Aerodynamic modeling and active sampling based on uncertainty prediction

Neural network methods, as an efficient and accurate modeling approach, have been widely used in various fields. However, the "black box" feature of neural networks, combined with the engineering problem of few-shot phenomenon, leads to insufficient model reliability and high uncertainty in the prediction results, severely limiting the use of neural network models. In order to enhance the engineering applicability of neural network models, this study focuses on the unsteady aerodynamic characteristic and utilizes time convolutional networks (TCN) to model the temporal unsteady aerodynamic forces in large-amplitude oscillatory wind tunnel tests. The MC-dropout technique is employed to evaluate the uncertainty of prediction results. Based on the uncertainty analysis results, active sampling of wind tunnel test samples is conducted. The results indicate that model uncertainty can be used as a prior evaluation of the prediction accuracy. There is a strong linear relationship between the model prediction error and the uncertainty. The active sampling strategy can reduce the required samples by up to 40% compared to the random sampling strategy. This validates the effectiveness of the present method in improving the trustworthiness of black-box models and reducing the number of modeling samples required.