Younghwan Jeong, Taemin Hwang, Won-gi Choi, Jinyoung Lee, Seolyoung Park, Sangshin Lee, Journal of Internet Computing and Services, Vol. 26, No. 3, pp. 83-97, Jun. 2025
10.7472/jksii.2025.26.3.83, Full Text:  HTML
Keywords: Digital Twin, Federated learning, similarity search, object tracking, system optimization

Abstract

Digital twin is a M&S (Modeling and Simulation) technology designed to quickly express physical changes in reality as synchronized virtual objects, and to precisely predict and infer through simulation to solve and optimize phenomena or problems. In order to apply digital twins to a more realistic environment and expand their scale, the computational load of the server must be offloaded to the device terminal, and the prediction process must be optimized to suit the operating method of the terminal. In this paper, we propose a common model federated training-inference method (DT-MAS) for accelerating real-virtual synchronization of digital twins. The proposed approach reduces the training cost through similarity-based search and noise injection in the federated learning process for distributing the training cost, and simplifies the inference procedure to reduce the inference load of the terminal, thereby securing an intelligent system of scalable digital twins. As a result of experiments with various scales and conditions, the proposed method is 14.13% better in training accuracy, 1.61 times better in training speed, and 7.68% better in inference efficiency than the existing method, proving that it is an effective intelligent synchronization method that can simultaneously satisfy scalability, resource efficiency, and real-time in a large-scale digital twin environment.

Statistics
Cite this article