Hysteretic behavior of column-base connection using coupled mechanical analysis and LSTM model

Accurately and reasonably capturing the hysteretic behavior of key components with complex nonlinearities, such as connections, is of significant importance for the dynamic response analysis of the overall structure. Compared with traditional numerical and experimental methods for hysteretic identification, the component hysteretic behavior recognition method based on deep neural networks (NNs) is expected to demonstrate superior overall advantages in terms of accuracy, efficiency, and cost. Most of the current related research is based on numerical simulation data. To overcome the difficulty of obtaining hysteretic test data with time-series information, this study first used the column-base connection, which can be quickly repaired and reused, as the test component. A series of tests on specimens with different structural detailing parameters were conducted at a low cost, systematically establishing a hysteretic behavior test database. Then, the hysteretic identification and prediction method based on the long short-term memory (LSTM) model for the same component was developed and validated using the hysteretic test data of the column-base connection. The case results indicated that after training on hysteretic data from low-cycle tests, the model could effectively predict the hysteretic response under time-history displacement sequences. Subsequently, to address the challenge of hysteretic prediction for similar components, this study combined mechanical analysis with a dimensionless approach to develop a method for predicting hysteretic behavior using coupled mechanical analysis and the LSTM model. The results showed that this method could progressively predict the hysteretic response of similar components with structural detailing parameters under various displacement sequences. Finally, the NN UniaxialMaterial that can be embedded with trained deep NN models were developed based on the OpenSees platform, which facilitates the invocation of local component hysteretic behavior surrogate models for overall structural analysis. The focus of this study is to develop the hysteretic prediction method based on deep NNs for the same and similar components, with verification conducted using the column-base connection. The deep NN-based hysteretic surrogate model reduces the reliance on physical derivations, improving the efficiency of hysteretic model development, but also reduces interpretability.