基于U-Net深度神经网络的地震数据去噪研究

在能源和矿产资源开采中,大规模采空区塌陷不仅威胁了矿区工人和设备的安全,同时也制约了矿区的经济效益。当前,地震波反射法是探测隐伏采空区的一种常用方法。但在实际探测过程中获取的地震数据往往含有大量随机噪声,这些噪声会为后续数据处理和成像环节带来较大干扰,因此对数据进行随机噪声压制和去除通常是地震数据处理的首要工作。采用改进后的U-Net深度神经网络作为去噪手段,在输入端除了加入一层含有高斯白噪声的地表激发地表接收的多道地震信号外,另一层则添加深度加权信息以充分挖掘深部反射信号。在中间层,通过压缩通道提取数据特征并通过扩展通道还原数据细节信息,构成由含噪数据到去噪数据的非线性映射,最终输出去噪结果。该网络的数据集由随机生成的大量地质模型正演数据组成,在GPU环境下使用Pytorch进行训练,并将最终结果与传统F-X滤波结果进行对比。结果表明:从多人主观打分评价以及基于结构相似性指标和信噪比指标的客观评价结果来看,采用U-Net深度神经网络得到的数据去噪效果明显优于传统F-X滤波结果。

Study on Seismic Data Denoising Based on U-Net Deep Neural Network

During the process of energy and mineral resources mining,a large number of goaf after long-term mining not only threaten the safety of workers and equipment,but also restricts the economic of mines.At present,seismic prospecting is a common method to detect concealed goaf.However,during actual detection process,seismic data contains a large mount of random noise normally,which could have a great impact on data processing and imaging results.Therefore,it is necessary to suppress and remove random noise from seismic data.The improved U-Net neural network is taken as the denoising method for seismic data.Besides using multi-channel seismic signals with Gaussian white noise by surface excitation and surface reception,another layer adds depth weighting information to fully extract deep reflection signals.As for middle layer,the data characteristics are extracted by compression channel and data details are restored by extension channel,to construct a nonlinear mapping from noisly data to denoised data.Finally,the denoising result is output.The data set of the U-Net neural network is composed of a mass of seismic wave fields which generated by random geologic models,and the neural network is trained in the GPU environment by using Pytorch,besides that,the denoising results of the proposed methed is compared with the ones of traditional F-X filtering method.The result show that the performance of U-Net deep neural network is better than the traditional F-X filtering method,whatever from subjective perspective(multiplayer scoring)or objective perspective(structural similarity index and SNR).