In the singular value decomposition (SVD) of signal, non-zero singular values are the important feature parameters of signal, and the variation of non-zero singular values has the important influence on the feature extraction results. The variation characteristic of non-zero singular values is studied, and it is found that with the increase of the matrix dimension, the non-zero singular values of the sinusoidal signal do not rise monotonically, and there is periodic fluctuation in the rising process of the singular values. Furthermore, the fluctuation of non-zero singular values is closely related to the frequency of the original signal, and the greater the frequency of the original signal, the faster the fluctuation speed of the non-zero singular values will be. The fluctuation of non-zero singular values is analyzed theoretically, and it is proved that the fluctuation speed of non-zero singular values is determined by the frequency of sinusoidal signal. In addition, the fundamental frequency of fluctuation is twice as much as that of the original signal frequency, and this characteristic is verified by the numerical simulation examples.