Fault Diagnosis of Axle Box Bearings via MPA-Optimized VMD and Lightweight Wavelet CNN

This study presents a novel lightweight convolutional neural network model designed for efficient and accurate fault diagnosis of high-speed train axle box bearings using raw vibration signals. The model has been extensively validated on a CRRC dataset containing 12,000 samples, using parameterized Morlet wavelet kernels to directly generate high-resolution two-dimensional time-frequency representations from the input at the initial layer, significantly enhancing the capture of early fault transient pulses. By combining depthwise separable convolution and 1×1 convolution compression strategy, this model achieved a compact size of 2.6 million parameters, requiring only 0.38 GFLOPs. The system, combined with optimized MPA-VMD denoising preprocessing steps, exhibited excellent robustness, maintaining a Pearson coefficient of 0.92 even under strong noise conditions with a signal-to-noise ratio of 5dB. Comprehensive evaluation showed the model significantly outperformed the standard convolutional neural network and MobileNetV2, achieving a test accuracy of 97.08%, precision of 0.97, and F1 Score of 0.97, representing a 2.22% accuracy gain over MobileNetV2. The entire system provided low latency performance, which was crucial for edge deployment. Preprocessing took 38ms, model inference was 320ms, and the total latency was 358ms, completely within the real-time limit of 500ms. The deployment on the Jetson Nano platform further optimized latency to 2,90ms and memory usage to 5.2MB. These quantitative results confirm the high precision, high efficiency, and practical feasibility of the model as a robust edge intelligent diagnostic system for railway bearing health monitoring.