An interpretable waveform segmentation model for bearing fault diagnosis

Abstract

Accurate extraction of fault features is crucial for successful bearing diagnosis. Existing signal analysis methods require substantial expert knowledge and are prone to failure when encountering strong background noise. Meanwhile, routine intelligent models focus on diagnosis accuracy merely and are ambiguous in feature mining. To address these issues, this work proposes a novel interpretable waveform segmentation model for bearing fault diagnosis. Specifically, a nested U-Net is built as the backbone of the proposed method, in which the signal segmentation technology is designed to enable the pixel-level mining of interpretable information related to fault features. Additionally, class weighted lovász-softmax loss is designed to address the problem of spatial recognition ability and imbalanced pixel distribution. Physics-informed denoising loss is integrated to make the model more robust and interpretable. We evaluate the proposed method by numerical simulation and bench experiments, in which the overall segmentation accuracy and individual segmentation features are presented. Methodically, we illustrate the advantages of this approach using three distinct evaluation criteria, achieving pixel accuracy surpassing 90% in both simulated and experimental scenarios.