Performance-Driven Distillation and Confident Pseudo Labeling for Semi-Supervised Industrial Soft-Sensor Application.

Area of Science:

• Industrial Process Monitoring
• Machine Learning Applications
• Data Science

Background:

• Labeled industrial data for soft-sensor development is often limited, hindering accurate process monitoring.
• Existing semi-supervised methods face challenges with feature effectiveness and pseudo-label confidence evaluation.

Purpose of the Study:

• To propose a novel semi-supervised soft-sensor framework for industrial quality variable prediction.
• To address data scarcity and improve the generalization of soft-sensor models.

Main Methods:

• Developed a performance-driven distillation strategy using a siameseLSTM structure for teacher models.
• Introduced a pseudo-label confidence evaluation strategy to select high-quality training samples.
• Integrated these strategies into a unified semi-supervised soft-sensor framework.

Main Results:

• The proposed framework demonstrated improved feature representation learning and model generalization.
• Validated on alumina production datasets, showing significant reductions in RMSE and MAE.
• Achieved an average RMSE reduction of 10.76% and MAE reduction of 11.18%.

Conclusions:

• The novel semi-supervised framework effectively enhances soft-sensor performance in industrial applications with limited labeled data.
• The combined distillation and pseudo-labeling strategies offer a robust solution for accurate industrial quality prediction.