L'objectif de cette thèse est la résolution d’un problème d’optimisation multi-objectif sous incertitudes en présence de simulations numériques coûteuses. Une validation est menée sur un cas test de thermique transitoire. Dans un premier temps, nous développons un algorithme d'optimisation multi-objectif basé sur le krigeage nécessitant peu d’appels aux fonctions objectif. L'approche est adaptée au calcul distribué et favorise la restitution d'une approximation régulière du front de Pareto complet. Le problème d’optimisation sous incertitudes est ensuite étudié en considérant des mesures de robustesse pires cas et probabilistes. Le superquantile intègre tous les évènements pour lesquels la valeur de la sortie se trouve entre le quantile et le pire cas mais cette mesure de risque nécessite un grand nombre d’appels à la fonction objectif incertaine pour atteindre une précision suffisante. Peu de méthodes permettent de calculer le superquantile de la distribution de la sortie de fonctions coûteuses. Nous développons donc un estimateur du superquantile basé sur une méthode d'échantillonnage préférentiel et le krigeage. Il permet d’approcher les superquantiles avec une faible erreur et une taille d’échantillon limitée. De plus, un couplage avec l’algorithme multi-objectif permet la réutilisation des évaluations. Dans une dernière partie, nous construisons des modèles de substitution spatio-temporels capables de prédire des phénomènes dynamiques non linéaires sur des temps longs et avec peu de trajectoires d’apprentissage. Les réseaux de neurones récurrents sont utilisés et une méthodologie de construction facilitant l’apprentissage est mise en place. / This work aims at solving multi-objective optimization problems in the presence of uncertainties and costly numerical simulations. A validation is carried out on a transient thermal test case. First of all, we develop a multi-objective optimization algorithm based on kriging and requiring few calls to the objective functions. This approach is adapted to the distribution of the computations and favors the restitution of a regular approximation of the complete Pareto front. The optimization problem under uncertainties is then studied by considering the worst-case and probabilistic robustness measures. The superquantile integrates every event on which the output value is between the quantile and the worst case. However, it requires an important number of calls to the uncertain objective function to be accurately evaluated. Few methods give the possibility to approach the superquantile of the output distribution of costly functions. To this end, we have developed an estimator based on importance sampling and kriging. It enables to approach superquantiles with little error and using a limited number of samples. Moreover, the setting up of a coupling with the multi-objective algorithm allows to reuse some of those evaluations. In the last part, we build spatio-temporal surrogate models capable of predicting non-linear, dynamic and long-term in time phenomena by using few learning trajectories. The construction is based on recurrent neural networks and a construction facilitating the learning is proposed.

[pt] A abstração matemática de um processo físico é essencial em problemas de engenharia, pois muitas vezes pode ser impraticável ou impossível realizar experimentos no sistema real. Além disso, modelos matemáticos são mais flexíveis que protótipos físicos, permitindo um rápido refinamento dos projetos do sistema para otimizar várias medidas de desempenho. As aplicações dos modelos podem ser divididas em quatro partes, a saber: projeto, estimativa, controle e monitoramento. Algumas aplicações específicas são i) simulações, ii) soft sensors, iii) avaliação de desempenho, iv) controle estatístico de qualidade e v) detecção e diagnóstico de falhas. Este trabalho visa então: i) desenvolver diferentes classes de modelos capazes de simular com precisão a variável de saída de um sistema, ii) avaliar a eficiência dos algoritmos de otimização utilizados na tarefa de estimação de parâmetros, iii) avaliar qual modelo de atrito é o mais adequado para descrever esse fenômeno em um sistema de posicionamento. Os resultados mostraram que o atrito no sistema de posicionamento apresenta comportamento não linear e assimétrico, já que alguns termos dos modelos de atrito relacionados às velocidades positiva e negativa são significativamente diferentes um do outro. O resultado final do processo de otimização que usou um algoritmo de busca local foi altamente dependente das condições iniciais e do número de parâmetros estimados, o que elevou o erro de simulação. Entretanto, melhores estimativas da variável de saída foram alcançadas quando essa abordagem foi combinada com outros modelos de diferentes classes. Através dessa última abordagem o erro relativo foi reduzido em mais de 20 porcento. As simulações realizadas com os parâmetros estimados pelos algoritmos evolucionários foram mais acuradas, eles foram capazes de reduzir o erro relativo em quase 30 porcento quando comparados com o algoritmo de busca local. Considerando o segundo estudo de caso, o otimizador baseado em árvores de decisão se mostrou igualmente eficaz se comparado aos algoritmos evolucionários. O erro relativo das simulações usando os parâmetros estimados por esses algoritmos foi inferior a 8 porcento. Além disso, a forma do atrito reconstruído na segunda junta do manipulador robótico através dos parâmetros estimados pelos algoritmos está de acordo com o esperado. / [en] The mathematical abstraction of a physical process is essential in engineering problems, as it can often be impractical or impossible to perform experiments on the real system. Besides, mathematical models are more flexible than physical prototypes, allowing for quick refinement of system designs to optimize various performance measures. The applications of the models can be divided into four parts, namely: design, estimation, control and monitoring. Some specific applications are i) simulations, ii) soft sensors, iii) performance evaluation, iv) statistical quality control and, v) fault detection and diagnosis. This work aims to: i) develop different classes of models capable of accurately simulating the output variable of a system, ii) evaluate the efficiency of optimization algorithms used in the parameter estimation task, iii) assess which friction model is the most appropriate to describe this phenomenon in a positioning system. The results showed that the friction in the positioning system presents a nonlinear and asymmetric behavior since some terms of the friction models related to the positive and negative velocities are significantly different from each other. The final result of the optimization process that used a local search algorithm was highly dependent on the initial conditions and the number of estimated parameters, which increased the simulation error. However, better estimates of the output variable were achieved when this approach was combined with other models of different classes. Through this last approach, the relative error was reduced by more than 20 percent. The simulations performed with the parameters estimated by the evolutionary algorithms were more accurate, they were able to reduce the relative error by almost 30 percent when compared with the local search algorithm. Considering the second case study, the decision tree-based optimizer proved to be equally effective compared to evolutionary algorithms. The relative error of the simulations using the parameters estimated by these algorithms was less than 8 percent. Besides, the shape of the friction reconstructed in the second joint of the robotic manipulator through the parameters estimated by the algorithms is in accordance with the expected.

No description available.

Le développement de la fabrication additive permet d'élaborer des pièces de forme extrêmement complexes, en particulier des structures alvéolaires ou "lattices", où l'allégement est recherché. Toutefois, cette technologie, en très forte croissance dans de nombreux secteurs d'activités, n'est pas encore totalement mature, ce qui ne facilite pas les corrélations entre les mesures expérimentales et les simulations déterministes. Afin de prendre en compte les variations de comportement, les approches multiparamétriques sont, de nos jours, des solutions pour tendre vers des conceptions fiables et robustes. L'objectif de cette thèse est d'intégrer des incertitudes matérielles et géométriques, quantifiées expérimentalement, dans des analyses de flambement. Pour y parvenir, nous avons, dans un premier temps, évalué différentes méthodes de substitution, basées sur des régressions et corrélations, et différentes réductions de modèles afin de réduire les temps de calcul prohibitifs. Les projections utilisent des modes issus soit de la décomposition orthogonale aux valeurs propres, soit de développements homotopiques ou encore des développements de Taylor. Dans un second temps, le modèle mathématique, ainsi créé, est exploité dans des analyses ensemblistes et probabilistes pour estimer les évolutions de la charge critique de flambement de structures lattices. / The development of additive manufacturing allows structures with highly complex shapes to be produced. Complex lattice shapes are particularly interesting in the context of lightweight structures. However, although the use of this technology is growing in numerous engineering domains, this one is not enough matured and the correlations between the experimental data and deterministic simulations are not obvious. To take into account observed variations of behavior, multiparametric approaches are nowadays efficient solutions to tend to robust and reliable designs. The aim of this thesis is to integrate material and geometric uncertainty, experimentally quantified, in buckling analyses. To achieve this objective, different surrogate models, based on regression and correlation techniques as well as different reduced order models have been first evaluated to reduce the prohibitive computational time. The selected projections rely on modes calculated either from Proper Orthogonal Decomposition, from homotopy developments or from Taylor series expansion. Second, the proposed mathematical model is integrated in fuzzy and probabilistic analyses to estimate the evolution of the critical buckling load for lattice structures.

Increasing the operational speed of trains has attracted a lot of interest in the last decades and has brought new challenges, especially in terms of infrastructure design methodology, as it may induce excessive vibrations. Such demands can damage bridges, which in turn increases maintenance costs, endangers the safety of passing trains and disrupts passenger comfort. Conventional design provisions should therefore be evaluated in the light of modern concerns; nevertheless, several previous studies have highlighted some of their shortcomings. It should be emphasized that most of these studies have neglected the uncertainties involved, which preventsthe reported results from representing a complete picture of the problem. In this respect, the present thesis is dedicated to evaluating the performance of conventional design methods, especially those related to running safety and passenger comfort, using probabilistic approaches. To achieve this objective, a preliminary study was carried out using the first-order reliability method for short/medium span bridges passed by trains at a wide range of operating speeds. Comparison of these results with the corresponding deterministic responses showed that applying a constant safety factor to the running safety threshold does not guarantee that the safety index will be identical for all bridges. It also shows that the conventional design approaches result in failure probabilities that are higher than the target values. This conclusion highlights the need to update the design methodology for running safety. However, it would be essential to determine whether running safety is the predominant design criterion before conducting further analysis. Therefore, a stochastic comparison between this criterion and passenger comfort was performed. Due to the significant computational cost of such investigations, subset simulation and crude Monte-Carlo (MC) simulation using meta-models based on polynomial chaos expansion were employed. Both methods were found to perform well, with running safety almost always dominating the passenger comfort limit state. Subsequently, classification-based meta-models, e.g. support vector machines, k-nearest neighbours and decision trees, were combined using ensemble techniques to investigate the influence of soil-structure interaction on the evaluated reliability of running safety. The obtained results showed a significant influence, highlighting the need for detailed investigations in further studies. Finally, a reliability-based design optimization was conducted to update the conventional design method of running safety by proposing minimum requirements for the mass per length and moment of inertia of bridges. It is worth mentioning that the inner loop of the method was solved by a crude MC simulation using adaptively trained Kriging meta-models. / Att öka tågens hastighet har väckt stort intresse under de senaste decennierna och har medfört nya utmaningar, särskilt när det gäller broanalyser, eftersom tågen inducerar stora vibrationer. Sådana vibrationer kan öka underhållskostnaderna, äventyra säkerheten för förbipasserande tåg och påverka passagerarkomforten. Konstruktionsbestämmelser bör därför utvärderas mot bakgrund av dessa problem; dock har flera tidigare studier belyst några av bristerna i dagens bestämmelser. Det bör understrykas att de flesta av dessa studier har försummat de osäkerheter som är involverade, vilket hindrar de rapporterade resultaten från att representera en fullständig bild av problemet. I detta avseende syftar denna avhandling till att utvärdera prestandan hos konventionella analysmetoder, särskilt de som rör körsäkerhet och passagerarkomfort, med hjälp av sannolikhetsmetoder. För att uppnå detta mål genomfördes en preliminär studie med första ordningens tillförlitlighetsnmetod för broar med kort/medellång spännvidd som passeras av tåg med ett brett hastighetsspektrum. Jämförelse av dessa resultat med motsvarande deterministiska respons visade att tillämpa en konstant säkerhetsfaktor för verifieringen av trafiksäkerhet inte garanterar att säkerhetsindexet kommer att vara identiskt för alla broar. Det visar också att de konventionella analysmetoderna resulterar i brottsannolikheter som är högre än målvärdena. Denna slutsats belyser behovet av att uppdatera analysmetoden för trafiksäkerhet. Det skulle emellertid vara viktigt att avgöra om trafiksäkerhet är det dominerande designkriteriet innan ytterligare analyser genomförs. Därför utfördes en stokastisk jämförelse mellan detta kriterium och kriteriet för passagerarkomfort. På grund av den betydande. analystiden för sådana beräkningar användes delmängdssimulering och Monte-Carlo (MC) simulering med metamodeller baserade på polynomisk kaosutvidgning. Båda metoderna visade sig fungera bra, med trafiksäkerhet som nästan alltid dominerade över gränsningstillståndet för passagerarkomfort. Därefter kombinerades klassificeringsbaserade metamodeller som stödvektormaskin och beslutsträd genom ensembletekniker, för att undersöka påverkan av jord-brointeraktion på den utvärderade tillförlitligheten gällande trafiksäkerhet. De erhållna resultaten visade en signifikant påverkan och betonade behovet av detaljerade undersökningar genom ytterligare studier. Slutligen genomfördes en tillförlitlighetsbaserad konstruktionsoptimering för att föreslå ett minimikrav på erforderlig bromassa per längdmeter och tröghetsmoment. Det är värt att nämna att metodens inre loop löstes med en MC-simulering med adaptivt tränade Kriging-metamodeller. / <p>QC 20210910</p>