大数据2026，Vol.12Issue(2)：111-128,18.DOI:10.11959/j.issn.2096-0271.2025070
基于深度卷积和自注意力机制的端到端地震波降噪方法
End-to-end seismic signals denoising via deep residual convolution and self-attention mechanisms
摘要
Abstract
Denoising of seismic waveform signals is crucial for seismic monitoring and seismological research.To this end,we propose an end-to-end deep learning method for denoising seismic waveforms.The method combines the deep convolutional network with the multi-head self-attention mechanism.We employ a residual encoder-decoder structure,which is particularly well-suited for processing signals with complex backgrounds and rich details.At the same time,the multi-head self-attention mechanism can capture long-range dependencies.By jointly constraining the model with a consistent correlation loss and a frequency-domain mean squared error loss,outstanding denoising performance is achieved in both the time and frequency domains.Evaluation of the publicly available dataset STEAD shows that our method outperforms traditional and existing deep learning methods in two key metrics:peak signal-to-noise ratio(PSNR)and signal correlation coefficient(CC),achieving a Pearson correlation of 0.918 and a PSNR of 36.79,reaching the state-of-the-art performance.Furthermore,we have further validated our method using the seismic waveform data recorded by the Beijing-Tianjin-Hebei earthquake early warning network,which started operating in 2021,and the results indicate that our approach can effectively suppress noise while better preserving seismic signals(achieving a CC of 0.70 and a PSNR of 35.26).关键词
地震波降噪/深度学习/卷积网络/自注意力机制/残差编解码器Key words
seismic signal denoising/deep learning/convolutional network/self-attention mechanism/residual encoder-decoder分类
信息技术与安全科学引用本文复制引用
赵博涛,王健宗,亢祖衡,贺亚运,彭俊清,张旭龙,瞿晓阳,谭毅培,陈雨乐,肖春光..基于深度卷积和自注意力机制的端到端地震波降噪方法[J].大数据,2026,12(2):111-128,18.基金项目
广东省重点领域研发计划"新一代人工智能"重大专项(No.2021B0101400003) The Key Research and Development Program of Guangdong Province(No.2021B0101400003) （No.2021B0101400003）
