ORCID

0000-0003-2750-0358

Keywords

Real-Time Crash Prediction System, Anomaly Detection Learning, Machine Learning, Large Language Models, I-4 Express Managed Lane

Abstract

This research develops a real-time crash prediction system that integrates machine learning techniques and real-time data to forecast crash likelihood across various road segments. The system particularly addresses newly constructed Interstate 4 Express managed lanes, modeling them separately due to their unique challenges, such as limited crash data and new road designs. By treating these segments individually, the system ensures more accurate predictions that account for their distinct traffic behaviors. By leveraging anomaly detection learning (ADL), the system identifies rare crash events by detecting deviations from normal traffic behavior, even with imbalanced data. ADL is applied specifically to the influenced segments of the I-4 Express Lanes, effectively addressing the challenges of limited crash data by identifying traffic flow anomalies and providing reliable predictions. Furthermore, the system’s real-time crash prediction capabilities are expanded by integrating multiple models for segment types with a multi-layered approach to primary, secondary, and severity crashes, offering a comprehensive view of traffic safety. Additionally, the system is designed to specialized modules. This modular approach ensures adaptability across different segment types, optimizing the system for large-scale, real-world traffic data. The inclusion of specialized modules, such as the user interface, further enhances the system's ability to deliver crash risk summary reports and visual updates for traffic operators. Additionally, fine-tuned large language models (LLMs) are introduced to improve the prediction's interpretability, generating textual explanations for predicted crashes and offering actionable insights for informed decision-making. The system is evaluated, demonstrating its ability to accurately forecast crashes and contribute to proactive traffic management. By combining machine learning, anomaly detection, and LLMs, this integrated approach offers a significant advancement in traffic safety, providing timely, accurate, and explainable predictions that improve overall transportation safety and management.

Completion Date

2025

Semester

Spring

Committee Chair

Abdel-Aty, Mohamed

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Department of Civil, Environmental and Construction Engineering

Identifier

DP0029418

Document Type

Dissertation/Thesis

Campus Location

Orlando (Main) Campus

Share

COinS