FPGA-Based Adaptive Digital Beamforming Using Machine Learning for MIMO Systems

Embargo until
Journal Title
Journal ISSN
Volume Title
Johns Hopkins University
In modern Multiple-Input and Multiple-Output (MIMO) systems, such as cellular and Wi-Fi technology, an array of antenna elements is used to spatially steer RF signals with the goal of changing the overall antenna gain pattern to achieve a higher Signal-to-interference-plus-noise ratio (SINR). Digital Beamforming (DBF) achieves this steering effect by applying weighted coefficients to antenna elements- similar to digital filtering- which adjust the phase and gain of the received, or transmitted, signals. Since real world MIMO systems are often used in dynamic environments, Adaptive Beamforming techniques have been used to overcome variable challenges to system SINR- such as dispersive channels or inter-device interference- by applying statistically-based algorithms to calculate weights adaptively. However, large element count array systems, with their high degrees of freedom (DOF), can face many challenges in real application of these adaptive algorithms. These statistical matrix methods can be either computationally prohibitive, or utilize non-optimal simplifications, in order to provide adaptive weights in time for an application, especially given a certain system's computational capability; for instance, MIMO communication devices with strict size, weight and power (SWaP) constraints often have processing limitations due to use of low-power processors or Field-Programmable Gate Arrays (FPGAs). Thus, this thesis research investigation will show novel progress in these adaptive MIMO challenges in a twofold approach. First, it will be shown that advances in Machine Learning (ML) and Deep Neural Networks (DNNs) can be directly applied to the computationally complex problem of calculating optimal adaptive beamforming weights via a custom Convolutional Neural Net (CNN). Secondly, the derived adaptive beamforming CNN will be shown to efficiently map to programmable logic FPGA resources which can update adaptive coefficients in real-time. This machine learning implementation is contrasted against the current state-of-the-art FPGA architecture for adaptive beamforming- which uses traditional, Recursive Least Squares (RLS) computation- and is shown to provide adaptive beamforming weights faster, and with fewer FPGA logic resources. The reduction in both processing latency and FPGA fabric utilization enables SWaP constrained MIMO processors to perform adaptive beamforming for higher channel count systems than currently possible with traditional computation methods.
FPGA, beamforming, deep learning, machine learning, adaptive, CNN, VHDL