Cyclo-AMC: Automatic Modulation Classification on Versal utilising Cyclostationary Features

Recently, it has been shown that cyclostationary features can greatly improve generalization of deep neural network (DNN) automatic modulation classification (AMC) systems. Unfortunately, its computation has a high computational cost, limiting its application in real-time settings. To address these issues, we present Cyclo-AMC, the first field-programmable gate array (FPGA)-based custom AMC accelerator that combines cyclostationary features with a DNN. Cyclo-AMC runs entirely on an AMD Versal VEK280 AI Engine-Machine Learning (AIE-ML) device and integrates an area-efficient implementation of the FFT accumulation method (FAM) algorithm to extract cyclostationary features, with a compact ResNet-18 model. Cyclo-AMC achieves a 32× reduction in energy consumption over an NVIDIA GeForce RTX 3090 graphics processing unit (GPU). On the CSPB.ML.2022 dataset, Cyclo-AMC achieves 93.8% classification accuracy compared to 81.7% for FINN-A.