Enhancing Power Distribution Defect Identification with Video- Based Computer Vision: A Two-Step Defect Detection Method

However, video and computer vision-based inspections face challenges. Many defects are small and hard to detect in videos, and some defects are rare, making it difficult to gather enough training data for accurate detection. To address these issues, this paper introduces a Two-Step

Defect Detection (TSDD) method. In the first step, TSDD identifies certain larger objects like poles or crossarms using a computer vision model, which can typically achieve high detection performance due to the objects’ large sizes. Each detected object is then cropped, and a second computer vision model is applied to detect defects within the cropped area. This method allows for a larger pool of training data for the second model, as it does not rely solely on video frames.

In this paper, the method of using TSDD for identifying distribution defects is discussed. The identification is focused on five typical defects in distribution systems. Some of these defects, like missing animal protection, are small and difficult to detect in video frames. Some defects, such as open cutouts and floating primary lines, are rare, making the preparation of training data challenging. We present novel and practical designs of TSDD for these defects. We demonstrate the efficacy and superior performance of TSDD and provide valuable lessons and insights gained throughout the process. By incorporating videos, computer vision, and TSDD into distribution defect identification, utilities can enhance efficiency, improve accuracy, and increase human safety, ultimately benefiting both utility companies and the broader society.