Structural condition assessment of existing buildings using machine learning-The role of SMOTE-based data augmentation

To ensure a safe environment for occupants, analyzing the structural safety of existing buildings is essential. The main objective of this paper is to combine machine learning (ML) algorithms with the Synthetic Minority Over Sampling Technique (SMOTE) to establish a comprehensive condition assessment model for structural safety in existing buildings and provide an interpretation of results. Firstly, a raw dataset comprising 18,090 existing buildings in a region of Southwest China was assembled, containing fundamental information about each building. This dataset was then pre18 processed and augmented using the SMOTE method. Subsequently, the analysis was performed using four different ML algorithms, including artificial neural network (ANN), decision tree (DT), random forests (RF), and Adaptive Boosting (AdaBoost). Hyperparameters for these models were optimized using a grid search method with 5-fold cross-validation. Finally, a feature importance analysis was conducted based on the best-performing algorithm. The SMOTE-based RF model demonstrated the highest performance, with the evaluation metric G-mean reaching 96.34%. Among all input features, Service Life, Function, and Location were identified as the three most important factors influencing the structural condition of existing buildings. This study represents a promising approach for assisting government regulators in making critical maintenance decisions more effectively and efficiently through a rapid screening model for identifying buildings with potential structural issues.