Abstract: In this paper,we choose the infrared spectral data of wheat flour as subject for the study,partial least squares dimension reduction technique is used to set analysis model for predicting the nature or the composition of unknown samples. NIR spectra of the district (700-2500nm) is mainly composed by the hydrogen-containing group and the frequency group of absorption peak,the absorption intensity is weak and sensitivity degree is relatively low, absorption range is wide and overlapping seriously,considering it belongs to the weak spectral signal analysis technology, information obtained is influenced by many factors, and as a source of information in the near-infrared spectroscopy there’s low rate of effective information, so we need effective ways to eliminate noise and other impacts or to reduce the dimension of spectral data for the establishment of calibration model and predict the nature or the composition of unknown samples. Currently used methods for modeling is stepwise regression, principal component regression and partial least-squares regression methods.This experiment use partial least squares (PLS)technology,and establish the multivariate calibration model for NIR quantitative analysis, and then use the model to predict the sample data. At the same time,implement modeling and forecasting use principal component regression method,in order to verify the merits of two different methods.