Dynamic Causal Entropy-Spatiotemporal Convolutional Network for Quality-Related Fault Diagnosis of Large-Scale...

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Quality Control
• Causal Inference in Engineering

Background:

• Complex industrial processes exhibit increasing data interdependence, challenging quality-related fault diagnosis.
• Traditional causal discovery methods struggle with hierarchical, dynamic, and spatiotemporal process features.
• Accurate fault diagnosis is crucial for maintaining product quality and operational efficiency.

Purpose of the Study:

• To develop an advanced method for quality-related fault diagnosis in complex industrial settings.
• To address limitations of existing causal discovery techniques in representing dynamic and spatiotemporal process data.
• To enhance the interpretability and accuracy of fault root cause identification.

Main Methods:

• A dynamic causal entropy (DCE) method to construct hierarchical dynamic causal graphs (CGs).
• A 3-D squeeze-and-excitation (SE) convolutional neural network for spatiotemporal feature analysis.
• A local-global fault detection approach with a causal anomaly vector for root cause recognition.

Main Results:

• The proposed DCE-spatiotemporal convolutional network accurately models dynamic interactions and spatiotemporal features.
• The method effectively mitigates confounding factors and enhances the interpretability of causal relationships.
• Demonstrated high effectiveness on numerical simulations and real-world hot strip mill process (HSMP) data.
• Achieved a superior fault detection accuracy of 95.78%.

Conclusions:

• The novel DCE-spatiotemporal convolutional network offers a robust solution for quality-related fault diagnosis in complex industrial processes.
• The approach enables precise recognition of fault root causes across multiple hierarchical levels.
• The method shows significant practical advantages and high accuracy, validated by real-world industrial data.