Collaborative Multiple Players to Address Label Sparsity in Quality Prediction of Batch Processes.

Area of Science:

• Chemical Engineering
• Data Science
• Process Control

Background:

• Soft sensors are crucial for real-time tracking of process variables in advanced control.
• Label sparsity in soft sensor modeling presents a significant challenge, limiting model performance.
• Existing methods struggle to effectively utilize limited labeled data.

Purpose of the Study:

• To develop an improved soft sensor modeling framework addressing label sparsity.
• To investigate the efficacy of a multi-player co-training approach for soft sensor development.
• To enhance the utilization of unlabeled data in batch process modeling.

Main Methods:

• A novel multi-player co-training technique is proposed, extending the conventional two-player scheme.
• A sliding window approach is employed to capture 2D correlations in batch process data.
• The framework is designed to effectively leverage a small ratio of labeled data.

Main Results:

• The proposed multi-player co-training framework demonstrates superior performance compared to existing methods.
• Effectiveness is particularly pronounced when the proportion of unlabeled data increases.
• Two case studies validate the practical applicability and robustness of the developed framework.

Conclusions:

• The multi-player co-training framework offers a promising solution for overcoming label sparsity in soft sensor modeling.
• This approach enhances the predictive accuracy and efficiency of soft sensors in industrial processes.
• The study highlights the potential of leveraging unlabeled data through advanced machine learning techniques.