A novel stacking-based algorithm for determining the extent of damage in braced-frame structures

It is well-understood that various machine learning algorithms can present different performances in solving complicated problems, and there are uncertainties in the machine learning-based predictions. This study proposes a stacking based seismic-induced damage detection method to reduce the uncertainties related to training individual machine learning models. To do so, three different machine learning models (i.e., support vector machine, K- nearest neighbor, and convolutional neural networks) are employed as the first-level learners. Then, the results of these predictors are combined using a decision tree algorithm as the meta-model. A series of 111 earthquake records, which were originally simulated/modified by the SAC project, is used to generate the dataset. These records are uniformly scaled from 0.05 g to 1.60 g to provide a wide range of earthquake intensities. The proposed framework uses a combination of feature extraction-based machine learning and deep learning models to implement the damage detection procedure. Combining the capabilities of feature-based and deep learning algorithms minimizes the errors related to relying on only one of these learning methods. Bayesian optimization algorithm was used in this study to tune the hyperparameters of all classification learners. A one-story chevron-braced frame and a five-story concentric braced frame structure are served to validate the proposed approach. Results show that the proposed technique increases the accuracy and reliability of predicting the extent of damage compared to individual models.