Variational Information Inference: An Interpretable Disentangled Transfer Learning Quality Prediction for Multirate...

Area of Science:

• Industrial Process Control
• Machine Learning
• Data Science

Background:

• Industrial processes often feature variables sampled at different rates due to sensor characteristics.
• Existing soft sensors typically assume uniform sampling, which is unrealistic and can impair production decisions.
• Quality variables are frequently sampled at lower rates than process variables, creating a multirate data challenge.

Purpose of the Study:

• To propose a novel interpretable disentangled transfer learning (IDTL) method for quality prediction in multirate industrial processes.
• To address the limitations of uniform sampling assumptions in current soft sensor models.
• To enhance the accuracy and reliability of quality prediction in complex industrial environments.

Main Methods:

• A Signal Conversion (SC) module was designed to diversify multirate data into multiple sets without information loss.
• A disentangled transfer learning (TL) approach was developed to extract domain-invariant and domain-specific representations.
• Information theory principles were applied to establish theoretical foundations for disentanglement and its link to TL.

Main Results:

• The proposed IDTL method effectively handles multirate industrial data, improving soft sensor performance.
• Theoretical analysis confirmed that IDTL achieves optimal disentangled representations.
• Validation on debutanizer column and polyester esterification datasets demonstrated the method's effectiveness.

Conclusions:

• IDTL provides a robust solution for quality prediction in multirate industrial settings.
• The method enhances understanding of intrinsic industrial process properties.
• IDTL offers improved decision-making capabilities for industrial production.