有限数据条件下空间变异岩土力学参数随机反演分析及比较

Stochastic back analysis and comparison of spatially varying geotechnical mechanical parameters based on limited data

现场和室内试验数据通常非常有限,基于有限数据难以确定岩土力学参数统计特征。随机反演方法可为有限数据条件下岩土力学参数统计特征及概率分布推断提供一条有效的途径。发展3种可解释岩土力学参数空间变异性的参数随机反演方法(DREAM(zs)、BUS和aBUS),并从随机样本产生方式、收敛判据、模型证据和后验失效概率计算等方面对这3种方法的基本原理进行比较。最后,通过2个边坡案例对这3种方法的收敛性、计算精度和效率等进行了系统比较,进而总结了针对不同岩土力学参数随机反演问题应优先推荐使用的方法。结果表明:DREAM(zs)方法对于低维问题计算精度和效率高。BUS方法在子集模拟运算之前需要提前确定似然函数乘子取值,适合分析考虑参数空间变异性和似然函数计算量较大的高维问题。aBUS方法不依赖于似然函数乘子取值,并且计算精度较高,适合求解考虑参数空间变异性和似然函数计算量相对较小的高维问题,但是该方法需要耗费一定的计算量来定量地判断计算是否收敛。

In-situ and laboratory test data are often quite limited,by which it is hard to determine the statistical characteristics of geomechanical parameters.Fortunately,the stochastic back analysis method provides an approach to overcome the shortcoming.In this paper,three stochastic back analysis methods(i.e.,DREAM(zs),BUS and aBUS)of geomechanical parameters accounting for the effect of spatial variation are developed,and the basic principles of the three methods are compared from the aspects of generation of random samples,convergence criterion,model evidence and estimation of posterior probability of failure.Two slope examples are investigated to further compare these three methods systematically on the convergence,computational accuracy and efficiency.Based on these,the DREAM(zs),BUS and aBUS methods are respectively recommended to give priority to tackle different stochastic back analysis problems.The results indicate that the DREAM(zs)method has good computational accuracy and efficiency only for dealing with low-dimensional problems,that the BUS method,in which the value of likelihood function multiplier has to be determined before the operation of subset simulation,is preferable to solve high-dimensional problems involving the spatial variability of mechanical parameters and intensive computations likelihood function,and that the aBUS method,which does not rely on the likelihood function multiplier and has good computational accuracy,is fairly suitable for analyzing high-dimensional problems involving the spatial variability of mechanical parameters and less computation of likelihood function,although it is time consuming to quantitatively determine whether the computations converge to the accurate results.