137091 research outputs found
Sort by
Accelerated Tensor Robust Algorithms for Hyperspectral Imaging and Video Processing
No full textIn recent years, the application of tensor-based methods to high-dimensional data has gained considerable attention, particularly for tasks involving denoising, classification, and compression of complex data structures such as hyperspectral images. This thesis presents novel approaches to enhance Tensor Robust Principal Component Analysis (TRPCA), addressing challenges such as computational efficiency, noise removal, and real-time processing.
Firstly, the thesis presents a new online robust principal component analysis (RPCA) algorithm that recursively decomposes incoming data into low-rank and sparse components. Unlike traditional approaches that operate on data vectors, this method preserves the multi-dimensional structure of data, such as video frames. It is based on the recently proposed tensor singular value decomposition (T-SVD) and incorporates a convex optimization-based approach for recovering the sparse component and updating the low-rank component using incremental T-SVD. An efficient tensor convolutional extension to the Fast Iterative Shrinkage Thresholding Algorithm (FISTA) is also proposed, significantly speeding up the optimization process. The effectiveness of this online tensor-RPCA is demonstrated through its application in background-foreground separation in video streams, where the foreground is modeled as a sparse signal and the background as a gradually changing low-rank subspace. Extensive experiments on real-world videos showcase the robustness and effectiveness of the proposed algorithm.
Secondly, the thesis proposes a randomized blocked algorithm for tensor singular-value thresholding (T-SVT), aimed at reducing the computational demands of TRPCA when applied to noisy hyperspectral images. TRPCA has been successfully employed to reduce noise by employing a minimization involving a tensor nuclear norm and a -norm to separate the low-rank hyperspectral image from the sparse noise. However, the high computational complexity of TRPCA is primarily due to the implementation of the T-SVT operator, which typically involves performing full tensor singular value decomposition (T-SVD) followed by shrinking the singular values of the frontal slices in the frequency domain. The proposed randomized blocked algorithm incrementally finds the singular values until they fall below the threshold, leveraging compression achieved by the fast Fourier transform (FFT) to accelerate TRPCA significantly. Numerical experiments indicate that this method is much faster than traditional TRPCA approaches while maintaining classification accuracy.
Finally, the tensor-robust CUR (TRCUR) method is introduced for hyperspectral data compression and denoising. This method heavily downsamples the input hyperspectral image to form small subtensors and performs TRPCA on these subtensors. The desired hyperspectral image is recovered by combining the low-rank solution of the subtensors using tensor CUR reconstruction. We provide theoretical guarantees showing that the desired low-rank tensor can be exactly recovered using our proposed TRCUR method. Numerical experiments demonstrate that our method is up to 14 times faster than performing TRPCA on the original input data, while maintaining the classification accuracy.Ph.D.Electrical and Computer Engineerin
Enabling Scalable, Versatile, and Robust Control for Robotic Exoskeletons
No full textLower-limb exoskeleton technologies—rigid or soft devices that provide assistance to users—show promise in restoring and augmenting human movement. However, current state-of-the-art exoskeleton control primarily addresses consistent, time-repeatable tasks and device-specific, state-machine-based transitions that stand in stark contrast with the fluidity and variability of natural human movement. As I demonstrate in this work, even at its theoretical best, the current control paradigm cannot handle the uncertain and ever-changing environment we live in. In this work, I expand controllers based on deep learning estimates of physiological state to operate in the expansive regime of human activities while also generalizing to novel activities. I show that, when deployed on a hip and knee exoskeleton, these controllers can augment human performance across tasks and time-varying conditions, promising task-agnostic and user-independent control. The process of training these models, however, is device-specific and highly costly in terms of resources and personnel. This threatens to negate its potential for real-world viability. In this work, I also present a novel framework that uses deep domain adaptation to reduce or eliminate the need for costly device-specific data. When deployed on an exoskeleton in real-time, these data-limited models still achieved performance comparable to models with complete access to costly data. These advances are a promising step toward enabling exoskeletons to break the critical task- and device-specific barriers to everyday, outside-laboratory use, and thereby achieve their transformative potential to aid ordinary people.Ph.D.Robotic
Slip in Bimanual Gripping of Deformable Objects with Gelsight Hybrid Adhesion
No full textRobotic object manipulation has increased exponentially over the last couple of decades. Detection and prevention of slip of objects plays a vital role in secure object grasping and manipulation. Through the sensory feedback provided by their skin, humans possess the remarkable ability to readily perceive slip. To attain a level of skill comparable to humans, robots must be equipped with artificial tactile sensing integrated into their system. In this work, object manipulation is studied within the context of Agility Robotics’ humanoid Digit robot. A custom mechanical bimanual gripper is designed to grip deformable objects with optical tactile Gelsight sensors equipped on each finger. The fabrication process is discussed in depth, along with the inverse kinematics model used to control gripper motion. After construction of the gripper, the problem of slip detection is decomposed into a classification problem by using the input from Gelsight sensors. The benefits and limitations of this novel design is discussed with future work on dynamic slip proposed.UndergraduateMechanical Engineerin
Evaluation and Impact of Mixing Phenomena & Injection Strategy on Ducted Fuel Injection Combustion
No full textDiesel engines have long served society as powerplants for both transportation and power generation. While these engines possess high simple-cycle thermal efficiencies due to their high compression ratios and low pumping work, they are plagued by relatively high production of harmful emissions—such as particulate matter and oxides of nitrogen—per unit energy released. This elevated emission production per unit heat release stems from the non-premixed nature of the combustion process.
One promising solution, demonstrated in early experiments to significantly reduce particulate matter formation, is Ducted Fuel Injection (DFI). DFI is believed to function akin to the nozzle of a Bunsen burner, where the fuel draws air into the fuel injector, thereby promoting leaner combustion. To this end, a hollow cylindrical body is positioned in front of the fuel injector within a diesel engine combustion chamber, and the fuel spray passes through it during the injection/combustion event. Existing literature has consistently shown that DFI increases ignition delay and lift-off length while concurrently decreasing particulate matter formation.
This thesis pursued three primary objectives: experimentally investigating DFI's soot mitigation mechanisms, experimentally examining DFI's effect on lift-off length (LOL), and assessing pilot injection impacts on DFI's ignition characteristics and soot reduction.
High-speed optical diagnostics under diesel engine conditions addressed the first two objectives, examining non-reacting mixing and reacting flow interactions across various duct geometries. Note that for the ducts tested and described herein, ``D'' indicates the inner diameter, ``L'' indicates the duct length, and ``G'' indicates the gap distance from the injector tip to the duct. D3L16G2.6, D2L16G2.6, D1L16G2.6, D2L8G2.6, and D1L8G2.6 ducts with a 90 µm injector orifice were used for this study.
Regarding non-reacting mixing, variations in injection pressure minimally influenced mixing fields. Larger duct diameters increased upstream air entrainment, and the influence of duct length depended on diameter. A one-dimensional mass and momentum conservation based jet-pumping model was developed, to better understand the measured results. This model showed that larger duct exit diameters entrain more air due to having lower dynamic pressures at duct exit. Jet-pumping air mass flow rate initially rose with inlet diameter before asymptotically declining due to decreasing vacuum levels at the duct inlet.
Reacting flow results aligned with literature, showing DFI increased LOL and ignition delay (ID), as well as decreased spatially integrated natural luminosity (SINL) compared to free-spray conditions. Three different flame stabilization modes were measured: detached, near-nozzle, and upstream of duct exit. The D1 and D3 configurations consistently stabilized flames in detached and upstream positions, respectively, while the D2 flame stabilization mode varied with injection pressures and chamber temperatures. Detached flames demonstrated nearly linear SINL reduction with increased LOL. When near-nozzle or upstream flames transitioned to detached, SINL dropped significantly.
Non-dimensional analysis was used to evaluate if the yielded flame stabilization mode could be predicted. Higher injection pressures and lower chamber pressures favored flame detachment. A scalar dissipation rate estimation and chemical timescale calculations via constant-pressure reactors was used to calculated an effective Damköhler number (). Plotting duct exit-to-LOL measurements against revealed that Da could effectively predict which flame stabilization mode would occur for a given configuration and condition.
The mixing data was then used to better interpret the reacting SINL measurements. A positive relationship was found between SINL and mass flow weighted cross-sectionally averaged equivalence ratio at LOL, as expected. Flames upstream of the duct exit did not align with the other configuration's results, which is likely due to in-duct combustion elevating the dynamic pressure at duct exit, causing the non-reacting mixing field to be non-indicative of the reacting one.
Addressing the third objective, an experimental and numerical study examined pilot injections' influence on DFI's premixed heat-release spike and soot reduction. Pilot injections shortened ID and decreased peak rate-of-heat-release (ROHR) in both free-spray and DFI, though DFI metrics only reduced to free-spray no-pilot levels due to no combustion recession. Pilot injections decreased peak SINL under DFI, coinciding with increased spray-head penetration rates, likely due to the spray propagating through the pilot injection's residual. Numerical studies corroborated the spray penetration rate finding, and showed that the use of pilot injections lowered the average mixture fraction between LOL and spray tip. This likely contributed to reduced peak SINL by limiting residence time for soot growth.Ph.D.Mechanical Engineerin
Durable Routing in Hyperconnected Logistic Networks
No full textIn the context of Physical Internet-based hyperconnected logistics, regional and gateway hubs promote freight consolidation between regions and urban areas. Several routing protocols, including dynamic directional routing, have been proposed to efficiently identify the best next-hop hubs based on the origin-to-destination direction. While these protocols improve hub-level decision-making, they often overlook the impact of road-level traffic dynamics, particularly in congested metropolitan regions where passenger vehicle interactions can disrupt freight movement. To address this gap, this paper introduces durable routing – a departure-time-dependent approach that leverages historical traffic patterns to predefine reliable, congestion-resilient paths. Unlike purely reactive protocols, durable routing anticipates predictable traffic cycles, ensuring stable transit times even under moderate congestion. This approach supports high-confidence dispatching, enhances route predictability, and promotes balanced modal distribution, reducing overall network congestion while maintaining service-level guarantees. We further present case study results from the state of Georgia, analyzing how factors like day-of-the-week, hour-of-the-day, and risk tolerance thresholds influence durable route selection, demonstrating the scalability and adaptability of this approach to diverse regional contexts
Computational models for bacterial dynamics in community and treatment contexts
No full textMicrobes are key players in human health and disease; however, there is much debate over the nature, consequences, and importance of interactions between bacteria and their environments on the population scale. Interactions in bacterial communities and infection environments are complex and present challenges for modeling, measurement, and inference. However, rising interest in microbiomes (multi-species microbial communities), increasing antimicrobial resistance, and the quest for novel therapeutic strategies to combat human bacterial infection, all center around being able to answer common questions: how do bacteria grow and interact with each other and their environments on the population level? How do they respond to external perturbation from antibiotic exposure or bacteriophage? Using a range of mathematical approaches, we address these questions by integrating forward models and data-driven methods to assess the impacts of underlying mechanisms, abiotic and biotic perturbations, and spatio-temporal heterogeneity as they relate to microbial dynamics in human infections.
Throughout this dissertation, we employ mathematical modeling as a tool to bridge gaps between theoretical and empirical microbiology, highlighting that many standard models and inference methods fail to capture qualitative and quantitative features of microbial dynamics. First, we challenge the received wisdom that antibiotic resistance genes always worsen treatment outcomes and should be strictly minimized. We mathematically explore the effects of ecological interactions on antibiotic treatment in a two lineage system of a pathogen and commensal, proposing an optimization approach to antibiotic resistance management. We define conditions for competitive release and “beneficial” commensal resistance—namely, when commensals inhibit pathogens—and demonstrate generality to resource explicit and spatially extended models. These results are conserved in a four-species experimental community with phage, showing that the addition of phage, targeting the dominant competitor in the community, leads to extinction of the dominant species, competitive release of the next strongest competitor, and maintenance of community diversity. Next, we present an iterative approach for understanding antibiotic and inoculum effects on bacterial growth and yield. Using fine-scale experimental data and a menu of standard population models, we conclude that both growth rate and yield are modified by antibiotic exposure and that populations exhibit distinct regimes of dynamical behavior given distinct exposure conditions. Finally, we expand our modeling into two-dimensional space, building an agent-based simulation of bacterial cells and aggregates to explore physical and socio-microbiology mechanisms underlying relationships between bacterial growth rate and aggregate size.
This work has important implications for both theoretical and empirical studies of microbial systems—evaluating and informing methods for sampling, inference, and modeling to efficiently capture underlying complexities of interactions between bacteria and their environments. In the study of human infection, we provide a baseline toolkit to develop improved treatment strategies for acute and chronic infections and to increase predictability of treatment outcomes.Ph.D.Quantitative Bioscience
Deep Learning for High-Dimensional Decision Making and Uncertainty Quantification
No full textThe advent of deep learning has enabled the solving of problems of increasing complexity and dimension. Such problems include statistical hypothesis testing, learning policies for sequential-decision-making agents, and providing solutions to high-dimensional inverse problems. These problems can be associated with societal and scientific applications, including patrol and dispatch policies for emergency services and data recovery in planetary geophysics. In all such settings, deep neural networks can act as function approximators embedded in some part of a larger system. The immense expressive capacity of neural networks, in conjunction with advanced optimization techniques and computational resources, facilitate more efficient and effective solution of these complex problems.
The research outlined in this dissertation develops novel deep learning methodologies for solving complex problems related to decision-making and uncertainty quantification. First, a model diagnosis procedure is outlined, based on goodness-of-fit assessment using neural network critic functions. Second, a technique for learning joint policies of multiple decision-making agents with shared goals is described, with a particular application to police emergency services dispatch and patrol. Finally, a method for conditional sampling from generative models to solve inverse problems is covered. Inverse problems are highly relevant in the natural sciences; such a problem related to recovering planetary topography data is highlighted in this dissertation. The work outlined in this dissertation approaches these complex problems using deep learning tools, leveraging insight from computational statistics and operations research. This work demonstrates that embedding neural networks into larger frameworks for problem solving proves an effective tool for modern applications in science and society.Ph.D.Machine Learnin
Fixed-Time Reinforcement Learning-based Control for Safe Autonomy
No full textExploiting the benefits of learning can enhance the performance and ensure the safety of autonomous systems in complex and unknown environments. Nonetheless, existing safe learning architectures lack finite time convergence guarantees, rendering these algorithms impractical for real-world applications. In this dissertation, we enable safe autonomy by endowing autonomous systems with safety-critical control frameworks predicated on online reinforcement learning mechanisms with fixed-time convergence guarantees. Specifically, we develop a safe pursuit-evasion game for enabling finite-time capture, optimal performance, and adaptation to an unknown cluttered environment. Then, we leverage ideas from behavioral game theory to construct a learning-based evader assignment algorithm to address the problem of multiple bounded rational pursuers against multiple bounded rational evaders, wherein the assignment is performed based on the agent rationality level. Subsequently, we design an online reinforcement learning architecture with fixed-time convergence guarantees to address the optimal fixed-time stabilization problem. Finally, we address a safety-critical control problem using reachability analysis and design an online reinforcement learning-based mechanism for learning the solution to the safety-critical control problem in a fixed time.Ph.D.Aerospace Engineerin
Adapting Random Tree Search to Fixed Wing Aerial Vehicles with Closed-Loop Prediction and Hybrid Control
No full textPath planning for mobile robots is a developing and evolving area of robotics research to improve their autonomous capabilities. Unmanned Aerial Vehicle (UAV)s present difficult challenges to path planning algorithms that require unique solutions to overcome. This work presents improvements to existing planning algorithms for fixed-wing UAVs. We propose three aims as follows: 1) utilize hybrid control to improve flight planning and performance capabilities, 2) leverage closed-loop prediction of fixed-wing aerial vehicles for path planning decisions, and 3) implement Reinforcement Learning (RL) techniques to learn control policies for fixed-wing aerial vehicles. This thesis will cover the methods to further these aims, including control architecture, planning algorithms, preliminary results, and the final objectives we hope to achieve. The key contributions to highlight in this work are 1) novel control schemes for fixed-wing UAVs, 2) improvements made to closed loop path planning that leverage the forward simulation for prediction, and 3) integrating reinforcement learning for improving the hybrid control performance.Ph.D.Robotic