DEEP LEARNING-BASED MODULATION CLASSIFICATION USING SYNTHETIC AND OVER-THE-AIR SDR SIGNALS

Abstract

This work presents a convolutional neural network (CNN) for automatic classification of dig ital modulation schemes using signals received via software-defined radios (SDRs). A synthetic dataset was created in MATLAB for BPSK, 8-PSK, 16-PSK, QAM, 16-QAM, and 64-QAM, with 1,500 messages per scheme at five SNR levels and added phase noise. A 12-layer CNN trained on this dataset achieved 97% accuracy in classifying modulation types. To evaluate real-world performance, over-the-air signals were captured and used for validation, yielding classification accuracies ranging from 72% to 91%. While performance on live signals showed variability, the results indicate strong potential for generalization with further refinement. Enhancing the synthetic dataset with additional channel impairments may improve model robustness and real-world appli cability. This research demonstrates the viability of using deep learning for signal classification in intelligent communication systems.