A cross-condition fault diagnosis method for bearings based on feature enhancement and semantic information constraints

Abstract Unsupervised domain adaptation has garnered widespread attention in the field of rolling bearings fault diagnosis (FD), thanks to its remarkable performance in transfer-based diagnostic tasks. However, in a noisy environment, the existing domain adaptive networks are difficult to accurately mine domain invariant features and effectively align marginal distribution and conditional distribution, which lead to the problem of low diagnostic accuracy in the existing models. In response to the above challenges, a cross-condition FD method for bearings based on feature enhancement and semantic information constraints (CC-FESIC) is proposed. Firstly, a networked blind deconvolution module is designed based on the time-domain autoencoder and frequency-domain filter, which enhances the expression of domain-invariant features in the fault signal domain of rolling bearings. Secondly, the fusion of domain-invariant features is accomplished through a multi-scale feature extractor. Again, a joint distribution framework with SIC (JDF-SIC) is constructed, the framework realizes the JD alignment of features respectively through multi-kernel maximum mean discrepancy (MK-MMD) and the improved MK-MMD (SMK-MMD). In addition, the JD alignment is constrained through a SIC module. Finally, extensive cross-condition diagnostic experiments are carried out on two public bearing datasets to verify the effectiveness of the proposed method.