Adaptive Data Processing Method Based On Edge Intelligence In Internet Of Vehicles

With the rapid development of cloud computing technology,the combination of in-vehicle network and cloud computing technology is increasingly close,resulting in a variety of Internet of Vehicles(Io V)and intelligent transportation applications.However,while using cloud platform to provide services for on-board network applications,there are also many problems in the current research of Io V.First of all,the data collection mode in the Io V is single and the data redundancy is large,so the transmission cost is high.Secondly,the cloud server is far away from the in-vehicle network,and the transmission delay is caused by the network bandwidth transmission error and other factors,so the real-time performance cannot be guaranteed.Finally,as user data is in a shared business environment,the risk of privacy leakage is extremely high.The characteristics of edge computing are local deployment,proximity to users,and low latency.The vehicle node resources can be efficiently and directly managed.This paper studied the adaptive data processing method based on edge intelligence in Io V.From the three aspects of Qo S assessment and data collection,data cleaning,privacy protection,the paper puts forward solutions to the problems of Io V data respectively,which realizes the real-time and efficient data processing of Io V,and provides new ideas for the future research of Io V data.The main work of this thesis is listed as follows:(1)In order to effectively solve the various requirements of service quality and data collection for Io V applications,an adaptive data collection method based on fuzzy logic is proposed.Firstly,a Qo S evaluation model based on intelligent fuzzy system is established,and a two-level fuzzy system is designed to carry out Qo S parameter reasoning and global Qo S evaluation respectively.Then,the detailed evaluation scheme is designed for three different service types and two different execution modes,and the execution situation closer to the real Io V scene is selected for simulation,and fuzzy rules are formulated according to the simulation cases.Finally,an adaptive data collection and uploading strategy based on edge computing is proposed according to the Qo S evaluation results.The experimental results show that the proposed scheme can realize the overall performance evaluation in the Io V system,and effectively improve the adaptability of data collection.(2)In order to effectively solve the problem of data redundancy in Io V system,a data cleaning method based on image fingerprint and deep learning is proposed.First of all,only t he low-frequency information of the image is retained after image compression and gray processing.Meanwhile,the compression technology is improved according to the data characteristics of the Io V.Then the image fingerprint is obtained by calculating the hash value,which is used to check the similarity of data,establish the similarity baseline,and flexibly control the screening of similar images.Finally,the semi-supervised learning algorithm is used to verify the data correlation.The convolutional neural network(CNN)is first constructed with labeled data sets,and then lowered to the edge nodes after multiple sets of data constant error correction and adjustment,so as to further reduce redundant data.The experimental results show that the recogni tion accuracy of the proposed system is nearly as good as that of the supervised learning system,but the delay is reduced to 11% of this system.Therefore,the data cleaning scheme based on edge computing and deep learning significantly reduces the amount of data uploaded to the cloud,and effectively reduces the delay of data processing and transmission.(3)In order to effectively solve the problem of data privacy leakage of Io V users,a layered homomorphic encryption method based on federated learning is proposed.The encryption and decryption parts of Paillier and RSA are split and deployed in the car node and edge layer respectively,and the corresponding keys are distributed.The additive homomorphism is completed on the vehicle node and the edge serve r,and the cloud server encrypts the multiplicative homomorphism on the part of the model parameters that need to be updated,which allows the non-local end to operate directly on the ciphertext without exposing the plaintext.It not only strengthens the s ecurity of private data,but also improves the learning efficiency.In addition,a new node selection strategy,namely distributed data stream sampling algorithm,is designed to ensure that the honest nodes can participate in the model training as much as possible,and improve the recognition accuracy of the system.The experimental results show that the scheme has a very outstanding performance in security and accuracy.