针对分布式光纤传感数据的残差注意力机制去噪方法研究OA

To address the limitations of traditional denoising methods for processing complex seismic data acquired via Distributed Acoustic Sensing(DAS),this paper proposes a deep learning-based denoising approach incorporating a residual attention mechanism.Traditional methods,such as the short-time Fourier transform and the wavelet transform,typically assume that noise is stationary or concentrated in specific frequency bands.However,background noise in real seismic data is often non-stationary,making traditional methods less effective under high-noise or complex conditions.To overcome these limitations,we propose the DAS-DnAttn model,which integrates a residual neural network(ResNet)with attention mechanisms.This model effectively captures noise features through residual learning and adaptively enhances useful signal features while suppressing complex background noise via channel and spatial attention mechanisms.In addition,the PhaseNet DAS tool is used to pick the P-and S-wave arrival times,providing a quantitative assessment of the proposed method's denoising performance and practical applicability.Experimental results demonstrate that the DAS-DnAttn model achieves superior denoising performance on synthetic data,real seismic records,and field-measured DAS earthquake data.Compared to traditional denoising methods(e.g.,DnCNN and BM3D),our approach significantly improves both signal-to-noise ratio(SNR)and peak signal-to-noise ratio(PSNR),while better preserving signal details.This study provides a more robust and effective deep learning solution for denoising DAS seismic data and holds substantial promise for enhancing the accuracy of seismic phase picking in practical applications.