Comparison Between Numerical Results and Operative Environment Data on Neural Network for Air Data Estimation

Several architectures exist to measure aerodynamic angles based on physical sensors. As far as Unmanned Aerial Vehicle (UAV) is concerned, traditional systems hardly comply with reliability and redundancy requirements due to size and weight limitations. A patented virtual sensor, based on Neural Network (NN) techniques, named Smart-ADAHRS (Smart-Air Data, Attitude and Heading Reference System) has been investigated as a good estimator for aerodynamic angles in simulated environment. This paper focuses on flight testing procedures in operative environment and data processing for the Smart-ADAHRS validation with real data. As many factors interfere during the generation of the NN training set, an accurate choice and integration of the FTI (Flight Test Instrumentation) system components becomes crucial. A comprehensive description has been included about the FTI equipment and its influence on the neural network performance. Differences between numerical simulation and operative environment data are detailed as final aim of this work. Finally, feasible solutions are suggested to solve the typical gap between virtual and real scenario, both in terms of data analysis and neural network architecture.