A new compact belief rule model for fault diagnosis

Abstract

Amidst growing scrutiny of “black box” models in fault diagnosis applications that demand high interpretability, rule-based models are regaining prominence due to their inherent interpretability. Belief rule-based (BRB) models with belief distributions stand out in fault diagnosis. Nevertheless, the incorporation of multiple fault features introduces a substantial challenge for the BRB modeling, due to the phenomenon of combinatorial explosion, which leads to an exponential escalation in the number of rules required. To tackle the root cause of the exponential growth of rules in the BRB model, a new compact belief rule model (CBRM) for fault diagnosis is proposed. A flexible rule antecedent that leverages the similarity of reference values, eschewing the Cartesian product structure, is proposed to effectively diminish the complexity and size of the fault diagnosis model. Consequently, a rule activation method leveraging the hash algorithm has been proposed, wherein the spatial relationship between a sample and a rule is translated into binary code. The rule closest to the sample in each space is activated, and its corresponding activation weight is calculated combining the matching degree, fault features weights and rules weights. Subsequently, the reasoning and optimization of the CBRM are conducted to grab the diagnostic results and update parameters. Eventually, the fault diagnosis for relays is utilized to validate the proposed method.