Efficient Bearing Fault Diagnosis for Edge Computing Using Grayscale Spectrograms and Hybrid Neural Model Compression

Abstract

Bearing elements play important roles in mechanical machining. Efficient monitoring and detection of bearing faults are critical for ensuring motor reliability. Deep learning techniques, particularly those that use convolutional neural networks, have emerged as prominent solutions for achieving high accuracy in bearing fault classification. However, the computational complexities of these models render them unsuitable for use in resource-constrained devices. In this study, we explored the application of a combination of multiple techniques, including knowledge distillation, layer fusion, and quantization-aware training, to develop lightweight models that accelerate inference speed while maintaining high accuracy. In addition, acoustic emission signals containing bearing health information were preprocessed using a short-time Fourier transform and converted into grayscale spectrograms before being fed into a neural network. The experimental results indicate that the proposed method can significantly reduce the inference latency and memory footprint while achieving high accuracy for different bearing fault types at variable rotational speeds.