基于一种注意力机制U-Net的地震数据去噪方法

Seismic data de⁃noising method based on an attention mechanism U⁃Net

受野外采集过程中设备和环境等多种因素影响,地震数据中往往存在表面波、鬼波、随机噪声等各种噪声,影响了地震数据处理和解释的可靠性和准确性。近年来,基于人工智能的方法以其计算效率高、数值效果好等优点成为地震数据去噪的研究热点。U型网络(U-Net)是一种经典的卷积神经网络结构,常用于图像分割任务;注意力机制(Attention Mechanism,AM)是一种能够让模型在学习过程中更加关注特定区域或特征的技术。通过在U-Net网络中添加AM模块,构建了一种具有注意力功能的U型网络(AU-Net),并将其运用到地震数据去噪。为解决去噪过程中产生的边界效应,使用膨胀填充的方法对数据进行切分,该方法通用性较高,可以用于其他网络模型。AU-Net和U-Net的去噪试验结果表明:AU-Net网络去噪的效果比U-Net更好,可更好地保留弱信号;同时,通过迁移学习使AU-Net去噪方法更具适应性。

Due to various factors such as instruments,equipment,and environment during field acquisition,there often exist various types of noise in seismic data,including surface waves,ghost waves,random noise,etc.,affecting the reliability and accuracy of seismic data processing and interpretation.Recently,methods based on artificial intelligence have become a research hotspot in seismic data denoising,as they have high com⁃puting efficiency and good numerical effects.U⁃Net is a classic convolutional neural network structure com⁃monly used in image segmentation tasks.Attention mechanism(AM)is a technique that allows models to fo⁃cus more on specific regions or features during the learning process.This paper constructs a U⁃Net with atten⁃tion function by adding an AM module to the U⁃Net network and applies it to seismic data denoising.To ad⁃dress the boundary effects generated during the denoising process,the expansion filling method is used to seg⁃ment data.This method has strong universality and can be used for other network models.By comparing the de⁃noising effect of AU⁃Net and U⁃Net,it has been proved that AU⁃Net network has better denoising effect than that of the U⁃Net,which can better preserve weak signals.Meanwhile,AU⁃Net denoising method is more adap⁃table by transfer learning.