Deep learning-based predication of the input energy spectra for self-centering systems

Abstract

The accurate prediction of input energy spectra is crucial for the application of energy-based design methodologies. In this study, a deep-learning based artificial neural network (ANN) is utilized to evaluate the input energy spectra for self-centering single-degree-of-freedom (SDOF) systems. A dataset comprising 225 self-centering systems with varying structural characteristics is created to generate input energy spectra under 360 ground motions, specifically selected in accordance with the soil types outlined in the Chinese code. The ANN model, which incorporates a multi-input module, is developed to simultaneously consider both seismic and structural features during the prediction process. Structural features, including energy ratio η, damping ratio ξ, and ductility factor μ are extracted and used as inputs, while different seismic response spectra are employed to derive seismic features for the ANN model. The effectiveness of utilizing various input features is examined in terms of the model's performance and generalization capability. Furthermore, sensitivity analyses are performed to investigate the importance of different structural features in predicting the input energy spectra for self-centering systems and to evaluate the model's generalization capability. The results demonstrate that the proposed ANN model reliably predicts the input energy spectra for self-centering systems, regardless of variations in structural features and input ground motions. Moreover, displacement response spectra are shown to yield better performance as input earthquake features for the ANN model. Sensitivity analyses further indicate that the model, when using only ξ and μ as input structural features, maintains satisfactory performance and generalization capability, whereas the influence of η on the input energy spectra for self-centering systems is found to be negligible.