Excited respectively by the light with wavelengths of 300, 400, 440 and 380 nm, the fluorescence spectra of synthetic food color ponceau 4R, amaranth, allurea red and industrial dye Sudan Ⅳ have been measured. For each sample, 60 emission wavelength values were selected. The fluorescence intensity corresponding to the selected wavelength was used as the network characteristic parameters, a probabilistic neural network for kind identification was trained and constructed. It was employed to identify the 32 kinds of color solution samples. Because the fluorescence spectra of these colors overlap, the identification rate is low. In order to solve this problem, a derivative fluorescence spectroscopy was introduced. The derivative fluorescence data was used as the network characteristic parameters, a probabilistic neural network was constructed and was employed to identify colors. The identification rate is up to 100%. Based on this, a new method is presented, which combines the derivative fluorescence spectroscopy and probabilistic neural network, and can identify synthetic colors easily, quickly and accurately. This method can provide support for food safety supervision and management.