Diagnostic et surveillance des processus complexes par réseaux bayésiens

This thesis is about the multivariate process monitoring (detection and diagnosis) withbayesian networks. It allows to unify in a same tool (a bayesian network) some monitoring ded-icated methods like multivariate control charts, discriminant analysis and the MYT method.The first chapter gives some essential points of the process monitoring, with a study of differ-ent approaches. Some supervised and non-supervised monitoring methods are presented and astudy of different classifiers for monitoring purpose is made. A classifier is then chosen : bayesiannetworks. The second chapter gives a more precise presentation of bayesian networks and theirpossible extensions in the context of process monitoring. After that, a state of the art of diagnosisand monitoring methods with bayesian networks is studied. The third chapter explains the con-tributions given to the topic of process monitoring with bayesian networks. These contributionsare in three groups : detection, supervised diagnosis and non-supervised diagnosis. Based onthese contributions, a complete structure of a bayesian network dedicated to process monitoringis given. The last chapter presents an application of the proposed method on a benchmark prob-lem : the Tennessee Eastman Process. Efficiency of the network is evaluated for detection andfor supervised and non-supervised diagnosis. Finally, conclusions and outlooks of the proposedapproach are given.