Smart Soft Sensor Design with Hierarchical Sampling Strategy of Ensemble Gaussian Process Regression for Fermentation...

Area of Science:

• Industrial Process Control
• Machine Learning Applications
• Soft Sensor Technology

Background:

• Real-time quality prediction is crucial for optimal process control in Industrial 4.0.
• Traditional ensemble soft sensors struggle with limited labeled data, affecting prediction accuracy.
• Existing methods require significant human resources for data labeling and model development.

Purpose of the Study:

• To develop a novel active learning (AL) framework for smart soft sensor design.
• To overcome the limitations of traditional ensemble models dependent on labeled data.
• To reduce human annotation costs while maintaining high prediction performance.

Main Methods:

• An ensemble Gaussian process regression (GPR) model integrated with an active learning (AL) strategy.
• Hierarchical sampling for iterative selection of informative unlabeled samples.
• Gaussian Mixture Model (GMM) for autonomous operation phase identification and local GPR model construction.
• Bayesian fusion strategy for integrating base predictors.

Main Results:

• The proposed AL framework effectively handles complex nonlinearities and dynamic changes in industrial processes.
• Demonstrated significant reduction in human annotation costs (by at least half).
• Achieved high prediction performance comparable to or exceeding traditional methods.

Conclusions:

• The novel AL framework for ensemble GPR soft sensors offers a reliable and superior solution for industrial quality prediction.
• This approach significantly saves human resources and improves performance in complex industrial settings.
• The method is validated through comparative studies on a penicillin fermentation process.