Industrial Anomaly Detection and Fault Grade Assessment for Railway Catenary Components Based on Diffusion Models

Area of Science:

• Engineering
• Artificial Intelligence
• Materials Science

Background:

• Railway catenary systems face degradation from vibration, fatigue, and environmental factors.
• Detecting faults is challenging due to ambiguous boundaries, component variations, and small defects.

Purpose of the Study:

• To develop an intelligent, vision-based framework for anomaly detection and condition assessment of railway catenary components.
• To improve the reliability and accuracy of fault identification in complex industrial settings.

Main Methods:

• Proposed a novel Railway Diffusion-based Anomaly Detection (Rail-DiffAD) model.
• Integrated residual feature mapping, Multi-scale Partial Convolutional Spatial-Channel Attention (MPSCA), and Log-Barrier Bi-directional Constraint Loss (LBBCL).
• Utilized conditional diffusion-based distribution modeling and severity-aware diffusion representation for defect characterization and quantitative grading.

Main Results:

• Achieved high performance with 0.953 image-level AUROC and 0.957 pixel-level AUROC on a real catenary dataset.
• Demonstrated robust cross-component generalization capabilities.
• Successfully provided quantitative fault grading for maintenance decision-making.

Conclusions:

• The proposed Rail-DiffAD framework offers a significant advancement in intelligent fault assessment for railway catenary systems.
• The multi-physics approach enables accurate defect characterization and grading, supporting proactive maintenance strategies.