相関型Just-In-Timeモデリングによるソフトセンサの設計

Abstract

Softsensors are widely used for estimating product quality or other key variables when on-line analyzers are not available. However, their estimation performance deteriorates when the process characteristics change. To cope with the changes in process characteristics and update the model, recursive methods such as recursive PLS and Just-In-Time (JIT) modeling have been developed. However, they do not always function well when process characteristics change abruptly. In the present work, a new method for constructing softsensors based on a JIT modeling technique is proposed. In the proposed method, referred to as correlation-based JIT modeling, the samples used for local modeling are selected on the basis of the correlation among variables. Q statistic is used as an index of the correlation. The proposed modeling procedure is as follows: 1) Divide samples stored in the database into some temporal datasets, 2) Apply Principal Component Analysis (PCA) to the datasets separately, 3) Calculate Q statistic of the query point against each dataset, 4) Select a dataset which provides the smallest Q statistic, and 5) Construct a temporary model from the selected dataset. The proposed method can adapt a model to changes in process characteristics even when operating condition is changed abruptly. It can also cope with process nonlinearity. The usefulness of the proposed method is demonstrated through a case study of a CSTR process whose catalyst deactivation and recovery are considered as changes in process characteristics.