Critical fault diagnosis method of small components of power transmission based on optimizational YOLOv3 algorithm

Abstract

In order to solve the problem that the image recognition algorithm in UAV (Unmanned Aerial Vehicle) inspection of power transmission has low recognition rate for small, indiscoverable and high-risk component failures, this paper proposes a critical fault diagnosis method based on the optimizational YOLOv3 algorithm. We add SPP net under the basic framework of the original YOLOv3 and extract multi-scale depth features in the same convolutional layer which can transform all the input images into a uniform size and solve the problem caused by different input image sizes and formats. An optimizational scale prediction framework of YOLOv3 is proposed, which expands the original three-scale prediction frameworks to four scales and carries out multi-scale integration. This improvement is used to strengthen the feature extraction of small and indiscoverable fault. The optimizational algorithm is verified by using the dataset of electrical components of transmission system in the actual scene which have small and indiscoverable faults. The results show that the optimizational YOLOv3 improves the recognition rate of the selected categories by 3.98% on average, and increases the recognition rate of the small size and indiscoverable fault by 11.95% on average.