A lightweight multi-scale time-frequency fusion model for motor fault diagnosis

Abstract Timely and accurate motor fault diagnosis is crucial for ensuring the reliability of industrial equipment. Analyzing fault signals and identifying fault categories in complex environments pose significant challenges for electric motor diagnostics. This study introduces a lightweight multi-scale time-frequency fusion model designed to enhance diagnostic accuracy while minimizing model parameters. The proposed model utilizes dual-branch feature extraction to capture distinct time-domain and frequency-domain characteristics across different fault scenarios, enhancing both performance and robustness. A key innovation is the frequency-attention-based adaptive filter, which dynamically emphasizes fault-specific frequency bands during training, facilitating precise feature extraction. Additionally, an advanced time-frequency fusion module, incorporating an improved cross-attention mechanism, efficiently integrates multi-domain information to enhance diagnostic capability. The model’s lightweight design, achieved through optimized convolutional attention mechanisms, results in the smallest parameter size among comparable models, with only 68,303 parameters. Comprehensive experiments conducted on synchronous motor and asynchronous induction motor datasets demonstrate superior classification accuracies of 98.83% and 99.09%, respectively, along with robust performance under high-noise conditions. Ablation studies further validate the effectiveness of the time-frequency fusion module across diverse scenarios. This model achieves an optimal balance between diagnostic performance and computational efficiency, representing a significant advancement in motor fault diagnosis.