基于多尺度增强级联残差网络的DAS地震资料背景噪声衰减方法

Background noise attenuation method of DAS seismic data based on multiscale enhanced cascade residual network

由于复杂强背景噪声的影响,分布式光纤声学传感(Distributed Optical Fiber Acoustic Sensing,DAS)采集的地震记录普遍信噪比较低。如何有效抑制背景噪声,恢复弱上行反射信息,切实提升DAS记录信噪比,已成为资料处理领域的热点问题之一。针对复杂DAS背景噪声消减问题,提出了一种多尺度增强级联残差网络(Multiscale Enhanced Cascade Residual Network,MECRN)。MECRN具有双路径级联残差网络结构,通过双路径机制提取DAS记录浅层信息。在此基础上,引入空洞卷积和多尺度模块提取DAS记录的多尺度特征,并通过跳跃连接导入浅层特征,在避免有效特征损失的同时,提升网络的特征提取能力。最后,通过残差学习整合局部和全局特征,并对重建特征细化,进一步提升了MECRN的去噪能力。模拟和实际DAS资料处理结果均表明,MECRN可以有效地压制DAS记录中的复杂背景噪声,准确恢复弱反射信号,显著提升处理DAS资料的能力。

Seismic records collected through distributed optical fiber acoustic sensing(DAS)typically exhibit a low signal-to-noise ratio(SNR)due to the pervasive influence of complex and intense background noise.How to effectively suppress background noise,restore weak upgoing reflection information,and substantially improve the SNR of DAS records havs become a prominent challenge in seismic data processing.To address the issue of complex DAS background noise attenuation,this paper proposes a multiscale enhanced cascade residual network(MECRN),which employs a dual-path cascade residual network structure to extract shallow information from DAS records.On this basis,dilated convolutional layers and multiscale modules are introduced to extract the multiscale features existing in DAS records.Additionally,skip connections are introduced to import shallow features,which enhances the feature extraction capability of MECRN and avoids effective feature loss.Finally,the local and global features are integrated by residual learning,and the reconstructed features are refined to improve the denoising capabilities of MECRN.The processing results from both simulated and field DAS data demonstrate that MECRN can effectively suppresses complex DAS background noise and accurately restores weak reflection signals,which enhances the processing capacity of DAS data significantly.