Nonlinear Traffic Data Reconstruction in Large-Scale Internet of Vehicle Systems: A Neural Network Approach

Abstract

Multi-Access Edge Computing (MEC) can reduce transmission costs and provide faster interactive responses in cloud server-centric Internet of Vehicle (IoV) scenarios. However, the resource space of MEC server is limited, and the large amount of data uploaded by RSU may limit the further development of MEC network. Compressed Sensing (CS) proves that for the data with certain internal structure, a complete data set can be reconstructed even if the sampling frequency is lower than the requirements of the sampling theorem. In the current works, traffic data compression is accomplished by directly reducing the sampling frequency in the process of collecting traffic data. However, CS reconstruction only utilizes the linear characteristics of data. Aiming to fully utilize the nonlinear characteristics of data, we propose a neural network based traffic data reconstruction model framework, called NRTD. NRTD instantiates the process of generalized matrix factorization and combines the potential features of multilayer perceptrons to improve the fitting degree of the model to the traffic flow data change curve. Parameter optimization and model comparison experiments are carried out with different missing rates and missing scenes. The experimental results show that the proposed method is superior to the conventional compressed sensing reconstruction model before improvement under different conditions.