A Blind Source Separation (BSS) based new method for multi-fault diagnosis of rotors is presented. The statistic variable based decorrelation approach is employed to analyze the signals of the rotors with crack and rub-impact, crack and oil-whirl multi-faults, the typical features of the frequency spectrum of each fault of the rotor are separated out. The separated results show that (1) the BSS technique is helpful to the multi-fault diagnosis of a rotor system. (2) Due to the signals in the same direction at different locations of a rotor is correlative in some extent, if the observed points are not chosen properly, the separated accuracy will be affected greatly. The analysis shows that: for a one shaft rotor, choosing the signals in x, y direction at a same point of a rotor will lead to a better separated result. (3) If want to applying the BSS to practical diagnosis of rotor system, some efforts and works still need to be done.