Identification and Classification of 16-QAM, 32-QAM, and 64-QAM Modulation in Radio Signals using DenseNet - Convolutional Neural Network Algorithm

This research focuses on the challenge of accurately classifying Quadrature Amplitude Modulation (QAM) signals, particularly in real-time scenarios where noise and interference can significantly affect performance. The primary problem addressed is the inefficiency and resource-intensive nature of traditional methods that convert radio signals into spectrogram images before classification. To solve this, the research employs an advanced artificial intelligence system using Convolutional Neural Networks (CNN), specifically the Densely Connected Convolutional Networks (DenseNet) architecture, to improve the accuracy and efficiency of QAM signal classification. The Densenet model is applied to classify 16-QAM, 32-QAM, and 64-QAM modulation types, offering better accuracy compared to standard CNN models. The process involves dataset preparation, feature extraction, and model training using the Keras framework with TensorFlow backend. After 80 epochs, the DenseNet model achieved a significant reduction in loss, with a best validation loss of $\mathbf{0. 2 1 2 2}$ and stable accuracy reaching $91.60 \%$. At high Signal-to-Noise Ratio (SNR) values of 20 dB and 30 dB, the model achieved perfect accuracy, precision, recall, and F1-Score. However, at low SNR levels ($\mathbf{2 0} \mathbf{d B}$ and $\mathbf{- 1 0} \mathbf{d B}$), the model’s performance declined, with accuracy dropping to $55 \%$ and lower precision and recall. This highlights the ongoing challenge of classifying QAM signals under high noise conditions. Overall, the research demonstrates that CNN, particularly DenseNet, can effectively classify QAM signals, though further improvements are needed for noisy environments.