Research On Obstacles Collision Avoidance Based On Enhanced Federated Learning In Intelligent Connected Vehicles

Intelligent Connected Vehicles(ICVs),as an emerging paradigm of intelligent transport,is expected to become an important part of the connected society in the future.The collision avoidance of intelligent vehicles is a key technology to ensure safe driving.Using deep reinforcement learning techniques,intelligent vehicles can adapt optimal vehicle behaviour based on decision networks to avoid obstacles,but single intelligent vehicles suffer from a lack of data and computing power when training decision networks.Federated learning techniques break the data silo dilemma in ICVs by sharing and aggregating model parameters,enabling the collaboration of computing power among intelligent vehicles,but vehicle mobility can directly affect the model performance of federated learning.To this end,this thesis presents a detailed study on the improvement of federated learning and federated learning-based collision avoidance in mobile node scenarios.An improved federated learning framework is proposed to address the impact of vehicle mobility on training speed and performance,and the framework is used to train a collision avoidance control decision network for intelligent vehicles to improve the training speed and performance of the decision network.The main research components of the thesis are as follows:Considering the influence of intelligent vehicle mobility on communication quality and data relevance,as well as the large time overhead caused by complex models in aggregation,this thesis designs an improved Mobility-aware Federated Reinforcement Learning(MFRL)framework.By mathematically modelling the effect of node mobility on communication quality and data relevance,a mobility-aware node selection algorithm is proposed that selects user nodes for federated learning based on model quality,communication quality and data relevance to accelerate the training speed and improve the model performance of federated learning.The MFRL framework also incorporates a knowledge-distillation based model compression method to reduce the communication overhead and model aggregation latency of federated learning.The experimental results show that MFRL can effectively accelerate the training speed and improve the performance of the decision network,and the compressed micro-decision network can effectively reduce the time overhead of model aggregation.In order to overcome the disadvantages of rule-based methods in vehicle collision avoidance control,such as complex rule design and unstable decision effect,this thesis develops a MFRL-based vehicle collision avoidance control method.The vehicle obstacle avoidance problem is modelled as an infinite state Markov decision process,and an intelligent vehicle obstacle avoidance algorithm based on Deep Deterministic Policy Gradient(DDPG)is proposed,while the distributed collaborative training of the intelligent vehicle obstacle avoidance decision network is implemented based on MFRL.The experimental results show that the DDPG model can guide intelligent vehicles to make reasonable safety collision avoidance control decisions,while MFRL can further improve the training speed and model performance of the decision network.