Formalization of Technical Knowledge in Dependability on an Industrial System

with a Probabilistic Relational Model (PRM)

G. Medina Oliva, P. Weber, E. Levrat, B. Iung

CRAN, Nancy-Université, CNRS, Boulevard des Aiguillettes B.P. 70239 F-54506 Vandœuvre lès Nancy (e-mail:
{gabriela.medina-oliva, benoit.iung, eric.levrat}@cran.uhp-nancy.fr,{philippe.weber}@esstin.uhp-nancy.fr

Abstract: This paper proposes a methodology to develop a decision-making aid system which purpose is
to evaluate dependability and performance in manufacturing processes. This model is based on a new
formalism, the probabilistic relational model (PRM) adapted to deal with large and complex systems. The
objective is to evaluate system’s performances in order to optimize the enterprise maintenance strategies.
The methodology is formalized from functional, dysfunctional and informational studies of the technical
industrial systems. This methodology is applied, to model the reliability and the output flow attributes of
a water heater system.
Keywords: Probabilistic relational model, maintenance, system performances

1. INTRODUCTION
The importance of the maintenance function has increased,
due to its role in improving availability, performance
efficiency, products quality, on-time deliveries, environment
and safety requirements, and total plant cost effectiveness at
high levels (Riis et al., 1997; Alsyouf, I. 2007). In order to
improve business performances, it is required to analyze
aspects such as maintenance, risk analysis and dependability
which allow to forecast the gaps between nominal and non-
nominal operation of the system (degradation, failure, etc).
Some papers such as Mahadevan (2001), Boudali and Dugan
(2005b) Langseth and Portinale (2007) show the increasing
interest on the use of Bayesian Networks to estimate and to
improve reliability and safety of systems over the last decade.
Nevertheless, this method is limited by the number of
variables in a model.
For this reason, this paper formalizes a methodology to
define an extension of Object Oriented Bayesian Networks,
the Probabilistic Relational Model to model technical
knowledge in order to help decision-making for risk,
maintenance and dependability domains. The network is not
defined by a graph but in a language. The proposed
methodology has originality on formalising, by means of
PRM, the models from prior knowledge on the system
functioning, malfunctioning and informational point of view
to estimate the overall performance of the process (reliability
and compliance of output flows).
2. PROBLEM STATEMENT
There are several works in the maintenance field especially in
the optimization of maintenance strategies, nonetheless this is
a complex issue since it is necessary to take into account
different conflicting criterion such as availability, safety and
costs, etc., (Kiker et al., 2005). As described by Crespo, 2007
there is a wide diversification in the maintenance problems.
Maintenance is composed of a set of activities for which it is
difficult to find procedures and information support systems
in one place to ease the improvement process. Also there are
some works that emphasis on the lack of maintenance
management models. There is a lack of models that could
improve the understanding of the underlying dimensions of
maintenance (Crespo, 2007). Maintenance is somewhat
“under-developed” (Jonsson, 1999; McKone et al., 2001)
with a lack of effective prevention methodologies and the
integration of standard methods in manufacturing companies
in most continents.
In that sense, the challenge is to formalize a generic model of
complex systems to evaluate maintenance strategies that
allows to evaluate the global performance integrating various
dimensions such as technical, organizational, human,
informational, decisional and financial dimensions correlated
with system's behavior and nowadays environmental factors
which are necessary to quantify failure scenarios and risky
situations. As a consequence, increasingly studies aim at
covering the whole of these factors. Nevertheless such
analyses are often difficult to achieve because of the
importance of required resources. This matter adds
complexity to the systems’ modelling due to the interaction
between the different kinds of knowledge.
Furthermore, while modelling these factors, it is required to
take into account the knowledge integration of diverse nature
such as qualitative (organizational and human analyses) and
quantitative (technical analyses) knowledge with several
abstraction levels (Muller, 2007).
This paper, explains the main concepts of an approach to
formalize a model required to evaluate risk analysis,
maintenance and dependability. The idea of this
formalization is to unified multiple and different kind of
knowledge in one model.
In the technical aspect the object is to be able to model the
impact on system global performances based on functional,
dysfunctional and informational analyses. The idea is to
formalize the interactions between an industrial system and
the maintenance system using treatment and data models
such as SADT, AMDEC, HAZOP, etc. (Figure 1).
Figure 1: Methodology to formalize knowledge within a
PRM network
Once there is a structure methodology to formalize this
knowledge, it is necessary to establish a coding semantic with
the required knowledge and programming rules to obtain a
generic maintenance model which enable to predict or
diagnostic the impact of the influencing factors on the system
global performances.
In that way, this research work is developed within the
SKOOB project sponsored by the French National Research
Agency. This project focuses on the development of a generic
model based on probabilistic relational model (PRM) (Getoor
et al., 2007) which enable to solve the formalism of complex
models in risk analysis, maintenance and dependability,
applied to various socio-economic systems of strategic
importance (nuclear, food industries, medical or social
organizations). The multi-disciplinary team to develop this
project involves some industrial partners (EDF, SOREDAB,
University Hospital of Nancy, INERIS), some scientific
partners (LIP6, CRAN, ERPI) and a software developer
partner (BAYESIA).
3.- PROPOSITION OF PRM APPROACH
To create a generic, complex and big size model to evaluate
risk, maintenance and dependability it is essential to use a
compact representation of knowledge. To make it possible,
the knowledge must be divided into groups of “local
knowledge”, that could be added according to the needs in
order to form a whole coherent model.
Concerning our problematic, bayesian networks (BN) appear
to be a solution to model complex systems. Nonetheless, one
of the weak points is that BN are not adequate for dealing
with very large complex systems. Such systems require an
explicit representation of additional types of structure: the
notion of an object, a complex structured domain entity with
its own properties; and the notion of a class of objects, that
captures properties common to an entire set of similar
[1] EDF: Electricity of France (www.edf.fr/)
[2] SOREDAB: Research and Development Society of Food Industry Bongrain
[3] CHU: University Hospital of Nancy (www.chu-nancy.fr/)
[4] INERIS: National Institute of the Industrial Environment and Risks (http://www.ineris.fr/)
[5] LIP6: Laboratory of Informatics of Paris 6 (www.lip6.fr/)
[6] CRAN: Nancy Research Center for Automatic Control( http://www.cran.uhp-nancy.fr/)
[7] ERPI: Research Team about Innovative Processes (http://www.inpl-nancy.fr/francais/rechvalo/r_labos/rp_erpi.php)
[8] BAYESIA: http://www.bayesia.com
objects. The Object-Oriented Bayesian Networks (OOBNs)
extended the language of BNs with these additional concepts
(Weber and Jouffe, 2006). However, the language of OOBNs
is quite restricted, in a way that is problematic in our domain.
Nevertheless, maintenance, dependability and risk analysis
require sometimes the representation of arbitrary relations
between objects and uncertainty over system structure. An
extension of the OOBNs, the Probabilistic Relational Model
(PRM) formalism allow to model real-world situations
(Pfeffer et al., SPOOK). This language allows to represent
important types of uncertainty that cannot be accommodated
within the framework of traditional BN or OOBN:
uncertainty over the set of entities present in our model, and
uncertainty about the relationships between these entities
(Koller and Pfeffer, 1998). Another advantage of the PRM is
their inference aspect. Actually, it is possible to make queries
so the inference is made in a specific compiled part of the
model. It is not necessary to calculate the whole model that is
never built but that is described in a language.
The development of a generic technical model with PRM for
risk, maintenance and dependability analysis aim at solving
the scientific issues in the domain such as: a.- improvement
on the engineering processes models, making it possible to
manage complexity and to re-use of generic components; b.-
increasing the expressive power of the current modeling
techniques, especially because the structure of model will be
known during the exploitation phase (the configuration and
number of components, etc).
4. KNOWLEDGE FORMALISATION
The proposed modeling approach consists, from functioning
systemic analysis, (a) in representing the abnormal operation
(malfunctioning) (Muller, 2007 and Weber P., Jouffe L.
2006), (b) in representing the informational point of view and
then (c) in formalizing and unifying these results in a unique
model by means of the language based on PRM.
Functioning system representation
According to the standard AFNOR X 50-151: “the functional
analysis is an approach that consists in searching, ordering,
characterizing, organizing into hierarchies and/or to develop
the functions of a product (system) expected by the user”.
The functional modeling of an industrial system consists in
formalizing, by qualitative causal relationships, the
interactions between the functions performed by each of the
sub-systems until the component level (elementary
functions), so it is possible to use a Structure Analysis and
Design Technique (SADT) as a functional approach method.
The system functioning modeling is based on the principle of
activity and sub-activities until elementary activities,
supported by components, are emerging. Each activity (Figure 2) fulfils finality, which is to modify a “product”
carried out by the manufacturing system. It produces or For this aspect, there are used the following dependability
consumes flows such as “Having to Do” (HD) materialising methods [Villemeur, 1988]:
the Input/Output (I/O) finality, “Knowing How to Do” - FMECA: to model failure modes of the functions,
(KHD) materialising the I/O knowledge, “being Able to Do” failure modes of the components, failure
(AD) representing I/O energies, resources, activity support consequences (impact on the flow and other
and finally “Wanting to Do” (WD) materialising the I/O functions) and the criticality of the failure.
triggers. - HAZOP: to model flow deviation, cause of flow
deviation and failure consequences (impact on the
WD having to KHD allowing to know
trigger activity how to do activity flow)
- Fault tree (FT), reliability block diagram or
AD having to be AD having to
be recycled
Bayesian networks (BN) to model the logical links
used by activity RWD Report on
the activity state
of events or logical links between components.
Activity in relation to WD
HD having to
be transformed HD transformed Informational system representation
by activity by activity

AD supporting the activity Each flow is characterized by the state of the variables related
Figure 2: Flows and Activity Representation to its morphologic, spatial or temporal properties of the
objects that composed it (i.e.: pieces and flow of pieces) and
For example, the output flow WD is a report (RWD) that by the flow variables that are express as the quantity of
represents the informational result of the Input HD product objects per time unit (such as a flow rate) [Mayer, 1995]. So,
flow transformed by the activity. the state variables and the flow variables can be regroup in
one denomination called flow attribute. In that way, to
Malfunctioning system representation measure the performance of a function, it is assume the
hypothesis that it can be evaluated directly from the flow
The functional model could be used to develop the attributes. The object representation allows to identify these
malfunctioning analysis, whose objectives are to identify the flow properties or attributes, and it can be represented as
degraded and failure states of the components and of the shown on the Figure 3
flows, and then to determine the causes and consequences of
these states on the industrial system behaviour.
The degradation is spread to the rest of the system through
the flow exchanging between processes, according to the
causality principle:
Figure 3: Entity-relationship diagram of a function and their
- The potential cause of the degradation of a process is flows
the deviation of an input flow attribute or the
deterioration of its support. Unification of Technical knowledge in a PRM model
- Contrarily, the potential effect of the degradation of
a process is the deviation of an attribute of its output To model the different aspects of a system in one model, it is
flows or its support. required to take into account the different types of knowledge
as shown on the Figure 4:
The industrial system is in degradation or failure mode when
there is a flow deviation and/or a deterioration of the supports
of the process.
- The flow deviation is linked to the qualitative or
quantitative deviation of a flow attribute compared
to its nominal value.
- The support deterioration is related to the apparition
of a physical mechanism of deterioration (Muller,
2007).

The dysfunctional analysis also involve the identification of

groups of elementary events or combination of events that
lead to a failure event, as well as, the identification of the
logical links between essential components to perform the
system mission. Figure 4: Knowledge integration in a CPT

[1] EDF: Electricity of France (www.edf.fr/)

[2] SOREDAB: Research and Development Society of Food Industry Bongrain
[3] CHU: University Hospital of Nancy (www.chu-nancy.fr/)
[4] INERIS: National Institute of the Industrial Environment and Risks (http://www.ineris.fr/)
[5] LIP6: Laboratory of Informatics of Paris 6 (www.lip6.fr/)
[6] CRAN: Nancy Research Center for Automatic Control( http://www.cran.uhp-nancy.fr/)
[7] ERPI: Research Team about Innovative Processes (http://www.inpl-nancy.fr/francais/rechvalo/r_labos/rp_erpi.php)
[8] BAYESIA: http://www.bayesia.com
In this figure it is shown how to integrate in a CPT the One reference is an order to send to another object whose
variables and the conditional probabilities according to the type is another class. The declaration of a reference in a class
different system’ point of view: functioning view, is made as a simple reference:
malfunctioning view and the informational view. Name_of_the_class reference_name;

The construction of a PRM model needs three steps: Or as a multiple reference:

- Formalisation of the network structure from the
functional analysis (input and output variables of a
process) and the system malfunctioning (logical
links between components)
- Definition of input and output variable states, as it is
described in the malfunctioning analysis.
- Definition of the conditional probabilities given in
the functional analysis, combinatory logic or
expertise.
Also, it is important to know that:
- To represent the input flow (energy, information or
material flow) of a function, there could be several
variables for each flow.
- To define an output flow, there are necessary several
CPT based on the input flows. There must be one
CPT for each output flow.
Name_of_the_class[]reference_name;
An attribute have the following characteristics:
° a type (enumeration, ticks or ranges);
° a name;
° a list, eventually empty, of attributes whose it depends on;
° a conditional probability table CPT ()
The declaration of an attribute follows the following scheme:
Name_of_the_type name_of_the_attribute_
dependson list_of_dependances {
declaration_of_the_cpt };
In the case that it exist a n-aire relationship, it is needed to
specify a quantificator (or aggregator). Here is a possible but
non-exhaustive list of the quantificators: min, max, exists,
forall, count, mean.

To integrate the different kind of knowledge in big and
complex models, the SKOOB project is developing a
language to represent PRM models.
6. STRUCTURE OF A PROJECT IN THE CURRENT
SKOOB LANGUAGE
This language is inspired in Java language (specially the
declarative part of the classes definitions), because it is one of
the more common languages used in the present. The
hierarchy of packages is that of the files of the project, and
each file includes the definition of a compilation unit. The
current characteristics of SKOOB language are described as
followed:
Variable declaration
Conditional probability table (CPT) declaration
In this case, the declaration order is essential. The first value
correspond to the probability of P(A[1]|P_1[1], P_2[1], …,
P_N[1]), the second value to the probability P(A[1]|P_1[1],
P_2[1], …, P_N[2]), and so on. The A[i] values correspond
to the label I of the attribute A, P_j[i] correspond to a label i
of the j-ieme attribute of the dependency list of A.
typeEtat perform_a_function dependson
exists(inputs.etat,OK), etat {
[1.0, 0.0, 0.0, 0.0, // P(OK | true,
OK), P(OK | true, NOK), P(OK | false,
OK), P(OK | false, NOK)
0.0, 1.0, 1.0, 1.0] // P(NOK | true,
OK), P(NOK | true, OK), P(NOK | false,
OK), P(NOK | false, NOK)
};

Three types of variables can be defined: enumerations, ticks Inheritance (specialisation)

and range variables. For example the enumeration variables
are random variable whose support is defined by the A class could inherit of another class, the syntaxe is the same
enumeration of its modalities: as that from Java:
type name_of_type mod1, mod2, ...; class name_of_the_class extends mother
class {
To declare a variable typeState it possible to write: // Body of the class
type typeState OK, NOK; }

Declaration of a class Model declaration

A class is declared by the following way:
class Name_of_the_class { A model is declared similar as a it is done with class:
// Body of the class model name_of_the_model {
} // Body of the model
The body of a class is composed by the declarations of }
attributes and references (or slot).
Instanciation
One can have an instance of an object or of a reference.
7. APPLICATION
The proposed method is applied to a classical example of a
water heater process. The objective of the thermal process
(show in Figure 5) is to ensure a constant water flow rate
with a given temperature. The process is composed of a tank
equipped with two heating resistors R1 and R2. The system
inputs are the water flow rate Qi, the water temperature Ti
and the heater electric power P that is controlled by a
computer. The outputs are the water flow rate Qo and the
temperature T that are regulated around an operating point
(Qi= Qo= 20 l min-1 and T= 50 °C). The input temperature
of the water T= 20 °C is assumed to be constant.
Figure 5: Water heater process
Functioning system representation: SADT model
Figure 6 presents the level A0 of the system SADT analysis.
This figure depicts the interaction between the system and the
external environment through the AD, HD and RHD flows.
The main functionality of the process is:
- to provide warm water.
Figure 6: Level A-0 of the SADT
The next figure presents the level A0 describing the four
functions that are necessary to perform the main task of the
system (Figure 7):
- to transform pressure into Qi (A1),
- to control V and P (A2),
- to transform Qi into H and Ti into T (A3),
- to transform H into Qo.
Figure 7: Level A0 of the SADT
Figure 8 formalises the function A3 ‘to transform Qi into H
and Ti into T’ from the elementary functions:
- to stock water supported by the component TANK,
- to heat water supported by the component HEATING
RESISTOR.
Figure 8: Level A3 of the SADT (First two functions)
Malfunctioning system representation: FMEA, HAZOP,
reliability block diagram, fault tree.
For this case, the study will be applied to the functions A31
and A32, so the components of these functions are indexed in
the FMEA analysis (Table 1). The failure modes of each
component are defined as well as their effects. The causes are
linked with the component states or the unavailability of the
electric energy required to supply the component. Therefore,
the loss of energy is a common cause of the six failure
modes.
Table 1: FMEA of the functions A31 and A32
The FMEA allow to determine the failure causes of a
component. For example, for two failure causes, the states in
a TPC (non-exhaustive list) would be:
- “OK”: if the component is available,
- “C1”: if the component is unavailable due to the failure
cause C1,
- “C2”: if the component is unavailable due to the failure
cause C2,
Then, it is necessary to study the possibilities of flow
deviation and their causalities through an HAZOP study
(Table 2). The flow deviation is linked to the qualitative or
quantitative variation of an attribute compared to its nominal
value.
Table 2: HAZOP of the function A31 and A32 of the water
heater process
Since R1 and R2 work on parallel to fulfill the function “to
heat water”, it is possible to obtain the reliability of the
support (AD support flow) of this function. The state of each class Abletodo extends Input_variable {
heating resistor was defined as follow: 80% is available, 5% }// AD- Support flow represented by material, energy, human or software flow.
works in a maximum level, 5% there is power loss in it and class Watingtodo extends Input_variable {
10% the heating resistor is unavailable. So with this }// WD- Materialising the trigger of the function
information, it is possible to build a dynamic bayesian type typeStateTank OK, Fissure, NOK;
network as shown in the Figure 9. class Tank extends Input_variable {
typeStateTank state_tank{
// OK Fissure NOK
[0.80, 0.1, 0.1] };
} // AD-Equipment that supports the function “To stock water Qi to H” / States of the heating
resistors coming from FMEA

type typeStateHR OK, Maximun_level, Power_loss, NOK;

class Heatingresistor extends Input_variable {
typeStateHR state_heatingresistor{
// OK Maximun_level Power_loss NOK
[0.92, 0.05, 0.0075, 0.0225] };
Figure 9: DBN of the parallel heating resistors } // AD-Equipment that supports the function “To heat water from Ti to T” / States of the
heating resistors coming from FMEA/ The conditional probabilities of this functions are taken
from the dynamic bayesian network showed in figure XX.
The result shows that the AD support flow of this function is
available 92% of the time, 5% works in a maximum level, class Function_RHD {
Flow Havingtodo;
0,75% there is power loss in them and 2,25% the heating Fllow Abletodo;
resistors are unavailable Fllow Watingtodo;
Support equipment;
boolean etat dependson Havingtodo.state, Abletodo.state, Watingtodo.state, equipment.state
{
Moreover, it could be possible to use fault trees or reliability [
block diagrams in these cases. They can describe the logical 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
links of events in order to obtain the reliability of the support 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0,
of a function in cases where there are redundancy or k/n 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
]
relations between components. Their limitation is that they };
represent boolean variables, that is why it was not }//All the input flows such as material, energies and informational flow must exist in order to
perform the function/ RHD- Represents the success of the function’ mission
appropriated for this example.
type typeStateRHDtostock OK, Spilled, Loss, NOK;
class RHDtostock extends Function_RHD {
Informational system representation: Entity-relationship typeStateRHDtostock state_RHDtostock{
diagram // OK Spilled Loss NOK
[0.80, 0.05, 0.05, 0.1] };
} // States of the function “To stock water Qi to H” coming from the FMEA
The informational point of view let identify the flow
type typeStateRHDtoheat OK, Maximun, Losspower, NOK;
properties and attributes. In our example, this aspect is class RHDtoheat extends Function_RHD {
represented within the following Figure 10and Figure 11: typeStateRHDtoheat state_RHDtoheat{
// OK Maximun Losspower NOK
[0.80, 0.05, 0.05, 0.1] };
} // States of the function “To heat water from Ti to T” coming from the FMEA

class Output_variable extends Flow{

Function_RHD source;
boolean etat dependson source.state { [1.0,0.0,0.0,1.0] };
Figure 10: Entity-relationship diagram of the A31 function } // To represent all the output variables such as AD, RHD and HD

type typeStateWL OK, More, Less, No;

class Waterlevel extends Output_variable {
typeStateWL state_waterlevel{
// OK More Less No
[0.85, 0.05, 0.05, 0.05] };
}// DF- Main outflow of the function “To stock water Qi to H” and input flow of the function
“To heat water from Ti to T”- The level is an attribute coming from the entity- relationship
Figure 11: Entity-relationship diagram of the A32 function diagram / States of the water rate flow coming from the HAZOP

type typeStateWT OK, More, Less;

Finally, it is shown the integration of different kind of class Watertemperature extends Output_variable {
knowledge in the SKOOB language as follow: typeStateWT state_watertemperature{
// OK More Less
[0.90, 0.05, 0.05] };
type typeState OK,NOK; }// DF- Main outflow of the function “To heat water from Ti to T”- The temperature is an
// a class to represent the “flow exchanges”, in order to link the output flow of a function as an attribute coming from the entity- relationship diagram / States of the water rate flow coming
input flow of the following function from the HAZOP
class Flow{
boolean state { [1.0,0.0] }; class HDoutput extends Output_variable {
} }// HD- Main flow transformed by the function
class Input_variable extends Flow{
} // To represent all the input variables such as AD, HD and WD. class ADoutput extends Output_variable {
}// AD- Flow of material or energy to be recycle
class Havingtodo extends Input_variable {
}// HD- Main flow. //INSTANCIATON
model m{
type typeStateWFR OK, More, Less, No;
class Waterflowrate extends Havingtodo{ //Function//
typeStateWFR state_flowrate{ RHDtostock _rhdtostock; //elementary function of the system //
// OK More Less No Waterflowrate waterflowrate;
[0.85,0.05, 0.05, 0.05] }; RHDtostock. Waterflowrate = waterflowrate;
}// DF- Main flow of the function- The flow rate is an attribute coming from the entity- Input_variable OrderHWD;
relationship diagram / States of the water rate flow coming from the HAZOP RHDtostock.Wantingtodo = OrderHWD;
Tank tank;
RHDtostock.Tank = tank;
// Definition of all output flows of the function “To stock water”//
Waterlevel waterlevel;
Waterlevel.source = RHDtostock;
RHDtoheat _rhdtoheat; //elementary function of the system //
// Definition of all input flows of the function “To heat water”//
RHDtoheat.Waterlevel = waterlevel;
Input_variable OrderHWD;
RHDtoheat.Wantingtodo = OrderHWD;
Heatingresistor heatingresistor;
RHDtoheat.Heatingresistor = heatingresistor;
// Definition of all output flows of the function “To heat water”//
Watertemperature watertemperature;
Watertemperature.source = RHDtoheat;
} AKNOWLEDGEMENT
The resulting network for our example is shown in Figure 12.
Figure 12: Bayesian network for the water heater process
In the resulting network it can be observe that the input
variables and the functions comes from the SADT model.
Also, it can be seen that the main flow is represented
according to its attributes in the entity-relationship diagram.
The deviation state of the flow attributes are defined in the
HAZOP study. In a similar way, it can be found the group of
possible component and functions states in the FMEA.
8. CONCLUSSION AND FURTHER WORK
The proposed model based on the functioning,
malfunctioning and informational studies is particularly
suitable to collect and to represent technical knowledge on
uncertain domains, but also enable to perform probabilistic
calculus and statistical analyses in an efficient and rigorous
manner. The difference of the PRM when comparing with
other classical methods is their capacity to deal with large
complex domains. Also this model allow to deal with issues
such as prediction or diagnostic optimization, data analysis of
feedback experience, deviation detection and model updating.
This methodology help to improve the automatic generation
of a model starting from SADT models, FMECA, HAZOP
analysis, fault trees, reliability block diagrams, 2TBN and
entity-relationship diagrams to evaluate global performance
of a technical system. The inference in this model is
performed through partial views of the global model which is
actually never built entirely as it is approached in SKOOB
project.
However, this methodology is a first step to gather technical
information in order to evaluate maintenance strategies. As
further work, some other factors should be incorporated in
the model such as human, organizational and environmental
factors.
It is also necessary to manipulate uncertainties within the
parameters and the knowledge of the model. The theory of
Dempster Shafer proposes an interesting formalism, and the
definition of the evidential networks developed by Simon et
al. (2009) are very interesting for decision making,
considering the imprecision on the utility calculus.
Finally, to validate the model by applying it to a real system
in order to show industrial feasibility and to confirm its added
value compared to the traditional computerised decision-
making systems.
The authors wish to express their gratitude to the French
National Research Agency ANR for the financial support of
the Structuring Knowledge with Object Oriented Bayesian
nets SKOOB project. Ref. ANR PROJET 07 TLOG 021
(http://skoob.lip6.fr). Special thanks is also paid to Paul
Munteanu and Frédéric Clerc from Bayesia and to Pierre-
Henri Wuillemin and Lionel Torti from LIP6 who are
developing the SKOOB language and the inference algorithm
for the PRM.
REFERENCES
Alsyouf, I. (2007) The role of maintenance in improving companies’ productivity
and profitability. International Journal of Production Economics, 105, 70–78.
Boudali H., Dugan J.B. (2005b). A discrete-time Bayesian network reliability
modeling and analysis framework. Reliability Engineering & System Safety.
87(3), 337-349.
Crespo-Márquez A, The maintenance management framework: models and
methods for complex systems maintenance, Springer Series in Reliability
Engineering (2008) ISBN-10:1846288207.
Jonsson P, (1999) Company-wide integration of strategic maintenance: an
empirical analysis. International Journal of Production Economics, 60-61:155–
164.
Kiker,G. A., Bridges, T. S., Varghese, A., Seager, T. P.,&Linkov, I. (2005).
Application of multicriteria decision analysis in environmental decision
making. Integrated Environmental Assessment and Management, 1, 95–108.
Koller D. and Pfeffer A. (1998), Probabilistic frame-based systems, Proceedings
of the 15th National Conference on Artificial Intelligence (AAAI), Madison,
Wisconsin..
Langseth H., Portinale L. (2007). Bayesian networks in reliability. Reliability
Engineering & System Safety. 92(1), 92-108.
Mahadevan S., Zhang R., Smith N. (2001). Bayesian networks for system
reliability reassessment. Structural Safety. 23(3), 231- 251.
McKone KE, Schroeder RG, Cua KO, (2001) The impact of Total Productive
Maintenance practices on manufacturing performance. Journal of Operations
Management, 19(1):39–58.
Muller A., Suhner M-C., Iung B. Formalisation of a new prognosis model for
supporting proactive maintenance implementation on industrial system.
Reliability Engineering & System Safety, In Press, Corrected Proof, Available
online 7 January 2007.
Pfeffer A., D. Koller, B. Milch and K.T. Takusagawa. SPOOK: A System for
Probabilistic Object-Oriented Knowledge Representation. Proceedings of the
14th Annual Conference on Uncertainty in AI (UAI), Stockholm, Sweden, July
1999.
Riis J., Luxhoj J. and Uffe T., A situational maintenance model, International
Journal of Quality and Reliability Management 14 (1997) (4), pp. 349–366.
Simon C., Weber P., Evsukoff A. (2008). Bayesian networks inference algorithm
to implement Dempster Shafer theory in reliability analysis. Reliability
Engineering and System Safety. Vol. 93, n° 7, p. 950-963.
Weber P., Jouffe L. Complex system reliability modelling with Dynamic Object
Oriented Bayesian Networks (DOOBN). Reliability Engineering and System
Safety, 91(2), February, 149-162, 2006.