Infrared defect detection in laser-MAG hybrid welding based on deep learning methods

Infrared thermal imaging technology enhances the visualization and identification of surface defects in welding joints by capturing the temperature field distribution and generating thermal images with color gradients. Compared with traditional visual inspection methods—which are significantly affected by factors such as lighting conditions, background complexity, and weak defect features—infrared imaging demonstrates superior capability in representing small-scale and diverse welding defects. Given the current lack of publicly available infrared welding defect datasets, this study develops an infrared image acquisition system tailored for penetration defects in laser-Metal Active Gas Arc (MAG) hybrid welding and constructs a corresponding dataset for training and evaluating non-destructive defect detection models based on infrared sensing. On this basis, a penetration defect detection model based on the YOLOv5s architecture is established, and a feature map visualization method is introduced to improve the interpretability of the detection results. The model is trained and tested on the constructed dataset, and the experimental results demonstrate that the model achieved strong performance in the task of penetration defect detection, with a precision of 97.2%, a recall of 96.6%, and a mean average precision (mAP) of 98.6% at an Intersection over Union (IoU) threshold of 0.5.