A new method which extend the partial least squares (PLS) method to three-dimensional (tri-PLS) and don’t need to care about the eigenvalue is presented to determinate the content of a variety of mineral oil mixture with overlapping spectra. This method is used to study the three-dimensional fluorescence spectra of the mixture of diesel, gasoline and kerosene. The 3D data matrix is constructed on the basis of the sample sequence, the excitation wavelength and the emission wavelength. The combination of 3D data matrix and the concentration matrix can be modeled by tri-PLS to predict the test samples.The experimental results show that the modeling precision of tri-PLS is superior to PARAFAC.