Identification and Validation of the Cessna Citation X Longitudinal Aerodynamic Coefficients in Stall Conditions using Multi-Layer Perceptrons and Recurrent Neural Networks

Abstract

The increased number of accidents in general aviation due to loss of aircraft control has necessitated the development of accurate aerodynamic airplane models. These models should indicate the linear variations of aerodynamic coefficients in steady flight and the highly nonlinear variations of the aerodynamic coefficients due to stall and post-stall conditions. This paper presents a detailed methodology to model the lift, drag, and pitching moment aerodynamic coefficients in the stall regime, using Neural Networks (NN). A system identification technique was used to develop aerodynamic coefficients models from flight data. These data were gathered from a level-D Research Aircraft Flight Simulator (RAFS) that was used to execute the stall maneuvers. Multilayer Perceptrons and Recurrent Neural Networks were used to learn from flight data and find correlations between aerodynamic coefficients and flight parameters. This methodology is employed in here to optimize neural network structures and find ideal hyperparameters: training algorithms and activation functions used to learn the data. The developed stall aerodynamic models were successfully validated by comparing the lift, drag, and pitching moment aerodynamic coefficients predicted for given pilot inputs with experimental data obtained from the Cessna Citation X RAFS for the same pilot inputs.