The invention relates to a rotor fault feature extraction method and device and belongs to the technical field of rotation mechanical equipment. The method is characterized by, to begin with, enabling mutually-perpendicular radial vibration and axial vibration signals of a rotor to form a tertiary signal; then, according to space envelope characteristics of the tertiary signal, adaptively decomposing the tertiary signal in the sequence of rotation speeds from high to low into the sum of a series of tertiary signals and a remaining signal; obtaining base frequency information based on rotation mechanical equipment rotation speed, and determining tertiary signals to be fused; and finally, fusing power spectrum of the corresponding tertiary signals in three vibration directions to obtain a vector power spectrum, and according to the vector power spectrum, extracting fault features. The method, by decomposing with the three mutually-perpendicular vibration signals as a whole, ensures same decomposition scale in each direction, fuses radial and axial fault characteristics of the rotor, and comprehensively reflects the radial and axial fault characteristics of the rotor; and the extracted fault features are more comprehensive and clearer.