This dissertation proposes the use of the shape of data as a new feature to improve and develop new in machine learning and deep learning algorithms utilized for different power systems applications. The new features are obtained through Shape Data Analysis (SDA), an emerging field in Statistics. SDA is used to obtain the shape of the data structure to observe different patterns developed under distribution networks abnormal conditions, as well as determining the shape of load curves to improve existing electrical load forecasting algorithms. Specifically, shape-based data analysis is implemented and developed for two different applications: electrical fault detection and electrical consumption short-term load forecasting. The algorithms proposed are implemented on data collected from Intelligent Electronic Devices (IEDs), Phasor Measurement Units (PMUs), and Supervisory Control and Data Acquisition (SCADA) systems in power distribution networks. / A Dissertation submitted to the Department of Electrical and Computer Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosophy. / Spring Semester 2019. / February 4, 2019. / Deep Learning, Fault Detection, Load Forecasting, Machine Learning, Shape Data Analysis / Includes bibliographical references. / Sastry Pamidi, Professor Directing Dissertation; Anuj Srivastava, University Representative; Eren Ozguven, Committee Member; Hui Li, Committee Member; Simon Foo, Committee Member.

With more renewable energy integrated into the existing power consumption, power electronics play an important part to convert and control the power. Power inverters employ power semiconductors to converter DC into AC, which is an essential part in the renewable energy utilization. The paralleled transformer-less inverters are well adopted in the industry for large capacity grid-tie application. Compared with the centralized inverter, inverters in parallel can offer higher power rating, higher reliability, and lower grid-side current harmonics. Transformers are commonly used in the grid-tie system to provide galvanic isolation and voltage ratio transformations. Eliminating transformers will be a great benefit to further improve the system efficiency, reduce the size and weight. However, removal of the transformer would result in ground leakage current between the DC input side and the grid ground. The emerging wide band gap (WBG) devices are bringing significant opportunities for inverters towards higher efficiency and higher power density, due to their substantial switching loss reduction over Si devices. Silicon carbide (SiC) adoption also brings new control challenges to the three-phase paralleled transformer-less inverters. The voltage slew rate can be as high as dozens or hundreds of volts per nanosecond and the harmonic frequency related with the turning-on and turning-off of the devices may be up to several hundreds of mega-hertz, these high dv/dt and di/dt can generate high frequency EMI noise that propagates to the whole system including the power stage and control circuits, and raise the issue of increased electromagnetic interference (EMI). With high switching frequency, it is more difficult to control the circulating current among paralleled inverters. The conventional carrier synchronization method cannot be applied due to the impact of communication and sample delay. Limited controller resource also prevents sophisticated control algorithms. In this research, a five-level T-type (5LT2) PV inverter paralleled through inter-cell transformer (ICT) is presented to elaborate the challenges and demonstrate the advantages in three-phase SiC inverter. There are three key current in the 5LT2 PV inverter: circulating current, grid current, and ground leakage current. Circulating current is suppressed by the ICT and further controlled by a current controller. With increased switching frequency and multilevel topology, it is possible for a SiC device based grid connected converter to achieve filter-less function and utilize the grid impedance for its switching harmonic attenuation. Analysis shows that the conventional control method with instantaneous grid voltage feedforward (IGVF) will significantly limit the bandwidth or stability margin of a filter-less grid-connected inverter, thus make the inverter sensitive to grid disturbance. Two proposed grid voltage feedforward control methods, which require little additional computation resources, are presented to suppress the grid voltage disturbance. The increased switching efficiency is beneficial to the high frequency (HF) ground leakage current suppression, since the common mode (CM) choke can be much smaller. The 5LT2 inverter has a significant common mode voltage (CMV) reduction compared to that of a 3-level T-type (3LT2) inverter. However, the low frequency (LF) ground leakage current caused by neutral point (NP) voltage oscillation becomes a new issue in larger power rating multi-level inverters. A LF CMV compensation method is proposed to suppress the LF CMV. In this research, a control system is developed for a 60 kW three-phase paralleled transformer-less filter-less SiC PV inverter, which achieves a power density of 27 W/in3 and 3 kW/kg with nature convection, and measured peak efficiency of 99.2%. / A Dissertation submitted to the Department of Electrical and Computer Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosophy. / Fall Semester 2018. / November 27, 2018. / Includes bibliographical references. / Hui Li, Professor Directing Dissertation; Jonathan Clark, University Representative; Thomas A. Lipo, Committee Member; Michael Steurer, Committee Member.

Electric power grids are evolving rapidly as increased emphasis is placed on integration of renewable resources into existing power infrastructures and as new paradigms for power production and distribution, such as microgrids, are developed. Resultant grid configurations must meet the needs and requirements of existing and evolving population distributions, feasible production facilities placement, and environmental stewardship associated with power transmission and distribution infrastructures. In most developed regions, large-scale transmission infrastructures are well established, and their geographic routing is increasingly difficult to alter or amend. Renewables integration, however, directs far more attention at the power distribution level. As more local power is produced, often intermittent in nature and sometimes by consumers themselves, power distribution becomes more problematic in several respects. Conceptually, “the grid” becomes less a fixed entity and more an ever-changing amalgam of sources, loads, and preferred routes among them. All such routes must meet certain fundamental physical requirements, such as current and voltage handling capabilities. For power quality and reliability reasons, however, they also need to be “stable” in several senses, and there is currently no comprehensive approach to selecting available or potential routes to optimize the resultant “stability” of the configuration, in any of the various senses. This work develops such an approach, applicable to the steady-state stability of grids subject to several simplifying constraints. That is, it provides a framework for analyzing the steady-state stabilities of all grid topologies for grids that meet those constraints. The approach is general and abstract in nature, as this work focuses not on particular commonly studied grids but instead on the characteristics of grid topologies that lend themselves to greater or lesser degrees of steady-state stability. As a baseline study, only grids having synchronous generators are considered, with the expectation that future work will adapt inertia-based models of renewable sources to this or a similar approach. Although the approach itself is the main contribution, several interesting discoveries have already been made regarding optimal configurations of some simple topologies and on quantifying how richness of grid interconnections influences grid steady-state stability. / A Dissertation submitted to the Department of Electrical and Computer Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosophy. / Spring Semester 2019. / April 15, 2019. / Includes bibliographical references. / Chris S. Edrington, Professor Directing Dissertation; William Oates, University Representative; Omar Faruque, Committee Member; Petru Andrei, Committee Member.

Speech is the most natural and fastest method of communication between humans. This fact compelled researchers to study acoustic signals as a fast and efficient means of interaction between humans and machines. For authentic human-machine interaction, the method requires that the machines should have the sufficient intelligence to recognize human voices and their emotional state. It is well-known that the emotional state of human drivers highly influences his/her driving performance. For example, there are many reports that describe road-rage incidents where drivers become emotionally enraged due to the actions of another driver. This anger may lead to a high-speed chase, tailgating, and sometimes even death due to a traffic crash or physical contact. If a car is ‘intelligent-enough’ to respond to a driver’s emotional state, it may be able to thwart negative outcomes of road-rage incidents. Speech emotion recognition, extracting the emotional state of speakers from acoustic data, plays an important role in enabling machines to be ‘intelligent’. Speech emotion recognition is an emerging field and presents many challenges. The set of most powerful features which can distinguish different emotions is not defined; hence, the selection of features is a critical task. Acoustic variability presented by numerous speech properties, such as length and complexity of human speech utterance, speaker’s gender, speaking styles and rate of speech, directly affects the most common speech features; thereby affecting the system performance. Most of the researchers used statistical approaches to recognize human speech; however statistical methods are complex and need more computational time. Moreover, emotion recognition being the developing field, researchers are exploring facial, gestural and acoustical features for emotion recognition. However, for vehicle-centric applications, audio and speech processing may provide better noninvasive and less distracting solutions than other interactive in-vehicle infotainment systems. Hence, acoustic feature extraction for emotion recognition in human drivers is a preferred design choice of this research. The goal of this research is to develop an optimal feature extraction algorithm for emotion recognition of four most common emotions (anger, happy, sad and no emotion). In this dissertation, six acoustic features are studied using decision-tree based algorithms to recognize speech-based human emotions and reduce the complexity of the system. The speech features used are pitch, intensity, frequency formants, jitter, shimmer and zero crossing rate. Pitch and intensity are qualitative voice feature, frequency formants and jitter provide the spectral features and zero-crossing rate and shimmer suffice as temporal features of human acoustical speech. The combination of different types of speech features is utilized to increase the accuracy of system. The decision tree-based algorithms are designed in MATLAB and are calculated using confidentiality-interval for each feature. For acoustic data visualization, PRAAT software is used. The system is designed for speaker-dependent emotion recognition since the accuracy of system is more as the utilized features are qualitative voice features; which are best-suited for emotion recognition. Data from two males and two females is analyzed for this dissertation. For the actual realization of system, noise analysis is performed using 5dB, and 15 dB signal-to-noise ratio levels. These are minimum and maximum noise levels experienced while driving on a freeway and parking lot. This dissertation is composed of five chapters. Chapter 1 presents the mechanism of human speech production and human emotions in speech. It comprises of various emotions and importance of acoustic signal for emotion recognition. Chapter 2 includes different local and global acoustic features, existing methods of speech recognition and emotion recognition and discusses the weaknesses of existing speech recognition systems for acoustical emotion recognition using various acoustic features and analysis algorithms. Chapter 3 outlines the proposed solution for acoustic emotion recognition using decision-tree based algorithm. It includes a description of each acoustical feature, data preparation techniques, data analysis methods, and algorithm design. Chapter 4 consists of results, discussion and comparison of proposed algorithm with state-of-the-art acoustic emotion recognition algorithms. Finally, conclusion, limitations and future work is discussed in chapter 5. / A Dissertation submitted to the Department of Electrical and Computer Engineering in partial fulfillment of the requirements for the degree of Doctor of Philosophy. / Fall Semester 2018. / November 29, 2018. / Acoustic Emotion Recognition, Acoustic Features, Speech Processing / Includes bibliographical references. / Shonda Bernadin, Professor Directing Dissertation; John O. Sobanjo, University Representative; Simon Foo, Committee Member; Bruce Harvey, Committee Member.

Semiconductor manufacturing companies aim to achieve shortest test times for products while maintaining the product quality. Achieving shortest test times for devices requires multiple updates to the test flow and test content. Test cost varies in direct proportion to production test time required to test chips and detect fails. This thesis presents a method to achieve shortest test times by determining when the updates are needed and what are the changes to the test flow and test content. This thesis introduces a new Adaptive Test Scheme (ATS). ATS estimates individual test fail rates dynamically, per die, and makes real-time modifications to test order and test contents. ATS computes data-driven test fail rate estimates and uses the estimates to identify the required changes and trigger updates to test flow and test content. ATS uses Bayesian statistics to model the per test fail rates and update the test orders. ATS achieves test time reductions by employing per wafer elimination. ATS also incorporates a simple quality monitor, by resetting the test content at the start of next wafer. This thesis evaluates the performance of ATS with synthetic data generated by a Monte Carlo method and with production wafer sort data for two manufactured products. The product data results show ATS reduced by 20% the total test-time for one product and by 40% for a second product, with changes in product quality level below industry targets.

Thesis (M.S.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 1990. / Includes bibliographical references (leaves 115-116). / by Stephen MacDonald Foster. / Thesis (M.S.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 1990.

Ultrasound waves have been used for imaging purposes for many years. However, a liquid interface has always been necessary between the transducer and the object being imaged due to a high mechanical resistance at the air-transducer interface. Recent advances in transducers have made it possible to omit the liquid interface, allowing imaging to be done through air interfaces. Because this is a relatively new field, research into ultrasound propagation in air is very limited. A comprehensive model of how an ultrasound wave propagates through air would expedite the study of air-coupled ultrasound for imaging. This thesis presents a mathematical model of two-dimensional linear acoustic wave propagation in air. The model takes as input the frequency and amplitude of an acoustic signal and outputs the pressure field over varying longitudinal and lateral distances from the source. The benefits of a mathematical model over a finite element model are first discussed, then the mathematical model for acoustic propagation in air is developed using both computer simulations and physical experiments in an anechoic chamber. Results are presented and compared to experimental data to confirm the validity of the mathematical model.

The modification of single crystal epitaxial ZnO thin films grown by Pulsed Laser Deposition on c-axis oriented sapphire substrates by Ion Beam Processing was investigated. Nitrogen doping of the films was attempted using nuclear transmutation using the ¹⁶O (³He, ⁴He) ¹⁵O reaction at 6.6 MeV. The ¹⁵O product is unstable and decays to ¹⁵N after several minutes by positron emission. There are several potential advantages to using nuclear transmutation including producing nitrogen atoms on the correct lattice site for doping and reduced crystal damage as compared to conventional ion beam implantation. In the experiments in this thesis the doping levels achieved ~10¹⁴ cm^-3 were too low to be expected to dope the films to p-type. However several beneficial effects due to the ion beam processing were observed, including large increases in resistivity, reduction of defect luminescence, and substantial increases in the response of photoconductive detectors. In addition to desired effects in some films it was also found that in some films bubble like structures approximately 10 ìm in diameter were formed where the thin film delaminated from the surface. It was assumed that mechanism for the bubble formation was the build up of helium gas at the sapphire/ZnO interface.

The design of a 4 Gbps serial link transceiver in 0.35µm CMOS process is presented. The major factors limiting the performance of high-speed links are transmission channel bandwidth, timing uncertainty and on-chip frequency limitations. The design uses a combination of multi-level signaling (4-PAM) and transmit pre-emphasis to overcome the channel low-pass characteristics. High on-chip frequency signals are avoided by multiplexing and de-multiplexing the data directly at the pads. Timing recovery is done through over-sampling the data using multi-phase clocks generated from a low-jitter PLL. The design achieves a 4 Gbps data transmission rate, with a transmit data jitter of 53.2 ps (p-p), while consuming 879.4 mW of power from a 3.3 V supply.

A series of free energy estimates can be calculated from the ribosome's progressive interaction with mRNA sequences during the process of translation elongation in eubacteria. A sinusoidal pattern of roughly constant phase has been detected in these free energy signals. Frameshifts of the +1 type occur when the ribosome skips an mRNA base in the 5'-3' direction, and can be associated with local phase-shifts in the free energy signal. We propose a mathematical model that captures the mechanism of frameshift based on the information content of the signal parameters and the relative abundance of tRNA in the bacterial cell. The model shows how translational speed can modulate translational accuracy to accomplish programmed +1 frameshifts and could have implications for the regulation of translational efficiency. Results are presented using experimentally verified frameshift genes across eubacteria.