L’identification des systèmes par modèle non entier a été initiée dans les années 1990 et de nombreux résultats ont été obtenus depuis. Néanmoins, la plupart de ces résultats utilise les méthodes de la famille des méthodes à erreur de prédiction, basées sur la minimisation de la norme ℓ2 de l’erreur d’estimation. Apparues en 1996, les méthodes des sous-espaces sont relativement nouvelles dans la théorie de l’identification de systèmes linéaires. Basées sur des projections géométriques et l’algèbre linéaire, elles présentent une alternative intéressante aux méthodes classiques basées sur la régression linéaire ou non linéaire. Elles permettent d’estimer les matrices d’un modèle à base d’une représentation d’état. Dans le contexte des systèmes non entiers, la notion de pseudo-représentation d’état généralise la notion de représentation d’état en introduisant un paramètre supplémentaire qui est l’ordre commensurable.Actuellement, la méthode des sous-espaces pour des systèmes non entiers n’a cependant été appliquée que dans le domaine temporel. Elle est alors développée dans cette thèse pour une telle classe de systèmes dans le domaine fréquentiel. De plus, comme les systèmes non entiers sont des systèmes à temps continu, un filtrage des données est nécessaire pour respecter la causalité des signaux et pour pouvoir réaliser l’identification. Une étude comparative des différentes méthodes de filtrage dans le contexte de l’identification pour déduire leurs avantages et inconvénients est réalisée dans le domaine temporel. Enfin,les méthodes développées ont été appliquées à un système réel en diffusion thermique.Les modèles obtenus sont généralisés à des matériaux soumis à plusieurs flux de chaleur en entrée tout en considérant leur température en plusieurs points de mesures. / The identification of systems by fractional models was initiated in the 1990s and various results have been obtained since. Nevertheless, most of these results are based on prediction error methods (PEM) of identification, based on the minimization of the norm of the estimation error. Apparent in 1996, the subspace methods are relatively new in the theory of the identification of linear systems. Based on geometric projections and linear algebra, they present an alternative to classical methods based on linear or nonlinear regression. They allow estimating the matrices of the state-space representation of a system. In the context of fractional systems, a pseudo-state-space representation generalizes the notion of state-space representation by introducing an additional parameter which is the commensurable order.Currently, the subspace method for non-integer systems has only been applied inthe time domain. It is then developed in this thesis for such a class of systems in the frequency domain. Moreover, since non-integer systems are continuous time systems, datapre-filtering is necessary to respect the causality of the signals and to be able to realize the identification. A study of the different filtering methods in the context of subspaceidentification is then carried out in order to deduce their advantages and disadvantages in the time domain. Finally, the method has been applied to a thermal diffusion system.The obtained models are generalized for several input heat flows, considering their temperature available at several measurement points.

The human genome contains tens of thousands of gene loci which code for an even greater number of protein and RNA products. The highly complex temporal and spatial expression of these genes makes possible all the biological processes of life. Altered gene expression by mutation or deregulation is fundamental for the development of many human diseases. The ultimate aim of this thesis was to identify gene expression changes relevant to cancer. The advent of genome-wide expression profiling techniques, such as microarrays, has provided powerful new tools to identify such changes and researchers are now faced with an explosion of gene expression data. Processing, comparing and integrating these data present major challenges. I approached these challenges by developing and assessing novel methods for cross-platform analysis of expression data, scalable subspace clustering, and curation of experimental gene regulation data from the published literature. I found that combining results from different expression platforms increases reliability of coexpression predictions. However, I also observed that global correlation between platforms was generally low, and few gene pairs reached reasonable thresholds for high-confidence coexpression. Therefore, I developed a novel subspace clustering algorithm, able to identify coexpressed genes in experimental subsets of very large gene expression datasets. Biological assessment against several metrics indicates that this algorithm performs well. I also developed a novel meta-analysis method to identify consistently reported genes from differential expression studies when raw data are unavailable. This method was applied to thyroid cancer, producing a ranked list of significantly over-represented genes. Tissue microarray analysis of some of these candidates and others identified a number of promising biomarkers for diagnostic and prognostic classification of thyroid cancer. Finally, I present ORegAnno (www.oreganno.org), a resource for the community-driven curation of experimentally verified regulatory sequences. This resource has proven a great success with ~30,000 sequences entered from over 900 publications by ~50 contributing users. These data, methods and resources contribute to our overall understanding of gene regulation, gene expression, and the changes that occur in cancer. Such an understanding should help identify new cancer mechanisms, potential treatment targets, and have significant diagnostic and prognostic implications. / Medicine, Faculty of / Medical Genetics, Department of / Graduate

The estimation of power spectral densities is a critical step in many speech enhancement algorithms. The demand for multi-channel speech enhancement systems is high with applications in teleconferencing, cellular phones, and hearing aids. The first objective of the thesis is to develop a general multi-channel framework to solve for the diffuse noise power spectral densities whenever the spatial correlation or coherence matrix is pre-estimated and the number of speakers is less than the number of microphones. The second objective is to develop closed-form analytical solutions. The performance of the developed algorithms is evaluated with pre-existing algorithms using prescribed performance measures.

This work focuses on the construction of functional representations in high-dimensional spaces.Attention is focused on the modeling of ignition phenomena using detailed kinetics, and on the ignition delay time as the primary quantity of interest (QoI). An iso-octane air mixture is first considered, using a detailed chemical mechanism with 3,811 elementary reactions. Uncertainty in all reaction rates is directly accounted for using associated uncertainty factors, assuming independent log-uniform priors. A Latin hypercube sample (LHS) of the ignition delay times was first generated, and the resulting database was then exploited to assess the possibility of constructing polynomial chaos (PC) representations in terms of the canonical random variables parametrizing the uncertain rates. We explored two avenues, namely sparse regression (SR) using LASSO, and a coordinate transform (CT) approach. Preconditioned variants of both approaches were also considered, namely using the logarithm of the ignition delay time as QoI. A global sensitivity analysis is performed using the representations constructed by SR and CT. Next, the tangent linear approximation is developed to estimate the sensitivity of the ignition delay time with respect to individual rate parameters in a detailed chemical mechanism. Attention is focused on a gas mixture reacting under adiabatic, constant-volume conditions. The approach is based on integrating the linearized system of equations governing the evolution of the partial derivatives of the state vector with respect to individual random variables, and a linearized approximation is developed to relate the ignition delay sensitivity to the scaled partial derivatives of temperature. In particular, the computations indicate that for detailed reaction mechanisms the TLA leads to robust local sensitivity predictions at a computational cost that is order-of-magnitude smaller than that incurred by finite-difference approaches based on one-at-a-time rate parameters perturbations. In the last part, we explore the potential of utilizing TLA-based sensitivities to identify active subspace and to construct suitable representations. Performance is assessed based contrasting experiences with CT-based machinery developed earlier.

Motion segmentation is part of the computer vision field and aims to find the moving parts in a video sequence. It is used in applications such as autonomous driving, surveillance, robotics, human motion analysis, and video indexing. Since there are so many applications, motion segmentation is ill-defined and the research field is vast. Despite the advances in the research over the years, the existing methods are still far behind human capabilities. Problems such as changes in illumination, camera motion, noise, mixtures of motion, missing data, and occlusion remain challenges. Feature-based approaches have grown in popularity over the years, especially manifold clustering methods due to their strong mathematical foundation. Methods exploiting sparse and low-rank representations are often used since the dimensionality of the data is reduced while useful information regarding the motion segments is extracted. However, these methods are unable to effectively handle large and complete occlusions as well as missing data since they tend to fail when the amount of missing data becomes too large. An algorithm based on Sparse Subspace Clustering (SSC) has been proposed to address the issue of occlusions and missing data so that SSC can handle these cases with high accuracy. A frame-to-frame analysis was adopted as a pre-processing step to identify motion segments between consecutive frames, called inter-frame motion segments. The pre-processing step is called Multiple Split-And-Merge (MSAM), which is based on the classic top-down split-and-merge algorithm. Only points present in both frame pairs are segmented. This means that a point undergoing an occlusion is only assigned to a motion class when it has been visible for two consecutive frames after re-entering the camera view. Once all the inter-frame segments have been extracted, the results are combined in a single matrix and used as the input for the classic SSC algorithm. Therefore, SSC segments inter-frame motion segments rather than point trajectories. The resulting algorithm is referred to as MSAM-SSC. MSAM-SSC outperformed some of the most popular manifold clustering methods on the Hopkins155 and KT3DMoSeg datasets. It was also able to handle complete occlusions and 50% missing data sequences, as well as outliers. The algorithm can handle mixtures of motions and different numbers of motions. However, it was found that MSAM-SSC is more suited for traffic and articulate motion scenes which are often used in applications such as robotics, surveillance, and autonomous driving. For future work, the algorithm can be optimised to reduce the execution time so that it can be used for real-time applications. Additionally, the number of moving objects in the scene can be estimated to obtain a method that does not rely on prior knowledge. / Dissertation (MEng (Computer Engineering))--University of Pretoria, 2021. / CSIR / Electrical, Electronic and Computer Engineering / MEng (Computer Engineering) / Unrestricted

This thesis proposes new algorithms for a group of sensing robots to learn a para- metric model for a dynamic spatio-temporal field, then based on the learned model trajectories are planned for sensing robots to best estimate the field. In this thesis we call these two parts learning and monitoring, respectively. For the learning, we first introduce a parametric model for the spatio-temporal field. We then propose a family of motion strategies that can be used by a group of mobile sensing robots to collect point measurements about the field. Our motion strategies are designed to collect enough information from enough locations at enough different times for the robots to learn the dynamics of the field. In conjunction with these motion strategies, we propose a new learning algorithm based on subspace identification to learn the parameters of the dynamical model. We prove that as the number of data collected by the robots goes to infinity, the parameters learned by our algorithm will converge to the true parameters. For the monitoring, based on the model learned from the learning part, three new informative trajectory planning algorithms are proposed for the robots to collect the most informative measurements for estimating the field. Kalman filter is used to calculate the estimate, and to compute the error covariance of the estimate. The goal is to find trajectories for sensing robots that minimize a cost metric on the error covariance matrix. We propose three algorithms to deal with this problem. First, we propose a new randomized path planning algorithm called Rapidly-exploring Random Cycles (RRC) and its variant RRC* to find periodic trajectories for the sensing robots that try to minimize the largest eigenvalue of the error covariance matrix over an infinite horizon. The algorithm is proven to find the minimum infinite horizon cost cycle in a graph, which grows by successively adding random points. Secondly, we apply kinodynamic RRT* to plan continuous trajectories to estimate the field. We formulate the evolution of the estimation error covariance matrix as a differential constraint and propose extended state space and task space sampling to fit this problem into classical RRT* setup. Thirdly, Pontryagin’s Minimum Principle is used to find a set of necessary conditions that must be satisfied by the optimal trajectory to estimate the field. We then consider a real physical spatio-temporal field, the surface water temper- ature in the Caribbean Sea. We first apply the learning algorithm to learn a linear dynamical model for the temperature. Then based on the learned model, RRC and RRC* are used to plan trajectories to estimate the temperature. The estimation performance of RRC and RRC* trajectories significantly outperform the trajectories planned by random search, greedy and receding horizon algorithms.

Radio frequency interference (RFI) mitigation enables radio astronomical observation in frequency bands that are shared with many modern satellite and ground based devices by filtering out the interference in corrupted bands. The present work documents the development of a beamformer (spatial filter) equipped with RFI mitigation capabilities. The beamformer is intended for systems with antenna arrays designed for large bandwidths. Because array data post processing on large bandwidths would require massive memory space beyond feasible limits, there is a need for a RFI mitigation system capable of doing processing on the data as it arrives in real-time; storing only a data reduced result into long term memory. The real-time system is designed to be implemented on both the FLAG phased array feed (PAF) on the Green Bank telescope in West Virginia, as well as future radio astronomy projects. It will also serve as the anti-jamming component in communications applications developed for the United States office of naval research (ONR). Implemented on a graphical processing unit (GPU), this beamformer demonstrates a working single step filter using nVidia's CUDA technology, technology with high-speed parallelism that makes real-time RFI mitigation possible.

We examine the related nature of two identification algorithms, subspace identification (SID) and Dynamic Mode Decomposition (DMD), and their correctness properties over a broad range of problems. This investigation begins by noting the strong relationship between the two algorithms, both drawing significantly on the pseudoinverse calculation using singular value decomposition, and ultimately revealing that DMD can be viewed as a substep of SID. We then perform extensive computational studies, characterizing the performance of SID on problems of various model orders and noise levels. Specifically, we generate 10,000 random systems for each model order and noise level, calculating the average identification error for each case, and then repeat the entire experiment to ensure the results are, in fact, consistent. The results both quantify the intrinsic algorithmic error at zero-noise, monotonically increasing with model complexity, as well as demonstrate an asymptotically linear degradation to noise intensity, at least for the range under study. Finally, we close by demonstrating DMD's ability to recover system matrices, because its access to full state measurements makes them identifiable. SID, on the other hand, can't possibly hope to recover the original system matrices, due to their fundamental unidentifiability from input-output data. This is true even when SID delivers excellent performance identifying a correct set of equivalent system matrices.

No description available.

No description available.