Abstract: Nuclear main pump warning method using single parametric threshold has the problem of repeatedly crossing the warning line in practice. Using multi characteristic parameters auto clustering and identification, nuclear main pump status warning method based on DPMM can deal with this problem. From normally vibrational data, warning threshold was studied. From real-time vibrational data, real-time warning index was calculated. Status warning was realized when real-time warning index beyond the warning threshold. Because of few fault cases and sparse alarm data, a method using correlation coefficient was proposed to locate abnormally vibrational data. Much fluctuant vibration data were obtained to develop and verify the warning method. The warning method has the accuracy of 97% and brings forward the warning time up to 4 h.