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

As a critical component of electric railways, catenary systems are prone to cracks, loosening, corrosion, and wear under long-term vibration, fatigue, and environmental erosion. However, ambiguous fault boundaries, large inter-component variations, and tiny defects severely hinder reliable anomaly detection and condition assessment. To address these challenges, this paper proposes a vision-based intelligent fault assessment framework for railway catenary components based on a novel Railway Diffusion-based Anomaly Detection (Rail-DiffAD) model. Specifically, Rail-DiffAD combines residual feature mapping, a Multi-scale Partial Convolutional Spatial-Channel Attention (MPSCA) module with Log-Barrier Bi-directional Constraint Loss (LBBCL), and conditional diffusion-based distribution modeling to achieve robust anomaly localization in complex industrial scenarios. Furthermore, a severity-aware diffusion representation is introduced to characterize structural defect evolution, and a multi-physics fault assessment framework integrating mechanical response, corrosion evolution, and stress concentration analysis is established for quantitative fault grading and maintenance decision-making. Experiments on a real catenary dataset covering 10 component categories demonstrate that the proposed framework achieves a 0.953 image-level AUROC and a 0.957 pixel-level AUROC, outperforming existing methods while maintaining strong cross-component generalization and providing quantitative fault grading support for intelligent railway catenary maintenance.