Accelerating physics-based simulations has been an evergreen topic across different scientific communities. This dissertation is devoted to this subject addressing bottlenecks in state-of-the-art approaches to the simulation of fluids of large-scale scenes, viscous threads, magnetic fluids, and the simulation of fibers and thin structures. The contributions within the thesis are rooted in mathematical modeling and numerical simulation as well as in machine learning. The first part deals with the simulation of incompressible flow in a multigrid fashion. For the variational viscous equation, geometric multigrid is inefficient. An Unsmoothed Aggregation Algebraic Multigrid method is devised with a multi-color Gauss-Seidel smoother, which consistently solves this equation in a few iterations for various material parameters. This framework is 2.0 to 14.6 times faster compared to the state-of-the-art adaptive octree solver in commercial software for the large-scale simulation of both non-viscous and viscous flow. In the second part, a new physical model is devised to accelerate the macroscopic simulation of magnetic fluids. Previous work is based on the classical Smoothed-Particle Hydrodynamics (SPH) method and a Kelvin force model. Unfortunately, this model results in a force pointing outwards causing significant levitation problems limiting the application of more advanced SPH frameworks such as Divergence-Free SPH (DFSPH) or Implicit Incompressible SPH (IISPH). This shortcoming has been addressed with this new current loop magnetic force model resulting in more stable and fast simulations of magnetic fluids using DFSPH and IISPH. Following a different trajectory, the third part of this thesis aims for the acceleration of iterative solvers widely used to accurately simulate physical systems. We speedup the simulation for rod dynamics with Graph Networks by predicting the initial guesses to reduce the number of iterations for the constraint projection part of a Position-based Dynamics solver. Compared to existing methods, this approach guarantees long-term stability and therefore leads to more accurate solutions.

Automated data collection, simulation and visualization can substantially enhance the process of designing, analysis, planning, and control of many engineering processes. In particular, managing processes that are dynamic in nature can significantly benefit from such techniques. Construction projects are good examples of such processes where a variety of equipment and resources constantly interact inside an evolving environment. Management of such settings requires a platform capable of providing decision-makers with updated information about the status of project entities and assisting site personnel making critical decisions under uncertainty. To this end, the current practice of using historical data or expert judgments as static inputs to create empirical formulations, bar chart schedules, and simulation networks to study project activities, resource operations, and the environment under which a project is taking place does not seem to offer reliable results. The presented research investigates the requirements and applicability of a data-driven modeling framework capable of collecting and analyzing real time field data from construction equipment. In the developed data collection scheme, a stream of real time data is continuously transferred to a data analysis module to calculate the input parameters required to create dynamic 3D visualizations of ongoing engineering activities, and update the contents of a discrete event simulation (DES) model representing the real engineering process. The generated data-driven simulation model is iv an effective tool for projecting future progress based on existing performance. Ultimately, the developed framework can be used by project decision-makers for shortterm project planning and control since the resulting simulation and visualization are completely based on the latest status of project entities.

The main objective of this research is to explore, investigate and develop the new data analysis techniques along with novel sensing technologies for structural health monitoring applications. The study has three main parts. First, a systematic comparative evaluation of some of the most common and promising methods is carried out along with a combined method proposed in this study for mitigating drawbacks of some of the techniques. Secondly, nonparametric methods are evaluated on a real life movable bridge. Finally, a hybrid approach for non-parametric and parametric method is proposed and demonstrated for more in depth understanding of the structural performance. In view of that, it is shown in the literature that four efficient non-parametric algorithms including, Cross Correlation Analysis (CCA), Robust Regression Analysis (RRA), Moving Cross Correlation Analysis (MCCA) and Moving Principal Component Analysis (MPCA) have shown promise with respect to the conducted numerical studies. As a result, these methods are selected for further systematic and comparative evaluation using experimental data. A comprehensive experimental test is designed utilizing Fiber Bragg Grating (FBG) sensors simulating some of the most critical and common damage scenarios on a unique experimental structure in the laboratory. Subsequently the SHM data, that is generated and collected under different damage scenarios, are employed for comparative study of the selected techniques based on critical criteria such as detectability, time to detection, effect of noise, computational time and size of the window. The observations indicate that while MPCA has the best detectability, it does not perform very reliable results in terms of time to detection. As a result, a machine-learning based algorithm is explored that not only reduces the associated delay with MPCA but further iii improves the detectability performance. Accordingly, the MPCA and MCCA are combined to introduce an improved algorithm named MPCA-CCA. The new algorithm is evaluated through both experimental and real-life studies. It is realized that while the methods identified above have failed to detect the simulated damage on a movable bridge, the MPCA-CCA algorithm successfully identified the induced damage. An investigative study for automated data processing method is developed using nonparametric data analysis methods for real-time condition maintenance monitoring of critical mechanical components of a movable bridge. A maintenance condition index is defined for identifying and tracking the critical maintenance issues. The efficiency of the maintenance condition index is then investigated and demonstrated against some of the corresponding maintenance problems that have been visually and independently identified for the bridge. Finally, a hybrid data interpretation framework is designed taking advantage of the benefits of both parametric and non-parametric approaches and mitigating their shortcomings. The proposed approach can then be employed not only to detect the damage but also to assess the identified abnormal behavior. This approach is also employed for optimized sensor number and locations on the structure.

The algorithms for the mathematical modeling to predict productivity of underground room-and-pillar mining systems are well-known and documented. These algorithms consider the time-varying relationships between mining equipment for a given geometry of operations as well as other constraints. This study presents a newly developed, user-friendly visual simulation computer tool for the Windows environment. This tool can be easily customized and utilized by field engineers and will help mine operators plan the optimum mining sequence for different mine geometries and equipment layouts. Program output includes monitoring of shift data, equipment utilization indices, and so forth. The simulation technique can be used with any environment. Use of the system is demonstrated in different mining equipment configurations. Development and validation of the system was aided by the Peabody Group. / Master of Science

Soil moisture is an important variable in many environmental researches and application areas as it affects the interactions between atmosphere and land surface by controlling the energy and water exchange. The current measurement techniques are insufficient to acquire accurate large-scale root zone soil moisture (RZSM) data at the spatial resolution of interest. Though assorted models have been successfully applied in relatively small areas to estimate RZSM, the large-scale estimation is still facing challenges as it requires the flexibility and practicality of the models for the applications under various conditions. Though physically based soil moisture models are widely used, the errors in model physics affect the flexibility of these models meanwhile their large demand of data and computational resources reduces the practicality. On the contrary, the statistical and data-driven methods have high potential but their applications for large-scale RZSM estimation have not been fully explored. To develop feasible models for large-scale RZSM estimation using the surface observations, artificial neural networks, specifically multilayer perceptrons (MLPs), were applied in this study to estimate RZSM at the depths of 20cm and 50cm, using the data of 557 stations in the United States. Two experiments including four models were developed and the input variables of the models were carefully selected. The sensitivity analysis found that surface soil moisture and the cumulative rainfall, snowfall, air temperature and surface soil temperature were important inputs. If given soil texture data as inputs, the models achieved better performance and were extremely sensitive to them. The results showed that the MLPs were effective and flexible for the estimation of soil moisture at 20cm under various climate types and were insensitive to the potential errors in soil moisture datasets. However, the results of the estimation at 50cm are not as good as that of the 20cm. / Thesis / Master of Science (MSc)

A formulation of a subspace identification method for multi-phase processes with applications to rotational moulding and suggestions for improvements and experimental applications. / This thesis focuses on the implementation of subspace identification (SSID) for nonlinear, chemical batch processes by introducing a model identification method for multi-phase processes. In this thesis, a multi-phase process refers to chemical or biological batch-like processes with properties that cause a change in the dynamics during the evolution of the process. This can occur, for example, when a process undergoes a change of state upon reaching a melting point. Existing SSID techniques are not designed to utilize any known, multiphase nature of a process in the model identification stage. The proposed approach, Multiphase Subspace Identification (MPSSID), is conducted by first splitting historical data into phases during the identification step and then building a subspace model for each phase. The phases are then connected via a partial least squares (PLS) model that transforms the states from one phase to the next. This approach makes use of existing SSID techniques that allow for model construction using batches of nonunifrom length. Here, MPSSID is applied to a uniaxial rotational moulding process. In rotational moulding, the dynamics switch as the process undergoes heating, melting, and sintering stages that are visibly distinct and recognizable upon a certain temperature (not time) being reached. Results demonstrate the ability of multiphase models to better predict the temperature trajectories and final product quality of validation batches. As an extension to this rotational moulding analysis, additional MPSSID methods of implementation are proposed and the results are compared. A MPSSID mixed integer linear program is then introduced for implementation within model predictive control. The applications to rotational moulding are presented within the context of plastics manufacturing and the impact of plastic on the global climate crisis, with suggestions for future work. / Thesis / Master of Applied Science (MASc) / The control of chemical processes is an important factor in achieving high quality products. To control a process well, the mathematical model of the system must be accurate. In the past, mathematical models for process control were designed based on engineering approximations. Now, with major advances in computing and sensor technology, it is possible to design a simulation of the entire process. These simulations can be designed using first-principles or black box approaches. First-principles approaches utilize rigorous models that are based on the complex chemical and physical formulas that govern a system. Black box approaches do not look at the first-principles dynamics. They only utilize the measured process inputs and outputs to form a model of the system. They are widely used because of their ease of implementation in comparison to first-principles approaches. In this thesis, a new black box process control model is proposed and is found to yield better theoretical results than existing techniques. This model is tested on data from a plastics manufacturing process called rotational moulding, which involves loading polymer powders into a mould that is simultaneously rotated and heated to yield seamless plastic parts. Lastly, a control framework that is compatible with the new black box model is proposed to be used for future experimental tests.

京都大学 / 新制・課程博士 / 博士(情報学) / 甲第24934号 / 情博第845号 / 新制||情||141(附属図書館) / 京都大学大学院情報学研究科社会情報学専攻 / (主査)教授 緒方 広明, 教授 伊藤 孝行, 教授 田島 敬史 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Data-driven model predictive control framework of semicontinuous distillation process / Distillation technology is one of the most sought-after operations in the chemical process industries. Countless research has been done in the past to reduce the cost associated with distillation technology. As a result of process intensification, a semicontinuous distillation system is proposed as an alternative for purifying the n-component mixture (n>=3) which has the advantage over both batch and continuous process for low to medium production rates. A traditional distillation setup requires n-1 columns to separate the components to the desired purity. However, a semicontinuous system performs the same task by integrating a distillation column with n-2 middle vessel (storage tank). Consequently, with lower capital cost, the total annualized cost (TAC) per tonne of feed processed is less for a semicontinuous system compared to a traditional setup for low to medium throughput. Yet, the operating cost of a semicontinuous system exceed those of the conventional continuous setup. Semicontinuous system exhibits a non-linear dynamic behavior with a cyclic steady state and has three modes of operation. The main goal of this thesis is to reduce the operating cost per tonne of feed processed which leads to lower TAC per tonne of feed processed using a model predictive control (MPC) scheme compared to the existing PI configuration This work proposes a novel multi-model technique using subspace identification to identify a linear model for each mode of operation without attaining discontinuity. Subsequently, the developed multi-model framework was implemented in a shrinking horizon MPC architecture to reduce the TAC/tonne of feed processed while maintaining the desired product purities at the end of each cycle. The work uses Aspen Plus Dynamics simulation as a test bed to simulate the semicontinuous system and the shrinking horizon MPC scheme is formulated in MATLAB. VBA is used to communicate the inputs from MPC in MATLAB to the process in Aspen Plus Dynamics. / Thesis / Master of Science in Chemical Engineering (MSChE)

The aim of this dissertation is to systematically construct a hierarchical framework that allows for robust multi-agent collaborative legged locomotion. More specifically, this work provides a detailed derivation of a torque controller that is theoretically justifiable in the context of Hybrid Zero Dynamics at the lowest level of control to produce highly robust locomotion, even when subject to uncertainty. The torque controller is based on virtual constraints and partial feedback linearization and is cast into the form of a strictly convex quadratic program. This partial feedback linearization is then relaxed through the use of a defect variable, where said defect variable is allowed only to change in a manner that is consistent with rapidly exponentially stable output dynamics through the use of a Control Lyapunov Function. The torque controller is validated in both simulation and on hardware to demonstrate the efficacy of the approach. In particular, the robot is subject to payload and push disturbances and is still able to remain stable. Furthermore, the continuity of the torque controller, in addition to robustness analysis of the periodic orbit, is also provided. At the next level of control, we consider emulating the Single Rigid Body model through the use of Behavioral Systems Theory, resulting in a data-driven model that adequately describes a quadruped at the reduced-order level. Still, due to the complexity and a considerable number of variables in the problem, the model further undergoes a $2$-norm approximation, resulting in a model that is computationally efficient enough to be used in a real-time manner for trajectory planning. In order to test the method rigorously, we consider a series of experiments to examine how the planner works when using different gait parameters than that which was used during data collection. Furthermore, the planner is compared to the traditional Single Rigid Body model to test its efficacy for reference tracking. This data-driven model is then extended to the multi-agent case, where each agent is rigidly holonomically constrained to one another. In this case, the model is used in a distributed manner using a one-step communication delay such that the coupling between agents can be adequately considered while spreading the computational demand. The trajectory planner is evaluated through various hardware experiments with three agents, and simulations are also used to display the scalability of the approach by considering five robots. Finally, this dissertation examines how traditional reduced-order models can be used in tandem with data-based models to reap the benefits of both methods. More specifically, an interconnected Single Rigid Body model is considered, where the interaction forces are described via a data-driven model. Simulations are provided to display the efficacy of this approach at the reduced order level and show that the interaction forces can be reduced by considering them in the trajectory planner. As in the previous cases, this is followed by experimental evaluation subject to external forces and different terrains. / Doctor of Philosophy / The goal of this dissertation is to create a layered control scheme for teams of quadrupeds that results in stable and robust locomotion, including a high-level trajectory planner and a low-level controller. More specifically, this work outlines an optimal torque-based whole-body controller that operates at the joint level to track desired trajectories. These trajectories are obtained by a high-level trajectory planner, which utilizes a data-driven predictive controller to create an optimal trajectory without explicitly requiring knowledge of a model. The hierarchical control scheme is then extended to consider collaborative locomotion. Namely, this work considers teams of quadrupeds that are rigidly connected to one another such that there is no relative motion between them. There are potentially large interaction forces that are applied between the robots that cannot be measured, which can result in instability. Furthermore, the models used to describe the interconnected system are prohibitively complex when being used for trajectory planning. For this reason, the data-driven model considered for a single robot is extended to create a centralized model that encapsulates not only the motion of a single robot but also its connection constraints. The resulting model is very large, making it difficult to use in a real-time manner. Therefore, this work outlines how to distribute the model such that each robot can locally plan for its own motion while also considering the coupling between them. Finally, this work provides one additional extension that combines a traditional physics-based model with a data-driven model to capitalize on the strengths of each. In particular, a physics-based model is considered as a baseline, while a data-driven model is used to describe the interaction forces between robots. In using this final extension, both improved solve times and smoother locomotion are achieved. Each of the aforementioned methods is tested thoroughly through both simulations and experiments.

With the recent advances in the field of unmanned aerial vehicles, many applications have been identified. In tasks that require high-payload-to-weight ratios, flight times in the order of days, reduced noise and/or hovering capabilities, lighter-than-air vehicles present themselves as a competitive platform compared to fixed-wing and rotor based vehicles. The limiting factor in their widespread use in autonomous applications comes from the complexity of the control task. The so-called airships are highly-susceptible to aerodynamic forces and pose complex nonlinear system dynamics that complicate their modeling and control. Model-free control lends itself well as a solution to this type of problem, as it derives its control policies using input-output data, and can therefore learn complex dynamics and handle uncertain or unknown parameters and disturbances. In this work, two multi-input multi-output algorithms are presented on the basis of optimal control theory. Leveraging results from reinforcement learning, a single layer, partially connected neural network is formulated as a value function appropriator in accordance with Weierstrass higher-order approximation theorem. The so-called critic-network is updated using gradient descent methods on the mean-squared error of the temporal difference equation. In the single-network controller, the control policy is formulated as a closed form equation that is parameterized on the weights of the critic-network. A second controller is proposed that uses a second single-layer partially connected neural network, the actor-network, to calculate the control action. The actor-network is also updated using gradient descent on the squared error of the temporal difference equation. The controllers are employed in a highly realistic simulation airship model in nominal conditions and in the presence of external disturbances in the form of turbulent wind. To verify the validity and test the sensitivity of the algorithms to design parameters (the initialization of certain terms), ablation studies are carried out with multiple initial parameters. Both of the proposed algorithms are able to track the desired waypoints in both the nominal and disturbed flight tests. Furthermore, the performance of the controllers is compared to a modern, state-of-the-art multi-input multi-output controller. The two proposed controllers outperform the comparison controller in all but one flight test, with up to four fold reduction in the integral absolute error and integral time absolute error metrics. On top of the quantitative improvements seen in the proposed controllers, both controllers demonstrate a reduction in system oscillation and actuator chattering with respect to the comparison algorithm.