Abstract: Aiming at the difficult problem of poor correlation between auxiliary variables, this paper proposes a soft sensor modeling method based on feature extraction sets and extreme learning machine (MLCRBMs-ELML). First of all, it divides the attribute of input into several classes by clustering algorithm. The classed data set enter to the MLCRBM_s feature extraction to perform synchronous feature extraction. Then, the extracted feature subsets are nonlinearly merged by the Blending layer and obtain the new feature set. Finally, the new feature set enter the ELM model to get the fitting results. The experimental results show that the soft sensor model is superior to the traditional method, and has higher prediction accuracy and generalization performance.