A Hybrid CNN/Poisson Fusion Based Power Transformer External Defect Detecting Method

Abstract

The external defects of power transformers always occurs, suffered from external forces, aging and long-term overload operation. With fast development of computer vision, inspection robot is replacing artificial patrol to detect external defects. In this paper, we propose an external defect detection framework for power transformer using a hybrid method of convolutional neural networks (CNNs) and Poisson fusion. Firstly, the YOLOv3 algorithm is introduced to locate and extract the power transformers, where the style transfer method is proposed to stylize the image and thus improve the robustness of the object detection model under image corruptions. Secondly, the improved LeNet-5 algorithm is proposed to detect the rust and oil leakage defects of the power transformer. To reduce the negative impact caused by the lack of defective images, Poisson fusion is introduced to generate defective images to improve the general ability of the defect detection model. Finally, the detection framework is validated with actual power transformer images.