Fault Tolerance and Extrapolation Stability of a Neural Network Air-Data Estimator

Abstract

Neural networks have been employed in signal-processing software used to translate the measured pressure distribution from the nose of an aircraft into estimates of freestream static pressure, vehicle speed, and vehicle attitude relative to the e owe eld. The performance of the resulting system under normal e ight conditions has been previously reported. This paper investigates the performance of the neural network system under more adverse conditions. Specie cally, the effect of partial signal failure from the pressure distribution measurements is investigated. Additionally, the stability of the system is tested for applications outside the original domain of the training set. The neural network air-data estimator was found to be both tolerant to faults in the input data and stable when applied to moderate distances of extrapolation.