Design of a deep learning-based framework for automatic modulation classification in wireless communication systems using neural networks

In modern communication networks, a set of modulation types can be employed by the transmitter to control both the data rate and bandwidth usage. While the transmitter selects the modulation type adaptively, the receiver may or may not know the modulation type. Thus, the Automatic Modulation Recognition (AMR) mechanism can be used to detect the type of incoming signal modulation, thereby eliminating any potential overhead in the network protocol. This research successfully explored the design, development, and evaluation of a hybrid deep learning model for Automatic Modulation Classification (AMC) using RadioML 2016.10a dataset. Through the fusion of two-dimensional Convolutional Neural Networks (2D CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) layers, the proposed architecture leveraged the strengths of both spatial and temporal pattern recognition to detect modulation types in various Signal-to-Noise Ratio (SNR) environments. The hybrid model was trained. Under the use of early stopping, label smoothing, and adaptive learning rate callbacks, the model had 89% training accuracy and 85% validation accuracy at the highest epoch. The validation loss was as low as 0.75 with very minimal overfitting. The experimental results obtained showed that the hybrid architecture, when compared to standalone model CNN (75%) and standalone LSTM (70%) models, the hybrid model outperformed the two at 85% modulation classification accuracy, The hybrid network also achieved a macro-averaged precision, recall, and F1-scores of 0.85, 0.84, and 0.85 respectively for the 11 classes of modulation at 18dB SNR, demonstrating the benefit of multi-dimensional feature learning. Model classification accuracy increased linearly with SNR to 85% for +18 dB and to 70% accuracy for 0 dB. Precision, recall, and F1-score were all greater than 0.80 for all classes, and confusion matrices indicated dominant diagonal patterns, which indicate good classification performance.