Research On Vehicle Scheduling Optimization For Federated Edge Learning In Internet Of Vehicles

With the rapid development of information and communication technology,road transport has entered an era of data abundance from an era of data scarcity,and vehicles are interconnected through vehicle networking technology to improve transportation efficiency and safety.Vehicles can process the data collected by on-board sensors and achieve massive information interaction with the surrounding environment,thus not only achieving rational traffic planning,but also meeting the needs of different vehicle applications such as autonomous driving,intelligent transportation,location awareness,object/gesture recognition,mobile biometrics and mobile augmented reality.However,vehicle information data is often private and different vehicles are not willing to share their data with other vehicles,so it is difficult to fully satisfy the needs of vehicle applications.To solve this problem,federated learning(FL)is introduced to protect vehicle privacy by sharing models rather than sharing the raw data of vehicles.Each vehicle obtains a local model by training data locally and then uploads the local model to the central cloud.The central cloud aggregates the uploaded local models to get a global model,and then sends the global model to the vehicle to complete the relevant vehicle application.In this way,vehicles can obtain accurate models while protecting the original data.However,local training in FL can impose a large computational burden on vehicles with limited computing power,and sharing models for communication between multiple vehicles and the central cloud can lead to problems such as high latency and network congestion.Therefore,it is necessary to consider applying mobile edge computing technology to FL to form a federated edge learning(FEEL)system that introduces a three-tier infrastructure including a central cloud,roadside units connected to edge servers,and vehicles to selectively schedule vehicles to maximize model accuracy while ensuring cache queue stability.This paper focuses on FEEL systems in Telematics to analyze and optimize the performance related to vehicle selection scheduling based on FEEL systems.The main research work included in this paper is as follows.(1)Research on vehicle selection scheduling for federated edge learning.In the FEEL system,the disrupted data uploaded by vehicles for training are temporarily stored in the cache queue of each road side unit(RSU),and the cache queue of each RSU is limited.For each RSU,if too many vehicles upload data,the cache queue will overflow and become unstable;if too few vehicles upload data,there will be insufficient data in the cache queue for training,thus reducing the accuracy of the model.Therefore,a FEEL vehicle selection and scheduling method is proposed to address the problem that it is difficult to train a highly accurate model while ensuring the stability of its own cache queue when the RSU selects and schedules vehicles.Finally,the priority is calculated based on the four resource status of each vehicle in the coverage of the RSU: remaining data amount,communication quality,remaining energy and survivability,so as to select the appropriate number of vehicles with better resource status to upload data for model training.The simulation experiments are compared with several other commonly used selection methods,and it is demonstrated that the proposed selection method can maximize the training accuracy and help more vehicles while ensuring the stability of the cache queue.(2)Research on vehicle selection scheduling method based on Cellular-Vehicle to Everything(C-V2X)communication in federated edge learning.The newly designed C-V2 X of the 3rd Generation Partnership Project(3GPP)is considered as an alternative communication technology to IEEE 802.11 p because it supports direct communication between vehicles and RSUs and vehicles using the PC5 interface(also called V2 X side-chain communication).Based on the previous chapter,considering direct communication between vehicles to RSU and vehicles to vehicles using C-V2 X,a vehicle selection scheduling method based on C-V2 X communication in FEEL is proposed for C-V2 X mode 4,where first,each vehicle predicts its own packet transmission delay;then,each vehicle selects radio resources and calculates its own packet collision probability based on a sense-based Semi-Persistent Scheduling(SPS)scheme defined in C-V2 X mode 4;next,packet transmission delay and packet collision probability are introduced into the priority of the selected vehicle,and the selected vehicle uploads data for model training;finally,the selected vehicle is judged If a collision occurs,the vehicle and its data are excluded.The performance of C-V2 X communication with different numbers of vehicles,packet transmission frequency,transmission power and number of subchannels is analyzed through simulation experiments,and compared with several other commonly used selection methods.It is proved that the proposed selection method can maintain the stability of the selected vehicle and further improve the training accuracy while ensuring the performance of C-V2 X communication.(3)Packet collision probability optimization scheme based on C-V2 X communication.For autonomous vehicles and intelligent transportation systems in vehicular networking,highly reliable data transmission and convergence are crucial.Based on the previous chapter,although the SPS process in C-V2 X mode 4 contains mechanisms such as random selection of resources and probabilistic reselection to avoid packet collisions,even so,the packet collision probability is still not negligible,especially in congested scenarios,i.e.,scenarios with more vehicles,and its performance is far from the expected result.Therefore,a scheme for vehicles to dynamically adjust their own packet transmission frequency to reduce packet collision probability is proposed.First,the optimization objective of the packet collision probability optimization scheme based on C-V2 X communication is constructed;then,each vehicle calculates its own optimal packet transmission frequency by greedy algorithm;finally,vehicles use the optimal packet transmission frequency for data transmission,so that the packet collision probability of each vehicle can be reduced to the The proposed scheme is analyzed through simulation experiments.The performance of the proposed scheme in optimizing the packet collision probability is analyzed through simulation experiments,and compared with the packet collision probability calculated by taking three values of packet transmission frequency defined in the 3GPP protocol,which proves that the proposed scheme can reduce the packet collision probability efficiently,which can enable more vehicles to transmit data successfully and improve the reliability of data transmission.