Research On Scenario Identification Algorithm Based On Deep Distributed Learning For Internet Of Vehicles

With the rapid development of traffic,the Intelligent Transport System(ITS)is becoming more and perfect.As a key hotspot in ITS research,vehicle communication has also received extensive attention.Vehicle networking technology and artificial intelligence technology promote the development of intelligent communication between vehicles;provide new ideas for applications such as autonomous driving and in-vehicle entertainment.In the Internet of Vehicles,recognizing the surrounding environment of the vehicle is essential to ensure road safety and avoid vehicle collisions.Moreover,from the technical point of view of the physical layer,the correct identification of the vehicle scenario can provide decision-making for the intelligent communication system,so that it can adjust the wireless channel model and transmission mode in time.At present,many researches apply deep learning technology to vehicle scenario identifications.Nevertheless,on the one hand,these methods are based on picture or video information,which consume too much computing resources.On the other hand,they rely on the cloud server of the Internet of Vehicles for centralized.The data collection process is complex and involves user privacy,while the training is not flexible.As a result,the efficiency is low,and the resource consumption is large.Therefore,this paper proposes a method of decentralized learning-based scenario identification(Decent SCI),which uses the wireless channel information as a data set,and adopts the neural network model to extract the inherent characteristic parameters of the wireless channel,without adding any additional photographing or detection equipment.In the process of decentralized learning,multiple in-vehicle terminals of the Internet of Vehicles use local data for independent and parallel model training and the server-side fuses the model to obtain a global model.Such a training process saves the cloud data collection process and improves training efficiency.However,multiple uploads and downloads of model parameters are required between the servers and the vehicles.Before decentralized learning,the model needs to be compressed to reduce model parameters,speed up model uploading and training.Finally,the Decent SCI method proposed in this paper can reduce model complexity,training time,computing storage space,and communication overhead,while accuracy is ensured.