Stepwise feature norm network with adaptive weighting for open set cross-domain intelligent fault diagnosis of bearings

Abstract

Cross-domain fault diagnosis of bearings has attracted significant attention. However, traditional cross-domain diagnostic methods have the following shortcomings: (1) when the trained model is applied to a new scenario, it leads to severe degradation of the model and a reduction in its generalisation ability. (2) The accuracy of the open-set fault diagnosis is affected by additional faults in the target domain data. To overcome these shortcomings, a stepwise feature norm network with adaptive weighting (SFNAW) is proposed for cross-domain open-set fault diagnosis. In SFNAW, two weight extractors are designed to adaptively calculate the sample weights such that a threshold can be set to mark the additional fault samples of the target domain as unknown faults using these weights. Transferable features are obtained by adaptively increasing the feature norm stepwise to alleviate model degradation and align the source and target domains. Finally, the fault diagnosis knowledge of the source domain is transferred to fault recognition in the target domain. The proposed SFNAW method was verified using two bearing datasets. The results show that the SFNAW can effectively detect additional faults in the target domain and reduce model degradation, thereby improving the fault diagnosis accuracy. Meanwhile, the SFNAW method has a higher accuracy than other traditional methods.