The research presented in this Honors Thesis provides development in machine learning models which predict future states of a system with unknown dynamics, based on observations of the system. Two case studies are presented for (1) a non-conservative pendulum and (2) a differential game dictating a two-car uncontrolled intersection scenario. In the paper we investigate how learning architectures can be manipulated for problem specific geometry. The result of this research provides that these problem specific models are valuable for accurate learning and predicting the dynamics of physics systems.<br/><br/>In order to properly model the physics of a real pendulum, modifications were made to a prior architecture which was sufficient in modeling an ideal pendulum. The necessary modifications to the previous network [13] were problem specific and not transferrable to all other non-conservative physics scenarios. The modified architecture successfully models real pendulum dynamics. This case study provides a basis for future research in augmenting the symplectic gradient of a Hamiltonian energy function to provide a generalized, non-conservative physics model.<br/><br/>A problem specific architecture was also utilized to create an accurate model for the two-car intersection case. The Costate Network proved to be an improvement from the previously used Value Network [17]. Note that this comparison is applied lightly due to slight implementation differences. The development of the Costate Network provides a basis for using characteristics to decompose functions and create a simplified learning problem.<br/><br/>This paper is successful in creating new opportunities to develop physics models, in which the sample cases should be used as a guide for modeling other real and pseudo physics. Although the focused models in this paper are not generalizable, it is important to note that these cases provide direction for future research.

High-entropy alloys possessing mechanical, chemical, and electrical properties that far exceed those of conventional alloys have the potential to make a significant impact on many areas of engineering. Identifying element combinations and configurations to form these alloys, however, is a difficult, time-consuming, computationally intensive task. Machine learning has revolutionized many different fields due to its ability to generalize well to different problems and produce computationally efficient, accurate predictions regarding the system of interest. In this thesis, we demonstrate the effectiveness of machine learning models applied to toy cases representative of simplified physics that are relevant to high-entropy alloy simulation. We show these models are effective at learning nonlinear dynamics for single and multi-particle cases and that more work is needed to accurately represent complex cases in which the system dynamics are chaotic. This thesis serves as a demonstration of the potential benefits of machine learning applied to high-entropy alloy simulations to generate fast, accurate predictions of nonlinear dynamics.

Bicycle stabilization has become a popular topic because of its complex dynamic behavior and the large body of bicycle modeling research. Riding a bicycle requires accurately performing several tasks, such as balancing and navigation which may be difficult for disabled people. Their problems could be partially reduced by providing steering assistance. For stabilization of these highly maneuverable and efficient machines, many control techniques have been applied – achieving interesting results, but with some limitations which includes strict environmental requirements. This thesis expands on the work of Randlov and Alstrom, using reinforcement learning for bicycle self-stabilization with robotic steering. This thesis applies the deep deterministic policy gradient algorithm, which can handle continuous action spaces which is not possible for Q-learning technique. The research involved algorithm training on virtual environments followed by simulations to assess its results. Furthermore, hardware testing was also conducted on Arizona State University’s RISE lab Smart bicycle platform for testing its self-balancing performance. Detailed analysis of the bicycle trial runs are presented. Validation of testing was done by plotting the real-time states and actions collected during the outdoor testing which included the roll angle of bicycle. Further improvements in regard to model training and hardware testing are also presented.

Daily collaborative tasks like pushing a table or a couch require haptic communication between the people doing the task. To design collaborative motion planning algorithms for such applications, it is important to understand human behavior. Collaborative tasks involve continuous adaptations and intent recognition between the people involved in the task. This thesis explores the coordination between the human-partners through a virtual setup involving continuous visual feedback. The interaction and coordination are modeled as a two-step process: 1) Collecting data for a collaborative couch-pushing task, where both the people doing the task have complete information about the goal but are unaware of each other's cost functions or intentions and 2) processing the emergent behavior from complete information and fitting a model for this behavior to validate a mathematical model of agent-behavior in multi-agent collaborative tasks. The baseline model is updated using different approaches to resemble the trajectories generated by these models to human trajectories. All these models are compared to each other. The action profiles of both the agents and the position and velocity of the manipulated object during a goal-oriented task is recorded and used as expert-demonstrations to fit models resembling human behaviors. Analysis through hypothesis teasing is also performed to identify the difference in behaviors when there are complete information and information asymmetry among agents regarding the goal position.

Almost all mechanical and electro-mechanical products are assemblies of multiple parts, either because of requirements for relative motion, or use of different materials, shape/size differences. Thus, assembly design is the very crux of engineering design. In addition to nominal design of an assembly, there is also tolerance design to determine allowable manufacturing variations to ensure proper functioning and assemblability. Most of the flexible assemblies are made by stamping sheet metal. Sheet metal stamping process involves plastically deforming sheet metals using dies. Sub-assemblies of two or more components are made with either spot-welding or riveting operations. Various sub-assemblies are finally joined, using spot-welds or rivets, to create the desired assembly. When two components are brought together for assembly, they do not align exactly; this causes gaps and irregularities in assemblies. As multiple parts are stacked, errors accumulate further. Stamping leads to variable deformations due to residual stresses and elastic recovery from plastic strain of metals; this is called as the ‘spring-back’ effect. When multiple components are stacked or assembled using spot welds, input parameters variations, such as sheet metal thickness, number and order of spot welds, cause variations in the exact shape of the final assembly in its free state. It is essential to understand the influence of these input parameters on the geometric variations of both the individual components and the assembly created using these components. Design of Experiment is used to generate principal effect study which evaluates the influence of input parameters on output parameters. The scope of this study is to quantify the geometric variations for a flexible assembly and evaluate their dependence on specific input variables. The 3 input variables considered are the thickness of the sheet material, the number of spot welds used and the spot-welding order to create the assembly. To quantify the geometric variations, sprung-back nodal points along lines, circular arcs, a combination of these, and a specific profile are reduced to metrologically simulated features.

Information exists in various forms and a better utilization of the available information can benefit the system awareness and response predictions. The focus of this dissertation is on the fusion of different types of information using Bayesian-Entropy method. The Maximum Entropy method in information theory introduces a unique way of handling information in the form of constraints. The Bayesian-Entropy (BE) principle is proposed to integrate the Bayes’ theorem and Maximum Entropy method to encode extra information. The posterior distribution in Bayesian-Entropy method has a Bayesian part to handle point observation data, and an Entropy part that encodes constraints, such as statistical moment information, range information and general function between variables. The proposed method is then extended to its network format as Bayesian Entropy Network (BEN), which serves as a generalized information fusion tool for diagnostics, prognostics, and surrogate modeling.

The proposed BEN is demonstrated and validated with extensive engineering applications. The BEN method is first demonstrated for diagnostics of gas pipelines and metal/composite plates for damage diagnostics. Both empirical knowledge and physics model are integrated with direct observations to improve the accuracy for diagnostics and to reduce the training samples. Next, the BEN is demonstrated in prognostics and safety assessment in air traffic management system. Various information types, such as human concepts, variable correlation functions, physical constraints, and tendency data, are fused in BEN to enhance the safety assessment and risk prediction in the National Airspace System (NAS). Following this, the BE principle is applied in surrogate modeling. Multiple algorithms are proposed based on different type of information encoding, such as Bayesian-Entropy Linear Regression (BELR), Bayesian-Entropy Semiparametric Gaussian Process (BESGP), and Bayesian-Entropy Gaussian Process (BEGP) are demonstrated with numerical toy problems and practical engineering analysis. The results show that the major benefits are the superior prediction/extrapolation performance and significant reduction of training samples by using additional physics/knowledge as constraints. The proposed BEN offers a systematic and rigorous way to incorporate various information sources. Several major conclusions are drawn based on the proposed study.

Visual navigation is a multi-disciplinary field across computer vision, machine learning and robotics. It is of great significance in both research and industrial applications. An intelligent agent with visual navigation ability will be capable of performing the following tasks: actively explore in environments, distinguish and localize a requested target and approach the target following acquired strategies. Despite a variety of advances in mobile robotics, enabling an autonomous with above-mentioned abilities is still a challenging and complex task. However, the solution to the task is very likely to accelerate the landing of assistive robots.

Reinforcement learning is a method that trains autonomous robot based on rewarding desired behaviors to help it obtain an action policy that maximizes rewards while the robot interacting with the environment. Through trial and error, an agent learns sophisticated and skillful strategies to handle complex tasks in the environment. Inspired by navigation procedures of human beings that when navigating through environments, humans reason about accessible spaces and geometry of the environment a lot based on first-person view, figure out the destination and then ease over, this work develops a model that maps from pixels to actions and inherently estimate the target as well as the free-space map. The model has three major constituents: (i) a cognitive mapper that maps the topologic free-space map from first-person view images, (ii) a target recognition network that locates a desired object and (iii) an action policy deep reinforcement learning network. Further, a planner model with cascade architecture based on multi-scale semantic top-down occupancy map input is proposed.

Machine learning has demonstrated great potential across a wide range of applications such as computer vision, robotics, speech recognition, drug discovery, material science, and physics simulation. Despite its current success, however, there are still two major challenges for machine learning algorithms: limited robustness and generalizability.

The robustness of a neural network is defined as the stability of the network output under small input perturbations. It has been shown that neural networks are very sensitive to input perturbations, and the prediction from convolutional neural networks can be totally different for input images that are visually indistinguishable to human eyes. Based on such property, hackers can reversely engineer the input to trick machine learning systems in targeted ways. These adversarial attacks have shown to be surprisingly effective, which has raised serious concerns over safety-critical applications like autonomous driving. In the meantime, many established defense mechanisms have shown to be vulnerable under more advanced attacks proposed later, and how to improve the robustness of neural networks is still an open question.

The generalizability of neural networks refers to the ability of networks to perform well on unseen data rather than just the data that they were trained on. Neural networks often fail to carry out reliable generalizations when the testing data is of different distribution compared with the training one, which will make autonomous driving systems risky under new environment. The generalizability of neural networks can also be limited whenever there is a scarcity of training data, while it can be expensive to acquire large datasets either experimentally or numerically for engineering applications, such as material and chemical design.

In this dissertation, we are thus motivated to improve the robustness and generalizability of neural networks. Firstly, unlike traditional bottom-up classifiers, we use a pre-trained generative model to perform top-down reasoning and infer the label information. The proposed generative classifier has shown to be promising in handling input distribution shifts. Secondly, we focus on improving the network robustness and propose an extension to adversarial training by considering the transformation invariance. Proposed method improves the robustness over state-of-the-art methods by 2.5% on MNIST and 3.7% on CIFAR-10. Thirdly, we focus on designing networks that generalize well at predicting physics response. Our physics prior knowledge is used to guide the designing of the network architecture, which enables efficient learning and inference. Proposed network is able to generalize well even when it is trained with a single image pair.

Aging-related damage and failure in structures, such as fatigue cracking, corrosion, and delamination, are critical for structural integrity. Most engineering structures have embedded defects such as voids, cracks, inclusions from manufacturing. The properties and locations of embedded defects are generally unknown and hard to detect in complex engineering structures. Therefore, early detection of damage is beneficial for prognosis and risk management of aging infrastructure system.

Non-destructive testing (NDT) and structural health monitoring (SHM) are widely used for this purpose. Different types of NDT techniques have been proposed for the damage detection, such as optical image, ultrasound wave, thermography, eddy current, and microwave. The focus in this study is on the wave-based detection method, which is grouped into two major categories: feature-based damage detection and model-assisted damage detection. Both damage detection approaches have their own pros and cons. Feature-based damage detection is usually very fast and doesn’t involve in the solution of the physical model. The key idea is the dimension reduction of signals to achieve efficient damage detection. The disadvantage is that the loss of information due to the feature extraction can induce significant uncertainties and reduces the resolution. The resolution of the feature-based approach highly depends on the sensing path density. Model-assisted damage detection is on the opposite side. Model-assisted damage detection has the ability for high resolution imaging with limited number of sensing paths since the entire signal histories are used for damage identification. Model-based methods are time-consuming due to the requirement for the inverse wave propagation solution, which is especially true for the large 3D structures.

The motivation of the proposed method is to develop efficient and accurate model-based damage imaging technique with limited data. The special focus is on the efficiency of the damage imaging algorithm as it is the major bottleneck of the model-assisted approach. The computational efficiency is achieved by two complimentary components. First, a fast forward wave propagation solver is developed, which is verified with the classical Finite Element(FEM) solution and the speed is 10-20 times faster. Next, efficient inverse wave propagation algorithms is proposed. Classical gradient-based optimization algorithms usually require finite difference method for gradient calculation, which is prohibitively expensive for large degree of freedoms. An adjoint method-based optimization algorithms is proposed, which avoids the repetitive finite difference calculations for every imaging variables. Thus, superior computational efficiency can be achieved by combining these two methods together for the damage imaging. A coupled Piezoelectric (PZT) damage imaging model is proposed to include the interaction between PZT and host structure. Following the formulation of the framework, experimental validation is performed on isotropic and anisotropic material with defects such as cracks, delamination, and voids. The results show that the proposed method can detect and reconstruct multiple damage simultaneously and efficiently, which is promising to be applied to complex large-scale engineering structures.

A defense-by-randomization framework is proposed as an effective defense mechanism against different types of adversarial attacks on neural networks. Experiments were conducted by selecting a combination of differently constructed image classification neural networks to observe which combinations applied to this framework were most effective in maximizing classification accuracy. Furthermore, the reasons why particular combinations were more effective than others is explored.