Blind Source Separation for Parameter Estimation Under Mixed Gaussian-Impulsive Noise: An U-net plus plus Based Method

In many practical applications, the communication system suffers from mixed Gaussian-impulsive noise consisting of both Gaussian white noise and non-Gaussian impulsive noise. Obtaining the optimal signal detection algorithm under these scenarios requires the estimate of the mixed Gaussian-impulsive noise parameters. Unfortunately, the estimation accuracy will deteriorate in the presence of the transmitted signal. To solve the problem, we propose a blind source separation method with a neural network, namely U-net++, to separate the transmitted signal from the mixed noise. Then, the parameter estimation can be implemented with the recovered noise samples. An adaptive clipping preprocessing module and a novel loss function of the network are designed based on the statistical property of the mixed noise. Results show that our algorithm outperforms existing baselines on both source separation and parameter estimation.