A neural network approach to the calibration of a flush air data system

A flush air data system uses surface pressure measurements to obtain speed and aerodynamic orientation of a flight vehicle. This paper investigates the use of a neural network to calibrate the air data system on a low-observable aircraft forebody with 22 pressure tappings. Wind tunnel data were obtained for between 0 and 25° angle of attack, 0 to 10° sideslip for speeds of 32, 37 and 45ms-1. Experimental data were used to train multilayer perceptron neural networks. A calibration accuracy of 0·322ms-1, 0·811° and 0·552° for speed, alpha and beta respectively was achieved just using the first five tappings on the nose cone. Increasing the number of tappings used as inputs to the neural network reduces the calibration error. A neural network with 22 inputs gives a best accuracy of 0·095ms-1, 0·15° and 0·085° for speed, α and β respectively. Trained networks show poor robustness to single sensor failures. Robustness is improved by preprocessing input data with an autoassociative network or by introducing sensor redundancy.