An ensemble principal component analysis is presented for online multivariable process fault detection on complicated conditions. In reproducing kernel Hilbert space(RKHS) spanned by the mapped samples, groups of basis(approximate) are found. Eigenvectors for principal component analysis problem are limited to the spaces spanned by
approximate basis. The eigenvectors and eigenvalues in different subspace are integrated to make up for the approximation. Hotelling ??2 and squared prediction error are calculated according to integrated eigenvectors and eigenvalues. Experiments on Tennessee Eastman is presented to demonstrate the effectiveness of the ensemble learning.