Recursive Gauss-Newton based training algorithm for neural network modelling of an unmanned rotorcraft dynamics

The ability to model the time varying dynamics of an unmanned rotorcraft is an important aspect in the development of adaptive flight controller. This paper presents a recursive Gauss-Newton based training algorithm to model the attitude dynamics of a small scale rotorcraft based unmanned aerial system using the neural network (NN) modelling approach. It focuses on selection of optimised network for recursive algorithm that offers good generalisation performance with the aid of the cross validation method proposed. The recursive method is then compared with the offline Levenberg-Marquardt (LM) training method to evaluate the generalisation performance and adaptability of the model. The results indicate that the recursive Gauss-Newton (rGN)method gives slightly lower generalisation performance compared with its offline counterpart but adapts well to the dynamic changes that occur during flight. The proposed recursive algorithm was found effective in representing helicopter dynamics with acceptable accuracy within the available computational time constraint.