Insulator defect detection in complex scenarios based on cascaded networks with lightweight attention mechanism

• Published in: Peer-to-peer networking and applications Vol. 17; no. 4; pp. 2123 - 2136
• Main Authors: Ning, Yang, Xiang, Li, Hongyuan, Jing, Xinna, Shang, Ping, Shen, Aidong, Chen
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2024; Springer Nature B.V
• Subjects: Accuracy; Agricultural production; Algorithms; Cluster analysis; Communications Engineering; Computer Communication Networks; Deep learning; Defects; Electric rates; Electricity distribution; Engineering; Fault detection; Fault location; Information Systems and Communication Service; Lightweight; Methods; Modules; Networks; Peer to peer computing; Signal,Image and Speech Processing; Special Issue on Affordable and Clean Energy; Target detection; Transmission towers; Unmanned aerial vehicles; Weight reduction; Complex environment target detection; UAV detection; Composite insulator; Defect detection
• ISSN: 1936-6442; 1936-6450
• DOI: 10.1007/s12083-024-01682-2

The power system stands as a crucial infrastructure pivotal to the country’s modern economic, security and social development. This paper addresses challenges in insulator fault detection on power transmission towers, leveraging the advancements in unmanned aerial vehicles equipped with target detection methods. We propose a novel method for insulator defect detection based on YOLOv5 (You Only Look Once), aiming to mitigate the issues associated with high missed detection rates. Small insulator faults and the limitation of unmanned aerial vehicle on-board capacity make it difficult to detect comprehensively. Firstly, the cluster analysis was carried out on the training data to obtain 9 kinds of better preset anchors for insulator detection, which improved the accuracy of the model to identify the location of targets. Secondly, the base-model is used to detect the insulator region, and the detection results are input into the sub-model to detect the location of faults, so as to form a cascade model, and make full use of the advantages of the two models to solve the problem of high missed detection rate. Finally, a lightweight attention module combining channel attention module and spatial attention module is added in YOLOv5 to improve the base-model’s attention to insulator region and suppress complex background features. Experimental results show that compared with the original model, the average precision of the proposed method for insulator detection is increased by 6.9%, and the missed detection rate of the fault location is 30% lower. Significant improvements in insulator detection performance have been achieved using the method proposed in this paper. It can not only effectively improve the detection accuracy, but also make the missed detection rate lower to meet the requirements of insulator defect detection and fault warning applications in complex environments, which proves that it has a wide range of application prospects in practice, especially in the field of power industry.