Deep Learning-Based Spectrum Sensing with Fused Time-Frequency Representations

In this letter, we design a spectrum sensing model based on multi-dimensional data. By leveraging the complementarity of spectrum-based multi-dimensional data, the separability of feature samples can be improved. We propose a Fusion Feature Spectrum Sensing (FFSS) method based deep learning. The model integrates a simplified Variational Autoencoder (VAE) for extracting time-domain features and a Residual Networks (ResNet) for extracting time-frequency features. By leveraging multi-dimensional data, it enhances feature separability and ultimately employs Support Vector Machine (SVM) for classification. Simulation results show that, compared to existing spectrum sensing models based on a single feature dimension, the FFSS model achieves better detection performance, especially in low signal-to-noise ratio (SNR) conditions.