You are on page 1of 4

l.

I)

Pergamon

Computers and Chemical Engineering Supplement (1999) S613-5616


C 1999 Elsevier Sci ence Ltd. All rights reserved
PH: 50098-1354199100147-7

Application of Bayesian Networks to Troubleshooting


Distillation Columns
D. Sampson and X.Z. Wang*
Department of Chemical Engineering, The University of Leeds, Leeds LS2 9JT, UK.
Abstract - Sophisticated computer simulation has revolutionised the design and control of distillation processes.
Despite this, statistics has shown that the number of column malfunctions is rising. It is believed that there is a good
amount of art in preventing and trouble-shooting column operational problems. Unlike principal knowledge which can
be described in equations and implemented in computer programs, the knowledge for trouble-shooting is often illstructured. This paper describes the application of Bayesian graphical models for representation of this type of
knowledge and for developing trouble-shooting systems.
Kiste~.s. Harrison and Prance" classify all distillation
Introduction
column malfunction symptoms into two classes, i.e.
Distillation is still the most popular separation
product off-specification and the expected capacity not
technique. World-wide, it represents 95% of all
delivered. The causes .are . categorised as external
separation processes I. As far as energy usage is
factors, e.g. up-stream operations, auxiliary equipment
concerned, it takes up some 15% of all industrial energy
including reboilers and condensers; as well as
consumption in the US 1 The use of high-speed
distillation column internals. The column internal
computers has revolutionised the design, control and
problems are further grouped into flooding, excessive
analysis of distillation processes. Despite this, a recent
pressure drop and low mass transfer efficiency. Kiste~.s
survey has shown that column malfunctions are not
found that similar malfunctions occur in different
becoming extinct', Their number is rising rapidly,
companies and repeat in the history, and therefore
showing no signs of decline. As an example, Kock
believed that a good strategy to prevent operational
problems is to learn from past experience. Towards this
Engineering has also noted an increase in requests for
solving operational problems). It is believed that there is
purpose he compiled 300 malfunction cases in 19905
a good amount of "art" involved in preventing and
and another 300 in 19972
A distinguished feature of the knowledge used for
solving column problems. Art is defined as personal,
trouble-shooting is that it has always been regarded as
unanalysable creative power. It is this power plus a
an art. The success of principal knowledge based on
good deal of experience that makes the difference
between success and failure in the design and operation
mass transfer and reaction kinetics can be attributed to
the fact that it is described in equations and therefore
of distillation equipment.
can be easily coded into computer programs, such as
Knowledge based expert systems (ESs) have been used
to capture the creative power and experience of human
Aspen, Provision and Hysis. In contrast, knowledge
experts. In ESs the causal relationship between the
used for trouble-shooting has always been compiled as
descriptive materials. While it is easy to compile, there
condition and the fact is very clear in such a rule as "If
are difficulties in computer implementation. Graphical
condition with certainty x then fact with certainty f(x)".
model based representation is able to transform such
In addition, the how explanation capability of an ES can
complex and ill-structured knowledge into an easy
give a causal picture on how a conclusion is reached.
understood form, is therefore considered as a potentially
However, ESs are still not satisfactory since the overall
powerful technology
causal relationships of all process variables are buried
in the rule base and are still far more complex to
Conceptualisation ofnodes using structural as well as
decision makers. A more powerful method of
functional and behavioural decomposition
representing the knowledge for trouble-shooting is using
A graphical model describes the causal relationships of
graphical models. In this paper we describe the
nodes. A node represents a concept or an event.
application of Bayesian networks to trouble-shooting
Properly defining the nodes in a graphical model is very
distillation columns.
important to the representation of trouble-shooting
knowledge. Jefferson et a\.6 and Cox et al.' have made
Knowledge Review and Representation
efforts to define key words for equipment fault
Trouble-shooting distillation columns
There have been a number of publications on troubleclassification. In a graphical model, it requires more
detailed classification, therefore a method for arranging
shooting distillation columns. The most comprehensive
the nodes is probably more important. The nodes for the
discussions are given by Harrison and France" and
Corresponding author. Email: x.z.wang@leeds.ac.uk

5614

Computers and Chemical Engineering Supplem ent (1999) S6J3-S616

graphical model construction are selected according to


structural as well as functional and behavioural
decomposition.
Structurally, process plants are composed of unit
operational equipment linked by pipelines and control
systems. The structure can be modelled using objectoriented paradigm. A piece of equipment is modelled
as an object. A group of similar objects form a class.
The relationships of equipment classes form the data
structure, which fall within one of the three categories,
(1) generalisation, i.e. "is-a" relationship, such as
Fig.l(a), (2) aggregation, i.e, "part-of" relationship
(Fig.ltbj), and (3) aggregation relationship, i.e. two
objects are independent(Fig.Ifcj).

shooting misbehaviour are called misbehaviour


networks,
Bm=Jnt0!1m, Y2m, ...)
The network models can be obtained by combining
process principal models with historical databases. An
exampleis given below.
Downcomerflooding =
J{[Geometryparameters], [Flow parameters],
[Systemproperties))
= f{[h ,..., HI> Lw, ha], [L, Uo], [<p, pv, pd}
Where, hw - exit weir height, HI - tray spacing, Lw - exit
weir length, h, -under downcomer clearance, L - liquid
flowrate, U, - vapour flow velocity through holes, <p foaming factor, Pv, PL - vapour and liquid densities.
The above analysis plus historical data help form the
networkin Fig. 2.
Causesof downcomer flooding
Excessive liquidflow...
Blockage of liquidflowareaunder
DC

--------1

Vapour leakageinto DC
Improper feed introduction ...
Bottom seal pan...
Foaming
Others...

Debrits (e.g. toolsand


loose tray parts)
Dislocatedinternals
Sidedraw is stopped
or restricted
Others

Fig. 2 The graphical model for downcomer flooding

(b) Au rfg atiol1relationship

ee) Association. relationship

Fig. I Structuraldecomposition of knowledge using


object-oriented paradigm

The function of a piece of equipment is the intended


purpose. For example, the function of a distillation
column is to separate a mixture into products of
specifications. The functionof a piece of equipmentcan
be described by a set of parameters
F =f(Xh X2, ...)
When a parameter or a number of parameters are out of
specification we say the system is in malfunction.
Networks developed based on functional (or
malfunction) descriptions are called malfunction
networkmodels:
Fm =fnt(Xlm, X2m, ...)
For example, the malfunction network model of a
distillationcolumn is
Fm fnt(Flooding, Internals_inefficiency,
Excessive-pressure_drop,
Low_reflux_ratio, Lownumberofstages,
Lowgradefeed, ...)
The behaviour of a piece of equipmentrefers to the
behaviourand states of the equipmentat normal
operation. It can be describedby a number of
parametersand their relationships.
B =j{yl, Yz, ...)
It =It{Yh Y2, ..)
Ji Ji(yh Yz, ...)

When the parameters are beyond the specified ranges or


the behaviour equations are violated, we say the
equipmentis in misbehaviour. The networksfor trouble-

Building Bayesian Networks in Data-poor


Application Domains

Bayesian networks
Graphical models described above fall short being able
to provide more precise information. Since malfunction
trouble-shooting involves time-consuming checking and
validation, Bayesian networks that provide quantitative
probabilities are useful. A Bayesian network consists of
a set of variables and a set of directed links between
variables'', If there is a link from node A to B we say B
is a child of A, and A is a parent of B. Combination of
the following three connections, serial (Fig. 3(a,
divergent (Fig.3(b
and convergent (Fig.3(c
connections can form any complicated Bayesian
networkssuch as Fig.3(d).

(a) SerialCOIUl<<O.o
(b) DivUJingconnection

ruuy llClWor1< is a
combinatioo of r.SU""(a~ (b) and (e)

(d) A causol

Fig. 3 Basicconnections in a Bayesian network

Suppose there is a connection from B to A, given the


probability P(B) and the conditional probability
P(AfB), then the probabilityofthe event A n B is
P(A/B)P(B)

= P(A, B)

(1)

5615

Computers and Chemical Engineering Supplement (1999) S613-5616

Given the event A, the conditional probability P(B/A)


can be calculatedusing the Baye's rule
P(B 1A) = P(A 1B)P(B)
(2)
P(A)

If P(Rebolier M) = 0.85 and P(Pump=M)=0.75


Then P(Column=M) = 1*0.85*0.75+1*0.85*0.25
+1*0.15*0.75+0*0.15*0.25 = 0.9625
and P(Column=N) = 0.0375

For convergent connections, for example both nodes A


and C are directed to B, specifications can be made to
calculate P(BI A, C). Reasoningrules are available in a
complex networks with some node having known
values'.

BuildingBayesian networks in data-poor domains


Probably the most difficult step in converting a
graphical model to a Bayesian network is the
calculation of conditional probabilities. If a large
volume of data is available there are algorithms to
calculate all conditional probabilities directly", Often
the large number of conditional probabilities that need
to be specified makes this approach impracticable, even
in such data-rich domain as the medical field. Fig. 4
illustrates the complexity. FigA is apparently a
simplified network which shows that malfunction of the
reboiler and the pump at the bottom of a column can
both cause the column to fail. There are eight
conditional probabilities that need to be specified. In
practise, we just need to specify four and the other four
can be calculated. The major difficulty comes from the
conditional probabilities. For example, it is difficult to
find data to specify P(C=M I R=N, P=M), the
conditionalprobability that the column is in malfunction
on the conditions that the reboiler is normal but the
pump is malfunctioning. If the child node has three
parent nodes, each takes three values then 3x3x3 = 27
conditionalprobabilities need to be specified.
Bayesian networks for riskassessment
We divide Bayesian network applications into two
types, i.e. risk assessment and malfunction troubleshooting. In Fig. 4, if we know the prior probabilities of
the reboiler malfunction, here 0.02, and the pump
malfunction, here, 0.05 as well as the conditional
probabilities shown in Fig. 4, the risk assessmenttask is
to find the probability of the column being in
malfunction. If each node takes only two values, i.e.
malfunction or normal, we can assume the conditional
probabilities being either 1 or 0, unless we have data or
knowledge to specify them more precisely. Under this
assumption, the conditional probabilities for the
example of FigA will be as shown in Table 1.
Table I. Simplified conditional probabilities for Fig.4
Reboiler=M
Pumo=M Pump=N
CoImm

IM

IN

I
0

Reboiler=N
Pumo=M Pumo=N
I
0
0
I

The following reasoning rules can be obtained which


are clearly reasonable.
IF P(Reboiler =M) = 1.0,
THEN regardless the pump, P(Column=M) 1.0
IF P(Pump=M) = 1.0,
THEN regardless the reboiler, P(Column=M) = 1.0

P(P=M) = 0.05
P(P=N) = 0.95
P(C=M I R=N,P= = 0.04
P(C=M I R=N.P=M = 0.80
P(C=M I R=M.P=N) 0.90
P[C=M I R=M.P=M)= .99

P(C=N =N.P=N)= 0.96


P(C: I R=N.P=M)=0.20
P(C I R=M.P=N) 0.10
P =N I R=M.P=M)=0.01

Fig.4. The conditional probabilities need to be specified

Bayesian networks for malfunction trouble-shooting


The task of trouble-shooting is the inverse of risk
assessment. For the example of Fig. 4, the troubleshooting task is to fmd the likelihood values of reboiler
M and Pump M, given Column = M. Theoretically
the likelihoodcan be calculatedas follows,

P(R=MIC=M)

P(C = MIR = M)P(R = M)


P(C=M)

P(p=MJC=M)= P(C = MIP = M)P(p = M)


P(C =M)

In this study the above equations are not used because

of two reasons. First, the conditional probabilities and


the prior probabilities are difficult to estimate
accurately. Secondly, the Bayesian network tool used in
this study does not allow to give values for the node
Column in FigA to reason backward. The approach in
this study is to redraw the network of FigA and the new
network is shown in Fig.5.

Fig. 5 Redrawof Fig.4for trouble-shooting


P(R=MIC=M) is the conditional probability that the

reboiler is in malfunction, given that the column is


malfunctioning. Similarly, P(P=MIC=M) is the
conditionalprobability that the pump is malfunctioning,
given that the column is malfunctioning. Network
developed in this way will normally be a decision tree a special network that has only serial and divergent
connections and no convergent connection. This
arrangement provides a useful way for making use of
past failure cases to calculate the conditional
probabilities, even the number of failure cases is small.
For example, suppose we have collected ten cases of
column malfunctions, seven are due to reboiler failure
and three are due to pump malfunctions. Then we have
the conditionalprobabilities, P(R=MIC=M) = 7/10 and
P(Pump=MIColumn=M)

=3/10.

5616

Computers and Chemical Engineering Supplement (/999) S613--S6/6

Case Study
A prototype system for distillation columns troubleshooting and risk assessment has been constructed using
a Bayesian network tool HUGIN JO8 Fig. 6 shows part
of the graph assessing column Capacity Reduction
(node C24). The left column in Fig. 6 displays the
values of nodes during the reasoning. It is also possible
to display the values of nodes on the graph, as shown
for nodes Cl, C20, C24, C7, C21, C23 and C4. The
node titles are shown in Table 2. If we read the graph

from node C24, it basically indicates that column


capacity reduction is generally caused by high pressure
drop in the column (C20) and flooding (C23). Flooding
(C23) can be divided into downcomer flooding (node
C22 ) and jet flooding (C21). Factors causing the two
types of flooding are different. In Fig. 6, it also shows a
reasoning
process.
When
P(Cl=Excessive
loading)=0.85
and
P(C7=Poor
vapour
distribution)=0.75, the probability of capacity reduction
is P(C24=Capacity reduction)= 0.9625.

Fig.6 Partof the Bayesian network

Table 2 Node titles of Fig. 6


C24Capaeityreduction
C23-Column flooding
C22-Dowcomer flooding
C21-Jetflooding
CIOExcessiveL flow
CII-L flowpath restriction
C16-C 19-Factors that can cause L flowpath restriction
Cl2-Vapour leakageinto downcomer
Cl3-lmproper feed location
Cl-l-Bortom seal-pan flooding CIS- Foaming
CS-Excessive loading
C6-Fouledtrays
C7Poor vapour distribution C8-Improper installations
C9-Foaming
Cl-Excessive loading
C2.Improperselection of high pressure drop trays
C3-Lowopeninghole area
C4-0ther factors that can cause poor vapourpath restriction
Final Remarks
Bayesian networks .provide a graphical model for
representing the knowledge for trouble-shooting
distillation columns. A method is described for
constructing graphical models using process structural
as well as functional and behavioural decomposition
and through combining principal knowledge and past
failure records. A method for converting a graphical
network to a Bayesian network in data-poor domains is

also discussed. A case study is given which shows part


of a prototype system developed using a commercial
tool HUGIN.
References
1. Zanetti, R. (1997), ChernEngng, 5(June).
2: Kister, H.Z., (I997), Trans. IChernE, 75A, 563.7.
3. Xu, S.X. and Bowman, J.D., (1997), Chern Engng,
8(December).
4. Harrison, M.E. and France, 1.1. t (1989), Chern
Engng, 116(March), 121(April), 126(May).
5. Kister, H.Z., (1990), Distillation Operation,
McGraw-Hili, New York.
6. Jefferson, M., Chung, P.W.H. and KIetz, T.A.,
Pro cd. Hazard XIII. Process Safety-The Future. April,
1997, Rugby, UK.
7. Cox, A.W., Lee, F.P. and Aug, M.L., (1990),
Classification ofHazardous Locations, IChemE.
8. Jensen, F.V., (1996), An introduction to Bayesian
networks, DCL Press.
9. Wang, X.Z. t Chen. RH. and McGreavy, C. (1997),
Trans. IChernE, 75B, 210.
10. Hugin, 1998, http://www.hugin.com

You might also like