Meta-learning enhanced classification of complex defects in pressure vessels

Abstract Weld defects in pressure vessels, prone to development under long-term operation under complex and harsh environments with alternating loads and corrosion, pose critical safety risks. Current defect classification primarily relies on manual interpretation of sensor images, which is inherently subjective. Therefore, deep learning (DL) has become a common way to classify images. However, insufficient quantity of defect sensor data limiting DL applicability, and the prevalence of small-scale, morphologically diverse defect instances that hinder accurate classification. To address these issues, we propose a two-stage meta-fine-tuning paradigm (TMF) for transfer learning in the classification of welding defects in pressure vessels. This paradigm migrates the weights of the source domain training to the RIAWELC dataset. The first stage employs meta-learning, where a limited representative sample is selected from multiple defect categories, allowing for rapid adaptation to new tasks with minimal supervision. The second phase performs fine-tuning with a small set of labeled target domain data and augments with MixUp to enhance generalization and robustness. Experimental results demonstrate that the TMF algorithm achieves an accuracy of over 92%, outperforming conventional manual inspection and DL approaches. The proposed paradigm demonstrates promising generalizability for defect classification tasks in analogous industrial scenarios.