HELICOPTER PARAMETER IDENTIFICATION USING A TRAINED MULTI-PERCEPTRON NEURAL NETWORK

A trained multi-perceptron is used to perform helicopter parameter identification. A learning algorithm based on error back-propagation is adapted in order to copy an input-output mapping into the network, which then realizes a parameter identification algorithm. The resulting architecture is used to reconstruct the stability and control derivatives of a UH-60A helicopter model, using simulated data at hover mode. The simulated data is generated with various system noise intensities to mimic realistic true helicopter data. The studied neural network identifier is found to be useful to the estimation task for the following reasons: (i) it does not require a-priori information, (ii) it is relatively insensitive to system uncertainties (e.g., noise levels, model structures), (iii) it demonstrates high relaxation on input designs and output samples, and (iv) it has self-learning ability to adjust the parameter estimate vector to compensate for incorrect model structure selection and to assure the fidelity of the identified model in comparison with the “true” system responses.