Enhancing mmWave Network Connectivity and Edge Intelligence through Reconfigurable Intelligent Surfaces for Communication and RF Processing

The next generation of wireless networks needs to support massive device deployments and ensure robust, widespread connectivity, particularly for edge devices in various Internet of Things (IoT) applications. To achieve this goal, two primary requirements should be addressed: (1) maintaining strong and reliable signal connectivity for a large number of devices, and (2) enabling real-time, on-device data processing for edge devices. For the first requirement, millimeter wave (mmWave) technology has been proposed as a solution to achieve ultra-high network throughput. However, inherent characteristics of mmWave signals, including limited propagation range and high penetration loss, restrict the effective coverage of mmWave networks. In particular, non-line-of-sight (NLoS) issues arise when obstacles block line-of-sight (LoS) paths between transmitters and receivers, further limiting connectivity. The second requirement, on-device data processing for edge devices, is driven by the need for data privacy, real-time processing, and reduced dependence on internet connectivity. Yet, edge devices often face challenges due to limited computational resources, especially for local machine learning (ML) inference tasks. This constraint is increasingly significant with the growing demand for artificial intelligence (AI) applications in IoT scenarios. In this dissertation, we investigate the use of reconfigurable intelligent surfaces (RISs) to address the two challenges above. First, we investigate the optimization of signal coverage in mmWave networks using RISs, since RISs can be leveraged to establish alternative links in the absence of LoS links. To this end, a coverage analysis framework is proposed based on stochastic geometry to evaluate the performance limits of RIS-assisted links, particularly in terms of signal sensitivity to obstacles. This framework enables evaluation of coverage performance for both single-RIS and multi-RIS links with different types of RISs, including reflective RISs and transmissive-reflective RISs. Through this comparative analysis, the proposed framework can provide insights to assist RIS deployment strategies. Then multi-RIS deployment strategies are proposed to enhance signal coverage in indoor environments. Two scenarios are investigated: (1) obstacles are randomly distributed in the room, and only stochastic information about their presence is available; and (2) obstacles are located at fixed positions, with full information about their locations and sizes provided. In the first scenario, analytical results for optimal RIS placements are derived to maximize connection probability when room dimensions and the number of available RISs are given. In the second scenario, a gradient-descent-based algorithm is developed to optimize the locations and orientations of RISs, aiming to minimize NLoS regions and thereby enhance coverage reliability. To address the second challenge of on-device data processing, we explore the possibility of offloading computations from digital processors to the radio frequency (RF) domain by using RISs. To be specific, RISs serve as a medium for over-the-air (OTA) computation, where the goal is to manipulate RF signals that carry sensing or communication data directly in the RF domain, enabling signal propagation to emulate specific mathematical operations. RISs enable OTA computation by adjusting the electromagnetic properties of each RIS element to control incoming RF signals and produce the desired outgoing signals. In this way, RIS-based OTA computation offers the potential to reduce memory and computational demands on edge devices. In this dissertation, RIS-based OTA computation is explored for both non-ML and ML tasks. For non-ML cases, a direction-of-arrival (DoA) estimator comprised of one transmissive intelligent surface (TIS) and two receive antennas is proposed. The signal from the target passes through the RIS and is received by the antennas, with the DoA estimated from the power ratio between received signals at two antennas. This estimator offers advantages of low computational and hardware complexity, as it relies solely on received power, unlike classic methods that require phase information. In ML cases, RIS-based RF neural networks are explored, where sequentially placed TISs are used to mimic neural network layers for ML inference in the RF domain. Two types of RIS-based RF neural networks are investigated: (1) RF fully connected (FC) layers, and (2) RF two-dimensional (2D) convolutional layers. For FC layers, TISs with 1-bit and 2-bit phase configurations are used to reduce hardware complexity compared to continuous phase configurations, and a training scheme for quantized complex-valued RF neural networks is proposed. For 2D convolutional layers, an RF convolutional layer using three TISs is introduced to perform 2D convolution in the RF domain. Both types of RF neural networks are validated through simulations.

Date

2025-04-29

Resource Type

Text

Resource Subtype

Dissertation

Full item page