Hybrid deep learning framework for enhancing seismic response prediction of slope pile-anchor composite reinforcement system

Abstract

In earthquake-prone mountainous regions, pile-anchor composite structures have been applied to practical slope reinforcement projects. Enhancing the capability to precisely forecast the seismic dynamic response of these support systems is crucial for safeguarding the security of personnel and assets. This article introduces an innovative CNN-LSTM-attention model that can handle long sequences more flexibly and effectively capture and utilize critical information from the input data. This study generates a series of random seismic motions using the Spectral Representation-Random Function Method and conducts consecutive dynamic centrifuge shaking table tests to obtain seismic response data for pile top displacement and anchor tension. A noise filtering process based on Discrete Wavelet Transform (DWT) was implemented, coupled with a Moving-Steps approach for expanding the dynamic response database of pile-anchor composite structures. Superior performance in seismic dynamic response analysis of slope pile-anchor composite structures was achieved through a hybrid architecture combining convolutional neural networks (CNNs), long short-term memory (LSTM), and attention mechanisms. This model shows strong concordance with centrifuge test outcomes and enhances adaptability in learning the pattern of seismic wave and the response of slope pile-anchor composite reinforcement system.