Dynamic Bayesian Network Decision Model for Improving Fault Detection Procedure

In Model-Based Diagnosis (MBD) approaches, the decision-making generally relies on a binary fault signature matrix which is systematically generated from structural diagnosability analysis. However, this task becomes complicated when considering hybrid systems undergoing discrete modes shift and variation of states which may increase false alarms rate during the fault indicators (i.e. residuals) evaluation stage. This paper proposes a generic computer-aided diagnosis approach based on Dynamic Bayesian Network (DBN) in order to enhance robustness with regards to discrete mode changes. The Hybrid Bond Graph (HBG) Model is used as a multidisciplinary and integrated tool for dynamic modeling of all modes. The originality of the proposed approach relies on its ability to integrate statistical monitoring scheme based on cumulative sum (CUSUM) control chart using historical available data and qualitative reasoning mechanism based on fault indicators generated on the basis of HBG structural analysis. A synthetic case study is used to show the effectiveness of the developed DBN-based approach and its superior performance with regards to traditional thresholds based approaches.