Abstract: Addressing the difficulty in obtaining training datasets for defect recognition in ultrasonic B-scan images using deep learning methods, a model training method combining simulation data driving and deep learning transfer mechanisms was designed. First, the CIVA nondestructive testing simulation platform was used to construct weld structure models containing various artificial defects. Different defect types, positions, sizes, and detection angle parameters were set to generate simulated ultrasonic B-scan images, building a simulated training dataset. Pre-training was conducted based on the YOLOv8x model to obtain an initial model with domain feature extraction capabilities. Finally, based on this initial model, transfer learning was performed by freezing some convolutional layers in the backbone network and inputting real detection images. The experimental results showed that the model exhibited good parameter metrics on the real image validation set and achieved satisfactory defect localization and recognition effects.