF RELEVANT ANALITICAL
TIME DELAY MODELS
2.1.Delay models structure
Structure of delay models is based on simplified
theoretical interpretation of traffic process on signa-
lized intersection. On the diagram that follows (Picture
1, Picture 2, and Picture 3) are shown tree main rea-
lization processes.
Picture 1. Shows stable process of queue formation
and queue discharge. In every cycle comes to complete

TECHNICS TRAFFIC (2014) 97

N. ELAR ANALITICAL DELAY MODELS FOR SIGNALIZED INTERSECTION
discharge of formed queue. These conditions of rea-
lization process are undersaturated and delays that
appear in these conditions are uniform delays
Figure 1 Theoretical case of stable realization
process in unsaturated conditions
Figure 2 Theoretical case of unstable realization
process in undersaturated conditions
Figure 3 Theoretical case of realization process
in oversaturated conditions
Last diagram represent traffic process in over-
saturated conditons. During the whole period of obser-
vation traffic demand is higher than intersection capa-
city. The number of unserved vehicles and queue len-
gth is successively increasing by the end of analysis
period. Delays that arise as result of constant appea-
rance of mentioned conditions are overflow delay. The
delay value depends upon the length of time that the
condition exists.
Therefore, in the analytical structure of delay
model, we distinguish three delay types:
98 TECHNICS TRAFFIC (2014)
Uniform delay base type of delay, is the
delay based on an assumption of uniform arrivals and
stable flow in unsaturated conditions;
Random delays is the additional delay, above and
beyond uniform delay, because flow is randomly
distributed rather than uniform;
Overflow delay is the additional delay that occurs
when the capacity is less than the demand or arrival
flow rate.
2.2. Webster delay model
Structurally, Webster [10] has developed two par-
tial models of delay:
Delay model for undersaturated conditions;
Delay model for oversaturated conditions (con-
gestion).
Total value of the average delay on the approach
of signalized intersection, in undersaturated conditio-
ns, represent the sum of uniform and random delay.
.
0.9 (1)
The average total delay represent the sum of
uniform delay component (case X=1) and overflow
delay.
d 0.5C 1 X 1 (2)
d average total delay (s/veh)
C- cycle lenght (s)
z effective green time (s)
X degree of saturation, or volume to capacity ratio (-)
q arrival flow rate (veh/s)
T analysis period (h)
Figure 4 Random and overflow delay models com-
pared
On diagram (Picture 4) are shown the zones of
application of different delay models. For the value of
degree of saturation less than 1 (X>1). delay is deter-
mined with delay model for undersaturated conditions
(uniform delay). The second term of expression
(random delay) in the denominator contains expression
1 X, which practically means that the curve of delay
value by increasing the values of saturation level (X)
asymptotically to an infinite value. When the q/c ratio
N. ELAR
(X) is greater than 1, an overflow delay model is
applied. The overflow delay model of Equation 2, ho-
wever, has an overflow delay of 0 when X = 1.00, and
increases uniformly with increasing values of X
thereafter.
Websters delay model has the few disputed facts:
Values of random time delay is not infinite, in
situation when the degree of saturation is close or
equal to 1;
Real congestion in situation X=1 doesnt exist,
although individual cycle failures due to random
arrivals do occur;
The overflow delay value 0 at X = 1 is also
unrealistic. The additional delay of individual cy-
cle failures due to the randomness of arrivals is not
reflected in this model.
2.3. Delay model in HCM 2010
Delay model in the HCM (Highway Capacity
Manual) 2010 [7] is identical to delay model in
previous edition (HCM 2000). Average total delay is
given in the next expression:
d-average total delay (s/veh)
PF- progression adjustment factor)
d1- uniform delay component (s/veh)
d2- overflow delay component (s/veh),
d3- delay due to pre-existing queue (s/veh).
Expression for uniform delay corresponds to
Websters formulation of this delay type.
0.5
, s- saturation flow rate (veh/s)
Overflow delay (d2) includes random and over-
flow delay. Their value depends on conditions in the
intersection approach (degree of saturation), analysis
period (T) and controller setting (k).
d 900T x 1 x 1 (5)
x- degree of saturation, or volume to capacity ratio (-) (-)
C- cycle lenght (s)
z effective green time (s)
T- analysis period (h)
k-incremental delay factor for controller settings (-)
l-upstream filtering/metering adjustment factor (-)
C-capacity (veh/h)
Beside the main expression terms, for the first time
is introduced the third term (d3), to evaluate delay
value that occurs as residual delay for oversaturation
queues, that may have existed before the analysis
period..
ANALITICAL DELAY MODELS FOR SIGNALIZED INTERSECTION
2.4. Akcelik delay model and delay model in HBS
2001
In the analyzed models two types of delay are
existed; uniform and overflow delay. Analytical for-
mulation for uniform delay, in both models, is taken
from Webster model. Overflow delay are defined in
the function of average oversaturated queue (number
of vehicle that remains in queue after expiration of
available green time). The queue defined in this way
can exist in few individual cycles or during whole
period of observation). Total overflow delay in Ake-
lik model [1] is computed using next equation:
(6)
Average overflow queue in vehicles is determined
with next expression [1]:
CT 12 x x
N x 1 x 1
4 CT
x x
N 0 (7)
D total overflow delay (veh s)
N0 average overflow queue (veh)
(3)
Ccapacity (veh/h)
T analysis period (h)
x- degree of saturation, or volume to capacity ratio (-)
xo threshold saturation value (-)
Threshold saturation value is defined as degree of
saturation value bellow witch the average overflow
queue is approximately zero. It is given with the
following expression [1].:
0.67 (8)
(4)
z effective green time (s)
Analyzing period duration (T) doesnt have
important influence on N0, for degree of saturation less
than 0.9. Appearance of the overflow queue, in this
zone, is possible only because of random variations of
flow. On the next diagram (Picture 5) is shown the
average vehicle number in the queue as a function of
the degree of saturation.
figure 5 Average overflow queue as a funcion of the
degree of saturation
TECHNICS TRAFFIC (2014) 99
N. ELAR ANALITICAL DELAY MODELS FOR SIGNALIZED INTERSECTION
Overflow delay, in HBS 2001 (Handbuch fr die
Bemessung von Straenverkehrsanlagen 2001) [4] is
given by following equation:
= ( ) (9)
d average overflow delay (s/veh)
C cycle lenght (s)
z effective green time (s)
S- saturation flow rate (veh/h)
NGE- average number of unserved vehicles during the cycle (veh)
In the model are defined three different scenarios
of oversaturated delay value determination. In the case
of X0.65, the overflow delay is negligible small or
equal to zero. For the degree of saturation between
0.65 and 0.90, delay has linear increasing trend. Over
0.90, the overflow delay becomes dependent from the
observation period, or it becomes dependent from
number of cycles in which is realized.
2.5. Evaluation of actual analytical delay models
Analyzed analytical models have unique approach
to delay determination. In model establishment are
used two basic parameters: vehicle arrival type and
traffic flow conditions. In this context, the total delay
is shown as realization of three delay types: uniform,
random and overflow delay..
Analytical formulation of uniform delay is identi-
cal in all analyzed models.
Websters model, formation of random and
overflow delay are formulated in two different models
(static and time dependent model). Problem of real
interpretation of delay, in the models, is in conditions
close to saturation. (X=1).
In other models formulation, random and overflow
delay, is described by unique expression term in the
model. Analysis period duration (T), as main parame-
ter, in different models have different function.
In HCM model time component has constant in-
fluence on additional delay (d2). Also, time, in fun-
ction of number of unserved vehicles in HCM model
exist in the third term of expression (delay due to pre-
existing queue).
In Akelik model formation of random delay is
connected to threshold value of the degree of saturation
(equation 7). Analysis period has important influence
on delay only in case X > 0.9. In the HBS 2001
formation of random delay is connected to fixed values
of the degree of saturation (X=0.65). Evaluation time,
depending on degree of saturation, has four step scale
of influence on delay.
Based on presented facts we can form next con-
clusions:
Random and overflow delay require unique model
formulation;
100 TECHNICS TRAFFIC (2014)
Analysis period, in the function of the degree of
saturation, has major impact on total delay;
Q/C ratio in the range of 0.8 to 1.2, in all models,
represent adjective zone for delay modeling.
3. COMPARATION RESULT OF DIFFERENT
ANALITICAL MODELS APPLICATION
In this part is shown comparative analysis of delay
that are result of different analytical models applica-
tion, considering determined real values.
Real delay values are formed by using experime-
ntal research methods (MEIS) [3]. Research method is
based on microscopic approach to determinate average
delay.
Experimental method is conducted at two signa-
lized intersections, in the undersaturated flow conditi-
ons (degree of saturation between 0.5-0.6). Process
parameters (traffic volume, saturation flow rate and
actual traffic signal parameters), obtained in research,
are used as input parameters for analytical models
Table 1. Delay values obtained by different analytical
models
Average delay (s/veh)
Int. Research W HCM HBS&Ak
1 32.5 27.8 32.3 29.2
2 25.4 19.0 22.1 19.3
W Websters model
HCM HCM model
HBS&AkAkceliks and HBS2001 model
Based on comparison of obtained results, follo-
wing facts are concluded:
Determined real delay values (delay values deter-
mined in research) are higher than those obtained
by using analytical models;
Different analytical models obtained similar delay
values in observed intersection. The results obta-
ined by using HCM delay models differ from this
conclusion.
Similar obtained delay values on intersections are
the result of similar/the same formulation of analytical
model for prevalent research conditions. (Uniform de-
lay model). Higher delay values obtained by using
HCM model, (the minimum deviation from values ob-
tained in research) are effect of HCM model structure
(second term of equation).
4. CONCLUSIONS AND RECCOMENDATIONS
Determination of delay, as measure effectiveness
and input parameter in evaluation procedure, are based
on usage of objective experimental and nonexperi-
mental method.
N. ELAR ANALITICAL DELAY MODELS FOR SIGNALIZED INTERSECTION

Analyzing actual analytical delay models, it can be REFERENCES

concluded that all models, on identical way interpret
delay in unsaturated conditions (uniform delay). Delay
modeling that appears as consequence of random
vehicle arriving and delay formulation for oversatu-
rated conditions (congestions), it differs in different
models.
Generally, in all models, analytical formulation of
process for degree of saturation in the range of 0.8 to
1.2 is uncertain, when there is expressed nonlinear
correlation between mentioned parameters and delay.
Directions of existing researches in the world, in the
field of analytical models development, are focused on
traffic process modeling in oversaturated conditions.
Real delay values are higher than values deter-
mined by analytical models. Obtained differences in
delay values go in range from 10 to 30%. The mini-
mum deviations are obtained by using HCM delay mo-
del (differnce about 10%).
Based on results it can be concluded that either of
analyzed models doesnt interpret real delay values, in
undersaturated conditions.
Direction of further research, in the domain of
application of analytical delay models in domestic
conditions, should be directed to calibration procedure
of relevant analytical delay model.
Taking in consideration restrictions from the
aspect of necessary research volume and time for
calibration procedure realization, recommendation in
domestic engineer practice, in the process of delay
determination, is applied analytical model from HCM
2010.
[1] Akcelik, R., Traffic signals: Capacity and timing
analysis, Australian Road Research Board, 1980.
[2] elar, N., Istraivanje vrednosti zasienog saobraaj-
nog toka na osnovnoj ulinoj mrei grada Beograda,
Tehnika , vol. 54, str.1-6, 2007.
[3] elar, N., Prilog analizi saobraajnog procesa na si-
gnalisanoj raskrsniciDoktorska disertacija, Saobra-
ajni fakultet, Beograd, 2013.
[4] FGSV., Handbuch fr die Bemessung von Stra-
enverkehrsanlagen, Forschungs -gesellschaft fr
Straen und Verkehrswesen. Kln: FGSV Verlag
GmbH., 2001
[5] Roess, R. P., Prassas, E. S., & McShane, W. R.
Traffic Engineering Fourth Edition, Pearson
Prentice Hall, 2011.
[6] Stanic B., Tubic, V., Celar, N.: Straight Lane Satura-
tion Flow and Its Rate in Serbian Cities, -Transport,
vol. 26, pp. 329-334, 2011.
[7] TRB, Highway Capacity Manual (HCM 2010), D.C:
National Research Council, Washington, 2010.
[8] Vukanovi, S., Proraun kapaciteta i nivoa usluge na
signalisanim raskrsnicama po metodi HCM iz 2000.
godine, Tehnika, vol. 51,str. 9-18, 2004.
[9] Vukanovi, S., Vremenski gubici, potronja goriva i
kapacitet kao kriterijumi za upravljanje saobraajem
putem svetlosnih signala Magistarska teza, Saobra-
ajni fakultet, Beograd 1980.
[10]Webster, F. V., Traffic signal settings, Her Majesty's
Stationery Office, London, 1958.

REZIME
ANALITIKI MODELI VREMENSKIH GUBITAKA ZA SIGNALISANU RASKRSNICU

Vremenski gubici, definisani kao dodatno utroeno vreme u realizaciji putovanja, predstavljaju osnovni
pokazatelj kvaliteta realizacije saobraajnog procesa na raskrsnici i ulazni parametar u postupak
optimizacije rada svetlosnih signala. Njihova vrednost se moe utvrditi primenom neeksperimentalnih
(analitikih i simulacionih) i eksperimentalnih metoda istraivanja. Analitiki modeli obezbeuju brz,
jednostavan postupak procene vremenskih gubitaka. U radu je izvren postupak uporeivanja vrednosti
vremenskih gubitaka dobijenih primenom relevantnih analitikih modela i realnih vrednosti utvrenih
istraivanjem. Na osnovu dobijenih rezultata zakljueno je da niti jedan od analiziranih modela u
potpunosti ne interpretira realne vrednosti vremenskih gubitaka, kao i da se najmanje odstupanje u
proceni vrednosti gubitaka ostvaruje primenom analitikog modela iz Prirunika HCM 2010.
Kljune rei: signalisana raskrsnica, vremenski gubici, modeli gubitaka

TECHNICS TRAFFIC (2014) 101

 TE H NI K A
asopis "TEHNIKA" je na vodei nauno-struni asopis iz oblasti tehnikih
nauka i struka koji, u izdanju Saveza inenjera i tehniara, Srbije kontinuirano izlazi
69 godina. U asopisu se objavljuju radovi svrstani po oblastima: "Novi materijali",
"Nae graevinarstvo", "Rudarstvo, geologija i metalurgija", "Mainstvo", "Elektro-
tehnika", "Saobraaj", "Menadment" i "Kvalitet-IMS, standardizacija i metrologija";
stalne rubrike: pregledi knjiga i asopisa, prikazi naunih i strunih skupova u zemlji i
inostranstvu, informacije o tehnikim novostima kao i propagandne poruke domaih i
inostranih organizacija.
Svi objavljeni radovi obavezno prolaze kroz strunu recenziju ime je obezbeen
visok nauni i struni nivo. Objavljeni radovi se referiu u uglednim svetskim
referalnim asopisima: Geotechnical Abstracts, Metals Abstracts, Chemical Abstracts,
Electrical and Electronics Abstracts, Science Abstracts, Ergonomics Abstracts i
referalnim urnalima VINITI-a.
Takoe se obrauju za inostrane baze podataka: SAIDC-el (ISKRA, SL), SAIDC-
gr (CTK, SL), INSPEC (IEEE, UK), METADEX (M. I., UK), CASEARCH (CA,
USA).
Ministarstvo prosvete, nauke i tehnolokog razvoja Republike Srbije svrstalo je
asopis Tehnika u prvu kategoriju asopisa ove vrste.
asopis izlazi dvomeseno na oko 180 strana, formata A4 i distribuira se u zemlji i
inostranstvu (u oko 30 zemalja) putem pretplate i razmene.

Redakcija
SAVEZ INENJERA I TEHNIARA SRBIJE
Kneza Miloa 7a/I, 11000 BEOGRAD
Telefon: (011) 32-35-891, Faks: (011) 32-30-067
www.sits.rs
e-mail: tehnika@sits.rs i office@sits.rs
 ISSN 1450-9911

TECHNICS

MANAGEMENT
Menadment Gestion Management -

YEAR 64 2014.

EDITOR-IN-CHEF:
Prof. Vujadin V e o v i , D.Sc, Engineering Academy
of Serbia, Belgrade

EDITORIAL COUNCIL:

Prof. Petar J o v a n o v i , D.Sc, Project Management

College, Belgrade

Academician Prof. Leonid Avramovi B a r a n o v,

Moscow State University, MIIT, Moscow

Prof. Sanja P e t r o v i , D.Sc, Nottingham University

Business School, Nottingham, UK

Prof. Dejan P e t r o v i , D.Sc, University of Belgrade,

Faculty of Organizational Sciences, Belgrade
Prof. Ilija osi, D.Sc, University of Novi Sad, Faculty of
Technical Sciences, Novi Sad
Prof. Neboja B o j o v i , D.Sc, University of Belgrade,
Faculty of Traffic and Transport Engineering,
Belgrade

Prof. Vesna Spasojevi B r k i , D.Sc, University of

Belgrade, Faculty of Mechanical Engineering,
Belgrade

Prof. Miodrag B u l a t o v i , D.Sc, University of

Montenegro, Faculty of Mechanical Engineering,
Podgorica

EDITORIAL OFFICE: Union of Engineers and

Technicians of Serbia, 11000 Belgrade, Kneza Miloa 7a/I,
Tel. +381/11/ 32 35 891, Fax +381/11/ 32 30 067
A. VUKOVI et al. STRATEGIC MANAGEMENT OF ENERGY EFFICIENCY PROJECT...

Strategic Management of Energy Efficiency Project

Portfolio in Public Buildings

ALEKSANDAR M. VUKOVI, University of Belgrade, Original scientific paper

Faculty of Organizational Sciences, Belgrade UDC: 005.8 ; 620.92:725
MARKO M. MIHI, University of Belgrade,
Faculty of Organizational Sciences, Belgrade
DEJAN . PETROVI, University of Belgrade,
Faculty of Organizational Sciences, Belgrade

Energy efficiency projects in public buildings contribute to the realization of many benefits, both for
users of public buildings, and for the community. In this regard, the implementation of these projects is
often associated with the realization of broader goals of the community. Also, achieving these goals
often includes projects in various fields, while their implementation is complementary and contribute to
the creation of synergy. This study identifies three models of the potential contribution of the portfolio
of energy efficiency projects in public buildings in the implementation of strategic goals. These models
include the linking of these portfolio with projects in other fields. The aim of this paper is to highlight
the importance and role of the energy efficiency project portfolios in public buildings in the
implementation of national and local strategies and achieving the goals of the community. The focus of
this paper is on the description of each model in terms of its objectives, organisation of projects, fields
that projects belong, and the like.
Key words: energy efficiency, public buildings, project portfolio, strategy.

1. INTRODUCTION folio of nergy efficiency projects in public buildings

Energy efficiency project in a public building is a
complex, unique endeavor of implementation of ene-
rgy efficiency measures on the selected building, whi-
ch are implemented within projected duration and cost,
and in order to achieve energy savings and reduction
of greenhouse gas emissions with the same or better
living and work conditions for users of the building [1].
Whereas public buildings are within the jurisdiction of
institutions, such as ministries, state agencies or local
governments, it often happens that the funds allocated
for the improvement of energy efficiency in public
buildings focus on the implementation of several
projects. Having this in mind, in order to maximize the
overall benefits of the projects, there is often
conducted the selection process of projects that will be
included in a joint funded portfolio. Value of the port-
Authors address: Aleksandar Vukovi, University
of Belgrade, Faculty of Organizational Sciences, Bel-
grade, Jove Ilia 54
Paper received: 05.07.2013.
Paper accepted: 08.07.2013.
can be reflected in the size of the energy savings,
reduction of greenhouse gas emissions, the level of
satisfaction of public building users by improved
living and working conditions, as well as other social
benefits [2]. The effectiveness of the energy efficiency
project portfolio in public buildings is the extent to
which the portfolio results correspond to the objectives
of public buildings users and the national objectives
[3]. Based on the above definitions it can be concluded
that the implementation of these portfolios is
inextricably linked to the goals of the community [4].
Achieving these goals contributes to the sustain-
nability and development of the community. The effe-
cts brought by the realization of energy efficiency
project portfolio in public buildings can be comple-
mentary to the effects of projects from other fields,
while those fields does not necessarily refers to the
energy. The combination of these effects through the
implementation of program or portfolio can be one of
the main tools of strategic management in a variety of
fields that are more or less associated with energy
consumption.

TECHNICS MANAGEMENT (2014) 105

A. VUKOVI et al. STRATEGIC MANAGEMENT OF ENERGY EFFICIENCY PROJECT...
2. STRATEGIC MANAGEMENT OF PROJECT
PORTFOLIO
An important element of any project is definition
of a project strategy. The project strategy is a direction
that project should be managed so it will bring
successful outcome with respect to characteristics of
the environment [5]. This strategy is a foundation of
strategic project management. The idea of strategic
project management is that the project has to be
managed in a way that will support the business stra-
tegy and sustainability of organization [6]. The project
results are not only aimed at achieving the objectives
of the project as such, but must contribute to the
achievement of strategic goals and business results of
the organization as a whole [7]. Therefore, project
management is seen as an important tool of im-
plementation of business strategies [8], and often as a
significant strategic competitive advantage [9]. Inte-
grating project management and strategic management
is possible regardless of the size of projects, the
methods, the number of implementation sites and type
of industry that an organization belongs [10]. As an
important factor of project success appears exploita-
tion of technical and market synergies between the
projects in the portfolio, bringing the total portfolio
success greater than the sum of individual project su-
ccesses [11]. This raises the need to extend the concept
of strategic project management in the strategic project
portfolio management.
Strategic project portfolio management is a new
approach to the development of project management
that allows a strong link between organizations stra-
tegy and the strategy of creation and implementation
of projects and programs that are included in project
portfolio [12]. Managing a group of projects should be
linked to the organization's strategy, in order to ensure
successful implementation of the strategy [13], and the
successful implementation of the portfolio [14]. The
business strategy defines the structure of project
portfolio and has a crucial influence on its success [11].
However, it should be added that the structure of the
portfolio is influenced, not only by the organization's
strategy and its objectives, but also by the constraints
from the environment. These constraints are regarded
to resources, duration, costs, geographical distribution
of projects, readiness for implementation, available
infrastructure, etc. [3]. In this regard, it is possible to
form some of the following types of energy efficiency
projects portfolios in public buildings [2]:
Portfolio of demonstration projects which invo-
lves the implementation of a relatively small nu-
mber of projects, but large (or full) utilization of
the potential for energy efficiency improvements
in public buildings. The measures applied in this
106 TECHNICS MANAGEMENT (2014)
case mostly belong to the type of expensive me-
asures.
Portfolio with low-cost projects is related to the
implementation of projects with partial utilization
of the potential for energy efficiency improve-
ments in the hundreds (and sometimes over a tho-
usand) of public buildings. Here is applied the so-
called low-cost measures.
The combined model of energy efficiency project
portfolio in public buildings means that in a num-
ber of buildings (several dozen) are implemented
the demonstration projects, while in other buildi-
ngs (several hundred) are implemented the low-
cost measures.
Implementation of the energy efficiency project
portfolio in public buildings may be suitable for achi-
eving the strategic goals associated with energy con-
sumption. The reason is that in some countries due to
lack of maintenance, public buildings were for years or
decades exposed to decay of the energy infrastructure,
and therefore have great potential for improving
energy efficiency and consequently achieving si-
gnificant benefits for the community.
3. ROLE OF ENERGY EFFICIENCY PROJECT
PORTFOLIO IN PUBLIC BUILDINGS IN
ACHIEVING STRATEGIC GOALS
As noted earlier, the effects and implementation of
energy efficiency projects in public buildings can be
complementary to the projects, programs or portfolios
from other fields. Having this in mind, it is possible to
associate these efforts in order to achieve strategic
goals of the city, region or even a country. Form of
association that will be shown in this paper, concerns
the selection of strategic portfolio that includes
projects in various fields, with respect to selection and
portfolio optimization procedure given in [15]. Based
on the name of the portfolio, it can be seen that
strategic portfolio is directed towards the realization of
strategic goals that exceeds the targets of energy
efficiency projects themselves. Energy efficiency pro-
ject portfolio in public buildings thus becomes a sub-
portfolio of strategic portfolio. In this way, there can
be utilized the advantage of synergy that is provided
by simultaneous realization of several energy effi-
ciency projects in public buildings, and also making a
contribution to achieving strategic goals.
Based on current practices in the implementation
of the energy efficiency project portfolios, it is possible
to identify three types of strategic portfolios related to
energy consumption, or, by the other words, three wa-
ys in which portfolios in public buildings contribute in
achieving the strategic goals. These strategic portfolios
are:
A. VUKOVI et al. STRATEGIC MANAGEMENT OF ENERGY EFFICIENCY PROJECT...
Strategic portfolio of efficiency improvement of
the overall energy system;
Strategic portfolio of efficiency improvements in
various fields of energy consumption;
Strategic portfolio, which involves the implemen-
tation of various projects and portfolios.
These strategic portfolios will be further explained
below.
3.1 Strategic portfolio of efficiency improvement of
the overall energy system (SP 1)
The energy system of a city, region or country is
comprised of three main elements, and those are the
production or supply of energy, its transmission to the
users and energy consumption [16]. The degree of
stability of the energy system can be defined as a th-
reshold that changes in the elements of the energy
system do not compromise the ability of unobstructed
meeting the energy needs of the users. A high degree
of energy stability is a prerequisite for economic de-
velopment and social welfare. The degree of energy
system stability can be increased by improving the
efficiency of the individual elements or by improving
the efficiency of the entire energy system. The form of
improvements depends on the technical characteristics
of certain elements, whereby it is primarily referred to
the level of energy loss that occurs in them. In addition
to establishing energy stability at the highest level
possible, it is necessary to establish equability of
technical capabilities of all the system elements.
Namely, the obsolescence of energy infrastructure for
production, supply or consumption, leads to an
increase of energy loss in those elements. The im-
provement of one of the elements of the energy system
does not necessarily lead to efficiency improvement of
the entire system. In fact, in time, equipment
obsolescence would increase the loss in other ele-
ments; thereby the previous efficiency improvement of
one element would lose its meaning. By simultaneous
improvement of all the system elements, the degree of
stability of the entire system is increased, which is then
capable of meeting energy needs of its users for a long
term [17].
Energy consumption in public buildings, as previ-
ously mentioned, has an important share in total energy
consumption. Efficiency improvement in this sector, in
some cases, may significantly contribute to increased
stability of the energy system of a city, region or
country. In this regard, the energy efficiency project
portfolio in public buildings may be included in the
strategic portfolio of improving energy stability, which
would include the improvements of two or three
elements of the energy system. The goal of the st-
rategic portfolio would be to increase the degree of
stability of the overall energy system of a city, region
TECHNICS MANAGEMENT (2014)
or country through the implementation of compleme-
ntary projects that improve efficiency, both in terms of
energy consumption, and its production and trans-
mission. In accordance with the strategic goal of in-
creasing the energy system stability, as well as taking
into account the constraints from the environment, it is
necessary to choose appropriate public buildings that
will be included in the portfolio, as well as other
elements of energy infrastructure, such as heating pla-
nts or sections of heating pipes system.
3.2 Strategic portfolio of efficiency improvements
in various fields of energy consumption (SP 2)
In the second decade of the 21st century, the pro-
blems related to energy sustainability are beginning to
take shape of the crisis that the world could be facing
in the future. The fossil energy sources are still the
dominant source of energy. Most of the world's re-
serves of fossil energy are concentrated in relatively
few countries. In this way, for other countries, which
are not rich in fossil energy, there is a problem of in-
sufficient energy security. The concept of energy se-
curity is associated with a variety of risks, including
the risk of unreliable and unsustainable sources of
energy supply, the risk of being unable to meet the
demand for energy and the risk of excessively high
energy prices [18]. Energy consumption is very im-
portant for the sustainability and development of a co-
untry [19]. Therefore, the dependence on energy im-
ports could potentially pose a prerequisite of economic
and political dependence. Also, the funds intended for
energy imports can be extremely large and hamper
overall economic development [20]. Taking into
account the above, it becomes understandable why
many countries nowadays are taking extensive
measures to reduce import dependence and the cost of
energy imports. Measures, in this regard, can be di-
vided into two types: measures aimed at increasing the
production of renewable energy and measures aimed at
reducing energy consumption at the national level.
Renewable energy sources are virtually unlimited and
available to all countries. Increase in production of
energy from renewable sources, contributes to the
reduction of energy imports [21]. The problem though
is the fact that the production of energy from some
renewable sources, due to the complex technology is
very expensive. Another alternative is to improve
energy efficiency and reduce energy consumption at
the national level [22]. Here the emphasis is on the
implementation of energy efficiency measures in the
various fields of energy consumption, such as
buildings, industry, transport and utilities. Im-
plementation of such measures contributes to the re-
duction in energy consumption and therefore contri-
butes to the reduction of energy import. Implementa-
tion of energy efficiency measures over the long term
107
A. VUKOVI et al. STRATEGIC MANAGEMENT OF ENERGY EFFICIENCY PROJECT...
may prove to be more cost effective than construction
of plants for the renewable energy production, as well
as energy import.
Pursuant to the share in total energy consumption,
and consistent with the potential for cost reduction,
implementation of energy efficiency measures in
public buildings can make a significant contribution to
the activities of the reduction of energy consumption
in the country. In this regard, it is possible to select and
link multiple project portfolios in different fields of
energy consumption that involves a portfolio of energy
efficiency in public buildings. Principle of selecting
the projects that should become a part of a strategic
portfolio is based on the total savings that this
endeavor would achieve.
3.3 Strategic portfolio which includes the
implementation of various projects and
portfolios (SP 3)
Energy efficiency projects in public buildings can
be complementary to the projects and activities of ot-
her non-energy fields and together with them to be a
part of a unique program. Some of the goals of such
programs are to improve the quality of public institu-
tions, to reduce the pollution, to raise the awareness of
rational use of energy, and so on.
Implementation of energy efficiency measures can
contribute in achieving energy savings, and thus
relieve the public sector of excess costs, allowing more
funds to be directed toward development. Also, by
implementation of the abovementioned measures, it is
possible to improve the living and working conditions
in public buildings [1]. The implementation of other
development projects, such as the equipment moder-
nization, training and development programs for em-
ployees, etc., together with energy efficiency projects,
makes it possible to achieve significant improvements
in the quality of public institutions.
Energy efficiency projects can contribute to the
reduction of greenhouse gas emissions through the re-
duction of energy consumption [3]. The potential for
achieving environmental benefits through reduction of
energy consumption in public buildings can be used in
implementation of national programs of improving
state of the environmental of the country. These
programs can relate only to reducing overall greenho-
use gas emission on the country level, but may also
include other aspects of environmental conservation.
When it comes to raising awareness of rational use
of energy, energy efficiency project portfolio can be
implemented in parallel with the projects of promotion
of energy-efficient behavior. In fact, in countries
where awareness of the previously mentioned issue is
not on high level, implementation of the portfolio of
demonstration projects of energy efficiency in public
108 TECHNICS MANAGEMENT (2014)
buildings can contribute in solving this problem [2]. As
its name implies, the objective of the demonstration
projects is to present the benefits of energy efficiency
improvements to a wider population. For this purpose,
as suitable sites for these projects are identified public
buildings with high potential for achieving energy
savings and other benefits. Demonstration of
significant positive effects of energy efficiency
projects tends to motivate people and organizations to
implement similar projects. The promotional activities
on the implementation of these projects may include
the media reports about the implemented energy
efficiency projects in public buildings, setting table in
plain sight with a description of the implemented
energy efficiency measures, lectures in schools about
energy efficiency, etc [1]. The aforementioned
activities may be a part of a wider campaign of
promotion of energy efficient behavior and
implementation of energy efficiency measures [23].
Promotional campaign can also includes other acti-
vities such as promotional videos, reports and articles
about energy efficiency, conferences and awards for
energy efficiency, and so on.
4. DISCUSSION
Linking energy efficiency project portfolio in
public buildings with strategic goals can be achieved
in a variety of ways, as shown. The interaction between
the energy efficiency project portfolio in public
buildings and strategy is mutual. The strategy has
crucial influence on the way of selecting the energy
efficiency project portfolio in public buildings, and the
potential energy savings, as well as potential problems
in achieving planned objectives of the mentioned
portfolio influence the definition or change of the
strategy.
When it comes to the type of the portfolio of ene-
rgy efficiency projects in public buildings that is se-
lected depending on the goals of the strategic portfolio,
it is possible to see some regularity. With SP1 and SP2
and in some cases with SP3, the emphasis is on
achieving the highest possible total energy savings.
Based on [2], it is possible to conclude that the greatest
total savings are made by the portfolio of
demonstration projects, because it tends to full using
of potentials for improving energy efficiency in se-
lected buildings. However, when it comes to the ef-
ficiency of investment, i.e. comparison between ene-
rgy savings and costs, a portfolio of low-cost projects
is a better solution. In this regard, the main factor of
choosing the type of portfolio is available budget. In
the case of a relatively large budget, priority is given
to the portfolio of demonstration projects, and alter-
native solution is a combined portfolio, while in the
case of a relatively small budget, preference should be
A. VUKOVI et al. STRATEGIC MANAGEMENT OF ENERGY EFFICIENCY PROJECT...
given to the portfolio with low-cost projects. On the
other hand, when it comes to the strategic portfolio
which should contribute to raising awareness of
rational use of energy, the emphasis is on the savings
in particular public buildings, and this is the primary
criterion, not the available budget. For strategic port-
folio that aims to improve the state of environment, the
emphasis is on reduction of greenhouse gas emissions
through implementation of energy efficiency projects
in public buildings. This reduction, which depends on
the type of energy used in public buildings, may be
disproportionate compared to the reduction of energy
consumption. In this regard, part of the energy
efficiency project portfolio in public buildings, will
become the buildings that use the energy produced
from the "dirty" fuels, such as petroleum products or
coal, but not the buildings that use natural gas or
biomass.
The potential savings in public buildings play an
important role in determining which fields of energy
consumption should be included in the strategic port-
folio. For example, the stability of the energy system
in the SP 1 can be raised to a satisfactory level, even if
on the consumption side is implemented only the
energy efficiency project portfolio in public buildings.
However, if the potential savings is not sufficient, it is
necessary to include in strategic portfolio, projects and
portfolios from other fields of energy consumption.
Implementation of presented strategic portfolios can
last for years, as it points to the necessity of constant
review of planned and achieved effects. Failure in
achieving planned savings in selected public buildings,
imposing the need to expand the strategic portfolio
with projects in other public buildings or energy
efficiency projects in other fields of energy
consumption in order to meet the strategic goals. This
case is particularly present in SP1 and SP2. In SP3,
with the focus on non-energy goals, energy saving
effects can be compensated by improvements or acti-
vities outside the energy consumption sector.
5. CONCLUSION
Numerous studies have confirmed the necessity of
linking project management with strategic goals. In
accordance with the above, there have been created the
new approach of portfolio management called the
strategic portfolio management. This approach is fo-
cused on achieving strategic goals. In this paper we
analyzed the role of energy efficiency project port-
folios in strategic portfolios, which should contribute
to achieving the goals of the community. Having this
in mind, we have identified three types of strategic
portfolios based on their goals, as well as types of
projects that are included in portfolio. These strategic
portfolios are related to the improvement of energy
TECHNICS MANAGEMENT (2014)
system stability, reduction of energy consumption at
the national level, and the achievement of national
goals through combination of projects from various
fields. Some of these goals are to raise the quality of
public institutions, to improve the state of environ-
ment, to raise the awareness of rational use of energy
and the like. Mainly through the reduction of energy
consumption, but also through other effects, an energy
efficiency project portfolio in public buildings achieve
complementarities with other projects within the
strategic portfolio and contribute to the achievement of
goals of the community. Linking energy efficiency
project portfolio in public buildings with the strategy
has a crucial influence on the portfolio selection, but
also, the strategy has been defined or changed,
influenced by the effects of the portfolio. Management
of energy efficiency projects in public buildings in
accordance with national strategies guarantees the full
effectiveness of these projects.
REFERENCES
[1] Mihi, M., Petrovi, D., Vukovi, A., Possibilities
of application of Cost-Benefit analysis in energy
efficiency projects in public buildings, Economic
Themes, 49, 3, p. 355-378, 2011.
[2] Mihi, M., Petrovi, D., Vukovi, A., Energy Effi-
ciency Project Portfolio Optimization for Public
Buildings, Metalurgia International, 17, 6, p. 166-
174, 2012.
[3] Mihi, M., Petrovi, D., Vukovi, A., Obradovi,
V., urovi, D., Application and Importance of Cost-
Benefit Analysis to Energy Efficiency Projects
Implemented in Public Buildings: The Case of Ser-
bia, Thermal Science, 16, 3, p. 915-929, 2012.
[4] Mihi, M., Vukovi, A., Vukovi, M., Benefits
management in energy efficiency projects in public
buildings, Management, 17, 62, p. 57-65, 2012.
[5] Artto, K., Kujala, J., Dietrich, P., Martinsuo, M.,
What is project strategy?, International Journal of
Project Management, 26, 1, p. 4-12, 2008.
[6] Patanakul, P., Shenhar, A. J., What project strategy
really is: The fundamental building block in strategic
project management, Project management Journal,
43, 1, p. 4-20, 2012.
[7] Todorovi, M., Beri, I., Spasi, ., Drobnjakovi,
S., Strategic management and strategic project ma-
nagement, VIII Proceedings, SPIN 2012., p. 317-
324, 2012.
[8] [Grundy, T., Strategic project management and stra-
tegic behaviour, International Journal of Project
Management, 18, 2, p. 93-103, 2000.
109
A. VUKOVI et al. STRATEGIC MANAGEMENT OF ENERGY EFFICIENCY PROJECT...

[9] Judgev, K., Mathur, G., Project management ele- [17] Rutherford, J. P., Scharpf, E. W., Carrington, C. G.,
ments as strategic assets: preliminary findings, Ma- Linking consumer energy efficiency with security of
nagement Research News, 29, 10, p. 604-617, 2006. supply, Energy Policy, 35, 5, p. 3025-3035, 2007.

[10] Mihi M, Strategic orientation of project manage- [18] Chester, L., Conceptualising energy security and
ment in Serbia, XIII International proceedings, making explicit its polysemic nature, Energy Policy,
YUPMA 2009, p. 17-21, 2009. 38, 2, p. 887-895, 2010.

[11] Meskendahl, M., The influence of business strategy
on project portfolio management and its success
A conceptual framework, International Journal of
Project Management, 28, 8, p. 807-817, 2010.
[12] Beri, I., Drobnjakovi, S., Spasi, ., Jovanovi, F.,
New aproach Strategic project portfolio
management, VIII Proceedings, SPIN 2012, p. 260-
265, 2012
[13] Dietrich, P., Lechtonen, P., Successful management
of strategic intentions through multiple projects
Reflection from empirical study, International
Journal of Project Management, 23, 5, p. 386-391,
2005.
[14] Mller, R., Martinsuo, M., Blomquist, T. Project
portfolio control and portfolio management perfor-
mance in different contexts, Project Management
Journal, 39, 3, p. 2842, 2008.
[15] Ghasemzadeh, F., Archer, N. P., Project portfolio
selection through decision support, Decision Support
Systems, 29, 1, p. 73-88, 2000.
[19] Tsani, S. Z., Energy consumption and economic
growth: A causality analysis for Greece, Energy
Economics, 32, 3, p. 582-590, 2010.
[20] Dowling, P., Russ, P., The benefit from reduced
energy import bills and the importance of energy
prices in GHG reduction scenarios, Energy Econo-
mics, 34, Supplement 3, p. S429-S435, 2012.
[21] Ramchandra, P., Addressing Indias energy security
and options for decreasing energy dependency, Re-
newable and Sustainable Energy Reviews, 14, 9, p.
3014-3022, 2010.
[22] Dixon, R. K., McGowan, E., Onysko, G., Sheer, R.
M., US energy conservation and efficiency policies:
Challenges and opportunities, Energy Policy, 38, 11,
p. 6398-6408, 2010.
[23]Boza-Kiss, B., Moles-Gruesco, S., Urge-Vorsatz, D.,
Evaluating policy instruments to foster energy
efficiency for the sustainable transformation of buil-
dings, Current Opinion in Environmental Sustaina-
bility, 5, 2, p. 163-176, 2013

[16] Ministery of mining and energy of the Republic of .

Serbia, Energy development strategy of the Republic
of Serbia by 2015., Belgrade, 2005.

REZIME
STRATEGIJSKO UPRAVLJANJE PORTFOLIOM PROJEKATA ENERGETSKE
EFIKASNOSTI U JAVNIM ZGRADAMA

Projekti energetske efikasnosti u javnim zgradama doprinose ostvarivanju mnogih koristi, kako za
korisnike javnih zgrada, tako i za iru drutvenu zajednicu. S tim u vezi, realizacija pomenutih projekata
je esto povezana sa ostvarivanjem irih drutvenih ciljeva. Takoe, u ostvarivanju ovih ciljeva esto su
ukljueni projekti iz razliitih oblasti, pri emu se nastoji da je njihova realizacija komplementarna i da
doprinosi stvaranju sinergetskog efekta. U ovom radu su identifikovana tri modela mogueg uea
portfolia projekata energetske efikasnosti u javnim zgradama u realizaciji stratekih ciljeva. Ti modeli
podrazumevaju povezivanje pomenutih portfolia sa projektima iz drugih oblasti. Cilj rada je da istakne
znaaj i ulogu portfolia projekata energetske efikasnosti u javnim zgradama u realizaciji nacionalnih ili
lokalnih strategija, te ostvarivanju ciljeva drutvene zajednice. Fokus rada je na opisu svakog od modela
u smislu ciljeva, organizovanja projekata, oblasti kojima projekti pripadaju, i sl.
Kljune rei: energetska efikasnosti, javne zgrade, portfolio projekata, strategija.

110 TECHNICS MANAGEMENT (2014)

D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...

Modification of Analytic Hierarchy Process (AHP) Method and its Application

in the Defense Decision-making

DARKO I. BOANI, University of Defense in Belgrade, Original scientific paper

Military Academy, Belgrade UDC: 303.732.3:355/359 ; 005:355/359
DRAGAN S. PAMUAR, University of Defense in Belgrade,
Military Academy, Belgrade
BOBAN D. OROVI, University of Defense in Belgrade,
Military Academy, Belgrade

This paper presents a new approach to modification of the method of analytic hierarchy processes (AHP)
using fuzzy logic. The main feature of the approach used in this paper is to determine the interval of
Saatys scale fuzzy numbers based on the decision-makers degree of certainty. This approach allows
the decision-maker with some degree of certainty (which is usually less than 100%) to define which
linguistic expression matches the comparison of optimality criteria. In this way, after the application of
AHP method, the values of criteria functions have been obtained for each of the considered alternatives,
which correspond to the determined value of the degree of certainty. Thus, it is ensured that generating
different sets of criteria functions can be made for more values of the degree of certainty. The model was
tested on a tactical level of decision-making in the defense regarding the selection of objects for
demolition by the Interdiction Group, as a procedure usually accompanied by a greater or lesser degree
of vagueness of the criteria that are necessary to make relevant decisions.
Key words: analytic hierarchy process (AHP), fuzzy logic, decision-making, defense system, Interdi-
ction Group

1. INTRODUCTION cal part of the system of military organization, the right

The decision-making process is a key aspect of
managers' success at all hierarchical levels. All deci-
sion-makers, regardless of the decisions they make,
must go through some form of this process [4]. Both in
theory and practice, one can find different approaches
to decision-making. Special tone to the nature of the
decision-making process is given to the character of
business activity in which the process occurs. In this
regard, decision-making is very distinctive in military
organizations, i.e. systems that represent forms of defe-
nse organization of societies defense systems.
The military organization is a combination of so-
ciological and technical system, as in some of its parts
it is characterized by the relationship man-machine,
and in others, it is a typical social organization. In
addition to the high degree of automation in the tecni-
Authors address: Darko Boani, University of
Defense in Belgrade, Military Academy, Belgrade, Pavla
Juriia truma 33
Paper received: 06.03.2013.
Paper accepted: 22.03.2013.
to make final decision is reserved for the man. This is
particularly true when it comes to the carrying out co-
mbat activities, where "bad decisions" could have far-
reaching consequences. Therefore, the importance of
scientific thinking in the field of decision-making in
the military is very important, as well as the subsequent
implementation of scientific findings into practical ac-
tion.
This paper aims to show the possibilities of applying
the modified method of analytic hierarchy process (AHP)
when making decisions in one part of the defense ope-
ration defining the tasks of the Interdiction Group. In
this paper, the aspects of the decision-making process in
the defense system are defined and displayed as a me-
thodological basis of AHP method. In addition, this paper
presents a new approach to fuzzification of Saaty's scale,
which is used to quantify the parameters of the AHP me-
thod. The fuzzy AHP approach is presented as a support
to the decision-making process at the tactical level of
the defense system on the example of the engagement
of the Interdiction Group (in some cases these deci-
sions are also made at the operational level in the same

TECHNICS MANAGEMENT 64 (2014) 111

D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...
manner). After the description of the position and role
of the Interdiction Group and manner of assigning
tasks to it, the criteria that influence the choice, i.e. ma-
king a decision and determining difficulty criteria, as
well as ranking the alternatives, is being described. The
last part includes discussion and conclusion, where the
findings from the previous units have been analyzed.
2. ASPECTS OF DECISION-MAKING PROCESS
IN THE DEFENSE SYSTEM
A specific characteristic of the defense system is
engagement in peacetime, war and state of emergency
[4]. These conditions are interconnected and require
the system's ability to switch quickly from one to the
other. The nature of the functioning of the defense sy-
stem and its organizational units in peacetime, war and
state of emergency is fundamentally different. Circu-
mstances of war and state of emergency impose a nu-
mber of restrictions and difficulties on decision-ma-
kers. These are the conditions in which the importance
of all decisions increasingly grows, due to the scope
and nature of the possible consequences of bad deci-
sions. Moreover, the time for decision-making is often
reduced, and thus, the speed of decision-making, in
addition to the quality of decision, becomes one of the
primary criteria. All other organizational resources, in
these conditions, present strong limiting factors in the
decision-making process, such as human and material
due to the destructive warfare of opponent, infor-
mation due to the unpredictable actions of the oppo-
nents, the dynamics of the situation and similar.
Due to the specific, predominantly line and staff
organizational structure, decision-making authorizati-
on in the defense system is awarded to the individual
who is the head of specific organizational unit (com-
mander, chief, etc.). This system of authority, in pra-
ctice, often does not leave much space for creativity
and self-initiative in decision-making at other levels of
decision-making. Acting on one's own decision is
usually undertaken only when, for whatever reason,
one cannot get a decision (order) from the superior le-
vel, and the situation requires the use of the unit or the
like. [6]. It is also important to notice one terminolo-
gical distinction between defense system and other sy-
stems, which is present when it comes to levels of the
hierarchy, and thus the decision-making levels. Three
hierarchical levels are identified in the defense systems
- strategic, operational and tactical, which are equi-
valent to the strategic, tactical and operational level of
hierarchy (successively) in business and other systems.
The review of the application of the multi-criteria
decision-making method is aimed at understanding the
impact of scientific methods on building of combat
competence and increasing the defensive power of the
112 TECHNICS MANAGEMENT 64 (2014)
military units and other elements of the defense sy-
stem.
3. METHODOLOGICAL BASES OF AHP
METHOD
AHP belongs to the group of methods for soft
optimization. Looking at the hierarchy from the top,
the goal is at the top, and it is not compared with any
of the other elements. The main problem with compa-
rison in pairs is how to quantify linguistically formu-
lated choices i.e. phrases [7]. In most of the methods
that use comparisons, this is achieved by using the ap-
propriate numerical values, usually expressed through
fractions with whole numbers. When comparison is to
be the expression of the similarity, and not the relative
ratio, the differences of integers may be used instead
of the fractions [16]. In developing the scales of eva-
luation, there are two approaches: linear [11] and expo-
nential [8]. Both approaches are based on certain the-
ories from the field of psychology and in this case, the
first approach is of interest since it is convincingly
dominant in application.
Since psychological experiments have shown that
an individual cannot simultaneously compare more
than 7 2 objects [12], Saaty defined the scale [14],
which has the highest value of 9, the lowest value is 1,
and the increment difference 1. The Saaty's scale is ge-
nerally considered the standard for AHP, and it is used
for comparison in pairs. In addition to the Saaty's scale,
Ma and Zheng [11] proposed the scale in which the
values are fairly distributed in the interval [1/9, 1], and
the other values are their reciprocal values.
Recent analysis has shown that there is neither an
all-purpose scale that is best for all the cases nor the
worst one [17]. Saaty's scale is predominantly applied
in spite of certain drawbacks. One of the drawbacks is
that a half of the scale is linear and the other half is
non-linear. This means that when a decision maker or
analyst performs comparisons in pairs, based on the
semantic preferences from the right column of Saaty's
scale or by means of direct association, numerical
values in the left column are entered in a square matrix
of comparison. Since a ji 1/ aij and aii 1 for each
i, j 1, 2,..., n , matrix A is positive, symmetrical and
reciprocal. Essential information about the preferences
of elements E1 , E2 ,..., En is found only in the upper
matrix triangle, but all methods for its further analysis
are using the reciprocal of the lower matrix triangle.
The key disadvantage of such approaches is the
hard defining of the importance intensity of the
linguistic expressions, and this is overcome by scale
fuzzification. The fuzzification of Saaty's scale is
described in numerous papers [3, 13, 19].
D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...

What is common for all approaches with fuzzified
Saaty's scale is sharp fuzzification of the scale ele-
ments. The term sharp fuzzification represents the
case when for certain fuzzy number T t1 , t2 , t3
certain interval of confidence is determined in adva-
nce, i.e., it is defined in advance, so that the value of
fuzzy number will not be bigger than t3 nor smaller
than t1 . In other words, we are sure that the value of
the linguistic expressions belongs to the closed interval
t1, t3 . The differences among fuzzy versions of
AHP/ANP are mostly those mirroring the way of the
scale fuzzification and the method of the result
defuzzification. Such an approach does not take into
Table 1. Fuzzified Saatys scale
Importance
Definition
intensity
account the effect that the level of uncertainty for the
assessment of the linguistic expressions has on the
final value of criterion functions. The model repre-
sented in this work takes into account the level of
uncertainty marked with parameter .
In this case, the greatest possible uncertainty is
described by the value 0 , while the value 1
corresponds to the situation in which we are sure that
the linguistic expression corresponds to given
comparisons of the optimality criteria. The value of the
parameter can be any number within the interval
[0,1]. Applying the described procedure means the
implementation of Saaty's scale fuzzification 1 [2].
Fuzzified values
Fuzzy number Inverse of fuzzy number

(1, 1, 1) compared with oneself,

1 Same importance 1/ (2 ),1,1/ in other cases
,1, 2 in other cases

3 Weak dominance 3 , 3, 2 3 1 2 3,1 / 3,1 3

5 Strong dominance 5 , 5, 2 5 1 2 5,1 / 5,1 5

7 Very strong dominance 7 , 7, 2 7 1 2 7 ,1/ 7,1 7
9 Absolute dominance 9 , 9, 9 1 2 9 ,1 / 9,1 9

2, 4, 6, 8 Inter-values
x , x, 2 x , 1 2 x ,1/ x,1 x
x 2, 4, 6, 8 x 2, 4, 6, 8

In this way, upper and lower limits of the con- 2 x, x 2 x 9

fidence interval of the fuzzy number are chosen ran- t3 2 x
9, 2 x 9
domly for the given value of the parameter , so that (4)
they are within the limits defined by the expression: Inverse of fuzzy number
t1 t2 , t1 , t2 1 / 9,9 T 1 / t1 ,1 / t 2 ,1 / t3 1 2 x ,1 / x,1 x ,
1
t1 t2 ,
(1)
T t1 , t2 , t3 t2 t2 , t2 1 / 9,9 x 1 / 9,1 is defined as:
t 2 t , t t , t , t 1 / 9,9
3 2 3 2 2 3
1 2 x , x 1 2 x 1 (5)
where value t2 represents value of linguistic expre- 1/ t1 1 2 x
1, 1 2 x 1
ssion in which membership function has the highest
value, i.e. t2 1 . 1/ t2 1/ x, 1/ x 1/ 9,1 (6)
T t1 , t2 , t3 x , x, 2 x 1 / x, 1/9 1 / x 1 / x (7)
Fuzzy number , 1 / t3 1 x
x 1,9 1 / 9, 1/ x 1 / 9
is defined as:
After fuzzification of Saaty's scale, the procedure
x , 1 x x for the implementation of AHP method is the same as
t1 x
1, x 1 with the classical AHP method. Detailed description of
(2) the AHP method phases is given in the literature
t2 x, x 1,9 [9,10]. Defuzzification of fuzzy numbers is performed
(3) by applying the centre of gravity method [3].

TECHNICS MANAGEMENT 64 (2014) 113

D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...
In this way, the values of criteria functions for
every considered alternative are obtained after
application of AHP method. A certain value of para-
meter corresponds to the obtained values of criteria
functions. It is possible to generate various sets of
values of the criteria functions for various values of
parameter . This work will deal with five sets for five
different values of parameter as follows 0,
0.3, 0.5, 0.7 and 1 . ating obstacles of combustible and chemical substa-
Numerous authors in their scientific papers, as well
as the increasing number of the real decision makers in
the field, stress that the quality decision-making based
only on personal thinking or intuition is very rare.
Recently, whenever there is an opportunity, decisions
have been made within the group, through the so-
called group sessions, where opinions are exchanged
and knowledge of individual participants improved,
different experiences of the participants exploited, pa-
ving the way for consensus and achieving the impo-
rtant psychological effect of common interest for the
success of the decision. Such an approach, through the
application of Delphi method [9] was used in this
paper.
Since experts from the related field act as advisors
in most cases of decision-making, described model wi-
ll often be applied in terms of group decision-making.
A detailed overview of application of AHP method in
a group decision-making is given in [15, 20].
4. FUZZY AHP APPROACH AS THE SUPPORT
TO DECISION-MAKING PROCESS IN THE
DEFENSE SYSTEM
The Serbian Armed Forces is the core of the de-
fense system of the Republic of Serbia. It is the pillar
of an armed opposition to military forms of enda-
ngering state's security. Functionally, the Serbian Ar-
med Forces are organized into commands, units and
institutions, and structurally into arms, branches and
services. Engineering is the arm of the Serbian Armed
Forces, whose troops carry out large spectrum of ac-
tivities in peacetime, war and state of emergency. En-
gineering units have a significant place and role in the
execution of combat operations (command and con-
trol, warfare and counter-warfare, and security), as the
essential content of the armed combat in war i.e. armed
conflict. One part of the warfare and counter-warfare
consists of engineering warfare and counter-warfare.
Engineering warfare causes losses to the enemy and
prevents, slows down and channels their movement
and maneuvers of forces and is carried out by
interdiction of facilities, districts and directions [5].
The term interdiction stands for the placement and
making of various artificial and strengthening of
114 TECHNICS MANAGEMENT 64 (2014)
natural barriers in order to slow the pace of opponents'
attacks (limiting and channeling of warfare primarily
of armored and mechanized forces, making landing of
air raids difficult, impeding traffic and supply, causing
losses) and creating favorable conditions for warfare of
one's own forces. Interdiction on land includes the de-
velopment and placement of mines and fortified bar-
riers, creating obstacles and destruction of commu-
nication facilities on them, by flooding the land, cre-
nces and strengthening of natural barriers [18]. The
former is accomplished by permanent and temporary
units composed of engineering, as well as other units.
The most significant temporary units that are orga-
nized out of engineering units are: Interdiction Group,
Group for Additional Interdiction, Group for the reha-
bilitation of roads and the Group for the rehabilitation
of the airport.
Interdiction Group is one of the most significant
temporary engineering structures whose position and
role are defined through the theory and practice of
conducting engineer operations and counter-warfare.
This paper covers the situation when the Interdiction
Group is formed during the combat preparation. The
second case, when the group is organized in the course
of combat operations, will not be considered as an
alternative in this situation has already been selected
(facilities are prepared for demolition). Rules and
regulations governing the forming and operation of the
Interdiction Group emphasize the need for the objects
to be assigned for demolition, but the method of se-
lection of such facilities in the available literature re-
mains undefined, and the solution to this problem re-
lies exclusively on the knowledge, experience and as-
sessment of decision-makers in specific situation. Nu-
merous criteria that influence the ranking and selection
of alternatives (facilities for demolition: bridges, roads,
culverts, overpasses, etc.) requires the use of multi-
criteria decision-making methods.
4.1. Defining criteria for selection of facilities for
demolition
Information about potential facilities that will be
prepared for demolition is collected through reconna-
issance. Based on the collected data, the decision ma-
ker formulates alternative solutions rejecting at the
same time those solutions that do not meet defined
criteria (the facilities that are, due to their characte-
ristics, immediately rejected). The next step for suc-
cessful decision-making is ranking evaluation of
formulated alternatives.
In terms of ranking the alternatives, Delphy met-
hod is used to define criteria and sub-criteria, as well
as to determine their weights. The criteria that the In-
D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...

terdiction Group uses to evaluate the alternati- cts must be considered in two time components, im-
ves/facilities to be prepared for demolition and des- mediately after demolition, as well as in the distant
truction are [1]: future.
Characteristics of the facility (C1) A comparison of the two elements of the hierarchy
By this criterion, the framework calculation of the (the model) is performed by using fuzzified Saaty's
need for human, material and time resources necessary scale. Linguistically expressed, preferences obtained
to prepare the facilities for demolition, which directly by comparing the criteria have been used to create a
depends on the type and characteristics of the facility matrix A.
that is being prepared for demolition and destruction, C1 C2 C3 C4
is done. This criterion considers the two sub-criteria: 1 2 2
1 1
C1 3
time that is necessary for the preparation of the facility
for demolition (C11) and necessary explosive ordnance A C2 3 1 2
2.5
1
needed for the preparation of the facility for demolition C3 2 1
2.5
1
1 2
(C12) C4 1

2 2 2 1
Characteristics of the direction in which the facility
is located in relation to enemy forces (C2) The weighting vector w of criteria which was sho-
By this criterion, the evaluation of the significance wn in the comparison matrix A was determined by
of the facility for the execution of military actions of using the expression (8)
the enemy is done. By demolition of some facilities in
a , w 1,
M
j 1 ij
K
0,1 , i 1, 2,..., M (8)
the zone of combat operations no or negligible effects wi i wi
a
M K

are achieved, while some others may have even crucial k 1 l 1 kl

i 1

effect to the planned course of combat operations of Weights of matrix A entries are the sum of ele-
the opponent. This criterion considers two sub-criteria: ments in the related row of matrix, divided by the sum
evaluation of the importance of the facility for carrying of all the elements of matrix A.
out combat operations by opponents forces (C21) and wC1 0.177 , 0.177, (2 ) 0.177
evaluation of the impact of the demolition on the
continuation of combat operations of the enemy forces wC2 0.434 , 0.434, (2 ) 0.434
wCi
(C22). w 0.130 , 0.130, (2 ) 0.130
C3
wC 0.259 , 0.259, (2 ) 0.259
Characteristics of the direction in which the facility is 4
located in relation to friendly forces (C3) In the following step, with the use of pair wise sub-
By this criterion, the evaluation of the significance criteria comparison, their weights were obtained.
of the facility for the execution of military actions of Weighting values of sub-criteria were obtained by
friendly units is done. In addition, warfare conseque- using the expression (8).
nces for our combat units are estimated, which could
wC11 0.683 , 0.683, (2 ) 0.683
occur in the event of demolishing certain facility. This wC1i
criterion considers three sub-criteria: coverage of dire- wC12 0.317 , 0.317, (2 ) 0.317
ction in which the object is located by friendly forces
wC21 0.259 , 0.259, (2 ) 0.259
(C31); assessment of fire protection possibilities of the wC2 i
destroyed barrier (C32) and negative consequences of wC22 0.741 , 0.741, (2 ) 0.741
demolition for the subsequent combat of friendly wC31 0.135 , 0.135, (2 ) 0.135
forces (C33).
wC3i wC32 0.481 , 0.481, (2 ) 0.481
The importance of the facility to the population w 0.384 , 0.384, (2 ) 0.384
(C4) C33
By this criterion, consequence to the population By multiplying the values of fuzzy weights of sub-
that uses the facility is evaluated, and which could criteria with weights of related criterion, the
occur in the event that the demolition is carried out. aggregation of weight criteria of the first level of the
Some of the facilities, in addition to the fact they sig- hierarchy is performed. Final weight of the criteria on
nificantly affect the performance of combat operations, the first level of the hierarchy is obtained by using the
also have great significance for the people, both local expression (9).
and wider community, and thus, their demolition wo-

kj
a (9)
p
w l 1 il w j , j 1, 2,..., M , p 1, 2,..., k j
uld have number of other negative effects. These effe- j
k j ail
kj
i 1 l 1

TECHNICS MANAGEMENT 64 (2014) 115

D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...

Since all indices of consistency are smaller than 3 4 5 2 3 3

1
2 1 4

0.10, it can be considered that comparison matrices are 3 1 2 3 2
1
1 5
1
2
1
3
1
3
consistent. After the successful formulation of criteria 1 1 1 1 1 1 1

4 2 2 3 2 6 2 4 1
and their weight values, all prerequisites for testing the 1 1 1 1 1 1 1 1 1

5 3 2 4 3 7 4 5 2
present model are obtained. Final criteria weights of 1
1 1
the first level are shown in the following section. AOi C11
2

2 3 4 2 4 1 2 3
1 1 1 1 1

3 1 2 3 2 5 2 3 2
wC wC


wC 1 wC1i
wC1 1 11 3 5 6 7 4 5 4 3 6

wC1
wC12 1 1 1
2 2 2 4 1 2 4 2 3
1
1 3 4 5 2 3 3 2 4
wC wC21 1
wC 2 wC2 i
1 1 1 1 1 1
wC2 2 4 3 1 2 3 2 6 3 4

wC2
wC22
wC3
wC31
wC3 wC3
wC 3 wC3i wC32
w racteristics of direction in which the facility is located
C3 wC33
In order to perform the ranking, it is necessary to
calculate the value of criterion functions for each of the
considered alternatives. For testing the described
model, illustrative data that describe ten facilities
available for selection were used,.
The said alternatives are not taken from the field
because that kind of work would initially require
setting up the situation with all the elements of
(operating) combat deployment, which is not the
subject of this study, and available situations, which
have already been studied, are graded as top secret.
Comparisons of alternatives in relation to the
observed criteria are presented in the following part of
the paper. The comparison of alternatives in to Time
necessary for the preparation of the facility for
demolition ( C11 ) is shown in matrix AO C . Compa-
i 11
rison of other alternatives in relation to the criteria was
performed in an identical manner.
Comparisons of alternatives from matrices AO C i i
are brought in connection with the weights of criteria
and sub-criteria by applying fuzzy arithmetic opera-
tions. By applying equation (10) the fuzzified values
of criteria functions for all ten alternatives were
obtained.
K
Fi x ij
j , i 1,2,..., N
w during studies and exercises in peacetime situations.
j 1
(10)
Criteria functions are presented in the form of
triangular fuzzy numbers. Left and right distribution of
confidence interval of the criterion functions will be
determined based on four levels of certainty
0.25, 0.50, 0.70 and 1.00 . values of criteria functions for all the observed values
Based on the obtained criteria functions, final
ranking of the alternatives is performed (Table 2).
5. DISCUSSION AND CONCLUSION
Calculation of weight values of the criteria stressed
expressly deviation of criteria C2 weight (Cha-
in relation to enemy forces) in relation to other
criteria. This relation is, to some extent, expected since
the essence of the Interdiction Group's activities is
expression of a certain impact on the enemy forces.
Therefore, it is expected that the criteria C1 (Cha-
racteristics of the facility), C3 (Characteristics of the
direction in which the facility is located in relation to
friendly forces) and C4 (The importance of the
facility to the population), have significantly lower
weight, which indicates that they are less important but
not negligible criteria.
Such a relation is conditioned by the reality of the
situation, i.e. striving to achieve the original goal of
forming the Interdiction Group contribution to the
successful execution of the defensive operations. To
some extent, weight of criteria C4 (The importance of
the facility to the population) is surprising, but also
confirms the importance of a scientific approach to
solving these problems. This criterion would probably
be largely ignored in ad hoc because it is a criterion
that looks far into the future, and contrary to the
objectives pursued, at the time of the decision on the
selection of facilities. This is particularly important
when the decision is made by a person with less
experience, which will often be a reality, because in
spite of the large number of exercises and simulations,
carried out in order to prepare for combat, the reality
has its differentia specifica, which is difficult to see
Testing of the developed model showed that fu-
zzified AHP method can be used to the described
problem. That is how the successful ranking of the al-
ternatives was done, which simplifies the final decision
regarding the selection of facilities, but also the nu-
mber and composition of the Interdiction Group. The
of the level of certainty indicate that the alternative
number seven is the best, but in the same manner, the

116 TECHNICS MANAGEMENT 64 (2014)

D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...

decision maker may also make a final decision that the Besides, ranking of alternatives indicates that facilities
Interdiction Group should prepare facilities number number one, five, four, and six are expressly unfa-
seven and number two, which could make him decide vorable, which is a fact that greatly facilitates the
to form two groups, because the facilities are in di- choice for the decision maker. Of course, the final
fferent directions, or in the case the facilities are in the decision will largely depend on the specific situation.
same direction, he could form one Interdiction Group.
Table 2. Final rank of alternatives
Level of certainty
Alternatives Final rank
=0.25 =0.50 =0.70 =1.00
A1 0.063 0.078 0.083 0.092 7
A2 0.102 0.126 0.134 0.142 2
A3 0.095 0.117 0.124 0.133 5
A4 0.045 0.055 0.058 0.067 9
A5 0.060 0.074 0.078 0.087 8
A6 0.034 0.042 0.045 0.053 10
A7 0.137 0.169 0.179 0.188 1
A8 0.083 0.102 0.108 0.117 6
A9 0.096 0.118 0.125 0.134 4
A10 0.097 0.119 0.126 0.135 3

In most cases, the decision-making process comes
down to the experience and knowledge of the decision
maker. However, despite his experience, the comple-
xity of the combat operations could largely obscure
the real picture when deciding and redirect and focus
the decision-maker on less important parameters. This
is particularly important when one takes into consi-
deration that this issue represents only part of the set of
decisions that decision maker should define in a com-
bat situation. Introduction of the model saves the time
needed for decision- making, reduces the stress of
decision-makers and provides opportunity for persons
with less experience to make less harmful and more
beneficial decision.
One of the problems that could occur is defining
certain criteria values for some of the alternatives, es-
pecially because it is an estimate, which might have
a subjective character. This problem is mostly present
in the evaluation of criterion C2 because, in real life
situation, the most difficult issue is to obtain infor-
mation on the schedule of enemy units and their plans
for the execution of combat operations. With other cri-
teria, it is possible to make much more reliable com-
parisons in pairs, because these are available data that
can be collected by reconnaissance (criterion C1 and
criterion C4), i.e., they are present with the decision
maker and the persons around him (criterion C3). This
is especially important because these three criteria
significantly influence the final decision (the sum of
criteria C1, C3 and C4 weights is 0,566). By testing the
new approach in fuzzification of Saaty's scale, the
uncertainties that appear when comparing in pairs are
eliminated to some extent and more realistic order al-
ternatives are obtained. The developed model takes
into account the probability (level of certainty) that
indicates the validity of the data obtained from the
experts participating in the decision-making.
By analyzing the obtained results, it can be co-
ncluded that selected facilities can be successfully eva-
luated and decision-making strategies formulated in
the course of selection of the facilities for demolition
by the application of fuzzy AHP method. The practical
value of the study lies in the fact that the experience of
several experts is collected and incorporated in the
defined model, thus avoiding the situation in which the
decision-maker is limited to their own knowledge. In
addition, the model enables the successful ranking of
alternatives, even if some data are insufficient, as will
usually be the case with the data as per the criterion C2.
REFERENCES
[1] Boani, D., Suknovi, M. and Sekulovi, D., Ana-
lytic hierarchy process as a support in decision ma-
king process to the Group for interdiction in defence
operation, Proceedings of the SYM-OP-IS, p. 465-
468, 2012.
[2] Boani, D., Pamuar, D., orovi, B., Mili, A. and
Lukovac, V., Application of fuzzy AHP method on
selection of actions direction of the Group for su-
pplementary obstacle placing, Proceedings of the
SYM-OP-IS, p. 556-559, 2011.
[3] Chen, S. M., A new method for tool steel materials
selection under fuzzy environment, Fuzzy Sets and
Systems, 92, p. 265274, 2007.

TECHNICS MANAGEMENT 64 (2014) 117

D. BOANI at al. MODIFICATION OF ANALYTIC HIERARCHY PROCESS (AHP)...

[4] upi M. and Suknovi M., Odluivanje, Faculty of
Organizational Sciences, Belgrade, 2008.
[5] Doktrina Vojske Srbije, Odbrana Media Center, Bel-
grade, 2010.
[6] orovi, B., Istraivanje projektovanja organizacio-
ne strukture upravnih organa saobraajne sluzbe,
PhD thesis, Military Academy, Belgrade, 2003.
[7] Kujai, M., Izbor najbolje varijante u projektovanju
organizacije potanskog saobraaja korienjem
analitikog mrenog procesa, PhD thesis, Faculty of
Transport and Traffic Engineering, Belgrade, 2001.
[8] Lootsma, F. A., Mensch, T. C. A. and Vos, F. A.,
Multicriteria analysis and budget reallocation in lo-
ngterm research planning, European Journal of
Operational Research 47, p. 293305, 1990.
[9] Lootsma, F. A., Numerical scaling of human judg-
ment in pairwisecomparison methods for fuzzy mu-
lticriteria decision analysis, Mathematical Models
for Decision Support. NATO ASI Series F, Computer
and System Sciences, SpringerVerlag, Berlin, Ger-
many, p. 5788, 1988.
[10]Lootsma, F. A., T French and the American school in
multicriteria decision analysis, Operations Rese-
arch, 24, 3, p. 263285, 1990.
[11]Ma, D., Zheng, X., 9/99/1 scale method of the AHP,
Proceedings of the 2nd International Symposium on
the AHP, Pittsburgh, p. 197202, 1991.
[12]Miller, G. A., The magical number seven plus or mi-
nus two: some limits on our capacity for processing
information, Psychological Review, 63, p. 8197,
1956.
[13]Ray, T. and Triantaphyllou, E., Procedures for the
evaluation of conflicts in rankings of alternatives,
Comput. Ind. Eng., 1, p. 35-44, 1999.
[14]Saaty, T. L., The analytic hierarchy process, Mc-
Graw-Hill, New York, 1980.
[15]Srevi, B. and Zoranovi, T., AHP in group decision
making with complete and incomplete information,
Proceedings of the SYM-OP-IS 2003, p. 727-730.
[16]Triantaphyllou, E. and Lin, C. T., Development and
evaluation of five multiattribute decision making me-
thods, International Journal of Approximate Reaso-
ning, 14, p. 281-310, 1996.
[17]Triantaphyllou, E., Shu, B., Snachez, S. N. and Ray,
T., Multicriteria decision making: An operations
research approach, Encyclopedia of Electrical and
Electronics Engineering, JohnNjilley & Sons, New
York, 1998.
[18]Vojni leksikon, Military paper office, Belgrade,
1981.
[19]Zhu, K. J, Jing, Y. and Chang, D. Y., A discussion on
extent analysis method and applications on fuzzy
AHP, European Journal of Operational Research,
116, p. 450456, 1999.
[20]Zoranovi T. and Srevi B., Example AHP appli-
cation in group decision making in agriculture, Pro-
ceedings of the SYM-OP-IS, p. 723-26, 2003.

REZIME
MODIFIKACIJA METODE ANALITIKOG HIJERARHIJSKOG PROCESA I NJENA
PRIMENA U DONOENJU ODLUKA U SISTEMU ODBRANE

U radu je prikazan novi pristup modifikacije metode Analitikog hijerarhijskog procesa primenom fuzzy
logike. Osnovna karakteristika pristupa koji je korien u ovom radu je odreivanje intervala fuzzy
brojeva Saaty-jeve skale na osnovu stepena uverenosti donosioca odluke. Ovakav pristup dozvoljava da
donosilac odluke sa odreenim stepenom uverenosti (koji je najee manji od 100%) definie koji
lingvistiki izraz odgovara poreenju kriterijuma optimalnosti. Na ovaj nain, nakon primene metode
Analitikog hijerarhijskog procesa, dobijene su vrednosti kriterijumskih funkcija za svaku od
posmatranih alternativa, kojima odgovara odreena vrednost stepena uverenosti. Tako je obezbeeno
da se za vie vrednosti stepena uverenosti moe izvriti generisanje razliitih skupova kriterijumskih
funkcija. Predstavljeni model testiran je na taktikom nivou donoenja odluka u sistemu odbrane na
izboru objekata za ruenje Grupe za zapreavanje, kao postupku koji je najee propraen veim ili
manjim stepenom neodreenosti kriterijuma koji su neophodni za donoenje relevantne odluke.
Kljune rei: analitiki hijerarhijski proces (AHP), fuzzy logika, odluivanje, sistem odbrane, grupa za
zapreavanje

118 TECHNICS MANAGEMENT 64 (2014)

 ISSN 2334-7368

TECHNICS

QUALITY- IMS,
STANDARDIZATION
AND METROLOGY
Kvalitet IMS, standardizacija i metrologija, Quality
IMS, Standardizaction and Metrology

YEAR 14 - 2014.

EDITOR-IN-CHEF

Zoran P e n d i , M.Sc, United Association of Serbia for

Quality, Belgrade

EDITORIAL COUNCIL:

Prof. Vidosav M a j s t o r o v i , D.Sc, University of

Belgrade, Faculty of Mechanical Engineering,
Belgrade
Prof. Dragutin S t a n i v u k o v i , D.Sc, University of
Novi Sad, Faculty of Technical Sciences, Novi Sad
Prof. Sran S t a n k o v i , D.Sc, University of Belgrade
School of Electrical Engineering, Belgrade
Docent Valentina M a r i n k o v i , D.Sc, Zdravlje,
Leskovac
Milovan L u k o v i , Sloboda, aak
Prof. Dejan K o s t i , D.Sc, EPFL, Lausanne, Switzerland
Prof. Herbert O s a n n a, D.Sc, Viena University of
Technology, Austria
Prof. Ton van der W i e l e, D.Sc, Erasmus Research
Institute of Management, University Rotterdam,
Netherlands

EDITORIAL OFFICE: Union of Egineers and Technicians

of Serbia, 11000 Belgrade, Kneza Miloa 7a/I, Tel. +381/11/
32 35 891, Fax +381/11/ 32 30 067
R. JEVTI SIMULATION OF THE SHOPPING CENTRE ZONA I EVACUATION

Simulation of the Shopping Centre Zona I Evacuation

RADOJE B. JEVTI, Electrotechnical school Nikola Tesla, Ni Previous announcement

UDC: 355.244.2:519.876.5

One of the most important and the most complex tasks in human protection and human safety in objects
is the projecting of the object evacuation. There are many factors that could effect on the opportune
living of object such as object assignment, arrangement of rooms, arrangement of furniture,
arrangement of exits, occupant speed and many other that human lives and material properties depend
on. This is very important for objects with great number of humans, such as high residential objects,
shopping centers, schools, hospitals etc. This paper has written to show the possible evacuation
situations and calculate minimal time for evacuation in case of the shopping centre Zona Iin Ni.
Key words: evacuation, simulation, shopping centre

1. INTRODUCTION Very often, it is not possible to realize the eva-

The problem of the object evacuation presents one cuation until the accident is complete, although the
of the most important and most problematic tasks in evacuation demands momentarily reaction. An exa-
object projection. It is very difficult to predict every mple for that case is the earthquake. An example of
possible situation that leads to safe and secure living of evacuation in case of earthquake is presented on figure
the object. The evacuation term is very complex, but, 2.
generally, it presents the safest, shortest and fastest
way of moving for people, animals and material
properties from endanger object or location to the
secure place. The causes of evacuation could be
different: fire, gas, bomb threat, earthquake, flood, ci-
vil disorders etc. Every of these causes demand proper
strategy of object or location projecting that must be
realized in order to increase of human and material
properties safety. An example of the evacuation in case
of flood is presented on figure 1.

Figure 2 - An example of evacuation in case of ear-

Figure 1 - An example of evacuation in case of flood
Authors address: Radoje Jevti, Electrotechnical
school Nikola Tesla, Ni, Aleksandra Medvedeva 18
Paper received: 10.04.2014.
Paper accepted: 24.04.2014.
thquake
Very complex case in the evacuation problematic
is evacuation of object with the great number of
humans inside in case of fire. These objects could be
buildings, schools, markets, shopping centers, hospi-
tals, theaters and similar objects. The fire presents very
unpredictable accident and could cause huge damage
to human lives and material properties. This is the fact
that significantly can complicate projecting and design
of evacuation routes and, if they dont realize correctly,
directly harm safety of humans inside them.

TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014) 121

R. JEVTI SIMULATION OF THE SHOPPING CENTRE ZONA I EVACUATION
Evacuation routes were projected as primary and
secondary. Primary evacuation route is the most fre-
quently, route for normal communication in object. For
example, these routes could be stairs, hallways,
corridors and other surfaces used for communication
in object or in separate floor. They are with the
different dimensions for every type of object. These
routes are the routes that fire services used for, in case
of fire. The secondary routes depend on objects pu-
rpose. These routes could involve windows, roofs etc.
Both types of evacuation routes must satisfied many
standard and no standard demands, according to the
number of people, type and purpose of object, speed of
people moving, necessary time of evacuation etc. The
evacuation routes must be clearly visible at the
evacuation plan. An example of evacuation plan with
evacuation routes is presented on figure 3.
Figure 3 - An example of evacuation plan with the
evacuation routes
One particularly interested aspect according to
object or location evacuation is panic and stress affects
to the occupants. It is known that calm occupants have
bigger chances to leave the determinate object or
location than occupants influenced by stress and panic.
In panic and stress situations, occupants move faster
but chaotic. That implies many unexpected situations
in object or at location. Evacuation routes in object
have its own maximal occupant flow. That is
especially important for stairs, elevators, hallways,
corridors etc. Very often case is, that in the state of
panic and stress, occupants shrilly try to get out from
the object or try to leave the location and cause tragedy.
There were lots of similar cases with tragic epilogue.
The occupants behavior, knowledge and education,
physical and psychic condition (occupant speed, panic
and stress influence, moral and human qualities)
should be studied in details. For example, the
occupants speeds in this paper were from 1,5 m/s to
4,5 m/s in steps of 1 m/s, which show situations with
calm occupants and situations with occupants in panic.
It is also important to note that panic state of occupants
122 TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)
pretty much determinate than the state that implies fast
moved occupants. An example of panic and stress
situation in Moscow metro, caused by fire, is presented
on figure 4.
Figure 4 - An example of panic and stress situation in
Moscow metro
The successful object evacuation and saving of
human lives and material properties can be signifi-
cantly improved with the usage of simulation progra-
ms, such as Pathfinder. Pathfinder is an agent based
egress and human movement simulator. This program
provides a graphical user interface for simulation de-
sign and execution as well as 2D and 3D visualization
tools for results analysis. This and similar programs
present inevitable and necessary tool in evacuation
predicting, planning and projecting [1, 2].
2. SIMULATION MODEL
As always, the first step in the solving of the pro-
blem of this kind is to create the simulation model of
the desire object, in this case, the shopping centre Zona
I. Simulation model was created according to its real
model, with its complete dimensions. Although this
simulation model was created in Pathfinder, it was
possible to create this model in some other program,
such as Auto Cad or PyroSim and inserted it into
Pathfinder program.
The shopping centre Zona I in Ni is an object
with approximately 56 m X 20.5 m. It is located in
Bulevar Nemanjia street, number 67. It was built with
idea to provide all types of stuff so that there were no
needs for living in the downtown. This shopping centre
is one of three shopping centers that were built in the
Bulevar Nemanjia street. They are located only few
hundred meters between each other. There are lot
rooms with different purpose in the object: shops,
markets, bank, pharmacies, cafes etc. It has basement
and two floors with four exits/inputs, two from the
front and two from the back side of the object. The
width of stairs was 1,28 m. The width of the exit/input
doors was 1,82 m. There was a lift in the object, but
R. JEVTI SIMULATION OF THE SHOPPING CENTRE ZONA I EVACUATION
this lift was cargo lift and it was irrelevant according
to the evacuation. The pictures of the shopping centre
Zona I in Ni and its inside detail from the first floor
are presented on figures 5 and 6 [3].
Figure 5 - Shopping centre Zona I in Ni
Figure 6 - Shopping centre Zona I in Ni, inside
detail from the first floor
Figure 7 - Shopping centre Zona I in Ni Pa-
thfinder presentation
According to exposed facts and dimensions, the
simulation model of shopping centre Zona I in Ni
in Pathfinder was realized. The presentation of the
shopping centre Zona I in Ni simulation model, in
Pathfinder, is presented on figure 7. The Pathfinder
presentation of the second floor (above view) of the
shopping centre Zona I in Ni is presented in figure
8.
TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)
Figure 8 - Shopping centre Zona I in Ni, secound
floor - above view Pathfinder presentation
The simulation of evacuation was realized for di-
fferent occupants speeds, with different occupants
number per floor. It was decided that the occupant
number per floor is, in the first case, 80 occupants per
floor; in the second case, 100 occupants per floor and
in the third case, 150 occupants per floor. The scenario
implied that all of four exits/inputs were available. The
occupants speeds were from 1,5 m/s to 4,5 m/s in steps
of 1 m/s. The main reason for this assortment of speeds
was in wish to simulate the situation of calm
evacuation, without stress and panic, and to simulate
situation with occupants in panic, where the stuck and
chaos would be presented. The situations where the
stuck is caused by panic and stresses (stairs for
example) are very often and real [4, 8].
3. SIMULATION AND SIMULATON RESULTS
The simulation results mean minimal time need
that all of the occupants inside the object live the
object, for all three different number of occupants per
floor. At the first case, the complete number of oc-
cupants was 240, in the second case, the complete nu-
mber of occupants was 300 and n the third case, the
complete number of occupants was 450. An examples
of the simulation for the case with 100 occupants per
floor, randomly positioned, at 15,4 seconds and 59,4
seconds are presented on figures 9 and 10. The simu-
lation results for 80 occupants per floor, 100 occupants
per floor and 150 occupants per floor are presented on
figures 11, 12 and 13. All of the occupants were
randomly positioned.
Figure 9 - An example of the simulation for the case
with 100 occupants per floor at 15,4 seconds
from the start of the simulation
123
R. JEVTI SIMULATION OF THE SHOPPING CENTRE ZONA I EVACUATION

4. ANALYSE OF THE RESULTS

Simulation results that were presented on figures
10, 11 and 12 showed the time need for the shopping
center evacuation. For the first and the second sce-
nario, where the 80 and 100 occupants per floor were
positioned, no matter on occupants speed, the minimal
times need for object evacuations were from 50 to 100
seconds. But, for the third scenario, where the 150
occupants per floor were positioned, the minimal times
for occupants speeds from 2,5 m/s to 4,5 m/s were from
1560 to almost 1900 seconds, which was pointed to the
Figure 10 - An example of the simulation for the case stuck situation. As it was shown, it takes a lot of time
with 100 occupants per floor at 59,4 seconds to resolve the stuck situation. It is very important to
from the start of the simulation note that these situations could be very unpredictable
according to time need for resolve and consequences
Number of occupants occupant speed 1,5 m,s
occupant speed 2,5 m/s
that could arise. But, according to the data in the last
occupant speed 3,5 m/s years, the shopping center Zona I, for the most of the
200 occupant speed 4,5 m/s
time, was half empty, which was considered mostly the
people which worked in its shops, which is appropriate
150
to the first or the second scenario. Of course, it is
100
important to know and to predict the possible situations
with more occupants in the object.
50
According to these results, it is possible to locate
0
potential evacuation routes which could be very useful
0 10 20 30 40 50 60 70 80
Time [s]
in some situation of evacuation [9].
Figure 11 - Simulation results for 80 occupants per 5. CONCLUSION
floor
This way of analyzing of evacuation times and
Number of occupants occupant speed 1,5 m/s evacuation routes is very appropriate because it gives
300
occupant speed 2,5 m/s
occupant speed 3,5 m/s
good presentation how available evacuation routes
250
occupant speed 4,5 m/s could be used for different accidents (fire, earthquake
etc) that could happen. It is also possible to locate new
200
evacuation routes that could be used in accident (lower
150 floors exits, lower windows etc). Testing these factors
100
for different occupants speeds and behaviors gives
good real presentation of potential evacuation scenario
50
in object and great advantages in projecting and
0 Time [s] installing of complete protection systems, such as, for
0 10 20 30 40 50 60 70 80 90
example, fire protection system. There are many new
Figure 12 - Simulation results for 100 occupants per approaches and algorithms about evacuation and
floor behavior under some accident that are constantly
improving, witch cognition significantly increase the
Number of occupants
450
occupant speed 1,5 m/s
occupant speed 2,5 m/s
evacuation safety level [10-14].
400 occupant speed 3,5 m/s
occupant speed 4,5 m/s REFERENCES
350

300
[1] Blagojevi, M.: Alarm systems, Faculty of occu-
250
pational safety, Ni, 2011.
200

150 [2] Jevti, B. R, Blagojevi, . M.: Simulation of the

100 school object evacuation, Tehnika, Vol 2, pp. 365-
50 370, ISSN 0040-2176, Beograd, Serbia, 2013.
0 Time [s]
0 200 400 600 800 1000 1200 1400 1600 1800 2000
[3] Thunderhead, Pathfinder Exmple Guide, 2012
Figure 13 - Simulation results for 150 occupants per
floor

124 TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)

R. JEVTI SIMULATION OF THE SHOPPING CENTRE ZONA I EVACUATION

[4] Helbing, D., Farkas, I., Viscsek T.: Simulating dina- [9] Jevti, B. R.: The importance of fire simulation in
mmical features of escape panic, Nature, Vol. 40, pp. fire prediction, Tehnika, Vol 1., pp. 153-158, ISSN
487-490, 2000. 0040-2176, Beograd, Serbia, 2014.

[5] Tabares, R. M.: Evacuation Process Versus Evacu- [10] www.iafss.org/publications/fss

ation Models: Quo Vadimus?, Fire Technology 45,
pp. 419-430, 2009. [11] www.nfpa.org/...buildings/faqs-about-building-e.

[6] Hasofere, A. M., Beck, V. R., Bennetts I. D.: Risk
Analysis in Building Fire Safety Engineering, Els-
evier, 2007.
[7] Thorton, C., OKonski, R. and Hardeman, B.: Intro-
ducing: An Agent-Based Egress Simulator, Proce-
edings of the Fourth International Symposium on
Human Behaviour in Fire, pp. 567-572. UK, 2009.
[8] Jevti, B. R, Nikovi, T.J.: Simulation of residential
object evacuation, 57th ETRAN CONFERENCE,
Zlatibor, ISNB 978-86-80509-68-6, Serbia, 2013.
[12] www.stonybrook.edu/asaforms/.../EHSD0091.pd
[13] Kuligowski, E. D.: Modeling Human Behavior Du-
ring Building Fires, NIST National Institute of Sta-
ndards and Technology, NIST Technical Note 1619,
USA, 2008.
[14] Galea, R. E.: The Developement and Validation of a
Rail Car Evacuation Model, Proceedings of the 13th
International Fire Science & Engineering Confere-
nce INTERFLAM 2013, Interscience, pp.1013-1034,
2013.

REZIME
SIMULACIJA EVAKUACIJE TRNOG CENTRA ZONA I

Jedan od najvanijih i najkompleksnijih zadataka vezanih za zatitu i bezbednost ljudi u objektima je

projektovanje evakuacije objekta. Postoji mnogo razliitih faktora koji mogu uticati na pravovremeno
napustanje objekta kao to su namena objekta, raspored prostorija, raspored nametaja, raspored
izlaza, brzina ljudi koji naputaju objekat i mnogi drugi koji mogu uticati na ljudske ivote i materijalna
dobra. Ovo je posebno vano za objekte sa velikim brojem ljudi u njima, kao to su stambeni
objekti,trgovaki centri, kole, bolnice itd. Ovaj rad je napisan da pokae mogue situacije prilikom
evakuacije i da izrauna minimalno vreme koje je potrebno za evakuaciju u sluaju trnog centra Zona
I u Niu.
Kljune rei: evakuacija, simulacija, trni centar

TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014) 125

G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...

Development of Methodology for Evaluation and Real Water Loss Reduction

in Water Supply Systems in Transition Countries

GORAN S. ORAANIN, Municipal Water Utility, East Sarajevo Professional paper

DRAGANA S. RISTI, REIC, Sarajevo UDC: 628.179
DALIBOR Z. VLAKI, Municipal Water Utility, East Sarajevo
BRANKO S. VUIJAK, University of Sarajevo, Mechanical Engineering Faculty, Sarajevo

Key strategy for water supply companies, aiming increasing their capacity and expanding the network,
should be to focus on reducing water losses and non revenue water in general. Water loss reduction,
after conveniently achieving initial progress, requires continuous efforts to maintain or improve the
achieved level of water losses. This paper proposes a methodology developed on the basis of collected
and analyzed knowledge and experience gained in water utilities in BiH and neighbouring countries.
The methodology is also based on the experiences presented in the literature relating to water losses in
developed countries, which are applicable or adaptable, so as affordable for the countries in transition.
Key words: methodology, real losses, district metered area, countries in transition

1. INTRODUCTION cially viable for transition countries. Water supply

Water supply companies should be efficient (ope-
rating with lowest possible effort and in optimal con-
ditions), as well as effective (to fully provide desired
services). Creation of strategy and action programme
in order to reduce water loss should become inevitable
integral part of management policy of a water supply
company. It is significant to determine tangible and
achievable goals for water loss reduction along with
the implementation of a suitable methodology. Apart
from conveniently applied initial efforts in this
direction, water loss reduction requires further
continuous efforts in order to maintain or improve
reached water loss level. It is of the utmost importance
for a water supply company to build and develop its
own capacities in charge of water loss issues [9].
Methodology for real water loss reduction is pre-
sented with this paper. The methodology has been
made in form of a summary of collected and analyzed
knowledge and experiences of water supply compa-
nies in B&H and surrounding countries. Additionally,
the methodology is based on experiences presented in
literature referring to water losses in developed cou-
ntries, which are applicable or adaptable, so as finan-
Authors address: Goran Oraanin, Municipal Water
Utility East Sarajevo, East Sarajevo, Nikole Tesle 53a
Paper received: 20.12.2013.
Paper accepted: 30.12.2013.
systems usually have specific methods for water loss
reduction. Those experiences and knowledge are con-
solidated into one operative methodology. Specific
suggestion of the methodology for real water loss
reduction for water supply systems in transition co-
untries follows.
2. METHODOLOGY
Generally, three types of measures for water loss
reduction are applicable:
Short-term measures are measures for leak dete-
ction at the beginning of water loss reduction process.
Those measures can be easily conducted and they give
good results. Tail, unfixed leaking (reported and
detected) that have been accumulated for years can be
relatively quickly reduced. Conducting of those
measures brings water supply company into the po-
sition that there are no visible leaks. Maintaining of
such position requires additional efforts.
Mid-term measures represent organizing of the
most appropriate District Metered Areas (DMA) and
introducing pressure control measures. Measuring and
controls in measuring area maintain attained improved
position regarding water loss. Conducting of active
leakage control (regularly measuring flow and
pressure, evaluation of the water balance, detection
and registration of leaks/failures and similar) reflects
and enhances the improved position. Pressure control

126 TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)

G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...
in DMAs may reduce water losses, but it cant fully
eliminate them.
Long-term measures present plans encompassing
replacement of pipes and connections where leaking is
frequently recorded.
Maintaining the database on leaks and failures
within the network allows their occurrence analysis, as
well as recognizing sections or connections (where
large losses occur) within the water supply system
characterized with frequent leakages.
Water loss reduction, along with fulfilling all th-
ree aforementioned measures, can be achieved by im-
plementing the following steps (Figure 1):
Step 1 Organization of the water losses control
department
In order to achieve efficient management of water
supply systems, especially regarding water loss aspect,
it is inevitable and crucial to reorganize water supply
companies. Determination of employees' number and
structure, dealing with water loss issues in water
supply systems, represents an important precondition
for water loss reduction in water supply companies.
This means that firstly, water supply companies
should have a special unit for water loss control,
needed for continuous following up and maintaining
of water losses at acceptable level [9].
In order to organize Water loss control Service it
is necessary to provide technical and material preco-
nditions for successful business running (vehicle, me-
asurement devices, sound leak detection equipment,
personal computers and other IT, and other equi-
pment) as well as adequate working space. Water loss
control equipment1 necessary for any water supply
company depends on the network condition and qua-
ntity of losses in the system, as well as on financial and
human resources of the company.
Step 2 Water balance
Periodical water balance evaluation should be
conducted for the entire system, due to the fact that
employees should have a sense how much water loss
can be expected in metering areas. Water losses should
be presented as liter/connection/day. It should be
calculated in accordance to IWA recommendations
and based on ILI (Infrastructure Leakage Index) as an
1Here it is necessary to emphasize that equipment could be
purchased, via water associations, for more than one water
supply company, or that equipment could be periodically rented
from some other water supply company. In case of small water
supply companies, it is recommended that few water supply
companies, geographically near, purchase equipment and
organize one service that would cover working activities for
each and every included company.
TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)
excellence indicator pointing at changes in efficiency
category [2, 5, 11].
Quantity of Non Revenue Water (NRW) is being
determined in a way that water audit is being condu-
cted and results are shown in the water balance. Water
balance is based on measuring or quantitative esti-
mations of produced water, exported, imported, used
and lost water. International Water Association (IWA)
recommends tabular review of water balance which
each and every water supply company should have in
order to start activities regarding water loss reduction.
Water balance table provides indicators and quantity
of the water used and water lost in any way.
Evaluation of the water balance2 has a goal to fo-
llow up and identify each and every water balance
component in a specific time period. Based on calcu-
lation method, all consumption components as well as
losses are being identified in one standardized form in
this water balance. Clearly defined water balance
represents a first step in the process of evaluation of
non revenue water, water losses and leakage
management in water supply systems [4].
Regarding evaluation of the water balance, it is
important to have in mind the fact that accuracy of data
on water loss quantity depends on accuracy and data
quality used during the process of evaluation. Thus,
reliable measuring of all inputs, as well as outputs,
represents a basic demand. Reliable data on water pipe
length, number of connections and other represent
second demand. However, in transition countries, it is
relatively unusual for a water supply company to have
exact data on water quantity entering water supply
system. Thus, those evaluations of the water balance
are considered as approximate ones, due to the
difficulties to estimate all components accurately.
Irrespective of those difficulties in estimation of water
balance components, it is always necessary to
calculate water balance.
Step 3 Determinations of District Metered Areas
(DMA)
DMA implementation enables dividing of the
whole water supply system into smaller and limited
zones (sub-systems) that can be easily controlled. In-
ternational Water Association roughly recommends
DMA from 500 to 3000 connections. However, IWA
also states that hydraulic, practical and economic fa-
ctors set size of district metered area DMA [8], what
leads to the conclusion that within one water supply
2www.waterloss.com.ba is a web page made in 2012. by
Mr. Djevad Koldzo, expert for Non Revenue Water. Apart from
other things, two software tools for evaluation of water balance
as well as for failure record on water supply network can be
found here.
127
G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...
system a DMA would differ from the other one in
number of connections.
Procedure is as follows: firstly, district metering
areas in the area of water supply system are being
determined; first area is being selected and measuring
is being performed within it; then accurate water
balance is being calculated, failures are being identi-
fied and fixed. Upon finishing of all aforementioned,
the procedure is being performed for the next DMA.
Testing in metering area should be conducted in three
steps [3]: (i) first testing represents preliminary de-
termination of condition (evaluation of water balance)
and narrowing down of the area where leakage appears
(night measuring method); (ii) second testing
represents direct testing of the failure/leak place or
location (sound leak detection and leakage spot loca-
tion) and (iii) Third testing is being performed upon
repairing of defined leakage spots in order to check
effects of repair to the water balance. The procedure is
iterative until achieving sufficient results, that is, new
water balance is being compared to previous water
balance and improvements are being determined. If
Figure 1 - Schematic overview of the methodology
The following chapter demonstrates the results
achieved for DMA of East Sarajevo water supply sy-
stem, where previously mentioned procedure has been
undertaken. Established department for water losses
control has performed preliminary measures in order
to determine condition, that is, Service has determined
water balance for the clearly defined metering area.
Following activities have been performed: at first,
failures have been detected, closely located and then
repaired; pressure has been regulated; repeated
measuring has been performed in order to determine
condition and repair effects. Achieved results are
presented in this paper, and were followed by active
128 TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)
the results are not satisfactory, the procedure should be
repeated starting from night measuring method.
Testing procedure should be conducted at all
DMAs in water supply system.
Step 4 Active loss control and system maintena-
nce
Upon establishing of all DMAs in water supply
system, active loss control represents a very signifi-
cant part of process of maintaining of attained im-
provement level [1]. Based on field work, detecting of
invisible leakages and failure record improvement
measures for certain measuring areas are being defi-
ned. There can be a case that for a certain measuring
areas in a certain time period there is no need to
conduct additional measures (there are no frequent
failures, acceptable pressure level in DMA, that is,
DMA losses are at acceptable level). However, there
can also be a case that in some metering areas pressure
should be controlled [6, 7, 10] as well as reservoir
leaking or pipe-lines and connections should be
replaced.
loss control and system maintaining in order to
maintain attained level of improvement.
3. EXAMPLE OF EAST SARAJEVO WATER
SUPPLY SYSTEM DMA
District Metered Area (DMA) selected for testing
of this methodology, represents a closed area with one
water entrance to the system, that is, looking from the
hydraulic aspect, water supplement is closed-water
towards settlement goes in one direction and does not
leave that settlement. Measuring area is being supplied
by water via pump station, that is, water is directly
being pumped up into the network. This represents an
G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...
interested fact regarding measuring area analyzing,
especially regarding the aspect of real water losses.
Pump station is located at altitude 550,5 m. Pressure at
the entrance to the pump station varies in range from
Figure 2 - District meter area
Water supply network of this measuring area,
along with data received by performing of testing at
the terrain during the process of measuring area de-
termination, is given in Figure 2. During the process
of terrain research, following data have been collected:
population number-about 700; number of co-
nnections-188 (households-177 and legal subjects-
11): network lenght-3.063 m (PE 63, PE 90, PE 110,
and PE 160).
Based on network characteristics, critical points
and possibility to set flow meter, three measuring areas
have been determined. At one measuring area (at
pump station) pressure and flow have been measured,
while on other two spots only pressure has been
measured. Measures have been conducted into two
phases per 7 days. During the first measuring, average
water quantity that entered into entire water supply
system was 108,83 l/s. Based on indicators given by
first measuring, three failures have been located and
repaired, so the pressure at pump station was reduced
Figure 3 - Comparison chart of measurements of flow
TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)
2,8 to 3,5 bars. Pump station is set to pressure from 5,5
to 7,5 bars. Ultimate and critical points are at 574 m
altitude (measuring spot - MM 2) and 570 m altitude
(measuring spot - MM 3).
for 0,5 bars. Upon conducting of those measures
second measuring has been performed. During the
second measuring, average water quantity that entered
water supply system was 115,06 l/s.
3.1.Measurements analysis
During flow meter setting, reading of values on
water meters has been performed in district meter area.
Additionally, 7 days later, values on water meters have
been read in that district meter area. That very day
measuring data have been downloaded. Values
recorded at water metres for 186 water metres and two
evaluated lump households were 578 m3. Results
showed that average flow in 7 days time period of
measuring was 1,5 l/s (Table 1).
The same procedure has been conducted during
the second measuring. Figure 3 shows results of flow
measuring during the first and second measuring.
Results showed that average flow in 7 days time period
of measuring was 1,54 l/s (Table 1).
129
G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...

Table 1 represents a comparison of average flow
for days during first and second measuring.
Flow increase in the second measuring was caused
by significantly lower temperatures in comparison to
temperatures during the first measuring. Recorded
flow of 1,5 l/s during the first measuring is a result of
smaller failures in the system as well as result of higher
pressure than it was during the second measuring.
Flow of 1,54 l/s at temperatures which reached -20 C
during the night represented increase of consumption
due to the fear of population that water in household
installations could froze. This conclusion is supported
by the fact that minimal night flow in those days is not
equal to zero (as it was the case during the first
measuring) but it is equal or higher than 1 l/s.
However, during the day flow equal to zero appears
what points to the fact that there are no bigger failures
in the system.

Table 1. Comparison of average flow

1 2 3 4 5 6 7 Average
Measuring 1 1,52 1,34 1,49 1,46 1,48 1,65 1,53 1,50
Measuring 2 1,58 1,56 1,49 1,52 1,45 1,62 1,58 1,54

Figure 4 shows comparative results of pressure minimal entrance pressure was decreased from 5,5
measuring during the first and second measuring at the bars to 4,76, and maximal pressure was decreased
entrance to the system. Upon finishing first measuring, from 7,5 bars to 6,87 bars.

Figure 4 - Comparison chart of measurements of pressure at the measuring spot 1

Figure 5 shows comparative results of pressure measure spot 2 was reduced from 2,9 bars to 2,57 bars,
measuring at measure spot 2 (MM 2). By reducing of while maximal pressure was reduced from 5,36 bars to
pressure at entrance to the system, minimal pressure at 4,68 bars.

Figure 5 - Comparison chart of measurements of pressure at the measuring spot 2

130 TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)

G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...

Figure 6 shows comparative results of pressure measure spot 3 was reduced from 2,3 bars to 1,94 bars,
measuring at measure spot 3 (MM 3).By reducing of while maximal pressure was reduced from 5,08 bars to
pressure at entrance to the system, minimal pressure at 4,54 bars.

Figure 6 - Comparison chart of measurements of pressure at the measuring spot 3

Table 2 shows average pressure for all 3 measure of lower pressure at DMA entrance (pump station).
spots during both measuring. Total average pressure in the system is lower for 0,55
The table clearly shows that average pressures in bars, that is, it is reduced from 4,68 to 4,13 bars.
second measuring are lesser for 0,5 bars. This is result
Table 2. Comparison of average pressure
Measuring spot 1
1 2 3 4 5 6 7 Average
Measuring 1 6,30 6,08 6,23 6,23 6,18 6,24 6,29 6,22
Measuring 2 5,73 5,68 5,61 5,66 5,62 5,77 5,71 5,68
Measuring spot 2
1 2 3 4 5 6 7 Average
Measuring 1 4,14 3,93 4,08 4,08 4,00 4,08 4,14 4,06
Measuring 2 3,55 3,51 3,44 3,49 3,45 3,59 3,54 3,51
Measuring spot 3
1 2 3 4 5 6 7 Average
Measuring 1 3,84 3,62 3,78 3,78 3,70 3,78 3,84 3,76
Measuring 2 3,23 3,19 3,13 3,16 3,13 3,27 3,22 3,19

In line with IWA recommendations, water balance Losses (CARL), Unavoidable Annual Real Losses
in DMA is calculated after first and second measuring. (UARL) and ILI (Infrastructure leakage Index) indi-
Based on network length (3.603.m), average cators are being calculated.
pressure in system during the course of the first and Connection density per network kilometre is 52
second measuring (4,68 bar and 4,13 bar), number of connections, so the indicators are presented as
connections (188) and average distance of water me- liter/connection/day (Table 3).
ters in private ownership (5m), Current Annual Real
Table 3. Comparison of systems condition indicators
CARL UARL Improvement potential NRW
ILI
(liter/connection/day) (liter/connection/day) (liter/connection/day) (%)
Measuring 1 218,2 53,6 164,6 4,1 36,1
Measuring 2 188,22 47,3 104,92 4 31
Result - 13,7 % - 11,7 % - 36,3 % - 2,4 % - 5,1

TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014) 131

G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...
Table 3 shows comparison of system's condition
indicators between the two measuring. Non-Revenue
Water (NRW) percentage is lower for 5,1 than it was
during the second measuring. Infrastructure Leakage
Index (ILI) indicator is decreased for 0,1. Due to the
pressure decrease, Unavoidable Annual Real Losses
(UARL) are lower for 6,3 liter/connection/day. Repair
of failures, pressure decrease and increased con-
sumption caused Current Annual Real Losses (CARL)
to be lower for 29,98 liter/connection/day, while
improvement potential (difference between CARL
and UARL) is lower for 59,68 litre/connection/day. By
CARL water decrease, economic or financial benefits
are the following ones: i) 205,72 euro/annually if using
the marginal water production cost (estimated cost of
water production is 0,10 Euros/m3); ii) 1542,92
euro/annually if using the water supply tariff and
sewage tariff (estimated tariff for water supply and
sewage is 0,75 Euros/m3).
4. CONCLUSION
In the frame of proposed methodology, described
in this paper, step by step approach has been de-
veloped which should be based on measuring and
dividing of water supply system to smaller sub-sy-
stems (DMAs). Implementation of this methodology
to any water supply system can be defined based on a
procedure, which is based on sequence of actions
through the following steps:
Organization of department for water losses co-
ntrol (delegation of employees in charge of this
activity);
Water balance evaluation for water supply system
(rising of consciousness what can be expected to
occur in water supply system regarding water
losses);
Determination of District Metered Areas and me-
asuring in those areas (determination of actual
condition in sub-systems);
Repair of failures, improvement measures and
measuring (determination of improvement mea-
sures success); and
Active loss control and infrastructure maintaining
in district meter areas and entire system
(maintaining of attained level of decreased real
water losses and improvement activities).
Results demonstrated in this paper can be also
achieved in other district metered areas, at the same
time thus proving that applied procedure for water loss
132 TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014)
decrease in one district meter area can be applied to
the entire water supply system.
REFERENCES
[1] Farley M., Liemberger R., Developing a non-
revenue water reduction strategy: planning and
implementing the strategy, Water Science and Tec-
hnology: Water Supply Vol. 5 No. 1 pp. 41-50, IWA
Publishing, 2005.
[2] Hamilton S., Mckenzie R., Seago C., A Review of
Performance Indicators for Real Losses from Water
Supply Systems, UK house of commons report, Juli,
2006.
[3] Koldo ., Prirunik za efikasno mjerenje i
otkrivanje gubitaka u vodovodnim sistemima, Insti-
tut za hidrotehniku Graevinskog fakulteta u Sara-
jevu, Sarajevo, 2004.
[4] Lambert A., Hirner W., IWA BLUE PAGES, Losses
from Water Supply Systems: Standard Terminology
and Recommended Performance Measures, London,
2000.
[5] Lambert A., Mckenzie R., Practical Experience in
using the Infrastructure Leakage Index. - IWA
Conference Leakage Management A Practical
Approach, Conference Proceedings, Cyprus, 2002.
[6] McKenzie R., Wegelin W. Implementation of
Pressure Management in Municipal Water Supply
Systems. IWA pres paper 0309, 2009.
[7] McKenzie R., Wegelin W. Pressure Management in
South Africa. Proceedings of the WISA Conference,
Durban, South Africa, 2010.
[8] Morrison J., Tooms S., Rogers D., District Metered
Areas, Guidance Notes, International Water
Association, 2007.
[9] Oraanin G., Vlaki D., Vuijak B., Improving the
water utilities' organizational structure in transition
countres as a precondition to reduce water losses,
Tehnika Vol. 67 No. 6, pp. 1037-1042, 2012.
[10] Thornton, J. Managing leakage by managing
pressure: a practical approach. Water 21, No. 43-44,
2003.
[11] Winarni W., Infrastructure Leakage Index (ILI) as
Water Losses Indicator, Civil Engineering
Dimension, Vol. 11, No 2, pp 126 134, 2009.
G. ORAANIN at al. DEVELOPMENT OF METHODOLOGY FOR EVALUATION AND REAL...

REZIME

RAZVOJ METODOLOGIJE ZA EVALUACIJU I SMANJENJE STVARNIH GUBITAKA

VODE ZA VODOVODNE SISTEME ZEMALJA U TRANZICIJI

Vodovodna preduzea kao kljunu strategiju za poveanje kapaciteta i irenja mree trebaju odabrati
usmjerenje ka smanjenju gubitaka vode i uopte na smanjenju neprihodovne vode. Smanjenje gubitaka
vode nakon poetnog napretka zahtijeva i dalje napore na odravanju dostignutog nivoa gubitaka vode.
U radu je predloena metodologija koja je nastala kao rezime prikupljenih i analiziranih znanja i
iskustava iz vodovodnih preduzea u BiH i zemalja u okruenju. Metodologija se, takoe, bazira na
iskustvima prikazanim u literaturi koja se odnose na gubitke vode u razvijenim zemljama, a koja je
primjenljiva ili prilagodljiva, kao i finansijski prihvatljiva za zemlje u tranziciji.
Kljune rijei: metodologija, stvarni gubici vode, ograniena mjerna oblast, zemlje u tranziciji

TECHNICS QUALITY IMS, STANDARDIZATION AND METROLOGY (2014) 133

 CO-PUBLISHERS
SERBIAN CHAMBER OF
ENGINEERS, BELGRADE
FACULTY OF CIVIL ENGINEERING,
BELGRADE
FACULTY OF TECHNICAL SCIENCES,
NOVI SAD
FACULTY OF TRANSPORT AND
TRAFFIC ENGINEERING, BELGRADE
SCHOOL OF ELECTRICAL
ENGINEERING, BELGRADE
FACULTY OF MECHANICAL
ENGINEERING, BELGRADE
FACULTY OF MINING AND GEOLOGY,
BELGRADE
FACULTY OF TECHNOLOGY AND
METALLURGY, BELGRADE
TECHNICAL FACULTY, BOR
TECHNICAL FACULTY, AAK
IPIN INSTITUTE FOR APPLIED
GEOLOGY AND HYDRO
ENGINEERING, BIJELJINA
SPONSORS
PUBLIC ENTERPRISE FOR ELECTRIC
ENERGY TRANSMITION SYSTEM
CONTROL ELEKTROMREZA SRBIJE,
BELGRADE
PUBLIC ENTERPRISE ELECTRIC
POWER INDUSTRY OF SERBIA,
BELGRADE
TERMAL POWER PLANTS NIKOLA
TESLA, OBRENOVAC
MAINOPROJEKT - COPRING,
BELGRADE
PE HIDRO POWER PLANTS DRINSKO
LIMSKE, BAJINA BATA
PUBLIC UTILITY COMPANY
BELGRADE WATERWORKS AND
SEWERAGE, BELGRADE
TRAFFIC INSTITUTE
"CIP", BELGRADE
INSTITUTE "MIHAJLO PUPIN",
BELGRADE
INSTITUTE "VINA" BELGRADE
BUSSINESS COMPANY THERMAL
POWER PLANTS AND MINES,
KOSTOLAC, KOSTOLAC
PE KOLUBARA MINING BASSIN
LAZAREVAC
PE JUGOISTOK, NI
BUSSINESS COMPANY CENTAR,
KRAGUJEVAC