A Deep Learning Approach for Aerodynamic Characteristics Identification Based on Time-Frequency Analysis

With the increasingly wide range of flight altitude and velocity, the complexity and uncertainty of aircraft dynamic characteristics become prominent. The difference between aerodynamic data obtained according to wind-tunnel testing or computational fluid dynamics and actual aerodynamic characteristics becomes increasingly significant, resulting in the inability to schedule control gains. Deep learning is used to identify the actual aerodynamic characteristics of the aircraft and adjust control gains in real time. First of all, aerodynamic characteristic uncertainties and inconsistencies are modeled by varying the aerodynamic data used for control system design. The digital simulation is carried out under different aerodynamic modes and types of overload commands. Then, classify the aerodynamic characteristics according to the gain margin. Secondly, continuous wavelet transform is used to obtain time-frequency images of angular velocity, control surface deflection, overload, and overload command. Finally, convolutional neural networks are trained to extract time-frequency image features and identify the type of aerodynamic characteristics. The simulation results show that the accuracy of online identification reaches 96.3%.