Iterative Reweighted DOA Estimation for Impulsive Noise Processing Based on Off-Grid Variational Bayesian Learning

Abstract

The performance of acoustic source localization with array system is limited by impulsive noise such as electromagnetic interference, car ignitions, bursting, and so on. The impulsive noise decays with heavy-tailed distribution which can be considered as outliers. In order to alleviate the performance degradation of traditional direction-of-arrival (DOA) estimation with impulsive noise, a novel iterative reweighted variational Bayesian learning algorithm based on off-grid model (OG-WVBL) is proposed. OG-WVBL employs impulsive noise as two independent components and models directly the outliers with sparse distribution in the time domain. OG-WVBL utilizes two iterative VBL to reconstruct signal and outliers matrix and then retrieves the DOA. Then, OG-WVBL also introduces the iteratively reweighted strategy to hyperparameters so that the more importance is given to those hyperparameters with non-zero entries over others which can encourage sparsity and achieve the consistent convergence. With the iteratively reweighted strategy, OG-WVBL can automatically identify the number of sources without any prior knowledge. Moreover, the proposed algorithm can use a coarse sampling grid to achieve the accurate DOA estimation with the off-grid model. The experiments and simulation results show that OG-WVBL possesses robust performance and outperforms several existing algorithms under impulsive noise environment.