Numéro : 2259 - Year : 1995
Neural architectures for ship modelling and identification
G. HARDIER, ONERA/CERT/DERA, Toulouse
From the beginning of the eighties, artificial neural networks have focused a number of applications in fields such as pattern recognition, classification. optimisation, diagnosis, adaptive filtering and more recently identification and control of nonlinear systems. The origin of this interest lies in their ability to approximate any nonlinear function by means of suitable learning methods. according to the network characteristics. Consequently, within the framework of sea vehicles modeling. they can provide an interesting altemative 10 more classical methods. the use of which could be tedious when no satisfying theoretical model exists. Therefore, this paper aims at analysing the implementation of these techniques for the identification of a manoeuvrability model. The recurrent neural architectures, which will be used, and the principles of the associated learning methods, will be detailed and profusely illustrated with a set of realistic simulations of ship manoeuvres (presently the Charles de Gaulle aircraft carrier). The difficulty linked with a rough blackbox approach will be highlighted. We will point out the benefits to be derived from a more specific methodology, allowing for a priori knowledge and stressing the different phenomena by means of proper tests. A comparison with classical methods will permit a better understanding of its interest and complementarity. and likewise to unmyslify such an approach.
Advanced search
Year