Ground motion inversion utilizing sensor-based structural response time histories with a multi-head attention-enhanced temporal convolutional network

Accurate measurement of ground motion (GM) time histories is essential for precise structural analysis and safety evaluation. However, high costs, limited availability, and maintenance demands of GM recording devices lead to several challenges. This paper aims to present the Multi-Head Attention-Enhanced Temporal Convolutional Network (MHA-TCN Net), a framework developed for inverting GM from recorded structural response (SR) data during earthquakes. A case study on complex multi-degree-of-freedom (MDOF) structures demonstrates the MHA-TCN Net's ability to predict GM time histories across elastic and various plastic deformation ranges, showcasing its potential for real-time GM inversion. By incorporating a multi-head attention mechanism along with an overlapping subset strategy, the MHA-TCN Net achieves high precision in GM inversion. It consistently demonstrates low validation losses, with the mean absolute error not exceeding 0.063, exhibits test correlation coefficients as high as 0.95, and maintains a mean square error below 0.4 on unseen test datasets. Comparative experiments against the Temporal Convolutional Network, Long Short-Term Memory, Convolutional Neural Network, Gated Recurrent Unit, and Transformer models demonstrate the MHA-TCN Net's faster convergence and superior accuracy for both training and validation phases. Importantly, the proposed approach can handle inherent sensor noise, ensuring reliable performance under challenging field conditions, thereby highlighting the contribution of advanced artificial intelligence methods for engineering applications with constrained data environments. Furthermore, without retuning the network, the model demonstrates high accuracy with limited real-time training datasets, proving its versatility across various real-world datasets, especially in areas with SR sensors but few seismic instrumentations.