20 samples of Carmine solution with different mass concentration were prepared and the emission spectra were measured by FLS920P fluorescence spectrometer. Experimental results showed that the optimum excitation wavelength and emission wavelength of Carmine solution were 300 and 440 nm respectively. The spectral data of ultrapure water were measured under the same condition and it was selected as the reference spectrum. The two-dimensional fluorescence correlation spectra were calculated under the perturbation of concentration. Sym8 wavelet based on the four-scale was selected for denoising. Partial least squares regression (PLSR) predictive models were built by using the synchronous correlation spectral data and auto correlation spectral data after the noise reduction. When Partial least square regression model was used combined with synchronous correlation spectral data for predicting the carmine contents in prediction set, the root mean square errors of prediction (RMSEP) was 0.414 μg·mL-1 and the coefficient correlation of actual values and predicted values was 99.863%. However, the model based on the Partial least squares regression model and auto correlation spectral data was better. The coefficient correlation of prediction set reached to 99.863%, and the root mean square errors of prediction (RMSEP) was 0.303 μg·mL-1. As can be seen from the results, the data of the autocorrelation spectra can effectively avoid information redundancy, and the effect is more reliable. The method simple operation does not rely on sample separation, , and can provide a new way of thinking for food safety testing.