Automatic Modulation Classification using a Deep Learning model based on ResNet

Abstract

Automatic Modulation Classification (AMC) serves as a foundational element in contemporary wireless communication systems, enabling adaptive signal processing and efficient spectrum management. This study explores deep learning-based approaches—particularly a Res-Net architecture—for robust AMC performance using the benchmark RADIOML 2018.01A dataset. The dataset comprises 24 modulation schemes across a wide SNR range from -20 dB to +30 dB. Comprehensive data preprocessing was performed, including normalization, as well as various augmentation methods like phase rotation, temporal shifting, and artificial noise addition to strengthen the model’s resilience and ability to generalize under challenging conditions. A Res-Net model was constructed and optimized with categorical cross-entropy as the loss function and Adam as the learning algorithm. The model achieved a test accuracy of 95.72% under high-SNR conditions (SNR > 8 dB) with a low training loss of 0.0933, demonstrating strong convergence and generalization capabilities. Confusion matrix analysis highlighted the model’s strengths in accurately classifying most modulation types, while revealing challenges in differentiating between similar schemes like 16QAM and 64QAM under low SNR conditions. The findings confirm that the proposed deep learning framework is capable of learning and distinguishing complex signal characteristics directly from unprocessed I/Q data without the need for manual feature crafting. Future work will focus on integrating Transformer-based architectures, wavelet transform features, and hybrid CNN–RNN models to improve performance in noisy environments. The results underscore the potential of deep learning for deploying AMC in cognitive radio, signal surveillance, and secure communication systems.