Analog system modeling based on a double modified complex valued neural network

Abstract

Abstract-The aim of this work is to present a novel technique for the identification of lumped circuit models of general distributed apparatus and devices. It is based on the use of a double modified complex value neural network. The method is not oriented to a unique class of electromagnetic systems, but it gives a procedure for the complete validation of the approximated lumped model and the extraction of the electrical parameter values. The inputs of the system are the geometrical (and/or manufacturing) parameters of the considered structure, while the outputs are the lumped circuit parameters. The method follows the Frequency Response Analysis (FRA) approach for elaborating the data presented to the network. I. INTRODUCTIONURING the analog circuit design process, in many cases the phase of modelization, parameter identification and simulation of distributed circuits is still an arduous challenge. The difficulties are inherent in locating the parameters that can be extracted and in obtaining the requested precision for them. On the other hand, this phase can become essential to solve several problems, as, for example, the study of the transient part of the response, the evaluation of the electromagnetic compatibility, the estimate of the harmonic content, the detection and location of faults, the influence of a single parameter over the output final behavior. In the last few years several soft computing algorithms dealing with this subject have been developed, using artificial neural networks (ANNs) [1][2][3], genetic algorithms (GAs) [4,5], particle swarm optimizers (PSOs) [6,7].In this paper we propose a new neural architecture able to accomplish the identification task. It is constituted by the union of a pair of multi-valued neuron neural networks with complex weights [8]. In this kind of architecture it is possible to use a set of multifrequency measurements or simulations made on the device, taken at different values of geometrical parameters, and train a MultiLayer Multi-Valued Neuron Network (MLMVN), able to estimate the electrical parameters of the lumped model. A single network is not A. Luchetta is with the Dipartimento di Ingegneria dell'Informazione (DINFO) of the University of Florence, Via S.Marta, 3, 50139 Firenze, Italy (phone: +39-055-4796461; fax: +39-055-4796442; e-mail: luchetta@ unifi.it).S. Manetti is with the Dipartimento di Ingegneria dell'Informazione (DINFO) of the University of Florence, Via S.Marta, 3, 50139 Firenze, Italy (e-mail: stefano.manetti@ unifi.it).M.C. Piccirilli is with the Dipartimento di Ingegneria dell'Informazione (DINFO) of the University of Florence, Via S.Marta, 3, 50139 Firenze, Italy (e-mail: mariacristina.piccirilli@ unifi.it). sufficient to achieve the goal, because, in general, the supervised reference of the lumped model is not available. In order to overcome this limitation, another network must be added, having the role of approximating and "inverting" a network function of the chosen lumped model. The output of this layer consists of the numerical estim...