Wireless techniques for next-generation edge applications : high-rate access, positioning, and learning

Abstract

With rapid evolution of wireless networks in the past few decades, the functions of wireless networks have shifted from providing high-rate data transmission to supporting a large variety of applications including high-rate access, accurate sensing (e.g., autonomous driving and positioning) and fast intelligence acquisition (e.g., machine learning and edge inference). The growth of the applications has prompted the emergence of the new paradigm of multi-functional networks, which dier from classic wireless networks viewing high-rate data transmission as the stringent goal. This motivates effective designs with high communication efficiency to fulfill the requirements of the emerging applications. The paradigm shift also opens opportunities for the integration of different wireless technologies to fully exploit their strength. This dissertation contributes novel task-specific designs for several representative emerging applications in wireless networks. For applications demanding high-rate data transmission, massive multiple-input multiple-output (MIMO) and device-to-device (D2D) communication are two promising technologies to boost the network throughput. Massive MIMO exploits the spatial degree-of-freedom by deploying a large number of antennas at the base station (BS). D2D communication establishes direct links between users at the cell edge to provide additional network throughput. Although both of them have been key technologies in the 5G standards, their integration still lacks investigation. In this dissertation, a comprehensive investigation on the interactions between the two technologies in the coexistence network is provided, followed by an effective rate adaptation design to control the outage probability. Next, as autonomous driving has come into reality in the past few years, vehicular sensing and communication has emerged as a key application to secure transportation and improve the traffic flow in congestion. For vehicular sensing, on-board positioning techniques have become an essential component to detect the surrounding environments independently, which improves the robustness of the system. Without assistance of the advanced communication technologies, the conventional RADAR system is unable to detect objects in non-line-of-sight (NLoS). To tackle this challenge, a so-called cooperative multi-point positioning (COMPOP) is proposed in this dissertation, which features positioning vehicles in both LoS and NLoS by exploiting novel waveform designs and multi-path propagations reflected by the nearby vehicles.Last, the enormous data generated by edge devices has contributed to the emergence of edge intelligence or edge AI. Considering the massive amount and dimensionality of the distributed data samples, data compression is necessary to relieve the transmission and computation burden, where principal component analysis (PCA) is viewed as a powerful tool. The principle of PCA is to find a subspace that preserves the most information of the original data. Without compromising the privacy, a distributed PCA algorithm, called as AirPCA, is proposed by federated training, where gradients are computed at the devices and uploaded to the server via over-the-air aggregation for the subspace update in each round. Moreover, by exploiting the channel noise effect, an adaptive power control design is provided to further accelerate the AirPCA.