Soft Sensing of Silicon Content via Bagging Local Semi-Supervised Models

Area of Science:

• Metallurgical Engineering
• Data Science
• Machine Learning

Background:

• Online measurement of silicon content in industrial blast furnaces presents significant challenges.
• Existing soft sensor models often fail to fully leverage process variables for accurate predictions.

Purpose of the Study:

• To develop an advanced soft sensing framework for online silicon content prediction in industrial blast furnaces.
• To enhance prediction accuracy by integrating unlabeled data through semi-supervised learning.

Main Methods:

• Proposed a novel Bagging Local Semi-Supervised Models (BLSM) framework.
• Integrated bagging strategy, just-in-time learning, and semi-supervised extreme learning machines.
• Employed online semi-supervised learning to utilize unlabeled process data.

Main Results:

• BLSM demonstrated superior prediction performance compared to traditional supervised soft sensors.
• The model effectively extracted valuable information from unlabeled data for improved accuracy.
• Successful application in an industrial blast furnace setting validated the approach.

Conclusions:

• The proposed BLSM framework offers a more effective approach for online silicon content prediction.
• Semi-supervised learning significantly enhances soft sensor capabilities by incorporating unlabeled data.
• BLSM provides a robust and accurate solution for critical industrial process monitoring.