The invention discloses a rotation machinery fault diagnosis method based on singular spectrum decomposition. The method comprises the following steps: 1) arranging a sensor near a key component of a rotation machinery, and collecting a measuring signal as a source signal; 2) carrying out decomposition on the source signal through singular spectrum decomposition to obtain a plurality of singular spectrum components, of which the instantaneous frequency has physical meaning; 3) according to characteristic energy factor maximum criterion, selecting a decomposed component having rich fault characteristic information as a main singular spectrum component; 4) carrying out Hilbert demodulation on the main singular spectrum component to obtain corresponding envelope spectrums; and 5) observing whether the fault characteristic frequency in the envelope spectrums has obvious peak, thereby realizing accurate judgment of fault type of the rotation machinery. The method is simple and practicable, and can realize more accurate fault diagnosis compared with other methods in the prior art.