A Novel Joint Class Subdomain Adaptive Fault Diagnosis Algorithm for Bearing Faults Under Variable Operating Conditions

Aiming at the poor transferability of bearing fault diagnosis models under variable operating conditions and the difficulty in acquiring labeled defective samples. In this article, a novel joint class subdomain adaptive network (JCSDAN) is proposed by fusing deep convolution and attention mechanisms for bearing fault diagnosis under variable working conditions. First, in order to capture the position information of periodic impulses in fault signal sequence, a relative attention convolution module (RACov) is proposed by introducing relative position coding. Then, to enhance the ability of the model to capture global fault features, the local features of fault periodic impulse are extracted by using a multipath convolution module (MPCov) and input to the designed RACov. Finally, a new domain adaptive (DA) loss function used to improve the ability to extract domain-invariant features is constructed by simultaneously considering the difference of fault features in feature space and probability space. In order to verify the performance of the proposed method, experiments of the variable conditions transfer tasks are conducted on two publicly available bearing datasets. Experimental results demonstrate that the proposed algorithm has higher accuracy and better generalization than the comparison algorithms.