Intelligent Framework for Bearing Fault Diagnosis in High-Noise Environments: A Location-Focused Soft Threshold Denoising Approach

In recent years, numerous bearing fault diagnosis systems tailored for special scenarios have emerged. However, when dealing with high-noise vibration data in real industrial environments, the model is unable to distinguish embedded noise effectively. Consequently, the noise persists in various feature extraction structures, leading to a lower fault diagnosis recognition rate. This paper presents an intelligent bearing fault diagnosis framework (ENC-LFST), comprising an instance enhancement batch normalization layer (IEBN), an improved channel attention block, and a soft threshold denoising structure (LFST) that prioritizes location information. IEBN serves to mitigate original noise as well as batch noise in features. Improved channel attention blocks enable the model to focus on important feature channels, assigning greater weight while suppressing noise information in the channels of small weights. LFST uses the concept of soft threshold denoising, establishing a threshold on feature information, filtering out data below the threshold, and addressing location information on feature information. This process eliminates noise while extracting more valuable feature information. To validate the effectiveness of the proposed framework, experiments are conducted on three distinct datasets. The results show that the method described in this paper effectively recognizes fault categories in high-noise industrial environments, exhibiting strong robustness.