Bias-aware degradation models for reinforced concrete bridges based on XAI

Abstract

Bridge management systems store condition data collected during visual inspections of bridges. However, rapidly evolving damage on older bridges still in operation was repaired before it could be documented in these databases, which were developed only 30 to 40 years ago. As a result, condition data is affected by survivorship bias. This work addresses this issue by developing a composite analysis based on explainable artificial intelligence techniques to analyze condition data and derive degradation models for existing bridge components. The analysis comprises four steps: (1) cluster analysis of damage transition times using the k-means algorithm to identify damage patterns with similar damage evolution rates (fast, normal, slow, corresponding to bridge components with a fragile, normal, and robust deterioration behavior); (2) Random Forest classification to predict the cluster based on bridge inventory data; (3) SHAP analysis to explain the predictions of the Random Forest classifier; (4) application of the gamma process to the grouped damage transition times to assess damage evolution. SHAP analysis reveals survivorship bias, indicating that fast-evolving damages primarily affect newly built bridges, while slow-evolving damages are associated with older bridges. By clarifying the composition of the clusters and the population of bridges to which the damage belongs, SHAP analysis also enables the development of bias-aware degradation models. The approach is demonstrated through the analysis of crack evolution in reinforced concrete bridges in Germany.