Detection of cotter pins missing of connection fittings on transmission lines of power system

Connection fittings are widely used in the transmission lines. However, the missing of cotter pins usually happens to the bolts of connection fittings due to the complex environment and aging. Cotter pins missing detection is one of the most time-consuming and labor-intensive parts of manual judgment. It belongs to the small target detection because the cotter pins with less features in the images are much smaller than other components on transmission lines. This paper designs a three-stage cotter pin missing detecting method based on pictures taken by UAVs. The first stage utilizes YOLOv4 to detect insulators and locate the connection fittings by extending the predicted bounding box of the insulator in the original pictures. The second stage is a small object detection model based on an object network called Faster R-CNN with ResNet-101 backbone. The network of this stage detects the bolts on the connection fittings and divides these bolts into two categories. The third stage utilizes the DenseNet121 classification network to identify the integrity or missing of the cotter pins. The experimental dataset consists of a public dataset and pictures captured by the UAVs from state grid. The result shows the accuracy of the bolts detection exceeds 90%, and the accuracy of front pins missing detector exceeds 85%. However, the lower accuracy of lateral pins detector due to the samples of dataset.