Abstract:An improved time-domain blind deconvolution algorithm was proposed, based on genetic algorithm (GA) and higher order statistics (HOS). A newly defined distance measure based on kurtosis was employed to improve the classification accuracy of independent components in the cluster analysis process, and a GA was applied to search for an optimal length of blind deconvolution filters. With the help of these enhancements, this improved algorithm leads to perfect convolutive source separation for acoustic-based machine diagnosis. Both numerical and experimental studies were carried out. The results show that this algorithm can efficiently extract acoustic signals of fault bearings in real-world situations.