Multi-Scale Channel Mixing Convolutional Network and Enhanced Residual Shrinkage Network for Rolling Bearing Fault Diagnosis

Abstract

Rolling bearing vibration signals in rotating machinery exhibit complex nonlinear and multi-scale features with redundant information interference. To address these challenges, this paper presents a multi-scale channel mixing convolutional network (MSCMN) and an enhanced deep residual shrinkage network (eDRSN) for improved feature learning and fault diagnosis accuracy in industrial settings. The MSCMN, applied in the initial and intermediate network layers, extracts multi-scale features from vibration signals, providing detailed information. By incorporating 1 × 1 convolutional blocks, the MSCMN mixes and reduces the feature dimensions, generating attention weights to suppress the interference from redundant information. Due to the high noise and nonlinear nature of industrial vibration signals, traditional linear layer representation is often inadequate. Thus, we propose an eDRSN with a Kolmogorov–Arnold Network–linear layer (KANLinear), which combines linear transformations with B-spline interpolation to capture both linear and nonlinear features, thereby enhancing threshold learning. Experiments on datasets from Case Western Reserve University and our laboratory validated the efficacy of the MSCMN-eDRSN model, which demonstrated improved diagnostic accuracy and robustness under noisy, real-world conditions.