Orientadora: Profa. Dra. Aline de Oliveira Neves Panazio / Dissertação (mestrado) - Universidade Federal do ABC, Programa de Pós-Graduação em Engenharia da Informação, Santo André, 2018. / Este trabalho dedica-se ao estudo de algoritmos de filtragem adaptativa autodidata no modo difusão, com aplicações em redes de sensores sem fio (RSSF). No modo difusão, os nós sensores da rede possuem poder de processamento local e trocam informações com seus vizinhos. Neste trabalho, propomos dois algoritmos utilizando como base o algoritmo CMA no modo Difusão (CMAD), com duas abordagens distintas da técnica Set-Membership. O primeiro baseia-se no algoritmo Set-Membership Least Mean Squares (SM-LMS), desenvolvido também no modo difusão. Estendemos o algoritmo para o contexto não supervisionado, denotando por Algoritmo Set-Membership CMA no modo Difusão (SM-CMAD). Mostramos que este algoritmo apresenta desempenho melhor ou similar ao CMAD, em termos de velocidade de convergência, patamar de interferência intersimbólica (IIS) e possuindo a importante vantagem de reduzir as trocas de informações entre os nós, economizando energia e recursos da rede. O segundo algoritmo proposto se baseia no Set-Membership do Módulo Constante (SM-CM), o qual estendemos para o contexto de redes de sensores sem fio no modo difusão. Tal algoritmo é denotado por Algoritmo Set-membership CMA no modo Difusão Square-root Gamma (SM-CMAD-SG). Novamente o algoritmo apresenta um bom desempenho quando comparado com o CMAD e, quando comparado ao SM-CMAD, vemos que sua principal vantagem está na economia em termos de atualizações dos coeficientes do filtro, que chega a valores acima de 70% em diversos cenários de simulação, sem grandes perdas de desempenho economizando energia. / This work is devoted to the study of unsupervised adaptive filtering algorithms in diffusion mode, with applications in wireless sensor networks (WSNs). In diffusion mode, network sensing nodes have local processing power and exchange information with their neighbors. In this work, we propose two algorithms based on the CMA algorithm in Diffusion mode (CMAD), with two different approaches to the Set-Membership technique. The first one is based on the Set-Membership Least Mean Squares (SM-LMS) algorithm, also developed in the diffusion mode. We extend the algorithm to the unsupervised context, denoting by Set-Membership CMA in Diffusion mode (SM-CMAD). We show that this algorithm presents better or similar performance to CMAD, in terms of convergence speed, intersymbol interference threshold (IIS), and has the important advantage of reducing the exchange of information between nodes, saving energy and network resources. The second proposed algorithm is based on the Set-Membership of the Constant Modulus (SM-CM), which we extend to the context of wireless sensor networks in the diffusion mode. This algorithm is denoted by the Set-membership CMA in Diffusion mode Square-root Gamma (SM-CMAD-SG). This algorithm performs well when compared to CMAD and, when compared to SM-CMAD, we see that its main advantage lies in the economy in terms of the update of the filter coefficients, which reaches values above 70% in several scenarios without loss of performance, saving energy.

Les premières recherches en détection, localisation et séparation de signaux remontent au début du 20ème siècle. Ces recherches sont d’actualité encore aujourd’hui, notamment du fait de la croissance rapide des systèmes de communications constatée ces deux dernières décennies. Par ailleurs, la littérature du domaine consacre très peu d’études relatives à certains contextes jugés difficiles dont certains sont traités dans cette thèse. Ce travail porte sur la localisation de signaux par détection des temps d’arrivée ou estimation des directions d’arrivée et sur la séparation de sources dépendantes ou à module constant. L’idée principale est de tirer profit de certaines informations a priori disponibles sur les signaux sources telles que la parcimonie, la cyclostationarité, la non-circularité, le module constant, la structure autoregressive et les séquences pilote dans un contexte coopératif. Une première partie détaille trois contributions : (i) un nouveau détecteur pour l’estimation des temps d’arrivée basé sur la minimisation de la probabilité d’erreur ; (ii) une estimation améliorée de la puissance du bruit, basée sur les statistiques d’ordre ; (iii) une quantification de la précision et de la résolution de l’estimation des directions d’arrivée au regard de certains a priori considérés sur les sources. Une deuxième partie est consacrée à la séparation de sources exploitant différentes informations sur celles-ci : (i) la séparation de signaux de communication à module constant ; (ii) la séparation de sources dépendantes connaissant la nature de la dépendance et (iii) la séparation de sources autorégressives dépendantes connaissant la structure autorégressive. / Signal detection, localization, and separation problems date back to the beginning of the twentieth century. Nowadays, this subject is still a hot topic receiving more and more attention, notably with the rapid growth of wireless communication systems that arose in the last two decades and it turns out that many challenging aspects remain poorly addressed by the available literature relative to this subject. This thesis deals with signal detection, localization using temporal or directional measurements, and separation of dependent source signals. The main objective is to make use of some available priors about the source signals such as sparsity, cyclo-stationarity, non-circularity, constant modulus, autoregressive structure or training sequences in a cooperative framework. The first part is devoted to the analysis of (i) signal’s time-of-arrival estimation using a new minimum error rate based detector, (ii) noise power estimation using an improved order-statistics estimator and (iii) side information impact on direction-of-arrival estimation accuracy and resolution. In the second part, the source separation problem is investigated at the light of different priors about the original sources. Three kinds of prior have been considered : (i) separation of constant modulus communication signals, (ii) separation of dependent source signals knowing their dependency structure and (iii) separation of dependent autoregressive sources knowing their autoregressive structure.