Date of Award

5-1-2024

Language

English

Document Type

Master's Thesis

Degree Name

Master of Science (MS)

College/School/Department

Department of Electrical and Computer Engineering

Dissertation/Thesis Chair

Hany Elgala

Committee Members

Dola Saha, Gary Saulnier

Keywords

Hybrid Waveform, Machine Learning, Model optimization, Multi-task learning, Task relationships, Visible Light Communication

Subject Categories

Electrical and Electronics

Abstract

Wireless communications have become ubiquitous, enabling seamless connectivity and driv- ing innovations across various domains. As we look to the future, visible light communication (VLC) is a promising technology that offers the potential to revolutionize how we transmit and receive data. It seamlessly integrates multiple functionalities, including localization, control/sensing, and high-speed data transmission.This thesis proposes a multi-task learning deep convolutional neural network approach to optimize a hybrid waveform for VLC-enabled networks. By integrating Beacon Posi- tion Modulation (BPM), Beacon Phase Shift Keying (BPSK), and OFDM symbols within a virtual Pulse Width Modulation (PWM) envelope, this waveform supports localization, control/sensing, and high-speed data transmission. The proposed multi-task learning (MTL) model employs a cross-stitch unit to learn shared representations between BPM and BPSK classification tasks. Extensive experiments demonstrate the model’s superior performance over single-task models, achieving high accuracy, fast convergence, and significant reductions in model complexity. Furthermore, this thesis quantifies task relationships within the multi- task learning framework by analyzing cross-stitch weights. The results reveal a high positive correlation between BPM and BPSK tasks, with values of 0.71 and 0.65, indicating a strong effective learning of shared representations for related tasks. These findings contribute to VLC-enabled networks by providing a unified learning framework that optimizes multiple functionalities while offering high computational efficiency.

Download

Share

COinS