Multisource Feature Aggregation Approach for Predicting Extreme Pitting Depth Distributions of In Situ Bridge Cable Wires

Extreme pitting depths in bridge cable steel wires serve as fatigue crack initiation sites, making accurate assessment crucial for failure risk prediction. Existing methods require rust removal and cable disassembly, which not only disrupts traffic but also permanently damages serviceable cables, resulting in resource waste. A method is proposed for predicting extreme pitting depth distributions from in situ images of steel wires without rust removal. A multisource collaborative representation approach integrating color, texture, edge, and deep learning features captures corrosion information associated with pitting statistical patterns. Multiple instance learning aggregates local image features into vectors reflecting macroscopic characteristics, resolving inconsistency between local observations and statistical patterns. An interpretable model predicts the location parameter μ, scale parameter σ, and 99th percentile Q99 of the extreme pitting depth distribution. Field validation on in-service bridge cables shows relative errors of 15.31%, 3.56%, and 5.80%, respectively. The predicted distribution parameters can be converted to strength loss estimates and mapped to National Cooperative Highway Research Program (NCHRP) Report 534 corrosion stages, enabling integration with established cable assessment frameworks. This approach demonstrates a practical pathway for in situ corrosion assessment.