Physical Graph-Based Spatiotemporal Fusion Approach for Process Fault Diagnosis.

Area of Science:

• Chemical Engineering
• Data Science
• Artificial Intelligence

Background:

• Big data and machine learning are crucial for complex chemical process fault diagnosis.
• Existing data-driven methods often neglect physical system correlations and lack interpretability.
• A robust and explainable fault diagnosis framework is needed for complex chemical processes.

Purpose of the Study:

• To propose a graph-based fault diagnosis model framework.
• To develop a dependable fault node diagnosis analysis method for enhanced interpretability.
• To improve the accuracy and trustworthiness of fault diagnosis in chemical processes.

Main Methods:

• Integrated a graph convolution network (GCN) for spatial feature extraction and a long short-term memory (LSTM) network for temporal dependencies.
• Constructed the adjacency matrix using a priori chemical process knowledge and Pearson correlation to capture physical correlations.
• Employed a dual-supervision strategy for stable model training and a multi-model voting strategy for robust inference.
• Developed a node masking method for interpretable fault node analysis.

Main Results:

• The proposed model achieved high accuracy in fault diagnosis on the Tennessee Eastman process.
• The average fault diagnosis rate reached 0.9844 across all fault types, demonstrating state-of-the-art performance.
• The node masking method effectively identified critical nodes contributing to system faults, enhancing interpretability.

Conclusions:

• The graph-based framework effectively integrates physical correlations and spatiotemporal data for accurate chemical process fault diagnosis.
• The proposed methods offer a robust and interpretable solution for complex industrial systems.
• The model demonstrates significant advancements in fault diagnosis accuracy and trustworthiness.