Improved automatic modulation recognition using deep learning with additive attention
Noureddine El-Haryqy, Anass Kharbouche, Hamza Ouamna, Zhour Madini, Younes Zouine · Results in Engineering · 2025
Automatic Modulation Recognition (AMR) is a critical task in modern communication systems, enabling applications such as cognitive radio, spectrum monitoring, and IoT networks. This paper proposes ICRNNA, a novel deep learning model that integrates Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (BiLSTM) networks, and an attention mechanism to achieve state-of-the-art performance in AMR tasks. The proposed model is evaluated on the RadioML2016.10a and RadioML2016.10b datasets, demonstrating superior accuracy, computational efficiency, and robustness, particularly in low Signal-to-Noise Ratio (SNR) environments. Through extensive ablation studies, we highlight the contributions of each component, showing that the combination of CNNs, BiLSTMs, and attention mechanisms significantly enhances performance. Comparative experiments against state-of-the-art models, including ResNet, MCLDNN, and CNN-BiLSTM-DNN, reveal that ICRNNA achieves the highest accuracy (63.24% on RadioML2016.10a and 65.39% on RadioML2016.10b) and outperforms baseline models in computational efficiency, with only 48.42 MFLOPs and 0.79 million parameters. The results underscore the model's suitability for real-time applications in dynamic and noisy environments. This work advances the field of AMR by providing a robust, efficient, and high-performing solution for modern communication systems. • Our model achieves 63.24% accuracy on RadioML2016.10a and 65.39% on RadioML2016.10b, outperforming advanced models. • It achieves 92.21% accuracy at 0 dB SNR on RadioML2016.10a, demonstrating strong robustness in low-SNR environments. • The model effectively classifies both digital (8PSK, 16-QAM, 64-QAM) and analog (AM-DSB, WBFM) modulations. • With 48.42 MFLOPs and 0.79M parameters, it is efficient for real-time deployment on resource-constrained devices. • Ablation studies show CNNs aid feature extraction, BiLSTMs handle temporal modeling, and attention enhances performance.