Development and Evaluation of Neural Network-Based Virtual Air Data Sensor for Estimation of Aerodynamic Angles

Aerodynamic angles of ight vehicles are necessary to pilot and automatically control of aircraft. These angles are usually measured using probes that protrude from the vehicle surface out into the ow eld. However, this arrangement was found to be unacceptable for modern unmanned airplanes whenever stealthiness features are required. In addition, redundant sensor arrangements, when dictated by safety regulations, were also critical because of the possible heavy impact on the airframe of small UAVs. New virtual software-based systems were therefore developed in order to nd a viable solution for reducing the number of traditional hardware-based air data sensors, and they oered the benet of simplifying air data system architectures. The aerodynamic angles were derived from inertial data and by exploiting the airspeed sourced by the Pitot-static system. The relationship between these parameters and the aerodynamic angles was a complex, non-linear function that was not easily described by means of aircraft models. The main goal of this work, which was aimed at UAV applications, was to analyze the aircraft system and develop virtual sensors by exploiting soft computing methods, such as neural prediction techniques, in order to assess the feasibility of this kind of neural system. The performance of virtual sensors were tested using real hardware in the simulation loop and to represent real-world ight conditions: wind gusts, air turbulence and internal sensor noise were simulated. A sensitivity analysis was carried out to study the performance of virtual sensors even when realistic accuracy of measured signals, processed by neural networks, and failure modes were simulated. Finally, neural networks resulted to be suited for aerodynamic angle estimation technique: the neural networks worked properly with the available vehicle data and demonstreted to be as accurate as traditional probes.