Probabilistic Seismic Landslide Hazard Analysis Based On Deep Neural Network And Mixed-effects Prediction Model

Earthquakes are a major cause of landslides.Studies have shown that earthquakes of magnitude 4 or higher will trigger landslides.As a result of a major earthquake,a large number of landslides can occur in an instant.These landslides can not only bury houses and buildings,but also cause significant damage to critical infrastructures such as roads and railways,resulting in injuries and property destruction.As a result of the study for the probabilistic seismic landslide hazard analysis of slopes to potential sources,potential permanent displacements and hazard curves for slopes can be generated.Providing technical support and scientific basis for the planning,construction,and operation of the slope,the study is of significant research significance and engineering value.A subset of the NGA-West2 ground motion database is selected in this study.Those low-quality ground motions are excluded from the database.The Newmark sliding block displacement method is used to calculate the permanent displacement of landslides.The uncertainties of slope parameters and critical accelerations are considered for regional landslides.First,based on the deep learning method,the permanent displacement prediction model is developed.Furthermore,the linear mixed-effects regression algorithm is used to optimize the displacement model.Finally,a full probability seismic hazard analysis of landslides is proposed based on correlations of ground motion parameters calculated,using the Monte-Carlo simulation.The main research contents and conclusions are as follows:(1)3316 pairs of ground motion records are selected from the NGA-West2 database,which are then used to calculate ground motion intensity measures(IMs).A comprehensive database for ground motions and ground-motion intensity measures is established.The distribution of slope parameters is used to generate a set of slope critical accelerations.The permanent displacements of slopes are determined by taking into account the vertical loading of ground motions on the slope.Consequently,a database consisting of 11,426,936displacements is established.(2)The correlation between IMs and displacements is first examined.On the basis of a proper vector of IMs that are selected from a series of comparison analyses,deep neural network-based(DNN)prediction models are then proposed.The DNN models are compared with three traditional regression-based prediction models using the same predictors.It was found that the DNN models increased R~2 to 0.877,0.831,and 0.918 from 0.88,0.769,and0.885,and decreased the residual standard deviation by 13%,14%,and 15%,respectively.(3)After comparing different combinations of soil and slope geometric parameters,a linear mixed-effects model for the permanent displacement residuals is chosen.In this model,soil effective internal friction and critical acceleration are considered fixed effects,while slope critical acceleration is considered a random effect.The residual model is then combined with the DNN prediction model to form the hybrid model which significantly improves prediction performance.The R~2 increases from 0.9400 to 0.9764,and the standard deviation of residuals decreases by more than 32%.The hybrid model significantly reduces the standard deviation by more than 58%compared to conventional regression models(SR08).(4)A logical tree framework is used to consider multiple ground motion prediction equations for IMs.The residuals of ground motion IMs are corrected through a random effect model.A correlation model of the ground motion IMs is examined based on their corrected residuals.This is followed by the production of a heat map for the correlation.It is found that the correlation between acceleration and velocity parameters is relatively positively high and that the duration-related parameters are relatively independent.(5)The probabilistic seismic landslide hazard assessment was carried out for three hypothetical slopes in the Longmenshan fault zone using earthquakes generated by a Monte Carlo simulation,taking into consideration different critical slope accelerations.As a result,hazard curves for permanent slope displacements are drawn.According to the results:the seismic landslide hazard of slopes at the near-fault site decreases significantly with the increase of ac value;the medium seismic landslide hazard of slopes will be underestimated if the correlation of the ground motion IMs is not taken into consideration;and the effective internal friction angle of soil(φ’)has a significant effect on the seismic landslide hazard.