The mechanical noise sources identification without source signal inputs was mainly studied in this paper with the theory of blind signal processing (BSP). In traditional noise sources identification methods, the preknowledge of noise source input signals and transmission paths was required in advance. In order to overcome this shortage, a blind sources separation/deconvolution model of mechanical noise sources identification was suggested, based on the analysis of the characteristics of vibration and acoustic signals’ production, transmission and mixing. And a natural gradient method of convolutive blind separation (CBS) was carried out based on minimal mutual information (MMI). Accordingly the validity of this method was confirmed by tank experiment.