Digital Twin-Enabled Efficient Federated Learning for Collision Warning in Intelligent Driving

• Published in: IEEE transactions on intelligent transportation systems 2024-03, Vol.25 (3), p.2573-2585
• Main Authors: Tang, Lun, Wen, Mingyan, Shan, Zhenzhen, Li, Li, Liu, Qinghai, Chen, Qianbin
• Format: Article
• Language: English
• Subjects: Adaptation models; Algorithms; asynchronous advantage actor-critic; Collision avoidance; Collision dynamics; Collision warning; Completion time; Computational modeling; Data models; Delays; digital twin; Digital twins; Federated learning; Machine learning; Mathematical models; Model accuracy; Parameters; Support vector machines; Training; training delay; Warning
• ISSN: 1524-9050; 1558-0016
• DOI: 10.1109/TITS.2023.3330938

Considering the limited resources, user mobility and unpredictable driving environment in intelligent driving, this paper studies the optimal training efficiency of federated learning for distributed training of collision warning services with the assistance of digital twin (DT). DT is emerging as one of the most promising technologies to make the digital representation of physical components for better prediction, analysis, and optimization of various services in intelligent driving. we first propose a DT-enabled collision warning framework, including physical network layer, digital twin layer, and application layer. Then, for the cooperative training of multi-level warning models combining gate recurrent unit (GRU) and support vector machine (SVM) in the digital twin layer, we propose semi-asynchronous federated learning with adaptive adjustment of parameters (SFLAAP) scheme. We aim at minimizing the training delay of collision warning model by dynamically adjusting the training parameters according to real-time training state and resource conditions of digital space, specifically the local training times and the number of local nodes participating in the aggregation, while ensuring the accuracy of the model. Considering the complexity of the target problem, we propose parameter adjustment algorithm based on asynchronous advantage actor-critic (A3C). Experiments on the classical dataset show high effectiveness of the proposed algorithms. Specifically, SFLAAP can reduce the completion time by about 12% and improve the learning accuracy by about 1%, compared with the state-of-the-art solutions.