Cette thèse traite de l’analyse et de la synthèse de multimodèles pour la simplification de modèles, l’estimation d’état et le diagnostic des systèmes non linéaires caractérisés par une ou plusieurs échelles de temps. Ces travaux visent, dans un premier temps, à développer une procédure systématique de transformation d’un système non-linéaire en le récrivant sous une forme multimodèle, en évitant quelques inconvénients majeurs : la transformation est réalisée sans perte d’information, le choix de différents points de fonctionnement n’est plus nécessaire, le choix de variables de prémisse est réalisé d’une façon systématique. De plus, la méthode offre le choix entre différents multimodèles. Ce degré de liberté sera utilisé pour faciliter les études de contrôlabilité, d’observabilité et d’analyse de stabilité. Dans un deuxième temps, l’obtention de la forme à perturbations singulières d’un système non linéaire est proposée, en éliminant quelques contraintes structurelles et en rendant l’identification et la séparation des échelles de temps indépendante de la structure du modèle. Dans un troisième temps, la synthèse de plusieurs observateurs robustes vis-à-vis des perturbations, des erreurs de modélisation et des entrées inconnues a été réalisée afin dereconstruire l’état et l’entrée inconnue du système. La difficulté de cette étude provient du fait que le multimodèle utilisé dépend de variables de prémisse non mesurables, situation qui n’est pas intensivement étudiée, alors qu’elle est naturellement issue de l’approche par transformation système non linéaire!multimodèle. Ensuite, le diagnostic de défauts de systèmes est réalisé au moyen de bancs d’observateur à entrées inconnues permettant la génération et la structuration de résidus indicateurs de défauts. Finalement, tous les travaux proposés sont appliqués au modèle d’une station d’´epuration, Activated Sludge Model No.1, qui est largement utilisé dans le domaine du traitement des eaux usées / This thesis deals with analysis and synthesis of multiple model structures for model simplification, state estimation and diagnosis of nonlinear systems represented by one or several time-scales. This work aims, at first, to develop a systematic procedure to transform a nonlinear system into a multiple model form, by avoiding some major drawbacks : the transformation causes no information loss, the choice of the different operating points is no more necessary, the choice of the premise variables is realized in a more systematic way. Furthermore, the method gives the possibility of choosing between different multiplemodel structures. This degree of freedom will be used to ease the controllability, observ-ability, stability analysis studies. Secondly, the derivation of a singularly perturbed form for a multiple time scale non linear system is proposed, by eliminating some structuralconstraints and by making the identification and the separation of the time-scales independent to the model structure. Thirdly, the robust observer synthesis with respect to perturbations, modeling errors and unknown inputs are presented for state and unknowninput estimation. The difficulty of these studies comes from the fact that the multiple model depends on unmeasurable premise variables, this case being not intensively studied, whereas it results naturally from the method of transformation nonlinear system - multiple model. Afterward, fault diagnosis is performed using banks of observer to generate andstructure residual signals. Finally, this works are applied to a model of wastewater treatment plant, Activated Sludge Model No.1 (ASM1) that is largely used in the concerned fiel

In this thesis study, single target tracking algorithms including IMM-PDA and IMM-IPDA algorithms / Optimal approaches in multitarget tracking including IMM-JPDA, IMM-IJPDA and IMM-JIPDA algorithms and an example of Linear Multi-target approaches in multitarget tracking including IMM-LMIPDA algorithm have been studied and implemented in MATLAB for comparison. Simulations were carried out in various realistic test scenarios including single target tracking, tracking of multiple targets moving in convoy fashion, two targets merging in a junction, two targets merging-departing in junctions and multitarget tracking under isolated tracks situations. RMSE performance, track loss and computational load evaluations were done for these algorithms under the test scenarios dealing with these situations. Benchmarkings are presented relying on these outcomes.

Target tracking is one the most fundamental elements of a radar system. The aim of target tracking is the reliable estimation of a target&#039 / s true state based on a time history of noisy sensor observations. In real life, the sensor data may include substantial noise. This noise can render the raw sensor data unsuitable to be used directly. Instead, we must filter the noise, preferably in an optimal manner. For land, air and surface marine vehicles, very successful filtering methods are developed. However, because of the significant differences in the underwater propagation environment and the associated differences in the corresponding sensors, the successful use of similar principles and techniques in an underwater scenario is still an active topic of research. A comparative study of the effects of the underwater environment on a number of tracking algorithms is the focus of the present thesis. The tracking algorithms inspected are: the Kalman Filter, the Extended Kalman Filter and the Particle Filter. We also investigate in particular the IMM extension to KF and EKF filters. These algorithms are tested under several underwater environment scenarios.

Manufacturing equipments in semiconductor factories (fabs) provide abundant data and opportunities for data-driven process monitoring and modeling. In particular, virtual metrology (VM) is an active area of research. Traditional monitoring techniques using univariate statistical process control charts do not provide immediate feedback to quality excursions, hindering the implementation of fab-wide advanced process control initiatives. VM models or inferential sensors aim to bridge this gap by predicting of quality measurements instantaneously using tool fault detection and classification (FDC) sensor measurements. The existing research in the field of inferential sensor and VM has focused on comparing regressions algorithms to demonstrate their feasibility in various applications. However, two important areas, data pretreatment and post-deployment model maintenance, are usually neglected in these discussions. Since it is well known that the industrial data collected is of poor quality, and that the semiconductor processes undergo drifts and periodic disturbances, these two issues are the roadblocks in furthering the adoption of inferential sensors and VM models. In data pretreatment, batch data collected from FDC systems usually contain inconsistent trajectories of various durations. Most analysis techniques requires the data from all batches to be of same duration with similar trajectory patterns. These inconsistencies, if unresolved, will propagate into the developed model and cause challenges in interpreting the modeling results and degrade model performance. To address this issue, a Constrained selective Derivative Dynamic Time Warping (CsDTW) method was developed to perform automatic alignment of trajectories. CsDTW is designed to preserve the key features that characterizes each batch and can be solved efficiently in polynomial time. Variable selection after trajectory alignment is another topic that requires improvement. To this end, the proposed Moving Window Variable Importance in Projection (MW-VIP) method yields a more robust set of variables with demonstrably more long-term correlation with the predicted output. In model maintenance, model adaptation has been the standard solution for dealing with drifting processes. However, most case studies have already preprocessed the model update data offline. This is an implicit assumption that the adaptation data is free of faults and outliers, which is often not true for practical implementations. To this end, a moving window scheme using Total Projection to Latent Structure (T-PLS) decomposition screens incoming updates to separate the harmless process noise from the outliers that negatively affects the model. The integrated approach was demonstrated to be more robust. In addition, model adaptation is very inefficient when there are multiplicities in the process, multiplicities could occur due to process nonlinearity, switches in product grade, or different operating conditions. A growing structure multiple model system using local PLS and PCA models have been proposed to improve model performance around process conditions with multiplicity. The use of local PLS and PCA models allows the method to handle a much larger set of inputs and overcome several challenges in mixture model systems. In addition, fault detection sensitivities are also improved by using the multivariate monitoring statistics of these local PLS/PCA models. These proposed methods are tested on two plasma etch data sets provided by Texas Instruments. In addition, a proof of concept using virtual metrology in a controller performance assessment application was also tested.

An analysis and simulation tool is developed for optimizing system parameters of the monopulse target tracking radar and observing effects of the system parameters on the performance of the system over different scenarios. A monopulse tracking radar is modeled for measuring the performance of the radar with given parameters, during the thesis studies. The radar model simulates the operation of a Class IA type monopulse automatic tracking radar, which uses a planar phased array. The interacting multiple model (IMM) estimator with the Probabilistic Data Association (PDA) technique is used as the tracking filter. In addition to modeling of the tracking radar model, an optimization tool is developed to optimize system parameters of this tracking radar model. The optimization tool implements a Genetic Algorithm (GA) belonging to a GA Toolbox distributed by Department of Automatic Control and System Engineering at University of Sheffield. The thesis presents optimization results over some given optimization scenarios and concludes on effect of tracking filter parameters, beamwidth and dwell interval for the confirmed track.

abstract: Tracking targets in the presence of clutter is inevitable, and presents many challenges. Additionally, rapid, drastic changes in clutter density between different environments or scenarios can make it even more difficult for tracking algorithms to adapt. A novel approach to target tracking in such dynamic clutter environments is proposed using a particle filter (PF) integrated with Interacting Multiple Models (IMMs) to compensate and adapt to the transition between different clutter densities. This model was implemented for the case of a monostatic sensor tracking a single target moving with constant velocity along a two-dimensional trajectory, which crossed between regions of drastically different clutter densities. Multiple combinations of clutter density transitions were considered, using up to three different clutter densities. It was shown that the integrated IMM PF algorithm outperforms traditional approaches such as the PF in terms of tracking results and performance. The minimal additional computational expense of including the IMM more than warrants the benefits of having it supplement and amplify the advantages of the PF. / Dissertation/Thesis / Masters Thesis Electrical Engineering 2015

Made available in DSpace on 2014-12-17T14:55:11Z (GMT). No. of bitstreams: 1 CarlosAGF_TESE.pdf: 1739972 bytes, checksum: 7401db4e68ede642dc9d65e00bd935e6 (MD5) Previous issue date: 2012-12-21 / In this work a modification on ANFIS (Adaptive Network Based Fuzzy Inference System) structure is proposed to find a systematic method for nonlinear plants, with large operational range, identification and control, using linear local systems: models and controllers. This method is based on multiple model approach. This way, linear local models are obtained and then those models are combined by the proposed neurofuzzy structure. A metric that allows a satisfactory combination of those models is obtained after the structure training. It results on plant s global identification. A controller is projected for each local model. The global control is obtained by mixing local controllers signals. This is done by the modified ANFIS. The modification on ANFIS architecture allows the two neurofuzzy structures knowledge sharing. So the same metric obtained to combine models can be used to combine controllers. Two cases study are used to validate the new ANFIS structure. The knowledge sharing is evaluated in the second case study. It shows that just one modified ANFIS structure is necessary to combine linear models to identify, a nonlinear plant, and combine linear controllers to control this plant. The proposed method allows the usage of any identification and control techniques for local models and local controllers obtaining. It also reduces the complexity of ANFIS usage for identification and control. This work has prioritized simpler techniques for the identification and control systems to simplify the use of the method / Neste trabalho prop?e-se uma modifica??o na estrutura neurofuzzy ANFIS (Adaptive Network Based Fuzzy Inference System) para a obten??o de um m?todo sistem?tico para identifica??o e controle de plantas com ampla faixa de opera??o e n?o linearidade acentuada, a partir de t?cnicas lineares de identifica??o e controle. Este m?todo se baseia na metodologia de m?ltiplos modelos. Dessa forma, obt?m-se modelos lineares locais e esses s?o combinados pela estrutura neurofuzzy proposta. Uma m?trica que permite combinar adequadamente esses modelos ? obtida ap?s o treinamento dessa estrutura, resultando na identifica??o global da planta. Para cada um desses modelos ? projetado um controlador. O controle global ? obtido a partir da combina??o dos sinais dos controladores locais. Essa mistura ? feita pelo ANFIS modificado. A modifica??o na arquitetura do ANFIS permite o compartilhamento do conhecimento adquirido pelo treinamento da estrutura empregada na combina??o de modelos locais. Assim n?o se faz necess?rio o treinamento da estrutura empregada na mistura de controladores. Avaliaram-se as estruturas modificadas atrav?s de dois estudos de caso. Verificou-se que ? poss?vel treinar apenas um ANFIS, para a obten??o de uma m?trica que permita a combina??o adequada dos modelos lineares, v?lidos localmente, e essa estrutura, j? ajustada, pode ser aplicada na combina??o de controladores lineares, projetados para cada um dos modelos, resultando em um sistema de controle que satisfaz as especifica??es de desempenho previamente estabelecidas. O m?todo proposto possibilita a utiliza??o de quaisquer t?cnicas de identifica??o e controle para a obten??o dos modelos e controladores locais, e a redu??o da complexidade de utiliza??o do ANFIS para identifica??o e controle. Neste trabalho priorizaram-se as t?cnicas mais simples de identifica??o e controle de sistemas de forma a simplificar a utiliza??o do m?todo

Dans le cadre de la surveillance maritime ou terrestre par radar aéroporté, l’un des principaux objectifs est de détecter et de poursuivre une grande diversité de cibles au cours du temps.Ces traitements s’appuient généralement sur l’utilisation d’un filtre Bayésien pour estimer récursivement les paramètres cinématiques (position, vitesse et accélération) des cibles. Il repose surla représentation dans l’espace d’état du système et plus particulièrement sur la modélisation a priori de l’évolution des cibles à partir d’un modèle de mouvement (mouvement rectiligne uniforme, mouvement uniformément accéléré, mouvement rotationnel, etc.). Si les cibles pistées sont manoeuvrantes, plusieurs modèles de mouvement, chacun avec une dynamique prédéfinie,sont classiquement combinés au sein d’une structure à modèles multiples. Même si ces approches s’avèrent pertinentes, des améliorations peuvent être apportées à plusieurs niveaux, notamment sur la manière de sélectionner et définir a priori les modèles utilisés.Dans ce contexte d’étude, plusieurs problématiques doivent être traitées.1/ Lors de l’utilisation d’une structure à modèles multiples, on considère en général deux à trois modèles. Ce choix est fait lors de la phase de conception de l’algorithme selon la connaissance du système et l’expertise de l’utilisateur. Cependant, il n’existe pas à notre connaissance d’outils ou de règles permettant de définir les types de mouvement à associer et leurs paramètres.2/ Il est préférable que le choix du ou des modèles de mouvement soit cohérent avec le type de cible pisté.3/ Lorsqu’un type de mouvement est utilisé, ses paramètres sont fixés a priori mais ces valeurs ne sont pas nécessairement adaptées à toutes les phases du mouvement. L’une des difficultés majeures réside dans la manière de définir et de faire évoluer la matrice de covariance du bruit de modèle. Le travail présenté dans ce mémoire vise à proposer des solutions algorithmiques aux problématiques précédentes afin d’améliorer l’estimation des trajectoires des cibles d’intérêt.Dans un premier temps, nous établissons une mesure de dissimilarité fondée sur la divergence de Jeffrey entre deux densités de probabilité associés à deux modèles d’état différents. Celle-ci est appliquée à la comparaison de modèles de mouvement. Elle est ensuite utilisée pour comparer un ensemble de plusieurs modèles d’état. Cette étude est alors mise à profit pour proposer une méthode de sélection a priori des modèles constituant des algorithmes à modèles multiples.Puis, nous présentons des modèles Bayésiens non-paramétriques (BNP) utilisant les processus de Dirichlet pour estimer les statistiques du bruit de modèle. Cette modélisation a l’avantage de pouvoir représenter des bruits multimodaux sans avoir à spécifier a priori le nombre de modes et leurs caractéristiques. Deux cas sont traités. Dans le premier, on estime la matrice de précision du bruit de modèle d’un unique modèle de mouvement sans émettre d’a priori sur sa structure.Dans le second, nous tirons profit des formes structurelles des matrices de précision associées aux modèles de mouvement pour n’estimer qu’un nombre réduit d’hyperparamètres. Pour les deux approches, l’estimation conjointe des paramètres cinématiques de la cible et de la matrice de précision du bruit de modèle est réalisée par filtrage particulaire. Les contributions apportées sont notamment le calcul de la distribution d’importance optimale dans chacun des cas.Enfin, nous tirons profit des méthodes dites de classification et pistage conjoints (joint tracking and classification -JTC-) pour mener simultanément la classification de la cible et l’inférence de ses paramètres. Dans ce cas, à chaque classe de cible est associé un ensemble de modèles d’évolution qui lui est propre. [...] / As part of the ground or maritime surveillance by using airborne radars, one of the mainobjectives is to detect and track a wide variety of targets over time. These treatments are generallybased on Bayesian filtering to estimate recursively the kinematic parameters (position,velocity and acceleration) of the targets. It is based on the state-space representation and moreparticularly on the prior modeling of the target evolutions (uniform motion, uniformly acceleratedmotion, movement rotational, etc.). If maneuvering targets are tracked, several motionmodels, each with a predefined dynamic, are typically combined in a multiple-model structure.Although these approaches are relevant, improvements can be made at several levels, includinghow to select and define a priori the models to be used.In this framework, several issues must be addressed.1 / When using a multiple-model structure, it is generally considered two to three models. Thischoice is made in the algorithm design stage according to the system knowledge and the userexpertise. However, it does not exist in our knowledge tools or/and rules to define the types ofmotions and their associated parameters.2 / It is preferable that the choice of the motion model(s) is consistent with the type of targetto be tracked.3 / When a type of motion model is used, its parameters are fixed a priori but these values ??arenot necessarily appropriate in all phases of the movement. One of the major challenges is theway to define the covariance matrix of the model noise and to model its evolution.The work presented in this thesis consists of algorithmic solutions to the previous problemsin order to improve the estimation of target trajectories.First, we establish a dissimilarity measure based on Jeffrey divergence between probability densitiesassociated with two different state models. It is applied to the comparison of motion models.It is then used to compare a set of several state models. This study is then harnessed to providea method for selecting a priori models constituting multiple-model algorithms.Then we present non-parametric Bayesian models (BNP) using the Dirichlet process to estimatemodel noise statistics. This model has the advantage of representing multimodal noises withoutspecifying a priori the number of modes and their features. Two cases are treated. In the firstone, the model noise precision matrix is estimated for a single motion model without issue ofany a priori on its structure. In the second one, we take advantage of the structural forms ofprecision matrices associated to motion models to estimate only a small number of hyperparameters.For both approaches, the joint estimation of the kinematic parameters of the target andthe precision matrix of the model noise is led by particle filtering. The contributions includecalculating the distribution optimal importance in each case.Finally, we take advantage of methods known as joint tracking and classification (JTC) forsimultaneously leading the classification of the target and the inference of its parameters. Inthis case, each target class is associated with a set of evolution models. In order to achievethe classification, we use the target position measurements and the target extent measurementscorresponding to the projection of the target length on the line of sight radar-target. Note that this approach is applied in a single target tracking context and a multiple-target environment.

Les travaux de cette thèse portent sur l’investigation d’approches de commande permettant d’aborder la stabilisation des réseaux électriques continus.En effet les interactions entre les différents éléments du réseau : sources, filtres et charges peuvent conduire à son instabilité. Ces interactions peuvent simplement être mises en évidence au moyen d’un cas d’étude de réseau contenant des charges à puissance constante (CPLs). Pour pallier les problèmes induits par les interconnexion de ces éléments, différentes approches de commande ont été évaluées afin d’assurer la stabilité et maintenir les performances du réseau dans tout son domaine de fonctionnement. La première approche concerne la synthèse par la méthode « backstepping », qui nécessite de reformuler le modèle du réseau sous une structure cascade. Toutefois, selon le moyen d’action disponible, cette approche peut se révéler difficile à mettre en œuvre lorsque plusieurs charges à puissance sont présentes. La deuxième approche fondée sur les méthodes de passivité synthétisant une commande par « injection d’amortissement ». Cette commande permet d’ajouter un amortisseur virtuel aux filtres d’entrée des charges afin de compenser l’effet d’impédance négative introduit par la CPL. Enfin, pour proposer une solution intégrée permettant de mieux répondre à la problématique de la stabilisation du réseau, une approche fondée sur une représentation sous forme multimodèle du système a été étudiée. Cette méthode permet aussi d’envisager la synthèse d’un observateur lorsque l’ensemble du vecteur d’état n’est pas mesuré. Afin de valider et de comparer les performances des différentes méthodes de commande, un réseau électrique DC type caractérisé par deux charges de nature différente et d’un organe de stockage réversible, a été défini où le seul actionneur considéré est l’organe de stockage utilisé ici dans un contexte de stabilisation. / This work investigates control approachs for the stabilization of DC electrical networks. Interactions between different elements of a network i.e, sources, filters and loads may lead to instability. These interactions may be identified by styding a network containing constant power loads (CPLs). To address the problems caused by the interconnection of these elements, different control methods could be evaluated to ensure the stability and maintain network performances throughout its operating range. The first approach uses « backstepping »method, which requiers cascade structure models. However, according to available control input, this approach may be difficult to implement when multiple power loads are present. The second approach is based on passivity theory using « damping injection » control. This control law adds a virtual damper to the input filter loads in order to compensate the negative impedance effect introduced by the CPL. Finally, in order to provide an integrated solution for the problem of network stabilization, an approach based on multiple model representation of the system was investigated. This method also allows to consider the design of an observer when the entire state vector is not measured. To validate and compare the performance of different control methods, a DC electrical network characterized by two loads of different natures and a reversible storage device was defined. The storage device is the only control input considered for the stabilization.

Depuis 1978, le système ARGOS assure à l’échelle mondiale la collecte de données et la localisation de plateformes pour des applications liées au suivi d’animaux, à l’océanographie et à la sécurité maritime. La localisation exploite le décalage Doppler affectant la fréquence de porteuse des messages émis par les plateformes et réceptionnés par des satellites dédiés. Au cours des vingt dernières années, les puissances d’émission des plateformes se sont réduites pour des conditions d’utilisation toujours plus extrêmes, augmentant le nombre de localisations de moindre qualité. Paradoxalement, les utilisateurs ont cherché à identifier des comportements à des échelles de plus en plus petites. L’objectif de ce projet est de développer un algorithme de localisation plus performant dans le contexte actuel afin de remplacer le traitement temps réel historique basé sur un ajustement par moindres carrés. Un service hors ligne, permettant de déterminer des localisations encore plus précises, est proposé dans un second temps.Le problème est reformulé comme l’estimation de l’état d’un système dynamique stochastique, tenant compte d’un ensemble de modèles de déplacement admissibles pour les plateformes. La détermination exacte de la loi a posteriori de l’état présente alors une complexité exponentiellement croissante avec le temps. Le filtre « Interacting Multiple Model » (IMM) est devenu l’outil standard pour approximer en temps réel la loi a posteriori avec un coût de calcul constant. Pour des applications hors ligne, de nombreuses solutions sous-optimales de lissage multi-modèle ont aussi été proposées. La première contribution méthodologique de ce travail présente l’extension du cadre initial de l’IMM à un ensemble de modèles hétérogènes, c.-à-d. dont les vecteurs d’état sont de tailles et de sémantiques différentes. En outre, nous proposons une nouvelle méthode pour le lissage multi-modèle qui offre une complexité réduite et de meilleures performances que les solutions existantes. L’algorithme de localisation ARGOS a été réécrit en y incorporant le filtre IMM en tant que traitement temps réel et le lisseur multi-modèle comme service hors ligne. Une étude, menée sur un panel de 200 plateformes munies d’un récepteur GPS utilisé comme vérité terrain, montre que ces stratégies améliorent significativement la précision de localisation quand peu de messages sont reçus. En outre, elles délivrent en moyenne 30% de localisations supplémentaires et permettent de caractériser systématiquement l’erreur de positionnement / The ARGOS service was launched in 1978 to serve environmental applications including oceanography, wildlife tracking and maritime safety. The system allows for worldwide positioning and data collection of Platform Terminal Transmitters (PTTs). The positioning is achieved by exploiting the Doppler shift in the carrier frequency of the messages transmitted by the PTTs and recorded by dedicated satellite-borne receivers. Over the last twenty years, the transmission power has decreased significantly and platforms have been used in increasingly harsh environments. This led to deliver a greater number of low quality locations while users sought to identify finer platform behavior. This work first focuses on the implementation of a more efficient location processing to replace the historical real time processing relying on a Least Squares adjustment. Secondly, an offline service to infer locations with even higher accuracy is proposed.The location problem is formulated as the estimation of the state vector of a dynamical system, accounting for a set of admissible movement models of the platform. The exact determination of the state posterior pdf displays a complexity growing exponentially with time. The Interacting Multiple Model (IMM) algorithm has become a standard online approach to derive an approximated solution with a constant computational complexity. For offline applications, many sub-optimal multiple model schemes have also been proposed. Our methodological contributions first focused on extending the framework of the IMM filter so as to handle a bank of models with state vectors of heterogeneous size and meaning. Second, we investigated a new sub-optimal solution for multiple model smoothing which proves to be less computationally expensive and displays markedly better performance than equivalent algorithms. The ARGOS location processing was rewritten to include the IMM filter as real time processing and the IMM smoother as offline service. We eventually analyzed their performances using a large dataset obtained from over 200 mobiles carrying both an ARGOS transmitter and a GPS receiver used as ground truth. The results show that the new approaches significantly improve the positioning accuracy, especially when few messages are received. Moreover, the algorithms deliver 30% more positions and give a systematic estimation of the location error