RSU-Assisted Digital Twin Framework with Failure Recovery in Vehicular Networks

Abstract

The evolution of autonomous driving and connected vehicles has increased the need for real-time monitoring, performance optimization, and proactive intervention across dynamic vehicular environments, including both vehicle status and surrounding environmental conditions. Digital Twin (DT) technology has emerged as a promising paradigm to achieve this goal by enabling dynamic, virtual representation of physical systems for data-driven decision making. Existing DT-enabled vehicular systems often utilize roadside units (RSUs) to assist in data collection to maintain high accuracy and synchronization; however, they remain susceptible to interruptions due to RSU failures, resulting in critical synchronization gaps between the physical system and its digital counterpart. This mismatch between the physical system and the digital twins may compromise decision accuracy and safety measures. To address these challenges, this thesis proposes a robust DT framework that ensures high synchronization with the physical system in the presence of infrastructure disruptions caused by RSU failures. The framework operates across three coordinated layers, including a Physical System Layer: Real-World Vehicles and Infrastructure, a Decentralized Twin Layer: RSU Zone Digital Twins, and a Centralized Twin Layer: Vehicular Network Digital Twin, allowing optimized data collection, processing, and real-time decision making. The thesis formulates the problem as a mixed integer program to maximize the synchronization between the digital twin and its physical counterpart and proposes a reinforcement learning-based solution using Twin Delayed Deep Deterministic Policy Gradient (TD3) method to address the high mobility and uncertainty of vehicular environments. Synchronization between the twins is measured in terms of the sum age of information collected from the various sensors along the highway segment covered by the failed RSU. Simulation results demonstrate the proposed TD3-based solution achieves close-to-optimal results, while outperforming baseline approaches by over 63%.