A deep ensemble learning-driven method for the intelligent construction of structural hysteresis models

Accurate force–deformation hysteretic models for structures, components, and materials are essential for structural analysis. The development of an explicit mathematical model for hysteresis poses challenges in the fields of artificial model selection and precise and efficient calibration of key parameters. In recent years, deep learning-based implicit hysteresis model construction methods have been widely recognized. However, such methods exhibit instability and lack of generalization. By combining deep neural network feature extraction with physics-based explicit hysteresis models, it is possible to efficiently construct an accurate and stable force–deformation hysteresis model purely based on test data. Therefore, in this study, a sensitivity-guided long short-term memory neural networks (LSTM) method is proposed to extract the features of structural behaviors and predict the key parameters of explicit hysteresis models accurately and efficiently. Furthermore, an adaptive ensemble learning method based on multiple explicit hysteresis models is proposed to improve the generalization ability. Case studies on test results demonstrate that this method can accurately predict structural behaviors, which is more efficient than manual construction and has better generalization ability than classical data-driven hysteresis models.