Partial Domain Bearing Fault Diagnosis Method Based on Weighted Bi-Classifier

The investigation of bearing fault diagnosis is a crucial and pressing core technology in the advancement of rotating mechanical equipment, as well as a significant national strategic requirement. In the realm of fault diagnosis, domain adaptation techniques have been extensively employed to effectively address the challenge of aligning source and target domains with dissimilar data distributions. Nevertheless, prevailing approaches typically presume that the source and target domains possess identical health conditions. However, in industrial applications, the target domain frequently constitutes a subset of the source domain. In light of this issue, this research paper presents a novel unsupervised partial domain fault diagnosis approach. Firstly, it incorporates a hybrid convolutional block, which is a lightweight network structure, to enhance the receptive field through wide kernel convolution and extract global information from various states. Secondly, it employs a dual classifier structure and incorporates an adaptive source domain class weighting method to dynamically select the most pertinent classes for target domain prediction. Additionally, an enhanced local maximum mean square discrepancy is utilized: A proposed approach is presented to attain alignment, and the experimental analysis of multiple datasets under varying working conditions demonstrates that this method exhibits a high level of recognition accuracy when dealing with asymmetry within certain domain classes.