The present invention belongs to the field of on-line monitoring and fault diagnosis of a transformation device in a power system, and relates to a new failure mode based method for calculating a failure probability of the transformation device. The method in particular comprises: according to an established device fault tree, calculating an occurrence probability Pi(i=1,2...n) of various failure modes, severity Ii(i=1,2...n) of various failure modes, and criticality of various failure modes Wi(i=1,2...n) on the device. According to the present invention, when the various failure modes are mutually independent, a device failure rate lambda caused by the various failure modes can be computed according to a series model of a failure reliability theory, i.e., (reference to the specification), wherein computing of the defect occurrence probability Pi of various failure modes of transformation device is based on a defect probability computing method of an online monitoring data statistic probability distribution regulation; computing of the severity Ii(i=1,2...n) of defect corresponds to various failure modes is obtained according to the development process of a device defect, and developmental and changing rules of a characteristic parameter; and the criticality of various failure modes Wi is obtained by using a fuzzy hierarchical analysis method.