An Adaptive CNN–BiLSTM Framework for Joint Energy and Detection Performance Optimization in 6G Cognitive Radio Spectrum Sensing

Abstract

The transition toward sixth-generation (6G) wireless systems intensifies the pressure on scarce radio spectrum and raises the energy burden of intelligent edge devices. Cognitive radio (CR) addresses spectrum scarcity through dynamic spectrum access, yet spectrum sensing remains costly because it must be repeated frequently and, in deep learning (DL) settings, also incurs inference overhead. This paper presents an adaptive hybrid CNN–BiLSTM framework that jointly improves sensing reliability and energy efficiency. The proposed design combines convolutional feature learning for robust local signal representation, bidirectional recurrent modeling for temporal dependencies, and a confidence-driven early-stopping mechanism that suppresses unnecessary inference when the decision confidence is high. A unified energy model is also introduced to account for both sensing and DL inference energy. Simulation results reported in the source study show that the proposed framework achieves ROC AUC values in the range 0.880–0.885, exceeding conventional energy detection (0.846) while reducing the average sensing energy by 17.8% and reaching up to 34% energy saving in the moderate-to-high SNR regime. Under AWGN and Rayleigh fading channels, the adaptive mechanism preserves detection accuracy while lowering redundant computation, indicating that accuracy and energy efficiency can be optimized jointly for practical 6G spectrum sensing.