The invention discloses a rotary machine fault characteristic extraction method based on order-holospectrum principle, wherein the ability of order analysis for processing non-stationary signal and the ability of holospectrum for comprehensively and reliably reflecting machine set status are utilized. The ability of order analysis and the ability of holospectrum are fused, so that fault diagnosis an be carried out by utilizing two-dimensional holospectrum when speed oscillation, and force and force couple decomposition and holographic waterfall chart analysis can be carried out conveniently on cycling start and stopping signals. According to the method, the application scope of the traditional holospectrum is greatly expanded, various characteristic parameters can be extracted for the cycling start and stopping signals, the traditional limitation of only carrying out Bode diagram analysis is broken through, the monitor and diagnosis level and range of rotary machine faults are widened, and powerful support is provided for rotary machines healthily and reliably.