Machine Learning for RF Cloud in Proximity Detection and Intelligent Spectrum Management: An Empirical Study

Su, Zhuoran

Etd

Machine Learning for RF Cloud in Proximity Detection and Intelligent Spectrum Management: An Empirical Study

Público Deposited

The RF cloud is a shared Radio Frequency (RF) database supported by various wireless technologies. The RF cloud, constituted by billions of Internet of Things (IoT) devices, Wi-Fi access points, and cellular towers, forms a complex ecosystem transmitting vast quantities of data via RF signals. Beyond the exchange of information packets, the RF cloud encompasses detailed signal characteristics such as the Receive Signal Strength Indicator (RSSI), Time of Arrival (TOA), and Channel State Information (CSI). This rich data environment enables the development of intelligent applications in cyberspace, ranging from enhancing wireless positioning accuracy, advancing motion detection capabilities, building security frameworks, to optimizing spectrum sharing techniques. Emerging techniques often outperform traditional estimation methods because they leverage Machine Learning (ML) and Deep Learning (DL) to utilize RF signal characteristics more effectively. However, incorporating ML and DL into RF Cloud applications introduces challenges related to complexity, interpretability, adaptability, and the cost of implementation. In this dissertation, we focus on two novel problems of RF cloud applications, proximity detection and mobility support for spectrum sharing, to mitigate these concerns with theoretical analysis and empirical study validation. 1) We evaluate the efficacy of Machine Learning (ML) algorithms versus classical estimation theory in addressing the proximity detection challenge, in both theoretical foundations and empirical performance evaluation. Utilizing the Mitre Range Angle Structured (MRAS) Private Automated Contact Tracing (PACT) dataset, we compare classical estimation methods and ML algorithms—Support Vector Machines, Random Forest, and Gradient Boosted Machines—on Bluetooth Low-Energy (BLE) Received Signal Strength Indicator (RSSI) data. We contrast the complexity, availability, and precision of RSSI-based BLE and Time of Arrival (TOA)-based Ultra-Wide Band (UWB) technologies for proximity detection during epidemics. We present a detailed performance evaluation for the theoretical precision and confidence limits, and the performance on a new empirical dataset collected from diverse environments. Our findings reveal the UWB TOA algorithm's superior precision and confidence, albeit with the advantages of BLE RSSI in smartphone integration and minimal computational requirements. 2) We conduct an empirical study on RF interference intensity (RII) in licensed and unlicensed bands along a downtown Worcester, MA route to demonstrate the spatiotemporal RII behavior, utilizing a mobile spectrum monitoring system. This study informs an ML approach to predict channel availability for vehicular network spectrum access with fourteen RII features. We present the accuracy in predicting channel availability and regenerating RII as a validation of the proposed theoretical foundations, highlighting the potential of explainable ML in intelligent spectrum access and mobility support in the context of next-generation wireless networks. This comprehensive analysis showcases the potential of utilizing classical theoretical foundations for enhancing the interpretability, adaptability, and efficiency of ML in proximity detection and spectrum management, offering significant implications for public health safety measures during epidemics and for the wireless communication efficiency of next-generation wireless networks.

Creator