Computationally Efficient Wideband Spectrum Sensing through Cumulative Distribution Function and Machine Learning

Abstract

Blind spectrum sensing (BSS) is crucial for identifying unknown signals in scenarios with limited prior knowledge. Traditional methods face challenges with unknown and timevarying signals, especially in the presence of noise interference. This paper addresses these issues by introducing a statistical signal processing framework that extends the use of machine learning (ML) features. Our approach improves BSS by incorporating cumulative distribution functions (CDFs) into unsupervised ML, enabling effective clustering of diverse transmission states without assumptions about specific noise distributions. Additionally, we introduce a temporal decomposition technique using shorter Fast Fourier Transforms (FFTs), enhancing the learning process, reducing system inertia, and minimizing data requirements for retraining under dynamic conditions. We evaluate our method, focusing on various features/approaches for incorporating CDFs into ML, including centroid, linear approximation, and low-order statistics. Simulation results demonstrate robust detection in a standard transmission scenario with a Gaussian pulse amidst additive white Gaussian noise, maintaining a consistently low false alarm rate. These findings highlight our BSS approach's effectiveness and practical potential in handling unknown signals in challenging environments. This research provides valuable insights, laying the groundwork for practical implementation in real-world scenarios.