Research on Lightweight Method of Image Deep Learning Model for Power Equipment

Abstract

Image information is the key element to judge the fault and defect of equipment in the power grid inspection. At present, all the automatic image defect analysis technology based on deep learning is to transfer the image to the server cloud for processing. However, due to the wide distribution of power equipment, large amount of image data and long time of data transmission, deep learning of massive image data will lead to severe problems of storage and inference speed. Therefore, it is necessary to carry out automatic analysis and research of power equipment based on edge computing, realize real-time defect analysis of power equipment image, significantly reduce the time of field detection, post-processing data and defect analysis, improve work efficiency and ensure the timeliness of defect detection. This paper analyzes the application scenarios of power equipment image recognition edge computing, summarizes the commonly used power equipment image recognition depth learning model, compares the computing power of existing edge computing chips, and puts forward two kinds of power equipment image recognition depth learning model lightweight schemes, which are network optimization, reasoning optimization and hardware optimization, and obtains the design of substation A scheme example of the application scene for visible light image recognition is provided. The results show that compared with NVIDIA TX2 chip which uses mobile edge computing directly, using TensorRT hardware architecture acceleration and low precision grid acceleration can reduce the recognition time by 48% ~ 72%. Compared with desktop GPU GTX1080TI, using TensorRT hardware architecture acceleration and low-precision grid acceleration recognition time is still longer, so further research on lightweight optimization scheme is needed.