Lightweight intelligent fault diagnosis method based on a multi-stage pruning distillation interleaving network

Abstract

Edge computing, a key technology in the Internet of Things, can help integrate real-time fault diagnosis into industrial applications. Lightweight and compression technologies are essential for deploying high-precision deep learning methods on resource-constrained edge computing systems. However, modeling accuracy is severely compromised by existing methods. To overcome this limitation, a new multi-stage pruning and distillation architecture was proposed in this study to compress a depthwise separable convolutional network for intelligent fault diagnosis of bearings in edge computing systems. The model was implemented on an NVIDIA Jetson Nano and verified using two bearing fault datasets. The results show that the proposed method can significantly reduce the calculation and reasoning time of the model and maintain high accuracy. The proposed method exhibits remarkable effectiveness, requires minimal memory, provides fast inference speeds, and is suitable for use in edge devices with less configuration.