A Multilevel Attitude-Aware Denoising Network for Bearing Fault Diagnosis

Abstract

The denoising of vibration signals is crucial for bearing fault diagnosis in harsh environments with strong noise. Nonetheless, the existing denoising approaches are insufficiently reliable to extract discriminative fault feature information from nonstationary signals. To address the issue, a multilevel attitude-aware denoising network (MADN) is proposed for bearing fault diagnosis with noise. First, an elicitation encoding structure is constructed to extract multiscale features. Then, the attitude-aware denoising modules are designed to mine the attitude information of features and learn the interdependencies among capsules. Finally, a multilevel capsule routing mechanism is proposed to accurately integrate the attitude information of features at different scales, alleviating fault information redundancy. The superiority of MADN is that multiscale feature attitude information is utilized to enhance the network's robustness. The comparison with state-of-the-art networks indicates a promising future for the proposed method under noisy conditions.