Abstract—In the context of energy-based spectrum sensing, effective noise reduction is paramount to enhance the Signalto- Noise Ratio (SNR) and improve the probability of detection, particularly in cognitive radio networks. This paper presents a novel approach that synergistically combines Fractional Bernstein Approximation, Fast Fourier Transform (FFT), and Convolutional Neural Networks (CNN) to robustly detect signals under noisy conditions. The Fractional Bernstein Approximation is employed as a preprocessing step to smooth the signal and mitigate noise effects. FFT is then utilized to transform the signal into the frequency domain, where CNN is applied to extract features that differentiate between signal and noise. Performance metrics such as SNR improvement, probability of detection, and computational efficiency are analysed under different noise scenarios. Our results demonstrate that the proposed method significantly outperforms other approaches, particularly in low SNR environments, offering a robust and scalable solution for spectrum sensing. The suggested approach is compared in this work against a number of other methods, including CNN with windowing segmentation for Quadrature Phase Shift Keying (QPSK) and 8PSK modulations and cyclisation features. The combination of CNN features and QPSK signals with the Bernstein polynomial approximation maintains a higher Pd (~96%) at lower SNRs (−10 dB) than cyclisation features.