L'instrumentation permet à notre environnement, maison ou bâtiment, de devenir intelligent en s'adaptant à nos modes de vie et en nous assistant au quotidien. Un environnement intelligent est sensible et réactif à nos activités, afin d'améliorer notre qualité de vie. La fiabilité d'identification des activités est ainsi essentielle pour cette intelligence ambiante : elle est directement dépendante du positionnement des capteurs au sein de l'environnement. Cette question essentielle du placement des capteurs est très peu considérée par les systèmes ambiants commercialisés ou même dans la littérature. Pourtant, elle est la source principale de leurs dysfonctionnements où une mauvaise reconnaissance des activités entraîne une mauvaise assistance fournie. Le placement de capteurs consiste à choisir et à positionner des capteurs pertinents pour une identification fiable des activités. Dans cette thèse, nous développons et détaillons une méthodologie de placement de capteurs axée sur l'identifiabilité des activités d'intérêt. Nous la qualifions en nous intéressant à deux évaluations différentes : la couverture des intérêts et l'incertitude de mesures. Dans un premier temps, nous proposons un modèle de l'activité où nous décomposons l'activité en actions caractérisées afin d'être indépendant de toute technologie ambiante (axée connaissances ou données). Nous représentons actions et capteurs par un modèle ensembliste unifiant, permettant de fusionner des informations homogènes de capteurs hétérogènes. Nous en évaluons l'identifiabilité des actions d'intérêt au regard des capteurs placés, par des notions de précision (performance d'identification) et de sensibilité (couverture des actions). Notre algorithme de placement des capteurs utilise la Pareto-optimalité pour proposer une large palette de placements-solutions pertinents et variés, pour ces multiples identifiabilités à maximiser. Nous illustrons notre méthodologie et notre évaluation en utilisant des capteurs de présence, et en choisissant optimalement la caractéristique à couvrir pour chaque action. Dans un deuxième temps, nous nous intéressons à la planification optimale des expériences où l'analyse de la matrice d'information permet de quantifier l'influence des sources d'incertitudes sur l'identification d'une caractéristique d'action. Nous représentons les capteurs continus et l'action caractérisée par un modèle analytique, et montrons que certaines incertitudes doivent être prises en compte et intégrées dans une nouvelle matrice d'information. Nous y appliquons les indices d'observabilité directement pour évaluer l'identifiabilité d'une action caractérisée (incertitude d'identification). Nous illustrons cette évaluation alternative en utilisant des capteurs d'angle, et nous la comparons à la matrice d'information classique. Nous discutons des deux évaluations abordées et de leur complémentarité pour la conception d’environnement instrumenté pour la veille à la personne. / Instrumentation enables our environment, house or building, to get smart through self-adjustment to our lifestyles and through assistance of our daily-life. A smart environment is sensitive and responsive to our activities, in order to improve our quality of life. Reliability of activities' identification is absolutely necessary to such ambient intelligence: it depends directly on sensors' positioning within the environment. This fundamental issue of sensor placement is hardly considered by marketed ambient systems or even into the literature. Yet, it is the main source of ambient systems' malfunctions and failures, because a bad activity recognition leads to a bad delivered assistance. Sensor placement is about choosing and positioning relevant sensors for a reliable identification of activities. In this thesis, we develop and detail a sensor placement methodology driven by identifiability of activities of interest. We quantify it by looking at two different evaluations: coverage of interests and uncertainty of measures. First, we present an activity model that decomposes each activity into characterised actions to be technology-free (either knowledge or data driven one). We depict actions and sensors by a set theoretic model, enabling to fuse homogeneous informations of heterogeneous sensors. We then evaluate each action of interest's identifiability regarding placed sensors, through notions of precision (identification's performance) and sensitivity (action's coverage). Our sensor placement algorithm use Pareto-optimality to offer a wide range of relevant solution-placements, for these multiple identifiabilities to maximise. We showcase our methodology and our evaluation through solving a problem featuring motion and binary sensors, by optimally choosing for each action the characteristic to cover. Finally, we look into optimal design of experiments by analysing the information matrix to quantify how sources of uncertainties influence the identification of an action's characteristic. We depict continuous sensors and the characterised action by an analytical model, and we show that some uncertainties should be considered and included in a new information matrix. We then apply directly observability indexes to evaluate identifiability of a characterised action (uncertainty of identification), and compare our new information matrix to the classical one. We showcase our alternate evaluation through solving a sensor placement problem featuring angular sensors. We discuss both covered evaluations and their complementarity towards the design of instrumented environment for human monitoring.

This work investigates the application of active noise control (ANC) to speech. ANC has had success reducing tonal noise. In this work, that success was extended to noise that is not completely tonal but has some tonal elements such as speech. Limitations such as causality were established on the active control of human speech. An optimal configuration for control actuators was developed for a sphere using a genetic algorithm. The optimal error sensor location was found from exploring the nulls associated with the magnitude of the radiated pressure with reference to the primary pressure field. Both numerically predicted and experimentally validated results for the attenuation of single frequency tones were shown. The differences between the numerically predicted results for attenuation with a sphere present in the pressure field and monopoles in the free-field are also discussed.The attenuation from ANC of both monotone and natural speech is shown and a discussion about the effect of causality on the results is given. The sentence “Joe took father’s shoe bench out” was used for both monotone and natural speech. Over this entire monotone speech sentence, the average attenuation was 8.6 dB with a peak attenuation of 10.6 dB for the syllable “Joe”. Natural speech attenuation was 1.1 dB for the sentence average with a peak attenuation on the syllable “bench” of 2.4 dB. In addition to the lower attenuation values for natural speech, the pressure level for the word “took” was increased by 2.3 dB. Also, the harmonic at 420 Hz in the word “father’s” of monotone speech was reduced globally up to 20 dB. Based on the results of the attenuation of monotone and natural speech, it was concluded that a reasonable amount of attenuation could be achieved on natural speech if its correlation could approach that of monotone speech.

A deep scientific understanding of complex physical systems, such as the atmosphere, can be achieved neither by direct measurements nor by numerical simulations alone. Data assimilation is a rigorous procedure to fuse information from a priori knowledge of the system state, the physical laws governing the evolution of the system, and real measurements, all with associated error statistics. Data assimilation produces best (a posteriori) estimates of model states and parameter values, and results in considerably improved computer simulations. The acquisition and use of observations in data assimilation raises several important scientific questions related to optimal sensor network design, quantification of data impact, pruning redundant data, and identifying the most beneficial additional observations. These questions originate in operational data assimilation practice, and have started to attract considerable interest in the recent past. This dissertation advances the state of knowledge in four dimensional variational (4D-Var) - data assimilation by developing, implementing, and validating a novel computational framework for estimating observation impact and for optimizing sensor networks. The framework builds on the powerful methodologies of second-order adjoint modeling and the 4D-Var sensitivity equations. Efficient computational approaches for quantifying the observation impact include matrix free linear algebra algorithms and low-rank approximations of the sensitivities to observations. The sensor network configuration problem is formulated as a meta-optimization problem. Best values for parameters such as sensor location are obtained by optimizing a performance criterion, subject to the constraint posed by the 4D-Var optimization. Tractable computational solutions to this "optimization-constrained" optimization problem are provided. The results of this work can be directly applied to the deployment of intelligent sensors and adaptive observations, as well as to reducing the operating costs of measuring networks, while preserving their ability to capture the essential features of the system under consideration. / Ph. D.

In large-scale civil structures, a limited number of sensors are placed to monitor the health of civil structures to reduce maintenance, communication and energy costs. In this thesis, the problem of optimal sensor location placement to infer the health of civil structures is explored. First, a comparative study of approaches from the fields of control engineering and civil engineering is conducted . The widely used civil engineering approaches such as effective independence (EI) and modal assurance criterion (MAC) have limitations because of the negligence of modes and damping parameters. On the other hand, control engineering approaches consider the entire system dynamics using impulse response-type sensor measurement data. Such inference can be formulated as an estimation problem, with the dynamics formulated as a second-order differential equation. The comparative study suggests that damping dynamics play significant impact to the selection of best sensor location---the civil engineering approaches that neglect the damping dynamics lead to very different sensor locations from those of the control engineering approaches. In the second part of the thesis, an initial attempt to directly connect the topological graph of the structure (that defines the damping and stiffness matrices) and the second-order dynamics is conducted.

Recent experimental achievements in active noise control (ANC) for cooling fans have used near-field error sensors whose locations are determined according to a theoretical condition of minimized sound power. A theoretical point source model, based on the condition previously stated, reveals the location of near-field pressure nulls that may be used to optimize error sensor placement. The actual locations of these near-field pressure nulls for both an axial cooling fan and a monopole loudspeaker were measured over a two-dimensional grid with a linear array of microphones. The achieved global attenuation for each case is measured over a hemisphere located in the acoustic far field of the ANC system. The experimental results are compared to the theoretical pressure null locations in order to determine the efficacy of the point source model. The results closely matched the point source model with a loudspeaker as the primary source, and the sound power reduction was greatly reduced when error sensors were placed in non-ideal locations. A weakness of the current near-field modeling process is that a point monopole source is used to characterize the acoustic noise from an axial cooling fan, which may have multipole characteristics. A more complete characterization of fan noise may be obtained using a procedure based on the work of Martin and Roure [J. Sound Vib. 201 (5), 577--593 (1997)]. Pressure values are obtained over a hemisphere in the far field of a primary source and the contributions from point source distributions up to the second order, centered at the primary source, may be calculated using a multipole expansion. The source information is then used in the aforementioned theoretical near-field calculation of pressure. The error sensors are positioned using the complete fan characterization. The global far-field attenuation for the multipole expansion model of fan noise is compared to that of previous experiments. Results show that the multipole expansion model yields a more accurate representation the near field, but is not successful in achieving greater sound power reductions in the far field.