Abstract: To solve the problem of selecting effective eigenvalues，the number law of effective eigenvalues was de-duced theoretically． That is，one frequency corresponds to two eigenvalues． The order rule of eigenvalues was de-duced as well． That is，the larger the amplitude of signal is，the larger the corresponding two eigenvalues are．The above two properties were collectively referred as amplitude filter characteristics of principal component analysis (PCA-AF)，and a novel signal separation algorithm based on this characteristic was proposed． Through the analy-sis of simulation signal and the actual rotor signal，the effectiveness of the algorithm for signal separation was veri-fied． Ｒesearch results show that the algorithm has excellent advantage in both extracting multiple and single fre-quency components，and the purified signal does not contain redundant components，nor does phase deviation oc-cur． Finally，the proposed algorithm was applied to purify rotor axis orbit of large sliding bearing test bed，and the misalignment fault of the rotor was identified successfully．

Key words: principal component analysis (PCA), effective eigenvalue, amplitude filter, feature extraction, fault diagnosis