<p> This thesis contributes to the development of adjoint variable methods (AVM) and space mapping (SM) technology for computer-aided electromagnetics (EM)-based modeling and design of microwave circuits and antennas.</p> <p> The AVM is known as an efficient approach to design sensitivity analysis for problems of high complexity. We propose a general self-adjoint approach to the sensitivity analysis of network parameters for an Method of Moments (MoM) solver. It requires neither an adjoint problem nor analytical system matrix derivatives. For the first time, we suggest practical and fast sensitivity solutions realized entirely outside the EM solver, which simplifies the implementation. We discuss: (1) features of commercial EM solvers which allow the user to compute network parameters and their sensitivities through a single full-wave simulation; (2) the accuracy of the computed derivatives; (3) the overhead of the sensitivity computation. Our approach is demonstrated by FEKO, which employs an MoM solver.</p> <p> One motivation for sensitivity analysis is gradient-based optimization. The sensitivity evaluation providing the Jacobian is a bottleneck of optimization with full-wave simulators. We propose an approach, which employs the self-adjoint sensitivity analysis of network parameters and Broyden's update for practical EM design optimization. The Broyden's update is carried out at the system matrix level, so that the computational overhead of the Jacobian is negligible while the accuracy is acceptable for optimization. To improve the robustness of the Broyden update in the sensitivity analysis, we propose a switching criterion between the Broyden and the finite-difference estimation of the system matrix derivatives.</p> <p> In the second part, we apply for the first time a space mapping technique to antenna design. We exploit a coarse mesh MoM solver as the coarse model and align it with the fine mesh MoM solution through space mapping. Two SM plans are employed: I. implicit SM and output SM, and II. input SM and output SM. A novel local meshing method is proposed to avoid inconsistencies in the coarse model. The proposed techniques are implemented through the new user-friendly SMF system. In a double annular ring antenna example, the S-parameter is optimized. The finite ground size effect for the MoM is efficiently solved by SM Plan I and the design specification is satisfied after only three iterations. In a patch antenna example, we optimize the impedance through both plans. Comparisons are made. Coarseness in the coarse model and its effect on the SM performance is also discussed.</p> / Thesis / Master of Applied Science (MASc)

This thesis investigates the application of latent variable methods to three combustion processes. Multivariate analysis of flame images and process data is performed to predict important quality parameters and monitor flame stability. The motivation behind this work is to decrease operational costs and greenhouse gases in these energy intensive processes. The three combustion processes studied are a lime kiln, a basic oxygen furnace and a coal-fired boiler. In lime kiln operation, the main goal is to stabilize final product temperature in order to reduce fouling and energy costs. Due to long process dynamics, prediction of product temperature is required at least one hour in advance for potential use in a control scheme. Several methods for extracting features from flame images were investigated for the prediction of the temperature. The best method is then combined with process data in a PLS model that also incorporates dynamic information. The analysis revealed that prediction one hour into the future is successful using latent variable methods. In the basic oxygen furnace analysis, the main goal is to predict end-point carbon of the batch process. Termination of the batch as soon as the desired carbon is attained reduces oxygen consumption and thus operational cost. Traditional image analysis is used to identify a constant field of view in the flame images. Multivariate image feature extraction methods were then used in combination with process data to successfully predict the final carbon content of the heat. The coal-fired boiler analysis focuses on monitoring of flame stability at different production and air to fuel levels of the boiler. Prediction of energy efficiency and off-gas chemistry from flame images is also investigated. An unexpected result was the ability to use the installed cameras for localized fouling monitoring. This thesis showed that the use of multivariate analysis of flame images and process data in combustion process is very promising. / Thesis / Master of Applied Science (MASc)

High-accuracy electromagnetic design and analysis of electric machines is enhanced by the use of various numerical methods. These methods solve Maxwell’s equations to determine the distribution of the electric and magnetic fields throughout the considered machine structure. Due to the complicated architectures of the machines and the nonlinearity of the utilized magnetic materials, it is a very challenging task to obtain an analytical solution and, in most cases, only a numerical solution is possible. The finite element method (FEM) is one of the standard numerical methods for electromagnetic field analysis. The considered machine domain is divided into finite elements to which the field equations are applied. FEM solvers are utilized to develop optimization procedures to assist in achieving a design that meets the required specifications without violating the design constraints. The design process of electric machines involves adjusting the machine parameters. This is usually done through experience, intuition, and heuristic approaches using FEM software which gives results for various parameter changes. There is no guarantee that the achieved design is the optimal one. An alternative approach to the design of electric machines exploits robust gradient-based optimization algorithms that are guaranteed to converge to a locally-optimal model. The gradient-based approaches utilize the sensitivities of the performance characteristics with respect to the design parameters. These sensitivities are classically calculated using finite difference approximations which require repeated simulations with perturbed parameter values. The cost of evaluating these sensitivities can be significant for a slow finite element simulation or when the number of parameters is large. The adjoint variable method (AVM) offers an alternative approach for efficiently estimating response sensitivities. Using at most one extra not-iterative simulation, the sensitivities of the response to all parameters are estimated. Here, a MATLAB tool has been developed to automate the design process of switched reluctance motors (SRMs). The tool extracts the mesh data of an initial motor model from a commercial FEM software, JMAG. It then solves for magnetic vector potential throughout the considered SRM domain using FEM taking into consideration the nonlinearity of the magnetic material and the motor dynamic performance. The tool calculates various electromagnetic quantities such as electromagnetic torque, torque ripple, phase flux linkage, x and y components of flux density, air-region stored magnetic energy, phase voltage, and phase dynamic currents. The tool uses a structural mapping technique to parametrize various design parameters of SRMs. These parameters are back iron thickness, teeth height, pole arc angle, and pole taper angle of both stator and rotor. Moreover, it calculates the sensitivities of various electromagnetic quantities with respect to all these geometric design parameters in addition to the number of turn per phase using the AVM method. The tool applies interior point optimization algorithm to simultaneously optimize the motor geometry, number of turns per phase, and the drive-circuit control parameters (reference current, and turn-on and turn-off angles) to increase the motor average dynamic torque. It also applies the ON/OFF topology optimization algorithm to optimize the geometries of the stator teeth for proper distribution of the magnetic material to reduce the RMS torque ripple. A 6/14 SRM has been automatically designed using the developed MATLAB tool to achieve the same performance specifications as 6110E Evergreen surface-mounted PM brushless DC motor which is commercially available for an HVAC system. / Thesis / Doctor of Philosophy (PhD)

Hundreds of steel products need to be scheduled on a single or parallel machine in a steel plant every week. A good feasible schedule may save the company millions of dollars compared to a bad one. Single and parallel machine scheduling are also encountered often in many other industries, making it a crucial research topic for both the process system engineering and operations research communities. Single or parallel machine scheduling can be a challenging combinatorial optimization problem when a large number of jobs are to be scheduled. Each job has unique job characteristics, resulting in different setup times/costs depending on the processing sequence. They also have specific release dates to follow and due dates to meet. This work presents both an exact method using mixed-integer quadratic programming, and an approximate method with metaheuristics to solve real-world large-scale single/parallel machine scheduling problems faced in a steel plant. More than 1000 or 350 jobs are to be scheduled within a one-hour time limit in the single or parallel machine problem, respectively. The objective of the single machine scheduling is to minimize a combined total changeover, total earliness, and total tardiness cost, whereas the objective of the parallel machine scheduling is to minimize an objective function comprising the gaps between jobs before a critical time in a schedule, the total changeover cost, and the total tardiness cost. The exact method is developed to benchmark computation time for a small-scale single machine problem, but is not practical for solving the actual large-scale problem. A metaheuristic algorithm centered on variable neighborhood descent is developed to address the large-scale single machine scheduling with a sliding-window decomposition strategy. The algorithm is extended and modified to solve the large-scale parallel machine problem. Statistical tests, including Student's t-test and ANOVA, are conducted to determine efficient solution strategies and good parameters to be used in the metaheuristics. / Thesis / Master of Applied Science (MASc)

In this dissertation, neural networks are designed to effectively model static non-linear aeroelastic problems in adaptive structures and linear dynamic aeroelastic systems with time varying stiffness. The use of adaptive materials in aircraft wings allows for the change of the contour or the configuration of a wing (morphing) in flight. The use of smart materials, to accomplish these deformations, can imply that the stiffness of the wing with a morphing contour changes as the contour changes. For a rapidly oscillating body in a fluid field, continuously adapting structural parameters may render the wing to behave as a time variant system. Even the internal spars/ribs of the aircraft wing which define the wing stiffness can be made adaptive, that is, their stiffness can be made to vary with time. The immediate effect on the structural dynamics of the wing, is that, the wing motion is governed by a differential equation with time varying coefficients. The study of this concept of a time varying torsional stiffness, made possible by the use of active materials and adaptive spars, in the dynamic aeroelastic behavior of an adaptable airfoil is performed here. A time marching technique is developed for solving linear structural dynamic problems with time-varying parameters. This time-marching technique borrows from the concept of Time-Finite Elements in the sense that for each time interval considered in the time-marching, an analytical solution is obtained. The analytical solution for each time interval is in the form of a matrix exponential and hence this technique is termed as Matrix Exponential time marching. Using this time marching technique, Artificial Neural Networks can be trained to represent the dynamic behavior of any linearly time varying system. In order to extend this methodology to dynamic aeroelasticity, it is also necessary to model the unsteady aerodynamic loads over an airfoil. Accordingly, an unsteady aerodynamic panel method is developed using a distributed set of doublet panels over the surface of the airfoil and along its wake. When the aerodynamic loads predicted by this panel method are made available to the Matrix Exponential time marching scheme for every time interval, a dynamic aeroelastic solver for a time varying aeroelastic system is obtained. This solver is now used to train an array of neural networks to represent the response of this two dimensional aeroelastic system with a time varying torsional stiffness. These neural networks are developed into a control system for flutter suppression. Another type of aeroelastic problem of an adaptive structure that is investigated here is the shape control of an adaptive bump situated on the leading edge of an airfoil. Such a bump is useful in achieving flow separation control for lateral directional maneuverability of the aircraft. Since actuators are being used to create this bump on the wing surface, the energy required to do so needs to be minimized. The adverse pressure drag as a result of this bump needs to be controlled so that the loss in lift over the wing is made minimal. The design of such a "spoiler bump" on the surface of the airfoil is an optimization problem of maximizing pressure drag due to flow separation while minimizing the loss in lift and energy required to deform the bump. One neural network is trained using the CFD code FLUENT to represent the aerodynamic loading over the bump. A second neural network is trained for calculating the actuator loads, bump displacement and lift, drag forces over the airfoil using the finite element solver, ANSYS and the previously trained neural network. This non-linear aeroelastic model of the deforming bump on an airfoil surface using neural networks can serve as a fore-runner for other non-linear aeroelastic problems. This work enhances the traditional aeroelastic modeling by introducing time varying parameters in the differential equations of motion. It investigates the calculation of non-conservative aerodynamic loads on morphing contours and the resulting structural deformation for non-linear aeroelastic problems through the use of neural networks. Geometric modeling of morphing contours is also addressed. / Ph. D.

The design of any mechanical components requires an understanding of the general statical, dynamical and environmental conditions where the components will be operating to give a satisfactory results in terms of performance and endurance. The premature failure of any components is undesirable and potentially catastrophic, therefore predictions on performances and endurances of components to proceed with repair or substitution is vital to the stability of the structure where the component is inserted. The capability of a component of withstanding fatigue loading conditions during service is called fatigue life and the designed predictions can be conservative or non conservative. Improvements to a strain based approach to fatigue were obtained in this study, studying the effects of mean stresses on fatigue life and investigating cyclic mean stress relaxation of two aluminum alloys, 7075-T6511 and 7249-T76511, used in structural aircraft applications. The two aluminum alloys were tested and their fatigue behavior characterized. The project, entirely funded by NAVAIR, Naval Air Systems Command, and jointly coordinated with TDA, Technical Data Analysis Inc., was aimed to obtain fatigue data for both aluminum alloys, with particular interest in 7249 alloy because of its enhanced corrosion resistance, and to give guidelines for improving the performances of FAMS, Fatigue Analysis of Metallic Structures, a life prediction software from the point of view of both mean stress effects and mean stress relaxation. The sensitivity of engineering materials to mean stresses is of high relevance in a strain based fatigue approach. The performance of the most common models used to calculate mean stress correction factors was studied for the two aluminum alloys 7075 and 7249 to give guidelines in the use of those for life predictions. Not only mean stresses have a high influence on fatigue life, but they are also subjected to transient cyclic behaviors. The following study considered both an empirical approach and a plasticity theory approach to simulate and include these transient effects in life calculations. Results will give valid directions to a successful modification of FAMS like any other life calculation software to include in the picture transient phenomena. / Ph. D.

The lithium disilicate (LS2) glass system provides the basis for a large number of useful glass-ceramic products. Microwave processing of materials such as glass-ceramics offers unique benefits over conventional processing techniques. Variable frequency microwave (VFM) processing is an advanced processing technique developed to overcome the hot spot and the arcing problems in microwave processing. In general, two main questions are addressed in this dissertation: 1. How does microwave energy couple with a ceramic material to create heat? and, 2. Is there a "microwave effect" and if so what are the possible explanations for the existence of that effect? The results of the present study show that VFM processing was successfully used to crystallize LS2 glass at a frequency other than 2.45 GHz and without the aid of other forms of energy (hybrid heating). Crystallization of LS2 glass using VFM heating occurred in a significantly shorter time and at a lower temperature as compared to conventional heating. Furthermore, the crystallization mechanism of LS2 glass in VFM heating was not exactly the same as in conventional heating. Although LS2 crystal phase (Orthorhombic Ccc2) was developed in the VFM crystallized samples as well as in the conventionally crystallized samples as x-ray diffraction (XRD) confirmed, the structural units of SiO4 tetrahedra (Q species) in the VFM crystallized samples were slightly different than the ones in conventionally crystallized samples as the Raman spectroscopy revealed. Moreover, the observed reduction in the crystallization time and apparent temperature in addition to the different crystallization mechanism observed in the VFM process both provided experimental evidence to support the presence of the microwave effect in the LS2 crystallization process. Also, the molecular orbital model was successfully used to predict the microwave absorption in LS2 glass and glass-ceramic. This model was consistent with experiments and indicated that microwave-material interactions were highly dependent on the structure of the material. Finally, a correlation between the Fourier transform infrared reflectance spectroscopy (FTIRRS) peak intensities and the volume fraction of crystals in partially crystallized LS2 glass samples was established. / Ph. D.

This dissertation provides three essays. In the first essay, a model with two linguistic variables is built to demonstrate the joint effect of multiple linguistic variables in a dynamic modeling context. Triangular membership function is used to represent the linguistic variables and the joint effect is captured through fuzzy inference method. In this essay, the results obtained by employing fuzzy concepts are compared with the results that one would obtain using generic lookup functions. The second essay develops a system dynamics model by which policy makers can assess the impact of various travel demand management interventions within a metropolitan area and as a consequence understand the complex behavior of affected transportation-socioeconomic systems. This essay builds on a previously formulated approach where fuzzy concepts are used to represent five linguistic variables used in the model. We also compare the level of traffic congestion under the scenarios with and without traffic congestion pricing. The third essay is based on the second essay where different scenarios of the travel demand management policies are evaluated and analyzed. There are two parts in this essay. The first part addresses the construction of a Management Flight Simulator (MFS) that is used to do policy analysis for travel demand management policies. By using the Management Flight Simulator, the second part of the essay describes the evaluation of alternative travel demand management policies. In this research, we found that the revenue generated from congestion pricing does increase mass transit capacity even with the aging of mass transit capacity. However, in the short term traffic congestion is mitigated while in the long term the proposed travel demand management policy actually deteriorates the traffic situation. / Ph. D.

This dissertation is intended to provide a statistical foundation for the IV models and shed lights on a number of issues related to the IV method. The first chapter shows that the theoretical Instrumental Variable model can be derived by reparameterization of a well-specified statistical model defined on the joint distribution of the involved random variables as the actual (local) data generation process. This reveals the covariance structure of the error terms of the usual theory-driven instrumental variable model. The revealed covariance structure of the IV model have important implications, particularly, for designing simulation studies. Monte Carlo simulations are used to reexamine the Nelson and Startz (1990a) findings regarding the performance of IV estimators when the instruments are weak. The results from the simulation exercises indicate that the sampling distribution of ^Î _IV is concentrated around ^Î _OLS. The second chapter considers the underlying joint distribution function of the instrumental variable (IV) model and presents an alternative definition for the exogenous and relevant instruments. The paper extracts a system of independent and orthogonal equations that covers up a non-orthogonal structural model and argues that the estimated IV regression is well-specified if the underlying system of equations is well-specified. It proposes a new instrument relevancy measure that does not suffer from the first-stage <i>R²</i> deficiencies. Third chapter argues the application of the IV method in estimation of models with omitted variable. The paper considers the implicit parametrization of statistical models and presents five conditions for an appropriate instruments. Two of them are empirically measurable and can be tested. This improves the literature by adding one more objective criterion for the selection of instruments. This chapter applies the IV method to estimate the rate of return to education in Iran. It argues that the education of two cohorts of Iranians was delayed or cut short by the Cultural Revolution. Therefore, the Cultural Revolution, as an exogenous shock to the supply of education, establishes the year of birth as the exogenous and relevant instrument for education. Using the standard Mincerian earnings function with control for experience, ethnicity, location of residence and sector of employment, the instrumental variable estimate of the return to schooling is equal to 5.6&#37;. The estimation results indicate that the Iranian labor market values degrees more than years of schooling. / Ph. D.

With the rapid development of modern high-throughput technologies, scientists can now collect high-dimensional complex data in different forms, such as medical images, genomics measurements. However, acquisition of more data does not automatically lead to better knowledge discovery. One needs efficient and reliable analytical tools to extract useful information from complex datasets. The main objective of this dissertation is to develop innovative Bayesian methodologies to enable effective and efficient knowledge discovery from complex high-dimensional data. It contains two parts—the development of computationally efficient functional mixed models and the modeling of data heterogeneity via Dirichlet Diffusion Tree. The first part focuses on tackling the computational bottleneck in Bayesian functional mixed models. We propose a computational framework called variational functional mixed model (VFMM). This new method facilitates efficient data compression and high-performance computing in basis space. We also propose a new multiple testing procedure in basis space, which can be used to detect significant local regions. The effectiveness of the proposed model is demonstrated through two datasets, a mass spectrometry dataset in a cancer study and a neuroimaging dataset in an Alzheimer's disease study. The second part is about modeling data heterogeneity by using Dirichlet Diffusion Trees. We propose a Bayesian latent tree model that incorporates covariates of subjects to characterize the heterogeneity and uncover the latent tree structure underlying data. This innovative model may reveal the hierarchical evolution process through branch structures and estimate systematic differences between groups of samples. We demonstrate the effectiveness of the model through the simulation study and a brain tumor real data. / Doctor of Philosophy / With the rapid development of modern high-throughput technologies, scientists can now collect high-dimensional data in different forms, such as engineering signals, medical images, and genomics measurements. However, acquisition of such data does not automatically lead to efficient knowledge discovery. The main objective of this dissertation is to develop novel Bayesian methods to extract useful knowledge from complex high-dimensional data. It has two parts—the development of an ultra-fast functional mixed model and the modeling of data heterogeneity via Dirichlet Diffusion Trees. The first part focuses on developing approximate Bayesian methods in functional mixed models to estimate parameters and detect significant regions. Two datasets demonstrate the effectiveness of proposed method—a mass spectrometry dataset in a cancer study and a neuroimaging dataset in an Alzheimer's disease study. The second part focuses on modeling data heterogeneity via Dirichlet Diffusion Trees. The method helps uncover the underlying hierarchical tree structures and estimate systematic differences between the group of samples. We demonstrate the effectiveness of the method through the brain tumor imaging data.