Causal Graph Spatial-Temporal Autoencoder for Reliable and Interpretable Process Monitoring.

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Causal Inference

Background:

• Traditional industrial process monitoring methods often lack reliability and interpretability.
• Accurate fault detection is crucial for operational efficiency and safety in industrial settings.
• Existing approaches may struggle to capture complex dynamic relationships within process data.

Purpose of the Study:

• To propose a novel Causal Graph Spatial-Temporal Autoencoder (CGSTAE) for enhanced industrial process monitoring.
• To improve the reliability and interpretability of monitoring systems.
• To enable effective fault detection in complex industrial processes.

Main Methods:

• Developed a CGSTAE network combining a spatial self-attention mechanism (SSAM) for correlation graph learning and a graph convolutional long short-term memory (GCLSTM) encoder-decoder.
• Introduced a three-step causal graph structure learning algorithm leveraging causal invariance principles.
• Utilized SSAM to capture dynamic variable relationships and GCLSTM for time-series data reconstruction.
• Employed feature and residual space statistics for monitoring and fault detection.

Main Results:

• The CGSTAE effectively learns correlation and causal graphs from industrial process data.
• The GCLSTM-based encoder-decoder accurately reconstructs time-series data.
• The proposed method demonstrated successful process monitoring and fault detection capabilities.
• Validation was performed on the Tennessee Eastman Process (TEP) and an air separation process (ASP).

Conclusions:

• The CGSTAE provides a reliable and interpretable framework for industrial process monitoring.
• The integration of causal graph learning and spatial-temporal modeling enhances fault detection accuracy.
• The proposed approach shows significant potential for real-world industrial applications.