System Probability Of Failure Assessment Of Soil Landslides Using Active-learning Surrogate Models And Strength Reduction Method

It is difficult to quantify uncertainties of physical and mechanical parameters for soil landslides using conventional deterministic stability model with factor of safety.In order to obtain more reasonable evaluation index of landslide stability,reliability methods provide a way to quantify those uncertainties,and provide probability of failure assessment for landslides.However,it is difficult for conventional reliability methods to obtain a balance between computational accuracy and efficiency.Particularly,when complex landslides are considered,the great amount of computation cost can't satisfy the demand of engineering application,which severely limits the application of failure probability evaluation method of landslide in practical engineering.Based on the existing research results of reliability theory of geotechnical engineering worldwide,this paper introduces and improves active-learning surrogate models,and further develops the finite difference strength reduction method.A binary classification method(BCM)for system probability of failure estimation of landslide is proposed by combining the aforementioned two techniques.Finally,a complete,efficient,concise and accurate evaluation process for failure probability of landslide is established,which can provide a comprehensive probabilistic evaluation index for the landslide,including system failure probability,failure probability along different slip surfaces and volume.The main contents of this paper are:(1)On the basis of systematically summarizing the related researches of system failure probability of landslide,the finite difference strength reduction method is taken to compute factor of safety,and three advanced surrogate models are introduced.On this basis,several important elements of active-learning strategy,including the initial sampling strategy,active-learning functions and the stopping criteria,are constructed,respectively.Then,three active-learning surrogate models,including active-learning support vector machine(ASVM),active-learning Kriging(AK)and active-learning radial basis function(ARBF),are established to estimate the systemfailure probability of soil slopes,by which the number of calculation times of the factor safety using strength reduction method can be reduced to minimum on the premise of ensuring accuracy.(2)On the basis of in-depth study on the theory of system failure probability analysis of landslide,a judgement-based strength reduction method is proposed to replace the traditional strength reduction method.And together with ASVM,a new method for estimating failure probability of landslide based on binary classification is proposed.This method can greatly improve the computational efficiency of failure probability estimation of landslide to meet the needs of engineering applications meanwhile ensuring accuracy.(3)Based on the predicted limit state function constructed by binary classification model and considering the correlation of different potential slip surfaces,a representative slip surfaces identification method based on finite difference model,together with its corresponding failure probability and failure volume,is proposed referring to previous research results.In this paper,advanced reliability methods are integrated to establish a complete set of evaluation process for failure probability of landslide,and some algorithms and procedures are extended and optimized to increase the operability and reproducibility of the whole process.Through the tests and validations of many soil slope cases,the evaluation method for failure probability of landslide presented in this paper can meet the requirements of practical engineering application in terms of computational efficiency and accuracy.The proposed method has important theoretical and practical significance for the development of landslide stability evaluation theory,which can assist the development of landslide quantitative risk assessment and guide the decision-making of actual engineering projects.