Rapid seismic damage state assessment of RC frames using machine learning methods

A rapid seismic damage state assessment of individual building is essential for a region-scale risk and vulnerability assessment that requires significant manpower, time, and computational efforts. In this study, three machine learning (ML) algorithms that exhibited high predictive accuracy in previous studies, namely random forest (RF), extremely gradient boosting (XGB), and active machine learning (AL) were used to develop models for rapidly assessing the seismic damage states of reinforced concrete (RC) frames after an earthquake. Compared to RF and XGB, the active machine learning develops an efficient model with a small number of instances by inter-actively selecting the valuable instances for desired outputs. Using these aforementioned algo-rithms, three predictive models were developed, tested, and validated using a comprehensive dataset which included a total of 9900 data points. The dataset was developed according to a non-linear time history analysis involving a combination of 199 RC frames and 50 ground motions. The results indicated that active machine learning predicted the damage states of RC frames with an accuracy of 84% in the testing dataset, followed by the XGB algorithm with an accuracy of 80%. These predictive models were also validated using actual damaged buildings in the Taiwan earthquake. Seismic design intensity (SDI) and spectrum intensity (SI) were the most important input features in the damage states of RC frames, with a relative importance factor exceeding 50% for the two features. Constructed periods have a non-negligible influence on the damage states of RC frames when these differ for regional buildings. Finally, an interactive and user-friendly graphical user interface (GUI) platform was created to provide a rapid seismic damage state assessment of RC frames. This study represents a pioneering step toward the application of AL in damage state assessment of existing RC frames.