自适应动态滤波网络地震随机噪声压制方法

Seismic random noise suppression method by adaptive dynamic filtering network

由于地质及环境条件复杂,导致地震勘探采集资料信噪比相对较低,对后续的研究带来不利影响,因此地震勘探数据处理中对随机噪声的压制一直备受关注。现有算法无法较好实现对噪声的有效压制和对有效信号的极大保留,为此,将传统方法和深度学习相结合,提出了基于自适应动态滤波网络(Adaptive Dynamic Filtering Net,ADFNet)的方法压制地震资料中的随机噪声。该网络以编码器—解码器为架构,首先引入通道注意力机制(Attention Mechanism,AM)的思想,通过通道AM对空洞卷积多尺度数据特征集成,为网络提供了精准且丰富的特征表示;然后,引入动态卷积,以较低的计算复杂度实现对地震资料高频特征的学习,从而保留更丰富的细节信息。合成数据和实际数据的实验结果均表明,ADFNet可有效压制地震资料中的随机噪声,同时保留更丰富的地震数据细节,处理后的地震数据信噪比得到显著提升。

Due to the complex geological and environmental conditions,the signal‑to‑noise ratio of seismic data is relatively low,which has a negative impact on subsequent research.Therefore,the suppression of random noise in seismic data processing has been of great concern.The existing algorithms are unable to effectively suppress noise and preserve the effective signal.Therefore,this paper combines traditional methods with deep learning and puts forward a method based on an adaptive dynamic filtering network to suppress random noise in seismic data.The network is based on an encoder‑decoder architecture.Firstly,the idea of channel attention mechanism(AM)is introduced to realize the feature integration of multi‑scale data formed by dilated convolution through channel AM,providing accurate and rich feature representation for the network.Then,dynamic convolution is introduced to achieve the learning of high‑frequency features of seismic data with low computational complexity,so as to preserve more detailed information.The experimental results of both synthetic data and actual data show that the adaptive dynamic filtering network can effectively suppress random noise in seismic data while retaining richer details of seismic data,and the signal‑to‑noise ratio of seismic data after processing is significantly improved.