The invention relates to a collision early warning method based on adaptive federated learning, and belongs to the technical field of intelligent driving. The method comprises the following steps: the RSU determines training parameters in the global update of the round, wherein the training parameters comprise the local training times tau and the number N of local models participating in aggregation; the ICV locally updates the collision early warning model for tau times by using the local model parameters and the local data set, and uploads the local model parameters to the RSU; when the number of the local models received by the RSU reaches N, executing model aggregation to obtain a global collision early warning model, otherwise, continuing to wait for other ICVs to upload local model parameters; when the global collision early warning model converges or reaches the specified model precision, training is ended; and the RSU adopts a deep reinforcement learning method of A3C to realize adaptive parameter adjustment. According to the invention, on the premise of ensuring the prediction precision, real-time training and resource conditions are fully considered, and the training parameters are adaptively adjusted, so that the training time delay of the collision early warning model is reduced.
本发明涉及一种基于自适应联邦学习的碰撞预警方法,属于智能驾驶技术领域。该方法包括:RSU确定本轮全局更新中的训练参数,包括本地训练次数τ和参与聚合的局部模型数量N;ICV利用本地模型参数和本地数据集进行τ次碰撞预警模型的本地更新,并将本地模型参数上传至RSU;当RSU接收到的局部模型数量达到N时,执行模型聚合获得全局碰撞预警模型,否则继续等待其他ICV上传本地模型参数;当全局碰撞预警模型收敛或达到指定模型精度,结束训练;RSU采用A3C的深度强化学习方法实现自适应调整参数。本发明能在保证预测精度的前提下,充分考虑实时的训练和资源情况,自适应调整训练参数,以降低碰撞预警模型训练时延。
Collision early warning method based on adaptive federated learning
一种基于自适应联邦学习的碰撞预警方法
2023-05-30
Patent
Electronic Resource
Chinese
European Patent Office | 2022
|Collision early warning method and collision early warning device
European Patent Office | 2021
|Anti-collision early warning system and anti-collision early warning method
European Patent Office | 2024
|Deep learning-based space target collision early warning method
European Patent Office | 2023
|