The roasting process of shaft furnace is one of the most important processes in the Minerals Processing Factory,the primary task of it is to provide the roasting ore of hematite with higher magnetism by high temperature deoxidization,in order to fit request of the integrated production indexes.Traditional math models cant describe the dynamics of many mineral industry processes due to their integrated complexities,and complex association of much technical equipment is in production process.This paper establishes a soft sensor model for predicting the reduction section temperature of shaft furnace roasting,and uses CS(Compressed Sensing)theory for data sampling,combining PCA(principal component analysis),case-based reasoning and RBF.The proposed approach can keep the stability of temperature away from some harmful effects,such as big time delay,nonlinearity,etc.Even it can avoid the coming faults during roasting.With production data of the hematite ore beneficiation process of a plant,the simulation experiments prove that the CS is effectiveness for low rate sampling,and industrial results prove the soft sensor model effectively.