Les récents développements des systèmes technologiques ont mené à une complexification des comportements des systèmes actuels. Une solution pour gérer cette complexité croissante consiste à les considérer comme un ensemble de sous-systèmes hétérogènes et à développer des techniques distribuées pour les contrôler et les gérer. Cette solution soulève plusieurs problèmes. Tout d’abord, l’augmentation de la taille et du nombre de composants entraîne immanquablement l’augmentation du nombre de fautes qui peuvent conduire le système dans un état de défaillance critique. De fait, parmi les fonctions opérationnelles, les tâches de détection et d’isolation des fautes (Fault Detection and Isolation ou FDI), de maintenance et de réparation sont devenues prédominantes et elles influent considérablement sur le coût total des produits finaux.Cette thèse porte sur la détection et l’isolation de fautes. Parmi les différentes méthodes pour générer des tests de diagnostic utilisant la redondance analytique, cette thèse adopte l’approche par espace de parité qui utilise les relations de redondance analytique (RRA). Étant donné un modèle du système sous la forme d’un ensemble d’équations différentielles, les RRA sont des relations obtenues à partir du modèle en éliminant les variables non mesurées. Ceci peut être effectué dans un cadre analytique en utilisant la théorie de l’élimination. Une autre solution consiste à utiliser l’analyse structurelle. L’analyse structurelle est basée sur une abstraction du modèle qui ne conserve que les liens entre variables et équations. Malgré son apparente simplicité, l’analyse structurelle fournit un ensemble d’outils puissants, s’appuyant sur la théorie des graphes, pour analyser et inférer des informations sur le système. Par ailleurs, elle a l’avantage de s’appliquer indifféremment sur les systèmes linéaires ou non linéaires.L’objectif de cette thèse est de développer des techniques efficaces basées sur l’analyse structurelle pour le diagnostic des systèmes continus distribué. Dans ce cadre, le système se décompose en un ensemble de sous-systèmes en fonction de contraintes fonctionnelles, géographiques ou de confidentialité. La thèse se divise principalement en deux parties :• la première partie cherche à mettre à lumière, à partir des modèles structurels obtenus au niveau des sous-systèmes, les redondances qui généreront des tests de diagnostic pertinents au niveau du système global,• la deuxième partie vise à formuler et résoudre le problème d’optimisation lié au choix d’un sous-ensemble de tests de diagnostic au niveau des sous-systèmes permettant une diagnosticabilité maximale pour le système global.La première partie utilise le concept d’ensemble minimal structurellement surdéterminé guidé par les fautes (Fault-Driven Minimal Structurally Overdetermined Set ou FMSO set). Ce concept est introduit dans la thèse. Il s’agit d’un sousensemble d’équations du modèle avec une redondance minimale à partir de laquelle une RRA sensible à un ensemble de fautes peut être obtenu. Deux solutions pour générer des ensembles FMSO pour le système global sont présentées, d’une part dans un cadre décentralisé avec des superviseurs imbriqués suivant une hiérarchie; d’autre part dans un cadre totalement distribué. Ces solutions sont basées sur les propriétés des ensembles FMSO au niveau des sous-systèmes qui sont présentées dans la thèse. La deuxième partie pose un problème d’optimisation dans le cadre d’une recherche heuristique et propose trois solutions basées sur un algorithme A* itératif combiné avec une fonction capable d’évaluer si un ensemble FMSO au niveau global peut être obtenu à partir des ensembles FMSO locaux sélectionnés. Les concepts introduits dans la thèse et les résultats sont appliqués à deux cas d’étude industriels. Le premier est une usine de désalinisation. Le second est un système de détermination et de contrôle d’attitude pour un satellite en orbite basse. / The recent development of technological systems implies a high complexity of behaviors for today’s systems. An answer to the increased system’s complexity is to look at them as a multitude of heterogeneous subsystems and develop distributed techniques to control and manage them. This raises a number of problems. Firstly, as the size and number of components increase, so does the number of fault occurrences that may drive the system to undergo critical failures. Fault detection and isolation (FDI), maintenance and repair are an increasing part of the operational everyday’s tasks and they impact drastically the total cost of final products.This thesis focuses on fault detection and isolation. Among the different methods to generate diagnosis tests by taking advantage of analytical redundancy, this thesis adopts the so-called parity space approach based on analytical redundancy relations (ARRs). Given a model of the system in the form of a set of differential equations, ARRs are relations that are obtained from the model by eliminating non measured variables. This can be performed in an analytical framework using elimination theory but another way of doing this is to use structural analysis. Structural analysis is based on a structural abstraction of the model that only retains a representation of which variables are involved in which equations. Despite the rusticity of the abstract model, structural analysis provides a set of powerful tools, relying on graph theory, to analyze and infer information about the system. Interestingly, it applies indifferently to linear or nonlinear systems. The goal of this thesis is to develop effective techniques based on structural analysis for diagnosis of distributed continuous systems. In this framework, the system is decomposed into a set of subsystems according to functional, geographical or privacy constraints. The thesis is organized in two parts:• highlighting the redundancies that are built into the global structural model and that can be used to generate diagnosis tests starting from the redundancies existing in the subsystem’s models,• formulating and solving the optimization problem linked to the choice of a subset of diagnosis tests at the subsystems level that can lead to a set of diagnosis tests achieving maximum diagnosability for the global system.The first part takes benefit of the concept of Fault-Driven Minimal Structurally Overdetermined Set (FMSO set) that is introduced in the thesis. An FMSO set determines a subset of equations of the model with minimal redundancy from which an ARR sensitive to a set of faults can be obtained. Two solutions for generating FMSOs for the global system are presented, in a decentralized framework with supervisors at each level of a hierarchy and in a totally distributed framework.These are based on the properties of the FMSO sets for the subsystems in relation to those of the global system derived in the thesis.

Made available in DSpace on 2014-06-11T19:27:14Z (GMT). No. of bitstreams: 0 Previous issue date: 2010-05-27Bitstream added on 2014-06-13T19:14:29Z : No. of bitstreams: 1 watanabe_l_me_ilha.pdf: 2840433 bytes, checksum: dfba204f1b82da88aa2da2eca31e1d20 (MD5) / Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / Universidade Estadual Paulista (UNESP) / Nos últimos anos, tem havido um grande interesse das indústrias no desenvolvimento de novas técnicas de detecção e localização de falhas, pois se preocupam cada vez mais com a segurança, havendo assim, a necessidade de supervisão e monitoramento dos sistemas para que as falhas sejam evitadas ou sanadas o mais rápido possível. Determinados parâmetros em sistemas reais como massa, rigidez e amortecimento, podem variar devido ao aparecimento de falhas ou ao próprio desgaste natural dos componentes. Um aparecimento de trincas pode provocar perdas econômicas ou até conduzir a situações perigosas com paradas abruptas das máquinas e/ou equipamentos. Através do auxílio de modelos teóricos bem definidos, métodos de identificação de parâmetros, observadores de estado e auxílio à decisão foi possível desenvolver uma metodologia para detecção e localização de trincas em estruturas reticuladas, dando ênfase às tridimensionais. Foi possível detectar e localizar a trinca já no seu início e acompanhar sua propagação para uma possível parada programada. Foi utilizada a metodologia dos observadores de estado, que pode reconstruir os estados não medidos ou os valores provenientes de pontos de difícil acesso no sistema. Foi construída uma estrutura reticulada no Laboratório para validação da metodologia desenvolvida e os resultados foram bastante satisfatórios / In recent years there has been a great interest of industry in developing new techniques for detection and location of faults, because they worry more about security, so there is the need for supervision and monitoring systems so that failures are avoided or remedied as soon as possible. Certain parameters in real systems such as mass, stiffness and damping can vary due to some failures to own or wear and tear of components. An appearance of cracks can cause economic loss or even lead to dangerous situations with abrupt stopping of machines and/or equipment. Through the aid of well-defined theoretical models, methods of parameter identification, state observers and aid the decision was possible to develop a methodology for detecting and locating cracks in frame structures, emphasizing the three-dimensional. It was possible to detect and locate the crack already in its early stages and monitor its spread to a possible shutdown. Methodology was applied for observer status, which can reconstruct the unmeasured states or values from points of difficult access in the system. A reticulated structure was built at the Laboratory for validation of the methodology and the results were very satisfactory

Le travail de recherche effectué dans cette thèse a été réalisé dans le cadre d'une convention CIFRE entre le laboratoire IMS de l'université Bordeaux 1 et la société Airbus Operations S.A.S. Cette thèse traite de la détection robuste et précoce des pannes oscillatoires de faible amplitude dans les systèmes de commandes de vol électriques. Une panne oscillatoire est une oscillation anormale d'une surface de contrôle due à un dysfonctionnement dans la chaîne d'asservissement de la servocommande d'une gouverne. Les pannes oscillatoires ont une influence sur la structure, l'aéroélasticité et la pilotabilité de l'avion, lorsqu'ils sont situés dans la bande passante de l'actionneur. La capacité à détecter ces pannes est très importante car elles ont un impact sur la conception structurale de l'avion. Au plan méthodologique, nous nous sommes focalisés sur l'estimation adaptative des paramètres et de l'état à base d'une technique de filtrage non linéaire local. Le mécanisme de filtrage opère sur un modèle non linaire de la chaine de contrôle-commande de l’actionneur hydraulique en amont des surfaces de contrôle. L'algorithme d'estimation est basé sur une interpolation polynomiale d'opérateurs linéaire, et offre l'avantage d'une implémentation relativement aisée. Un problème crucial et sous-jacent est la détermination des hyper-paramètres de réglage de cet algorithme. Nous avons proposé une démarche hors-ligne dédiée, en intégrant un critère de sensibilité vis-à-vis des pannes que nous devons détecter. La technique proposée a été implémentée et testée: les résultats expérimentaux obtenus sur banc essai et sur un simulateur A380 ont clairement mis en évidence l'apport de la nouvelle approche en termes de performances, tout en gardant le même niveau de robustesse. / The research work done in this PhD has been caried out in the frame of an industrial convention (CIFRE) between the IMS laboratory and Airbus Operations S.A.S. The thesis deals with robust and early detection of oscillatory failures (OFC: Oscillatory Failure Case) in the Electrical Flight Control System. An oscillatory failure is an abnormal oscillation of a control surface due to component malfunction in control surface servoloops. OFCs have an influence on structural loads, aeroelasticity and controllability when located within the actuator bandwidth. The ability to detect these failures is very important because they have an impact on the structural design of the aircraft. Usual monitoring techniques cannot always guarantee to remain within an envelope with acceptable robustness. In this work, we develop a model based strategy to detect such failures with small amplitude at a very early stage. The monitoring strategy is based on dedicated non linear local filtering for on-line joint parameter/state estimation, allowing for model parameter variations during A/C flight. This strategy is associated with the same decision making rules as currently used for in-service Airbus A380. We propose a method for adjusting the tuning parameters so that various design goals and trades-off can be easily formulated and managed. The performance of the proposed fault detection scheme is measured by its detection delay, its propensity to issue false alarms and whether it permits a failure to go undetected. The proposed technique has been implemented and tested with success on Airbus test facilities including an A380 flight simulator.

No description available.

Bit flips are known to be a source of strange system behavior, failures, and crashes. They can cause dramatic financial loss, security breaches, or even harm human life. Caused by energized particles arising from, e.g., cosmic rays or heat, they are hardly avoidable. Due to transistor sizes becoming smaller and smaller, modern hardware becomes more and more prone to bit flips. This yields a high scientific interest, and many techniques to make systems more resilient against bit flips are developed. Fault-tolerance techniques are techniques that detect and react to bit flips or their effects. Before using these techniques, they typically need to be configured for the particular system they shall protect, the grade of resilience that shall be achieved, and the environment. State-of-the-art configuration approaches have a high risk of being imprecise, of being affected by undesired side effects, and of yielding questionable resilience measures. In this thesis we encourage the usage of formal methods for resiliency configuration, point out advantages and investigate difficulties. We exemplarily investigate two systems that are equipped with fault-tolerance techniques, and we apply parametric variants of probabilistic model checking to obtain optimal configurations for pre-defined resilience criteria. Probabilistic model checking is an automated formal method that operates on Markov models, i.e., state-based models with probabilistic transitions, where costs or rewards can be assigned to states and transitions. Probabilistic model checking can be used to compute, e.g., the probability of having a failure, the conditional probability of detecting an error in case of bit-flip occurrence, or the overhead that arises due to error detection and correction. Parametric variants of probabilistic model checking allow parameters in the transition probabilities and in the costs and rewards. Instead of computing values for probabilities and overhead, parametric variants compute rational functions. These functions can then be analyzed for optimality. The considered fault-tolerant systems are inspired by the work of project partners. The first system is an inter-process communication protocol as it is used in the Fiasco.OC microkernel. The communication structures provided by the kernel are protected against bit flips by a fault-tolerance technique. The second system is inspired by the redo-based fault-tolerance technique \haft. This technique protects an application against bit flips by partitioning the application's instruction flow into transaction, adding redundance, and redoing single transactions in case of error detection. Driven by these examples, we study challenges when using probabilistic model checking for fault-tolerance configuration and present solutions. We show that small transition probabilities, as they arise in error models, can be a cause of previously known accuracy issues, when using numeric solver in probabilistic model checking. We argue that the use of non-iterative methods is an acceptable alternative. We debate on the usability of the rational functions for finding optimal configurations, and show that for relatively short rational functions the usage of mathematical methods is appropriate. The redo-based fault-tolerance model suffers from the well-known state-explosion problem. We present a new technique, counter-based factorization, that tackles this problem for system models that do not scale because of a counter, as it is the case for this fault-tolerance model. This technique utilizes the chain-like structure that arises from the counter, splits the model into several parts, and computes local characteristics (in terms of rational functions) for these parts. These local characteristics can then be combined to retrieve global resiliency and overhead measures. The rational functions retrieved for the redo-based fault-tolerance model are huge - for small model instances they already have the size of more than one gigabyte. We therefor can not apply precise mathematic methods to these functions. Instead, we use the short, matrix-based representation, that arises from factorization, to point-wise evaluate the functions. Using this approach, we systematically explore the design space of the redo-based fault-tolerance model and retrieve sweet-spot configurations.

No description available.