A generalized data-driven framework for guided wave-based damage localization and severity determination across different structures using small-sample data

Data-driven approaches have been extensively investigated for damage detection using ultrasonic guided waves (GW). However, most existing frameworks assume the availability of large training datasets from the targeted structure. In practice, GW datasets are often scarce for an individual structure. This limited data availability poses a challenge to the effectiveness of many existing data-driven models. To address this limitation, this paper presents a small-sample data-driven framework for determining crack location and severity across different structures. A progressive hybrid dilated convolution and efficient channel attention assisted transformer encoder-based variant (PHDC-ECA-TEV) network is proposed to achieve the purpose of generalization with limited datasets. Different isotropic beam-like structures are employed to systematically examine the applicability and generalizability of the proposed framework with limited datasets. For efficient data preparation, a validated higher-order beam model (HOBM) is adopted. Multiple experimentally collected datasets from different beam models are also adopted to further examine the generalizability. Furthermore, SHapley Additive exPlanations (SHAP)-based method is introduced to interpret the internal decision-making process of the deep learning (DL) network in predictions, offering insights into feature contributions. The numerical results demonstrate that the proposed framework can generalize damage detection across different beam-like structures under small-sample conditions, achieving a MAPE of 1.46% for damage localization and a Macro F1 score of 99.29% for damage severity determination, where only 150 samples are available for each individual structure. Experimental results further support its robustness, with most cases exhibiting APD values below 10% for damage localization, and all aluminium and steel beams correctly classified in severity. Meanwhile, the SHAP-based analysis shows the proposed approach can capture physically meaningful patterns consistent with domain knowledge, increasing the fidelity and trustworthiness of the application.