Seismic Performance Prediction of a Slope-Pile-Anchor Coupled Reinforcement System Using Recurrent Neural Networks

Abstract

Seismic performance prediction of slope reinforcement measures is an essential and significant problem in structure design and monitoring stage. Enlightened by the advancement of Recurrent Neural Network (RNN) in geotechnical engineering, this paper develops RNN models integrated with discrete wavelet transform and Dung Beetle Optimization (DBO) algorithms for predicting the dynamic response of slope-pile-anchor coupled reinforcement systems. Centrifuge shaking table tests and moving-steps strategy is used to create database for model training. Results shows that DBO-Bidirectional Long-Short Term Memory (BiLSTM) demonstrate superiority than DBO-simple RNN, DBO-LSTM and DBO-Gated Recurrent Unit (GRU) due to its unique bidirectional learning ability. Moreover, four cases are predicted using the DBO-BiLSTM model and compared with measured data for verification. It is concluded that the proposed DBO-BiLSTM model is suitable for the time-series prediction of seismic responses of slope stabilization structure, offering solutions for slope engineering structures performance design and health monitoring under limited monitoring data scenarios.

Keywords: Recurrent Neural Networks, Bidirectional Long-Short Term Memory, Slope-pile-anchor coupled reinforcement system, Dynamic centrifuge tests, Seismic performance prediction