CNN-Based Automatic Modulation Classification of Radio Frequency Signals Under Varying SNR Conditions

Abstract

This paper presents modulation classification using the RadioML 2016.10A dataset under varying SNR conditions, with a primary focus on high-SNR scenarios (≥10 dB). Automatic modulation classification (AMC) plays a vital role in modern cognitive radio and wireless communication systems by enabling dynamic spectrum access, interference mitigation, and intelligent resource allocation. A residual Convolutional Neural Network (CNN) is designed and implemented using regularization techniques and a high-SNR-focused training strategy to improve classification performance. To provide a comparative benchmark, two classical machine learning models, Random Forest (RF) and Support Vector Machine (SVM), are trained using statistical features derived from the I/Q signal representations. Experimental results show that the proposed CNN consistently achieves higher classification accuracy in high-SNR scenarios with an overall accuracy of 88%, outperforming the RF and SVM baselines, which achieve 72.29% and 70.31%, respectively. Moreover, performance analysis across different SNR levels illustrates the resilience of deep learning approaches in noisy environments compared to traditional methods. t-SNE and Grad-CAM tools visualize how the model groups data and highlight key signal features. The results validate the potential of deep residual architectures for AMC in practical communication systems and provide a reproducible framework for future research in high-SNR signal classification.