Neural network-based hysteresis modeling for structural steel components

Abstract

In this study, a novel approach for constructing a neural network-based hysteresis model is presented. As an alternative to conventional differential equation-based models that require extensive parameter identification and considerable computational effort, the proposed method directly generates hysteresis loops from input datasets, eliminating the need for predefined mathematical formulations. The model was trained using both numerical data from the Bouc-Wen (BW) and Bouc-Wen-Noori (BWBN) models, as well as experimental data obtained from pseudo-static cyclic tests of composite beam-to-column connections and brace components. These experimental datasets were selected to evaluate the generalization performance of the neural network model against cases exhibiting severe asymmetry and pinching in the hysteresis loops. A key innovation of this study is the dataset augmentation process, which incorporates both original hysteretic data and datasets with randomly sampled intervals to improve model robustness. The proposed neural network-based model demonstrates high accuracy in predicting hysteresis loops, effectively capturing the behavior of both Bouc-Wen class models and the experimental datasets. Based on the hysteresis generation performance comparison for BW datasets with coarser sampling rates, the numerical stability of the proposed neural network model was evaluated. Additionally, a comprehensive analysis of the network structure and hyperparameters was performed, highlighting the importance of structural design and hyperparameter tuning in optimizing model performance. By eliminating the need for the sequential numerical analysis steps required in conventional modeling approaches, the proposed method contributes to reducing the computational effort involved in the complex nonlinear dynamic analysis of structural systems.