04/08/2025
By Ruifeng Liu

The Richard A. Miner School of Computer and Information Sciences invites you to attend a doctoral dissertation defense by Ruifeng Liu entitled: "Anomaly Detection and Monitoring in Dynamic Network Systems, A Representation Learning Approach."

Ph.D. candidate name: Ruifeng Liu
Defense Date: Tuesday, April 8th, 2025
Time: 3 to 4:30 p.m. EST
Location: This will be a virtual defense via Zoom, Meeting ID: 994 9792 2710

Committee Members:
Tingjian Ge, Ph.D. (Advisor), Professor, Graduate Coordinator for MS Programs, Miner School of Computer & Information Sciences, CHORDS
Cindy Chen, Ph.D. (Member), Associate Professor, Associate Chair, Miner School of Computer & Information Sciences
Yuanchang Xie, Ph.D. (Member), Professor, Civil and Environmental Engineering

Dynamic network systems, such as urban traffic networks, are increasingly managed through Intelligent Transportation Systems (ITS), which significantly improve safety and efficiency. By utilizing data from sources such as the Department of Transportation (DOT) and other transportation agencies, ITS can adapt dynamically to complex and changing demands. Traditional ITS approaches have typically relied on historical data; however, recent advancements in real-time data collection and immediate feedback enable smarter and more responsive decision-making. Modern ITS, empowered by deep learning, can monitor traffic flow, detect incidents and anomalies, alleviate congestion, prevent accidents, enhance work zone safety, and supply critical information for autonomous vehicles.

This Ph.D. thesis focuses on detecting and monitoring traffic incidents and anomalies within dynamic network systems, emphasizing arterial roads. The research integrates advanced data simulation techniques with deep learning models to accurately predict changes in traffic flow and identify deviations from expected patterns. Data was simulated for major intersections in Waterwill, Maine, located near highways and urban centers. This simulated data encompasses traffic flows from different directions and lanes, as well as traffic control signals. These signals effectively regulate vehicle movement on roads, creating various traffic flow patterns. Additionally, an analysis of data provided by the Centracs Mobility Platform was conducted, highlighting its limitations.

An innovative representation learning approach was developed, incorporating state-of-the-art graph neural networks (GNN), recurrent neural networks (RNN), attention mechanisms, and transformer encoders to predict traffic from both temporal and spatial dimensions. The goal of this research is to achieve efficient, accurate, and reliable traffic predictions, while providing precise anomaly detection capabilities.

By rapidly pinpointing the time and location of traffic anomalies, this study enhances the understanding of how temporal, spatial, and signal control factors influence traffic dynamics. Furthermore, it seeks to provide actionable insights for transportation agencies, with the potential to effectively improve traffic conditions in urban dynamic network systems.