Research On Insulator Detection And Self-detonation Defect Identification Method Based On Improved YOLOX Model

Insulators are key components that provide electrical insulation and mechanical support for current-carrying conductors on high-voltage transmission lines.Defects are likely to occur due to various factors such as transient loads,mechanical stress,atmospheric conditions,etc.Furthermore,they might then threaten the stable operation of transmission lines which highly impacts the security of the power system.A UAV(Unmanned Aerial Vehicle)is more efficient and convenient as it offers visual assessments of structures.Therefore,it has gradually replaced the traditional manual inspection method.The detection of insulator defects based on aerial images consequently has become popular.However,insulator defects in aerial images often exhibit small targets and complex backgrounds in the dataset.However,insulator defects are often characterized by small target and complex background,which makes it difficult to detect insulator defects quickly and accurately.To address the above issues,after analyzing and comparing several existing classical target detection algorithms,this paper selects the YOLOX algorithm as the benchmark model for this paper,and improves the YOLOX algorithm by means of ablation experiments to enhance the performance of the algorithm in insulator detection and self-explosion defect identification tasks.The main research is as follows.(1)The advantages and disadvantages of several classical Two-stage algorithms and One-stage algorithms are compared and analyzed,and the YOLOX-S model of YOLOX algorithm is selected as the benchmark model in this paper,and comparative experiments are designed.The experimental results show that the YOLOX-S model performs better in the insulator detection and self-explosion defect identification tasks.(2)In response to the characteristics of insulator image data such as sparse self-exploding fault image data,single light condition and low percentage of targets in the image,pre-processing methods such as geometric transformation and contrast transformation are used to increase the number and diversity of insulator image data,and Label Img is used for annotation to construct an image data set of insulators and their self-exploding defects.(3)The shortcomings of the YOLOX-S model are analyzed and corresponding solutions are proposed for the small target and background complexity problems of insulator detection and self-exploding defect identification tasks.First,to address the problem of small targets,the shortcomings of the regression loss function IoU(Intersection over Union)of the YOLOX-S model are analyzed in the context of the sensitivity of small targets to the model regression accuracy,and the influence law of the regression direction of the prediction frame on the model accuracy is studied,and the SIoU-d regression loss function is proposed.The experimental results show that the regression loss function can effectively improve the accuracy of the model in detecting small targets of insulator self-detonation defects with an AP of 95.39%,an improvement of 2.63%.(4)For the complex background problem,an attention mechanism is embedded between the backbone feature network and the feature fusion layer of the YOLOX-S model to reduce the influence of redundant features on the detection accuracy and improve the detection accuracy of the model in complex backgrounds.The experimental results show that the embedded ECA attention mechanism can effectively improve the accuracy of the model in detecting targets in complex backgrounds with an m AP of 95.17%,which is an improvement of 0.73%.In this paper,the insulator image dataset is constructed by preprocessing methods such as geometric transformation and contrast transformation,and the regression loss function of the YOLOX-S model is improved to reduce the influence of complex background by embedding the attention mechanism.The experimental results show that the improved YOLOX-S model improves the detection accuracy of insulators by 0.45%,the recognition accuracy of self-burst defects by 5.03%,and the detection speed is 71 FPS,which verifies the effectiveness of the method in this paper.