Abstract: In order to establish a non-destructive and rapid method of foundation fluid testing and classification,50 randomly collected foundation fluids with different colors,different origins and different prices on the market were selected in a wavelength range of 400～600nm for NDT in hyperspectral mode.Random points of each sample were selected for 3 times,and a total of 150 sets of data were measured.The spectral data were smoothed by Savitzky-Golay,and the feature values were extracted by principal component analysis,and the top 7 principal components were selected as subsequent experimental data sets;150 groups of data were classified by K-Means clustering method,50 foundation were divided into 5 categories,and the classification effect was verified by Fisher discriminant;finally,BP neural network model(BPNN)was established based on the clustering results,and the classification accuracy reached 88.89％with good results.The research shows that BP neural network combined with hyperspectral technology can be feasible and scientific,and can play an effective role in the inspection of material evidence in public security organs and court science.