Small moving window calibration models for soft sensing processes with limited history

Abstract

While many soft-sensing strategies have been developed for monitoring chemical processes, most require extensive historical data to train the soft sensor calibration. Many of these strategies involve very complex models to account for drift and other effects in the process. When a process changes, perhaps because of some upset or a change in process parameters, it is often necessary to rebuild the soft-sensing model, a process that can take considerable time. This paper focuses on use of simple models that can be used to provide soft-sensing with minimal calibration data, and with minimal training of hyperparameters to permit process monitoring while a more complex model is rebuilt. The goal of this study was to explore simple soft sensing strategies for the case where a new process has been put online with limited history. Five soft sensor methodologies with two update conditions were compared on two experimentally-obtained datasets and one simulated dataset. The soft sensors investigated were moving window partial least squares regression (and a recursive variant), moving window random forest regression, the mean moving window of the property y, and a novel random forest partial least squares regression ensemble (RF-PLS), all of which can be used with small sample sizes so that they can be rapidly placed online. It was found that, on two of the datasets studied, small window sizes led to the lowest prediction errors for all of the moving window methods studied. On the majority of datasets studied, the RF-PLS calibration method offered the lowest one-step-ahead prediction errors compared to those of the other methods, and it demonstrated greater predictive stability at larger time delays than moving window PLS alone. Both the random forest and RF-PLS methods most adequately modeled datasets that did not feature purely monotonic increases in property values, but both methods performed more poorly than moving window PLS models on one dataset with purely monotonic property values. Other data dependent findings are presented and discussed.