Detection and quantification of tube-melting defects in tube-to-tubesheet welding based on angle patch embedding and Transformer

Tube-to-tubesheet welding is a critical connection technique in heat exchangers, but tube melting defects may occur during welding process due to abnormal current fluctuations and other factors. However, most existing defect detection approaches fail to achieve accurate tube melting defect localisation and quantitative measurement. To address this challenge, a new approach is designed to detect and quantify the tube melting defects in tube-to-tubesheet welding based on angle patch embedding and Transformer. The main work is as follows: (1) An angle patch embedding module (APEM) is designed, which applies semantic segmentation, polar coordinate transformation, and patch embedding to convert annular welding images into circumferential feature sequence suitable for Transformer models; (2) A tube melting defect quantitative detection module is developed, utilising Transformer to model global circumferential dependencies, enabling high-precision regression of inner and outer tube boundaries and tube melting defect detection, based on which the angle-wise tube melting rates (TMRs) are calculated; (3) A tube-to-tubesheet welding image acquisition system is constructed, and a tube melting defect dataset is established. Experiments indicate that the proposed method achieves accurate identification of tube melting defects and provides quantitative detection results with angular resolution, offering a novel non-destructive testing approach for welding quality assessment.