A self-adjusting transformer network for detecting transmission line defects

作者：Li, Weitao; Tong, Qianqian; Gu, Jiaqin^*; Li, Junchen; Sun, Wei; Li, Qiyue

来源：NEURAL COMPUTING & APPLICATIONS, 2023.

DOI：10.1007/s00521-023-09319-w

• Electric power defect recognition
• Electric power depth vision
• Feedback mechanism
• Intelligent detection
• Reinforcement Learning
• Transformer

To address the limitations of traditional defect detection methods for power transmission lines, this paper proposes an intelligent defect recognition method based on self-adjusting Transformer. Firstly, a deterministic networking with a large receptive field is used to extract features from the defect images obtained during power transmission line inspections. Subsequently, a DQN is employed to select important regions containing foreground information. Secondly, a bilinear attention mechanism is utilized to project the background region feature vectors, compressing their contribution in the fused feature vectors of the foreground and background regions. Furthermore, the fused feature vectors are input into a Transformer network based on adaptive encoding layers, enabling better focus on the target region. Position-scale constraints are added to the decoding layers of the Transformer to enhance the attention's emphasis on position-scale information, thereby accelerating the convergence speed of the Transformer. Finally, gate units are introduced in each decoding layer to adaptively adjust the structure of the Transformer decoding layers to accommodate the feature extraction requirements of different inputs. Experimental studies on aerial images of power transmission line defects were conducted, and the proposed method achieved an average detection accuracy of 89.9%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document}. Compared with other commonly used algorithms, it demonstrated superior detection accuracy and generalization ability.