A fault mechanism-based model for bearing fault diagnosis under non-stationary conditions without target condition samples

Abstract

Fault diagnostic technique with high adaptability to industrial environments is important to engineering. Based on the assumption that samples from the training set obey the identical distribution as signals from the industrial equipment, deep learning-based methods achieved high diagnostic accuracy. However, the assumption is not always held in the industrial environment of non-stationary working conditions. Hence, a novel model named Fault Response Network (FRN) is proposed, which is based on the bearing fault mechanism for diagnosis under variable conditions. Firstly, we calculated the fault feature that does not change with working conditions. Secondly, Fault Response Convolutional Layer (FRCL) is proposed based on that feature. Finally, the FRN is constructed with FRCL and improved soft threshold function. Four diagnostic cases are used to verify the superiority of FRN. The FRN can obtain high diagnostic accuracy when working conditions change largely without samples from unknown conditions.