Therefore, this paper seeks to develop a soft sensor based on the MRF. First, the joint probability distribution between the primary variable and auxiliary variables is obtained by MRF. Then, the Gaussian mixture model (GMM) is used to approximate the joint probability distribution, and the unknown parameters in the GMM are estimated by the expectation maximization (EM) algorithm. Finally, a KPI expression represented by the joint probability distribution is obtained based on the maximization coefficient of determination criterion. The proposed approach will be tested on determining the alumina concentration for a real aluminum electrolysis process.