Landslide Displacement Interval Prediction And Stability Analysis Based On Machine Learning

Landslide disaster forecasting and stability analysis are significantly important in engineering.Machine learning,which is the most popular artificial intelligence method at present,has been widely used in various industries.It has also been gradually introduced into landslide prevention.However,machine learning algorithms are widely used in the landslide displacement deterministic prediction and landslide susceptibility,but they have limited applications in landslide displacement interval prediction and landslide stability analysis.Therefore,this paper takes machine learning as the main research method and uses the advantages of machine learning algorithms in regression and clustering to explore new possibilities for landslide displacement interval prediction and stability analysis.Based on the nonlinear characteristics of landslide deformation,three landslide displacement interval prediction models are proposed.Then,machine learning algorithms are used to study the global sensitivity analysis of the hydraulic parameters of the reservoir landslide,bayesian back analysis of landslides considering slip surface uncertainty and automatic identification of critical slip surface of slopes.The study reasons,processes and results are described below:(1)Accurate and reliable landslide displacement predictions are important for landslide early warning.Machine learning methods are widely used for point predictions of landslide displacement because of their powerful nonlinear ability.However,due to the uncertainties involved in landslide systems,prediction errors are unavoidable in traditional point prediction methods.To quantify the uncertainties associated with point forecasting,three interval prediction methods,namely B-LSSVM,DS-LSSVM,DESPSO-ELM,are proposed.The B-LSSVM consists of three steps: First,the LSSVM and bootstrapping are combined to estimate the true regression means of landslide displacement and the variance with respect to model misspecification uncertainties.Second,a new LSSVM model optimized by a genetic algorithm(GA)is implemented to estimate the noise variance.Finally,the point prediction is derived from the regression means and the PIs are constructed by combining the regression mean,the model variance,and the noise variance.The DS-LSSVM is a direct interval prediction approach for landslide displacements which is proposed based on the framework of the original lower upper bound estimation(LUBE)method.Two LSSVM models are directly implemented to generate the lower and upper bounds of future displacements,and the optimal model parameters are derived by minimizing a PI-based cost function via the differential search(DS)algorithms.The proposed DES-PSO-ELM method consists of three steps.First,DES is applied to predict the linear component of the cumulative displacement of the landslide.Second,the partial autocorrelation function(PACF)and maximum information coefficient(MIC)are used to select the optimal variables that influence the nonlinear component(residuals from the first step);then,these variables are used as inputs for the PSO-ELM method to construct the PIs of the nonlinear component.An ensemble technique is also applied to improve the stability and accuracy of PSO-ELM.Finally,the PIs of cumulative displacement are obtained by adding the predicted linear component and the PIs of the nonlinear component.The Baishuihe landslide and Shuping landslide and Tanjiahe landslide in the Three Gorges Reservoir area are selected to test the effectiveness of the proposed methods.The comparison of the results shows that the proposed methods perform better and can provide high-quality PIs of landslide displacement.(2)Hydraulic parameters are key data for calculating groundwater level,which is critical for assessing reservoir landslide stability.However,the quantitative understanding the influence of hydraulic parameters on the groundwater level of a reservoir landslide remains limited.In this paper,a novel global sensitivity analysis method,PAWN,is applied to quantify the sensitivity of hydraulic properties.The Shuping landslide,which is a typical reservoir colluvial landslide located in the Three Gorges Reservoir area,China,is used as a study case.The hydraulic parameters are first sampled within their entire feasibility space by the Latin hypercube sampling method.These samples are then used as inputs into a nonintrusive finite element program to automatically compute the corresponding groundwater level outputs.Finally,sensitivity indices are calculated based on the input–output dataset via the PAWN method.The global sensitivity analysis results provide useful guidelines for site investigation and reservoir colluvial landslide model simplification and calibration.(3)Although numerical methods based on strength reduction are becoming popular in slope stability analysis,they fail to provide a crisp critical slip surface but only a shear band.The widely used visualization techniques for defining the critical slip surface are susceptible to subjective judgment and are inefficient in batch analysis and three-dimensional analysis.When a slope reaches failure,the displacements on two sides of the critical slip surface will be significantly different.Based on this observation,an automatic identification method for locating the critical slip surface is proposed.The k-means clustering algorithm is firstly applied to automatically separate the nodal displacements into two categories representing the sliding mass and stable block,respectively.Then the scatters near the separation surface are obtained by constructing the alpha shape of the sliding mass.Finally,the critical slip surface is obtained by fitting the extracted scatters.A homogeneous slope and a slope with a thin weak layer are used to test the effectiveness of the proposed method.Results show that the proposed method can automatically and accurately identify the two-dimensional and three-dimensional critical slip surface.(4)Previous studies about probabilistic back analysis for shear strength parameters of landslides generally adopted a fixed slip surface.This setting may lead to unreliable results due to the uncertainty of slip surface location speculated by limited observations.Based on Bayes' theorem,this paper proposes a probabilistic framework for the back analysis of landslides considering slip surface uncertainty.The posterior distributions of shear strength parameters in Bayesian inference are solved by Markov chain Monte Carlo simulation method.To improve computational efficiency,a response surface function based on extreme learning machine is constructed to approximate the relationship between shear strength parameters and the corresponding factor of safety and critical slip surface.A synthetic slope for which the actual shear strength parameters and slip surface are known,is used to compare the proposed and traditional methods.The effects of measurement error of slip surface and prior distribution of shear strength parameters on probabilistic back analysis results are also investigated.Results show that the shear strength parameters obtained from traditional probabilistic back analyses neglecting slip surface uncertainty significantly deviate from actual values,and are greatly affected by prior mean of shear strength parameters.The proposed method performs better than traditional method,and is less affected by the prior distributions of shear strength parameters,and the smaller the measurement error of slip surface,the higher the Bayesian back analysis accuracy.