Research On Key Problems Of Distributed Deep Learning For Internet Of Vehicles Application

Edge caching and vehicle driving behavior prediction are important applications of Internet of vehicles,and distributed deep learning provides important support for Internet of vehicles applications.Traditional distributed deep learning is usually implemented in two ways.The first is that the edge device uploads all the collected data to the cloud server for neural network training,and then the cloud server deploys the trained model in the edge devices.The other is that the edge device and cloud server conduct neural network training at the same time in the way of parameter synchronization.When the neural network training is completed,the edge device provides services directly.Edge caching and vehicle driving behavior prediction requires high performance of neural network to provide service support,however,there are some problems in the above two types of neural network training methods.On the one hand,sending all data to the cloud server needs to transmit large amounts of raw data,not only will cause great pressure to communication network,at the same time can also lead to privacy;On the other hand,the training mode based on parameter synchronization needs to communicate with each iteration,and also needs to transmit a large amount of parameter data and cause high delay.In addition,edge devices and cloud devices need to maintain the same scale neural network,resulting in poor network flexibility.The above problems restrict the development of distributed deep learning in the Internet of vehicles.Therefore,it is necessary to conduct in-depth research on the key issues of distributed deep learning in edge caching and vehicle driving behavior prediction.The specific expansion is the following three aspects.1)Research on efficient framework of distributed deep learning communication based on knowledge distillationA distributed deep learning framework Ubi NN based on knowledge distillation is proposed to solve the problem of communication effectiveness of distributed deep learning.The framework uses knowledge distillation technology for model aggregation.Each edge device collects real-time data locally,trains the local neural network model independently,and flexibly controls the scale of the neural network on the premise of ensuring the performance of the neural network.An algorithm KDC based on knowledge distillation and covariance is proposed,so that each edge device only needs to send part of the data to the cloud to build a public data set to complete the training of the final neural network in the cloud,without sending all the data or synchronizing the parameters of the neural network with the cloud regularly,so as to reduce the amount of transmission data and communication delay.2)Research on distributed cache location decision framework based on convolutional neural networkA distributed framework CNN-Ubi NN based on convolutional neural network is proposed to solve the problem of cache location decision.A cache location prediction algorithm CLP is proposed,which enables the convolutional neural network to predict the corresponding cache location list by using the vehicle request information and edge device information,realize sequential download,improve the hit rate of cache,and reduce the delay of user request.3)Research on distributed prediction framework of vehicle driving behavior based on edge enhanced graph convolution neural networkA distributed framework EGCN-Ubi NN based on edge enhanced graph convolution neural network is proposed to solve the problem of vehicle driving behavior prediction.The framework takes advantage of two important characteristics of edge enhanced graph convolution neural network,namely edge enhanced attention mechanism and feature transfer mechanism of graph convolution neural network.The edge enhanced attention mechanism enriches the edge features and makes the distribution of attention coefficients more accurate,so as to better obtain the interactive features of vehicles;The feature transfer mechanism of graph convolution neural network makes the interactive features of vehicles transfer in the form of graph,which fully represents the interactive relationship between vehicles,so that the prediction of vehicle driving behavior has higher accuracy and lower calculation time.