송전철탑 및 전력설비 자동검출을 위한 영상 분석 알고리즘 개발

Abstract

Over the past few decades the capacity of the power transmission system is growing more and more rapidly, and the power systems have also been expanding in line with continuous expansion of infrastructure. Efficient maintenance of the transmission infrastructure is critical for stable power system operation. However, since the transmission infrastructure is installed at a high location from the ground, the maintenance and diagnosis on transmission towers always involves the possibility of the safety accident of workers. Since 2016, Korea Electric Power Corporation has been developing technologies for monitoring and diagnosis of power lines using drones and has verified the usefulness of technologies through on-site demonstration. 3GBytes of video will be acquired by one flight. And video analysis is performed by below process. ① A worker finds an image frame that includes a transmission tower and power facilities, ② and expands the area of interest in the image, ③ finally the visually observes the facilities and determines whether defects are. These process requires the concentration of the operator and causes very high fatigue. To solve these problems, it is necessary to develop an image analysis algorithm for automatic detection of transmission towers and power facilities to automatically find image frames containing facilities and automatically enlarge the area of interest in the frame to provide diagnosis operator. In this paper describes the process of building a learning and testing image database for detecting transmission towers and facilities carried out for the development of these algorithms and the process of developing algorithms using deep learning CNN-YOLO v2(Convolutional Neural Network - You Only Look Once v2) and the improvement of performance aspects including accuracy and speed with existing algorithms. The recall and precision rate of YOLO v2 for detecting transmission infrastructure is 96.30% and 95.65% respectively, which is more accurate and faster than YOLO v1.