Quality-Driven Regularization for Deep Learning Networks and Its Application to Industrial Soft Sensors

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Chemical Engineering

Background:

• Industrial processes generate vast data, necessitating advanced data-driven soft sensors.
• Effective feature representation is crucial for soft sensor accuracy and reliability.
• Traditional deep networks may lose quality-relevant information during unsupervised pretraining.

Purpose of the Study:

• To develop a novel deep learning approach for enhanced feature representation in soft sensors.
• To introduce a quality-driven regularization (QR) technique for deep networks.
• To improve the prediction accuracy of industrial process quality variables.

Main Methods:

• A quality-driven regularization (QR) technique was developed for deep networks.
• A QR-based stacked auto-encoder (QR-SAE) was designed, modifying the loss function to control input variable weights.
• The QR-SAE model was applied to predict quality in a real industrial hydrocracking process.

Main Results:

• The QR-SAE model successfully learned quality-related features from industrial process data.
• The proposed method demonstrated accurate prediction performance for industrial process quality.
• Comparative experiments validated the effectiveness of QR-SAE over traditional methods.

Conclusions:

• The developed QR-SAE effectively extracts quality-relevant features, improving soft sensor performance.
• This approach enhances the reliability and accuracy of predicting industrial process quality.
• Quality-driven regularization offers a promising direction for advanced soft sensor development.