Feature Recognition Technology of Substation Inspection Image

Yi Zhang,Xiaoqing Han,Baafi Zhang,Chunguang Ren,Lei Wang,Yue Qin

DOI: https://doi.org/10.1109/ei252483.2021.9713309

2021-10-22

Abstract:The task of target detection is to locate and classify the detected objects in the image. Applying it to the power inspection reduces the manpower and time consumption, improves the efficiency of the inspection work, and is of great significance to ensure the reliability of the power system operation. In this paper, the Faster R-CNN method is applied to the state detection of cotter pins. And the detection process is divided into two parts. The first part is the image preprocessing part, which will expand the date set by adding noise, sharpening, and shadowing on the original date. The cotter pins can be marked by the labeling tool Labelimage. The second part realizes the state detection of the cotter pin by Faster R-CNN model. The target candidate area is closer to the true value because of the unique RPN technology of the Faster R-CNN network. On the basis of the Faster R-CNN network, the update weight adopts L2 regularity to reduce the impact of over-fitting, which also solves the problems of gradient dispersion and gradient explosion to a certain extent. The Tensorflow's deep learning framework is applied in the experiment test to realize the location and classification of the four states: normal cotter pin, cotter pin not split, cotter pin single leg out, and cotter pin missing. The experimental results verify the correctness and effectiveness of the proposed method.