Lightweight Disconnector State Detection Model Based on Improved yolov5s
Jishen Peng, DongJie Xin, HaiMing He, Liye Song · 2022 9th International Forum on Electrical Engineering and Automation (IFEEA) · 2022
Aiming at the problems of complex environmental factors in substations, low accuracy, poor robustness and large amount of detection network parameters, an improved yolov5s lightweight network model is proposed for the location and state detection of six types of disconnector. Firstly, aiming at the problem of sample imbalance, use the albumentations library to perform data enhancement on the disconnector image; Secondly, the lightweight network MobileNetV3 is used for backbone feature extraction, and the Coordinate Attention (CA) mechanism is introduced to improve its inverted residual structure to more accurately locate the target; Change the PANet of the neck to BiFPN, a bidirectional feature pyramid, to enhance the feature transfer to solve the problem of the target being occluded; Finally, the weight coefficient $\alpha$ is introduced to improve the frame loss function and optimize the convergence effect of the model. The experimental results show that the improved network model obtained through iterative training can overcome environmental factors and accurately identify the different states of various disconnectors. The average detection accuracy ([email protected]) can reach 97.2%, which is 3.4% higher than the original model.