Robust Forward Control of Aircraft Wing Shaking using Neural Networks

This study comprehensively reviews the challenges of wing flutter prediction and suppression under high-speed flight conditions. At the same time, a robust feed-forward control strategy based on radial basis function neural network (RBFNN) is proposed. The strategy first uses the finite element method to establish a dynamic model of wing elastic vibration and aerodynamic coupling. At the same time, the flutter critical speed is determined by combining the simplified unsteady aerodynamic theory. On this basis, RBFNN is used to perform online approximation of the unknown nonlinear dynamic terms in the system. This operation can reduce the dependence on accurate physical modeling. In order to offset the neural network approximation error and external disturbances, a sliding mode robust term is introduced in the feed-forward compensation. At the same time, the asymptotic stability of the closed-loop system is rigorously proved by Lyapunov stability analysis. The experimental part, based on the MATLAB/Simulink environment, compared the performance of PID control, sliding mode control (SMC), neural network feed-forward control (NNFC), and the proposed method (R-NNFC) over a flight speed range of $100-400 \mathrm{~m} / \mathrm{s}$. The results showed that R-NNFC significantly outperformed other methods in terms of vibration peak suppression across the entire speed range. Numerically, for example, at $300 \mathrm{~m} / \mathrm{s}$, its vibration peak was approximately $72.4 \%$ lower than that of PID. Frequency response analysis showed that the main modal frequency of each method remained at 15.2 Hz. The modal damping ratio of R-NNFC reached $12.3 \%$, and the vibration decay time was 1.9 s. Neural network approximation error analysis showed that the root mean square error of RBFNN remained within 0.01 across the entire speed range. External disturbance suppression experiments further verified the results. Experimental results showed that at a disturbance amplitude of 50 N, the vibration peak of the R-NNFC was only 8.30 mm, significantly lower than the $\mathbf{3 3. 5 0 ~ m m}$ of PID.