Abstract:The underdetermined blind source separation(UBSS) algorithm at present can separate sparse signals well, but can’t separate non-sparse signals successfully. Classical ICA algorithms such as extended Infomax can separate both super-Gaussian and sub-Gaussian signals, but it is only used in the over-determined BSS(OBSS). Combining extended Infomax, an underdetermined ICA algorithm is proposed in this text. By generating hidden data, the UBSS problem is transformed into OBSS problem, and then extended Informax algorithm is used to analyze the signal. This method can separated both super-Gaussian and sub-Gaussian signals in the UBSS problem. Through the analysis of transient signal on gearbox by use of underdetermined ICA combined with order envelope spectrum analysis, the fault features are fully detected and its effectiveness is verified.