Pseudo-Labeling Optimization Based Ensemble Semi-Supervised Soft Sensor in the Process Industry.

Area of Science:

• Process Analytical Technology
• Machine Learning
• Industrial Chemistry

Background:

• Soft sensor techniques enable real-time estimation of industrial process quality variables.
• Limited labeled data in real-world applications hinders the development of accurate soft sensor models.
• Existing methods struggle with data scarcity for effective soft sensor development.

Purpose of the Study:

• To propose a novel semi-supervised soft sensor method, EnSSNCLELM, for industrial processes with limited labeled data.
• To enhance the accuracy and reliability of soft sensor models by leveraging both labeled and unlabeled data.
• To address the challenge of data scarcity in developing robust soft sensor solutions.

Main Methods:

• Developed an improved supervised regression algorithm, Negative Correlation Learning Extreme Learning Machine (NCLELM).
• Proposed a multi-learner pseudo-labeling optimization approach to generate high-confidence pseudo-labeled data.
• Created diverse semi-supervised NCLELM models (SSNCLELM) using combined labeled and pseudo-labeled data.
• Employed a stacking strategy to ensemble SSNCLELM models with superior predictive accuracy.

Main Results:

• The proposed EnSSNCLELM method effectively utilizes both labeled and unlabeled data.
• EnSSNCLELM demonstrates superior prediction performance compared to traditional supervised and semi-supervised soft sensor methods.
• The method's effectiveness was validated through two practical chemical process applications.

Conclusions:

• The EnSSNCLELM method offers a powerful approach for soft sensor development in data-scarce industrial environments.
• Combining semi-supervised learning and ensemble learning paradigms enhances soft sensor model performance.
• The proposed technique provides a significant advancement for real-time quality estimation in industrial processes.