Multi-objective Probabilistic Back Analysis Of Slope Under Rainfall Condition

Rainfall-induced landslide is the serious geological disaster in the world.Traditional back analysis method usually uses one type of monitoring data to inverse analysis model parameters for the geotechnical problem,which is unable to calibrate the model effectively considering different types of data in the field.Because of the limitation of the back analysis method for the rainfall-induced slope,a multi-objective back analysis method for the problem is proposed based on the Bayesian theory.A coupled probabilistic back analysis model of a slope under rainfall condition is developed.The multi-objective back analysis method is used to the case of highway subgrade and slope.The main research findings are summarized as follows:(1)Based on Pareto optimization theory and consolidation theory of unsaturated soils,a two-dimensional coupled model is constructed.Using the multi-algorithm genetically adaptive multi-objective method,model parameters are back analyzed using two types of measurements(pore pressure,displacement)in the case of highway subgrade and slope simultaneously.The results obtained by multi-objective back analysis are compared with the results obtained by single objective back analysis.The results show that the two cases have a sharp Pareto front,so the compromise solution can be obtained.The simulation using the compromise solution can meet the two measurements well.(2)Based on Bayesian theory,a multi-objective probabilistic back analysis method using time-varied data is proposed,and a hydro-mechanical coupled slope model under rainfall condition is constructed In this paper,the model uncertainty and posterior distribution of the input parameters are estimated,and the effect of the different types of measurement to the model inversion is studied.The results show that the standard deviation of the posterior distribution is significantly reduced compared with the results obtained by single objective back analysis.The 95% uncertainty bonds of the model parameters obtained by the multi-objective back analysis are thinner than the results of single objective.(3)Model error is usually assumed to be independent and identically distributed according to a normal distribution,results in the standard least square(SLS)approach parameter estimation.A general likelihood(GL)function for correlated,heteroscedastic,and non-Gaussian,residual errors is adopted.A case study of a well instrumented natural terrain slope in Hongkong under rainfall condition is developed to illustrate the effects of residual errors on model calibration.The results show the volatility of the MPD(Maximum Posterior Density)parameters obtained by GL method reduces obviously compared with SLS method.A correlated,heteroscedastic,and non-Gaussian error assumption can describe the model error more accurately.