Automatic Inspection of Power System Operations Based on Lightweight Neural Network
Bin Li, Shuang Wu, Shengjie Wang, Liang Zhang · 2022 7th Asia Conference on Power and Electrical Engineering (ACPEE) · 2022
In the automatic inspection task of the power system industry, it is of great significance to employ artificial intelligence algorithms to automatically detect the illegal operation action of workers. Although many existing object detection algorithms can achieve good detection rate, due to the large scale and large amounts of parameters of the model, it can hardly be directly used in edge devices for real-time detection. Based on two frequency employed baseline lightweight models, this paper presents a novel deep learning algorithm for illegal operation action detection which uses the 5×5 deep detachable convolution kernel. The proposed lightweight network could not only extracts discriminative features by the two special convolutional layers, but also better detect small targets in the scene. The proposed network can be implemented on edge devices such as Huawei Atlas 200 DK. Experimental results on our collected dataset of the surveillance videos shot during the project of Ningxia Electric Power Co., LTD. Show that the proposed method could achieve comparable accuracy with only very few number of parameters, thus can be applied for real-time automatic inspection.