Joint data processing and model maintenance for digital twin-enabled vehicular crowdsensing systems

Vehicular crowdsensing (VCS) leverages vehicles equipped with onboard sensors to collaboratively collect large-scale urban data, enabling applications such as traffic monitoring and environmental sensing. However, traditional VCS systems suffer from challenges in data reliability, resource allocation, and system scalability due to the lack of real-time global coordination. To address these issues, digital twin (DT)-enabled VCS has emerged as a promising paradigm by establishing dynamic virtual replicas of physical vehicles to enhance data utilization and decision-making. However, existing DT-enabled VCS studies mainly focus on incentive design and data aggregation, while overlooking the joint optimization of data processing and model maintenance, which is critical for maintaining synchronization between physical and virtual entities. In this paper, we formulate a Joint Data Processing and Model maintenance optimization (JDPM) problem aiming to maximize system utility under limited computing resources and latency constraints. We prove the NP-hardness of the JDPM problem and transform it into a multi-agent Markov decision process (MDP). To solve it, we propose a Resource Allocation strategy of joint Data processing and Model maintenance based on multi-agent Deep Reinforcement Learning (RADM-DRL) with an actor–critic architecture and an attention-enhanced critic network to capture inter-agent dependencies. Extensive experiments demonstrate that RADM-DRL significantly outperforms baseline algorithms in terms of system utility, latency, and resource utilization rate, offering an efficient and intelligent framework for resource management in DT-enabled VCS.

© 2026 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.