A Multi-Domain Data-Based Attentive Residual Network for Bearing Fault Diagnosis

Abstract

 Methods of data-driven fault diagnosis are of great significance to ensure the stability and reliability of bearings systems. However, the existing methods still encounter many challenges. From the perspective of data, a single domain signal cannot fully reveal complex industrial processes. From the perspective of the model, abstract features at different levels contain fault information with various importance, which affects the model performance. In this paper, an adaptive residual network based on multi-domain data is proposed to make up for the shortcomings of existing fault diagnosis methods. Firstly, FFT frequency domain analysis is conducted on time domain signals, and multi-domain data are constructed together. Secondly, an adaptive attention mechanism is introduced into the residual block based on one-dimensional convolution to fuse shallow and deep features, so as to extract features more effectively. Finally, experiments on rolling bearings at Case Western Reserve University manifest that the proposed method is superior to other comparative methods in fault diagnosis.