Numerical Study on the Robustness of a Neural Network Model for Gross Weight Estimation of Tiltrotor Aircraft

Abstract

The objective of this study is to investigate the robustness of a neural network model for gross weight estimation of a tiltrotor aircraft in both airplane and helicopter modes. To overcome the limitations in available flight test data for training and validating a neural network model, the Comprehensive Analytical Model of Rotorcraft Aerodynamics and Dynamics II (CAMRAD II) analysis program was used to simulate various flight conditions for a generic tiltrotor aircraft. A total of 315 different flight conditions in airplane mode and 330 different flight conditions in helicopter mode were simulated. An equilibrium solution for each flight condition was achieved by CAMRAD II through adjusting power settings, aircraft states, control surface deflections and rotor controls. Two comprehensive databases were created with parameters required for aircraft trim; one for the airplane mode and one for the helicopter mode. The training and validation datasets were selected randomly from these databases and are exclusive of one another. The Back-Propagation Network and Radial Basis Function Network paradigms were utilized respectively in the airplane mode and in the helicopter mode to estimate aircraft gross weight. Results show that the neural network model has a good capability of estimating the gross weight for a tiltrotor aircraft in both airplane and helicopter modes. An RMS error of 168 lbs was achieved in airplane mode and the RMS error in helicopter mode was 189 lbs. These results were very robust within the boundaries of the training dataset. The robustness of the neural network model was also investigated by examining data points on the borders of training dataset domain. In this case, it was found that caution should be used when neural network estimation is applied to flight conditions on the border of the training dataset domain. In addition, different data densities for gross weight and airspeed values in the training dataset were investigated. The results can be used as guidance in the selection of a training dataset required for achieving acceptable neural network estimation accuracy. Neural network performance was also investigated when flight conditions in the validation dataset are completely outside the training dataset. Results indicate that neural network estimation accuracy declines rapidly for this case.