Abstract: Objective Addressing the issues of low manual efficiency, high detection difficulty, and high missed detection rates in traditional weld defect detection, this study proposes a defect recognition method for TOFD D-scan images based on experimental simulation and deep learning algorithms. Methods Different types of butt weld defects (slag inclusions, cracks, lack of penetration, porosity, and lack of fusion) are simulated through indoor experiments. A feature image library of actual TOFD detection data is conducted by various data augmentation methods. TOFD D-scan image datasets are trained and detected by the YOLOv5 deep learning network structure to enhance the model’s capability to recognize weld defects and automatically output intelligent detection results. Results The experimental results indicate that this method possesses excellent model generalization ability, achieving an accuracy of 98.05% with an IoU threshold of 0.5. For five types of welding defects, the classification confidence during recognition exceeds 95%, significantly improving the efficiency of weld defect recognition and meeting the requirements for online recognition in practical scenarios. Conclusion The proposed weld defect recognition method demonstrates high accuracy and can be widely used for constructing various defect detection models, providing effective technical support for welding quality control.