Physical Simulation-Driven Hilbert Convolutional Transfer Network for Few-Shot Defect Detection by Ultrasonic Guided Wave

In real-world applications, machine learning-based ultrasonic guided wave structural health monitoring (SHM) often faces challenges due to insufficient labelled data. In recent years, pre-trained models have been widely adopted to address the issue of limited sample sizes. However, existing pre-trained models struggle when there is a distribution mismatch between the source and target domains. To tackle this problem, we propose a physical simulation-driven Hilbert Convolutional Transfer Network (HCTN) framework for few-shot defect detection by ultrasonic guided wave. Within this framework, we introduce a Hilbert convolution module to extract the signal envelope, enabling HCTN to capture intrinsic features of damage signals in both the source and target domains and enhancing its ability to learn domain-invariant representations. The framework adopts a simulation-pre-trained transfer strategy and introduces a domain discrepancy-based dynamic fine-tuning approach, which reduces the risk of overfitting while minimizing the number of trainable parameters. Experimental results demonstrate that incorporating the proposed Hilbert convolution module leads to more distinct inter-class boundaries and tighter intra-class clustering in the feature space. Compared with other transfer learning, the proposed approach achieves superior performance in terms of both prediction accuracy and stability.