A Frequency-Based Ground Motion Clustering Approach for Data-Driven Surrogate Modeling of Bridges

Machine learning-based methods, especially deep learning methods, have achieved great success in seismic response modeling due to their exceptional performance in capturing nonlinear features. However, imbalanced features of a limited training data set can significantly decrease the prediction accuracy of machine learning models. Therefore, this study proposes a novel frequency-based clustering approach for ground motion selection to generate a balanced training data set to improve the data-driven surrogate modeling of bridges. The hierarchical clustering method was developed to suppress the redundant information on the basis of a wavelet analysis of ground motion records. The proposed method was validated by a benchmark finite-element model of a girder bridge, in which long short-term memory (LSTM) neural network was used to predict the seismic responses given ground motion excitations. Specifically, the prediction performances of LSTM surrogate models trained using different data sets have been compared, while the influence of time-frequency characteristics of ground motions has been discussed in detail. The results indicated that the proposed method can provide a balanced training data set with a uniform distribution of time-frequency characteristics and effectively improve the prediction accuracy of deep learning-based surrogate models.