BSLM-CDAN: Batch-Norm-Stabilized and Locality-Weighted Conditional Adaptation for Cross-Condition Fault Diagnosis

Variable operating conditions induce covariate and conditional shifts that degrade bearing fault diagnosis. To stabilize conditional-marginal alignment within a CDAN backbone, an enhanced variant, BSLM-CDAN, which integrates a logit-level BN-stabilized classifier and a locality-weighted MK-MMD into CDAN, is introduced. The method integrates a logit-level Batch Normalization (BN)-regularized classifier to suppress gradient spikes, a locality-weighted multi-kernel maximum mean discrepancy (MK-MMD) to emphasize in-cluster target samples and mitigate outliers and pseudo-label noise, and adaptive loss scheduling that balances adversarial (conditional) and discrepancy (marginal) objectives. On CWRU cross-load transfers, BSLM-CDAN consistently outperforms DANN, ACDANN, and CDANN-JMMD; ablation results confirm the necessity of BN stabilization and locality weighting, and t-SNE visualizations display tighter, better-aligned class clusters. These findings indicate a practical, stability-oriented upgrade that makes CDAN-style adaptation more reliable for diagnosis under variable conditions.