Fault Detection Method of Infrared Image for Circulating Pump Motor in Valve Cooling System Based on Improved YOLOv3

Abstract

Timely maintenance of the key equipment in valve cooling system plays an important part in maintaining stable operation of a flexible DC converter station. In order to accurately locate and recognize the defects from infrared images of a motor of circulating pump, a motor fault detection method based on improved YOLOv3 is proposed in this paper. First, an improved Multi-Scale Retinex with Chromaticity Preservation (MSRCP) image enhancement algorithm based on Y component is proposed to increase the infrared image contrast, which makes the target more prominent. Then the convolutional block attention module (CBAM) is applied to feature pyramid network (FPN) to improve the YOLOv3 network. In order to improve the accuracy of model as much as possible, various kinds of training strategies are employed, which include mosaic data augmentation, mixup data augmentation, label smoothing, exponential moving average (EMA) and transfer learning. Finally, comparative experiments are carried out to test the effectiveness of the employed methods. The experiment results show that the network improvement methods could effectively increase the detection accuracy of the model. The mean of average precision (mAP) of the final model reaches 96.09%, and the average fault detection accuracy improves to 94.98%. The detection speed of the improved model can reach 42 frames per second (FPS), which meets the real-time monitoring requirements of the valve cooling system equipment.