The relationship among the effective eigenvalue, the frequency and its amplitude in principal component analysis(PCA) is studied. It is found that the number of valid eigenvalues, which are irrelevant size of the amplitude, phase and frequency of the signal, are determined by the number of signal frequency components. The sequence of the eigenvalues in the covariance matrix C characteristic distribution chart depend on the amplitude of the signal frequency. Each characteristic frequency in the signal produces two valid eigenvalues. A new characteristic frequency separation method is proposed, based on the above-mentioned discoveries of PCA theory. This method can be well applied to filter the single frequency and extract the characteristic frequency, which is used in the application to purify the axis orbit of the large rotor vibration test bed. The simulation and experimental results manifest that the proposed algorithm has advantages in feature frequency extraction.