As the result of vibration emission in air, the mechanical noise signal carries affluent information about the working condition of machinery and it can be used in mechanical fault diagnosis. But in practice, the measured sound signal is usually the mixing of condition signal and other uncorrelated signals. And the signal received is usually of very low SNR. Therefore, to obtain the features of original signals, the mixed signals have to be separated and the uncorrelated signals have to be removed by means of the blind source separation technique. The BSS is a class of signal processing method that can recover the original signals according to the observed mixing signals. In application of BSS algorithms, it is generally supposed that the number of sources is known. But unfortunately, this is not the case in application. Then, before applying the BSS method, the singular-value analysis method is introduced to estimate the number of sound sources at first. On the other hand, to avoid the ill-conditioned problem caused by environment noise and/or measuring noise in applying BSS method, the partial singular-value analysis method is employed to select those signals with maximum information entropy from mixed signals. This method significantly reduces the distortion of separated signals. Afterward, the second order blind identification (SOBI) algorithm, one of the BSS methods, which only relies on the second order statistics of measuring signals, is utilized and it is modified, in this paper, especially for purpose of spectra separation. Finally, the spectra separation results obtained from the mixed signals measured in a semi-anechoic chamber demonstrate the availability of the presented method.
Improved Singular Value Decomposition (TopSVD) for Source Number Estimation of Low SNR in Blind Source Separation