Abstract

Over the past two decades, communication technologies have advanced significantly, but the growing use of various communication methods has led to a shortage of available spectrum. Cognitive Radio (CR) emerges as a solution to this challenge. A crucial aspect of CR is spectrum sensing, which detects available spectrum gaps. However, current spectrum sensing methods have limitations, including insufficient signal representation, inefficiency, and sensitivity to noise. To address these issues, this study embraces a deep learning approach and introduces an innovative spectrum detection architecture for cognitive radio networks. The method combines deep learning and reinforcement learning, leveraging deep learning for energy and energy correlation feature extraction. Additionally, a Recurrent Neural Network (RNN) module is used to capture time-shifted signal correlation. To enhance feature extraction, Short-Time Fourier Transform (STFT) feature extraction is incorporated. The combined feature vector is processed through a reinforcement learning module. Finally, these features are used to train the deep learning classifier which uses residual blocks for better representation of feature while learning. The highest prediction score is considered as the decision threshold in this work. The outcome of this work is compared with other deep learning methods in terms of <em>P_d</em>, <em>P_f</em>  and sensing error for varied sample size and modulation schemes. The comparative analysis shows the robustness of proposed approach by achieving the <em>P_f</em>  as 0.32 and 0.06 for QAM16 and QPSK modulations.