Traffic Data Reconstruction Based On Neighbor Node Information

Multi-access edge computing(MEC)technology can reduce the transmission cost of cloud-centered Internet of Vehicles scene and provide faster interactive response.However,the amount of data between the roadside unit(RSU)detector and the MEC server has not decreased.MEC server resource space is limited,and the large amount of traffic data uploaded by RSU may limit the development of MEC network.We achieve the compression of traffic data by reducing the sampling frequency directly during the collection of traffic data,to reducing the "last mile",which from the detectors to MEC servers,transmission data of the edge computing network.Compressed Sensing technology proposes if the sampling frequency is lower than the requirement of the sampling theorem,the complete data set can also be recovered when the data has a certain structure.In this paper,SIMM1 and SIMM2 are proposed to reconstruct traffic data based on compressed sensing by adding the range constraint of neighbor node information into the ordinary compressed sensing model.These two methods respectively use different neighbor data regularization method in the objective function of the reconstruction model to improve the accuracy of the basic Compressed Sensing reconstruction model and make the model more suitable for the high data loss scenario.The method of CS technology to reconstruct complete data through matrix decomposition has the limitation of using only linear modeling.In this paper,in order to increase the degree of non-linearity of the model,a model framework NRTD based on neural network is proposed to reconstruct traffic data.The process of Generalized Matrix Factorization is instantiated under NRTD framework,and the latent features of Multi-Layer Perceptron are combined to improve the fitting results of the model.We use the traffic flow data collected by the highway RSU detector node network.We classify the RSU detector nodes and design a low sampling frequency working mode for some detectors based on the classification results to filter redundant information in the data.This mode reduced the amount of data from RSU to MEC server.In the experiment of reconstructing data,we perform parameter optimization and model comparison experiments with different missing rates and missing scenarios.The experimental results show that the proposed method can outperform the basic Compressed Sensing reconstruction model with different conditions.