Intelligent Edge Gearbox Faults Diagnosis System via Multiscale Depthwise Separable Convolution Network

The health status monitoring of key components in gearbox is of great significance to ensure the safe operation as well as the stability of production efficiency. However, the realtime requirements of the fault diagnosis system and the model size of the deep learning approach limit the practical application of edge intelligent fault diagnosis in gearboxes. Therefore, by combining deep learning methods method and edge computing technology, this study proposes an intelligent edge diagnosis system for gearbox. First, the architecture of the edge fault diagnosis system is designed based on wireless sensor networks. Second, a lightweight model multiscale depthwise separable convolution network (MKDS_1DCNN) is proposed to enhance the extraction of vibration signal features. Finally, the MKDS_1DCNN is deployed on the edge node with X-CUBE-AI edge inference framework, including signal acquisition circuit and STM32. The conducted experimental studies show that this implementation can achieve an inference time of less than 20 ms and accuracy of more than 98 %. The effectiveness of the proposed gearbox edge fault diagnosis system and its capability in practical applications at the edge are verified in the experiments.