Deep learning-based ground motion inversion through recursive structural acceleration response using DRA-LSTM Net

Known ground motion time history is crucial for the precise numerical analysis of structural response and evaluation of its safety. This study introduces the Deep Recursive Attentive Long Short-Term Memory Network (DRA-LSTM Net), a framework designed for the inversion of ground motion (GM) from recorded structural response (SR) during an earthquake. By integrating an attention mechanism in network along with a strategic recursive approach for dataset management, the DRA-LSTM Net achieves high precision in GM inversion. The method incorporates a novel pre-processed matrix format to capture unique recursive behaviour patterns in GM data, enhancing the model's training, validation efficiency, and prediction accuracy on unseen test datasets. Case studies on single-degree-of-freedom (SDOF) and multi-degree-of-freedom (MDOF) structures highlight the method's potential for real-time and high-precision GM inversion. Additionally, transfer learning was implemented to fine-tune the pre-trained model with real-world sensor data and structural responses from different floors, demonstrating the model's adaptability and ability to retain high prediction accuracy across diverse datasets. This approach is particularly useful for regions with limited seismic instrumentation but equipped with structural response sensors.