Académique Documents
Professionnel Documents
Culture Documents
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 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
The importance of the maintenance function has increased, improve the understanding of the underlying dimensions of
due to its role in improving availability, performance maintenance (Crespo, 2007). Maintenance is somewhat
efficiency, products quality, on-time deliveries, environment “under-developed” (Jonsson, 1999; McKone et al., 2001)
and safety requirements, and total plant cost effectiveness at with a lack of effective prevention methodologies and the
high levels (Riis et al., 1997; Alsyouf, I. 2007). In order to integration of standard methods in manufacturing companies
improve business performances, it is required to analyze in most continents.
aspects such as maintenance, risk analysis and dependability
which allow to forecast the gaps between nominal and non- In that sense, the challenge is to formalize a generic model of
nominal operation of the system (degradation, failure, etc). complex systems to evaluate maintenance strategies that
allows to evaluate the global performance integrating various
Some papers such as Mahadevan (2001), Boudali and Dugan dimensions such as technical, organizational, human,
(2005b) Langseth and Portinale (2007) show the increasing informational, decisional and financial dimensions correlated
interest on the use of Bayesian Networks to estimate and to with system's behavior and nowadays environmental factors
improve reliability and safety of systems over the last decade. which are necessary to quantify failure scenarios and risky
Nevertheless, this method is limited by the number of situations. As a consequence, increasingly studies aim at
variables in a model. covering the whole of these factors. Nevertheless such
analyses are often difficult to achieve because of the
For this reason, this paper formalizes a methodology to importance of required resources. This matter adds
define an extension of Object Oriented Bayesian Networks, complexity to the systems’ modelling due to the interaction
the Probabilistic Relational Model to model technical between the different kinds of knowledge.
knowledge in order to help decision-making for risk,
maintenance and dependability domains. The network is not
Furthermore, while modelling these factors, it is required to
defined by a graph but in a language. The proposed
take into account the knowledge integration of diverse nature
methodology has originality on formalising, by means of
such as qualitative (organizational and human analyses) and
PRM, the models from prior knowledge on the system
quantitative (technical analyses) knowledge with several
functioning, malfunctioning and informational point of view
abstraction levels (Muller, 2007).
to estimate the overall performance of the process (reliability
and compliance of output flows).
This paper, explains the main concepts of an approach to
2. PROBLEM STATEMENT formalize a model required to evaluate risk analysis,
maintenance and dependability. The idea of this
There are several works in the maintenance field especially in
formalization is to unified multiple and different kind of
the optimization of maintenance strategies, nonetheless this is
knowledge in one model.
a complex issue since it is necessary to take into account
different conflicting criterion such as availability, safety and
In the technical aspect the object is to be able to model the
costs, etc., (Kiker et al., 2005). As described by Crespo, 2007
impact on system global performances based on functional,
there is a wide diversification in the maintenance problems.
dysfunctional and informational analyses. The idea is to
Maintenance is composed of a set of activities for which it is
formalize the interactions between an industrial system and
difficult to find procedures and information support systems
the maintenance system using treatment and data models objects. The Object-Oriented Bayesian Networks (OOBNs)
such as SADT, AMDEC, HAZOP, etc. (Figure 1). 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.
To create a generic, complex and big size model to evaluate According to the standard AFNOR X 50-151: “the functional
risk, maintenance and dependability it is essential to use a analysis is an approach that consists in searching, ordering,
compact representation of knowledge. To make it possible, characterizing, organizing into hierarchies and/or to develop
the knowledge must be divided into groups of “local the functions of a product (system) expected by the user”.
knowledge”, that could be added according to the needs in
order to form a whole coherent model. The functional modeling of an industrial system consists in
formalizing, by qualitative causal relationships, the
Concerning our problematic, bayesian networks (BN) appear interactions between the functions performed by each of the
to be a solution to model complex systems. Nonetheless, one sub-systems until the component level (elementary
of the weak points is that BN are not adequate for dealing functions), so it is possible to use a Structure Analysis and
with very large complex systems. Such systems require an Design Technique (SADT) as a functional approach method.
explicit representation of additional types of structure: the
notion of an object, a complex structured domain entity with The system functioning modeling is based on the principle of
its own properties; and the notion of a class of objects, that activity and sub-activities until elementary activities,
captures properties common to an entire set of similar
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).
To integrate the different kind of knowledge in big and Conditional probability table (CPT) declaration
complex models, the SKOOB project is developing a In this case, the declaration order is essential. The first value
language to represent PRM models. 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],
6. STRUCTURE OF A PROJECT IN THE CURRENT P_2[1], …, P_N[2]), and so on. The A[i] values correspond
SKOOB LANGUAGE 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.
This language is inspired in Java language (specially the typeEtat perform_a_function dependson
declarative part of the classes definitions), because it is one of exists(inputs.etat,OK), etat {
the more common languages used in the present. The [1.0, 0.0, 0.0, 0.0, // P(OK | true,
hierarchy of packages is that of the files of the project, and OK), P(OK | true, NOK), P(OK | false,
each file includes the definition of a compilation unit. The OK), P(OK | false, NOK)
current characteristics of SKOOB language are described as 0.0, 1.0, 1.0, 1.0] // P(NOK | true,
followed: OK), P(NOK | true, OK), P(NOK | false,
OK), P(NOK | false, NOK)
Variable declaration };
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
Figure 5: Water heater process component are defined as well as their effects. The causes are
linked with the component states or the unavailability of the
Functioning system representation: SADT model electric energy required to supply the component. Therefore,
the loss of energy is a common cause of the six failure
Figure 6 presents the level A0 of the system SADT analysis. modes.
This figure depicts the interaction between the system and the
external environment through the AD, HD and RHD flows.
The resulting network for our example is shown in Figure 12. 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.
Figure 12: Bayesian network for the water heater process
REFERENCES
In the resulting network it can be observe that the input Alsyouf, I. (2007) The role of maintenance in improving companies’ productivity
and profitability. International Journal of Production Economics, 105, 70–78.
variables and the functions comes from the SADT model. Boudali H., Dugan J.B. (2005b). A discrete-time Bayesian network reliability
Also, it can be seen that the main flow is represented modeling and analysis framework. Reliability Engineering & System Safety.
according to its attributes in the entity-relationship diagram. 87(3), 337-349.
The deviation state of the flow attributes are defined in the Crespo-Márquez A, The maintenance management framework: models and
methods for complex systems maintenance, Springer Series in Reliability
HAZOP study. In a similar way, it can be found the group of Engineering (2008) ISBN-10:1846288207.
possible component and functions states in the FMEA. Jonsson P, (1999) Company-wide integration of strategic maintenance: an
empirical analysis. International Journal of Production Economics, 60-61:155–
164.
8. CONCLUSSION AND FURTHER WORK Kiker,G. A., Bridges, T. S., Varghese, A., Seager, T. P.,&Linkov, I. (2005).
The proposed model based on the functioning, Application of multicriteria decision analysis in environmental decision
malfunctioning and informational studies is particularly making. Integrated Environmental Assessment and Management, 1, 95–108.
Koller D. and Pfeffer A. (1998), Probabilistic frame-based systems, Proceedings
suitable to collect and to represent technical knowledge on of the 15th National Conference on Artificial Intelligence (AAAI), Madison,
uncertain domains, but also enable to perform probabilistic Wisconsin..
Langseth H., Portinale L. (2007). Bayesian networks in reliability. Reliability
calculus and statistical analyses in an efficient and rigorous Engineering & System Safety. 92(1), 92-108.
manner. The difference of the PRM when comparing with Mahadevan S., Zhang R., Smith N. (2001). Bayesian networks for system
other classical methods is their capacity to deal with large reliability reassessment. Structural Safety. 23(3), 231- 251.
complex domains. Also this model allow to deal with issues McKone KE, Schroeder RG, Cua KO, (2001) The impact of Total Productive
Maintenance practices on manufacturing performance. Journal of Operations
such as prediction or diagnostic optimization, data analysis of Management, 19(1):39–58.
feedback experience, deviation detection and model updating. 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
This methodology help to improve the automatic generation online 7 January 2007.
of a model starting from SADT models, FMECA, HAZOP Pfeffer A., D. Koller, B. Milch and K.T. Takusagawa. SPOOK: A System for
analysis, fault trees, reliability block diagrams, 2TBN and Probabilistic Object-Oriented Knowledge Representation. Proceedings of the
14th Annual Conference on Uncertainty in AI (UAI), Stockholm, Sweden, July
entity-relationship diagrams to evaluate global performance 1999.
of a technical system. The inference in this model is Riis J., Luxhoj J. and Uffe T., A situational maintenance model, International
Journal of Quality and Reliability Management 14 (1997) (4), pp. 349–366.
performed through partial views of the global model which is Simon C., Weber P., Evsukoff A. (2008). Bayesian networks inference algorithm
actually never built entirely as it is approached in SKOOB to implement Dempster Shafer theory in reliability analysis. Reliability
project. 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
However, this methodology is a first step to gather technical Safety, 91(2), February, 149-162, 2006.
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.