Chengchen Mao

Graduation Semester and Year

2023

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Electrical Engineering

Department

Electrical Engineering

First Advisor

Qilian Liang

Abstract

ABSTRACT: Machine learning refers to a machine or an algorithm that draws experience from data. A certain pattern is found to build a model, which is used to solve real problems. Deep learning, an important branch and extension of machine learning, employs a neural network structure containing multiple hidden layers. It learns critical features of the data by combining lower-level features to form more abstract higher-level representations of attribute categories or features. In this dissertation, deep learning network models were applied to sense-through-foliage target detection and extended with Rake structure. The deep learning network models had a large number of redundant parameters from the convolutional layer to the fully-connected layer, and a large number of neuron activation values converged to zero. The challenging task was to reduce parameter redundancy while maintaining model accuracy. In Chapter 2, an approach based on stacked autoencoders (SAE) was proposed for ultra wide band radar for sense-through-foliage target detection. SAE, as one of the widely used deep learning structures, could learn representations of data with multiple levels of abstraction automatically. The SAE-based target detection approach performed well in processing poor signal collections in some positions. In other positions, a single radar target detection performed under satisfaction. Rake structure was applied in radar sensor networks with maximum ratio combining and equal combining to combine radar echoes from different radar cluster-members. In Chapter 3, pruning in deep learning network models was investigated. Pruning presented significant opportunities for compression and acceleration in deep neural networks by eliminating redundant parameters. Structured pruning gained popularity in the edge computing research area, especially with more terminal chips integrated with AI accelerators for Internet of Things (IoT) devices. Stripe-wise pruning (SWP), which conducted pruning at the level of stripes in each filter, was different from filter pruning and group-wise pruning. The existing SWP method introduced filter skeleton (FS) to each stripe, setting an absolute threshold for the values in FS, and removing stripes whose corresponding values in FS could not meet the threshold. The research involved investigating the process of stripe-wise convolution and using the statistical properties of the weights located on each stripe to learn the importance between those stripes in a filter and remove stripes with low importance. In Chapter 4, the conception of a deep energy autoencoder (EA) for a noncoherent multicarrier single-input and multiple-output (SIMO) system operating amidst multipath channels was explored. The multicarrier SIMO structure involved a single-antenna sender and a multi-antenna receiver, both depicted via neural networks. The encoder generated a real-valued vector for each subcarrier, while the decoder received the combination of energy from all the receiving antennas. To address the major challenge of mitigating intersymbol interference (ISI) caused by multipath channels without relying on delicate designs common in traditional communication systems, two different types of neural networks, namely DNN (Deep Neural Network) and RNN (Recurrent Neural Network), were adopted for the demodulation rule at the receiver. Simulation results demonstrated that, with adequate training, RNN efficiently recovered the transmitted data even in the absence of channel state information, which was often required in traditional communication systems.

Keywords

Machine learning, Deep learning

Disciplines

Electrical and Computer Engineering | Engineering

Comments

Degree granted by The University of Texas at Arlington

Download

Share

COinS