بررسی اجمالی برنامه های کاربردی شبکه بیزی برای وابستگی؛ تجزیه و تحلیل ریسک و مناطق تعمیر و نگهداری
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|29172||2012||12 صفحه PDF||سفارش دهید||9219 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Engineering Applications of Artificial Intelligence, Volume 25, Issue 4, June 2012, Pages 671–682
In this paper, a bibliographical review over the last decade is presented on the application of Bayesian networks to dependability, risk analysis and maintenance. It is shown an increasing trend of the literature related to these domains. This trend is due to the benefits that Bayesian networks provide in contrast with other classical methods of dependability analysis such as Markov Chains, Fault Trees and Petri Nets. Some of these benefits are the capability to model complex systems, to make predictions as well as diagnostics, to compute exactly the occurrence probability of an event, to update the calculations according to evidences, to represent multi-modal variables and to help modeling user-friendly by a graphical and compact approach. This review is based on an extraction of 200 specific references in dependability, risk analysis and maintenance applications among a database with 7000 Bayesian network references. The most representatives are presented, then discussed and some perspectives of work are provided.
The management of complex industrial systems contributes to higher competitiveness and higher performances at lower costs. In that way, the relevance of the maintenance and dependability analyses increased due to their role in improving availability, performance efficiency, products quality, on-time delivery, environment and safety requirements, and total plant cost effectiveness at high levels (Alsyouf, 2007 and Kutucuoglu et al., 2001). Nowadays, one of the major problems in the dependability field is addressing the system modeling in relation to the increasing of its complexity. This modeling task underlines issues concerning the quantification of the model parameters and the representation, propagation and quantification of the uncertainty in the system behavior (Zio, 2009). In previous years, the reliability and risk analysis of systems were studied by making assumptions simplifying the study. One of these assumptions is to focus the study only on the technical part of the system. This assumption is no longer valid, since it has been shown the importance of organizational and human factors contributions (Leveson et al., 2009). Indeed, if studies were centered on technical aspects of systems until seventies (Villemeur, 1992), several major accidents, such as the Three Miles Island nuclear accident and the Bhopal catastrophe have pointed out cause operator errors and organizational malfunctions. These accidents allowed the scientific community to present and develop, in eighties, first methods centered on the analysis of these human errors. It led to the expansion of the human reliability analysis (HRA). But other accidents (Challenger explosion, Chernobyl nuclear accident …) have emphasized, in nineties, the importance of organizational malfunctions in their occurrences and, have contributed to the emergence of different theories for the study of these organizational issues: normal accident (Perrow, 1990 and Weick et al., 2001 2001) and high reliability organizations (Robert, 1990, Léger et al., 2008 and Léger et al., 2009). As a consequence, innovative studies aim at covering the whole of these causes (technical, human and organizational). Nevertheless, such analyses are often difficult to achieve because they require a lot of resources. This matter adds complexity to the systems’ modeling due to the interaction between different technical, human, organizational and nowadays environmental factors which are necessary to quantify failure scenarios and risky situations. Thus, the challenge is to formalize a model of a complex system integrating all these aspects (Trucco et al., 2008 and Kim et al., 2006) (Fig. 1). Full-size image (28 K) Fig. 1. Context of the complex system to be modeled. Figure options Furthermore, while modeling these factors, it is required to take into account the knowledge integration of diverse natures such as qualitative and quantitative with several abstraction levels. The organization and human analyses are more naturally modeled with a qualitative knowledge (to describe situations, scenarios…) such as knowledge represented in failure mode, effects, and criticality analysis (FMECA), HAZard OPerability (HAZOP), probabilistic risk assessment (PRA) analysis, etc.; and in other hand, the technical level is usually known with quantitative information (failure rates, unavailability level, Mean Time To Failure (MTTF), etc.) (Røed et al., 2008). A complementary point of view to be modeled for the system is the temporal dimension (system dynamics) which consists in describing phenomenon such as: sequences in scenarios, degradations of components, evolution of symptoms corresponding to deterioration mechanisms, impact of preventive maintenance actions on the degradation, influence of environmental conditions and effects of the operation conditions on the evolution of the component states. Once assessed the failure probability and risk associated to a system situation, the information is provided to support the decision making process. It implies to quantify the uncertainty and imprecision on parameters, for example, the uncertainty of the failure occurrence and its consequences (Zio, 2009). Therefore, the main characteristics to be modeled in a system for assessing dependability and maintenance aspects are: • the complexity and size of the system (large-scale systems) (Zio, 2009), • the temporal aspects (Labeau et al., 2000), • the integration of qualitative information with quantitative knowledge on different abstraction levels (Papazoglou et al., 2003 and Delmotte, 2003), • the nature of multi-state components (Griffith, 1980), • the dependences between events such as failures (Torres-Toledano and Sucar, 1998), • uncertainties on the parameter estimation (Zio, 2009). For modeling these requirements, there are some classical dependability methods such as fault trees, Markov chains, dynamic fault trees, Petri nets and Bayesian networks (BN). In the recent literature, it is observed a growing interest focused on BN. This modeling method is not the solution to all problems, but it seems to be very relevant in the context of complex systems (Langseth, 2008). Indeed some papers such as Mahadevan et al. (2001), Boudali and Dugan (2005b), Langseth and Portinale (2007) and Langseth (2008) show the increasing interest on the use of BN to estimate and to improve reliability and safety of systems over the last decade. For example, during the period 1999–2009, RESS journal (Reliability Engineering and System Safety), well known in dependability area, shows an increment of 100% of a ratio consisting on the paper number dedicated to the application of BN to reliability (or risk) divided by the total amount of papers. This type of ratio has strengthened our interest to analyze the evolution of the literature about BN and their applications on dependability, risk analysis and maintenance. For this purpose, we have built a database of references from 1990 to 2008 with different bibliographical research tools (i.e. google scholar, Sciencedirect, Web of Knowledge …). In this paper, the most relevant articles according to their citation number were referenced until 2008. Nonetheless, some citations on “hot topics” of research until 2009 are also given. The rest of this paper is organized as follow. Section 2 is introducing the bases of BN and explaining why they are suitable to model complex systems. Section 3 shows a bibliographical review of the relevant research directions for modeling dependability, risk analysis and maintenance problems with BN. Section 4 presents a comparison of the BN modeling capabilities with other modeling methods such as Fault Tree, Markov Chains and Petri Nets. Finally, the conclusions are given by integrating also highlights future research directions.
نتیجه گیری انگلیسی
The research works and applications of Bayesian Networks in risk analysis, dependability and maintenance have shown a significant upward trend since 2000, especially in dependability. Recently, there have been about 30 articles per year and, an increase of 800% of publications between 2000 and 2008. BN in reliability, risk and maintenance areas are chosen since they are easy to use with domain experts. BN are particularly suitable for collecting and representing knowledge on uncertain domains but also enable to perform probabilistic calculus and statistical analyses in an efficient manner. The difference of BN, in comparison with other classical methods, is their polyvalence. They allow dealing with issues such as prediction or diagnosis, optimization, data analysis of feedback experience, deviation detection and model updating. The graphical representation is interesting since the model complexity is understandable in a single view. In the case of large size model, object oriented representation OOBN or probabilistic relational descriptions (PRM) provide manageable models. One of the weak points of BN is that there is no specific semantic to guide the model development and to guarantee the model coherence. Therefore, a relevant issue is the use of tools for the formalization of BN models in order to integrate various dimensions (technical, organization, information, decision and finance) correlated with system’s behavior in reliability, risk analysis and maintenance fields (Øien, 2001, Kim et al., 2006 and Trucco et al., 2008). For solving this issue, the research can follow two directions: The first one concerns the translation of the classical dependability model into a BN model. The second one is to define new methodologies of model development. The first solution leads to a coherent model but is limited by the conditions and hypotheses related to the classical dependability model translated in BN. In opposition, the second approach is more innovative because it leads to a model exploiting all the flexibility of BN formalism but it is difficult to prove the result consistence by comparison with other methods classically based on restrictive hypotheses. In addition, since there is no specific semantic to build a BN, it is necessary to verify the models and to validate them in accordance with the system reality. One aspect to be developed is formalizing some methods for the sensibility analysis of a model in order to investigate its robustness according to the problem studied (Pollino et al., 2007). When exploiting a DBN model, there are several inference algorithms that are appropriated to different situations. For example, with the exact inference algorithm proposed by Jensen (1996), the 2TBN model is similar to a Markovian model with dynamic independent variables. It means that when calculating variables at step (i+1), the past before step (i) is forgotten thanks to the Markov property. Thus, the inference using junction tree computes the exact distribution if the variable of the dynamic processes respect the Markov property and no dependency exists between the processes. In this particular case, the results are only exactly the same as the computation in the unroll-up BN model. In that sense, one of the research directions is to guide the use of BN taking into account the limitations of the current inference algorithms in order to warn the community on the possible erroneous use in the models with the temporal aspect. For these representations, several inference algorithms exist and are still in development. Their efficiency depends on the model complexity ( Murphy, 2002). In the dependability analysis there are different phenomena of diverse natures that should be considered i.e. discrete and continuous variables. For this reason a lot of work has been developed in this area in order to integrate continuous variables in BN models. As a result, a significant part of the community is directing its efforts on the development of inference algorithms for hybrid BN (Boudali and Dugan, 2006, Neil et al., 2009 and Langseth et al., 2009). An interesting issue would be to deal with large systems (several hundred variables) in order to formalize complex models. For example, the SKOOB project is developing a generic model based on PRM (Getoor et al., 2007), which enables a better understanding of complexity and the reutilization of generic parts of a model to represent systems. The network is not defined by a graph but in a language. The inference is performed through partial views of the global model which is actually never built entirely as it is approached in SKOOB project (SKOOB, 2008). Another interesting issue is the manipulation of the imprecision within the parameters and the knowledge of the model (uncertainty). The theory of Dempster Shafer proposes a relevant formalism, and the definition of evidential networks developed by Simon and Weber, 2009a and Simon and Weber, 2009b are suitable for decision making, considering the imprecision on the utility computation. As a final point, BN are limited by the modeling aspects that they can deal with. Thus, it is necessary to make BN interoperable with other dependability/risk tools in order to complement the capabilities of BN to better represent the characteristics of a system.