In order to deal with problems of noisy source mixing and unknown number failure sources in complex sound field of rotating machines, an acoustic fault detection method based on blind deconvolution is proposed. The Cosine measure of spectral envelope is employed as distance measure between independent components (ICAs), impulse signal’s kurtosis value is applied to optimal selecting the ICAs. Combining with frequency-domain sparse component analysis (FD-SCA), the combined failure signals of rolling bearing are extracted perfectly when the number of observations is less than the number of source signal. Acoustical transient impulse signal extraction of combined failure bearing in real-world situation shows the validity of the proposed method．