Decoupled Graph Attention Modeling and Anomaly Traceability Method for Multisystem Coupling in SLM Equipment

Area of Science:

• Cyber-Physical Systems Engineering
• Advanced Manufacturing
• Artificial Intelligence in Industrial Applications

Background:

• Selective Laser Melting (SLM) equipment functions as a complex cyber-physical system with challenging implicit subsystem couplings for monitoring and fault diagnosis.
• Existing deep learning methods struggle with feature submersion in multi-source heterogeneous data and lack system-level causal inference capabilities.

Purpose of the Study:

• To develop a multisystem coupling modeling and anomaly traceability method for enhanced condition monitoring and fault diagnosis in SLM equipment.
• To address limitations of current deep learning approaches in handling heterogeneous data and inferring system-level causal relationships.

Main Methods:

• Proposed a decoupled graph attention network (ST-DBGAE) featuring independent local spatiotemporal feature alignment modules for unified latent space mapping of heterogeneous sensory data.
• Implemented a dynamic graph attention mechanism with sparse priors to adaptively capture time-varying coupling weights from implicit interactions.
• Designed a dual-branch two-stage decoupled optimization architecture for continuous equipment health index (HI) output and topological difference matrix inference for root-cause analysis.

Main Results:

• Achieved a comprehensive diagnostic Macro-F1 score of 96.5% across eight operating states using over 310,000 real operational records from industrial SLM equipment.
• Significantly improved single-class detection rates (ACCs) for specific underlying anomalies.
• Demonstrated high-precision equipment health warnings and physically interpretable fault propagation mapping.

Conclusions:

• The ST-DBGAE method effectively models multisystem couplings and enables accurate anomaly traceability in complex SLM equipment.
• The approach overcomes feature submersion and provides system-level causal inference, enhancing predictive maintenance capabilities.
• The method offers a robust solution for real-time health monitoring and microscopic fault analysis in industrial cyber-physical systems.