Fault Diagnosis of Large-scale Electric Pumping and Irrigation Electromechanical Equipment by Integrating CWT and Swin Transformer

Jin Chen^*

• Gansu Province Jingtaichuan Electric Power Lifting Irrigation Water Resources Utilization Center, Baiyin, China

Abstract: This paper proposes a fault diagnosis framework integrating Continuous Wavelet Transform (CWT), Swin Transformer, and Cross-Attention for large-scale electromechanical equipment. CWT with optimized complex Morlet wavelet (scale factors 0.82 for impact sharpness, 0.67 for frequency focus) maps vibration signals into high-resolution time-frequency images. A four-stage Swin Transformer hierarchically extracts multi-scale features (56×56×96 to 7×7×768), while stride convolution aligns shallow features for Cross-Attention fusion, where shallow features serve as Query and deep features as Key/Value to dynamically weight and integrate cross-stage information. Experimental results on 10,190 samples demonstrate a classification accuracy of 98.7% (macro F1: 98.5%), with perfect identification of Motor Rotor Eccentricity (MRE). Ablation studies confirm the contribution of each component: removing Cross-Attention reduces accuracy to 97.7%, replacing Swin with standard Transformer drops it to 96.4%, and excluding CWT preprocessing further degrades performance to 95.6%. t-SNE visualization verifies enhanced intra-class compactness and inter-class separability, particularly between MRE and Gear Tooth Wear (GTW). The model maintains robustness under noise, sustaining accuracy from 82.3% (SNR=10 dB) to 98.7% (SNR=90 dB), validating its effectiveness and reliability for intelligent maintenance in complex industrial environments.