You are on page 1of 20

ARTICLE IN PRESS

Reliability Engineering and System Safety 93 (2008) 234253


www.elsevier.com/locate/ress

Formalisation of a new prognosis model for supporting proactive


maintenance implementation on industrial system
Alexandre Muller, Marie-Christine Suhner, Beno t Iung
Centre de Recherche en Automatique de Nancy (CRAN), Nancy Universite, UMR 7039 CNRSUHPINPL, Faculte des Sciences1er CycleBP239,
54506 Vandoeuvre-Les-Nancy Cedex, France
Received 13 February 2006; received in revised form 10 October 2006; accepted 6 December 2006
Available online 7 January 2007

Abstract

The importance of the maintenance function has increased because of its role in keeping and improving system availability and safety,
as well as product quality. To support this role, the maintenance concept has undergone several major developments that have led to
proactive considerations mainly based on a prognosis process, which normally allows selection of the best maintenance action to be
carried out. This paper proposes the deployment and experimentation of a prognosis process within an e-maintenance architecture. The
deployment follows a methodology based on the combination of both a probabilistic approach for modelling the degradation mechanism
and of an event one for dynamical degradation monitoring. The feasibility and benets of this new prognosis process is investigated with
an experiment using a manufacturing TELMA (TELe-MAintenance) platform supporting the unwinding of metal bobbins.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Proactive maintenance; Prognosis; Degradation; Process approach; Dysfunctional analysis; Dynamic Bayesian network

1. Introduction is regarded as the key element for acting on the opera-


tional requirements to carry out the global system
The growth in awareness of sustainable development performance [2].
practices in industry has imposed on industrial companies Nevertheless, it is difcult to control global performance
a need to integrate into their development strategy not only because the system environment is changing, its function-
the current economical nality but also some social and ing mode is dependent on product ow and component
environmental requirements [1]. These requirements are ageing continuously modies the system characteristics.
then, at least, passed on to the life cycle of the product and To address these new constraints, the current SMOC
production system. For example, in the operation phase of structure must evolve into new more cooperative and
the production system, the sustainability objective is proactive forms that are capable of anticipating the
mainly materialised at the manufacturing execution system performance deviation. Thus, the maintenance process
(MES) level through the concept of system maintaining in has undergone several major developments that have led to
operational conditions (SMOC). This concept is close to proactive considerations and the transformation of the
that of integrated logistic support and its purpose is to traditional fail and x practices into the predict and
satisfy the primary mission for which the system was prevent e-maintenance methodology [3]. The anticipation
designed and used (i.e., to maintain the expected perfor- action, which characterises this proactive maintenance
mance). Thus, SMOC engineering is considered as an strategy, is mainly based on monitoring, diagnosis,
integrated part of the system engineering process /http:// prognosis and decision-making modules. The implementa-
www.incose.org/S. SMOC implements concretely a set of tion of these modules within a unique system ends in an
logistics means and processes such as maintenance, which integrated system of proactive maintenance (ISPM) [4].
Nevertheless, despite being facilitated by the emergence of
Corresponding author. Tel.: +33 3 83684438. information and communication technology (ICT), the
E-mail address: Benoit.iung@cran.uhp-nancy.fr (B. Iung). ISPM development encounters difculties with regards to

0951-8320/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2006.12.004
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 235

industrial constraints because of the lack of methods and of the existing prognosis methods are component oriented
tools for modelling and integrating these fundamental without really taking into account system performance.
maintenance processes. This is even true for the prognosis Thus, on this basis, we propose to consider prognosis as
process, given its prediction capabilities [5]. Thus, our the ability to perform the future of an item from its
contribution aims at formalising the generic prognosis present, its past, its degradation laws and the maintenance
process for a methodology that combines both a probabil- actions to be investigated. This denition implies a degree
istic approach for (system) degradation mechanism model- of uncertainty because prognosis involves a projection into
ling and an event-based one for dynamical degradation the future and the future cannot be predicted with
monitoring. Indeed, prognosis has to analyse the impact of certainty. Therefore, the result of the prognosis process
degradation on the component itself and on the other can be summarised as follows:
elements of the production system to predict system
failures and investigate (future maintenance) actions for  If the forecast situation of a degraded system is
satisfying the primary mission (availability, productivity, considered as safe or satisfactory to perform the
safety, etc.). Accordingly, it enables one to offer a system goals, no action needs to be planned. The system
maintenance-aided decision-making tool capable of assist- will evolve to another acceptable degraded state and it
ing in selecting the best maintenance strategy. can drift under control.
Following this introduction, Section 2 outlines a clear  If the forecast situation is not acceptable, several
statement on the prognosis process. From this state of the maintenance alternatives must be investigated through
art, Section 3 describes the attack we proposed for a new the prognosis process.
formalised prognosis in terms of prognosis modelling
concepts and its implementation mechanisms within an e- In the latter case, it means that the current situation
maintenance architecture. The application and feasibility identied by the monitoring process will be used to assess
of this prognosis deployment on a subset of the TELMA new forecast situations resulting in the virtual deploy-
experimental platform are shown in Section 4. Finally, a ment of each alternative (combining simultaneously the
review and conclusions are developed in Section 5. degradation and the restoration of the system). Then,
the results of the different assessments will be used by
2. Statement on the prognosis process the aided-decision-making module to select, by compar-
ison, the most efcient maintenance policy. The execution
Prognosis is usually dened as the ability to perform a of the prognosis process is therefore related to its
reliable and sufciently accurate prediction of the remain- integration with the monitoring and decision-making
ing useful life of equipment in service [6]. Consequently, processes (Fig. 1).
the primary function of the prognostic is to project the From an engineering point of view, the deployment of
current health state of equipment into the future by taking the prognosis process can come from three types of
into account estimates of future usage proles. To support approaches as classied by Byington et al. [7]:
the SMOC purpose, this denition has to be generalised by
extending the concept of equipment at different abstraction  Experience-based prognostics: use statistical reliability
levels of a system (function, subsystem, etc.). Indeed, most to predict probability of failure at any point in time.

Functioning Monitoring
data process
Conceivable
maintenance plan
Current state of Prognosis
the process
process

Operating Estimation of the system future


knowledge performances

Direct/indirect Aided-decision
maintenance making process
costs

Selected
maintenance
plan

Fig. 1. The integrated system of proactive maintenance (ISPM).


ARTICLE IN PRESS
236 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

 Evolutionary/statistical trending prognostics: multi-


c

Co
variable analysis of system response and error patterns

nv
en
compared to known fault patterns. Model-Based

tio
 Model-based prognostics: fully developed functional

na
Causality Prognostics

lP
and physics-of-failure models to predict degradation Model

ro
b

gn
rates at given loads and conditions.

os
Event Model

tic
Evolutionary or

s
Degradation indicators Trending Models

Ap
Even if these approaches allow the development of

pr
Maintenance actions

oa
operational prognosis, they are not efcient enough to a

hc
support and address the new industrial constraints and the Probabilistic Model
complex assessment of the future system situations (i.e., Experience-based
Behavioural model
performance related to multi-criterion, multi-levels, multi- Prognostics
Degradation process
actors, etc.) as requested in an SMOC environment. In that
way, the prognosis process is still considered the Achilles Fig. 2. Positioning of the proposed methodology.
heel of ISPM, even though its goal is fundamental for
implementing prediction capabilities [8]. This is because
most of the time: 3. Methodology proposed for prognosis process development

 The approaches are component oriented For developing the right models at generic and particular
 The degradation concept and its impact are not taken levels, the methodology proposed is structured in three
into account stages, as detailed below:
 A gap exists between the complexity of the model
proposed and the real system.  The R&D phase aims at formalizing a set of generic
construction mechanisms required in the next steps for
developing the specic prognosis process. These me-
To be more efcient, the three different approaches are chanisms are fully independent of support tools and of
often used in combination as proposed in our work [9] for application areas.
exploiting prognosis not only at the component level but  The Engineering stage is intended to select a set of
also at the function and system levels (function is methods or tools for supporting in a concrete way the
supported by several components). In general, a compo- generic mechanisms previously dened. Thus, these
nent is subject to degradation phenomena such as ageing, mechanisms can be used by an engineer (a specialist)
wear or erosion. It implies that a function is subject to the for designing prognosis processes related to this
consequences of degradation occurring in all of its industrial system.
components. The degradation mechanism obtained is  The Operational stage denes the resources required for
complex but the function state can be determined from the integration and implementation of the previously
the combination of the states of each component. There- designed prognosis process on a given ISPM architecture.
fore, the methodology developed for the prognosis process
modelling must be generic regardless of the level considered This paper does not aim at developing the Engineering
(component, function or system); however, the degradation and Operational phases which are detailed and justied in
attributes and the meaning of degraded functioning mode [23] but more focus is on the R&D phase. However, a brief
will be dedicated to each level. description of the two last stages is given at the end of
To address this issue, our investigations led us to Section 3.6.
propose [4] the implementation of a prognosis process that The generic (i.e., independent of the tools selected and
is able to operate on different system levels within a holistic independent of the application) R&D mechanisms are
maintenance approach in the enterprise. In order to dened through ve modelling steps of the industrial
specify, design, develop and integrate the prognosis process system in general (Fig. 3):
into the ISPM, a complete methodology is dened. The
benet of this methodology is the combination of different  Mechanisms related to the functional modelling of the
types of knowledge for the prognosis process development. industrial system.
Indeed, the methodology is based on a combination of (a)  Mechanisms related to the dynamical modelling of the
stochastic models of degradation based on prior knowledge industrial system.
and expert judgement (experience-based prognostics), (b)  Mechanisms related to the elaboration of a behavioural
degradation indicators usable by the monitoring process model based on the rst two steps.
(data-driven prognostics), and (c) the causal relationships  Mechanisms related to the elaboration of an event
expressed by the physical laws on system evolution (model- model from the operational data issued from the
based prognostics) (Fig. 2). monitoring and the decision-making processes.
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 237

Prognosis
model Goal
Event model building Probabilistic and Process
event coupling
formalize P Event model
Maintenance integrates
Input Flow Function
actions Output flow

Degradation Behavioral (Mass, (Mass,


indicators model Energy or Energy or
Information) Information)
Probabilistic model building Support
Generic building
define results from...
mechanisms Dynamical
Functional model Fig. 4. Process formalisation.
model

results from...
results from...
particular (goal) process (Fig. 4) is thus broken up into
several sub-processes, and the same mechanism is
applied from the sub-processes until the expected level
Functional
of abstraction is reached (e.g., component).
Dysfunctional Stochastic
analysis analysis processes The proposed denition of a process relies on the four
concepts (goal, function, behaviour and structure)
Fig. 3. The proposed methodology (NIAM /ORM formalism11 [10]). introduced by Lind [12] and interpreted as follows
 Support (or Structure) represents an arrangement of
physical components.
 Mechanisms related to the prognosis model denition  Goal represents the purpose of the process (what the
through the integration of the event model with the process is intended to do in relation to the product
behavioural model. being transformed).
 Function is a desired action without considering
3.1. Functional modelling of the industrial system why or by what means this action is performed.
 Behaviour explains how a system does what it is
The rst modelling step consists of formalising the intended to do. The process behaviour is desc-
interactions between the system components by means of ribed by a rule-based approach, which takes into
causal qualitative relationships. To perform this action, a account the causal relationships between its input
dual functioningmalfunctioning reasoning of the indus- ows, the support and output ows. It can be
trial system is proposed. It is based on: identied as nominal (function fullled), degraded
(partial loss of the function) or failing (total loss of
(a) A process approach, which breaks down the whole the function).
system into processes and identies the components A process is linked to another by the ows it
supporting the processes and the ows consumed and produces or consumes. By denition, a ow (classed
produced by the processes. as Mass, Energy or Information) is composed of
(b) Causal modelling, this studies the degraded and failed objects characterized by attributes (Fig. 5).
states of the process model objects (components and Let us assume that the performance (goal achieve-
ows) and their consequences on the behaviour of the ment) of a process is directly measured on the
industrial system. produced ows and, more precisely, on the value of
its attributes.
(b) Secondly, the causal modelling determines the non-
(a) Firstly, the industrial system to be considered is nominal states of each support and each ow from
modelled through a functional analysis such as the the previous model. Then, the causes and effects of
process approach [11]. This step allows dening all the these states on the industrial system behaviour are
transformation processes of the systems at all the studied. Actually, this step consists of qualitatively
abstraction levels. The abstraction levels of a system dening all the causal relationships: object degradation
allow to study rst, the why (goal) and then, the statecauseconsequences Ri, which are collated in
what (the function performing the goal) until the Table 1.
how (the component supporting the function). A  R1 is associated to Nominal Behaviour
 R2, R3 and R4 are associated to Degraded
1
NIAM (natural language information analysis method)/ORM (object- Behaviour
role modeling).  R5, R6 and R7 are associated to Failing Behaviour.
ARTICLE IN PRESS
238 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

Current cycle

Product-cycle temporal specification

Previous cycle
is of

has
Information Flow

Energy Flow Flow


is characterized by characterizes Flow variable
(flow)

Mass Flow
is constituted by
constitutes Attribute

Object is characterized by characterizes State variable


(O)

Fig. 5. Flow concept formalisation.

Table 1
Causal relationships typology (adapted from [13])

F Flow Xn Nominal state


SP Support iF Input ow Xd Degraded state
R Relation oF Output ow Xf Failing state
Input ow Support Output ow Behaviour Causal relationship Type

IFn SPn OFn Nominal iFn4SPn-oFn R1


IFd SPn OFd Degraded iFd4SPn-oFd R2
IFn SPd OFd Degraded iFn4SPd-oFd R3
IFd SPd OFd Degraded iFd4SPd-oFd R4
IFd SPd OFf Failing iFd4SPd-oFf R5
IFf SPn/SPd or SPf OFf Failing iFf4(SPn3SPd3SPf)-oFf R6
IFn or IFd SPf OFf Failing (Fen3Fed)4SPf-oFf R7
Interaction between deterioration processes SP1d3SP1f-(SP2d3SP2f)4(SP3d3SP3f) R8

 For example, R4 and R5 materialise the same This modelling phase begins with the selection of the
relationships but for a behaviour going from critical degradation/failure modes and the identication of
degraded to failing state. the dependencies between the deterioration processes
 R8 is related to the common-mode degradation responsible for the appearance of modes (formalisation of
(failure) when the degradation of a component may the causal relationships R8). In a complex system, the
induce the degradation of other components. dependencies are represented by many interactions between
the deterioration mechanisms such as common cause
3.2. Dynamic modelling of the industrial system failure, cascade failure, redundancy,y Usually, these
mechanisms are modelled by stochastic processes and
The previous functional modelling, which represents the often by Markovian or semi-Markovian ones [14]. In
system from a synchronic point of view, does not allow the this study, the selected deterioration mechanisms were
elaboration of the behavioural model required by the formalised using discrete time Markov processes (DTMP)
prognosis process by itself. To build this behavioural with nite state space. This choice is justied in view
model, it is also necessary to model the system degradation of the computation constraints met in the continuous
evolution (diachronic vision). This stage consists of domain [15].
modelling all the physical mechanisms of deterioration, The representation of an independent deterioration
such as wear, cracks and corrosion, affecting the process process {Xk} is made up of the identication of its discrete
supports previously identied. state space wX and, then, the elicitation of its transition
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 239

matrix PX. If {Xk} is a semi-Markovian process, the phase- legibility and, on the other hand, the quality of the results,
type method is implemented. It allows substituting the the execution time and the available tools.
initial process by a DTMP whose behaviour can be Introduced by [24], DBN are directed graphical models
considered as equivalent to the semi-Markovian process of stochastic processes (deterioration ones, for instance).
one [17]. The transition probabilities pij (k) p(Xk xj/ The DBN graphical structure provides an easy way to
Xk1 xi) can be independent (homogeneous case) or not specify these conditional independencies and, hence, to
independent (non-homogeneous case) of the value of k. provide a compact parameterization of the model. In
In order to model stochastic dependent processes, we addition, DBN allows the merging, in the same model, of
have chosen a simple method, which consists of aggregat- sources of knowledge different to those depicted in the rst
ing the interactive process into a single model [16]. For step of the methodology: operative knowledge database
example, the combination of two processes {At} and {Bt} (OKD), expertise (expressed under logic rules, equations or
leads to the elaboration of a macro-process {ABt} subjective probabilities), observations, etc.
dened by the following:
3.3.1. BN/DBN generalities
 The states wAB result from the combination of the states Before describing the behavioural modelling step, let us
wA and wB (with the creation of a failing macro-state introduce some BN/DBN notations:
Af4Bf and the removal of the non-reachable states).
 The parameters pik,jl evolution result from the product
 B (z, y) is a Bayesian network if z (X, E) is a
of the parameters pij and phl according to Eq. 1(a)
directed acyclic graph (DAG) of nodes representing a set
(Change of state for A) and Eq. 1(b) (Change of state
of variables X {X1,y, Xn}, and a set of arcs E
for B). X representing dependence relations among the variables.
pih;il 1  pim phl , (1a)
mai
 The conditional probability table (CPT) yXi associated
with each node contains the probability of each state of
X
pih;jh 1  phn pij . (1b) the variable given every possible combination of states
nah of its parents: p(Xi/XPa(Xi)) in z.
 Considering a variable X[k], which evolves temporally,
the DBN represents the probability distribution over the
states of this variable during the period [0, T]. It can be
3.3. Behavioural modelling of the industrial system modelled by a static BN with T+1 variables (X[0], X[1],
y, X[T]) or by a compact form according to the
The denition of the (probabilistic) behavioural model canonical formalism introduced by Dean and Kanasawa
results from the integration of the functional (see Section [24], and then resumed in the dynamic reliability context
3.1) and dynamic (see Section 3.2) models in a unied by Welch et al. [25]. In this second case, the DBN is
formalism. For this reason, the selected formalism must be dened as a pair, (Xk; Xk+1), where Xk is a BN that
able to integrate simultaneously the structure of the denes the prior P(X[0]), and Xk+1 is a two-slice
industrial system (described by the processes and the temporal Bayes net (2TBN) that denes P(X[k]/
ows), the causal mechanisms (qualitative/quantitative X[k1]). With this model, the future slice (k+1) is
equations) and the deterioration processes (DTMP). In conditionally independent of the past, given the present
addition, it would be better that a computation method (k) (i.e., Markov properties).
(and/or a tool) be interfaced with the selected formalism in
order to make the quantitative analysis of the system Because of this canonical form, DBNs allow the
behaviour easier. decreasing effect of combinatorial explosion in reliability
Among the methods generally used in (dynamic) estimation by a synthetic description of the system.
reliability engineering, ve solutions seem to full all the Explosion means that the number of states needed to
previous requirements: model a complex system increases exponentially (a state for
each combination of elementary states) [26].
 Markov chains (MC) [18]. As the DBN representation (Fig. 6) is based on the
 Stochastic Petri nets (SPN) [19]. modelling of process entities, the obtained model is more
 Dynamic Bayesian networks (DBN) [20]. compact and readable than the corresponding MC model
 Dynamic fault trees (DFT) [21]. (Fig. 7). However, Weber et al. demonstrate the equiva-
 Boolean logic driven Markov processes formalism lence of the DBN and the MC models from a computa-
(BDMP) recently introduced by [22]. tional point of view [26].

We decided to use DBN [23] based on the following 3.3.2. DBN-building mechanisms
criteria: on the one hand, the strength of this formalism, From these general DBN concepts, the proposed
the difculty of construction and the graphical model methodology introduces at the R&D level a set of generic
ARTICLE IN PRESS
240 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

Flows are represented by m static nodes related to


p(Xk+1/Xk)
their m identied attributes. Thus, the state of a
Xk Xk+1 ow F is dened as a reachable combination of the
states of its m attributes (e.g., water ow is chara-
p(Xk+1/Xk) X1 X2 X3 X4 cterized by its volume, ow rate and pressure). The
X1 - p12 p13 0 static variables Xj representing the ow attributes
X2 0 - p23 P24 are generally quantitative and continuous. Never-
X3 0 0 - p34 theless, as the continuous running of DBN is difcult,
X4 0 0 0 - the static variables must be studied at discrete
times. This operation consists of cutting Xj into
Fig. 6. DBN modelling.
intervals with the same range by determining the
discrete pitch dXj adapted to the continuous state
space wXj.
p13 (b) The network structure is constructed by creating the
p12 p23 p34 directed edges. An edge is indicative of a conditional
X3 dependency between the variable that it links (e.g.,
X1 X2 X3 X4
the impact of the supports degradation on a pro-
p24 duced ow). As the processes are linked together by
their common ow (an output ow for the upstream
Fig. 7. DMC modelling.
process can be an input ow for the down-
stream process), the model structure consists of
specifying the oriented edges E (directed edges repre-
mechanisms used from the process model to develop the sent the dependencies (causality) between the variables)
probabilistic behavioural model in three times: that represent the causal relationships Ri between
the input ows, the support, and the output ows
(a) Identication of the DBN variables, (Table 1). For example, an edge [Xi[k]-Xj] linking
(b) Elaboration of the network structure by creating the Xi[k] and Xj, denes the impact of the deterioration
directed edges, mechanism i on the ow attribute j. By default, each
(c) DBN parameters elicitation. output ow attribute can have for parents every
attribute of the input ow and every degradation mode
(a) The main rule of the DBN building is to associate a of the related process. The result of steps (a) and (b) is
dynamic variable Xi[k] (dynamic node) to each shown in Fig. 8.
deterioration mechanism and a static variable Xj to (c) The last step of the DBN building is probably the most
each ow attribute. difcult; it is the denition of all the CPTs. Each CPT
As previously dened, the deterioration processes, contains conditional probabilities of a child node
which affect the support, can be considered indepen- (e.g., an output ow) being in a specic state given the
dent (i.e., there is no relation between the degradation states of its parents (e.g., support and input ows).
mechanisms) or dependent. In the former case, a
support of the process is represented by l dynamic
variables associated to their l degradation modes. In
the latter case, it is necessary to amalgamate the s
correlated deterioration processes into a complex
deterioration process (consequently, the support is
represented by ls+1 dynamic variables). This manip-
ulation is due to the difculties that persist in the exact
inference of DBN. However, the suitability of BN to
represent complex dependencies is not questioned.
The state space wX i k of each dynamic variable has
already been dened in the dynamic modelling step.
When the phase-type method is implemented, the states
xi do not have a physical meaning. In contrast, in the
case of a step-down deterioration, the states xi
correspond to degradation thresholds of the concerned
support. These thresholds are so dened either by their
physical meaning (e.g., oxidation) or by the conse-
quence of the degraded state on the process perfor-
mance (e.g., loss of conductivity). Fig. 8. DBN construction.
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 241

Event model
Monitoring process
Ix (t) Indicators
Degradation 3
event
Ix=I3
modeling SPn 2
SPd SPd
Ix=I2
1
Ix=I1 SPf

Time

Conceivable interventions
Aided-decision process Maintenance
event Support Type Date / Periodicity
modeling Cmp A Repair -
Cmp B Substitution Every 100days

System C Check Each day

Fig. 9. Event modelling.

To carry out this task, two solutions (eventually 3.4. Event modelling of the industrial system
combined) can be used:
 Learning the parameters on the dataset, The event model aims at ensuring the consistency of the
 Calling on an experts judgment. prognosis regarding the systems real evolution. Actually,
The choice depends on the sources of knowledge the event model is dedicated to providing the state of the
available for the studied industrial system. variables dened in the previous DBN behavioural model.
Therefore, it contains two types of variables (Fig. 9):
The rst alternative requires the possession of an OKD
supporting the implementation of learning algorithms. If (a) A set of degradation variables related to the indicators
all the DBN variables are observable, the learning amounts provided by the monitoring process.
to a simple counting. By contrast, the learning algorithms (b) A set of maintenance variables related to the con-
become more complex when several variables are hidden ceivable maintenance actions (the Future) on the
[27]. Obviously, the quality of learning depends on the support of each process provided by the aided-decision
quantity and the efciently of the available data. Among process.
the conceivable algorithms for this, the simplest consists of
parametric learning (i.e., probabilities resulting from the (a) The degradation variables dene the current situation
count of the individuals found in the database). of the system (the Present). They can be monitored on
With the second alternative, the transition probabilities line through the monitoring process of the ISPM.
of Markovian processes may be derived a priori from any Normally, the behavioural model has been used for
combination of process knowledge, statistical correlations designing the monitoring process: as far as possible,
and expert judgment, depending on the type of information each DBN variable Xi should be associated with a real-
available about that particular mechanism. The efciency time indicator Z X i . The set of these indicators
of this solution is totally dependent on the expert abilities. constitutes the degradation part of the event model,
However, because of the Bayesian approach, the collected and so the current values of all the indicators determine
operating data allow periodic updating of the conditional the current state of the degradation event model.
probabilities [28]. For static variables, the CPT elicitation (b) The maintenance part of the event model allows the
can be realized by using the physical law governing the prognosis process to take into account the maintenance
causality between the ow attributes and their parents. plan (the Future), as dened by the aided-decision
In that case, the law, which can be described by a process. It synthesizes the conceivable maintenance
determinist equation, is directly encoded in the CPT (see action applicable to the process supports (elementary
Section 4.1.3). components or systems). In practice, a maintenance
This elicitation phase is the last in the construction of the action is integrated within the event model by means of
DBN behavioural model, which will be also used to design a Boolean variable ICmp. This variable is associated
the event model (see Section 3.4) and, then, the prognosis with the intervention I applicable on the component
model (see Section 3.5). (support) Cmp. When this action is pre-programmed at
ARTICLE IN PRESS
242 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

the time l, the event variable ICmp is equal to 1 at time These modications are directly related to the efciency
t l, Else, the variable ICmp is equal to 0. of the maintenance action from which three main
categories are usually distinguished [29]:

3.5. Prognosis model of the industrial system  Perfect maintenance (as good as newAGAN): this
means that the component is considered new after each
The prognosis model results from the integration of the maintenance intervention. This maintenance is repre-
event model into the behavioural model. This integration sented in the DBN by hard evidence, which restores the
consists of creating a set of event DBN variables that component to its nominal state at intervention time (k).
correspond to the degradation (a) and maintenance (b) Then, the inference process updates the probabilities of
events formalized in the last phase (3.4). all the variables connected directly, or indirectly, to the
component.
(a) An observation variable Z(Xi) reects (eventually) the  Minimal maintenance (as bad as oldABAO): a
current value of each indicator Z X i provided by the minimal maintenance action leaves the system in the
event model. Z(Xi) is dened in the same way as Z X i . same state as it was before failure. With regard to the
An edge [Z(Xi)-Xj] links this new variable to its DBN model, it consists of doing nothing because, rstly,
observed variable Xi (Fig. 10.). In addition, the CPT of the state of the component before and after the
Xi is updated according to the following rule: If intervention is the same and, secondly, the degradation
Z(Xi)6unspecied, then Xi Z(Xi). process does not change.
(b) A maintenance variable ICmp is dened for each  Imperfect maintenance: generally, a maintenance action
planned/unplanned intervention carried out to main- is somewhere between AGAN and ABAO. Imperfect
tain or restore a component/function to acceptable maintenance materialises this fact and reduces either the
working conditions. ICmp has an impact on each age or the ageing rate of the component, better
deterioration process that affects the component than it was but worse than it could be as new. This
(support) Cmp (Fig. 10). The impact quantication maintenance is represented in the DBN by soft evidence
on system performances (i.e., impact on the support materialised by a vector of the probability distribution
degradation) can be achieved by two ways in the CPT over the variable possible values (component states).
of the selected deterioration process. Firstly, the repair For example, if it is impossible to know the exact
is considered as a modication of the probability component state after repairing but, if there is a
density associated with the states of the degraded/failed probability of X that the maintenance will be perfect
component. Secondly, the repair acts on the deteriora- and a probability of Y that the component will be
tion process whose parameters are updated every time restored to its rst degraded state, then a soft evidence
an intervention on the component occurs. (likelihood) [X Y y] is entered in the model [9]. In

Behavioral model
Ix (t) Indicators
3
Ix=I3 Degradation
SPn 2 event If Z(Xi) Unspecified,
SPd integration ThenXi = Z(Xi)
Ix=I2 SPd
1
Ix=I1 SPf
Time Z(Xi) Xi

REPCMP A A deterioration

Conceivable interventions
Maintenance
Support Type Date /Periodicity event
CMP A Repair - integration
SUBCMP A A deterioration[k-1]
CMP A Substitution Every 100 days
C deterioration
System C Check Each day
CHSYST C

C deterioration[k-1]

Fig. 10. Event model integration.


ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 243

summary, we have introduced (through the construction Inter-operability is seen as the ability for a system or a
of several models) a set of generic mechanisms that product to work with other systems or products without
allow the development of a prognosis process. However, special effort of the part of the customer /http://interop-noe.
these mechanisms should be still supported by a set of org/INTEROP/presentationS. This is why we propose
concrete methods and tools to be implemented by an at the operational stage to develop such a technical
engineer (a specialist). architecture [9]. This technical architecture is built using
a set of various technical solutions such as PLCs, OPC
3.6. Implementation of the R&D mechanisms server, eld buses, real-time database, Ethernet, Oracle
database,y (Fig. 12) and materialises a rst appli-
The industrial deployment of the previous prognosis cation based on ICT components of the e-maintenance
approach is supported today, in engineering phase, by concept [35].
different tools to help engineer for developing particular With these architectures, engineers will have at their
prognosis related to the system to be maintained. The tools disposal a formalised guide based on method combina-
are the following: tion for helping them in the design of the prognosis process
models, as well as the execution of the prognostic.
 MEGA tool /http://www.mega.comS to support pro-
cess approach (process model of the system),
 CASIP tool /http://www.predict.frS to support causal 4. Implementation of the proposed methodology on a
relationship studies through the implementation of TELMA platform
failure modes and effects analysis (FMEA) [30] and
hazard and operability studies (HAZOP) [31] methods. In relation to the operational phase, the deployment of
 ACCESS database to support the event model, the prognosis process is made today according to
 BAYESIALAB tool /http://www.bayesia.comS to experimentation in order to demonstrate prognosis feasi-
support the development and the implementation of bility and potential benets. This experimentation con-
the probabilistic models (among the prognosis ones). cerns the TELMA /http://www.aip-primeca.netS platform
whose technical architecture is built on the industrial
The results of this engineering phase are synthesized in CASIP product.
the application architecture diagram depicted in Fig. 11. TELMA is a platform materialising a physical process
However, the concrete implementation of this applica- dedicated to unwinding metal strip. This process is similar
tion architecture required the integration of its elements in to concrete industrial applications such as sheet metal
a unique software/hardware infrastructure [32]. Assem- cutting and paper bobbin cutting. The physical process is
bling the previous tools to form a whole (efcient) ISPM is divided into four parts: bobbin changing, strip accumula-
a difcult task because of the problems of inter-operability. tion, punchingcutting and advance system (Fig. 12). Each

Dysfunctional analysis
C
CASIP
CASIP of supports
FMEA

Dysfunctional
CASIP analysis of flows
HAZOP

MEGA Functional Deterioration


decomposition of Functional
Process processes
the system model

Knowledge of Causal relationships Dynamic


the maintenance model
actions
ACCESS 7
Probabilistic Markovian
Event model processes
model
Knowledge of
the degradation Event variables Probabilistic variables
indicators

BAYESIA
Prognosis Lab
model

Fig. 11. Application architecture.


ARTICLE IN PRESS
244 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

Internet Network
EMPACix Customer
CASIP FMECA HAZOP CMMS EMPACix CASIP Customer
CASIP Customer

TDMS SQL TDMS Oracle


Server CASIP EMPACix

Server PC Server PC CMMS PC (Sciences


CASIP Ethernet AIP Lorrain EMPACIX Campus or ESSTIN)
Switch AIP Lorrain
Touch screen API_DEF : Failure generator API_CDE : Control/Command OPC & SAM
MAGELIS TSX Premium PLC TSX Premium PLC Servers

Vlan
Industrial Ethernet RLI
Web
Inputs/Outputs Altivar 58
Altivar 58
Momentum Cam 240v-370W
2
240v - 370W

System consol Bobbin Strip Punching- Advance


Changing accumulation cutting WEB
Break cam
Pneumatic
Bobbins tour cylinder

Accumulation motor Advance motor

Fig. 12. TELMA platform description.

part is composed of several components such as pneumatic  Its function, to punch/cut the strip.
cylinder, chuck and a marking system.  Its behaviour, described by all relationships existing
between the input ows (Advance output rotation and
4.1. Prognosis process development Advanced provided by the Advance process) and the output
ows (Finished product, Scraps and Punching state).
The prognosis process development, detailed hereafter, is
focused on the punchingcutting part of TELMA and Fig. 14 shows the breakdown of this process into ve
more precisely on the subset whose nal purpose is to elementary sub-processes, (respectively, supported by the
punch the strip. The translation movement (used to operate Punching belt and pulley system, the Camshaft system, the
the punching/cutting) is induced by a camshaft that drives Sensors SQ101 & SQ102, the Punching split and slide
a crosshead. Two springs allow the punch to remain in system and the Punching/cutting tool).
contact with the cam and so to follow the up and down This process breakdown is then used as the common core
motion. of the dysfunctional analysis (FMEA and HAZOP).
Firstly, the FMEA determines, for each process support,
4.1.1. Punchingcutting functional modelling the causes and the consequences of the listed degradation
As explained in Section 3.1, the rst modelling action is modes. An occurrence of a degradation mode (caused by a
intended to design the process model of the punching deterioration process) generates one or more ow devia-
cutting part of TELMA. tions. Fig. 15 presents, for example, the FMEA of the
The process to punch/cut the strip is therefore dened Punching belt & pulley system, which supports the process
as described in Fig. 13: To transmit the punching rotation. According to this
study, the support is subjected to:
 Its support, the Punching subsystem.
 Its goal, directly measured on the output ow attributes,  Mechanical wear of the belt. This wear process leads to
i.e., the properties of the Finished product. the occurrence of the degradation mode Belt slip or/
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 245

Punching in top
Camshaft position
system Cam shaft SQ101
Punching Spilt system
DT282
& Slide system
Punching Spilt
& Slide system PLC part
Punching/ KA262 10.1
Cutting
tool Punching in low
Advanced strip To punch/cut Scraps SQ102 position
Advance the strip Finished products
Punching state DT283
output rotation
Punching sub-
PLC part
system KA263
10.2

Fig. 13. To punch/cut the strip process and system.

Advance output rotation

To transmitt the To transform the


punching rotation Punching rotation rotation in punching
movement

Punching belt & Camshaft system


pulley system To report the
punching position
Punching state
(low/high)

Output cam movement Sensors SQ101 &


SQ102
Slide translation (previous cycle)
To operate the
punching Slide translation

Punching split &


slide system

To execute the
Scraps
punching
Advanced strip Finished products

Punching/cutting
tool

Fig. 14. Punchingcutting process model.

Cause Degradation / Failure Consequence


Mechanical wear of the
Slip MORE of start time for the Shaft punching rotation
belt
NO start time for the Shaft punching rotation
Belt break
LESS of stop time for the Shaft punching rotation
Pulley wear Pulley misalignment NO start time for the Shaft punching rotation

Fig. 15. Extract of FMEA related to the Punching belt & pulley system.

then the occurrence of the failure mode Belt break. In misalignment. It causes the deviation NO start time for
case of slip, this deterioration process causes the the Shaft punching rotation ow.
deviation MORE of start time for the Shaft punching
rotation ow. In case of break, it causes the deviation Secondly, the HAZOP table determines for each output
LESS of stop time or NO start time for the Shaft ow the causes and consequences of the deviations of their
punching rotation ow. attributes. The occurrence of the deviation of an output
 Wear process of the pulley. This deterioration process ow is due to either the deviation of an input ow or the
leads to the occurrence of the degradation mode Pulley occurrence of a support degradation/failure mode (cause).
ARTICLE IN PRESS
246 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

Attribute Deviation Cause Consequence


Torque LESS LESS of torque for the Advance output rotation LESS of torque for the Output cam movement

MORE MORE of torque for the Advance output rotation MORE of torque for the Output cam movement

Angular LESS of angular velocity for the Advance output LESS of angular velocity for the Output cam
LESS rotation movement
Shaft punching rotation velocity
MORE of angular velocity for the Advance output MORE of angular velocity for the Output cam
MORE rotation movement
LESS LESS of start time for the Advance output rotation LESS of start time for the Output cam movement
NO start time for the Advance output rotation
Start NO NO start time for the Output cam movement
Belt break of the Punching belt & pulley system
Belt slip of the Punching belt & pulley system
MORE MORE of start time for the Output cam movement
MORE of start time for the Advance output rotation
LESS of stop time for the Advance output rotation
LESS LESS of stop time for the Output cam movement
Belt break of the Punching belt & pulley system
Stop
NO NO stop time for the Advance output rotation NO stop time for the Output cam movement

MORE MORE of stop time for the Advance output rotation MORE of stop time for the Output cam movement

Fig. 16. Extract of HAZOP related to the Shaft punching rotation ow.

It leads to the deviation of one or more output ows fOK; Degradation; HSg and a transition matrix
consumed by the downstream processes (consequences). PU bearings , which contains the bearings wear process
Fig. 16. presents, as an example, the HAZOP analysis of parameters, i.e., the transition probabilities pij
the Shaft punching rotation ow produced by the process p(Ubearings k Uj/Ubearings k1 Ui).
To transmit the punching rotation. The temporal attributes 2 3
 pOK; Degradation pOK; HS
start time and stop time (already introduced in the FMEA) 6 7
and two new shape attributes, the angular velocity and PU bearings 4 0  pDegradation; HS 5
torque of the rotation, are studied. They can be subjected 0 0 
to LESS or MORE deviations. PU bearings is a stochastic matrix, i.e., the entries in each
From the causality principle, the causes of these column are nonnegative real numbers whose sum is 1.
deviations are either the degradation/failure modes of the In the same way, the fretting wear process of the cam is
Punching belt & pulley system, which supports the To formalized by a Markovian process {Ccam} character-
transmit the punching rotation process, or the deviation of ized by a state space wC cam fOK; Degradationg and a
an attribute of the Advance output rotation ow consumed transition matrix PC cam .
by the same process. In the same way, the effects of the
Advance output rotation ow deviations concern the  
 pOK; Degradation
attributes of the Output cam movement ow produced by PC cam
0 
the To transform the rotation in punching movement process
below the studied process. PC cam is a stochastic matrix. There is no failed state for
The use of DBN formalism avoids the translation of this process because the probability of reaching it is
causal relationships issued from the FMEA and HAZOP considered negligible.
studies under causal equations. (b) The Punching belt & pulley system modelling
(Table 2) is related to the stochastic dependence case.
4.1.2. Punchingcutting dynamic modelling This support is subjected to two stochastically depen-
The dynamic modelling of the processes to transmit the dent deterioration mechanisms:
punching rotation and to transform the rotation in  Fatigue wear of the belt, which is formalized by a
punching movement details successively the stochastic (a) Markovian process {Fbelt} with three states (OK,
dependence and (b) independence situations depicted in Lengthening and Break) and three transition prob-
Section 3.2. abilities [p1, p2, p3], is dened as:
o p1 p(Fbelt k Lengthening/Fbelt k1 OK).
(a) The dynamic modelling of the camshaft system o p2 p(Fbelt k Break/Fbelt k1 Lengthening).
(support of the process To transform the rotation into o p3 p(Fbelt k Break/Fbelt k1 OK).
punching movement) illustrates the stochastic indepen-  A misalignment mechanism of the pulley, which is
dence. Actually, this support is subjected to two formalized by the Markovian process {Dpulley} with
deterioration mechanisms wear of bearings and three states (OK, Soft misalignment and Hard
fretting wear of the cam, which are assumed to be Misalignment) and two parameters [p4, p5], is
independent of each other. The bearings wear process is dened as:
formalized by a Markovian process {Ubearings} o p4 p(Dpulley k Soft M./Dpulley k1 OK).
characterized by a state space wU bearings o p5 p(Dpulley k Hard M./Dpulley k1 Soft M.).
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 247

Table 2 represented in Fig. 17. For example, the pro-


Punching belt & pulley system modelling cess {Ubearings} associated with the dynamic variable
p1 p(Fbelt k Lengthening/Fbelt k1 OK)
Ubearings[k] is represented by the nodes U bearings [k1]
p2 p(Fbelt k Break/Fbelt k1 Lengthening) and U bearings. The Punching rotation ow is, in
p3 p(Fbelt k Break/Fbelt k1 OK) contrast, represented by four static variables, and so
p4 p(Dpulley k Soft M./Dpulley k1 OK) four nodes are associated with its attributes: Punching
p5 p(Dpulley k Hard M./Dpulley k1 Soft M.) rotation.Start, Punching rotation.Stop, Punching rota-
tion.Torque and Punching rotation.Angular velocity.
(b) Afterwards, the DBN structure is established because
The stochastic dependence is dened as follows: of the causal relationships, which have been formalized
during the functional/dysfunctional analysis. For ex-
 When the pulley reaches the soft misalignment state, ample, the causal relationships: LESS of Advance
wear of the belt is accelerated. The set of parameters [p1, output rotation.torque-LESS of Punching rotation.-
p2, p3] are then substituted by another set [p0 1, p0 2, p0 3] torque and MORE of Advance output rotation.tor-
(p0 i4pi). que-MORE of Punching rotation.torque, leads to the
 When the pulley reaches the hard misalignment state, creation of the A1 edge (Fig. 17). Similarly, the relation
the belt slip and the rotation transmission are broken. Belt break-NO Rotation punching.start lead to the
The Punching belt & pulley system is thus failing. creation of the A2 edge.
(c) Finally, the last step of the DBN construction concerns
The aggregation of {Fbelt} and {Dpulley} leads to the the elicitation of various parameters. A well-suited
elaboration of a macro-process {FDpunching} characterised learning algorithm included in BAYESIA Lab allows
by a state space wFDpunching composed of: the estimation of the transition probabilities pi from the
operational learning database available on TELMA.
 Nominal state: OK The existing learning methods guarantee to produce
 A degraded state with only lengthening of the belt : very good results given sufciently large data sets.
Length. However, in the case of small data sets (this situation
 A degraded state with only soft misalignment of Pulley : concerns several deterioration processes on TELMA),
Soft M. the parameters learning is not good and it is necessary
 A state with the simultaneous presence of two degrada- to combine existing frequency data with expert judg-
tions: Length. & Soft M. ment in order to obtain realistic results.
 A failing macro state: HS.

The evolution parameters of the process {FDpunching} are For a dynamic variable, such as FDpunching [k], the wear
obtained by combining the parameters of the initial processes are dened by a parameter set [p1, p2, p3, p4, p5,
processes. The result is given by the following transition p0 1, p0 2, p0 3] obtained from the operating data. These
matrix: PFDpunching transition probabilities are associated with the following
2 3 numerical values:
 p1 1  p4 p4 1  p1  p3 p1 p4 p3
60  0 p4 1  p2 p2 7 Fatigue wear of the belt (generalized Erlang distribution)
6 7
6 7 Without misalignment With a pulley misalignment
60 0  p0 1 1  p5 p5 p 3 1  p5 7
0
6 7 p1 0.026 p0 1 0.1
6 7
40 0 0  p5 p0 2  p5 p0 2 5
p2 0.026 p0 2 0.1
0 0 0 0  p3 0.047 p0 3 0.05
Misalignment process of the pulley
All the parameters pi will be determined during the DBN
development.
p4 0.003 p5 0.03

4.1.3. Punchingcutting behavioural modelling As explained in Section 3.3, the CPT yFD punching
The DBN model of the process To punch/cut the strip depicted in Table 3 is obtained with these parameters.
results from the implementation of the generic building If the operational data related to the relevant deteriora-
mechanisms by using the knowledge formalized in the two tion process are insufcient, such as in the {Ccam} case, the
previous steps (Sections 4.1.1 and 4.1.2). transition probabilities are obtained from knowledge of the
deterioration law. In this specic case, the degradation
(a) Each deterioration process identied during the FMEA apparition time of the cam follows a Weibull distribution
study is associated with a dynamic variable, whereas with a scale parameter of Z 12000 and a shape parameter
each ow attribute identied during the HAZOP study of b 5. Therefore, the CPT of the node C cam[k+1] can
is associated with a static variable. This action leads to be established by using the equation that denes the hazard
the denition of six dynamic and 18 static variables rate evolution (Table 4).
ARTICLE IN PRESS
248 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

Fig. 17. DBN structure of to punch/cut the strip process.

Table 3
CPT yFDpunching

FDpunching k+1 OK Length. Soft M. Length. & Soft M. HS

FDpunching k OK 99.924 0.026 0.003 0 0.047


FDpunching k Length. 0 99.971 0 0.003 0.026
FDpunching k Soft M. 0 0 99.901 0.025 0.074
FDpunching k Length. & Soft M. 0 0 0 99.945 0.055
FDpunching k HS 0 0 0 0 100

Table 4
CPT of yCcam

Ccam k+1/Ccam k OK HS

OK 5
 k
4  b1  4
1  lk1  12000 12000 lk bZ kg 5
lk 12000 k
Z 12000
HS 0 1

As introduced in Section 3.3, the CPT of a static variable 4.1.4. Punchingcutting event modelling
(e.g., a ow attribute) can be determined according to a The event model interfaces the prognosis process with
physical law. This law, which governs the interactions the monitoring and aided-decision-making processes. As
between the ow attribute and its parents, is described explained in Section 3.4, two stages are identied in the
by means of deterministic equations. Each of these development of the Punchingcutting event model:
equations species the variable value with respect to the
value/state of their parent variables. For example, the (a) The rst one is relating to the Punching subsystem
fullment of the CPT yPunching rotation.start is done from the system current state.
deterministic equation Punching rotation.start f (b) The second one aims at formalizing the maintenance
(FDpunching, Advance output rotation.start) as illustrated actions and their impacts on the support of the process
by Fig. 18 and Table 5. To punch/cut the strip.
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 249

Similarly, the belt substitution is a non-periodic main-


tenance action represented by the Boolean variable SUBBelt
that is by default in the False state.
FD punching [k-1] By contrast, it is in the True state when this action is
scheduled at a date l:
Punching rotation.start =
f (FD punching
Advance output rotation.start)  p (SUBBelt k True) 1 if k l
 p (SUBBelt k True) 0 if k6l
FD punching
4.1.5. Prognosis model of the punchingcutting system
Punching rotation.start The prognosis model of the process To punch/cut the
Advance output rotation.start Punching rotation.start strip results from the integration of the event variables
(indicator and maintenance action variables) in the
Fig. 18. Punching rotation.start CPT determination. previous behavioural model as depicted in Section 3.4.
Table 5
Punching rotation.start determination (a) The integration of the Punching period indicator (Flow
deviation indicator case) causes, for example, the
Punching rotation.start creation of a new DBN variable: Z(Punching period)
If Advance output rotation.start NO
which updates the state of the variable Punching
then Punching rotation.start NO, Else period (Fig. 19).
If FDpunching (OK or Pulley soft misalignment),
then Punching rotation.start Advance output rotation.start, The states of Z(Punching period) are based on the states
If FDpunching Belt Lengthening, of the variable Punching period in which the state
then Punching rotation.start Advance output rotation.start+0.5,
If FDpunching Belt Lengthening+Pulley soft misalignment,
Undetermined (i.e., no available information) is added.
then Punching rotation.start Advance output rotation.start+1, The impact of the punching period observation on the ow
Else Punching rotation.start NO. attribute Punching period is formalized in the CPT
yPunching by the following equation:

(a) The degradation part of the event model associated  If Z(Punching period)6Undetermined, Then Punch-
with the process To punch/cut the strip is composed of: ing period Z(Punching period).
 An indicator of the Punching period formalised by
the discrete variable Z (Punching period), continu- Similarly, the integration of the belt wear indicator
ously updating using the instrumentation available (deterioration indicator of a component case) leads to the
on TELMA. creation of the variable, Z(Fbelt), which updates the state of
 A failure indicator of the belt associated with the the variable FDpunching (Fig. 20).
Boolean variable Z (Fbelt), updated as soon as an As previously, the states of the variable Z(Fbelt) are
operator visually inspects the state of this component. based on the states of ZFbelt in which the state Un-
(b) The maintenance part of the event model falls into the determined is added.
maintenance actions suggested by the aided-decision- The impact of observing the belt on the deterioration
making process to improve the state of the process To process {FDpunching} is difcult to establish because of the
punch/cut the strip. These potential maintenance
actions can be periodic (period: p) or not:

 Belt substitution  Camshaft repair


 Bearings lubrication (p)  Slide system repair
 Bearings substitution  Cutting tool sharpening (p) Slide translation.start
 Pulley balancing  Cutting tool substitution
 Springs substitution

The lubrication of the bearings is a periodic (T)


maintenance intervention. It is represented by the Boolean Slide translation.stop Punching period
variable Lbearings. This variable is in the state True when
the periodicity is reached:

 p (LBearings k True) 1 if k Tn with n natural Slide translation.linear speed Z (Punching period)


integer 40
 p (LBearings k True) 0 if k6Tn Fig. 19. Integration of the punching period indicator.
ARTICLE IN PRESS
250 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

FD punching [k-1] SUB Belt


FD punching [k-1]

Z (F belt)
FD punching
BAL pulley Z (F belt)
FD punching
Fig. 20. Support degradation indicator integration.

Fig. 21. Maintenance event model integration.

lack of direct correspondence between the states of the


p (FDpunching (k+1) OK/SUB Belt k & FDpunching k
indicator ZFbelt and those of the variable FDpunching.
HS) 1*
Actually, the knowledge of the state of belt degradation
* When a failure of the Punching belt/pulley system occurs,
does not allow the exact determination of the state of the
there is a complete substitution of this system, and so the
process {FDpunching}. It also depends on the pulley
system comes back to its nominal state (OK).
misalignment process. The update of the FDpunching state
is therefore partial. It is integrated in the CPT y FD punching
according to the deterministic equation FDpunching f Z On the other hand, the pulley repair is considered an
(Fbelt): imperfect maintenance action. The expertise says that, after
this intervention, there is a probability of 0.8 that the pulley
FDpunching will be restored to its nominal state and a probability of 0.2
If Z (Fbelt) OK,
that it stays in the degraded state. The impact of the pulley
If FDpunching [k1] (OK4Length.),
repair on the deterioration process {FDpunching} is also
then FDpunching [k] OK,
formalized by the following conditional probabilities:
If FDpunching [k1] (Soft M.4Length &
Soft M.) P (FDpunching k+1 OK/BAL Pulleyk & FDpunching k
then FDpunching [k] Soft M., OK) 1
If FDpunching [k1] HS then FDpunching P (FDpunching k+1 Length./BAL Pulley k & FDpunching k
[k] HS, Length.) 1
If Z(Fbelt) Degradation, P (FDpunching k+1 OK/BAL Pulley k & FDpunching k
If FDpunching [k1] (OK4Length.), Soft M.) 0,8
then FDpunching [k] Length, P (FDpunching k+1 Soft M./BAL Pulley k & FDpunching k
If FDpunching [k1] (Soft M.4Length. & Soft M.) 0,2
Soft M.), P (FDpunching k+1 Length./BAL Pulley k & FDpunching k
then FDpunching [k] Length. & Soft M., Length. & Soft M.) 0,8
If FDpunching [k1] HS then FD punching P (FDpunching k+1 Length. & Soft M. /BAL Pulley k &
[k] HS, FDpunching k Length. & Soft M.) 0,2
If Z(Fbelt) HS, then FDpunching [k] HS. P (FDpunching k+1 OK/BAL Pulley k & FDpunching k
HS) 1*
*When a failure of the punching belt/pulley system occurs,
The integration of the maintenance actions (Belt
there is a complete substitution of this system, and so the
substitution and Pulley repair), which can be implemented
system comes back to its nominal state (OK).
on the Punching belt/pulley system, enables the creation of
two DBN variables: SUBBelt and BALPulley (Fig. 21). These
variables update the state of the variable FDpunching, at each This was the last step concerning the design of the
maintenance intervention. prognosis model. Now, we are going to run the prognosis
On one hand, the belt substitution is considered a perfect process in the operational phase.
maintenance action that restores this component in its
nominal state after the intervention. The maintenance
4.2. Prognosis process operation
impact on {FDpunching} is formalized into the CPT
yFDpunching by means of the following conditional prob-
By using the prognosis model (DBN), the prognosis goal
abilities:
is, from the current state of the punching/cutting subsystem
p (FDpunching (k+1) OK/SUB Belt k & FDpunching k and the maintenance alternative dened by the aided-
OK4Length.) 1 decision-making process, to assess its future situation and
p (FDpunching (k+1) Soft M./SUB Belt k & FDpunching k its related performance X time units later. This operation is
Soft M.4Length. & Soft M.) 1 supported by the tool BAYESIA Lab, which runs the DBN
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 251

simulation required by the prognosis process using its To highlight this fact, Fig. 22 depicts the evolution curve
iterative inference algorithm. The simulation usually of these three criteria for the period [0; 1000]. These curves
unfolds in two stages: provide evidence of the impact of each bearings lubricant
performed at each 100 time units in the process. These
(a) Initialization of the prognosis model. The event model results, which constitute the estimation of the systems
state is integrated within the prognosis model by means future performance, are then transferred to the aided-
of an observation le (text le) generated by the Event decision-making process, as dened in Section 2 (Fig. 1).
model database (Fig. 11). Indeed it is mandatory to take into account their
(b) Simulation stage. The simulation is made on the basis maintenance costs before analysing the expected perfor-
of the system evolution from its current state to a mance provided by the prognosis process. Otherwise, the
future time (k X) and by assessing the state reached interpretation has no real meaning.
at this time with implementation of the maintenance In this specic case, the prognosis process is launched a
strategy dened in the event model. More precisely, second time because analysis of the previous results shows
BAYESIA Lab updates at every time slice (k) the that the degradation of the process Punching the strip
probability distributions of all the DBN variables performances is mainly caused by deterioration of the
during the requested period [0; X]. As soon as the camshaft bearings and the punching split/slide system.
simulation is over, the state of all the variables at time Therefore, it was decided to replace these components. This
(k X) is available. decision implies the modication of the state of the event
model (i.e., new maintenance alternative). However, in this
simple case, the integration of these actions in the
To punch/cut the strip prognostic prognostic can be done by injecting two hard evidences
Assume that it is necessary to realize the prognostic of the into the prognosis model at time (k 1000). The new
process To punch/cut the strip at time (k 1000). The results are again transferred to the aided-decision-making
goal of the prognosis process is to determine at this time: process, which can repeat the action until a satisfactory
maintenance strategy with regards to the expected perfor-
 The deterioration state of the punching system compo- mances is obtained.
nents
 The state of the ows, which are consumed/produced by
the punching system. More precisely, the goals achieve- 5. Conclusion
ment of the process is measured by the state of the
produced ows: Punching period, Finished products.- The main contribution of this paper is a methodology
cutting state and Finished products. surface appearance. structured by generic modelling concepts for prognosis
(a) First, the export (and then the integration) of the event process formalization and deployment based on a combi-
model state into BAYESIA Lab allows the initialisa- nation of both a probabilistic model and a dynamic
tion of the prognosis model. In this example, the monitoring architecture. Moreover, experimentation on an
observation le contains the following information: industrial system demonstrated the feasibility of this
 Current state of the system at time (k 0) methodology. Based on these results, further developments
o Z(Punching period)0 45.5 should be implemented to extend the experimentation
o p(F belt Degradation)0 p(F belt HS)0 0 context to the whole platform. It is also expected to
 Maintenance alternative on period [k 0; k 1000] validate the added value of this methodology by comparing
o Periodic lubrication of the bearing each 100 time
units
1
o Gbearings 100 Gbearings 200 y G bearings 1000
0.9
True
Probability of remaining in the

0.8
(b) Then, the simulation is launched for the requested
nominal state (OK)

period (i.e., 1000 time units). The duration of the 0.7

simulation is less than 3 min 30 s. As soon as it is ended, 0.6


the states of all the variables at time (k 1000) are 0.5
provided by BAYESIA Lab in terms of probability 0.4
distribution. From the results, there is a probability of 0.3
Finished products. cutting state
0.78 that the punching process would not be able to 0.2 Finished products. surface appearence
full its nal goal 1000 time units later. Besides, the 0.1 Punching period

attributes cutting state and surface appearance of the 0


ow Finished products are subjected to an important 0 100 200 300 400 500 600 700 800 900 1000
Time
degradation. They have, respectively, probabilities of
0.71 and 0.45 of being in the failing state HS at Fig. 22. Evolution of the punching system performance for the period [0;
(k 1000). 1000].
ARTICLE IN PRESS
252 A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253

its performance with performance data already obtained by Transition matrix: PX Transition probability: pij
conventional approaches. Event modelling step
In addition, other future developments will be devoted in Maintenance action:
Degradation indicator: ZX
the short term to take into account in the prognosis model Action typeComponent
the inuence of several operational/environmental condi- Dynamic Bayesian Network
tions on the deteriorations processes evolution [33], and in Dynamic variable: Xi[k] Static variable: Xj
the medium term to reinforce the weight of the event model Dynamic nodes: Xi & Static node: Xj
in the prognostic by extrapolating the values of the Xi1]
indicators in the future (using a data-driven approach Edge: [X-Y] CPT: yX
[34]) and by integrating it into the prognostic. Indicator variable: Z(X) Time: (time)
Finally, some improvements should be made to improve Cmp : Component ICmp Intervention on the
the degree of assistance of the prognosis process designer, component
notably during the ow attribute denition. One solution
would consist of adapting the dysfunctional analyses
FMEA/HAZOP to the treatment of temporal properties 2TBN two slice temporal Bayesian network
for ow attributes and to the formalisation of the ABAO as bald as old
dependencies between the deterioration processes. AGAN as good as new
BDMP Boolean logic driven Markov processes
Acknowledgements BN Bayesian networks
CPT conditional probability table
This research has been carried out within the group DAG directed acyclic graph
PRODEMAS (Maintenance decision-making processes for DBN dynamic Bayesian networks
controlling the dependability and productivity of industrial DFT dynamic fault trees
systems) of the CRAN. The authors thank all the DTMP discrete time Markov processes
PRODEMAS members for their comments and especially FMEA failure modes and effect analysis
Dr. Philippe Weber, for his support on the formalisation of HAZOP hazard and operability studies
the DBN model. We would also like to thank the ICT information and communication technology
PREDICT company and the AIPL people for their help ISPM integrated system of proactive maintenance
during the development of the ISPM technical and MC Markov chains
operational architectures. MES manufacturing execution system
The continuation of this work is devoted to the NIAM natural language information analysis method
implementation of the proposed approach at a business OKD operating knowledge database
level for a more strategic decision-making purpose. These ORM object-role modelling
future developments will be done with the support of SMOC system maintaining operational conditions
Professor Adolfo Crespo Marquez in the context of a new SPN stochastic Petri nets
project setting up between our university and the Uni- Telma tele-maintenance
versity of Sevilla. This project is called: Modeling policies
for the improvement of production systems dependability
(Project number DPI 2004-01843) and it is sponsored by References
the Spanish government.
[1] Venkatasubramanian V. Prognostic and diagnostic monitoring of
complex systems for product lifecycle management: challenges and
Notation opportunities. Comput Chem Eng 2005;29(6):125363.
[2] Al-Najjar B, Alsyouf I. Selecting the most efcient maintenance
Functional modelling step approach using fuzzy multiple criteria decision making. Int J Prod
Process: Process Function: Function Econ 2003;84:85100.
[3] Lee J, Ni J. Editorial of advanced engineering informatics. 2003;
Flow: Flow Finality: Finality 17(34): 107.
Support: Support [4] Leger J-B, Morel G. Integration of maintenance in the enterprise:
Dysfunctional modelling step towards an enterprise modelling-based framework compliant with
Degradation Mode: Causal relation: proactive maintenance strategy. Prod Plann Control 2001;12(2):
Degradation mode Cause-Consequence 17687.
[5] Provan G. Prognosis and condition-based monitoring: an open
Failure Mode: Failure Deviation type: systems architecture. In: Proceedings of the fth IFAC symposium on
mode DEVIATION fault detection, supervision and safety of technical processes.
Flow attributes: Attribute Washington, USA, 2003. p. 5762.
Dynamic modelling step [6] Lebold M, Thurston M. Open standards for condition-based
maintenance and prognostic systems. In: Proceedings of the fth
Deterioration process:
State space: wX annual maintenance and reliability conference. MARCON 2001,
{Deterioration process} Gatlinburg, USA, 2001.
ARTICLE IN PRESS
A. Muller et al. / Reliability Engineering and System Safety 93 (2008) 234253 253

[7] Byington C, Watson M, Roemer M, Galie T. Prognostic enhance- [21] Dugan JB, Bavuso S, Boyd M. Dynamic fault-tree models for fault-
ments to gas turbine diagnostic systems. In: Proceedings of the IEEE tolerant computer systems. IEEE Trans Reliab 1992;41(3):36377.
aerospace conference. Big Sky, USA, 2001. [22] Bouissou M, Bon JL. A new formalism that combines advantages of
[8] Wang P, Vachtsevanos G. Fault Prognosis using dynamic wavelet fault-trees and Markov models: Boolean logic driven Markov
neural networks. In: Proceedings of the AAAI spring symposium on processes. Reliab Eng Syst Saf 2003;82(2):14963.
equipment maintenance service and support. Polo Alto, USA, 1999. [23] Muller A. Contribution to the proactive maintenance of manufactur-
[9] Muller A, Suhner M-C, Iung B. Probabilistic vs. dynamical prognosis ing system: formalisation of the prognosis process. PhD thesis, Henri
process-based e-maintenance system. In: Proceedings of IFAC- Poincare University, Nancy, France, 2005 [in French].
INCOM04, information control in manufacturing. Salvador, Brazil, [24] Dean T, Kanazawa K. A model for reasoning about persistence and
2004. causation. Comput Intell 1989;5:14250.
[10] Halpin TA. Object-role modelling (ORM/NIAM). In: Handbook [25] Welch R, Thelen T. Dynamic reliability analysis in an operational
on architectures of information systems. Berlin: Springer; 1998 context: the Bayesian network perspective. In: Proceedings of the
[Chapter 4]. fourth workshop on dynamic reliability: future directions. 2000. p.
[11] ISO International Organisation for Standardisation. ISO 9000:2000 277307.
quality management systemsrequirements. Geneva, Switzerland; [26] Weber P, Jouffe L. Reliability modelling with dynamic Bayesian
2000. networks. In: Proceedings of the fth IFAC symposium on fault
[12] Lind M. Modeling goals and functions of complex industrial plant. detection, supervision and safety of technical processes. Washington,
Appl Artif Intell 1994;8:25983. USA, 2003. p. 5762.
[13] Muller A, Suhner M-C, Iung B. Bayesian network based proactive [27] Neapolitan RE. Learning Bayesian networks. New Jersey: Prentice
maintenance. In: Proceedings of the seventh international conference Hall, Upper Saddle River; 2003.
on probabilistic safety assessment and management. PSAM 7 and [28] Celeux G, Corset F, Lannoy A, Ricard B. Designing a Bayesian
ESREL04, Berlin, Germany, 2004. network for preventive maintenance from expert opinions in a rapid
[14] Grall A, Berenguer C, Dieulle L. A condition-based maintenance and reliable way. Reliab Eng Syst Saf 2006;91(7):84956.
policy for stochastically deteriorating systems. Reliab Eng Syst Saf [29] Pham H, Wang H. Imperfect maintenance. Eur J Oper Res
2002;76:16780. 1996;94(3):42538.
[15] Janssen J, Blasi A, Manca R. Discrete time homogeneous and non- [30] Lawley HG. Operability studies and hazard analysis. Chem Eng Prog
homogeneous semi-Markov reliability models. In: Proceedings of the 1974;70:10516.
third international conference on mathematical methods in reliability. [31] IEC 60812 standard. Analysis techniques for system reliabilitypro-
MMR 2002, Trondheim, Norway, 2002. cedure for failure mode and effects analysis (FMEA). 1985.
[16] Muppala JK, Trivedi KS. System dependencies in Markov depend- [32] Leger J-B. A case study of remote diagnosis and e-maintenance
ability modelling. In: Fault-tolerant systems and software, FTS-95. information system. Keynote speech of IMS2004, international
New Delhi, India: Narosa Publishing House; 1995. p. 3847. conference on intelligent maintenance systems. Arles, France, 2004.
[17] Limnios N. Semi-Markov processes and reliability. Boston: Birkhau- [33] Weber P, Munteanu P, Jouffe L. Dynamic Bayesian networks
ser; 2001. modelling the dependability of systems with degradations and
[18] Billington R, Allan RN. Reliability evaluation of engineering exogenous constraints. In: Proceedings of the 11th IFAC symposium
systems: concept and technics. New York: Plenum Press; 1983. on information control problems in manufacturing. Salvador, Brasil,
[19] Dutuit Y, Chatelet E, Signoret J-P, Thomas P. Dependability 2004.
modelling and evaluation by using stochastic Petri nets: application [34] Schwabacher M. A survey of data-driven prognostics. In: Proceed-
to two test cases. Reliab Eng Syst Saf 1997;55(2):11724. ings of AIAA infotech@aerospace conference. Arlington, USA,
[20] Weber P, Jouffe L. Complex system reliability modelling with 2005.
dynamic object oriented Bayesian networks (DOOBN). Reliab Eng [35] Iung B. From remote maintenance to MAS-based e-maintenance of
Syst Saf 2006;91(2):14962. an industrial process. J Intell Manuf 2003;14(1):5982.

You might also like