Deep-FusionNet: A prior-guided feature fusion framework for high-fidelity aerodynamic noise prediction

Although wind tunnel tests can provide high-precision aerodynamic noise data, their costs are limited by the number of test runs and the realization of high-resolution frequency sampling. Numerical simulations, on the other hand, are restricted by numerical accuracy and model reliability, making it difficult to independently achieve high-confidence noise prediction. To balance the contradiction between data acquisition cost and prediction accuracy, this paper proposes a deep learning framework that integrates sampling optimization and feature learning—Deep-FusionNet. The framework combines Proper Orthogonal Decomposition (POD) and Autoencoder (AE) techniques to enhance multi-scale flow feature representation and construct a modal space. In addition, a Genetic Algorithm (GA) is introduced to determine the frequency sampling points in wind tunnel experiments based on numerical simulation data, thereby achieving high-fidelity reconstruction of the full-spectrum and aerodynamic noise spectra across the entire state space. To verify the prediction accuracy and generalization capability of the proposed method, it is compared with the Compressive Sensing based on Proper Orthogonal Decomposition (POD_CS). Research shows that Deep-FusionNet requires only 35% of sparse frequency sampling points to control the reconstruction error of the aerodynamic noise spectrum within 1.0×10⁻³, which is two orders of magnitude lower error than the POD_CS method. Furthermore, the proposed method demonstrates strong generalization and robustness across different flow conditions and spatial positions, significantly reducing the dependence on the number of wind tunnel runs, long-duration stable test conditions, and dense sampling equipment. It thus provides a practical tool for conducting high-precision aerodynamic noise measurements under resource-constrained conditions.