Dynamic Risk Inference Method for Chemical Industrial Inspection Based on Spatio-Temporal Scene Graphs

Area of Science:

• Robotics
• Computer Vision
• Chemical Engineering

Background:

• Dynamic viewpoint noise in mobile chemical inspections leads to high false alarm rates, compromising industrial process safety.
• Existing methods struggle to effectively mitigate complex visual noise and topological mutations.
• Edge computing deployment requires efficient and high-confidence sensing solutions.

Purpose of the Study:

• To develop a highly robust adaptive dynamic risk inference model for mobile chemical inspections.
• To significantly reduce false alarm rates and improve the reliability of safety monitoring systems.
• To meet the demands of edge computing with low processing times and high accuracy.

Main Methods:

• Integration of spatio-temporal semantic constraints within an inference framework.
• Construction of a heterogeneous dynamic scene graph with a kinematic-aware anisotropic dynamic field.
• Design of an uncertainty-aware adaptive hysteresis filter with dynamically adjusting thresholds.

Main Results:

• Achieved a peak F1-Score of 93.1% on a real-world chemical dataset.
• Reduced the false alarm rate to 1.3 false alarms per hour.
• Demonstrated a single-frame processing time of only 24.8 ms, suitable for edge computing.

Conclusions:

• The proposed model effectively reduces spatio-temporal dynamic noise and mitigates alarm chattering.
• The method provides a high-confidence sensing decision hub for industrial process safety.
• The model meets edge computing deployment requirements for mobile chemical inspections.