Most of the blind source separation methods were dependent on the fact that the observing sensors are more than or equal to the number of signal source. The algorithm in this paper is aimed to research into the blind source separation of noisy harmonic signals from rotating machine when the underlying system is underdetermined, that is, the cases in which the separation of n sources is made from m mixtures while m is smaller than n. Based on the assumption that the input distribution is sparse and the mixture procedure is instantaneous, a potential-function clustering method for estimating the number of information sources was proposed and a calculation on the attenuation matrix and the delay matrix were made. Signal simulation in experiment shows the applicability and reliability of the method under discussion.