CALIBRATION METHOD OF SOFT SENSOR BASED ON BAYESIAN GAUSSIAN PROCESS REGRESSION

Soft sensor model is basically an approximation of the actual objective process model. As the process model is almost time-variant, soft sensor shall be calibrated regularly such that it keeps pace with the changes of the process. However, the sampling interval of the hard-to-measure variable is usually far longer than that of the easy-to-measure variables. Consequently soft sensor cannot make calibration timely due to the lack of estimation error, and the performance of soft sensor inevitably degrades. To solve this problem, we proposed a soft sensor calibration method based on the Bayesian Gaussian process regression (GPR). When soft sensor deteriorates, the target variable is estimated by GPR-based interpolation over the sparse history data. Then we obtain a missing data zone of the target variable. By selecting several datasets from the data zone, we can train several candidate models based on the soft sensor model. The soft sensor is finally calibrated by weighted combination of the trained candidate models. The feasibility and effectiveness of the proposed calibration method is verified by experiments on a pH neutralization facility and comparative simulation experiments on a continuous stirred tank reactor with a Kalman filter based calibration method.