Convolutive blind source separation in frequency domain with kurtosis maximization by modified conjugate gradient

To efficiently and accurately separate sources from the measured signals and align their permutation, a convolutive blind source separation (BSS) in frequency domain with kurtosis maximization (CBSS_FDKM) by the modified conjugate gradient is proposed. Firstly, Short Time Fourier Transform (STFT) is used to convert the convolutive BSS problem in time domain into an instantaneous BSS in frequency domain. Secondly, CBSS_FDKM is adopted to separate complex signals in each frequency bin, and kurtosis maximization is used as the objective function and optimized by a modified conjugate gradient with iterative step sizes optimized by an efficient algebraic method, and thus the accuracy and convergence speeds are significantly enhanced. Thirdly, the power spectra are obtained from the separated signals and the cosine distances based on power spectra are calculated for permutation alignment. Finally, inverse Short Time Fourier Transform (ISTFT) is adopted to reconstruct the separated signals in time domain. Furthermore, the accuracy and robustness of the proposed method are comparatively studied according to typical numerical studies and experimental studies on a cylindrical test bed. Generally, this paper provides a more accurate and robust BSS for source separation and identification, which can provide significant evidences for vibration & noise monitoring and control. (C) 2019 Published by Elsevier Ltd.