A dual-weight mechanism-based neural network for partial domain adaptation fault diagnosis of bearings under different working conditions

Abstract

The intelligent fault diagnosis of bearings often entails a distribution discrepancy between training data and real data, which significantly limits the generalization ability of fault diagnosis models. Traditional domain adaptation (DA) methods aim to achieve knowledge transfer from the source domain to the target domain. However, in reality, the incompleteness of the target domain dataset leads to the issue of missing classes in the target domain, rendering the assumption of identical label categories in traditional partial domain adaptation (PDA) methods difficult to implement. This study proposes a dual-weighted mechanism for fault diagnosis models in the context of PDA (PDA method-based dual-weight mechanism) to address the issue of weight bias in target domain labels due to category absence. This approach employs a sample-based weighting mechanism and a similarity-based weighting mechanism to assign weights to different label categories, referred to as the dual-weight mechanism based on similarity and sample number (WSS). This method can promote positive transfer while suppressing negative transfer, thereby resolving the issue of weight bias in the target domain label space. Experiments that use the WSS for fault transfer tasks under identical conditions, with multiple categories and multiple conditions employed for the same category, demonstrate the superior performance of the WSS.