Multi-perspective deep transfer learning model: A promising tool for bearing intelligent fault diagnosis under varying working conditions

Abstract

Taking advantage of multi-perspective information such as space, channel and sequence into the deep transfer learning (DTL) model is beneficial to extract discriminative features from the original bearing vibration data. However, the idea with multi-perspectives is not totally considered in the existing DTL models, which might cause insufficient characterizing capabilities of the DTL-based fault diagnosis models with acquiring domain invariant features. To this end, a multi-perspective DTL (MPDTL) model including the perspectives of space, channel and sequence is proposed for the bearing fault diagnosis under varying working conditions in this study. Specifically, the proposed MPDTL model consists of three following parts: (1) A feature enhancement network (FENet) is first established to improve the quality of characteristics under the perspectives of space and channel, in which a residual block attention model with the space and channel attention mechanisms is designed to reduce the redundant features and avoid the gradient vanishing. (2) A bidirectional long-short term memory (BiLSTM) network is further introduced to extract the high-level discriminative features from the output of FENet under the perspective of sequence. Hence, the sequence-related information contained in the features learned by the FENet will be disclosed via feeding the forward and backward sequences into the BiLSTM. (3) The module of optimization objective is designed to update our model under the constraint of domain adaptation learning and domain shared classifier training. Analytical results of three experimental cases show the effectiveness and superiority of the proposed method in enhancing the bearing fault diagnosis under varying working conditions.