Hybrid Attention-Driven Convolutional Adversarial Network for Cross-Domain Bearing Fault Diagnosis Using Acoustic Signals

Acoustic features have gained widespread popularity in machinery fault diagnosis owing to their benefits of non-contact measurements and low cost. However, most acoustic diagnostic methods focus only on specific operating conditions and show insufficient feature extraction capability when facing domain shift issues in complex scenarios. To address these problems, this paper proposes a novel hybrid attention-driven convolutional adversarial network (HA-CAN) for cross-domain fault diagnosis. First, HA-CAN employs dense convolution to enhance feature propagation, ensuring efficient information flow. Then, HA-CAN combines dual convolution attention (DCA) and squeeze-and-excitation (SE) modules to achieve global recalibration of spatial and channel features, while minimizing inter-domain differences using Wasserstein distance. Finally, a domain adversarial strategy is designed for training the features for classification, and two sets of bearing acoustic signal data are used for validating the proposed HA-CAN model. The results demonstrate that HA-CAN achieves an average diagnostic accuracy exceeding 97% in both groups of experiments, outperforming other methods in classification accuracy and fault feature extraction capability. Meanwhile, it also exhibits strong noise robustness under noisy conditions.