Modeling Aircraft Wing Loads from Flight Data Using Neural Networks
This paper documents input data conditioning, input parameter selection, structure, training, and validation of neural network models of the Active Aeroelastic Wing aircraft. Neural networks can account for uncharacterized nonlinear effects and retain generalization capability. Model inputs include aircraft rates, accelerations, and control surface positions. Linear loads models were developed for network training starting points. The models were trained with rolls, loaded reversals, windup turns, and individual control surface doublets for load excitation. Data results from all loads models at Mach 0.90 and altitude of 15,000 ft. show an average model prediction error reduction of 18.6 percent.