Enhancing interpretable soft sensing with embedded hybrid modeling: the GraphTrans approach for industrial processes.

Area of Science:

• Chemical Engineering
• Data Science
• Process Systems Engineering

Background:

• Accurate process models are crucial for optimization and control.
• Existing models often lack a balance between accuracy and interpretability, limiting industrial use.
• There is a need for advanced modeling techniques that offer both precision and understanding.

Purpose of the Study:

• To propose an interpretable embedded hybrid modeling framework to address the limitations of current process models.
• To introduce GraphTrans, a novel data-driven architecture for enhanced process modeling.
• To enable accurate and interpretable soft sensing of key process indicators.

Main Methods:

• Developed GraphTrans, integrating graph convolutional networks, graph-masked multi-head attention, and a kernel projection module.
• Embedded GraphTrans into a mechanism model to generate dynamic mechanism parameters for soft sensing.
• Utilized simulation experiments on zinc purification reactor datasets.

Main Results:

• GraphTrans-EMM reduced mean absolute error by 35.90% and improved R-squared by 31.43% compared to the best mechanism modeling strategy.
• Demonstrated superior predictive accuracy over deep learning baselines while maintaining interpretability.
• Identified mechanism parameters remained physically meaningful, offering insights into reaction states and atypical samples.

Conclusions:

• The proposed GraphTrans-EMM framework achieves a superior balance of accuracy and interpretability in process modeling.
• The framework enables accurate soft sensing and provides physically meaningful, interpretable outputs.
• GraphTrans-EMM shows potential for industrial applications like fault diagnosis through the analysis of abnormal parameter deviations.