This research work focuses at first on the intelligent model development for process state (special for fault) detection, behavior prediction and process control for complex industrial processes. In the model architecture, Fuzzy Neural Networks (FNNs) are employed as process state classifiers for process state (fault) detection; other (different) Neural Networks (NNs) models are applied for system identification of process characteristics in different process states. The model detects process states (faults) and predicts process behavior according to process input and historical behaviors, whose combination of influences generates the final results of process state (fault) detection and quantitative prediction. The whole model is constructed based on Fuzzy TS NARX models. Secondly, an optimal model is designed to two purposes, one is for optimal process diagnosis and another is for optimal prediction. To time varying processes, an adaptive strategy and algorithm, applying the Least Squares algorithm, has been developed for model adaptability to cover time depending process changes. Thirdly, a specific state space equation of discrete time varying system is being derived from the model. In the state space equation, the state transition matrix A is determined by the fuzzy degree of process state classification produced by process historical behavior in time t instant, and the input transition matrix B by process real input in time t instant. The state observer vector H is determined by optimization results generated by model adaptive or optimal scheme. Finally, to confirm the validity of the theoretical results from above, an application case has been studied for supply forecasting. The study and application results indicate that the model not only has good performance for fault detection, but also provides excellent quantitative prediction of process output. It can be applied in process state (fault) detection, diagnosis and prediction for process behavior, as well as fault predictive control.

Cette thèse porte sur la caractérisation de structures mécaniques, au travers de leurs paramètres structuraux, à partir d'observations perturbées par des bruits de mesure, supposés additifs blancs gaussiens et centrés. Pour cela, nous proposons d'utiliser des modèles à temps discret à parties linéaire et non-linéaire séparables. La première permet de retrouver les paramètres recherchés tandis que la seconde renseigne sur la non-linéarité présente. Dans le cadre d'une modélisation non-récursive par des séries de Volterra, nous présentons une approche à erreurs-dans-les-variables lorsque les variances des bruits ne sont pas connues ainsi qu'un algorithme adaptatif du type LMS nécessitant la connaissance de la variance du bruit d'entrée. Dans le cadre d'une modélisation par un modèle récursif polynomial, nous proposons deux méthodes à partir d'algorithmes évolutionnaires. La première inclut un protocole d'arrêt tenant compte de la variance du bruit de sortie. Dans la seconde, les fonctions fitness sont fondées sur des fonctions de corrélation dans lesquelles l'influence du bruit est supprimée ou compensée. / This PhD deals with the caracterisation of mechanical structures, by its structural parameters, when only noisy observations disturbed by additive measurement noises, assumed to be zero-mean white and Gaussian, are available. For this purpose, we suggest using discrete-time models with distinct linear and nonlinear parts. The first one allows the structural parameters to be retrieved whereas the second one gives information on the nonlinearity. When dealing with non-recursive Volterra series, we propose an errors-in-variables (EIV) method to jointly estimate the noise variances and the Volterra kernels. We also suggest a modified unbiased LMS algorithm to estimate the model parameters provided that the input-noise variance is known. When dealing with recursive polynomial model, we propose two methods using evolutionary algorithms. The first includes a stop protocol that takes into account the output-noise variance. In the second one, the fitness functions are based on correlation criteria in which the noise influence is removed or compensated.