Research On Deep Learning-based Geo-exploration Deep Reflections Enhancement And Multi-target Separation Of Borehole Seismic Data

From the mineral survey,detailed investigation,fine measurement to oil and gas field development stages,seismic exploration technology,as one of the applied geophysical technologies,has played an extremely important role.However,with the development of oil and gas resource exploration to deeper and unconventional fields,the seismic exploration records collected in the field often show the characteristics of weak effective reflections,strong noise level and complex noise wave fields,which makes it difficult to achieve high signal-to-noise ratio,high resolution and high fidelity.Therefore,effectively separating signal and noise in seismic data is one of the important technologies of seismic exploration.Deep learning(DL),as a new data-driven technology,has many advantages,such as high universality,no prior assumptions,and strong feature extraction capability,making it outstanding in attenuating seismic noise.The focus of this dissertation is to establish the research schemes of geo-exploration deep reflections enhancement and multi-target data separation in wells based on deep learning,including the processing method of surface seismic exploration data based on convolutional neural network(CNN)and generative adversarial net(GAN),and GAN-based multi-target data separation for borehole seismic data.Affected by the complex surface environment,deep targets and various styles of oil and gas formations caused by multiple tectonic actions.The quality of deep reflections collected by ground seismic exploration technology is generally low,mainly manifested in weak reflection signals,strong random noise and ground roll interference,which are difficult for subsequent data analysis and interpretation tasks.In contrast,borehole exploration works downhole in a quiet environment and the receiver is closer to the exploration targets.Therefore,the harsh surface environment has little impact on the quality of the collected data,and the collected effective signal energy is strong,which is an effective way to avoid the above-mentioned “strong noise level,weak signal” exploration problem.In recent years,distributed fiber optic acoustic sensing(DAS)technology has developed rapidly in borehole seismic exploration.The seismic data collected by this technology has high resolution and strong effective signal energy,which makes this technology have the potential to support high-resolution imaging.However,there is serious instrument interference in the DAS-VSP data,and its unique long-period noise,coupled noise and horizontal noise are strong,resulting in a new technical bottleneck in signal-noise separation tasks.Therefore,in view of the “strong noise level and weak reflection”characteristics of ground seismic exploration data,as well as a variety of new types of noise contained in DAS-VSP data.Based on the deep learning theory,this thesis proposes the research schemes of enhancing the effective reflections of ground exploration and separating multi-target data of borehole exploration.The main structure and research contributions of this dissertation are as follows:(1)Focusing on the “strong noise level” characteristic of ground seismic exploration data,this dissertation studies how to attenuate seismic noise without damaging the effective signal.Based on deep learning and sparse decomposition,a multimodal residual convolution neural network(MRCNN)is proposed.This thesis studies from the following three aspects: 1)Due to the false reflections caused by the “strong noise level” in DnCNN's denoised result,this thesis proposes to weaken the “strong noise” characteristics of seismic data by using variational modal decomposition(VMD)algorithm,so as to improve the accuracy of deep learning algorithm in distinguishing signal and noise;2)In the aspect of retaining effective signal energy,this dissertation proposes to retain all modal components of VMD algorithm by high channel(16 channels)training,so as to reduce the leakage of effective signal energy caused by abandoning modal components;3)In terms of efficiency,MRCNN algorithm introduces down-sampling operator and sub-pixel convolution to improve the training speed of the network.(2)Aiming at the “weak signal” of the ground seismic exploration data,research on the recovery of deep weak reflections is carried out,and an attribute-based double-constrained denoising network(Att-DCDN)is proposed.Conduct research: 1)functions of adversarial loss,reconstruction loss and classification loss in seismic noise attenuation and effective reflection recovery tasks,and build an attribute-guided double-constrained denoising network;2)research the construction method of attribute training set,put forward the concept of “weak label”,and build an over-complete seismic attribute training set suitable for denoising tasks;3)using residual loss to establish dual constraint training mode to improve the amplitude preservation ability of Att-DCDN algorithms.Furthermore,the robustness experiments are used to demonstrate that Att-DCDN can recover weak reflections at extremely low signal-to-noise ratios.In addition,by processing synthetic and actual seismic data,it is verified that the Att-DCDN can recover weak reflections that cannot be identified by traditional denoising algorithms(DnCNN algorithm,band-pass filter,etc.).Aiming at various new types of noise interference in DAS-VSP data,an attribute-guided target data separation network(Att-TDSN)is proposed to separate multiple target data(effective signal and some kinds of noise).This research innovatively proposes the concept of signal-noise-noise separation,that is,on the basis of the traditional signal-noise separation task,the noise-noise separation task is further realized.Specifically,this study extends the idea of“weak label” in the above Att-DCDN algorithm,and raises the two-dimensional weak labels in the Att-DCDN algorithm to multi-dimensional weak labels,where each dimension corresponds to a target data contained in the DAS-VSP data.Then,a training mode of “one-way matching and two-way constraint” is proposed to drive the update of network parameters.“One-way matching” can ensure the uniqueness of extraction results,and “two-way constraint” can ensure the complete separation of each kind of target data.This dissertation uses ablation experiments to prove the effectiveness and necessity of the “one-way matching and two-way constraint” training mode.In addition,a new method for constructing a training set with horizontal noise is proposed,which is very convenient and accurate,so that this method can be easily applied to this study and other studies.To sum up,this dissertation studies the method of attenuating noise and enhancing effective reflections for the deep reflection seismic data,and studies the signal-noise-noise separation method for the DAS-VSP data.The research results of this thesis provide technical support for many geophysical tasks such as finding oil and gas,reserve calculation,reservoir reconstruction,etc.