Skip to Main content Skip to Navigation

Identification de systèmes dynamiques non linéaires à l'aide de représentations multi-modèles

Abstract : The problem of system modelling is classical for process control. The aim is to present, with an optimal accuracy, the system behaviour. In general, real processes are non linear, multi-variables and time-varying parameters. It is therefore quite difficult to geta global system representation so that it is valid for any operating regime. The multi-models approach is based on the design of several quite simple models, called local models. Each local model is valid for a given operating point and around this point in a region of the operating domain defined within a weight function. These local models are further integrated using a weight expression in order to get an algebraic expression. This expression links the process inputs and outputs so that a global description is obtained. Many multi-model structures can be defined in order to describe the actual behaviour of more complex processes but, whatever the strategy is, the main problem relies on the increasing number of parameters to be identified. That is why, we develop a multi-model structure called "Hammerstein generalisé" (generalized Hammerstein) that allows us to get a smaller (in the number of parameters sense) representation of a given process. Such a multi-models structure depends , m parameter within non linear functions and we propose accordingly algorithms in order to estimate such parameters. In particular, we develop an iterative global method based upon sensitivity functions so that ail the model parameters are obtained: the parameters of the weight functions, of the local models and of the dynamic part. The previous global estimation method is applied on a multi-variables activated sludge process. We then compare the performances and the robustness of the developed identification algorithm on a simulated example. Finally, an esearch of the optimal multi-model structure of the process is performed with the search of the more significant inputs of the process behaviour, the optimal number of local models but also of the order of the dynamic model based on an extension of statistical tools already available for linear models.
Document type :
File URL :
Complete list of metadata
Contributor : Thèses Ul <>
Submitted on : Thursday, March 29, 2018 - 12:49:57 PM
Last modification on : Wednesday, January 13, 2021 - 11:26:01 AM


  • HAL Id : tel-01750589, version 1



Corinne Loverini. Identification de systèmes dynamiques non linéaires à l'aide de représentations multi-modèles. Sciences de la Terre. Institut National Polytechnique de Lorraine, 1999. Français. ⟨NNT : 1999INPL041N⟩. ⟨tel-01750589⟩



Record views