Résumé
This thesis considers the problem of model change detection in complex dynamic systems. The goal is to develop statistical methods able to detect possible change of parameters in the model describing the system, while keeping a low rate of false alarms. This type of method is applied to the detection of anomaly or failure in many systems (navigation system, quality control ...).<br />The methods developed take into account the characteristics of biotechnological processes, which are the main application of this work. Thus, the development of a CUSUM type procedure, based on estimation of conditional likelihoods enable to treat, first, the case where a part of the model is unknown by using a nonparametric approach to estimate this component, and second, the case frequently met in practice where the system is observed indirectly. In this second case, approaches such as particle filtering are used.<br />Several results of optimality under classical constraints are established for the proposed approaches which are applied to a real problem, a bioreactor for wastewater treatment.