Research On Aerial Insulator Defect Detection Based On Deep Learning

UAV power inspection is a new type of inspection method that utilizes the advantages of unmanned aircraft to operate independently and complete maintenance of overhead transmission lines.The drone acquires high-definition aerial image through the image acquisition device carried by the fuselage and the remote radio control system,and analyzes it.Different from the traditional manual inspection method,the UAV inspection has the advantages of strong terrain adaptability,high safety factor,low risk,low cost,accuracy and reliability,and has become one of the key development directions of transmission line operation and maintenance technology.However,due to the complexity of aerial imagery and the limitations of visual recognition technology,how to implement intelligent power inspection system is still a difficult task.In this paper,the insulator defect detection in aerial image is taken as the research object,and the method of insulator defect detection based on deep learning is studied in combination with the requirements of power inspection.A real-time single-stage insulator defect detector and a cascaded insulator defect detector for detecting very small defect blocks are proposed respectively.The main work of this paper is as follows:Firstly,based on the requirements of online inspection,based on the deep learning target detection method,a real-time lightweight YOLOV3 insulator defect detection method is proposed.The lightweight network is a reasonable optimization of the compressed network MobileNetV1,and greatly improves the inference speed of the network while maintaining the basic performance.In order to improve the stability and accuracy of anchor frame generation before training,a K-means++ algorithm based on Euclidean distance is proposed.For the problem of insufficient image data of the insulator image,the network parameters trained by the PASCALVOC2007 data set are pre-loaded into the insulator defect detection network to complete the weight initialization.The problem of slow network convergence and low detection accuracy caused by insufficient data volume is improved.Secondly,due to the safety distance regulations of the UAV power inspection,for some strict control areas,the aerial image taken by the UAV inspection will have a very small proportion of the power equipment.In this paper,a data fusion method is designed.By combining the reduced insulator with the background image of the transmission line,a minimum insulator defect data set that satisfies the condition is fabricated.Aiming at the problem that the defect is very small and the detection is difficult,a cascade network structure is proposed to detect the very small defect block.The structure establishes a two-level detection network.The first-level network is responsible for capturing the position of the insulator.After cutting,it inputs the secondary network to identify the insulator defect,strengthens the extraction and representation of the defect block feature,and improves the accuracy and recall rate of the defective block.In summary,this paper studies the defect detection method based on deep learning by in-depth analysis of insulator image information.For the characteristics of online and offline inspection,the two dimensions of speed and precision are optimized respectively.The proposed method can effectively It is applied to UAV power inspection tasks and has certain theoretical significance and research value.