Rolling bearing fault diagnosis based on multiple wavelet coefficient dimensionality reduction and improved residual network

Abstract

In response to limitations in traditional intelligent fault diagnosis methods, such as accuracy, robustness, generality, and susceptibility to noise, this article proposes Multiple Wavelet Coefficient Dimensionality Reduction and Improved Residual Network (MWCDR-IResNet). This novel approach represents the first combination of multi-wavelet transform for extracting wavelet coefficients and Principal Component Analysis (PCA) for dimensionality reduction. It incorporates a Residual Network enhanced by the Squeeze and Excitation (SE) module to focus on relevant channels and capture critical fault-related features, followed by fault classification. Validation is performed using rolling bearing datasets from Paderborn University(PBU) and Southeast University (SEU). Experimental results consistently demonstrate MWCDR-IResNet superior performance in terms of robustness, generality, and average accuracy when compared to existing methods.