Multi-type data fusion via transfer learning surrogate modeling and its engineering application

Abstract

In engineering problem, there exists many mixed datasets composed of multi-type data, including data without noise, data with homoscedastic noise (known noise variance), and data with heteroscedastic noise (unknown noise variance). To construct accurate predictive models, it is essential to fully utilize the information provided by the diverse dataset, particularly when confronted with limitations in training samples. Motivated by this, this paper introduces a novel transfer learning surrogate modeling approach to fuse multi-type data for prediction purpose. Numerical cases and two engineering applications validate the accuracy and robustness of the proposed model. Comparative analyses reveal superior predictive accuracy and robustness compared to other models. Additionally, by varying the levels of homoscedastic and heteroscedastic noise, the impact of noise on the proposed model is studied. Results indicate that this model has superior applicability and stability compared to other models. The proposed approach is capable of fusing multi-type data and adapting to different levels of noise, making it a promising approach for engineering application.