Dans cette thèse, nous présentons des modèles appartenant au domaine de l’intelligence artificielle afin de prédire l’indice magnétique global am à partir des paramètres du vent solaire. Ceci est fait dans l’optique de fournir des modèles opérationnels basés sur les données enregistrées par le satellite ACE situé au point de Lagrange L1. L’indice am ne possède pas à l’heure actuelle de modèles de prédiction. Pour prédire cet indice, nous avons fait appel à des modèles non-linéaires que sont les réseaux de neurones, permettant de modéliser le comportement complexe et non-linéaire de la magnétosphère terrestre. Nous avons dans un premier temps travaillé sur le développement et l’optimisation des modèles de réseaux classiques comme le perceptron multi-couche. Ces modèles ont fait leurs preuves en météorologie spatiale pour prédire aussi bien des indices magnétiques spécifiques à des systèmes de courant comme l’indice Dst, caractéristique du courant annulaire, que des indices globaux comme l’indice Kp. Nous avons en particulier étudié un réseau temporel appelé Time Delay Neural Network (TDNN) et évalué sa capacité à prédire l’indice magnétique am à une heure, uniquement à partir des paramètres du vent solaire. Nous avons analysé la sensibilité des performances des réseaux de neurones en considérant d’une part les données fournies par la base OMNI au niveau de l’onde de choc, et d’autre part des données obtenues par le satellite ACE en L1. Après avoir étudié la capacité de ces réseaux à prédire am, nous avons développé un réseau de neurones encore jamais utilisé en météorologie spatiale, le réseau Long Short Term Mermory ou LSTM. Ce réseau possède une mémoire à court et à long terme, et comme le TDNN, fournit des prédictions de l’indice am uniquement à partir des paramètres du vent solaire. Nous l’avons optimisé afin de modéliser au mieux le comportement de la magnétosphère et avons ainsi obtenu de meilleures performances de prédiction de l'indice am par rapport à celles obtenues avec le TDNN. Nous avons souhaité continuer le développement et l’optimisation du LSTM en travaillant sur l’utilisation de fonctions de couplage en entrée de ce réseau de neurones, et sur le développement de réseaux multisorties pour prédire les indices magnétiques am sectoriels ou aσ, spécifiques à chaque secteur Temps Magnétique Local. Enfin, nous avons développé une nouvelle technique combinant réseau LSTM et processus gaussiens, afin de fournir une prédiction probabiliste jusqu’à six heures des indices magnétiques Dst et am. Cette méthode a été dans un premier temps développée pour l’indice magnétique Dst afin de pouvoir comparer les performances du modèle hybride à des modèles de référence, puis appliquée à l’indice magnétique am. / In this thesis, we present models which belongs to the field of artificial intelligence to predict the geomagnetic index am based on solar wind parameters. This is done in terms to provide operational models based on data recorded by the ACE satellite located at the Lagrangian point L1. Currently, there is no model providing predictions of the geomagnetic index am. To predict this index, we have relied on nonlinear models called neural networks, allowing to model the complex and nonlinear dynamic of the Earth’s magnetosphere. First, we have worked on the development and the optimisation of basics neural networks like the multilayer perceptron. These models have proven in space weather to predict geomagnetic index specific to current systems like the Dst index, characteristic of the ring current, as well as the global geomagnetic index Kp. In particular, we have studied a temporal network, called the Time Delay Neural Network (TDNN) and we assessed its ability to predict the geomagnetic index am within one hour, base only on solar wind parameters. We have analysed the sensitivity of neural network performance when considering on one hand data from the OMNI database at the bow shock, and on the other hand data from the ACE satellite at the L1 point. After studying the ability of neural networks to predict the geomagnetic index am, we have developped a neural network which has never been used before in Space Weather, the Long Short Term Memory or LSTM. Like the TDNN, this network provides am prediction based only on solar wind parameters. We have optimised this network to model at best the magnetosphere behaviour and obtained better performance than the one obtained with the TDNN. We continued the development and the optimisation of the LSTM network by using coupling functions as neural network features, and by developing multioutput networks to predict the sectorial am also called aσ, specific to each Magnetical Local Time sector. Finally, we developped a brand new technique combining the LSTM network and gaussian process, to provide probabilistic predictions up to six hours ahead of geomagnetic index Dst and am. This method has been first developped to predict Dst to be able to compare the performance of this model with reference models, and then applied to the geomagnetic index am.

Cette thèse présente un framework d’optimisation pour la conception hydrodynamique de forme d’hydrofoils. L’optimisation d’hydrofoil par simulation implique des objectifs d’optimisation divergents et impose des compromis contraignants en raison du coût des simulations numériques et des budgets limités généralement alloués à la conception des navires. Le framework fait appel à l’échantillonnage séquentiel et aux modèles de substitution. Un modèle prédictif est construit en utilisant la Régression par Processus Gaussien (RPG) à partir des données issues de simulations fluides effectuées sur différentes géométries d’hydrofoils. Le modèle est ensuite combiné à d’autres critères dans une fonction d’acquisition qui est évaluée sur l’espace de conception afin de définir une nouvelle géométrie qui est testée et dont les paramètres et la réponse sont ajoutés au jeu de données, améliorant ainsi le modèle. Une nouvelle fonction d’acquisition a été développée, basée sur la variance RPG et la validation croisée des données. Un modeleur géométrique a également été développé afin de créer automatiquement les géométries d’hydrofoil a partir des paramètres déterminés par l’optimiseur. Pour compléter la boucle d’optimisation,FINE/Marine, un solveur fluide RANS, a été intégré dans le framework pour exécuter les simulations fluides. Les capacités d’optimisation ont été testées sur des cas tests analytiques montrant que la nouvelle fonction d’acquisition offre plus de robustesse que d’autres fonctions d’acquisition existantes. L’ensemble du framework a ensuite été testé sur des optimisations de sections 2Dd’hydrofoil ainsi que d’hydrofoil 3D avec surface libre. Dans les deux cas, le processus d’optimisation fonctionne, permettant d’optimiser les géométries d’hydrofoils et confirmant les performances obtenues sur les cas test analytiques. Les optima semblent cependant être assez sensibles aux conditions opérationnelles. / This thesis presents a practical hydrodynamic optimization framework for hydrofoil shape design. Automated simulation based optimization of hydrofoil is a challenging process. It may involve conflicting optimization objectives, but also impose a trade-off between the cost of numerical simulations and the limited budgets available for ship design. The optimization frameworkis based on sequential sampling and surrogate modeling. Gaussian Process Regression (GPR) is used to build a predictive model based on data issued from fluid simulations of selected hydrofoil geometries. The GPR model is then combined with other criteria into an acquisition function that isevaluated over the design space, to define new querypoints that are added to the data set in order to improve the model. A custom acquisition function is developed, based on GPR variance and cross validation of the data.A hydrofoil geometric modeler is also developed to automatically create the hydrofoil shapes based on the parameters determined by the optimizer. To complete the optimization loop, FINE/Marine, a RANS flow solver, is embedded into the framework to perform the fluid simulations. Optimization capabilities are tested on analytical test cases. The results show that the custom function is more robust than other existing acquisition functions when tested on difficult functions. The entire optimization framework is then tested on 2D hydrofoil sections and 3D hydrofoil optimization cases with free surface. In both cases, the optimization process performs well, resulting in optimized hydrofoil shapes and confirming the results obtained from the analytical test cases. However, the optimum is shown to be sensitive to operating conditions.

Geometric quality variables of wafers, such as BOW and WARP, are critical in their applications. A large variation of these quality variables reduces the number of conforming products in the downstream production. Therefore, it is important to reduce the variation by variation modeling, analysis and control for multistage wafer manufacturing processes (MWMPs). First, an intermediate feedforward control strategy is developed to adjust and update the control actions based on the online measurements of intermediate wafer quality measurements. The control performance is evaluated in a MWMP to transform ingots into polished wafers. However, in a complex multistage manufacturing process, the quality variables may have nonlinear relationship with the parameters of the predictors. In this case, piecewise linear regression tree (PLRT) models are used to address nonlinear relationships in MWMP to improve the model prediction performance. The obtained PLRT model is further reconfigured to be complied with the physical layout of the MWMP for feedforward control purposes. The procedure and effectiveness of the proposed method is shown in a case study of a MWMP. Furthermore, as the geometric profiles and quality variables are important quality features for a wafer, fast and accurate measurements of those features are crucial for variation reduction and feedforward control. A sequential measurement strategy is proposed to reduce the number of samples measured in a wafer, yet provide adequate accuracy for the quality feature estimation. A Gaussian process model is used to estimate the true profile of a wafer with improved sensing efficiency. Finally, we study the multistage multimode process monitoring problem. We propose to use PLRTs to inter-relate the variables in a multistage multimode process. A unified charting system is developed. We further study the run length distribution, and optimize the control chart system by considering the modeling uncertainties. Finally, we compare the proposed method with the risk adjustment type of control chart systems based on global regression models, for both simulation study and a wafer manufacturing process.

Gaussian Processes (GPs) are promising Bayesian methods for classification and regression problems. They have also been used for semi-supervised classification tasks. In this thesis, we propose new algorithms for solving semi-supervised binary classification problem using GP regression (GPR) models. The algorithms are closely related to semi-supervised classification based on support vector regression (SVR) and maximum margin clustering. The proposed algorithms are simple and easy to implement. Also, the hyper-parameters are estimated without resorting to expensive cross-validation technique. The algorithm based on sparse GPR model gives a sparse solution directly unlike the SVR based algorithm. Use of sparse GPR model helps in making the proposed algorithm scalable. The results of experiments on synthetic and real-world datasets demonstrate the efficacy of proposed sparse GP based algorithm for semi-supervised classification.

This thesis consists of two parts. The first part is on fractional factorial design, and the second part is on computer experiment. The first part has two chapters. In the first chapter, we use the concept of conditional main effect, and propose the CME analysis to solve the problem of effect aliasing in two-level fractional factorial design. In the second chapter, we study the conversion rates of a system of webpages with the proposed funnel testing method, by using directed graph to represent the system, fractional factorial design to conduct the experiment, and a method to optimize the total conversion rate with respect to all the webpages in the system. The second part also has two chapters. In the third chapter, we use regression models to quantify the model form uncertainties in the Perez model in building energy simulations. In the last chapter, we propose a new Gaussian process that can jointly model both point and integral responses.

Le domaine de l’assurance automobile fonctionne par cycles présentant des phases de profitabilité et d’autres de non-profitabilité. Dans les phases de non-profitabilité, les compagnies d’assurance ont généralement le réflexe d’augmenter le coût des primes afin de tenter de réduire les pertes. Par contre, de très grandes augmentations peuvent avoir pour effet de massivement faire fuir la clientèle vers les compétiteurs. Un trop haut taux d’attrition pourrait avoir un effet négatif sur la profitabilité à long terme de la compagnie. Une bonne gestion des augmentations de taux se révèle donc primordiale pour une compagnie d’assurance. Ce mémoire a pour but de construire un outil de simulation de l’allure du porte- feuille d’assurance détenu par un assureur en fonction du changement de taux proposé à chacun des assurés. Une procédure utilisant des régressions à l’aide de processus gaus- siens univariés est développée. Cette procédure offre une performance supérieure à la régression logistique, le modèle généralement utilisé pour effectuer ce genre de tâche. / The field of auto insurance is working by cycles with phases of profitability and other of non-profitability. In the phases of non-profitability, insurance companies generally have the reflex to increase the cost of premiums in an attempt to reduce losses. For cons, very large increases may have the effect of massive attrition of the customers. A too high attrition rate could have a negative effect on long-term profitability of the company. Proper management of rate increases thus appears crucial to an insurance company. This thesis aims to build a simulation tool to predict the content of the insurance portfolio held by an insurer based on the rate change proposed to each insured. A proce- dure using univariate Gaussian Processes regression is developed. This procedure offers a superior performance than the logistic regression model typically used to perform such tasks.

The main objective of this dissertation is to apply Bayesian modeling to different complex and high-dimensional spatial data sets. I develop Bayesian hierarchical spatial models for both the observed location and the observation variable. Throughout this dissertation I execute the inference of the posterior distributions using Markov chain Monte Carlo by developing computational strategies that can reduce the computational cost. I start with a "high level" image analysis by modeling the pixels with a Gaussian process and the objects with a marked-point process. The proposed method is an automatic image segmentation and classification procedure which simultaneously detects the boundaries and classifies the objects in the image into one of the predetermined shape families. Next, I move my attention to the piecewise non-stationary Gaussian process models and their computational challenges for very large data sets. I simultaneously model the non-stationarity and reduce the computational cost by using the innovative technique of full-scale approximation. I successfully demonstrate the proposed reduction technique to the Total Ozone Matrix Spectrometer (TOMS) data. Furthermore, I extend the reduction method for the non-stationary Gaussian process models to a dynamic partition of the space by using a modified Treed Gaussian Model. This modification is based on the use of a non-stationary function and the full-scale approximation. The proposed model can deal with piecewise non-stationary geostatistical data with unknown partitions. Finally, I apply the method to the TOMS data to explore the non-stationary nature of the data.

Submitted by Aelson Maciera (aelsoncm@terra.com.br) on 2017-08-29T19:13:47Z No. of bitstreams: 1 TeseLHMM.pdf: 2605456 bytes, checksum: b07c268a8fc9a1af8f14ac26deeec97e (MD5) / Approved for entry into archive by Ronildo Prado (ronisp@ufscar.br) on 2017-09-25T18:22:48Z (GMT) No. of bitstreams: 1 TeseLHMM.pdf: 2605456 bytes, checksum: b07c268a8fc9a1af8f14ac26deeec97e (MD5) / Approved for entry into archive by Ronildo Prado (ronisp@ufscar.br) on 2017-09-25T18:22:55Z (GMT) No. of bitstreams: 1 TeseLHMM.pdf: 2605456 bytes, checksum: b07c268a8fc9a1af8f14ac26deeec97e (MD5) / Made available in DSpace on 2017-09-25T18:27:54Z (GMT). No. of bitstreams: 1 TeseLHMM.pdf: 2605456 bytes, checksum: b07c268a8fc9a1af8f14ac26deeec97e (MD5) Previous issue date: 2017-04-07 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / Conventional reliability analysis techniques are focused on the occurrence of failures over time. However, in certain situations where the occurrence of failures is tiny or almost null, the estimation of the quantities that describe the failure process is compromised. In this context the degradation models were developed, which have as experimental data not the failure, but some quality characteristic attached to it. Degradation analysis can provide information about the components lifetime distribution without actually observing failures. In this thesis we proposed different methodologies for degradation data based on the inverse Gaussian distribution. Initially, we introduced the inverse Gaussian deterioration rate model for degradation data and a study of its asymptotic properties with simulated data. We then proposed an inverse Gaussian process model with frailty as a feasible tool to explore the influence of unobserved covariates, and a comparative study with the traditional inverse Gaussian process based on simulated data was made. We also presented a mixture inverse Gaussian process model in burn-in tests, whose main interest is to determine the burn-in time and the optimal cutoff point that screen out the weak units from the normal ones in a production row, and a misspecification study was carried out with the Wiener and gamma processes. Finally, we considered a more flexible model with a set of cutoff points, wherein the misclassification probabilities are obtained by the exact method with the bivariate inverse Gaussian distribution or an approximate method based on copula theory. The application of the methodology was based on three real datasets in the literature: the degradation of LASER components, locomotive wheels and cracks in metals. / As técnicas convencionais de análise de confiabilidade são voltadas para a ocorrência de falhas ao longo do tempo. Contudo, em determinadas situações nas quais a ocorrência de falhas é pequena ou quase nula, a estimação das quantidades que descrevem os tempos de falha fica comprometida. Neste contexto foram desenvolvidos os modelos de degradação, que possuem como dado experimental não a falha, mas sim alguma característica mensurável a ela atrelada. A análise de degradação pode fornecer informações sobre a distribuição de vida dos componentes sem realmente observar falhas. Assim, nesta tese nós propusemos diferentes metodologias para dados de degradação baseados na distribuição gaussiana inversa. Inicialmente, nós introduzimos o modelo de taxa de deterioração gaussiana inversa para dados de degradação e um estudo de suas propriedades assintóticas com dados simulados. Em seguida, nós apresentamos um modelo de processo gaussiano inverso com fragilidade considerando que a fragilidade é uma boa ferramenta para explorar a influência de covariáveis não observadas, e um estudo comparativo com o processo gaussiano inverso usual baseado em dados simulados foi realizado. Também mostramos um modelo de mistura de processos gaussianos inversos em testes de burn-in, onde o principal interesse é determinar o tempo de burn-in e o ponto de corte ótimo para separar os itens bons dos itens ruins em uma linha de produção, e foi realizado um estudo de má especificação com os processos de Wiener e gamma. Por fim, nós consideramos um modelo mais flexível com um conjunto de pontos de corte, em que as probabilidades de má classificação são estimadas através do método exato com distribuição gaussiana inversa bivariada ou em um método aproximado baseado na teoria de cópulas. A aplicação da metodologia foi realizada com três conjuntos de dados reais de degradação de componentes de LASER, rodas de locomotivas e trincas em metais.

Dans cette thèse, nous proposons de construire de meilleurs modèles Processus Gaussiens (GPs) en intégrant les connaissances antérieures avec des données expérimentales. En raison du coût élevé de l’exécution d’expériences sur les systèmes physiques, les modèles numériques deviennent un moyen évident de concevoir des systèmes physiques. Traditionnellement, ces modèles ont été construits expérimentalement et itérativement; une méthode plus rentable de construction de modèles consiste à utiliser des algorithmes d’apprentissage automatique. Nous démontrons comment créer des modèles en intégrant une connaissance antérieure en modifiant les fonctions de covariance. Nous proposons des modèles GP pour différents phénomènes physiques en mécanique des fluides.De même, les lois physiques entre plusieurs sorties peuvent être appliquées en manipulant les fonctions de covariance. Pour chaque application, nous comparons le modèle proposé avec le modèle de l’état de l’art et démontrons les gains de coût ou de performance obtenus. / In this thesis, we propose to build better Gaussian Process (GP) modelsby integrating the prior knowledge of Aircraft design with experimental data. Due tothe high cost of performing experiments on physical systems, models become an efficientmeans to designing physical systems. We demonstrate how to create efficient models byincorporating the prior information from engineering design, mainly by changing the covariancefunctions of the GP.We propose GP models to detect onset of non-linearity, detectmodal parameters and interpolate position of shock in aerodynamic experiments. Similarly,physical laws between multiple outputs can be enforced by manipulating the covariancefunctions, we propose to integrate flight-mechanics to better identify loads using thesemodels. For each application we compare the proposed model with the state-of-the-artmodel and demonstrate the cost or performance gains achieved.

Made available in DSpace on 2016-06-02T20:06:51Z (GMT). No. of bitstreams: 1 6609.pdf: 3049383 bytes, checksum: 33c7f1618f776ca50cf4694aaba80ea5 (MD5) Previous issue date: 2015-03-09 / In this work we propose a Bayesian nonparametric approach for modeling extreme value data. We treat the location parameter _ of the generalized extreme value distribution as a random function following a Gaussian process model (Rasmussem & Williams 2006). This configuration leads to no closed-form expressions for the highdimensional posterior distribution. To tackle this problem we use the Riemannian Manifold Hamiltonian Monte Carlo algorithm which allows samples from the posterior distribution with complex form and non-usual correlation structure (Calderhead & Girolami 2011). Moreover, we propose an autoregressive time series model assuming the generalized extreme value distribution for the noise and obtained its Fisher information matrix. Throughout this work we employ some computational simulation studies to assess the performance of the algorithm in its variants and show many examples with simulated and real data-sets. / Neste trabalho propomos uma abordagem Bayesiana não-paramétrica para a modelagem de dados com comportamento extremo. Tratamos o parâmetro de locação _ da distribuição generalizada de valor extremo como uma função aleatória e assumimos um processo Gaussiano para tal função (Rasmussem & Williams 2006). Esta situação leva à intratabilidade analítica da distribuição a posteriori de alta dimensão. Para lidar com este problema fazemos uso do método Hamiltoniano de Monte Carlo em variedade Riemanniana que permite a simulação de valores da distribuição a posteriori com forma complexa e estrutura de correlação incomum (Calderhead & Girolami 2011). Além disso, propomos um modelo de série temporal autoregressivo de ordem p, assumindo a distribuição generalizada de valor extremo para o ruído e determinamos a respectiva matriz de informação de Fisher. No decorrer de todo o trabalho, estudamos a qualidade do algoritmo em suas variantes através de simulações computacionais e apresentamos vários exemplos com dados reais e simulados.