@article{WANG2025, 
author = {Pengfei WANG and Lifang ZENG and Xueming SHAO and Jun LI},
title = {Multi-source data fusion modeling method for aerodynamic load of aircraft wing based on pre-training and fine-tuning},
year = {2025},
journal = {Acta Aeronautica et Astronautica Sinica},
volume = {46},
number = {19},
keywords = {vehicles, transfer learning, digital twins, distributed loads, data fusion},
url = {https://www.sciopen.com/article/10.7527/S1000-6893.2025.32297},
doi = {10.7527/S1000-6893.2025.32297},
abstract = {Accurate and rapid prediction of aerodynamic loads is an important part of the vehicle digital twinning technology, and is an important link between the real vehicle and its digital twin. At present, building aerodynamic load proxy model based on data modeling method to obtain aerodynamic data efficiently has become an important research direction in vehicle design. However, data modeling methods using a single source are difficult to break the upper limit of accuracy of the existing model predictions. Based on sparse and limited wind tunnel test data, a multi-source data fusion method of wing aerodynamic loads based on pre-training fine-tuning is proposed for the CRM-WB wing body assembly. Considering the difference in prediction accuracy caused by the pressure distribution characteristics on the upper and lower surfaces of the wing, the pre-training grouped fine-tuning strategy is further adopted to construct the aerodynamic load fusion model. The test results show that the average prediction error of the model is 3.17%, and compared with the prediction model based on single data training (an average error of 5.70%), the combined depth neural network fusion modeling method (an average error of 5.11%), and the Gauss process regression uncertainty weighted fusion modeling method (an average error of 6.16%), the multi-source data fusion method proposed in this paper achieves higher accuracy prediction. Generalizability tests show that the pre-training fine-tuning model proposed in this paper has good generalized ability, and the average error of the prediction model is reduced by 11.19% compared to the single data source in the extrapolation case.}
}