Aiming at the problem that the spectral data of prediction model are not fully utilized in near infrared spectroscopy, a new method which can make full use of spectral data and effectively predict the proportion of silk content is proposed. The proportion of silk content in 145 samples of 5 types and the corresponding spectral data of all protein bases are taken as the research objects. The samples are divided into calibration and verification sets respectively. The partial least squares regression method and the partial least squares regression multi-model method are respectively used to establish the prediction model, and the prediction effects of the two methods are compared and observed. Taking the silk samples of type 2 as an example, 13 principal components are selected and the two models are compared. It is found that the correlation coefficient of multi-PLSR increases from 0.594 to 0.9784, and the average relative error decreases from 0.4866 to 0.1384.The experimental results show that the new method makes full use of the information in the spectral data and improves the accuracy of the prediction model of silk content proportion, which provides a new thought for the establishment of the near infrared spectrum prediction model.