The physiological noise such as heart-beating and respiration can cause great interference to the subsequent process of data analysis. Because these noise has considerable autocorrelation characteristics. Combined with the characteristics of physiological noise in time and space, through the method of canonical correlation analysis, the noise components can be steadily identified and removed from the nerve tissues in the residual data after regressing the useful signal components. In this paper, the proposed method does not require any priori information about experiments, realizing the unsupervised reduction of fMRI physiological noise. Through the analysis on the real fMRI data, we illustrate the effectiveness and the reliability of our method.