基于时空神经网络的探地雷达层状介质参数反演

Parameters Inversion of Layered Medium for Ground Penetrating Radar Based on Spatio-temporal Neural Network

探地雷达(GPR)凭借其无损、高效、准确等优势已被广泛应用于道路检测、市政及环境工程等领域。GPR数据反演是获取地下层状介质参数的一种有效手段,但传统基于模型驱动的GPR反演存在线性反演精度不足,非线性反演求解难度大、时效性差等问题,无法满足现有工程勘探需求。近年来,基于深度学习的数据驱动反问题求解算法得到飞速发展,为GPR层状介质参数反演提供了一种新的思路。考虑到GPR传播的时空特性,文中提出一种基于时空神经网络的层状介质参数反演方法。首先,利用随机模拟算法生成大量一维介质参数模型,并结合有限差分算法计算得到对应GPR数据,构建出神经网络训练所需的输入/输出数据对;然后,设计合适的层状介质参数反演网络模型,并利用构建的训练数据对训练神经网络模型,得到GPR数据与地下介质参数之间映射关系的网络参数;最后,利用训练好的神经网络模型对不同层状介质模型进行反演测试,验证了所提方法的有效性和准确性。

Ground penetrating radar(GPR)is widely used in the fields of road inspection,municipal and environmental engineering due to its non-destructive,efficient and accurate advantages.GPR inversion is an effective means for obtaining parameters of the subsurface layered medium.However,the traditional model-driven based GPR inversion has the problems of insufficient accuracy of linear inversion,difficulty in solving nonlinear inversion,poor timeliness,etc.,which cannot meet the existing engineering exploration needs.Recently,data-driven inverse problem solving algorithms based on deep learning have been developed rapidly,which provides a new way of thinking for the inversion of layered medium parameters for GPR.Considering the spatio-temporal characteristics of GPR propagation,a spatio-temporal neural network-based inversion method for the parameters of layered media is proposed.First,the stochastic simulation algorithm is used to generate a large number of 1D models,and the corresponding GPR data are calculated by combining with the finite-difference algorithm to construct the input/output data pairs required for neural network training.Next,a suitable inversion network for layered medium parameters is designed and the constructed training data pairs are used to train the network to obtain the network parameters for the mapping relationship between GPR data and subsurface medium parameters.Finally,the effectiveness and accuracy of the proposed method is verified by inversion tests of different layered medium models using the trained neural network.