Data-driven ultrasonic defect identification is widely used in aerospace and industrial manufacturing. However, obtaining a large amount of experimental data remains challenging. While software-generated simulation data is easier to acquire, it differs significantly from experimental data, leading to suboptimal performance when applied directly. In this paper, the ultrasonic defect recognition method based on deep transfer learning is proposed. Firstly, ultrasonic testing experiments and ultrasonic testing simulations were carried out for defects of different shapes, sizes and depths in the specimen at the same time, generating both experimental and simulation data. Furthermore, a deep learning model for ultrasonic defect recognition was established based on simulation data. Then, a small amount of experimental data was employed to fine-tune the pre-trained model through transfer learning. So as to established a model that can achieve accurate defect recognition on the experimental data. Finally, the prediction effect of the built model was verified through experiments. The results show that after transfer learning, the accuracy and precision of the ultrasonic defect recognition model significantly improved, both achieving a value of 0.956.