Multi-Structure Fusion Conditional Domain Adaptation for Bearing Fault Diagnosis

To address the limited feature extraction capability of single-architecture deep models, pronounced distribution discrepancies in bearing vibration signals, and instability during cross-domain training under varying operating conditions, a multi-structure fusion conditional domain adaptation (MSFCDA) method is proposed for bearing fault diagnosis. First, a multi-structure feature extractor is constructed by integrating a multi-channel multi-scale convolutional neural network, a bidirectional long short-term memory network, and an efficient channel attention mechanism to comprehensively extract multi-scale features from vibration signals. On this basis, a class-conditional feature distribution alignment strategy is developed, in which class prediction information is coupled with the cross-domain alignment process to achieve feature alignment under class-conditional constraints. In addition, a dynamic adversarial weighting factor is introduced to adaptively regulate the strength of adversarial training, thereby balancing feature discriminability and domain invariance. Extensive cross-condition transfer experiments are conducted on the publicly available CWRU and PU bearing datasets. The results demonstrate that the proposed method consistently achieves high diagnostic accuracy under different operating conditions, validating its strong cross-domain transfer capability and training stability under variable operating conditions.