Improved YOLOv4 algorithm for safety management of on-site power system work. (November 2022)

• Record Type: Journal Article
• Authors: Li, Qiang; Zhao, Feng; Xu, Zhongping; Li, Kexin; Wang, Jing; Liu, Haofeng; Qin, Liang; Liu, Kaipei
• Is Part Of: Energy reports. Volume 8(2022)Supplement 13
• Page Start: 739
• Page End: 746
• Publication Date: 2022-11
• Subjects: Edge computing device -- Lightweight convolutional neural network -- Improved bidirectional feature fusion network -- Improved confidence loss function; Power resources -- Periodicals; Energy industries -- Periodicals; Power resources; Periodicals; Electronic journals; 621.04205
• DOI: 10.1016/j.egyr.2022.08.170
• Languages: English
• ISSNs: 2352-4847

Abstract

Abstract: To achieve real-time safety management of on-site power system work, the object detection algorithm of the cloud platform is transplanted into the edge computing device. However, due to the limited capacity and arithmetic power of the AI chip carried in the edge computing device, the algorithm needs to be studied in a lightweight manner. Therefore, this paper proposes an improved edge-side lightweight YOLOv4 (You Only Look Once version 4) algorithm, which is compressed and accelerated by improving the network structure and the confidence loss function of the algorithm without losing too much detection accuracy. Firstly, a lightweight convolutional neural network based on depthwise separable convolution and mobile inverted bottleneck convolution is constructed as the backbone network of YOLOv4, and the depthwise separable convolution is used instead of the conventional convolution, which can greatly reduce the parameters number and computation of the algorithm. Secondly, an improved bidirectional feature fusion network is constructed as the neck network of YOLOv4 algorithm to improve the information availability of multi-semantic features and enhance the extraction capability of small-scale features. Then, the confidence loss function of the algorithm is improved to alleviate the problems of uneven distribution of positive samples and negative samples and uneven distribution of easy-to-classify samples and hard-to-classify samples. Finally, the improved YOLOv4 algorithm is tested on an edge computing device equipped with an NVIDIA Jetson Xavier NX chip, and the dataset based on the compliance of on-site dress code in the electric power system field is tested. The experimental results show that, compared with the original YOLOv4 algorithm, the improved YOLOv4 algorithm reduces the parameters number by 93.11%, and the real-time video stream detection speed on the edge computing device is increased by 22.00%, reaching 15.05 FPS . The lightweight improved YOLOv4 algorithm is suitable for deployment at the edge end and can achieve real-time accurate identification of risks in the on-site work.