Identifying the parameters with the largest influence on the predicted outputs of a model revealswhich parameters need to be known more precisely to reduce the overall uncertainty on themodel output. A large improvement of such models would result when uncertainties in the keymodel parameters are reduced. To achieve this, new experiments could be very helpful,especially if the measurements are taken at the spatio-temporal locations that allow estimate the parameters in an optimal way. After evaluating the methodologies available for optimal sensor location, a few observations were drawn. The method based on the Gram determinant evolution can report results not according to what should be expected. This method is strongly dependent of the sensitivity coefficients behaviour. The approach based on the maximum angle between subspaces, in some cases, produced more that one optimal solution. It was observed that this method depends on the magnitude of outputs values and report the measurement positions where the outputs reached their extrema values. The D-optimal design method produces number and locations of the optimal measurements and it depends strongly of the sensitivity coefficients, but mostly of their behaviours. In general it was observed that the measurements should be taken at the locations where the extrema values (sensitivity coefficients, POD modes and/or outputs values) are reached. Further improvements can be obtained when a reduced model of the system is employed. This is computationally less expensive and the best estimation of the parameter is obtained, even with experimental data contaminated with noise. A new approach to calculate the time coefficients belonging to an empirical approximator based on the POD-modes derived from experimental data is introduced. Additionally, an artificial neural network can be used to calculate the derivatives but only for systems without complex nonlinear behaviour. The latter two approximations are very valuable and useful especially if the model of the system is unknown.

Infinite-Dimensional Systems (IDSs) which have been made possible by recent advances in mathematical and computational tools can be used to model complex real phenomena. However, due to physical, economic, or stringent non-invasive constraints on real systems, the underlying characteristics for mathematical models in general (and IDSs in particular) are often missing or subject to uncertainty. Therefore, developing efficient estimation techniques to extract missing pieces of information from available measurements is essential. The human brain is an example of IDSs with severe constraints on information collection from controlled experiments and invasive sensors. Investigating the intriguing modeling potential of the brain is, in fact, the main motivation for this work. Here, we will characterize the hemodynamic behavior of the brain using functional magnetic resonance imaging data. In this regard, we propose efficient estimation methods for two classes of IDSs, namely Partial Differential Equations (PDEs) and Fractional Differential Equations (FDEs). This work is divided into two parts. The first part addresses the joint estimation problem of the state, parameters, and input for a coupled second-order hyperbolic PDE and an infinite-dimensional ordinary differential equation using sampled-in-space measurements. Two estimation techniques are proposed: a Kalman-based algorithm that relies on a reduced finite-dimensional model of the IDS, and an infinite-dimensional adaptive estimator whose convergence proof is based on the Lyapunov approach. We study and discuss the identifiability of the unknown variables for both cases. The second part contributes to the development of estimation methods for FDEs where major challenges arise in estimating fractional differentiation orders and non-smooth pointwise inputs. First, we propose a fractional high-order sliding mode observer to jointly estimate the pseudo-state and input of commensurate FDEs. Second, we propose a modulating function-based algorithm for the joint estimation of the parameters and fractional differentiation orders of non-commensurate FDEs. Sufficient conditions ensuring the local convergence of the proposed algorithm are provided. Subsequently, we extend the latter technique to estimate smooth and non-smooth pointwise inputs. The performance of the proposed estimation techniques is illustrated on a neurovascular-hemodynamic response model. However, the formulations are efficiently generic to be applied to a wide set of additional applications.

[pt] Objetiva-se, nesse trabalho, formular e testar numericamente uma perspectiva computacional, baseada em elementos de controle robusto e indices de desempenho H2, para a sintese de controladores de dimensão finita (DF) para sistemas lineares de dimensão infinita (DI) correspondentes a certas equações de evolução parabólicas e, especialmente, a uma versão simplificada da equação do calor. A abordagem aqui utilizada é a de usar modelos aproximantes de DF (modelos nominais) e limitantes superiores nas normas H (infinito) dos erros de aproximação correspondentes nas funções de transferência de DI em questão, de modo que o procedimento de síntese baseie-se apenas em funções de transferência racionais e os controladores resultantes sejam de DF.Mais especificamente, uma classe de controladores que asseguram a estabilidade do sistema em malha fechada envolvendo o sistema de DI em questão e definida tomando-se as soluções ótimas de problemas H2 (infinito) nos quais o funcional de custo (H2) nominal é minimizado sobre os controladores nominalmente estabilizantes, sob uma restrição (H infinito) de margem de estabilidade mínima definida por um parâmetro de projeto escalar denotado por micro. A obtenção de um controlador é então feita pela escolha do valor de micro de modo a minimizar um limitante superior (calculado apenas com base em funçõess racionais) sobre o funcional de custo calculado no sistema de DI original. Esse procedimento é ilustrado por exemplos numéricos envolvendo a versão simplificada da equação do calor. / [en] A computational perspective based on robust control tools is presented for the H2 synthesis of finite-dimensional controllers for linear, stable, distributed-parameter systems corresponding to certain evolution equations. The approach pursued here relies on finite-dimensional approximations and error bounds on the H (infinite) norms of the corresponding errors on transfer functions so that the resulting synthesis procedure solely depends on rational transfer functions, thereby yielding finite-dimensional controllers. More specifically, a class of stabilizing controllers for a given infinite-dimensional system is defined taking optimal solutions of H2 / H (infinite) problems - i.e., a nominal H2 cost is minimized over controllers which satisfy a nominal stability margin defined by a scalar parameter micro. A controller is then obtained by choosing micro in such a way as to minimize an upper bound on the value taken by the cost functional on the original infinite-dimensional system. This procedure is illustrated by simple numerical examples involving the (simplified) heat equation in one dimension.

Du fait du manque de capteurs fiables et bon marché, la problématique de la reconstruction d'état est cruciale pour les bioprocédés, particulièrement pour les (bio)réacteurs à paramètres répartis. C'est pourquoi il est essentiel d'étudier leurs propriétés d'observabilité, qui peuvent notamment dépendre des dimensions spatiales des capteurs. Dans cette thèse, on étudie tout d'abord l'observabilité d'une classe de réacteurs tubulaires à lit fixe mettant en oeuvre une seule réaction biochimique, où un substrat est dégradé suivant une cinétique non linéaire. Plus précisément, l'analyse porte sur un modèle linéarisé, consistant notamment en une Équation aux Dérivées Partielles (EDP) parabolique linéaire comportant un coefficient non uniforme (c.-à-d. dépendant de la variable spatiale). Ce modèle entre dans le cadre d'une classe particulière de systèmes : les systèmes de Sturm-Liouville. Il s'en déduit que tout nombre fini de modes dominants du système est observable par un capteur (de concentration en substrat) situé en sortie et de largeur suffisamment petite. En outre, considérant un minorant et un majorant du coefficient non uniforme, on détermine une expression numérique, fonction du nombre de modes à observer, qui minore cette largeur. La pertinence de ce résultat est confirmée par un exemple numérique : un biofiltre de dénitrification. Cette étude est étendue à un procédé-pilote de digestion anaérobie de l'INRA-Narbonne. On montre l'existence d'un état d'équilibre, autour duquel le comportement du système est modélisé par deux EDP linéaires, dont une à coefficient non uniforme. La démarche précédente est généralisée et on calcule des expressions des largeurs de deux capteurs en fonction du nombre de modes dont on veut s'assurer qu'ils sont observables. Ce résultat s'applique notamment à la conception d'un estimateur d'état.

Grey-box models are constructed by combining model components that are derived from first principles with components that are identified empirically from data. In this thesis a grey-box modelling method for describing distributed parameter systems is presented. The method combines partial differential equations with a multi-layer perceptron network in order to incorporate prior knowledge about the system while identifying unknown dynamics from data. A gradient-based optimization scheme which relies on the reverse mode of automatic differentiation is used to train the network. The method is presented in the context of modelling the dynamics of a chemical reaction in a fluid. Lastly, the grey-box modelling method is evaluated on a one-dimensional and two-dimensional instance of the reaction system. The results indicate that the grey-box model was able to accurately capture the dynamics of the reaction system and identify the underlying reaction. / Hybridmodeller konstrueras genom att kombinera modellkomponenter som härleds från grundläggande principer med modelkomponenter som bestäms empiriskt från data. I den här uppsatsen presenteras en metod för att beskriva distribuerade parametersystem genom hybridmodellering. Metoden kombinerar partiella differentialekvationer med ett neuronnätverk för att inkorporera tidigare känd kunskap om systemet samt identifiera okänd dynamik från data. Neuronnätverket tränas genom en gradientbaserad optimeringsmetod som använder sig av bakåt-läget av automatisk differentiering. För att demonstrera metoden används den för att modellera kemiska reaktioner i en fluid. Metoden appliceras slutligen på ett en-dimensionellt och ett två-dimensionellt exempel av reaktions-systemet. Resultaten indikerar att hybridmodellen lyckades återskapa beteendet hos systemet med god precision samt identifiera den underliggande reaktionen.