A novel multi-channel data-driven cross-domain bearing fault diagnosis method based on central domain maximum mean discrepancy and local cosine similarity

Abstract The safety and reliability of mechanical equipment are ensured by rolling bearings, which play a crucial role as vital elements of rotating machinery. However, the complex fault features of bearings cannot be fully characterized by single-channel data, and the feature distribution is significantly varied under varying operating conditions, leading to a substantial decline in model diagnostic performance. To address these issues, a multi-channel data-driven domain adaptation (DA) method based on central domain maximum mean discrepancy and local cosine similarity (MCLDA) is proposed. Firstly, a multi-channel composite indicator weighted fusion strategy (MCW) is applied to integrate frequency-domain signals from the X, Y, and Z channels of the acceleration sensor, offering more comprehensive fault insights. Then, a DA method based on central domain maximum mean discrepancy and local cosine similarity (CLDA) is developed to achieve both global feature and local feature alignment between the source domain (SD) and the target domain (TD), reducing global distribution differences and capturing fine-grained relationships within the same classes. As a result, the feature space consistency is optimized, enhancing the model's generalization and robustness in cross-domain fault diagnosis (CDFD). Finally, experimental verification is conducted, and it is shown that excellent diagnostic performance is exhibited by MCLDA in multiple cross-domain (CD) tasks, with an average accuracy exceeding 99%, significantly outperforming the comparative methods, demonstrating its effectiveness in rolling bearing CDFD.