Novel multi-scale lightweight network cable defect diagnosis based on adaptive pruning

Intelligent fault diagnosis of high-voltage cable defects based on edge computing is a safe and reliable technology for high-voltage cable operation, which can provide better real-time performance and lower communication cost for fault diagnosis systems. However, existing lightweight diagnostic networks lack lightweight learning mechanisms for edge computing scenarios. To this end, this paper proposes a new multi-scale lightweight network based on adaptive pruning. First, a precision current sensor (TMR) based on tunnel magnetoresistance effect technology is used to capture the weak pulse current signal generated by partial discharge in high-voltage cables, and a weight-shared multi-scale convolution (WSMSC) is proposed, which can effectively extract the scales in partial discharge. Features and almost no increase in computing cost. This mechanism not only improves the completeness of feature extraction, but also improves fault diagnosis accuracy, which is especially suitable for resource-constrained edge computing environments. Then an inverse separable convolution (ISC) block is constructed to adjust the convolution order to further capture high-level features, achieve one-to-one matching of convolution kernels, and simplify the pruning process. The adaptive pruning technology proposed finally integrates the evaluation and pruning process into the training, avoiding the repeated evaluation, pruning and fine-tuning cycles in the traditional method, and can simultaneously train and optimize the network structure without manual intervention, which is beneficial to Deployment on an expanded spectrum of peripheral units. Empirical outcomes demonstrate that the precision and intricacy of the devised framework surpass those of other cutting-edge cable anomaly detection systems. Furthermore, the proposed network can be deployed on more general edge devices.