An improved cross-machine transfer strategy based on multi-source domain knowledge for abnormal sample recognition

Cross-machine transfer has garnered significant attention owing to its capacity to transfer diagnostic knowledge across different machines and address unforeseen operating conditions. Nevertheless, the personalized sample biases arising from industrial-specific conditions reduce the generalization performance of traditional cross-machine diagnostic methods. To address this, an enhanced cross-machine transfer strategy with multi-source domain knowledge (CMMK) is proposed for bearing fault diagnosis. Specifically, targeted training of model parameters is conducted to address the task challenges encountered in cross-device diagnosis. Multiple sets of source domain data are introduced for collaborative training, effectively mitigating feature discrepancies between samples from different distributions. To address ambiguous fault samples at class boundaries, adversarial training between independent task classifiers is incorporated, enabling precise fault identification under consistent working conditions. Furthermore, we introduce the custom threshold module and propose a novel residual block structure, which makes each residual block generate its own adversarial mechanism. Note that as the training progresses, the network parameters gradually evolve in a direction that aligns with the requirements of cross-device diagnosis. Finally, comprehensive experiments on extensive bearing fault datasets validate the superior diagnostic accuracy and generalization ability of the proposed CMMK compared to state-of-the-art methods.