A Model Compression Based Framework for Electrical Equipment Intelligent Inspection on Edge Computing Environment

The edge services including electrical equipment intelligent inspection on power IoTs often adopt deep neural networks (DNNs) to accurately recognize abnormal equipment by image classification and object detection for reducing work workload on the power cloud. However, the high computation complexity based on neural network models poses great challenges to real-world edge services and applications due to the limited computational ability and storage space on edge devices. In this paper, we propose a framework for electrical equipment intelligent inspection based on deep neural network model compression where a combination approach is used to prune and quantize the DNNs automatically without using any hyperparameters to manually set the compression rate for each layer, which is applied in the edge services to handle and analyze the real-time massive data acquired by a number of power devices for reducing the computational complexity and the workload of edge computing services. The related experiments were made, and the results show that the electrical equipment intelligent inspection based on the proposed framework has superior classification accuracy, in particularly maintaining a competitive compression rate in comparison with the popular deep compression approach.