Air Data Computation Using Neural Networks

The paper deals with the use of neural networks for the determination of pressure altitude and Mach number of a fly-by-wire high-performance aircraft during flight. In previous works the authors developed a methodology based on polynomial calibration functions for the determination of such flight parameters, together with the angles of attack and sideslip. Such an approach provided successful results, but the use of different polynomial functions in different areas was needed to map the entire flight envelope. The fading methodologies for the management of polynomial functions overlap and considerably increased both procedure complexity and the time to spent for the procedure tuning. In particular, the calibration functions related to the Mach number and static-pressure estimation are susceptible to these problems because of their high nonlinearity. The alternative approach studied in this paper, based on neural networks, provides a level of accuracy comparable with that of polynomial functions. However, such an approach is simpler, because it allows the entire flight envelope to be mapped by means of a single network for each output parameter, and so it eliminates the fading problems. In addition, the new procedure is extremely easier to tune when new data from flight tests are available. This is a very important point, because several versions of the air data computation algorithms are generally to be developed in parallel with the flight-envelope enlargement of a new aircraft.