基于多尺度边缘特征的深度学习电阻率反演方法

Deep learning resistivity inversion method based on multi-scale edge features

直流电阻率法是一种经济、高效的工程地球物理探测手段,对含水构造敏感。线性电阻率反演是实际探测中的主流方法,但其反演结果容易陷入局部最优,产生错误的地质解译。与之相比,无监督反演方法能够采用物理规律和数据挖掘双驱动训练网络,摆脱对真实模型的依赖,具备在实际数据中全局搜索的可行性。在无监督反演方法的基础上,创新了基于多尺度边缘特征的深度学习边界刻画方法。针对反演成像边界模糊的问题,借鉴地震、电磁勘探中多尺度反演的经验,提出了一种电阻率多尺度反演方法,以多尺度反演目标函数作为损失函数修正网络梯度,有效提高了无监督学习反演的边界刻画能力。在上海市域铁路机场联络线1号风井工程开展现场试验,以5号基坑地连墙渗漏点探测为例,探明了15处低阻异常,指导基坑补强作业,验证了方法的可行性和有效性。

The DC resistivity method is an economical and efficient engineering geophysical detection method.It has been widely applied in engineering detection due to its high sensitivity to a water-bearing geological body.The linear inversion is now mainly used in the data imaging and interpretation of the DC resistivity method in real applications.However,its inversion results are easy to fall into the local optimum,which may lead to a wrong imaging result as well as the geological interpretation.The unsupervised inversion method can use both the physical laws and data mining.It could get rid of the dependence on the real model,and can retain the feasibility of global search ability in the real data.Based on the unsupervised inversion method,we developed a deep learning boundary characterization method based on multi-scale edge features.To solve the problem of the fuzzy boundary of inversion imaging,we proposed a multi-scale inversion method of resistivity based on convolution wavelet transform by drawing on the experience of multi-scale inversion in seismic and electromagnetic exploration.On this basis,we used the multi-scale inversion objective function as the loss function to correct the network gradient,and thus effectively improved its boundary characterization ability.We carried out field tests in the No.1 air shaft project of the Shanghai regional airport connecting line.Taking the detection of the leakage point of the diaphragm wall of the 5th foundation pit as an example,we verified 15 low resistance anomalies to guide the reinforcement of the foundation pit,which proves that the method is feasible and effective.