天然气工业2025，Vol.45Issue(4)：60-69,10.DOI:10.3787/j.issn.1000-0976.2025.04.005
基于轻量化网络和多域损失函数的随机噪声衰减方法
Random noise attenuation based on lightweight network and multi-domain loss function
摘要
Abstract
Random noise seriously affects the quality of seismic data,severely interfering seismic interpretation and inversion analysis,especially when the effective signal is relatively weak in the process of deep oil and gas exploration.In order to improve the signal to noise ratio(RSN)of seismic data,a lightweight network architecture with multi-scale feature extraction capability was designed,a parallel multi-scale large kernel convolution module was used to capture local features across scales,and the global feature associations were established on the basis of channel-space attention mechanism.Then the time-frequency domain joint optimization objective function was constructed,and time domain mean square error and frequency domain energy loss were balanced by adaptive weight coefficient,to effectively removing random noise while reducing effective signal loss.Next,following the data block training strategy,the large-scale seismic data was segmented into training sample sets that can be processed in parallel to improve the generalization ability of the model.Finally,a method combining lightweight network and multi-domain loss function was developed to remove random noise from seismic data.The following results are obtained.First,by optimizing the loss values in the time and frequency domains,the multi-domain loss function ensures that the integrity and local details of the original signal are protected to the maximum extent while noise is suppressed,and the RSN of the data is effectively improved.Second,the proposed method performs better in params(down about 14.29%),floating-point operations per second(FLOPS,down about 15%),and training duration(down about 40.88%),compared with feedforward denoising convolutional neural networks(DnCNN).Third,multi-scale parallel large-kernel convolution realizes the collaborative extraction of local features across scales by performing three types of parallel dilated convolutions,which can better handle complex seismic data.In conclusion,the proposed method is effective in removing random noise in seismic data,achieving low computing cost and faster training by optimizing the network structure,which is of great significance to effective seismic survey.关键词
地震资料处理/地震数据去噪/随机噪声/深度学习/多尺度/轻量化Key words
Seismic data processing/Seismic data denoising/Random noise/Deep learning/Multi-scale/Lightweight分类
地质学引用本文复制引用
陈伟,李安禹,李韵竹,未晛,张庆臣,金彦,魏龙海..基于轻量化网络和多域损失函数的随机噪声衰减方法[J].天然气工业,2025,45(4):60-69,10.基金项目
国家自然科学基金项目"基于经验模态分解的自由表面多次波衰减方法研究"(编号:41804140)、长江大学非常规油气省部共建协同创新中心开放基金项目"鄂西页岩储层地震弱信号智能检测方法研究"(编号:UOG2024-19)、武汉市科技计划重点项目"基于水平钻的隧道围岩性态全域探测技术与装备研发"(编号:2023010402010609)、深地国家科技重大专项项目"地球深部探测与矿产资源勘查"(编号:2024ZD1001003). （编号:41804140）
