Identification of welding defects from ultrasonic signals using an improved ACGAN and WOA-ShuffleNet V1

In the ultrasonic detection of weld defects, addressing issues such as small-sample multi-class imbalanced distribution of echo signals and lightweight requirements for classification models, a welding defect ultrasonic signal recognition method based on Gramian angular summation field (GASF), improved auxiliary classifier generative adversarial network (ACGAN), and WOA-ShuffleNet V1 is proposed. First, ACGAN is enhanced by integrating a frequency-aware module, latent space optimisation, and optimised loss functions to improve generated sample quality. Then, through comparative analysis of expansion ratios and the impact of generative models on classification results, the effectiveness of the improved ACGAN in augmenting multi-class imbalanced small-sample data is validated. Finally, the whale optimisation algorithm (WOA) is employed to optimise hyperparameters of ShuffleNet V1. Experimental results show that on the expanded data, the method maintains 90.75% recognition accuracy while reducing FLOPs, model size and number of parameters, thus achieving a balance between lightweighting and classification performance.