The Defect Detection Method for Cross-Environment Power Transmission Line Based on ER-YOLO Algorithm

The defect detection algorithm for power transmission lines based on deep learning heavily relies on the dataset used. The learned features of line defects and image features are highly correlated with the feature distribution of the dataset used. Currently, most datasets used to train defect detection algorithms are captured in a single environment, leading to the specific defect recognition algorithm performing well only in a particular environment. However, its generalization ability is inadequate in different shooting environments, such as low light, haze, and blur, resulting in poor algorithm robustness. To improve the robustness of the intelligent defect detection algorithm for power transmission lines in different environmental conditions and overcome the issue of decreased recognition performance in different environments, this paper proposes the cross-environment robust YOLO algorithm (ER-YOLO). Firstly, based on the you only look once v8 (YOLOv8) algorithm, ER-YOLO enhances the algorithm's long-distance modeling capability by employing the Transformer attention mechanism based on the generalized attention theory. Secondly, ER-YOLO improves the path aggregation network (PANet) in the YOLOv8 algorithm by increasing the convolutional kernel size and introducing an efficient attention mechanism in the cross stage partial layer (CSPLayer), enhancing the network's object detection capability. Finally, ER-YOLO uses a multiple attention mechanism detection head network to strengthen the algorithm's multi-scale, spatial location, and multitask perception capabilities, enabling the network to focus on crucial target information. Additionally, to validate the cross-environment detection performance of the defect recognition algorithm, a certain number of corresponding datasets of power transmission line defects in different environments are required. However, due to environmental and practical engineering constraints, collecting a large number of datasets in specific environments is challenging. To address the issue of the test dataset, this paper proposes common methods for generating adverse environment data for power transmission line defect recognition. High-fidelity test datasets were generated based on normal environment data. The paper explores methods for simulating adverse environments, including dark light environment simulation using exposure fusion algorithms and brightness reduction methods, haze environment simulation using Cycle GAN networks, and imaging blur environment simulation using mean filtering methods. The effectiveness of each method was evaluated and compared with other methods, providing conditions for testing the robustness of defect recognition algorithms across different environments. Through ablation experiments and comparative analysis, ER-YOLO demonstrated higher defect recognition accuracy and robustness in cross-environment testing. The average mAP value under various test datasets was 0.726, showing an improvement of 0.069 compared to the previous version. The algorithm's effectiveness was further validated in real environments. This study proposes a defect recognition method for power transmission lines across shooting environments, exhibiting excellent performance in cross-environment recognition. It also explores methods for generating cross-environment images, providing insights for future virtual dataset generation techniques. Future research directions may focus on cross-environment defect recognition studies for other types of defects and explore other effective methods for generating multi-environment virtual datasets. © 2024 China Machine Press. All rights reserved.