Application of multivariate analysis techniques in understanding complex industrial processes: a pulp mill example
The financial and process benefits of improving the mill fiber line are widely acknowledged. However, process optimization of the fiber line is difficult due to the complex behavior of pulp and paper systems. The dissertation project focused on application of multivariate analysis techniques for understanding and improving fiber line performance. Models for prediction and variability analysis of kappa number and total bleaching cost were developed using data generated by pulping and bleaching operations. The research project led to refinement of earlier methods of data preprocessing and development of algorithmic solution for the data time shifting problem.Multivariate statistical techniques were used to analyze sources of kappa and bleaching cost variability for Weyerhaeuser Longview mill and Georgia Pacific Ashdown mill. For the Weyerhaeuser Longview mill, factor analysis allowed development of models that successfully predict kappa number out of a continuous digester and O2 delignification stage. The most important cause of kappa variability in the continuous digester was found to be mischarges in alkali. Variations in kappa number can be reduced by 45% in the digester and 40% in the O2 delignification reactor if variables correlating with the important factors are brought under control.None of the multivariate techniques were successful in predicting K-number for the Georgia Pacific Ashdown mill. The main reason for poor prediction was that the digester was already under tight control as evident from low (6.12%) coefficient of variation of K-number. Processes under tight control appear to generate datasets with minimal correlation structures. Such datasets are not suitable candidates for the purposes of predicting output variables such as K-number.In the bleaching study, principal component analysis as well as factor analysis models with fourteen upstream variables successfully predicted bleaching cost trend. However, neural networks bleaching cost predictions were poor. Factor analysis and PCA models of the bleaching cost indicated that most of the bleaching cost variability was either due to lignin factor (which represents pulping and washing variables) or due to digester column stability represented by outlet device amperage. A method to compare results from various multivariate methodologies was also developed. The factor model with fourteen variables achieved the highest score on comparison scale for bleaching cost study.Both the lignin and digester stability factor point at the digester being the major source of bleaching cost variability. It appears that there are variations in pulp lignin content (or some latent variable) that are not measured by the K number test at the Decker, but results in changes in the chlorine requirements at the D/C stage. In this situation, bleaching cost predicted by the model may be used as a soft sensor to manipulate temperature, steam flow in digester to produce pulp with uniform bleach chemical requirements (i.e., consistent latent variable variation). This way cost variability will be reduced, as presumably the variation in lignin content will be minimized.
- Forestry