De nombreux travaux de recherches sont menés depuis quelques années dans le but de fournir une solution précise et intègre au problème de la localisation de véhicules routiers. Ces recherches sont en majorité fondées sur la théorie probabiliste de l’estimation. Elles utilisent la fusion multi-capteurs et le filtrage de Kalman mono-modèle, au travers de variantes adaptées aux systèmes non linéaires ; l’unique modèle complexe étant supposé décrire toute la dynamique du véhicule. Nous proposons dans cette thèse une approche multi-modèles. Cette étude dérive d’une analyse modulaire de la dynamique du véhicule, c’est-à-dire que l’espace d’évolution est pris comme un espace discret : plusieurs modèles simples et dédiés chacun à une manœuvre particulière sont générés, ce qui améliore la robustesse face aux défauts de modélisation du système. Il s’agit d’une variante de l’algorithme IMM, qui prend en compte l’asynchronisme des capteurs embarqués dans le processus d’estimation de l’état du véhicule. Pour cela, une nouvelle modélisation sous contraintes est développée, ce qui permet de mettre à jour la vraisemblance des modèles intégrés même en l’absence de mesures provenant de capteurs extéroceptifs. Toutefois, la performance d’un tel système nécessite d’utiliser des données capteurs de bonne qualité. Plusieurs opérations sont présentées, illustrant la correction du biais des capteurs, des bruits de mesures ainsi que la prise en compte de l’angle de dévers de la chaussée. La méthodologie développée est validée à travers une comparaison avec les algorithmes de fusion probabilistes EKF, UKF, DD1, DD2 et le filtrage particulaire. Cette comparaison est fondée sur des mesures courantes de précision et de confiance, puis sur l’utilisation de critères statistiques de consistance et de crédibilité, à partir de scénarios synthétiques et ensuite des données réelles. / Many research works have been devoted in the last years in order to provide an accurate and high integrity solution to the problem outdoor vehicles localization. These research efforts are mainly based on the probability estimation theory. They use multi-sensor fusion approach and a single-model based Kalman filtering, through some variants adapted to nonlinear systems. The single complex model that is used is assumed to describe the dynamics of the vehicle. We rather propose a multiple model approach in this thesis. The presented study derives from a modular analysis of the dynamics of the vehicle, ie the evolution of the vehicle is considered as a discrete process, which combines several simple models. Each model is dedicated to a particular manoeuvre of the vehicle. This evolution space discretizing will improves the system robustness to modelling defects. Our approach is a variant of the IMM algorithm, which takes into account the asynchronism of the embedded sensors. In order to achieve this goal, a new system constrained modelling is developed, which allows to update the various models likelihood even in absence of exteroceptive sensors. However, the performance of such a system requires the use of good quality data. Several operations are presented, illustrating the corrections on the sensors bias, measurements noise and taking into account the road bank angle. The developed methodology is validated through a comparison with the probabilistic fusion algorithms EKF, UKF, DD1, DD2 and particle filtering. This comparison is based on measurements of accuracy and confidence, then the use of statistical consistency and credibility measures, from simulation scenarios and then real data.

<p> This thesis investigates the application of multiple model estimation algorithms to the problem of channel equalization for digital data transmission and channel tracking for space-time block coded systems with non-Gaussian additive noise. Recently, a network of Kalman filters (NKF) has been reported for the equalization of digital communication channels based on the approximation of the a posteriori probability density function of a sequence of delayed symbols by a weighted Gaussian sum. A serious drawback of this approach is that the number of Gaussian terms in the sum increases exponentially through iterations. In this thesis, firstly, we have shown that the NKF-based equalizer can be further improved by considering the interactions between the parallel filters in an efficient way. To this end, we take resort to the Interacting Multiple Model (IMM) estimator widely used in the area of multiple target tracking. The IMM is a very effective approach when the system exhibits discrete uncertainties in the dynamic or measurement model as well as continuous uncertainties in state values. A computationally feasible implementation based on a weighted sum of Gaussian approximation of the density functions of the data signals is introduced. Next, we present an adaptive multiple model blind equalization algorithm based on the IMM estimator to estimate the channel and the transmitted sequence corrupted by intersymbol interference and noise. It is shown through simulations that the proposed IMM-based equalizer offers substantially improved performance relative to the blind equalizer based on a (static or non-interacting) network of extended Kalman filters. It obviates the exponential growth of the state complexity caused by increasing channel memory length. The proposed approaches avoid the exponential growth of the number of terms used in the weighted Gaussian sum approximation of the plant noise making it practical for real-time processing.</p> <p> Finally, we consider the problem of channel estimation and tracking for space-time block coded systems contaminated by additive non-Gaussian noise. In many practical wireless channels in which space-time block coding techniques may be applied, the ambient noise is likely to have an impulsive component that gives rise to larger tail probabilities than is predicted by the Gaussian model. Although Kalman filters are often used in practice to track the channel variation, they are notoriously sensitive to heavy-tailed outliers and model mismatches resulting from the presence of impulsive noise. Non-Gaussian noise environments require the modification of standard filters to perform acceptably. Based on the coding/decoding technique, we propose a robust IMM algorithm approach in estimating time-selective fading channels when the measurements are perturbed by the presence of impulsive noise. The impulsive noise is modeled by a two terms Gaussian mixture distribution. Simulations demonstrate that the proposed method yields substantially improved performance compared to the conventional Kalman filter algorithm using the clipping or localization approaches to handle impulses in the observation. It is also shown that IMM-based approach performs robustly even when the prior information about the impulsive noise is not known exactly.</p> / Thesis / Master of Applied Science (MASc)

Hybrid systems have been identified as one of the main directions in control theory and attracted increasing attention in recent years due to their huge diversity of engineering applications. Multiplemodel (MM) estimation is the state-of-the-art approach to many hybrid estimation problems. Existing MM methods with fixed structure usually perform well for problems that can be handled by a small set of models. However, their performance is limited when the required number of models to achieve a satisfactory accuracy is large due to time evolution of the true mode over a large continuous space. In this research, variable-structure multiple model (VSMM) estimation was investigated, further developed and evaluated. A fundamental solution for on-line adaptation of model sets was developed as well as several VSMM algorithms. These algorithms have been successfully applied to the fields of fault detection and identification as well as target tracking in this thesis. In particular, an integrated framework to detect, identify and estimate failures is developed based on the VSMM. It can handle sequential failures and multiple failures by sensors or actuators. Fault detection and target maneuver detection can be formulated as change-point detection problems in statistics. It is of great importance to have the quickest detection of such mode changes in a hybrid system. Traditional maneuver detectors based on simplistic models are not optimal and are computationally demanding due to the requirement of batch processing. In this presentation, a general sequential testing procedure is proposed for maneuver detection based on advanced sequential tests. It uses a likelihood marginalization technique to cope with the difficulty that the target accelerations are unknown. The approach essentially utilizes a priori information about the accelerations in typical tracking engagements and thus allows improved detection performance. The proposed approach is applicable to change-point detection problems under similar formulation, such as fault detection.