Evolving chemometric models for predicting dynamic process parameters in viscose production.

Area of Science:

• Chemometrics
• Process Analytical Technology (PAT)
• Materials Science

Background:

• Accurate monitoring of H(2)SO(4), Na(2)SO(4), and ZnSO(4) concentrations is crucial for high-quality viscose production.
• Conventional chemometric models fail to adapt to the dynamic process variations common in on-line fiber manufacturing.
• Fixed models lead to imprecise and unreliable predictions when applied to new on-line data.

Purpose of the Study:

• To demonstrate evolving chemometric models capable of automatic adaptation to varying process dynamics.
• To improve the prediction accuracy of key process parameters in real-time viscose production.
• To overcome the limitations of traditional, static chemometric approaches.

Main Methods:

• Utilized Takagi-Sugeno fuzzy model architecture for flexible non-linearity modeling.
• Implemented single-pass incremental learning for updating model structures and parameters.
• Employed distance-based and similarity criteria for evolving/merging local linear predictors and gradual forgetting mechanisms.

Main Results:

• Achieved high correlations (0.95-0.98) between observed and predicted target values over a 3-month period.
• Maintained relative error below 3%, significantly outperforming state-of-the-art models.
• Demonstrated superior performance compared to off-line techniques, which showed correlations below 0.5 and higher error rates.

Conclusions:

• Evolving chemometric models offer a robust solution for real-time monitoring in dynamic industrial processes like viscose production.
• The adaptive nature of these models ensures sustained accuracy and reliability, crucial for quality assurance.
• This approach significantly enhances process control and product quality compared to conventional methods.