Identification of Nonlinear Corrections to Multicopter Flight Simulation Model Using Machine Learning

This study presents a novel methodology for identifying nonlinear corrections to improve the accuracy of a physicsbased simulation model of a hexacopter using flight test data. Two distinct models are employed to capture the dynamics of the hexacopter in hover: one model is identified from flight data, and the other utilizes a physics-based blade element model with a 10-state Peter-He inflow. An input filter is extracted based on the difference between the flight test data and frequency response from the physics-based model to make corrections. This improves the predictions over the entire frequency range. For time domain analysis, the nonlinear corrections are identified by analyzing correlations between different flight variables and utilizing a filtered dataset with high normalized correlation. Regularized version of partial least squares is applied for identifying the correction terms. The performance of the updated linearized model is compared with the physics-based model and system ID model in the time domain for all four axes. It was observed that for low amplitude maneuvers, the performance of model with nonlinear corrections is comparable to the model identified from flight test data and sometimes slightly better (2-5%). Predictions obtained using the corrected model exhibit superior performance to those generated by the physics-based model, particularly in the vicinity of peak values (8%-14%). For large amplitude maneuvers, the distinction is even more pronounced, and the model with nonlinear corrections surpasses all other models in terms of accuracy. Notably, while the physics-based model predictions exhibit an average error of 32% when compared to the flight test data, and the model identified using flight test data generated predictions with an average error of 24%, the model with nonlinear corrections yielded predictions with an average error of less than 10%.