Real-Time Automatic Modulation Classification Using Deep Learning on Software-Defined Radio Testbeds

Abstract

Automatic Modulation Classification (AMC) plays a critical role in cognitive radio and spectrum monitoring applications. Although deep learning (DL) techniques have demonstrated strong performance in AMC, most existing studies are confined to simulation-based evaluations. This work addresses this limitation by presenting an end-to-end, real-time AMC system implemented on Software-Defined Radios (SDRs). A dual-USRP testbed is employed for over-the-air (OTA) transmission and reception of four digital modulation schemes: QPSK, 8PSK, 16QAM, and 64QAM. Rather than relying on raw in-phase and quadrature (I/Q) samples, the proposed system adopts a feature engineering strategy based on robust higher-order cumulants (up to sixth order) and phase-domain statistical features. Three deep learning architectures—Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Residual Network (ResNet)—are trained and evaluated under real-time OTA conditions. Experimental results reveal that the MLP model achieves the most balanced and consistent performance across all modulation types, exceeding 90% classification accuracy in real-time operation with an inference latency below 1 ms. While the CNN exhibits strong performance for PSK modulations, and the ResNet achieves the highest accuracy for QPSK, the MLP attains peak accuracies of 98.6% for 64QAM and 98.2% for 8PSK. These results demonstrate the robustness of the proposed system to real-world channel impairments and highlight its suitability for practical deployment in intelligent wireless communication systems.