Electronic Theses and Dissertations

Author

Rabin Nepal

Date

2025

Document Type

Thesis

Degree Name

Master of Science

Department

Electrical & Computer Engineering

Committee Chair

Bonny Banerjee

Committee Member

Peter Lau

Committee Member

Shelby Peel

Abstract

Our goal is to develop a uniform, relatively lightweight model that can yield high accuracy on a wide variety of multimodal time-series classification problems. To that end, we propose a model that fuses multiple convolutional neural networks (CNNs), one for each modality, at the decision level using a multilayer perceptron. The model is trained on randomized (as opposed to sequential) signal windows, which enhances generalization across diverse time segments. We evaluate the model on two real-world problems: emotion recognition using physiological signals and action recognition using wearable sensor data. In both cases, our model achieves accuracy comparable to state-of-the-art approaches while being considerably smaller in size. This efficiency makes it well-suited for edge computing scenarios, where memory and compute resources are limited. The modular design also allows for seamless scaling across additional modalities or domains without requiring much architectural changes. Overall, our method provides a practical and adaptable framework for multimodal time-series classification.

Comments

Data is provided by the student

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Embargoed until 2027-07-15

Available for download on Thursday, July 15, 2027

Share

COinS