Estimation of predictive accuracy of soft sensor models based on data density

• Our goal is estimation of the prediction accuracy of a soft sensor for new data.
• Relationships between data density and prediction accuracy are formed quantitatively.
• We use chemometrics methods to estimate data density.
• The index to compare the methods estimating prediction accuracy was developed.
• The performance of our methods was confirmed with numerical data and industrial data.

Soft sensors are widely used to predict process variables that are difficult to measure online. By using soft sensors, analyzer faults can be detected when the difference between a measured value and a predicted value is large. However, it is difficult to detect abnormal data and determine the reasons for the abnormality because prediction errors increase not only because of analyzer faults but also because of variations caused by changes in the state of the chemical plants. To separate these factors, we previously applied applicability domains to the soft sensors and proposed construction of the relationships between the distances to soft sensor models (DMs) and the prediction accuracy of the models quantitatively, and estimated the prediction accuracy, i.e. the error bar, for new data online. In this paper, we use k-nearest-neighbor method and a one-class support vector machine (OCSVM) to estimate the data density and use the average of the distances from the k nearest data and the output of an OCSVM as DMs, respectively. The proposed method was applied to both simulation data and real industrial data, and the superiority of the proposed DMs compared with the traditional models was demonstrated by comparison of their results.