Learning-Enhanced Belief Rule Base for Sensor-Based Bearing Health Condition Assessment

The Belief Rule Base (BRB) provides explainable reasoning and uncertainty modeling capabilities, offering significant advantages in complex system modeling. However, the scope of its knowledge base is constrained by expert prior knowledge, rendering it incapable of encompassing fault categories that appear in the training data but lack expert-defined rules, thereby preventing post-construction expansion in line with data evolution. To this end, this paper proposes a learning-enhanced BRB model (LE-BRB), achieving a verifiable extension of the rule conclusion space within the BRB framework. First, association rules are introduced to mine classes with missing expert rules and generate candidate rules; secondly, a quality verification module is constructed through Pareto-based filtering, and rule fusion is implemented based on antecedent similarity and conclusion conflict degree; Finally, evidential reasoning is employed to achieve traceable diagnosis. Experiments on the CWRU local bearing fault dataset show that, while preserving the original interpretability of the BRB, LE-BRB completes the missing rolling element fault (Ball) rules in the knowledge base, achieving an average accuracy of 91.02%. Experiments on the MAFAULDA bearing dataset show that, while preserving the original interpretability of the BRB, LE-BRB completes the missing rotor unbalance fault in the knowledge base. This fills the coverage gap of the original BRB for this fault category, improving the diagnostic accuracy from 43.69% to 94.17%. Compared with models such as neural networks and decision trees, it achieves comparable or superior accuracy while exhibiting relatively higher stability.