Model-based fault diagnosis using artificial intelligence techniques often deals with uncertain knowledge and incomplete information. Probability reasoning is a method to deal with uncertain or incomplete information, and Bayesian network is a tool that brings it into the real world application. A novel approach for constructing the Bayesian network structure on the basis of a bond graph model is proposed. Specification of prior and conditional probability distributions (CPDs) for the Bayesian network can be completed by expert knowledge and learning from historical data. The resulting Bayesian network is then applied for diagnosing faulty components from physical systems. The performance of the proposed fault diagnosis scheme based on bond graph derived Bayesian network is demonstrated through simulation studies.
The growing demand for safety and reliability of modern engineering systems motivate the development of robust fault diagnosis algorithm. Early approaches, such as, fault tree analysis, expert systems, to fault diagnosis are inherently rule-based. They are proved to be inflexible, incomplete, and required comprehensive a prior knowledge of the fault characteristics, rather than actually deducing the fault themselves. Most advanced fault diagnosis algorithms now concern of using model that is derived from system's structure and behavior in order to establish the cause of system malfunction. This model-based fault diagnosis [1], [8] and [11] enables more complex cause–effect reasoning and hence a more robust diagnostic system can be developed.
A number of different model-based fault diagnosis algorithms have been proposed in the past decades, capable of dealing with different diagnostic problems. Quantitative and qualitative are the two major approaches to model-based fault diagnosis. In quantitative fault diagnosis, precise mathematical model is used to monitor system states, detect abnormal behaviors and diagnose the failures. The main problems with such methodologies are the intricacy and overheads of obtaining precise numerical models and the sensitivity of the diagnostic system to modeling error. Usually, the effects of modeling errors obscure the effects of faults and cause false alarms [7] and [11]. Qualitative fault diagnosis which dominates in the AI community, without the use of precise numerical model and capable of dealing incomplete information, alleviate some problems encountered by quantitative approach. However, the lack of precision in the representation, and ambiguities introduced during the inference process, limit the application of the qualitative approach to complex systems [3].
Fault diagnosis based on AI techniques often deals with uncertain knowledge and incomplete input data. Probability reasoning is a method to deal with uncertain information, and Bayesian network is a tool that brings it into the real world applications [12], [13] and [14]. In this paper, we proposed an alternative approach to model-based fault diagnosis, where Bayesian network is adopted to model the system and diagnose faulty system components. Bayesian network is a directed, acyclic graph (DAG), which embeds cause–effect relationship between variables (nodes). The representation framework of Bayesian network allows reasoning under uncertainty. Component failure probability of a system is computed by sequential evidence-propagation inference among conditional probability distributions that have been specified at each variable (node) [2].
The goal of the model-based fault diagnosis is to detect and localize faulty components in a system. Hence, the model used should incorporate structural information about the system and bond graph is such a representation. In this paper, a general procedure for constructing a Bayesian network structure on the basis of a bond graph model is proposed. Some researchers have proposed to learn the Bayesian network structure from data [4] and [9]. However, the accuracy of the learned Bayesian network is largely affected by the ‘richness’ of the data and the prior knowledge of the network ordering. There are several advantages of using bond graph model as the skeleton to construct the Bayesian network for fault diagnosis. The task of identifying system variables to construct Bayesian network is completed and the localization of faulty components from Bayesian network is enhanced since they are already represented in the bond graph model. Bayesian network based fault diagnosis contributes to the possibility of ranking possible failures, handling multiple simultaneous failures and uncertainty symptoms of certain faults.
The paper is organized as follows. In Section 2, fundamental knowledge of Bayesian network is provided. Section 3 describes the construction of Bayesian network on the basis of bond graph model. In Section 4, fault diagnostic scheme based on Bayesian network and its results are presented. Finally, the paper is concluded in Section 5.
This contribution presents a novel approach on constructing a Bayesian network from bond graph model. Information and hypothesis variables are first identified and the causal links between variables are generated from the qualitative interpretation through the set of qualitative equations. Specification of prior and CPDs can be completed by expert knowledge and learning from historical data. Simulation studies on the single tank and coupled-tank systems show that the proposed fault diagnosis based on Bayesian network is feasible. Faulty components can be localized correctly without extensive computation which is a major criterion of on-line diagnosis.
