Yolov7-DROT: Rotation Mechanism Based Infrared Object Fault Detection for Substation Isolator

Abstract

Fault detection of isolators plays a significant role for the safety of power systems. The majority of existing object detection algorithms only achieve the object discrimination of infrared images, without being able to identify failure of the object. Furthermore, the interference of complex backgrounds and the large aspect ratio structure of the isolators pose challenges to the detection model. To address the above issues, an infrared object detection method incorporating rotation mechanism, called Yolov7-DROT, is proposed. By fusing the rotation mechanism with the prediction part, the interference of the complex background is greatly reduced and the quality of the prediction box is improved. A deformable convolution is introduced for the structure of isolators with large aspect ratios, which strengthens the feature extraction capability of the model for isolators and improves the detection accuracy. Additionally, a global-local distribution detection strategy for isolator faults is proposed, where the global detection results are fed into a local detection model to learn the fault features of isolators. Experimental results show that the proposed method accurately identifies isolators and knife switches in object detection, achieving an average detection accuracy of 96.28%. For thermal fault recognition in knife switches, the fault identification rate reaches 96%.