Automatic image recognition and defect grading system of converter valve

Abstract

With the growing demand for power reliability in electrical equipment and ongoing advancements in infrared thermal imaging, infrared detection methods have become increasingly prevalent in the operation and maintenance of converter valves. However, due to the influence of heat diffusion, equipment noise, environmental noise, and weak intelligent means, the contrast of infrared images and the inspection arrival rate are low, and the inspection timeliness cannot be guaranteed. To solve the above problems, a new system of automatic image recognition and defect grading of converter valve is created in this paper. Firstly, the converter valve infrared images collected by infrared imaging equipment are optimized and enhanced by constructing image enhancement module. Then, the image region segmentation module is constructed. And pixel-level infrared image segmentation is realized by combining 2D U-Net network and coordinate attention mechanism. Finally, classification recognition and grading module is constructed. And the cascade of LSTM-BP network and multi-classification algorithm are used to realize infrared image classification and defect grading. The results indicate that the recognition accuracy of eight types of devices is the lowest is 87.33%, the highest can reach 96.99%, and the average recognition accuracy is 93.47%. The overall performance of the algorithm is good with high reliability.