Year: 2025
Type: article
Abstract: Automatic modulation recognition plays a critical role in both civilian and military communication systems. While traditional approaches rely on manual feature extraction with limited accuracy, deep learning methods offer promising alternatives for this pattern recognition task. This paper presents a systematic performance evaluation of classical deep learning models for automatic modulation classification, aiming to establish baseline references for future research. Through comparative experiments using the RadioML2018.01a dataset containing 24 modulation types across SNR levels from -20dB to 20dB, we demonstrate that modulation signals exhibit multidimensional characteristics with temporal dependencies. Our analysis reveals that the proposed Multi-Scale Contextual Attention Network (MCNet) outperforms conventional CNN and ResNet architectures, achieving 82.39% accuracy at high SNR conditions. The network's superior performance stems from its ability to extract multiscale spatiotemporal features through parallel asymmetric convolutions, preserve signal correlations via attention mechanisms, and maintain computational efficiency through optimized layer configurations. These findings provide two key contributions: quantitative benchmarks for model selection in practical implementations, and architectural insights for developing next-generation recognition systems. The study particularly highlights MCNet's robustness in processing high-order QAM/PSK modulations, though challenges remain for low-SNR scenarios.
Author: Zifeng Zhang
Institution: University of Shanghai for Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411041680