Multi-Grained Recognition of Facility Surface Defects and Safety Assessment Based on CAE-SHCIF

Abstract:

A multi-granularity defect recognition and safety assessment method is proposed for facility surfaces， using bridges as a representative example. The approach integrates a CAE_ViT network model with a sequential hierarchical coupled information framework （SHCIF） and a fuzzy comprehensive evaluation （FCE） system. First， the SHCIF and three corresponding granularity-specific recognition models are established， with datasets constructed and augmented for each granularity level. The SHCIF and cross-granularity classification strategy are designed to enhance defect severity recognition accuracy by leveraging information from both bridge component and defect type granularities. Second， transfer learning is applied to fine-tune the CAE_ViT pre-trained model for bridge defect detection， with classification performance further improved through cross-granularity decision-making. Finally， an analytic hierarchy process-entropy weight method （AHP-EWM） weighting system is incorporated into the FCE to achieve quantitative safety assessment of bridges based on apparent surface conditions， considering bridge locations， components， defect types， and severity levels. Experimental results show macro-average F1-scores of 94.1%， 81.6%， and 75.3% for the three granularity levels， respectively， with cross-granularity classification reaching 82% accuracy. A case study on bridge safety evaluation demonstrates the effectiveness， systematicness， and extensibility of the method.