A Correlation-Decoupled Interval Belief Rule Base for Interpretable Cross-Condition Bearing Fault Diagnosis

Cross-condition bearing fault diagnosis requires models that remain reliable under load-induced distribution shifts while providing transparent and traceable reasoning. Conventional belief rule bases (BRBs) may repeatedly use correlated vibration evidence during inference, and their Cartesian-product rule construction can rapidly increase rule-base complexity. This study proposes a correlation-decoupled interval belief rule base (CD-IBRB) for cross-condition bearing fault diagnosis. Seven diagnostically relevant time-domain features are selected using XGBoost and transformed into a less-correlated feature space through a Kendall-rank-correlation-guided matrix estimated exclusively from the source training data. Attribute-wise referential points and intervals are then constructed from the transformed training attributes, allowing the rule base to grow additively rather than combinatorially. Initial belief distributions are obtained from interval-level class distributions. The projection covariance matrix adaptation evolution strategy (P-CMA-ES) jointly optimizes the belief degrees, rule reliabilities, and rule weights, while evidential reasoning aggregates the activated interval rules to produce the final diagnostic result. In the primary cross-load bearing experiment, CD-IBRB achieved an accuracy of 0.9702 and a macro-averaged F1 score of 0.9703. It outperformed the strongest BRB variant and data-driven baseline by 7.70 and 6.10 percentage points in accuracy, respectively. Ablation experiments confirmed that removing parameter optimization or attribute decoupling reduced accuracy to 0.9053 and 0.9303, respectively. Additional cross-load and noise-injection experiments further demonstrated the stability of CD-IBRB under load shifts and input disturbances. Across five public multiclass datasets, CD-IBRB achieved a mean accuracy of 0.9004 and consistently outperformed the compared BRB variants. These results demonstrate that CD-IBRB provides a compact, uncertainty-aware, and traceable framework for cross-condition bearing fault diagnosis.