UMD Theses and Dissertations
Permanent URI for this collectionhttp://hdl.handle.net/1903/3
New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a given thesis/dissertation in DRUM.
More information is available at Theses and Dissertations at University of Maryland Libraries.
Browse
9 results
Search Results
Item Learning-based Physics Simulation with Collision Handling(2023) Tan, Qingyang; Manocha, Dinesh; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Numerous physics-based simulation approaches have been proposed to generate realistic and vivid deformations for 3D models. These systems include the mass-spring system, the finite element approach, the thin-shell model, and others. However, previous systems based on analytic and numerical methods tend to be computationally intensive. Achieving an ideal balance between simulation accuracy and efficiency still poses several challenges. In this dissertation, we present novel learning-based physics simulations and collision-handling algorithms that leverage the benefits of neural networks and optimization techniques. We use neural networks to compress high-dimensional 3D deformable models and accelerate the processing time. We also employ algorithms such as reinforcement learning, active learning, and imitation learning to capture complex physical behaviors that lack closed-form analytic models. We propose multiple novel approaches for novel learning-based physics simulation. First, we train a learning-based collision detector for 3D deformable models and utilize the detector as a surrogate constraint in an optimization-based collision handler. Our focus is on collisions between topologically disjoint triangles in triangular meshes. Traditional geometric-based search methods for collision detection are computationally expensive, with costs ranging from $O(n\log n)$ to $O(n^2)$. In comparison, our neural collision detector is $80\times$ faster. To perform stable collision prediction performance in large and unseen spaces, we employ active learning by progressively incorporating new collision data based on network inferences, reaching a collision detection accuracy of up to $98.1\%$. Second, we present an approach to accelerate collision response computations by incorporating an additional repulsive force unit in the learning-based pipeline. Our experiments demonstrate that backbone networks trained with the repulsive force unit can significantly decrease the number of collisions, boosting collision-free models from $49\%$ to $77\%$, while maintaining real-time performance, adding only $2$ milliseconds to the inference system. Third, we present a neural volumetric deformable object simulator with collision detection and handling based on an actor-critic neural architecture. Our critic network learns to estimate collision penetrations, while our actor network learns to minimize the penalty function through a series of gradient descent steps, resulting in nearly collision-free quasistatic deformable object poses. Finally, we introduce a novel framework for randomly reposing 3D humans to arbitrary poses based on a geometric optimization regularization that incorporates control information into diffusion-based inpainting. Our geometric inpainting algorithm reduces errors by $93\%$ when moving different body parts to random locations. In practice, our learning-based physics simulation systems can generate realistic 3D models that satisfy various constraints. We have tested our method using several large-scale datasets, including AMASS for humans and TailorNet for garments. Our approach can generate plausible results, and we observe $100-300\times$ speedups over numerical or analytic methods.Item NEXT GENERATION HEAT PUMP SYSTEM EVALUATION METHODOLOGIES(2021) Wan, Hanlong; Radermacher, Reinhard K.; Mechanical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Energy consumption of heat pump (HP) systems plays a significant role in the global residential building energy sector. The conventional HP system evaluation method focused on the energy efficiency during a given time scale (e.g., hourly, seasonally, or annually). Nevertheless, these evaluation methods or test metrics are unable to fully reflect the thermodynamic characteristics of the system (e.g., the start-up process). In addition, previous researchers typically conducted HP field tests no longer than one year period. Only limited studies conducted the system performance tests over multiple years. Furthermore, the climate is changing faster than previously predicted beyond the irreversible and catastrophic tipping point. HP systems are the main contributor to global warming due to the increased demands but also can be a part of the solution by replacing fossil fuel burning heating systems. A holistic evaluation of the HP system’s global warming impact during its life cycle needs to account for the direct greenhouse gas (GHG) emissions from the refrigerant leakage, indirect GHG emissions from the power consumption and embodied equipment emissions. This dissertation leverages machine learning, deep learning, data digging, and Life Cycle Climate Performance (LCCP) approaches to develop next generation HP system evaluation methodologies with three thrusts: 1) field test data analysis, 2) data-driven modeling, and 3) enhanced life cycle climate performance (En-LCCP) analysis. This study made following observations: First, time-average performance metrics can save time in extensive data calculation, while quasi-steady-state performance metrics can elucidate some details of the studied system. Second, deep-learning-based algorithms have higher accuracy than conventional modeling approaches and can be used to analyze the system's dynamic performance. However, the complicated structure of the networks, numerous parameters needing optimization, and longer training time are the main challenges for these methods. Third, this dissertation improved current environmental impact evaluation method considering ambient conditions variation, local grid source structure, and next-generation low-GWP refrigerants, which led the LCCP results closer to reality and provided alternative methods for determining LCCP input parameters with limited-data cases. Future work could be studying the uncertainty within the deep learning networks and finding a general process for modeling settings. People may also develop a multi-objective optimization model for HP system design while considering both the LCCP and cost.Item Estimating Biomechanical Risk Factors of Knee Osteoarthritis in Gait Using Instrumented Shoe Insole and Deep Learning Approaches(2021) Snyder, Samantha Jane; Miller, Ross; Shim, Jae Kun; Kinesiology; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)This study aims to implement an alternative to the cost-ineffective and time consuming current inverse dynamics approaches and predict knee adduction moments, a known predictor of knee osteoarthritis, through deep learning neural networks and a custom instrumented insole. Feed-forward, convolutional, and recurrent neural networks are applied to the data extracted from five piezo-resistive force sensors attached to the insole of a shoe. All models predicted knee adduction moment variables during walking with high correlation coefficients, greater than 0.72, and low root mean squared errors, ranging from 0.6-1.2%. The convolutional neural network is the most accurate predictor followed by the recurrent and feed-forward neural networks. These findings and the methods presented in the current study are expected to facilitate a cost-effective clinical analysis of knee adduction moments and to simplify future research studying the relationship between knee adduction moments and knee osteoarthritis.Item ESTIMATION AND CONTROL OF NONLINEAR SYSTEMS: MODEL-BASED AND MODEL-FREE APPROACHES(2020) Goswami, Debdipta; Paley, Derek A.; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)State estimation and subsequent controller design for a general nonlinear system is an important problem that have been studied over the past decades. Many applications, e.g., atmospheric and oceanic sampling or lift control of an airfoil, display strongly nonlinear dynamics with very high dimensionality. Some of these applications use smaller underwater or aerial sensing platforms with insufficient on-board computation power to use a Monte-Carlo approach of particle filters. Hence, they need a computationally efficient filtering method for state-estimation without a severe penalty on the performance. On the other hand, the difficulty of obtaining a reliable model of the underlying system, e.g., a high-dimensional fluid dynamical environment or vehicle flow in a complex traffic network, calls for the design of a data-driven estimation and controller when abundant measurements are present from a variety of sensors. This dissertation places these problems in two broad categories: model-based and model-free estimation and output feedback. In the first part of the dissertation, a semi-parametric method with Gaussian mixture model (GMM) is used to approximate the unknown density of states. Then a Kalman filter and its nonlinear variants are employed to propagate and update each Gaussian mode with a Bayesian update rule. The linear observation model permits a Kalman filter covariance update for each Gaussian mode. The estimation error is shown to be stochastically bounded and this is illustrated numerically. The estimate is used in an observer-based feedback control to stabilize a general closed-loop system. A transferoperator- based approach is then proposed for the motion update for Bayesian filtering of a nonlinear system. A finite-dimensional approximation of the Perron-Frobenius (PF) operator yields a method called constrained Ulam dynamic mode decomposition (CUDMD). This algorithm is applied for output feedback of a pitching airfoil in unsteady flow. For the second part, an echo-state network (ESN) based approach equipped with an ensemble Kalman filter is proposed for data-driven estimation of a nonlinear system from a time series. A random reservoir of recurrent neural connections with the echo-state property (ESP) is trained from a time-series data. It is then used as a model-predictor for an ensemble Kalman filter for sparse estimation. The proposed data-driven estimation method is applied to predict the traffic flow from a set of mobility data of the UMD campus. A data-driven model-identification and controller design is also developed for control-affine nonlinear systems that are ubiquitous in several aerospace applications. We seek to find an approximate linear/bilinear representation of these nonlinear systems from data using the extended dynamic mode decomposition algorithm (EDMD) and apply Liealgebraic methods to analyze the controllability and design a controller. The proposed method utilizes the Koopman canonical transform (KCT) to approximate the dynamics into a bilinear system (Koopman bilinear form) under certain assumptions. The accuracy of this approximation is then analytically justified with the universal approximation property of the Koopman eigenfunctions. The resulting bilinear system is then subjected to controllability analysis using the Myhill semigroup and Lie algebraic structures, and a fixed endpoint optimal controller is designed using the Pontryagin’s principle.Item DYNAMICS OF LARGE SYSTEMS OF NONLINEARLY EVOLVING UNITS(2017) Lu, Zhixin; Ott, Edward; Chemical Physics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)The dynamics of large systems of many nonlinearly evolving units is a general research area that has great importance for many areas in science and technology, including biology, computation by artificial neural networks, statistical mechanics, flocking in animal groups, the dynamics of coupled neurons in the brain, and many others. While universal principles and techniques are largely lacking in this broad area of research, there is still one particular phenomenon that seems to be broadly applicable. In particular, this is the idea of emergence, by which is meant macroscopic behaviors that “emerge” from a large system of many “smaller or simpler entities such that ... large entities” [i.e., macroscopic behaviors] arise which “exhibit properties the smaller/simpler entities do not exhibit.” [1]. In this thesis we investigate mechanisms and manifestations of emergence in four dynamical systems consisting many nonlinearly evolving units. These four systems are as follows. (a) We first study the motion of a large ensemble of many noninteracting particles in a slowly changing Hamiltonian system that undergoes a separatrix crossing. In such systems, we find that separatrix-crossing induces a counterintuitive effect. Specifically, numerical simulation of two sets of densely sprinkled initial conditions on two energy curves appears to suggest that the two energy curves, one originally enclosing the other, seemingly interchange their positions. This, however, is topologically forbidden. We resolve this paradox by introducing a numerical simulation method we call “robust” and study its consequences. (b) We next study the collective dynamics of oscillatory pacemaker neurons in Suprachiasmatic Nucleus (SCN), which, through synchrony, govern the circadian rhythm of mammals. We start from a high-dimensional description of the many coupled oscillatory neuronal units within the SCN. This description is based on a forced Kuramoto model. We then reduce the system dimensionality by using the Ott Antonsen Ansatz and obtain a low-dimensional macroscopic description. Using this reduced macroscopic system, we explain the east-west asymmetry of jet-lag recovery and discus the consequences of our findings. (c) Thirdly, we study neuron firing in integrate-and-fire neural networks. We build a discrete-state/discrete-time model with both excitatory and inhibitory neurons and find a phase transition between avalanching dynamics and ceaseless firing dynamics. Power-law firing avalanche size/duration distributions are observed at critical parameter values. Furthermore, in this critical regime we find the same power law exponents as those observed from experiments and previous, more restricted, simulation studies. We also employ a mean-field method and show that inhibitory neurons in this system promote robustness of the criticality (i.e., an enhanced range of system parameter where power-law avalanche statistics applies). (d) Lastly, we study the dynamics of “reservoir computing networks” (RCN’s), which is a recurrent neural network (RNN) scheme for machine learning. The ad- vantage of RCN’s over traditional RNN’s is that the training is done only on the output layer, usually via a simple least-square method. We show that RCN’s are very effective for inferring unmeasured state variables of dynamical systems whose system state is only partially measured. Using the examples of the Lorenz system and the Rossler system we demonstrate the potential of an RCN to perform as an universal model-free “observer”.Item Real-Time Pose Based Human Detection and Re-Identification with a Single Camera for Robot Person Following(2017) Welsh, John Bradford; Blankenship, Gilmer; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)In this work we address the challenge of following a person with a mobile robot, with a focus on the image processing aspect. We overview different historical approaches for person following and outline the advantages and disadvantages of each. We then show that recent convolutional neural networks trained for human pose detection are suitable for person detection as it relates to the robot following problem. We extend one such pose detection network to spatially embed the identity of individuals in the image, utilizing the pose features already computed. The proposed identity embedding allows the system to robustly track individuals in consecutive frames even in long term occlusion or absence. The final system provides a robust person tracking scheme which is suitable for person following.Item Nonlinear Analysis of Phase Retrieval and Deep Learning(2017) Zou, Dongmian; Balan, Radu V; Applied Mathematics and Scientific Computation; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Nonlinearity causes information loss. The phase retrieval problem, or the phaseless reconstruction problem, seeks to reconstruct a signal from the magnitudes of linear measurements. With a more complicated design, convolutional neural networks use nonlinearity to extract useful features. We can model both problems in a frame-theoretic setting. With the existence of a noise, it is important to study the stability of the phaseless reconstruction and the feature extraction part of the convolutional neural networks. We prove the Lipschitz properties in both cases. In the phaseless reconstruction problem, we show that phase retrievability implies a bi-Lipschitz reconstruction map, which can be extended to the Euclidean space to accommodate noises while remaining to be stable. In the deep learning problem, we set up a general framework for the convolutional neural networks and provide an approach for computing the Lipschitz constants.Item Mutual Information-based RBM Neural Networks(2016) Peng, Kang-Hao; Chellappa, Rama; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)(Deep) neural networks are increasingly being used for various computer vision and pattern recognition tasks due to their strong ability to learn highly discriminative features. However, quantitative analysis of their classication ability and design philosophies are still nebulous. In this work, we use information theory to analyze the concatenated restricted Boltzmann machines (RBMs) and propose a mutual information-based RBM neural networks (MI-RBM). We develop a novel pretraining algorithm to maximize the mutual information between RBMs. Extensive experimental results on various classication tasks show the eectiveness of the proposed approach.Item Adapting Swarm Intelligence For The Self-Assembly And Optimization Of Networks(2011) Martin, Charles E; Reggia, James A; Mathematics; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)While self-assembly is a fairly active area of research in swarm intelligence and robotics, relatively little attention has been paid to the issues surrounding the construction of network structures. Here, methods developed previously for modeling and controlling the collective movements of groups of agents are extended to serve as the basis for self-assembly or "growth" of networks, using neural networks as a concrete application to evaluate this novel approach. One of the central innovations incorporated into the model presented here is having network connections arise as persistent "trails" left behind moving agents, trails that are reminiscent of pheromone deposits made by agents in ant colony optimization models. The resulting network connections are thus essentially a record of agent movements. The model's effectiveness is demonstrated by using it to produce two large networks that support subsequent learning of topographic and feature maps. Improvements produced by the incorporation of collective movements are also examined through computational experiments. These results indicate that methods for directing collective movements can be extended to support and facilitate network self-assembly. Additionally, the traditional self-assembly problem is extended to include the generation of network structures based on optimality criteria, rather than on target structures that are specified a priori. It is demonstrated that endowing the network components involved in the self-assembly process with the ability to engage in collective movements can be an effective means of generating computationally optimal network structures. This is confirmed on a number of challenging test problems from the domains of trajectory generation, time-series forecasting, and control. Further, this extension of the model is used to illuminate an important relationship between particle swarm optimization, which usually occurs in high dimensional abstract spaces, and self-assembly, which is normally grounded in real and simulated 2D and 3D physical spaces.