Enhanced Soft Sensor with Qualified Augmented Samples for Quality Prediction of the Polyethylene Process.

Area of Science:

• Chemical Engineering
• Machine Learning
• Process Control

Background:

• Data-driven soft sensors are crucial for industrial polymerization quality measurement.
• Limited labeled data due to costly assays hinders accurate model development.
• Existing methods struggle with data scarcity and quality variations.

Purpose of the Study:

• To propose a novel soft sensor, SWGAN-SVR, for enhanced quality prediction with limited training samples.
• To address the challenge of insufficient labeled data in industrial polymerization processes.
• To improve the accuracy and reliability of soft sensors in quality measurement.

Main Methods:

• Utilizing Wasserstein generative adversarial network with gradient penalty (WGAN-GP) to generate virtual data from limited samples.
• Developing a data-selection strategy combining centroid metric and statistical criteria to filter generated data.
• Constructing a support vector regression (SVR) model using selected augmented data for quality prediction.

Main Results:

• The proposed SWGAN-SVR method effectively enhances quality prediction accuracy despite limited training data.
• The data-selection strategy successfully mitigates issues from varied-quality generated samples.
• Demonstrated superior performance compared to traditional methods in numerical and industrial polyethylene process examples.

Conclusions:

• SWGAN-SVR offers a robust solution for accurate quality prediction in data-scarce industrial polymerization.
• The combined approach of generative adversarial networks and data selection is effective for soft sensor development.
• This methodology provides a pathway for more efficient and cost-effective process quality monitoring.