Méthodes statistiques pour la prédiction de température dans les composants hyperfréquences. (Statistical methods for temperature prediction in hyperfrequency components)

This thesis is focused on the application of statistical learning methods for the temperature prediction of anelectronic component embedded in a radar. We study a simplified case of real systems, the system under study islimited to a single component mounted on a reduced cooling system. The first chapter is devoted to heat transfermodelisation. After presenting the major mechanisms of thermal agitation transmission, analytical and numericalmodels are studied. Using this knowledge, the second chapter offers a survey on the methods of temperaturemeasurement, choosing the fittest according to the specifications and the constraints of the chosen application.Once databases have been established, we can use in the third chapter statistical learning techniques to build adynamic model. After a brief reminder about the ins and outs of statistical modeling, four families of methods willbe presented : linear models, neural networks, dynamic bayesian networks and support vector machines (SVM).The fourth chapter is an opportunity to present a novel method of modeling. Indeed, after a presentation of themethods for the identification of state representation, we see how to take into account theoretical apriorism duringlearning of this model type, ie a stability constraint.