(239e) Outlier Detection for a Recursive Soft Sensor and Its Application to Digester Control

Kraft pulping is the commonly applied chemical pulping process that usually utilizes a Kamyr continuous digester. A direct indicator of the extent of delignification and pulp quality known as Kappa number is the most important variable in pulping process. In the pulp making process it is desired to minimize the variations of Kappa number in the pulp. In our previous work [1-2], we proposed a new dynamic soft sensor approach, termed reduced order PLS model, to predict kappa number and showed that the soft sensor works well on closed loop data. Additionally, to cope with time-varying parameters such as woodchip composition and white liquor concentration; we also investigated different adaptive approaches to recursively update the soft sensor model on-line, and derived an adaptive scheme which showed significant improvement over the static model. In this work, we focus on online outlier detection and its effect on soft sensor prediction performance. Because of the ?messy? composition in a digester, industrial data often exhibit substantial noise and outlier measurements, which have negative impacts on the adaptive model update approach. Therefore, if outlier measurements can be identified and eliminated from updating the model, the online soft sensor prediction performance would be improved. In this work, we developed an automatic outlier detector by making use of studentized residuals, leverage [3-4] and multivariate monitoring statistics such as the Hotellings T2, SPEx and SPEy [5-7] at different stages. In addition, we developed an adaptive approach to update the confidence limits of the monitoring statistics. The proposed outlier detection approach combined with the adaptive soft sensor was tested using data from a digester simulator and a real industrial Kamyr digester. In the simulation case study, the outlier detector can effectively identify abnormal measurements and provide significantly improved prediction performance. In the industrial case study, (data provided by MeadWestvaco Corporation) showed that both prediction performance and process trend tracking were improved significantly compared to the case without outlier detection.

References:

[1] Galicia H.; He Q.P.; Wang J. A Subspace Identification Based Dynamic Soft Sensor Approach for Digester Control, AIChE annual meeting, 2008, Philadelphia, PA [2] Galicia H.; He Q.P.; Wang J. Hodges R.; Krishnagopalan G.; Cullinan H.; A Multi-Model Recursive Subspace Identification Based Dynamic Soft Sensor Approach for Digester Control, AIChE annual meeting, 2009, Nashville, TN [3] Walczak, B. & Massart, D. L. Robust principal components regression as a detection tool for outliers. Chemometrics and Intelligent Laboratory Systems, 1995, 27, 41-54 [4] Martens Harald and Naes Tormod Multivariate calibration. Jhon Wiley and Sons Ltd. 1989 [5] MacGregor, J. F.; Jaeckle, C.; Kiparissides, C. & Koutoudi, M. Process monitoring and diagnosis by multiblock PLS methods AIChE Journal, 1994, 40, 826-838 [6] MacGregor, J. F. & Kourti, T. Statistical process control of multivariate processes Control Engineering Practice, 1995, 3, 403-414 [7] Wang, X.; Kruger, U. & Lennox, B. Recursive partial least squares algorithms for monitoring complex industrial processes Control Engineering Practice, 2003, 11, 613-632 [8] Wisnewski, P., Doyle III F.J. and Kayihan F., ?A fundamental continuous pulp digester model for simulation and control?, AIChE J., Vol. 43(12) p3175-3193, 1997