YOLO-oil: A Real-time Transformer Fault Detector toward Small Dataset

Abstract

In electrical substations, fault detection of the transformer widely relies on human eye, which is low efficiency and costly. With under-oil robot and deep learning algorithms, the fault detection will be done without draining the transformer oil. As we know, deep learning methods for computer vision have achieved incredible results on some tasks such as object detection. However, such success greatly relies on the huge dataset, which is extremely high-cost and unavailable in some industry application. Deep learning algorithm, such as YOLO series, often fails on small dataset, and the test accuracy decreases significantly due to the neural network overfitting on the small dataset. In this paper, the YOLO-oil network for transformer fault detector based on YOLOv5 is proposed to mitigate the overfitting problem on small dataset: First, we shrink the network depth and get a light weight backbone. Second, we improved the network architecture by decoupling the detect head network. Since no open dataset exists for transformer fault detection before, the author creates a brand-new training dataset and a test dataset. Experimental results on the test set show that our algorithm achieves surprising results for the transformer fault detection task and surpasses YOLOv5, which is a great help to industry application.