The significant amount of data contained in digital images present barriers to methods of learning from the information they hold. Noise and the subjectivity of image evaluation further complicate such automated processes. In this thesis, we examine a particular area in which these difficulties are experienced. We attempt to control the parameters of a multi-step algorithm that processes visual information. A framework for approaching the parameter selection problem using reinforcement learning agents is presented as the main contribution of this research. We focus on the generation of state and action space, as well as task-dependent reward. We first discuss the automatic determination of fuzzy membership functions as a specific case of the above problem. Entropy of a fuzzy event is used as a reinforcement signal. Membership functions representing brightness have been automatically generated for several images. The results show that the reinforcement learning approach is superior to an existing simulated annealing-based approach. The framework has also been evaluated by optimizing ten parameters of the text detection for semantic indexing algorithm proposed by Wolf et al. Image features are defined and extracted to construct the state space. Generalization to reduce the state space is performed with the fuzzy ARTMAP neural network, offering much faster learning than in the previous tabular implementation, despite a much larger state and action space. Difficulties in using a continuous action space are overcome by employing the DIRECT method for global optimization without derivatives. The chosen parameters are evaluated using metrics of recall and precision, and are shown to be superior to the parameters previously recommended. We further discuss the interplay between intermediate and terminal reinforcement.

The past two decades have witnessed tremendous advances in the field of high accuracy, high mass resolution data acquisition of complex samples such as crude oils and the human proteome. With the development of Fourier transform ion cyclotron resonance mass spectrometry, the rapidly growing field of petroleomics has emerged, whose goal is to process and analyse the large volumes of complex and often poorly understood data on crude oils generated by mass spectrometry. As global oil resources deplete, oil companies are increasingly moving towards the extraction and refining of the still plentiful reserves of heavy, carbon rich and highly contaminated crude oil. It is essential that the oil industry gather the maximum possible amount of information about the crude oil prior to setting up the drilling infrastructure, in order to reduce processing costs. This project describes how machine learning can be used as a novel way to extract critical information from complex mass spectra which will aid in the processing of crude oils. The thesis discusses the experimental methods involved in acquiring high accuracy mass spectral data for a large and key industry-standard set of crude oil samples. These data are subsequently analysed to identify possible links between the raw mass spectra and certain physical properties of the oils, such as pour point and sulphur content. Methods including artificial neural networks and self organising maps are described and the use of spectral clustering and pattern recognition to classify crude oils is investigated. The main focus of the research, the creation of an original simulated annealing genetic algorithm hybrid technique (SAGA), is discussed in detail and the successes of modelling a number of different datasets using all described methods are outlined. Despite the complexity of the underlying mass spectrometry data, which reflects the considerable chemical diversity of the samples themselves, the results show that physical properties can be modelled with varying degrees of success. When modelling pour point temperatures, the artificial neural network achieved an average prediction error of less than 10% while SAGA predicted the same values with an average accuracy of more than 85%. It did not prove possible to model any of the other properties with such statistical significance; however improvements to feature extraction and pre-processing of the spectral data as well as enhancement of the modelling techniques should yield more consistent and statistically reliable results. These should in due course lead to a comprehensive model which the oil industry can use to process crude oil data using rapid and cost effective analytical methods.

<p>U okviru disertacije, kori&scaron;ćenjem ve&scaron;tačkih neuronskih mreža razvijeni su modeli za predviđanje količina ambalažnog i biorazgradivog komunalnog otpada u Republici Srbiji do kraja 2030. godine. Razvoj modela baziran je na zavisnosti između ukupne potro&scaron;nje domaćinstva i generisane količine dva posmatrana toka otpada. Pored toga, na bazi zavisnosti sa bruto domaćim proizvodom (BDP), definisan je i model za projekciju zastupljenosti osnovnih opcija tretmana komunalnog otpada u Republici Srbiji za isti period. Na osnovu dobijenih rezultata, stvorene su polazne osnove za procenu potencijala za reciklažu ambalažnog otpada, kao i za procenu u kojoj meri se može očekivati da određene količine biorazgradivog otpada u narednom periodu ne budu odložene na deponije, &scaron;to je u skladu sa savremenim principima upravljanja otpadom i postojećim zahtevima EU u ovoj oblasti.</p> / <p>By using artificial neural networks, models for prediction of the quantity of<br />packaging and biodegradable municipal waste in the Republic of Serbia by<br />the end of 2030, were developed. Models were based on dependence<br />between total household consumption and generated quantities of two<br />observed waste streams. In addition, based on dependence with the Gross<br />Domestic Product (GDP), a model for the projection of share of different<br />municipal solid waste treatment options in the Republic of Serbia for the same<br />period, was created. Obtained results represent a starting point for assessing<br />the potential for recycling of packaging waste, and determination of<br />biodegradable municipal waste quantities which expected that in the future<br />period will not be disposed at landfills, in accordance with modern principles<br />of waste management and existing EU requirements in this area.</p>

<p>Mogućnost primene polivinilhloridnih (PVC) podnih obloga je određena krajnjim svojstvima koja zavise od sastava obloge i načina proizvodnje. Zbog složenog sastava i različitih načina pripreme PVC podnih obloga, veoma je te&scaron;ko tačno proceniti uticaj pojedinačnog procesnog parametara na svojstva dobijenog proizvoda. U ovom radu, da bi se ispitao uticaj sastava polivinilhloridnih sme&scaron;a na svojstva PVC podnih obloga pripremljeno je 27 receptura u kojima su varirane: koncentracije kalcijumkarbonata (40, 70 i 100 phr), koncentracije sredstva za ekspanziju, azodikarbonamida ADC (0,8, 1,0 i 1,2 mas. % u odnosu na ukupnu masu) kao i odnos &bdquo;kikeraˮ i sredstva za ekspanziju, ZnO/ADC (0,33; 0,50 i 0,67). Da bi se proučio uticaj procesnih parametara na svojstva PVC podnih obloga menjani su tehnolo&scaron;ki uslovi proizvodnje: temperatura (180, 184, 188, 192 i 196 &deg;C) i vreme (90, 120 i 150 sekundi) ekspanzije poleđinskog sloja PVC podne obloge. Na taj način je od svake PVC paste dobijeno 15 uzoraka. Za svaki uzorak određena su sledeća svojstva: stepen ekspanzije, zatezna sila pri kidanju, prekidna sila kidanja, zatezno i prekidno izduženje, početni otpor cepanju, otpor cepanju, ukupna deformacija, zaostala deformacija, povratna elastičnost, gustina pene i indeks žućenja. S obzirom na ovako veliki broj podataka urađena je statistička obrada dobijenih eksperimentalnih podataka metodom vi&scaron;estruke linearne regresione analize, kako bi procenili uticaji pojedinačnih procesnih parametara na ispitivana svojstva. Napisan je originalni programski kod primenom Garson-ovog i Yoon-ovog modela u programu Matlab koji omogućava formiranje neuronske mreže i njenu upotrebu u cilju fitovanja eksperimentalnih podataka. Rezultati dobijeni primenom modela po Garson-u nisu pogodni za određivanje uticaja sastava PVC sme&scaron;e i uslova prerade na konačna svojstva proizvoda, jer ne pokazuju pravac uticaja. Dok je ve&scaron;tačka neuronska mreža koja se zasniva na Yoon-ovom modelu uspe&scaron;no primenjena u razvoju novih i pobolj&scaron;anju postojećih svojstava heterogenih PVC proizvoda. Ispitan je i uticaj veličine čestice ZnO koji je upotrebljen kao &bdquo;kikerˮ (sredstvo koje utiče na smanjenje temperature raspada ADC) na svojstva penastih podnih obloga. Napravljene su dve PVC paste, jedna sa komercijalnim ZnO, a druga sa nano ZnO, pri čemu je sastav ostalih komponenti bio identičan. Na osnovu dobijenih eksperimentalnih podatka može se zaključiti da kod uzoraka u kojima je upotrebljen nano ZnO dolazi do intenzivnijeg &bdquo;curenjaˮ gasa nastalog raspadom ADC i međusobnog povezivanja pora. Kao posledica ovakve strukture pora uzorci sa nano ZnO imaju lo&scaron;ija mehanička svojstva. Na osnovu dobijenih karakterističnih temperatura na TG krivama nije uočen uticaj veličine čestica ZnO na termičku stabilnost ispitivanih uzoraka.</p> / <p>The application of PVC floor coverings is strongly connected with their end-use properties which depend on the composition and processing conditions. It is very difficult to estimate the proper influence of production parameters on the characteristics of PVC floor coverings due to their complex composition and various preparation procedures. In this paper, in order to investigate the effect of the PVC mixture composition on the properties of PVC floor coverings, 27 formulations are prepared varying concentration of calcium carbonate (40, 70 and 100 phr), concentration of blowing agent, azodicarbonamide ADC (0.8, 1.0 and 1.2 wt. % relative to the total weight) and the ratio of &quot;kicker&quot; and blowing agent, ZnO/ADC (0.33; 0.50 and 0.67). To study the influence of process parameters on the properties of PVC floor coverings technological production conditions are varied: expansion temperature (180, 184, 188, 192 and 196 &deg;C) and expansion time (90, 120 and 150 seconds) of the PVC floor covering back layer. In this way, 15 samples are made of each PVC paste. The following properties are determined for each sample: expansion ratio, tensile strength, braking strength, tensile extension, breaking extension, initial resistance to tearing, tearing resistance, identation, residual identation, elasticity, density foam and yellowing index. Having such a large amount of data, statistical analysis of experimental data are made with multiple linear regression analysis in order to assess the effects of process parameters on investigated properties. The original program code is written using the Garson&#39;s and Yoon&#39;s models in the Matlab programme that allows the formation of neural networks and its use for the purpose of fitting the experimental data. Results obtained by using the Garson model are not suitable for determining the influence of composition of the PVC mixture and processing conditions on the properties of the final product because it does not show the direction of impact. While the artificial neural network based on Yoon&#39;s model is successfully applied to the development of new as well as to the improvement of the existing properties of the heterogeneous PVC products. The influence of ZnO particle size, used as a&ldquo;kicker&ldquo; (this material reduces the decomposition temperature of ADC) is examined on the properties of the foam flooring. Two PVC pastes are made, one with commercial ZnO and the other with nano ZnO, with the other components of the compositions identical. Based on the obtained experimental data, it can be concluded that there is more intensive gas &bdquo;leak&ldquo; resulting from disintegration of the ADC and also more intensive interconnection of pores in the samples where nano ZnO is used. As a result of this structure of pores, samples with nano ZnO have inferior mechanical properties. Based on the characteristic temperature obtained on TG curves, the influence of ZnO particle size on the thermal stability of the investigated samples is not observed.</p>

The research work presented in this thesis describes the development and testing of a new data analysis methodology for the automated near real-time detection and approximate location of pipe bursts and other events which induce similar abnormal pressure/flow variations (e.g., unauthorised consumptions, equipment failures, etc.) in Water Distribution Systems (WDSs). This methodology makes synergistic use of several self-learning Artificial Intelligence (AI) and statistical/geostatistical techniques for the analysis of the stream of data (i.e., signals) collected and communicated on-line by the hydraulic sensors deployed in a WDS. These techniques include: (i) wavelets for the de-noising of the recorded pressure/flow signals, (ii) Artificial Neural Networks (ANNs) for the short-term forecasting of future pressure/flow signal values, (iii) Evolutionary Algorithms (EAs) for the selection of optimal ANN input structure and parameters sets, (iv) Statistical Process Control (SPC) techniques for the short and long term analysis of the burst/other event-induced pressure/flow variations, (v) Bayesian Inference Systems (BISs) for inferring the probability of a burst/other event occurrence and raising the detection alarms, and (vi) geostatistical techniques for determining the approximate location of a detected burst/other event. The results of applying the new methodology to the pressure/flow data from several District Metered Areas (DMAs) in the United Kingdom (UK) with real-life bursts/other events and simulated (i.e., engineered) burst events are also reported in this thesis. The results obtained illustrate that the developed methodology allowed detecting the aforementioned events in a fast and reliable manner and also successfully determining their approximate location within a DMA. The results obtained additionally show the potential of the methodology presented here to yield substantial improvements to the state-of-the-art in near real-time WDS incident management by enabling the water companies to save water, energy, money, achieve higher levels of operational efficiency and improve their customer service. The new data analysis methodology developed and tested as part of the research work presented in this thesis has been patented (International Application Number: PCT/GB2010/000961).

A l'heure actuelle, les seuls moyens d'informer les usagers de la ville de l'environnement sonore dans lequel ils vivent consistent en des indicateurs de niveaux sonores moyens et annuels obtenus par modélisation acoustique des principales infrastructures de transports. Or, ces indicateurs sont difficilement compris et de ce fait mal interprétés par les usagers de la ville car ils ne reflètent pas la signification des bruits perçus et la diversité des situations que les citadins rencontrent. Le but de ce travail de recherche est donc d'analyser la façon dont les usagers de la ville perçoivent le paysage sonore urbain afin de définir des indicateurs de qualité sonore qui pourront être à terme intégrés dans une représentation territoriale cartographique. Pour ce faire, il a tout d'abord été nécessaire de s'attacher à déterminer un pas temporel et spatial de mesure permettant de caractériser des ambiances urbaines d'un point de vue acoustique. A partir d'enregistrements longue durée (trois mois environs) en six points fixes à Paris, il a été possible de déterminer à travers des classifications ascendantes hiérarchiques de Ward associées à des cartes auto-organisatrices de Kohonen qu'une durée de dix minutes semble dans la plupart des cas être suffisante pour caractériser différentes ambiances sonores. Grâce aux mêmes méthodes de classification, l'analyse du maillage spatial a permis de définir quatre zones homogènes qui correspondent (1) au parc, (2) au boulevard, (3) à la rue piétonne puis (4) une zone que l'on qualifiera de zone de transition. La suite de l'étude s'est attachée à construire des modèles de prédiction de la qualité sonore. A partir d'enquêtes de terrain réalisées à Paris et à Lyon, il a été possible d'établir des modèles à la fois locaux (caractérisant le lieu même où le questionnaire a été évalué) et globaux basés d'une part sur des régressions linéaires multiples et d'autre part sur des réseaux de neurones artificiels. La comparaison de ces deux types de modèles a permis entre autre de mettre en évidence l'apport des réseaux de neurones artificiels devant les régressions linéaires multiples en termes de prédiction. Par ailleurs il est ressorti de ces modèles l'importance de variables telles que le silence, l'agrément visuel ou encore la présence de sources sonores particulières comme les véhicules légers pour expliquer la qualité sonore de l'environnement. / At present, the only ways to inform city dwellers about the sound environment in which they live are annual and average sound level indicators using acoustic modeling of main transport infrastructure. However, these indicators are difficult to understand and therefore misinterpreted by city dwellers because they do not reflect the significance of perceived noise and the diversity of the situations. The aim of this research is therefore to analyze how the city dwellers perceive the urban soundscape in order to characterize sound quality indicators which can be used into mapping. To do this, it was first of all necessary to determine a temporal and spatial resolution to characterize urban environment from an acoustic point of view. From long period recordings (almost three months) at six locations in Paris it was possible to determine through hierarchical ascendant Ward classifications combined with self-organizing Kohonen maps that duration of ten minutes for measurements seems to be enough to characterize in most cases different acoustic environments. Thanks to the same classification methods, spatial study made it possible to define four homogeneous areas which correspond (1) to the park, (2) to the boulevard, (3) to the pedestrian street and (4) to an area which can be considered as a transition one. Then this study focused on building sound quality predictive models. Thanks to field surveys in Paris and Lyon, it was possible to establish local models (characterizing the location where the questionnaire has been evaluated) and overall models based on one hand on multiple linear regressions and on the other hand on artificial neural networks. The comparison of both models highlighted the advantages of artificial neural networks compared to multiple linear regressions in terms of prediction. Moreover, according to these models, variables such as silence, the visual pleasantness or even the presence of specific sound sources as light vehicles explain the sound quality of the environment.

Un robot autonome collaborant avec des humains doit être capable d'apprendre à se déplacer et à manipuler des objets dans la même tâche. Dans une approche classique, on considère des modules fonctionnels indépendants gérant les différents aspects de la tâche (navigation, contrôle du bras...). A l'opposé, l'objectif de cette thèse est de montrer que l'apprentissage de tâches de natures différentes peut être abordé comme un problème d'apprentissage d'attracteurs sensorimoteurs à partir d'un petit nombre de structures non spécifiques à une tâche donnée. Nous avons donc proposé une architecture qui permet l'apprentissage et l'encodage d'attracteurs pour réaliser aussi bien des tâches de navigation que de contrôle d'un bras.Comme point de départ, nous nous sommes appuyés sur un modèle inspiré des cellules de lieu pour la navigation d'un robot autonome. Des apprentissages en ligne et interactifs de couples lieu/action sont suffisants pour faire émerger des bassins d'attraction permettant à un robot autonome de suivre une trajectoire. En interagissant avec le robot, on peut corriger ou orienter son comportement. Les corrections successives et leur encodage sensorimoteur permettent de définir le bassin d'attraction de la trajectoire. Ma première contribution a été d'étendre ce principe de construction d'attracteurs sensorimoteurs à un contrôle en impédance pour un bras robotique. Lors du maintien d'une posture proprioceptive, les mouvements du bras peuvent être corrigés par une modification en-ligne des commandes motrices exprimées sous la forme d'activations musculaires. Les attracteurs moteurs résultent alors des associations simples entre l'information proprioceptive du bras et ces commandes motrices. Dans un second temps, j'ai montré que le robot pouvait apprendre des attracteursvisuo-moteurs en combinant les informations proprioceptives et visuelles. Le contrôle visuo-moteur correspond à un homéostat qui essaie de maintenir un équilibre entre ces deux informations. Dans le cas d'une information visuelle ambiguë, le robot peut percevoir un stimulus externe (e.g. la main d'un humain) comme étant sa propre pince. Suivant le principe d'homéostasie, le robot agira pour réduire l'incohérence entre cette information externe et son information proprioceptive. Il exhibera alors un comportement d'imitation immédiate des gestes observés. Ce mécanisme d'homéostasie, complété par une mémoire des séquences observées et l'inhibition des actions durant l'observation, permet au robot de réaliser des imitations différées et d'apprendre par observation. Pour des tâches plus complexes, nous avons aussi montré que l'apprentissage de transitions peut servir de support pour l'apprentissage de séquences de gestes, comme c'était le cas pour l'apprentissage de cartes cognitives en navigation. L'utilisation de contextes motivationnels permet alors le choix entre les différentes séquences apprises.Nous avons ensuite abordé le problème de l'intégration dans une même architecture de comportements impliquant une navigation visuomotrice et le contrôle d'un bras robotique pour la préhension d'objets. La difficulté est de pouvoir synchroniser les différentes actions afin que le robot agisse de manière cohérente. Les comportements erronés du robot sont détectés grâce à l'évaluation des actions proposées par le modèle vis à vis des corrections imposées par le professeur humain. Un apprentissage de ces situations sous la forme de contextes multimodaux modulant la sélection d'action permet alors d'adapter le comportement afin que le robot reproduise la tâche désirée.Pour finir, nous présentons les perspectives de ce travail en terme de contrôle sensorimoteur, pour la navigation comme pour le contrôle d'un bras robotique, et son extension aux questions d'interface homme/robot. Nous insistons sur le fait que différents types d'imitation peuvent être le fruit des propriétés émergentes d'une architecture de contrôle sensorimotrice. / An autonomous robot collaborating with humans should be able to learn how to navigate and manipulate objects in the same task. In a classical approach, independent functional modules are considered to manage the different aspects of the task (navigation, arm control,...) . To the contrary, the goal of this thesis is to show that learning tasks of different kinds can be tackled by learning sensorimotor attractors from a few task nonspecific structures. We thus proposed an architecture which can learn and encode attractors to perform navigation tasks as well as arm control.We started by considering a model inspired from place-cells for navigation of autonomous robots. On-line and interactive learning of place-action couples can let attraction basins emerge, allowing an autonomous robot to follow a trajectory. The robot behavior can be corrected and guided by interacting with it. The successive corrections and their sensorimotor coding enables to define the attraction basin of the trajectory. My first contribution was to adapt this principle of sensorimotor attractor building for the impedance control of a robot arm. While a proprioceptive posture is maintained, the arm movements can be corrected by modifying on-line the motor command expressed as muscular activations. The resulting motor attractors are simple associations between the proprioceptive information of the arm and these motor commands. I then showed that the robot could learn visuomotor attractors by combining the proprioceptive and visual information with the motor attractors. The visuomotor control corresponds to a homeostatic system trying to maintain an equilibrium between the two kinds of information. In the case of ambiguous visual information, the robot may perceive an external stimulus (e.g. a human hand) as its own hand. According to the principle of homeostasis, the robot will act to reduce the incoherence between this external information and its proprioceptive information. It then displays a behavior of immediately observed gestures imitation. This mechanism of homeostasis, completed by a memory of the observed sequences and action inhibition capability during the observation phase, enables a robot to perform deferred imitation and learn by observation. In the case of more complex tasks, we also showed that learning transitions can be the basis for learning sequences of gestures, like in the case of cognitive map learning in navigation. The use of motivational contexts then enables to choose between different learned sequences.We then addressed the issue of integrating in the same architecture behaviors involving visuomotor navigation and robotic arm control to grab objects. The difficulty is to be able to synchronize the different actions so the robot act coherently. Erroneous behaviors of the robot are detected by evaluating the actions predicted by the model with respect to corrections forced by the human teacher. These situations can be learned as multimodal contexts modulating the action selection process in order to adapt the behavior so the robot reproduces the desired task.Finally, we will present the perspectives of this work in terms of sensorimotor control, for both navigation and robotic arm control, and its link to human robot interface issues. We will also insist on the fact that different kinds of imitation behavior can result from the emergent properties of a sensorimotor control architecture.

In the area of image recognition, the so-called deep neural networks belong to the most promising models these days. They often achieve considerably better results than traditional techniques even without the necessity of any excessive task-oriented preprocessing. This thesis is devoted to the study and analysis of three basic variants of deep neural networks-namely the neocognitron, convolutional neural networks, and deep belief networks. Based on extensive testing of the described models on the standard task of handwritten digit recognition, the convolutional neural networks seem to be most suitable for the recognition of general image data. Therefore, we have used them also to classify images from two very large data sets-CIFAR-10 and ImageNet. In order to optimize the architecture of the applied networks, we have proposed a new pruning algorithm based on the Principal Component Analysis. Powered by TCPDF (www.tcpdf.org)

La Radio Logicielle (SDR : Software Defined Radio) et la Radio Cognitive (CR : Cognitive Radio) deviennent d'un usage courant car elles répondent à plusieurs enjeux technico-économiques majeurs dans le domaine des télécommunications. Ces systèmes radio permettent de combler l'écart de développement technologique qui existe entre la partie matérielle et la partie logicielle des systèmes de communication, en permettant la gestion optimale des bandes de fréquences sous-utilisées par la commutation en temps réel d'une configuration radio à une autre. Dans ce cadre, cette thèse présente la mise en œuvre d'une chaîne de traitements Radio Logicielle (appelée SDR waveform) dans un Système Multiprocesseurs sur Puce (MPSoC) à usage général (implémenté dans un FPGA de type Xilinx Virtex-4). Cette plateforme est basée autour d'un Réseau sur Puce (NoC) interconnectant 16 processeurs élémentaires (appelés PE) disposant de quatre blocs-mémoires externes DDR2. Nous avons proposé des implémentations temps réel et embarquées sur MPSoC de différentes briques de traitements d'une chaîne SDR, en concevant une stratégie efficace de parallélisation et de synchronisation pour chaque composante élémentaire de la « waveform ». Nous avons amélioré la fonctionnalité de la chaîne de traitement Radio Logicielle, en intégrant un Transceiver reconfigurable basé sur différents modèles de Réseaux de Neurones Artificiels (RNA) : les Cartes Auto-Organisatrices (SOM), les Réseaux de Neurones Compétitifs (LVQ) et enfin les Réseaux Multi-Couches de Perceptrons (MLP). Ces trois RNA permettent la reconnaissance de la norme spécifique basée sur les paramètres d'entrée extraits du signal et la reconfiguration du Transceiver de CR. La solution adaptative que nous avons proposée commute vers le RNA le plus approprié en fonction des caractéristiques du signal d'entrée détecté. Il est important de pouvoir prendre en compte des signaux complexes et multi-porteuses. Dans ce cadre, nous avons adressé le cas d'un signal complexe composé de plusieurs porteuses, ainsi en divisant les PEs en différents groupes indépendants, nous affectons chaque groupe de PEs au traitement d'une nouvelle porteuse. Nous avons conçu une stratégie efficace de synchronisation et de parallélisation de ces trois RNA pour CR Transceiver. Nous l'avons appliquée, par la suite pour l'implantation de nos algorithmes sur le MPSoC déjà cité. L'accélération que nous obtenons pour la SDR waveform et pour les algorithmes de Transceiver de CR démontre que les MPSoC à usage général sont une réponse pertinente, entre autres, aux contraintes de performances sur une telle plateforme. Le système que nous proposons apporte une réponse aux défis technico-économiques des grandes entreprises qui investissent ou prévoient d'investir dans des équipements basés sur des SDR ou des CR, puisqu'il permet d'éviter de recourir à des équipements d'accélération coûteux. Nous avons, par la suite, ajouté d'autres fonctionnalités à notre waveform avec un « CR Transceiver multinormes », en proposant une nouvelle approche pour la gestion du spectre radio. Ceci étant l'aspect le plus important de CR. Nous rendons ainsi notre waveform spectralement efficace en modélisant les caractéristiques radiofréquences (RF) du signal utilisateur primaire sous la forme d'une série temporelle multi-variée. Cette série temporelle est ensuite fournie comme entrée dans un Réseau de Neurones Récurrent d'Elman (ERNN) qui prédit l'évolution de la série temporelle de RF pour déterminer si l'utilisateur secondaire peut exploiter la bande de fréquences. Nous avons exploité la cyclo-stationnarité inhérente des signaux primaires pour la Modélisation Non-Linéaire Autorégressive Exogène (NARX : Non-linear AutoRegressive Exogenous) des séries temporelles des caractéristiques RF, car la prédiction d'une caractéristique RF demande d'abord de connaître les autres caractéristiques radios pertinentes. Nous avons observé une tendance similaire pour les valeurs prédites et observées. En résumé, nous avons proposé des algorithmes pour SDR waveform à efficacité spectrale avec un Transceiver Universel, ainsi que leurs implantations parallèles sur MPSoC. Notre conception de waveform répond aux exigences en performances et aux contraintes de ressources embarquées des applications dans le domaine / Software Defined Radio (SDR) and Cognitive Radio (CR) are entering mainstream. These high performance and high adaptability requiring devices with agile frequency operations hold promise to :1. address the inconsistency between hardware and software advancements, 2. real time mode switching from one radio configuration to another and3. efficient spectrum management in under-utilized spectrum bands. Framed within this statement, in this thesis we have implemented a SDR waveform on 16 Processing Element (PE) Network on chip (NoC) based general purpose Multiprocessors System on chip (MPSoC), with access to four external DDR2 memory banks, which is implemented on a single chip Xilinx Virtex-4 FPGA. We shifted short term development of a waveform into software domain by designing an efficient parallelization and synchronization strategy for each waveform component, individually. We enhance our designed waveform functionality by proposing and implementing three Artificial Neural Networks Schemes : Self Organizing Maps, Linear Vector Quantization and Multi-Layer Perceptrons as effective techniques for reconfiguring CR Transceiver after recognizing the specific standard based on input parameters, pertaining to different layers, extracted from the signal. Our proposed adaptive solution switches to appropriate Artificial Neural Network, based on the features of input signal sensed. We designed an efficient synchronization and parallelization strategy to implement the Artificial Neural Networks based CR Transceiver Algorithms on the aforementioned MPSoC chip. The speed up we obtained for our SDR waveform and CR Transceiver algorithms demonstrated that the general purpose MPSoC devices are the most efficient answer to the acquisition challenge for major organizations that invest or plan to invest in SDR and CR based devices, thereby allowing us to avoid expensive hardware accelerators. We address the case of a complex signal composed of many modulated carriers by dividing the PEs in individual groups, thus received signal with more than one Standard is processed efficiently. We add further functionality in our designed Multi-standard CR Transceiver possessing SDR Waveform by proposing a new approach for radio spectrum management, perhaps the most important aspect of CR. We make our designed waveform Spectrum efficient by modelling the primary user signal Radio Frequency features as a multivariate time series, which is then given as input to Elman Recurrent Neural Network that predicts the evolution of Radio Frequency Time Series to decide if the secondary user can exploit the Spectrum band. We exploit the inherent cyclostationary in primary signals for Non-linear Autoregressive Exogenous Time Series Modeling of Radio Frequency features, as predicting one RF feature needs the previous knowledge of other relevant RF features. We observe a similar trend between predicted and actual values. Ensemble, our designed Spectrum Efficient SDR waveform with a Universal Multi-standard Transceiver answers the SDR and CR performance requirements under resource constraints by efficient algorithm design and implementation using lateral thinking that seeks a greater cross-domain interaction

Dans cette thèse, on s'est intéressé à trois problématiques des réseaux de capteurs sans fil (WSN). Dans un premier temps nous avons analysé l'impact de la chute de tension dans la batterie du nœud sur la puissance du signal en réception. On propose alors une méthode pour compenser l'augmentation apparente de la distance calculée entre les nœuds due à la diminution de l'énergie de la batterie. Pour les nœuds passant par deux états principaux endormi et actif, on propose d'étudier, la relation entre la diminution de la tension de la batterie en fonction du temps passé par un nœud dans l'état actif. Ensuite, on calcule le rapport entre la RSS et la distance entre les nœuds connectés avec des batteries complètement chargées. Après on mesure la RSS en faisant varier la tension de la batterie du nœud émetteur et en gardant le nœud récepteur à une distance constante. Finalement, on propose une relation entre la RSS observée et la tension actuelle de la batterie du nœud émetteur. Cette fonction permet de calculer la valeur corrigée de la RSS qui correspond à la distance réelle entre les nœuds connectés. Ainsi l'efficacité des méthodes de la localisation basée sur la RSS se trouvent améliorées. Dans la deuxième partie de cette thèse on propose une méthode d'estimation des positions des nœuds dans un WSN. Dans l'algorithme de localisation proposé, on utilise des nœuds ancres comme des points de référence. On a utilisé une approche heuristique pour trouver la topologie relative avec l'aide de la matrice de distance. Le but de la matrice de distance est d'indiquer s'il existe une connexion entre une paire de nœuds donnée et en cas de connectivité, la distance estimée entre ces nœuds. En utilisant les informations de connectivité entre les nœuds et leurs distances, on obtient la topologie du réseau. La méthode proposée utilise la solution de l'intersection de deux cercles au lieu de la méthode classique de triangulation, où un système quadratique de trois équations avec deux variables est utilisé ce qui rend la complexité de calcul augmentée. Lorsque deux nœuds connectés ont un autre nœud en commun, puis en utilisant les informations de distances entre ces nœuds interconnectés, nous pouvons calculer deux positions possibles pour le troisième nœud. La présence ou l'absence d'un lien entre le troisième nœud et un quatrième nœud, permet de trouver la position précise. Ce processus est réitéré jusqu'à ce que toutes les positions des nœuds aient été obtenues. Une fois la topologie relative calculée, il faut trouver la symétrie, l'orientation et la position de cette topologie dans le plan. C'est à ce moment que la connaissance des positions des trois nœuds entre en action. La topologie donne les coordonnées temporaires des nœuds. En ayant une comparaison de certaines caractéristiques entre les coordonnées temporaires et les coordonnées exactes, on trouve d'abord la symétrie de la topologie relative qui correspondrait à la topologie originale. En d'autres termes on vérifie si oui ou non la topologie relative est une image miroir de la topologie originale. Des opérateurs géométriques sont alors utilisés pour corriger la topologie relative par rapport à la topologie réelle. Ainsi, on localise tous les nœuds dans un WSN en utilisant exactement trois ancres. Dans la dernière partie de cette thèse, on propose une méthode pour la détection de défauts dans un WSN. Il y a toujours une possibilité qu'un capteur d'un nœud ne donne pas toujours des mesures précises. On utilise des systèmes récurrents et non récurrents pour la modélisation et on prend comme variable d'entrée, en plus des variables du nœud en question, les informations des capteurs voisins. La différence entre la valeur estimée et celle mesurée est utilisée pour déterminer la possibilité de défaillance d'un nœud / In this thesis three themes related to wireless sensor networks (WSNs) are covered. The first one concerns the power loss in a node signal due to voltage droop in the battery of the node. In the first part of the thesis a method is proposed to compensate for the apparent increase in the calculated distance between the related nodes due to decrease in the energy of the signal sending node battery. A function is proposed whose arguments are the apparently observed RSS and the current voltage of the emitter node battery. The return of the function is the corrected RSS that corresponds to the actual distance amongst the connected nodes. Hence increasing the efficiency of the RSS based localization methods in WSNs. In the second part of the thesis a position estimation method for localization of nodes in a WSN is proposed. In the proposed localization algorithm anchor nodes are used as landmark points. The localization method proposed here does not require any constraint on the placement of the anchors; rather any three randomly chosen nodes can serve as anchors. A heuristic approach is used to find the relative topology with the help of distance matrix. The purpose of the distance matrix is to indicate whether or not a pair of nodes has a connection between them and in case of connectivity it gives the estimated distance between the nodes. By using the information of connectivity between the nodes and their respective distances the topology of the nodes is calculated. This method is heuristic because it uses the point solution from the intersection of two circles instead of conventional triangulation method, where a system of three quadratic equations in two variables is used whereby the computational complexity of the position estimation method is increased. When two connected nodes have another node in common, then by using the information of distances between these interconnected nodes, two possible positions are calculated for the third node. The presence or absence of a connection between the third node and a fourth node helps in finding the accurate possibility out of the two. This process is iterated till all the nodes have been relatively placed. Once the relative topology has been calculated, we need to find the exact symmetry, orientation, and position of this topology in the plane. It is at this moment the knowledge of three nodes positions comes into action. From the relative topology we know the temporary coordinates of the nodes. By having a comparison of certain characteristics between the temporary coordinates and the exact coordinates; first the symmetry of relative topology is obtained that would correspond to the original topology. In other words it tells whether or not the relative topology is a mirror image of the original topology. Some geometrical operators are used to correct the topology position and orientation. Thus, all the nodes in the WSN are localized using exactly three anchors. The last part of the thesis focuses on the detection of faults in a WSN. There is always a possibility that a sensor of a node is not giving accurate measurements all of the time. Therefore, it is necessary to find if a node has developed a faulty sensor. With the precise information about the sensor health, one can determine the extent of reliance on its sensor measurement. To equip a node with multiple sensors is not an economical solution. Thus the sensor measurements of a node are modeled with the help of the fuzzy inference system (FIS). For each node, both recurrent and non-recurrent systems are used to model its sensor measurement. An FIS for a particular node is trained with input variables as the actual sensor measurements of the neighbor nodes and with output variable as the real sensor measurements of that node. The difference between the FIS approximated value and the actual measurement of the sensor is used as an indication for whether or not to declare a node as faulty