Robust spectrum sensing for unknown heteroscedastic noise via covariance-based convolutional neural network

This paper addresses the problem of spectrum sensing using multi-antenna cognitive receivers in unknown heteroscedastic noise environment, where the noise variances may vary in space and time. Specifically, we propose a robust data-driven spectrum sensing approach using a covariance-based deep convolutional neural network (CNN). In particular, we take the sample covariance matrix (SCM) with its unknown noise variances being well suppressed as the input of CNN to train a robust and generalized test statistic against the heteroscedastic noise. Meanwhile, we design a CNN architecture with a strided convolution layer to retain detailed feature information of the noise-suppressed SCM and a batch normalization layer to accelerate the CNN training. Various simulation results demonstrate that the proposed method attains an accurate detection performance and adapts well to different types of heteroscedastic noise. Particularly, the proposed approach achieves detection probabilities exceeding 99% and 95% under worst noise power ratios of 5 and 80, respectively, when the signal-to-noise ratio is $-18$ dB with a false alarm probability of 10%.