Modulation classification for RFSoC showcasing streaming-CNN architectures

MacLellan, Andrew and Crockett, Louise H. and Stewart, Robert W. (2022) Modulation classification for RFSoC showcasing streaming-CNN architectures. In: 2022 IEEE SPS - EURASIP Summer School, 2022-08-29 - 2022-09-02, Linköping University.

[thumbnail of MacLellan-etal-EURASIP-2022-Modulation-classification-for-RFSoC-showcasing-streaming-CNN-architectures]
Preview
 Text. Filename: MacLellan_etal_EURASIP_2022_Modulation_classification_for_RFSoC_showcasing_streaming_CNN_architectures.pdf
Final Published Version
License: Strathprints license 1.0
Download (5MB)| Preview

Abstract

In wireless communications, the amount of available spectrum is reducing as the demand for connectivity increases. The increase of devices such as mobile phones, Internet of Things, smart vehicles, and wearable devices are making sections of the radio spectrum quite congested. Shared spectrum is an approach that aims to optimise the use of wireless communications channels to be shared among multiple users. A core aspect of shared spectrum is spectrum sensing, where a radio device detects other users within the nearby channels in order to transmit with minimal interference. Knowledge of how data has been transmitted from nearby users can assist in the transmission decisions for a radio device. In this work, we present a modulation scheme classifier Convolutional Neural Network (CNN) running on an AMD-Xilinx Zynq RFSoC 2x2 development board utilising a streaming-based architecture. Our CNN is implemented entirely on the Zynq RFSoC chip including all weights and activations and uses 18-bit fixed point arithmetic. The user control and visualisation functionality is designed using the PYNQ software framework. The full system can be controlled on a web browser and interface with the development board using Python and interactive widgets. The project is open-source.

CORE (COnnecting REpositories)