A Multi-Scale Depth-wise Separable Convolution Swin Transformer for Fault Diagnosis of In-Wheel Motor Bearings

Fault diagnosis of in-wheel motor bearings is challenging due to weak fault features and non-stationary vibration signals under complex operating conditions. To address the limitations of conventional models in transient impact modeling and feature representation, this study proposes an enhanced Swin Transformer–based fault diagnosis framework. The proposed method integrates a multi-scale convolutional feature-enhanced feed-forward network (MSCF-EFFN) to improve shallow cross-scale representation, a unified deformable shifted window multi-head self-attention (DSW-MSA) framework to adaptively capture irregular transient impact features, and a depth-wise convolution attention module (DWCAM) to refine deep feature selection. The model is validated on a self-built dynamic test bench covering nine bearing health states and 28 operating conditions, achieving an average accuracy of 98.7% and a peak accuracy of 99.12%. Comparative and ablation studies demonstrate superior accuracy, robustness, and convergence performance over existing models.