RFSoC Modulation Classification With Streaming CNN: Data Set Generation & Quantized-Aware Training

• Published in: IEEE open journal of circuits and systems Vol. 6; pp. 38 - 49
• Main Authors: Maclellan, Andrew, Crockett, Louise H., Stewart, Robert W.
• Format: Journal Article
• Language: English
• Published: New York IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; AMD; artificial intelligence (AI); Artificial neural networks; Classification; convolutional neural network (CNN); Convolutional neural networks; Data processing; data set generation; data-flow; Datasets; Deep learning; Deep learning (DL); Field programmable gate arrays; Fixed point arithmetic; Fixed points (mathematics); FPGA; general matrix multiplication (GEMM) transform; inference; Machine learning; Modulation; modulation classification; Multiplication; Pipelines; PYNQ; quantized-aware training; Radio; Real time; Real-time systems; Receivers; RFSoC; System on chip; Training; Wireless communication; Wireless communications
• ISSN: 2644-1225; 2644-1225
• DOI: 10.1109/OJCAS.2024.3509627

This paper introduces a novel FPGA-based Convolutional Neural Network (CNN) architecture for continuous radio data processing, specifically targeting modulation classification on the Zynq UltraScale+ Radio Frequency System on Chip (RFSoC) operating in real-time. Evaluated on AMD's RFSoC2x2 development board, the design integrates General Matrix Multiplication (GEMM) optimisations and fixed-point arithmetic. We also present a method for creating Deep Learning (DL) data sets for wireless communications, incorporating the RFSoC into the data generation loop. Furthermore, we explore quantised-aware training, producing three modulation classification models with different fixed-point weight precisions (16-bit, 8-bit, and 4-bit). We interface with the implemented hardware through the open-source PYNQ project, which combines Python with programmable logic interaction, enabling real-time modulation prediction via a PYNQ-enabled Jupyter app. The three models, operating at a 128 MHz sampling rate prior to the decimation stage, were evaluated for accuracy and resource consumption. The 16-bit model achieved the highest accuracy with minimal additional resource usage compared to the 8-bit and 4-bit models, making it the optimal choice for deploying a modulation classifier at the receiver.