Development of a Cyclostationary Spectrum Sensing Framework Using Deep Learning Techniques for Cognitive Radio Networks in mmWave Band

Spectrum sensing is a fundamental component of Dynamic Spectrum Access (DSA) and Cognitive Radio (CR) technology. Accurate and rapid detection of signal presence is essential for the dynamic allocation of spectrum resources to specific devices. This capability is especially important in millimeter-Wave (mmWave) and wideband communication systems. These systems experience significant frequency-selective fading, severe multipath effects, and rapid signal variations over their wide bandwidths; therefore, reliable spectrum sensing methods are required. This paper proposes a spectrum sensing framework based on wavelet-derived cyclostationary feature extraction and a hybrid Convolutional Neural Network–Graph Neural Network (CNN-GNN) model. The extracted wavelet-based cyclostationary features are used to train a Multi-Layer Perceptron (MLP) for nonlinear feature learning. A CNN is trained to learn a time-frequency representation of the signals, whereas the GNN is used to learn both spatial and structural properties of the signals. The proposed model is compared with three different benchmark methods: energy detection, Cyclostationary Feature Detection (CFD) based on the Spectral Correlation Function (SCF), and a CNN-MLP baseline, across multiple evaluation metrics, including the probability of detection ( ), the probability of false alarm ( ), Area Under the Curve (AUC), accuracy, and runtime. Simulation results show that the proposed model achieves a above 0.95, an accuracy above 0.95, a below 0.05, and lower runtime than the other methods, demonstrating its effectiveness in supporting DSA and Software-Defined Access (SDA) in mmWave systems.