Abstract: Aircraft are prone to structural damage under high-temperature and high-speed flight conditions, posing a severe threat to flight safety. Given the complexity of aircraft structural models and the diversity of damage types, traditional prediction methods struggle to achieve high-precision dynamic characteristics prediction of damage. This paper proposes a composite material defect combination shell dynamic characteristics prediction method based on multi-source data transfer learning Transformer neural networks (MTT): First, experimental data and finite element data of simplified combination shell models are fused to extract data fusion features and construct a pre-trained model. Physical constraints are applied to the loss function of the pre-trained model to fine-tune its parameters, and then the pre-trained model is transferred to the target prediction model via transfer learning. Verification experiments were conducted using two types of defective combination shells as test objects. The results demonstrate that the proposed method significantly improves prediction accuracy and efficiency compared to traditional baseline models.