Identification of weld defects from ultrasonic signals using GASF and an improved DCGAN-ResNet network

In this paper, a method for weld defect identification from ultrasonic signals using the Gramian Angular Summation Field (GASF) and an improved deep convolutional generative adversarial network and residual network (DCGAN-ResNet) is proposed to overcome the problems of small-sample imbalance of echo signals as well as the low identification accuracy and poor efficiency of traditional convolutional neural networks (CNN). Firstly, the DCGAN model is improved based on Wasserstein distance and spectral normalisation, and the augmented dataset is used to validate its effectiveness. Then, the residual block for the ResNet model is improved using group convolution to enhance the nonlinear representation of the network and reduce the number of parameters and computations. Finally, the squeeze-and-excitation (SE) attention mechanism is introduced for feature recalibration to enhance attention to important features and recognition efficiency. Experimental results show that the improved DCGAN-ResNet method outperforms other commonly used methods in terms of feature extraction, recognition accuracy and efficiency for weld defects, and its test accuracy reaches 91.99%, which is 14.36% higher than that before dataset augmentation. Thus, the proposed method is effective and feasible for weld defect recognition from ultrasonic signals under small-sample imbalance conditions, and can also be applied to other pattern recognition fields.