Screening for Anomalous Safety Condition Among Existing Buildings Using Explainable Machine Learning

To ensure a safe environment for occupants, evaluating the physical status and service performance of existing buildings is essential. However, large-scale building condition assessment usually relies on the expertise and judgment of inspectors, which can be costly and laborious due to unclear priorities, ambiguous procedures, and ineffective operations. To address these challenges, this study proposes an explainable machine learning-based screening model for the anomalous safety condition among existing buildings, narrowing down the scope of buildings requiring further and detailed inspection and monitoring. Initially, an imbalanced dataset of 18,090 survey reports of existing buildings of safe and unsafe labels is collected. Then, the synthetic minority oversampling technique (SMOTE) is conducted to balance the dataset. Subsequently, seven machine learning models are trained utilizing 10-fold cross-validation with grid search. Findings reveal that, based on the balanced dataset, the performance of ensemble learning models is significantly better than that of individual machine learning models. Specifically, the XGBoost model achieves the highest performance, with a macro-F1 of 98.49%, G-mean value of 98.49%, and accuracy of 98.49%. The final predictive model (the SMOTE-based XGBoost model) is explained using the SHapley Additive exPlanations (SHAP). Service year, structure, and location are the three most important features influencing building structural safety. This study represents a promising approach for automated screening of the anomalous safety condition among buildings, optimizing resource allocation, and enhancing the effectiveness in decision-making for construction and maintenance.