DAS VSP Data Noise Suppression Based On Multi-scale Parallel Attention Network

The development of the petroleum industry and the practice of oil and gas exploration have fully proved that seismic exploration is an irreplaceable technical means to find underground oil and gas resources.Distributed optical fiber acoustic sensing(DAS)is an emerging sensing technology for seismic exploration data collection.Compared with traditional geophones,DAS has many advantages,such as high spatial and temporal resolution,wide measurement range,high efficiency and accuracy,low cost,high temperature and high pressure resistance,and has been widely used in vertical seismic profile(VSP)data collection in wells in recent years.However,the actual seismic data collected by the DAS array inevitably contains a large number of different types of background noise.These complex background noises with different characteristics make DAS data have a lower signal to noise ratio(SNR)than traditional detector data,and the data quality is seriously affected,which brings great difficulties to subsequent data processing and data interpretation.Some traditional noise reduction methods for seismic exploration data commonly used at home and abroad have achieved good results to a certain extent,but they have application limitations for noise reduction and weak effective recovery of DAS data in wells.Although these methods have certain effects on some/some DAS noises,they require some applicability assumptions or a large number of parameter tuning based on actual data.When some noises are reduced,there are serious effective energy losses and inefficiencies caused by large data volumes,and multiple noises cannot be effectively eliminated at the same time.In recent years,with the rapid development of artificial intelligence and big data technology,Deep Convolutional Neural Network(CNN)has been successfully extended to the field of seismic data denoising.Due to the mixed variety of complex noises in DAS exploration data,the morphological distribution and feature attributes are quite different,making feature extraction and separation of weakly reflected signals and various noises extremely difficult.Therefore,based on deep network technology,combined with multi-scale structure and attention mechanism,this master’s thesis designs and proposes a multiscale parallel attention residual network(PA-MRNet)method that can suppress the background noise of DAS in various wells and recover weak effective signals.The new method uses the idea of multi-scale to realize the sampling process of noisy DAS seismic data at different scales,extract its shallow features at different resolutions and different receptive fields,and then use the multi-scale residual module(AMRB)to gradually extract deep features.In each AMRB,the following sampling methods sample shallow features at different scales to expand the receptive field of the model.In addition,in different multi-scale convolution branches,parallel attention modules are designed to extract feature information in the channel dimension and spatial dimension of DAS seismic data at the same time.The channel attention module is used to extract channel dimension features and capture dependencies between channels,and the spatial attention module is used to extract spatial dimension features and focus on the most informative part of the features in space.Finally,features at different scales are resampled,and the parallel attention module is used for multi-scale feature fusion.The whole network adopts recursive residual structure design to simplify the network training process,and uses Charbonnier Loss to optimize.Both the synthetic DAS records and the experimental results of the actual DAS seismic data collected in the field show that the method on this master’s thesis can achieve high precision denoising of DAS data.Compared with the traditional methods of F-X deconvolution,singular value decomposition,variational modal decomposition,and deep learning methods ADNet,Att U-Net,RIDNet,and MIRNet,the PA-MRNet algorithm proposed in this master’s thesis has stronger denoising ability and weak signal recovery ability.