Defect signal intelligent recognition of weld radiographs based on YOLO V5-IMPROVEMENT

Abstract

Internal pipeline weld defects cause pipeline cracking accidents, whereas X-ray detection can detect these defects. The deep learning-based intelligent defect identification model of weld radiographs extracted weld defects automatically through a convolutional neural network, thereby eliminating the subjective interference of human factors and improving the quality and speed of film evaluation. By proposing the YOLO V5-IMPROVEMENT model and adding the CA attention mechanism, SIOU loss function, and FReLU activation function, this paper improved the ability to detect small targets, capture low-sensitivity spatial information, and perform global optimization. A total of 7500 radiographs containing weld defects of a Chinese oil and gas long-distance pipeline were selected for training, verifying, and testing the model developed in the paper. Precision and recall of the YOLO V5-improvement presented in this paper reached 92.2 % and 92.3 %, which were 10.7 % and 12.5 % higher than YOLO V4, and 9 % and 11.2 % higher than the unimproved YOLO V5 model, respectively. It is confirmed that YOLO V5-IMPROVEMENT has high accuracy and high robustness and that applying this model to the intelligent defect identification of weld ray images can significantly improve detection efficiency and reduce the misjudgment rate.