Hybrid deep learning for robust spectrum sensing over AWGN and fading channels

Accurate and reliable spectrum sensing (SS) is crucial for optimal spectrum usage in cognitive radio (CR) networks under challenging low signal-to-noise ratio (SNR) and fading channel conditions. The Bi-CLNetwork, a hybrid deep learning (DL) architecture, was proposed to enhance the performance of SS in both fading and non-fading environments. The proposed hybrid DL framework integrates bidirectional long short-term memory (Bi-LSTM) layers with two parallel one-dimensional convolutional neural network-LSTM (1D-CNN-LSTM) branches, accompanied by a global average pooling layer to reduce parameter complexity and strengthen generalization. Our approach was assessed using bit error rate (BER) and receiver operating characteristic (ROC) metrics across multiple channel models, such as additive white Gaussian noise (AWGN), Nakagami-m, Rayleigh, and Rician fading channels. The hybrid DL model achieves reductions BER of 36.55%, 40.44%, 42.72%, and 41.77% under AWGN, Nakagami-m, Rayleigh, and Rician channels, respectively, compared to existing DL models. Moreover, it increases the probability of detection by 65.1%, 46.8%, 42.9%, and 101.24% for a false alarm probability of 0.1 under identical channel conditions.