Machine learning enhanced grey box soft sensor for melt viscosity prediction in polymer extrusion processes

Area of Science:

• Polymer Science and Engineering
• Process Control and Automation
• Machine Learning Applications

Background:

• Melt viscosity is a critical quality indicator in polymer extrusion.
• Existing rheometers face limitations in real-time monitoring due to flow disturbances and measurement delays.
• Soft sensors offer a viable alternative for monitoring difficult-to-measure physical parameters.

Purpose of the Study:

• To develop a real-time soft sensor for predicting melt viscosity in polymer extrusion.
• To combine physics-based knowledge with machine learning for enhanced prediction accuracy.
• To overcome the limitations of conventional rheometers in dynamic process conditions.

Main Methods:

• A grey-box soft sensing approach was implemented, integrating a physics-based mathematical model with a deep neural network.
• The physics-based model provides initial melt viscosity predictions.
• A deep neural network compensates for the prediction errors of the physics-based model.

Main Results:

• The proposed soft sensor achieved a normalized root mean square error of 2.2 x 10^-3 (0.22%).
• It outperformed fully data-driven models (MLP, LSTM) and a previous RBFNN-based soft sensor by approximately 95%.
• The soft sensor effectively monitors viscosity changes due to operating condition variations.

Conclusions:

• The developed grey-box soft sensor provides accurate, real-time melt viscosity prediction for polymer extrusion.
• This approach enhances process monitoring and control capabilities.
• The system is effective for detecting viscosity changes from operational shifts but not material property variations.