Data based construction of Bayesian network for fault diagnosis of event-driven systems

This paper presents a decentralized fault diagnosis strategy of event-driven systems based on probabilistic inference and a method to construct the inference network, Bayesian network (BN), structure. First of all, the controlled plant is decomposed into some subsystems, and the global diagnosis is formulated using the Bayesian Network, which represents the causal relationship between the fault and observation in subsystems. Second, the local diagnoser is developed using the conventional Timed Markov Model (TMM), and the local diagnosis results are used to specify the conditional probability assigned to each arc in the BN. The structure of BN is essential since the computational burden and the fault diagnosis performance greatly depend on it. Accordingly, we propose a data based construction strategy of BN for fault diagnosis of event-driven systems. Finally, the usefulness of the proposed strategy is verified through some experimental results of an automatic transfer line.