Digital Over-the-Air Computation with Software-defined Radio : A Practical System Design

Sammanfattning

Traditional network schemes treat communication and computation as independent tasks, which brings the benefits of minimalist system complexity, high freedom of network architecture, and incredible flexibility with regard to applications. However, with the rapid growth of Artificial Intelligence (AI) and Machine Learning (ML), computation-oriented tasks are more and more popular for giant internet companies, which calls for efficient function transmission. Over-the-air Computation (AirComp) is a new research trend in Wireless Sensor Network (WSN) and distributed learning networks. The goal of AirComp is to compute the function of messages by exploiting electromagnetic interference. Within AirComp, the vast majority of the work is based on theoretical methods and numerical simulation. There is a call for the implementation of these methods over real wireless channels, using transceiver chains implemented in hardware. In this thesis, an AirComp scheme to calculate the Majority Voting (MV) is proposed for Federated Edge Learning (FEEL), and a practical implementation of AirComp with the Software-defined Radio (SDR) device ADALM-PLUTO is demonstrated. This thesis addresses the theoretical study of digital AirComp by building an experimental platform and verifying the theory with the SDR. Through simulations and experiments, it is proved that the proposed AirComp scheme can provide a good MV performance even when the time-synchronization and the power control are not ideal under heterogeneous data distribution scenarios. This thesis also addresses and analyzes the challenges encountered during the construction of the hardware experimental platform, proposing alternative solutions. Additionally, it discusses the platform's limitations and suggests directions for future work for AirComp.