As three-dimensional fluorescence spectroscopy has many advantages, such as good selectivity, high sensitivity, fast analysis, it has been widely used in many fields．As one of the characteristics of traditional Chinese medicine(TCM), Chinese herbal medicine property (CHMP) is the core of TCM. Objective discrimination of the properties of TCM is the key issues of modernization of TCM. The identification of traditional Chinese medicine property is of great significance in the theoretical study of Chinese medicine. Most of the molecules in traditional Chinese medicine have the ability to generate fluorescence. According to the characteristics of the three-dimensional fluorescence spectrum of traditional Chinese medicines, the classification and recognition were studied from the perspective of the properties of traditional Chinese medicines. Firstly, the three-dimensional fluorescence spectral data of 5 different concentrations of 23 cold and warm Chinese medicinal solutions were acquired by FS920 fluorescence spectrometer. Then, the ensemble empirical mode decomposition (EEMD) algorithm is applied to denoise the spectrogram, based on the analysis of noise in different excitation and emission wavelength ranges of different samples. Based on the local linear embedding (LLE) algorithm, feature extraction of spectral data is carried out. The extracted eigenvectors are input into the random forest (RF) to construct LLE-RF classification model. The classification effect of LLE-RF classification model on fluorescence spectrum data of cold and warm Chinese medicines was analyzed under different parameters. The sample ratio of the training set and test set in RF classifier is set to 3∶1 and 2∶1. The correct rate of LLE classification is analyzed when the nearest neighbor points k is 7～18 and the eigenvalue dimension d is 6, 7, 8, 9 and 10. When the nearest neighbor points k is 12 and the eigenvalue dimension d is 7, the accuracy of LLE-RF model for classification of Chinese herbal medicines was 96.6%. Finally, the classification effect of SVM classifier constructed with different kernels on fluorescence spectrum data of cold and warm Chinese medicines was compared under the same ratio of r. When multi-layer perceptron is used as the kernel function, the classification effect is the worst. When r=3/4 and radial basis function is used as the kernel function, the classification accuracy is 82.1%. The results show that the method of combining fluorescence spectroscopy with LLE-RF can effectively recognize cold and warm Chinese medicines, and the classification effect is better than LLE-SVM.