Algorithm for determining the tilt of support of electric power supplies by methods of deep learning with video data

Abstract

The uninterrupted operation of power lines can only be ensured by constant monitoring of the state of power lines. The most promising direction in the automation of the diagnostics of power transmission lines are robotic systems, such as UAVs with video cameras mounted on them or robots crawling through wires, used for a detailed analysis of the state of power transmission line elements. The video data received by them is further analyzed by the methods of machine vision and neural networks. One of the defects of power line elements is the deviation of the supports from the vertical. Despite the fact that in a number of works the state of power transmission line supports was determined (fallen supports or not), the angle of inclination of the supports was not calculated. The goal of this work is to develop an algorithm for calculating the angle of inclination of power transmission line supports for the module of the diagnostic system of the CableWalker complex, which is an unmanned aerial platform with the ability to land on a wire and move along it. The YOLO v3 neural network was used to select objects in images. To identify the slope of concrete supports in the obtained bounding rectangles, machine vision methods (Canny, Hough Line, GrabCut algorithm) were used. The algorithm has been developed to identify the objects contained in the image using the YOLO v3 neural network and determine the edges of the supports by building straight lines along them. The angles between the found edges of the support and the horizon (the lower edge of the image) have been calculated as the arctangent of the ratio of the coordinates of the point on the edge of the support. The algorithm has passed laboratory tests which have shown its suitability for use in the diagnostic system of the CableWalker complex. The application of the algorithm reduces the image processing time from several days (in the case of an expert) to several minutes and eliminates the error associated with the human factor.