Dynamic Prediction of Defective Composite Shells via Physics-Informed Data-Fusion Transformer

Composite open-shell structures manufactured using low-cost wet layup processes in electric aircraft are prone to resin unevenness and interlayer voids, resulting in multiple manufacturing defects. These defects induce localized stiffness discontinuities and complex nonlinear interactions, which pose severe challenges for accurate dynamic prediction. To address this issue, this study develops a Multi-source Data Fusion Enhanced Physics-Informed Transformer (MDFEPT) framework that integrates multisource data with physics-based stiffness boundary constraints that regularize the network and enforce physical consistency. This hybrid strategy enables accurate modeling of structures with arbitrary defect configurations. Trained on combined numerical–experimental data sets covering isolated and interacting defects, MDFEPT effectively captures nonlinear vibration features and mode coupling under varied operating conditions. The prediction accuracy exceeds 95%, and the mean square error is on the order of [Formula: see text], clearly outperforming conventional neural networks in both fidelity and stability. By embedding interpretable physical knowledge into deep learning, the framework bridges idealized models and real manufacturing scenarios, providing a robust tool for structural health monitoring and safety assurance of composite components in electric aircraft.