Robustness of the closed-loop system has repercussions on both stability and performance, making the study of robustness very important. Fundamentally, the performance and stability of closed-loop systems utilizing state-feedback are tied to that of the observers. The primary goal of this thesis is to develop a robust nonlinear observer and closely examine the usefulness of the observer in the presence of non-collocation and parametric uncertainty and as an integral component in closed-loop control. The usefulness of the observer being investigated depends on robustness, accuracy, computational burden, tunability, ease of design, and ease of implementation on an actual flexible motion system. The design and subsequent integration of the Kalman filter, an optimal observer, into a closed-loop system is well known and systematic. However, there are shortcomings of the Kalman filter in the presence of model uncertainty which are highlighted in this work. Simulation studies are conducted using the Simulation Module in National Instruments LabVIEW 8.5 and experiments are conducted on a physical system consisting of a single flexible link with non-collocation of actuators and sensors using LabVIEW Real Time 8.5. Simulations serve as a means to analyze the performance of the optimal observer and the robust observer by analyzing their dynamic behavior as well as that of the closed-loop system with each observer in place. The focus of experiments is on investigating implementation of the robust observer, including initialization and tuning of observer design parameters off-line and on-line. Simulations verify the robustness properties of the sliding mode observer while experiments show that the robust observer can be implemented at fast control rates and that replacing the Kalman filter with a robust observer has direct ramifications on closed-loop performance.

Dans ce travail, des méthodes de diagnostics et des techniques de développement de modèles sous-maille sont proposées pour la simulation des grandes échelles (SGE) du mélange turbulent. Plusieurs modèles sous-maille issus de ces stratégies sont ainsi présentés pour illustrer ces méthodes.Le principe de la SGE est de résoudre les grandes échelles de l'écoulement responsables des transferts principaux et de modéliser l'action des petites échelles de l'écoulement sur les échelles résolues. Au cours de ce travail, nous nous sommes appuyés sur le classement des modèles sous-maille en deux catégories. Les modèles "fonctionnels" qui s'attachent à reproduire les transferts énergétiques entre les échelles résolues et les échelles modélisées et les modèles "structurels" qui cherchent à bien reproduire le terme sous-maille. Le premier enjeu important a été d'évaluer la performance des modèles sous-maille en prenant en compte leur comportement à la fois fonctionnel (capacité à reproduire les transferts d'énergie) et structurel (capacité à reproduire le terme sous-maille exact). Des diagnosctics des modèles sous-maille ont pu être conduits avec l'utilisation de la notion d'estimateur optimal ce qui permet de connaitre le potentiel d'amélioration structurelle des modèles. Ces principes ont dans un premier temps servi au développement d'une première famille de modèles sous-maille algébrique appelée DRGM pour "Dynamic Regularized Gradient Model". Cette famille de modèles s'appuie sur le diagnostic structurel des termes issus de la régularisation des modèles de la famille du gradient. D'après les tests menés, cette nouvelle famille de modèle structurel a de meilleures performances fonctionnelles et structurelles que les modèles de la famille du gradient. L'amélioration des performances fonctionnelles consiste à supprimer la prédiction excessive de transferts inverses d'énergie (backscatter) observés dans les modèles de la famille du gradient. Cela permet ainsi de supprimer le comportement instable classiquement observé pour cette famille de modèles. La suite de ce travail propose ensuite d'utiliser l'estimateur optimal directement comme modèle sous-maille. Comme l'estimateur optimal fournit le modèle ayant la meilleure performance structurelle pour un jeu de variables donné, nous avons recherché le jeu de variable optimisant cette performance. Puisque ce jeu comporte un nombre élevé de variables, nous avons utilisé les fonctions d'approximation de type réseaux de neurones pour estimer cet estimateur optimal. Ce travail a mené au nouveau modèle substitut ANNM pour "Artificial Neural Network Model". Ces fonctions de substitution se construisent à partir de bases de données servant à émuler les termes exacts nécessaire à la détermination de l'estimateur optimal. Les tests de ce modèle ont montré qu'il avait de très bonnes perfomances pour des configurations de simulation peu éloignées de la base de données servant à son apprentissage, mais qu'il pouvait manquer d'universalité. Pour lever ce dernier verrou, nous avons proposé une utilisation hybride des modèles algébriques et des modèles de substitution à base de réseaux de neurones. La base de cette nouvelle famille de modèles ACM pour "Adaptative Coefficient Model" s'appuie sur les décompositions vectorielles et tensorielles des termes sous-maille exacts. Ces décompositions nécessitent le calcul de coefficients dynamiques qui sont modélisés par les réseaux de neurones. Ces réseaux bénéficient d'une méthode d'apprentissage permettant d'optimiser directement les performances structurelles et fonctionnelles des modèles ACM. Ces modèles hybrides allient l'universalité des modèles algébriques avec la performance élevée mais spécialisée des fonctions de substitution. Le résultat conduit à des modèles plus universels que l'ANNM. / This work develops subgrid model techniques and proposes methods of diagnosis for Large Eddy Simulation (LES) of turbulent mixing.Several models from these strategies are thus presented to illustrate these methods.The principle of LES is to solve the largest scales of the turbulent flow responsible for major transfers and to model the action of small scales of flowon the resolved scales. Formally, this operation leads to filter equations describing turbulent mixing. Subgrid terms then appear and must bemodeled to close the equations. In this work, we rely on the classification of subgrid models into two categories. "Functional" models whichreproduces the energy transfers between the resolved scales and modeled scales and "Structural" models that seek to reproduce the exact subgrid termitself. The first major challenge is to evaluate the performance of subgrid models taking into account their functional behavior (ability to reproduce theenergy transfers) and structural behaviour (ability to reproduce the term subgrid exactly). Diagnostics of subgrid models have been enabled with theuse of the optimal estimator theory which allows the potential of structural improvement of the model to be evaluated.These methods were initially involved for the development of a first family of models called algebraic subgrid $DRGM$ for "Dynamic Regularized GradientModel". This family of models is based on the structural diagnostic of terms given by the regularization of the gradient model family.According to the tests performed, this new structural model's family has better functional and structural performance than original model's family of thegradient. The improved functional performance is due to the vanishing of inverse energy transfer (backscatter) observed in models of thegradient family. This allows the removal of the unstable behavior typically observed for this family of models.In this work, we then propose the use of the optimal estimator directly as a subgrid scale model. Since the optimal estimator provides the modelwith the best structural performance for a given set of variables, we looked for the set of variables which optimize that performance. Since this set of variablesis large, we use surrogate functions of artificial neural networks type to estimate the optimal estimator. This leads to the "Artificial Neural Network Model"(ANNM). These alternative functions are built from databases in order to emulate the exact terms needed to determine the optimal estimator. The tests of this modelshow that he it has very good performance for simulation configurations not very far from its database used for learning, so these findings may fail thetest of universality.To overcome this difficulty, we propose a hybrid method using an algebraic model and a surrogate model based on artificial neural networks. Thebasis of this new model family $ACM$ for "Adaptive Coefficient Model" is based on vector and tensor decomposition of the exact subgrid terms. Thesedecompositions require the calculation of dynamic coefficients which are modeled by artificial neural networks. These networks have a learning method designedto directlyoptimize the structural and functional performances of $ACM$. These hybrids models combine the universality of algebraic model with high performance butvery specialized performance of surrogate models. The result give models which are more universal than ANNM.

Estudamos o problema de obtenção de estimadores/preditores ótimos para combinações lineares de respostas coletadas de uma população finita por meio de amostragem aleatória simples. Nesse contexto, estendemos o modelo misto para populações finitas proposto por Stanek, Singer & Lencina (2004, Journal of Statistical Planning and Inference) para casos em que se incluem erros de medida (endógenos e exógenos) e informação auxiliar. Admitindo que as variâncias são conhecidas, mostramos que os estimadores/preditores propostos têm erro quadrático médio menor dentro da classe dos estimadores lineares não viciados. Por meio de estudos de simulação, comparamos o desempenho desses estimadores/preditores empíricos, i.e., obtidos com a substituição das componentes de variância por estimativas, com aquele de competidores tradicionais. Também, estendemos esses modelos para análise de estudos com estrutura do tipo pré-teste/pós-teste. Também por intermédio de simulação, comparamos o desempenho dos estimadores empíricos com o desempenho do estimador obtido por meio de técnicas clássicas de análise de medidas repetidas e com o desempenho do estimador obtido via análise de covariância por meio de mínimos quadrados, concluindo que os estimadores/ preditores empíricos apresentaram um menor erro quadrático médio e menor vício. Em geral, sugerimos o emprego dos estimadores/preditores empíricos propostos para dados com distribuição assimétrica ou amostras pequenas. / We consider optimal estimation of finite population parameters with data obtained via simple random samples. In this context, we extend a finite population mixed model proposed by Stanek, Singer & Lencina (2004, Journal of Statistical Planning and Inference) by including measurement errors (endogenous or exogenous) and auxiliary information. Assuming that variance components are known, we show that the proposed estimators/predictors have the smallest mean squared error in the class of unbiased estimators. Using simulation studies, we compare the performance of the empirical estimators/predictors obtained by replacing variance components with estimates with the performance of a traditional estimator. We also extend the finite population mixed model to data obtained via pretest-posttest designs. Through simulation studies, we compare the performance of the empirical estimator of the difference in gain between groups with the performance of the usual repeated measures estimator and with the performance of the usual analysis of covariance estimator obtained via ordinary least squares. The empirical estimator has smaller mean squared error and bias than the alternative estimators under consideration. In general, we recommend the use of the proposed estimators/ predictors for either asymmetric response distributions or small samples.

Estudamos o problema de obtenção de estimadores/preditores ótimos para combinações lineares de respostas coletadas de uma população finita por meio de amostragem aleatória simples. Nesse contexto, estendemos o modelo misto para populações finitas proposto por Stanek, Singer & Lencina (2004, Journal of Statistical Planning and Inference) para casos em que se incluem erros de medida (endógenos e exógenos) e informação auxiliar. Admitindo que as variâncias são conhecidas, mostramos que os estimadores/preditores propostos têm erro quadrático médio menor dentro da classe dos estimadores lineares não viciados. Por meio de estudos de simulação, comparamos o desempenho desses estimadores/preditores empíricos, i.e., obtidos com a substituição das componentes de variância por estimativas, com aquele de competidores tradicionais. Também, estendemos esses modelos para análise de estudos com estrutura do tipo pré-teste/pós-teste. Também por intermédio de simulação, comparamos o desempenho dos estimadores empíricos com o desempenho do estimador obtido por meio de técnicas clássicas de análise de medidas repetidas e com o desempenho do estimador obtido via análise de covariância por meio de mínimos quadrados, concluindo que os estimadores/ preditores empíricos apresentaram um menor erro quadrático médio e menor vício. Em geral, sugerimos o emprego dos estimadores/preditores empíricos propostos para dados com distribuição assimétrica ou amostras pequenas. / We consider optimal estimation of finite population parameters with data obtained via simple random samples. In this context, we extend a finite population mixed model proposed by Stanek, Singer & Lencina (2004, Journal of Statistical Planning and Inference) by including measurement errors (endogenous or exogenous) and auxiliary information. Assuming that variance components are known, we show that the proposed estimators/predictors have the smallest mean squared error in the class of unbiased estimators. Using simulation studies, we compare the performance of the empirical estimators/predictors obtained by replacing variance components with estimates with the performance of a traditional estimator. We also extend the finite population mixed model to data obtained via pretest-posttest designs. Through simulation studies, we compare the performance of the empirical estimator of the difference in gain between groups with the performance of the usual repeated measures estimator and with the performance of the usual analysis of covariance estimator obtained via ordinary least squares. The empirical estimator has smaller mean squared error and bias than the alternative estimators under consideration. In general, we recommend the use of the proposed estimators/ predictors for either asymmetric response distributions or small samples.