Inspection Data‐based Prediction on Fatigue Crack of Orthotropic Steel Deck Using Interpretable Machine Learning Method
Yihu Ma,Benjin Wang,Airong Chen
DOI: https://doi.org/10.1111/ffe.14377
IF: 3.373
2024-01-01
Fatigue & Fracture of Engineering Materials & Structures
Abstract:The prediction of fatigue cracks on orthotropic steel decks is of great significance to the maintenance of bridges. However, fatigue cracks are affected by various uncertainties in reality, which encourages a data‐driven study for the sake of reliability and accuracy of predictions. Based on the crack inspection data from orthotropic steel decks on actual bridges in China, the feature engineering is conducted considering fatigue crack behaviors, and the machine learning models are trained and tested for predicting cracks, including XGBoost, random forest, and multiple decision trees. According to the receiver operating characteristic curves of the three models, the XGBoost model has the best performance, whereas the average AUC is about 0.75, limited by the insufficient data volume of positive samples. With the SHAP values of all features, the interpretation of the machine learning model is presented, indicating that the global effects, that is, the longitudinal position, the loading condition, and the bridge age, are always influential factors for fatigue cracks. The local features concerning the interactions between cracks have an effect on crack behaviors to a certain extent, but less important. Accordingly, the interpretable machine learning model can provide conservative predictions in a rather transparent way on this issue, which can benefit decision‐making in bridge designs, maintenance, and management.