A method for determining the attitude of a small unmanned aerial vehicle using sensor error information and a neural network

Abstract

The study has sought to identify a means for estimating the attitude of a small unmanned aerial vehicle by exploiting the fundamental operating principle of a neural network, using information about its micro-electro-mechanical systems sensors’ error properties. Individual attitude solutions have been derived from the accelerometer, gyroscope and magnetometer using traditional techniques. These solutions have been then fused together using a weighting strategy that ensures that each sensor’s error is accounted for. The developed algorithm has drawn upon knowledge from statistical noise parameter estimation and voting strategies applied in redundant aircraft systems to ensure that the system output closely matches the platform’s actual attitude. Experiments that have been carried out to determine the efficacy of the developed algorithm using data from real world flights have shown that the algorithm has improved accuracy compared to the individual sensor solutions, with root-mean-square-error of below 10 in several situations.