Adaptive multivariate regression modeling based on model performance assessment

Multivariate regression techniques have been successfully used in the modeling of chemical processes for the purpose of monitoring and diagnosis. However, when an operation mode changes, the regression model should be updated to reflect the process changes. The root mean squared error of prediction (RMSEP) based batch-type remodeling and recursive regression modeling techniques have been widely employed for updating the model. However, these methods induce a heavy load on the part of the model manager or computational machine. Also, since it is difficult to discriminate between disturbances and process shifts in real-time, the models can sometimes be updated with disqualified data resulting from disturbances. In this paper, we propose a novel adaptive modeling method based on model performance assessment. The key idea of the proposed method is for it to be invoked only when significant model degradation is detected. Robust multiple cumulative-sum SPC charts are used to monitor model degradation in a moving window manner. Dependent on the assessed model performance results, the adaptive remodeling algorithm utilizes either a partial or a complete adaptation style. The proposed method with partial least squares (PLS) was used to predict NOX emissions from an industrial fired heater. The performance assessment method provided a more robust update than the RMSEP based approach. The proposed adaptive modeling method based on the above criteria showed better updating performance and a lower updating frequency, as compared to the block-wise recursive PLS modeling technique.