The Study Of Joint Processing Method Of Seismic Data From Multiple Sources Based On Convolutional Neural Networks

Seismic exploration is one of the important tools for resource development of oil and gas reservoirs.It can be divided into active source seismic exploration and passive source seismic exploration from the perspective of seismic source.With the increasing degree of oil and gas exploration and development,the focus of oil and gas exploration is gradually shifting to more tectonically complex areas,which adds more disadvantages to the development of active source seismic exploration.Due to the complex conditions of the surface and geology,active source seismic records acquired by conventional means may be accompanied by low signal-to-noise ratio,weak effective signal energy and missing reception channels,which will adversely affect the inversion,imaging and interpretation of subsequent seismic data.In recent years,passive-source seismic surveys using environmental noise data have received much attention because of their low cost,wide source of seismic sources,and high adaptability.Passive source seismic data contains rich subsurface medium information,and seismic interferometry can retrieve effective deterministic information from the chaotic and disordered noise data,which can provide multi-angle compensation for active source seismic exploration.Therefore,joint multi-source exploration has gradually become a popular research direction in the field of seismic exploration,showing good development prospects in the directions of offset imaging and velocity modeling.In order to adapt to the new situation,new methods and techniques have been introduced in the field of seismic exploration while continuously optimizing the existing technologies,and one of the hottest topics is deep learning.Deep learning is a driving technique based on a large amount of training data,which can efficiently process high-dimensional and high-noise seismic data,and is widely used in the field of seismic exploration because of its outstanding advantages.Convolutional neural network is one of the most representative feedforward neural networks in deep learning,which uses local connectivity and weight sharing to reduce the number of weights while the network model is simple enough to avoid the risk of overfitting,so it has been widely used in various fields.To address these problems,this paper designs four convolutional neural networks using deep learning techniques for the characteristics of seismic data and the correlation between multi-source seismic data,namely,active source first-break picking convolutional neural network,multi-source seismic data joint denoising convolutional neural network,multi-source seismic data joint reconstruction convolutional neural network,and multi-source seismic data joint low-frequency reconstructed convolutional neural network.The main findings of this paper are as follows:(1)A convolutional neural network-based method is proposed to pick up the first arrivals of large offset distance seismic data with strong noise from active sources.Compared with the traditional first-arrival pickup method,the first-breaking picking of seismic data with large offset distance and strong noise is significantly improved by using this convolutional neural network,especially when there is a large amount of strong noise interference,because of the weak energy of the effective signal at the far offset distance of deep reflection seismic data.In order to overcome the negative impact of strong noise on the effective signal of the far offset data,an attention mechanism module is added to the conventional convolutional neural network to increase the weight of the effective features by performing attention analysis on the input data before the feature extraction work,meaning the weight of the data near the boundary between the noise and the effective signal,which enhances the noise immunity of the network while improving the accuracy of the far offset first arrivals pickup.The convolutional neural network,after training and learning on part of the actual data,predicts the test data and obtains the first-break picking results with a good fit to the manually picked first-break picking results,and then performs the laminar inversion of the first-break picking results predicted by the network,industrially obtained first-break picking results and manually picked first-break picking results respectively,and compares the laminar speed model of the three,further proving that the excellent performance of the convolutional neural network has great potential for production applications.(2)A passive virtual source seismic data denoising method based on convolutional neural network is proposed.For the problem of strong interference noise and complex components of passive virtual source seismic data,a denoising convolutional neural network based on passive virtual source seismic data is constructed based on the similarity of the homophase axis features between the virtual source data and the active source data.The active source seismic data are used as labels to train the neural network,which can effectively suppress the interference noise in the passive virtual source seismic data and improve the effective signal resolution.After experimental tests on different labeled data sets,integrating labeled data with full-wavefield active source seismic data can better match the wavefield signals in passive virtual source data and improve the denoising effect and offset imaging quality.(3)A joint reconstruction method of multi-source seismic data based on twodimensional convolutional neural network is proposed.For the phenomenon of missing seismic data in active source seismic data due to different reasons,the effective information in passive virtual source data is used to compensate for the missing data in active sources.The network performs well in numerical simulation experiments under noise-free conditions,and accurately recovers the missing information in active source data while suppressing the background noise in passive virtual source data.On this basis,a joint wavefield reconstruction method for multisource seismic data based on multi-channel input single-pass output convolutional neural network is proposed.In order to overcome the influence of noise in active source seismic data on the reconstruction effect,the noisy active source missing seismic data is superimposed with passive virtual source data in channels,which becomes 3D data as the input of the convolutional neural network,and the output is still a single-channel complete active source seismic data.The network saves the huge workload of pre-processing data and can also be regarded as a denoising effect on the active source seismic data to a certain extent.The results obtained from the prediction of this convolutional neural network on the test data show the successful performance of the network on the simulated data.(4)A combined multi-source low-frequency reconstruction method based on convolutional neural network is proposed.Using the broad frequency characteristics of the passive source seismic data and combining the matching of the interferometric signal with the active source signal,the passive virtual source low-frequency seismic data is proposed as the input,and the high quality low-frequency data is output by the convolutional neural network.A strategy of tandem denoising network and lowfrequency reconstruction network is proposed to suppress the interference noise in the passive virtual source seismic data and improve the accuracy of predicted lowfrequency information.The use of primary scattered wavefield active source seismic data as label data is proposed,which can effectively suppress multiple waves on the free surface and improve the accuracy of full waveform inversion speed modeling.The use of predicted low-frequency data is proposed to provide an initial model for low-frequency deficient active source full-waveform inversion to improve modeling stability and accuracy.