地面微地震资料震源定位的贝叶斯反演方法

Bayesian inversion method for surface monitoring microseismic data

针对地面微地震资料信噪比低、初至拾取不准、速度模型难以准确建立等问题,以及地面微地震资料多条测线测量和浅地表地层速度变化复杂特点,研究了地面微地震资料震源定位的贝叶斯反演方法,把所有测线反演结果设定为一个全概率事件,每条测线反演问题设定为一个划分,讨论利用贝叶斯最大后验方法反演震源位置。在反演时浅部采用横向变速模型,中深部采用水平横向均匀速度模型模型。对目标函数的后验概率密度函数、加权函数后验密度函数、速度参数方差的后验概率密度函数进行理论模型拟合,并取拟合后结果作为估计概率密度。采用极快速模拟退火方法加网格法的混合算法作为搜索方法,以网格算法为先导使搜索落入最优解所在的凸区间,再利用极快速模拟退火算法搜索最优解,这样既可以防止算法收敛于局部极值点,又极大地提高了算法的收敛速度。通过理论模型和实际资料验证了该方法的应用效果,即对随机跳动误差较大初至反演能够保证反演结果的精度。

The surface monitoring microseismic data is characterized by low S/N,inaccurate first-break picking-up,difficult to establish accurate velocity model.Meanwhile,surface monitoring microseismic data is measured by multi survey lines and its shallow surface layers have complex strata velocity.In order to solve the above problems,the Bayesian inversion method for the source positioning of surface monitoring microseismic data was probed.We assume the inversion results of all lines as a total probability and the inversion of every line as a classification,and then discuss the Bayesian maximum posterior method for source position inversion.During inversion,lateral velocity-variable model is adopted in shallow layers and horizon lateral even velocity model in middle-deep layers.The theoretical models are fitted on posterior probability density function of objective function,weighted function and velocity parameter variance.The fitting results are regarded as estimation probability density.Taking the hybrid algorithm of extremely fast simulated annealing grid method as searching method,and grid algorithm as guide to make searching fall into the convex interval of the optimal solution.Then,extremely fast simulated annealing algorithm is used to search the optimal solution.This is to prevent the algorithm converges to a local extreme point,and greatly improve the speed of convergence of the algorithm.The application results on theoretical model and actual data indicate the validation of the method that is to guarantee the accuracy of first-break inversion with large random error.