Ultrasonic evaluation of stainless steel resistance spot welded based on deep learning

Resistance spot welding (RSW) is a core process for joining thin sheets in aerospace, rail transportation and automotive manufacturing. The size of the weld nugget directly determines the strength of the joint, significantly affecting the reliability and service life of the entire structure. This study addresses the challenge of accurately measuring the true nugget size in stainless steel RSW joints due to the difficulty in distinguishing ultrasonic time-domain signals between the corona bond and the actual nugget. A deep learning-based ultrasonic spiral C-scan capsule detection method is proposed. First, a spiral C-scan capsule detection platform was designed and constructed, and typical joints were tested to obtain three-dimensional ultrasonic signals from different regions (non-welded, corona bond, nugget and defects). Then, a data processing model combining 3D Res-Unet and Transformer was used for semantic classification of the ultrasonic signals. The classification results were reconstructed into a two-dimensional imaging grid under the spiral scanning path, and image processing techniques were employed to extract the corona bond morphology and calculate the true nugget size. The results show that the Dice value for the nugget is 0.95, with a measurement error of less than 0.06 mm. This method provides a new perspective for in-situ detection of spot-welded joints and holds promise for engineering applications.