Cross-Condition Bearing Fault Diagnosis via DANN Enhanced with MSCNN and Minimum Class Confusion

To address performance degradation caused by distribution shifts in vibration signals across operating conditions, we enhance the Domain-Adversarial Neural Network (DANN) with a lightweight multi-scale 1D convolutional feature extractor (MSCNN1D) and a Minimum Class Confusion (MCC) regularizer. MSCNN1D uses three parallel branches (kernel sizes 3/5/7) followed by 1 × 1 fusion, capturing short-and mid-range temporal patterns at low computational cost. Complementary to adversarial alignment, MCC is applied only to target-domain predictions and minimizes inter-class correlation via class-wise normalization, thereby reducing class mixing and encouraging clearer decision boundaries. On cross-condition transfers of the CWRU dataset, the proposed method improves target-domain accuracy over baseline DANN under the reported protocol while preserving real-time feasibility. Qualitative analyses (t-SNE, confusion matrices, and class-correlation heatmaps) indicate tighter and better-separated target-domain clusters. These results suggest that coupling multi-scale temporal representation with alignment-and-discriminability constraints yields robust cross-condition generalization under limited labels and varying operating settings.