Research on fault diagnosis of metro traction motor bearings based on MCDRSN-DSC and information fusion

Abstract In response to the limitations of traditional fault diagnosis methods in feature extraction and the poor performance of single-sensor diagnostics under complex working conditions, this study proposes a multi-channel deep residual shrinkage network optimized with depthwise separable convolution (MCDRSN-DSC) and information fusion for diagnosing faults in subway traction motor bearings (model NU216). First, vibration and acoustic emission signals from the bearing were collected. Their time-frequency features were then extracted using continuous wavelet transform. An attention-optimized convolutional neural network image fusion module (IFCNN) was employed to perform weighted data fusion and construct the dataset. Subsequently, a fault identification model based on MCDRSN-DSC and information fusion was established. Finally, diagnostic experiments were conducted under single, variable, and composite working conditions, comparing the performance of single versus fused signals, as well as different models and fusion methods. The results demonstrate that the proposed method achieves superior diagnostic performance compared to single-signal approaches and outperforms other traditional models across all tested working conditions, verifying its effectiveness and feasibility.