Defense Notices

Data classification is a methodology of data mining used to organize data by relevant categories to obtain meaningful information. A model is generated from the input training set which is used to classify the test data into predetermined groups or classes. One of the most widely used models is a decision tree which uses a tree like structure to list all possible outcomes. Decision tree is an important predictive analysis method in Data Mining as it requires minimum effort from the users for data interpretation. 

This project implements ID3, an algorithm for building decision tree using information gain metric. Furthermore, through illustrating the basic ideas of ID3, this project also addresses the inefficiency of ID3 in handling continuous numerical attributes. A cut point approach is presented to discretize the numeric attributes into discrete intervals and enable ID3 functionality for them. Experiments show that such decision trees contain fewer number of nodes and branches in contrast to a tree obtained by basic ID3 algorithm. This modified algorithm can be used to classify real valued domains containing symbolic and numeric attributes with multiple discrete outcomes. 

Data sets may have instances where multiple values are possible which are described as set-value attributes. The established LEM2 algorithm does not handle data sets with set-value attributes. To solve this problem, a parallel approach was used during LEM2's execution to avoid preprocessing data. Changing the creation of characteristic sets and attribute-value blocks to include all values for each case allows LEM2 to induce rules on data sets with set-value attributes. The ability to create a single local covering for set-value data sets increases the variety of data LEM2 can process.

Machine learning is being applied in many domains of research. One such research area is the automation of gender prediction. The goal of this project is to determine a person’s gender based on his/her voice. Although it may seem like a simple task for any human to recognize this, the difficulty lies in the process of training a computer to do this job for us. This project is implemented by training models based on input data of voice samples from both male and female voices. The voice samples considered were from different datasets, with varying frequencies, noise ratios etc. This input data is passed through various machine learning models, with/without parameter tuning, to compare results. A comparative analysis of multiple machine learning algorithms was conducted, and the prediction with the highest accuracy is displayed as output for the given input voice sample.

Every student comes to the class with a smart phone, and they are constantly distracted. It has become a tough challenge for the instructors to keep the students focused on the lectures. The idea of this project is to build a hybrid responsive web application which helps the instructors to post questions between their discussions. The students can give their responses through their smart phones instantly. This enables the instructor to analyze the understanding of the students on the current topic through various statistics which are generated instantly. The instructors can improve their teaching methods while the students who are less interactive can give their voice along with others in the class and check their understanding. 

This application allows the instructor to add or edit courses in their account, add students to their courses, create or edit quizzes beforehand, post questions in different formats to the students, and analyze results through various kinds of plots. On the otherhand, a Student can view the courses he is added in to by his/her instructor, submit his/her responses for the quizzes posted. This application simplifies the process of conducting in-class quizzes and offers the students and the instructors an enhanced classroom experience. 

The dynamic thinning of fast-flowing glaciers is so poorly understood that its potential impact on sea level rise remains unpredictable. Therefore, there is a dire need to predict the behavior of these ice bodies by understanding their bed topography and basal conditions, particularly near their grounding lines (the limit between grounded ice and floating ice). The ability to detect previous VHF radar returns in some key glacier regions is limited by strong clutter caused by severe ice surface roughness, volume scatter, and increased attenuation induced by water inclusions and debris. 
The work completed in the context of this thesis encompasses the design, integration, and field testing of a new compact light-weight radar and antenna system suitable for low-frequency operation onboard Uninhabited Aerial Systems (UASs). Specifically, this thesis presents the development of two tapered dipole antennas compatible with a 4-meter wingspan UAS. The bow-tie shaped antenna resonates at 35 MHz, and the meandering and resistively loaded element radiates at 14 MHz. Also discussed are the methods and tools used to achieve the necessary bandwidth while mitigating the electromagnetic coupling between the antennas and on-board avionics in a fully populated UAS. The influence of EM coupling on the 14 MHz antenna was nominal due to relatively longer wavelength. However, its input impedance had to be modified by resistive loading in order to avoid high power reflections back to the transmitter. The antenna bandwidths were further enhanced by employing impedance matching networks that resulted in 17.3% and 7.1% bandwidths at 35 MHz and 14 MHz, respectively. 
Finally, a compact 4 lbs. system was validated during the 2013-2014 Antarctic deployment, which led to echo sounding of more challenging temperate ice in the Arctic Circle. The thesis provides results obtained from data collected during a field test campaign over the Russell glacier in Greenland compared with previous data obtained with a VHF depth sounder system operated onboard a manned aircraft. 

Technological advancements in analog and digital systems have enabled new approaches to study networks of physical and artificial neurons. In biological systems, a standard method to record neuronal activity is through cortically implanted micro electrode arrays (MEAs). As advances in hardware continue to push channel counts of commercial MEAs upwards, it becomes imperative to develop automatic methods for data acquisition and analysis with high accuracy and throughput. Reliable, low latency methods are critical in closed-loop neuroprosthetic paradigms such as spike-timing dependent applications where the activity of a single neuron triggers specific stimuli with millisecond precision. This work presents an adapted version of an online spike detection algorithm, previously employed successfully on in vitro recordings, that has been improved to work under more stringent in vivo environments subject to additional sources of variability and noise. The algorithm’s performance was compared with other commonly employed detection techniques for neural data on a newly developed and highly tunable extracellular recording model that features variable firing rates, adjustable SNRs, and multiple waveform characteristics. The testing framework was created from in vivo recordings collected during quiescence and electrical stimulation periods. The algorithm presents superior performance and efficiency in all evaluated conditions. Furthermore, we propose a methodology for online signal integrity analysis from MEA recordings and quantification of neuronal variability across different experimental settings. This work constitutes a stepping stone toward the creation of large scale neural data processing pipelines and aims to facilitate reproducibility in activity dependent experiments by offering a method for unifying various metrics calculated from single unit activity. Precise spike detection becomes crucial for experiments studying temporal in addition to rate coding mechanisms. To further study and exploit the potential of temporal coding, a delay-feedback-based reservoir (DFB) has been implemented in software. This artificial network is found to be capable of processing spikes encoded from a benchmark task with performance comparable to that of more complex networks. This work allows us to corroborate the capabilities of temporal coding in a minimally-complex system suitable for implementation in physical hardware and inclusion in low-power circuit applications where computational power is also necessary.

The worldwide energy demand is being increased day by day, anticipated to increase for 48% from 2012 to 2040. The distributed generation (DG) including renewable energy resources such as wind and solar are part of the solution in terms of lowering electricity cost, power reliability, and environmental concerns and therefore must function efficiently. Designing a robust maximum power point tracking (MPPT) technique can ensure maximized energy harvesting from PV solar systems and increases conversion efficiency which is the significant hindrance for their growth. The maximum power point (MPP) varies with intrinsic and climate changes nonlinearly. Thus, MPPT methods are expected to seek the MPP regardless of the solar module and ambient changes. The proposed method is based on the concept of Model Predictive Control (MPC) with unique properties. MPC is a powerful class of controllers that uses a system modeling to predict future behavior and optimize performance objectives. Unlike the traditional techniques that are prone to lose a tracking direction and their consequences on the stability, the proposed technique treats the photovoltaic (PV) module as a plant and uses a digital observer for predicting the behavior of the PV module and tracking the MPP. Further, it unifies the simplicity of implementation, enhances the overall dynamics performance and is robust against atmosphere changes.

Historical subsurface sensing methods applied to planar ice and snow sheets rely on underlying assumptions about the physical situation governing volumetric backscatter. Namely, the stratification of the natural medium under investigation consists of layered material with distinctly different dielectric properties. While appropriate for recovering sharp spatial discontinuities in the relative permittivity, the layer stripping approach [1] is not applicable to smooth permittivity variations about a common mean. In this project we developed techniques to model both the forward scattering from one-dimensional permittivity variation and the inverse problem - estimating the permittivity profile from the reflected energy. The underlying assumption is that smoothly varying inhomogeneities may be decomposed into wavelet basis functions which efficiently represent natural perturbations about an effective mean. Potential applications for this method are in ground penetrating radar, ionospheric sounding, nondestructive evaluation, and medical imaging.

Simulating the mechanics of a beating heart requires the numerical solution of partial differential equations. An application like this is a good candidate for high order computational methods that deliver higher solution accuracy at a lower cost than their low order counterparts. 
To fully leverage these high order computational methods, they must be paired with an accurate discretization of the domain. For a geometry like the heart, this requires a high order mesh. Thus robust high order mesh generation is a critical component to the widespread adoption of high order computational methods for numerically solving partial differential equations. Toward this end, we are developing high order mesh generation and untangling methods. As our first step, we have developed an optimization-based second order mesh generation method that employs triangles and tetrahedra. We will also develop generation methods for quadrilateral and hexahedral elements. Finally, we will develop untangling methods that can be used to untangle our generated meshes, as well as untangle any tangled elements that occur during motion (e.g. the beating of the heart). 

Unmanned Aerial Vehicles (UAVs) are increasingly demanded in civil, military and research purposes. However, they also possess serious threats to the society because faults in UAVs can lead to physical damage or even loss of life. While increasing their intelligence, for example, adding vision-based sense-and-avoid capability, has a potential to reduce the safety threats, increased software complexity and the need for higher 
computing performance create additional challenges—software bugs and transient hardware faults—that must be addressed to realize intelligent and safe UAV systems. 
This work present a fault tolerant system design for UAVs. Our proposal is to use two heterogeneous hardware and software platforms with distinct reliability and performance characteristics: High-Assurance (HA) and High-Performance (HP) platforms. The HA platform focuses on simplicity and 
verfiability in software and uses a simple and transient fault tolerant processor, while the HP platform focuses on intelligence and functionality in software and uses a complex and high performance processor. During the normal operation, the HP platform is responsible for controlling the UAV. However, if it fails due to transient hardware faults or software bugs, the HA platform will take over until the HP platform recovers. 
We have implemented the proposed design on an actual UAV using a low-cost Arduino and a high-performance Tegra TK1 multicore platform. Our case-studies show that our design can improve safety without compromising performance and intelligence of the UAV.