基于改良U-Net卷积神经网络的复杂地质构造智能识别

Automatic Geological Structure Detection Based on Seismic Method and Improved U-Net CNN

浅地层中的地质异常体给地下工程带来了极大安全隐患,可能导致经济和生命损失.浅层地震法是开展施工场地勘察的一种无损高效的手段.但是地震勘探在浅地层问题中面临着信噪比低、信号衰减强、波场复杂等问题,结果存在多解性和主观性.针对地震勘探问题,提出了一种改良的卷积神经网络地质速度模型预测模型,提供了一种无需初始速度模型的浅层地质模型反演方案,形成了一套完整的浅层地震勘探信号处理处理流程.在训练样本方面,采用了随机地质模型方法构建多种类地质模型,并形成了地质模型-地震信号数据库.在传统U-Net卷积神经网络上进行了改良,以更好地适应浅地层弹性波叠前信号数据的反演任务.结果表明,神经网络的反演结果直观准确该模型能够精确地预测出地层分界线、褶皱、起伏、断层滑移线等的位置和大小等参数,所采用的SSIM和PSNR两个定量化评价指标均表示,所提出的改良神经网络可以实现高精度反演.预测结果与真实模型相比较,得到的SSIM平均值为0.91,PSNR平均值为39.0.同时该神经网络模型能够向三维问题扩展,能够极大地提高地震信号处理的效率和解译精度.

Geological anomalies in near surface are extremely harmful to engineering.Unexplored structures may lead to construction difficulty,economic loss and even safety problems.Therefore,the investigation of the construction site in near surface is a key procedure before construction.Shallow seismic method is an efficient method to carry out the investigation of the construction site.However,seismic exploration faces problems such as low signal-to-noise ratio,strong signal attenuation,and complex wave field in shallow strata,which requires a lot of human experts’resources to interpret while the results are ambiguous and subjective.Aiming at the problems of shallow seismic exploration.In this paper,an improved convolutional neural network geological velocity model prediction model is proposed for seismic exploration,which provides a shallow geological model inversion scheme.This workflow provides a complete set of near surface seismic exploration signal processing and inversion process scheme without initial velocity model.This paper uses the stochastic geological model method to construct a variety of geological models,and forms a geological model-seismic signal database.An improved U-Net convolutional neural network to better adapt to the inversion task of near surface pre-stack signal data of elastic wave.The results show that the inversion results of the neural network are intuitive and accurate.Two quantitative evaluation indicators of SSIM and PSNR indicate that the improved neural network proposed in this paper can achieve high-precision inversion.The neural network model proposed in this paper can be extended to three-dimensional application,which can greatly improve the efficiency and interpretation accuracy of seismic signal processing.