Wideband Spectrum Sensing Utilizing Cumulative Distribution Function and Machine Learning

Blind spectrum sensing (BSS) is a valuable technique for identifying unknown signals in scenarios where prior knowledge is limited. However, traditional methods encounter difficulties when dealing with unknown and time-varying signals in the presence of noise. This paper addresses these challenges by enhancing machine learning (ML) features through a novel statistical signal processing approach. The proposed BSS approach integrates cumulative distribution functions (CDFs) into an unsupervised ML process, allowing for the effective clustering of distinct transmission states without making assumptions about specific noise distributions. Furthermore, the paper introduces a temporal decomposition technique that utilizes shorter Fast Fourier Transforms (FFTs) to enhance learning, reduce system inertia, and minimize the amount of data required for retraining in changing conditions. Simulation results presented in this paper demonstrate a good detection rate in a generic transmission scenario (i.e., receiving a Gaussian pulse disturbed by additive white Gaussian noise) while maintaining a constant false alarm rate. These findings indicate the efficacy of the proposed BSS approach in handling unknown signals and its potential for practical implementation.