VeNet: Hybrid Stacked Autoencoder Learning for Cooperative Edge Intelligence in IoV

Emerging applications of the Internet of Vehicles (IoV) require the wireless transmission of growing amounts of data, e.g., vehicle location and sensor data, over unreliable and increasingly congested wireless links between the mobile vehicles and the Road Side Units (RSUs); also, urban areas are becoming increasingly congested with vehicle road traffic. Road traffic management and data network traffic management to address these challenges require accurate representations of the road and network traffic, which are difficult due to the wide temporal and spatial correlations in the road and network traffic. We address this representation problem by designing, implementing, and evaluating the VeNet deep learning system to exploit the wirelessly transmitted data to predict future vehicle locations and network traffic. We develop the novel VeNet hybrid learning system that employs a stacked autoencoder (AE) consisting of a central AE and multiple local AEs that jointly feed into a Long-Short Term Memory (LSTM). We propose a new training algorithm for the hybrid VeNet learning system. The novel VeNet hybrid learning system conducts spatial learning that accounts for the spatial and temporal correlations in the dataset gathered from the mobile vehicles. Evaluations that involve measurements with custom-made Raspberry Pi vehicles indicate that the VeNet learning model significantly reduces the required signalling network traffic and prediction errors (down to approx. three quarters) compared to existing prediction models. At the same time, VeNet reduces the energy consumption on the vehicles as well as the learning delay.