Research on machine learning-assisted prediction methods for aircraft surface noise

This study aims to overcome the limitations of low prediction accuracy, long computation time, and limited computational states in aircraft surface noise analysis by introducing a machine learning-based approach. Using data obtained from high-fidelity numerical simulations, a deep feedforward neural network (DNN) model is developed to predict aircraft surface noise considering multiple influencing parameters. The model's robustness and predictive capability are enhanced through random search-based hyperparameter optimization, enabling efficient and accurate noise prediction within the training data range. Moreover, the incorporation of the softmax activation and cross-entropy loss functions facilitates automatic classification of flow states at monitoring points. An integrated neural network model, Physical-Model-Embedding Ensemble Neural Networks (PENN), incorporating empirical formulas for aircraft surface noise calculation is established, achieving noise fusion prediction combining simulation and engineering empirical formulas. The results demonstrate that the PENN model effectively addresses the limitations of low accuracy and poor generalization in high-precision modeling with small and sparse samples. Compared with the conventional DNN model, the PENN model achieves a 48.8% reduction in the root mean square error of the overall sound pressure level prediction.