Bearing Cooperative Fault Diagnosis Based on Multi-adversarial Multi-domain Adaptive Network

The existing methods for diagnosing faults in bearings require a considerable number of labeled signals. However, practical engineering scenarios often involve a significant amount of unlabeled signals. To address this issue, this study introduces a novel approach called the Multi-Adversarial Multi-Domain Adaptive Network (MMAN). To capture the temporal correlation present in vibration signals, a shared feature extractor is devised using multi-head self-attention encoder. An adversarial module is devised to calculate the adversarial loss of multi-source domain. Additionally, a source-target domain adaptive subnet is formulated, utilizing the local maximum mean discrepancy (LMMD) loss function, to aid in class alignment between the source and target domains.. The paper concludes by presenting a collaborative decision-making strategy involving multiple subnets. This approach employs the LMMD loss to calculate confidence scores, aiding the multi-subnet classifiers in making well-informed fusion decisions. This contributes to the improved accuracy of collaborative fault diagnosis. The introduced method is assessed using bearing fault data gathered under variable speed conditions, and the outcomes from comparative experiments emphasize the excellence of the proposed method.