Adaptive two-dimensional subspace identification for monitoring batch processes with limited batch data.

Area of Science:

• Chemical Engineering
• Process Control
• Data Science

Background:

• Data-driven batch process monitoring is crucial for stable operations and product quality.
• Traditional statistical models require extensive data, which is often impractical for long-duration processes.
• Modeling complex dynamics like nonlinearity and time-varying characteristics in batch processes is challenging.

Purpose of the Study:

• To propose a novel local learning-based two-dimensional subspace identification (LL-2D-SID) scheme for batch process monitoring.
• To address the limitations of conventional models by utilizing limited batch data effectively.
• To accurately model batch-wise and variable-wise dynamics, nonlinearity, and time-varying characteristics.

Main Methods:

• Developed an LL-2D-SID scheme leveraging the similarity between ongoing and previous batches.
• Employed extended extrapolative time-warping to estimate batch similarity.
• Utilized an online optimizing mechanism for model training with limited data.

Main Results:

• The LL-2D-SID scheme demonstrated good prediction performance with limited batch data.
• Successfully applied to monitor the sintering process in polytetrafluoroethylene production.
• Accurately tracked temperature changes and provided timely fault alarms.

Conclusions:

• LL-2D-SID offers an effective approach for data-driven batch process monitoring.
• The proposed scheme outperforms other subspace identification methods in terms of accuracy and alarm rate.
• Enables stable process operation and consistent product quality even with limited data.