Bayesian-Based Causal Structure Inference With a Domain Knowledge Prior for Stable and Interpretable Soft Sensing.

Area of Science:

• Industrial Process Control
• Machine Learning
• Causal Inference

Background:

• Industrial processes require stable and interpretable soft sensors for high-stakes operations.
• Existing causality-inspired methods lack temporal modeling and domain knowledge integration, limiting real-world application.
• Soft sensors are crucial for real-time quality variable estimation in industrial settings.

Purpose of the Study:

• To propose a novel causality-inspired Stable-LSTM for enhanced soft sensor stability and interpretability.
• To integrate domain knowledge and causal structure inference for improved soft sensing performance.
• To address limitations of existing methods by incorporating temporal modeling.

Main Methods:

• Leveraging Long Short-Term Memory (LSTM) for temporal feature extraction.
• Employing Bayesian-based causal structure inference with variational inference.
• Incorporating domain knowledge as a prior for causal structure.
• Utilizing a global sample reweighting strategy to remove spurious correlations.

Main Results:

• The Stable-LSTM achieved the highest soft sensing accuracy, especially under distribution shift.
• The inferred causal structure demonstrated strong consistency with domain knowledge.
• The method effectively revealed causal effects between hidden features and quality variables.

Conclusions:

• The proposed Stable-LSTM significantly enhances the stability and physical interpretability of industrial soft sensors.
• Integrating domain knowledge and causal inference is key to overcoming limitations of traditional soft sensing methods.
• The approach offers a promising solution for reliable and understandable industrial process monitoring.