Etude des structures microbandes par la technique des reseaux de neurones application a differentes excitations

Abstract

In this work a representative approach of the networks type of neurons is proposed to model three structures microbandes with various excitations starting from a unit of data obtained by two different methods: modified cavity and transmission line. The input impedance of a rectangular microstrip antenna on an isotropic substrate with or without air gap is calculated starting from the losses by radiation, by conduction and by dielectric. Network MLP is selected like model neuronal and the calculation of the input impedance is made by the application of ten algorithms of optimization.A comparison made between these algorithms shows that the Levenberg-Marquardth algorithm is most powerful with a good precision. We also showed the effectiveness of the reversed neuronal modeling in the determination of the geometrical paramétres of the microstrip antenna with air gap for a frequency predefined. The input impedance and the reflection coefficient of an aperture coupled rectangular microstrip antenna are calculated by applying the multilayer perceptron. The influence of various parameters of a structure supplied with coupling and by opening was detailed.The results obtained are compared with those of the literature. The neuronal model adopted in our own approach of modeling (RBF, GRNN) is exposed as well as the base of training which is elaborate starting from the results provided by TLM method on the rectangular antenna microbande coupled by proximity. A comparison between neuronal models MLP, RBF and GRNN shows that the RBF is characterized by its effectiveness and its performance on the speed level of convergence. The validation of the simulation results of the models is made on the Simulink™ in Matlab®, which permits to simulate the behavior of the system in time. The very good agreement between neural network values of the input impedance and those calculated (TLM and cavity) confirms the validity of the neuronal model.