Recognition Confidence of Welding Seam Defects in TOFD Images Based on Artificial Intelligence

Abstract

In this paper, a target detection approach (Faster R-CNN) based on convolutional neural networks is applied to the training and recognition of typical defects in TOFD welding seam images. Before training and recognition, a total of 162 ultrasonic TOFD welding seam images containing five typical defects are collected. The ultrasonic TOFD welding seam image dataset required for neural network model training is established on the basis of the collected images. The neural network model is trained including pre-training on ImageNet, RPN training alone, Faster R-CNN training alone, and joint RPN and Faster R-CNN training. During training, the parameters in the program are adjusted, and, then, the convergence of the neural network models and recognition performance are compared after the same training iterations. It is found that the neural network model has a tendency to converge only when the batch size is 10 and the learning rate is 0.001. Under this parameter configuration condition, the program conducts training with more iterations and is used to identify welding seam defects. The results show that the program is accurate in locating typical defects in the images, and the recognition confidence for all kinds of defects is more than 0.9. Compared with the other parameter configuration conditions after the same training iterations, the program has the highest recognition confidence in identifying all types of defects when the batch size is 10 and the learning rate is 0.001.