An Efficient Hybrid Deep Learning Model Based on Bi-LSTM and CNN for Automatic Modulation Classification

Automatic modulation classification (AMC) is vital in cognitive radio, spectrum management, and military applications. The rapid evolution of wireless networks demands reliable and effective techniques. Traditional AMC methods, like likelihood-based and feature-based algorithms, face limits in computational complexity and generalization across different channel conditions. This paper combines Bidirectional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Networks (CNN) to form a novel hybrid deep learning model. The model exploits each branch’s strengths for temporal and spatial feature extraction. The Bi-LSTM branch identifies long-term temporal relationships, such as phase transitions, in signal sequences. In contrast, the CNN branch captures local spatial patterns like constellation structures. The proposed model outperforms baseline models—CLDNN, CNN-GRU, ResNet-LSTM, and VTCNN2—while maintaining lower computational complexity (410,762 parameters) on the RML2016.10b dataset. It achieves a classification accuracy of 92.32% at 18 dB SNR and performs robustly across low-to-high SNR regimes. These results reveal its potential for deployment in real-time wireless scenarios that are both noisy and dynamic.