Accurate, real-time access to key performance indicators (KPIs) is critical to the overall performance of industrial processes. However, in many cases, it is difficult to obtain accurate and timely measurements, due to time delays or external disturbances in industrial processes. Soft sensors are one solution that can provide the necessary process information. This paper proposes a new approach for soft sensor design using Markov random fields (MRF). In which, a Gaussian mixture model (GMM) is firstly used to approximate the joint probability distribution in the soft sensor model, then the expectation maximization (EM) algorithm estimates the GMM parameters. Using this approach, a soft sensor is developed using industrial data for the alumina concentration process in the aluminum electrolysis industry, to show our proposed approach provides accurate estimation of the alumina concentration.