Hybrid deep learning based spectrum sensing with Neyman Pearson threshold calibration for cognitive radio networks

The dynamic activity of primary users (PUs) in cognitive radio networks creates temporary spectrum holes that can be opportunistically accessed by secondary users (SUs) without causing harmful interference. Reliable spectrum sensing is therefore essential for identifying available frequency bands, where the probability of detection (Pd) and the probability of false alarm (Pfa) are key performance metrics. However, conventional sensing methods often struggle to achieve a favorable Pd–Pfa trade-off, especially at low signal-to-noise ratio (SNR) conditions. To address this limitation, this work proposes a hybrid CNN-BiLSTM model that captures both local signal patterns and temporal dependencies in noisy environments. Experiments conducted on the RadioML2018.01A dataset show that the proposed approach achieves strong sensing performance compared with benchmark and previously reported models. In particular, the proposed CNN-BiLSTM achieves an area under the ROC curve (AUC) of 0.9836 while maintaining a competitive inference time of 9.16 ms. In addition, the decision threshold is calibrated using the Neyman–Pearson criterion to keep the false alarm probability close to the target value of \(P_{fa}=0.1\) $P_{\mathit{fa}}=0.1$ . These results confirm the effectiveness of the proposed framework under the considered benchmark conditions.