A Soft Sensor Modeling Method Based on EGMM Using Gaussian Process Regression

In this paper, we propose a Gaussian process regression (GPR)-based error-Gaussian-mixture-model (EGMM) soft sensor. First, we select appropriate variables to establish the error data and determine the optimal number of Gaussian components using a Bayesian information criterion. Next, we construct the EGMM based on the suitable error data to obtain the mathematical expressions for the conditional error mean and conditional error variance. When a new sample is available, the constructed GPR model can be used for output prediction. Then the conditional error mean of the new sample is computed using the EGMM model to compensate the prediction output in order to achieve a more accurate prediction. We performed a numerical simulation and soft sensor prediction of the H₂S concentrations of a sulfur recovery unit (SRU), and the results demonstrate the feasibility and effectiveness of the proposed approach.