Development of AI-Based Method for Predicting Potential Damage Modes Using a Database of Past Damage Cases

抄録

Equipment in energy facilities often faces damages such as corrosion, which accounts for over 25% of US pipeline failures between 2002 to 2018 by the Pipeline Hazardous Material Safety Administration. Japan is also experiencing a shortage of skilled engineers due to an aging population, with a decline of 644,000 population in 2020-2021. In 2023, 89% of Japanese companies reported challenges in staffing issues, making it harder to manage damage modes in these facilities. Not limited to a specific industry, AI models trained on data from a variety of industries are being developed to predict damage mechanism, helping ensure safe operations despite workforce shortages for extensive application. This research compares AI models for classifying eleven damage mechanisms across literal field datasets with nineteen parameter component combinations, such as plant type, system processes, material specifications, and chemical environmental conditions. The study evaluates ten machine learning models, including Logistic Regression, Support Vector Machine, Naïve Bayes, K-Nearest Neighbor, Random Forest, Extreme Gradient Boosting, Light Gradient Boosting Machine, Gradient Boosting, Category Boosting, and Adaptive Boosting. These models are tested on damage mechanism results combinations of past field failures 691 datasets, with a One-vs-Rest approach for multiclass classification. Performance is measured using accuracy, precision, recall, F1-Score, and ROC-AUC. The study also uses Shapley additive explanations and feature importance to rank parameters influencing damage classification. Logistic Regression, SVM, KNN, and Naïve Bayes scored below 80% in ROC-AUC, while tree-based methods, such as XGBoost, CatBoost, and Gradient Boosting, scored above 85%. CatBoost demonstrated the best performance with an ROC-AUC score of 89.85%. Stress anomaly conditions had the highest SHAP value of 1.281, indicating that stress anomaly is the parameter that most influences damage mechanism classification. This research shows that machine learning is effective in accurately classifying damage mechanisms based on the condition of various components from past cases.