Date of Award

August 2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering

First Advisor

Naima Kaabouch

Abstract

Unmanned aerial vehicle (UAV) systems are increasingly being adopted across commercial, industrial, and autonomous applications, creating a growing demand for intelligent monitoring techniques that improve operational safety, situational awareness, and airspace management. Among recent research directions in UAV systems, trajectory analysis and behavior monitoring have become essential for supporting reliable and scalable UAV operations. This thesis investigates two complementary challenges in UAV monitoring: trajectory prediction and abnormal motion-deviation detection.The first part of this thesis focuses on UAV trajectory prediction under incomplete observation conditions. A deep learning framework was developed to estimate future UAV positions using historical flight information and maintain trajectory continuity during periods of missing or interrupted telemetry data. Multiple deep learning models, including a Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Transformer, were implemented and evaluated using real-world UAV trajectory datasets. Experimental results demonstrated that sequence-based models consistently outperformed non-temporal models by more effectively capturing temporal dependencies, leading to more accurate future trajectory prediction.

The second part of this thesis presents a comparative framework for detecting abnormal UAV motion deviations using large-scale Remote ID telemetry data. Motion descriptors representing speed, acceleration, altitude, heading, and their temporal variations were extracted and organized into fixed-length sliding windows to characterize short-term UAV dynamics. Because the RID datasets do not contain verified abnormal motion labels, a controlled motion-deviation injection methodology was developed to generate reproducible benchmark datasets while preserving realistic flight behavior. Within this unified framework, three complementary detection approaches were evaluated: a statistical threshold-based detector, an MLP Autoencoder, and a Transformer Autoencoder. The statistical detector provides an interpretable and computationally efficient baseline, whereas the AI-based autoencoders learn normal UAV motion patterns and detect deviations using reconstruction error. Experimental results demonstrated that the MLP Autoencoder achieved the highest overall F1-score, while the Transformer Autoencoder provided the highest detection sensitivity through temporal sequence modeling. Although the statistical approach remained computationally efficient and highly interpretable, its reliance on fixed thresholds limited its ability to detect complex motion deviations, demonstrating the advantages of learning-based methods for UAV motion-deviation detection.

Together, the two proposed frameworks address both predictive and diagnostic aspects of intelligent UAV monitoring by estimating future UAV motion and detecting abnormal motion deviations from Remote ID telemetry. The combined outcomes of this thesis establish a comprehensive data-driven approach that improves UAV situational awareness and operational reliability while supporting future applications in Remote ID surveillance, autonomous traffic management, collision avoidance, and large-scale airspace monitoring.

Download

Share

COinS