.png){kind=icon}
.png){kind=icon}
Return
Seismic response prediction method for prefabricated bridge piers based on multi-head self-attention mechanism and frequency feature fusion
.png){kind=icon}
PRE
AI
Z
Y
Y
S
DOI：10.1016/j.istruc.2026.112395.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In cold seismic regions, freeze-thaw damage severely degrades the seismic performance of prefabricated bridge piers with grouted sleeve connections, leading to exacerbated deformations and accelerated failure under seismic loads, which threatens structural safety. To enable rapid and accurate seismic response prediction for such freeze-thaw-affected piers, a prediction model based on a multi-head self-attention mechanism and frequency feature fusion is proposed in this paper. First, frequency features correlated with structural responses are extracted from seismic response spectra using multiple one-dimensional convolutional neural networks (1D-CNNs), which accurately capture the frequency energy distribution and evolutionary characteristics of seismic waves. Subsequently, an SENet-MHSA-SENet feature fusion module, composed of Squeeze-and-Excitation (SENet) networks and Multi-Head Self-Attention (MHSA), is constructed to substitute traditional concatenation methods. This module dynamically recalibrates, globally interacts with, and enhances the frequency features extracted by the convolutions, thereby achieving a deep-level feature representation. Finally, the fused deep features, alongside ground motion intensity measures (PGA, PGV, PGD) and the number of freeze-thaw cycles (n), are input into fully connected layers to predict the maximum curvature response of the bridge pier's critical sections. This prediction model integrates the channel selection capability of the SENet mechanism with the global interaction advantages of MHSA and enhances key features at the local level while considering contextual dependencies at the global level, significantly improving feature representation capabilities while maintaining computational efficiency. A case study was conducted on prefabricated test piers with grouted sleeve connections from the Ruoqiang–Hetian Railway. The results indicate that the proposed 1D-CNN-SENet-MHSA-SENet model reduces the mean absolute error (MAE) by up to 25.5% compared to a traditional 1D-CNN model. Performance is consistent across training and test sets, with MSE values of 0.0267 and 0.0236 and R² values of 0.973 and 0.974, indicating no overfitting. The distributions and variation trends of the predicted versus actual values, as well as the residual distributions, are consistent across both datasets. Furthermore, 80% of the samples exhibit an absolute relative residual of less than 20%, demonstrating high predictive accuracy and strong generalization capability. This research provides a reliable and efficient deep learning model for post-disaster seismic risk assessment of bridges.
Journal
IF:
4.3
Papers:
1.2W
Citations:
2.7W
Organization
No organization information available
Cited Papers