Seismic Data Enhancement And The Multi-domain Feature Mapping Based Deep Learning Denoising Algorithms

Recently,the imbalance of supply-and-demand for gas and oil reservoirs' has become a major problem.In order to exploit more resources,the seismic exploration must be deeper for the unconventional resources.As a typical unconventional resource area,the Tarim Desert is known for its ultra-deep underground resources,harsh exploration environment,and complex geological structure.Especially,the thick sand layer on the surface makes it difficult to excite the seismic source,meanwhile,the energy of the reflected signals are absorbed seriously.As a result,the seismic data are of low signal-to-noise ratio(SNR),which has a serious impact on data interpretation.In addition,the new detection technology-distributed acoustic sensor(DAS)also brings some challenges to seismic data processing.Although the DAS is of high temperature and pressure-resistance that can work well in the ultra-deep environment,the DAS seismic data are vulnerable to being polluted by a variety of optical noise caused by the working mechanism in the system.Therefore,DAS seismic data are also of low quality.In conclusion,the low SNR seismic data caused by the above unconventional exploration are the primary problem of resource exploitation.Therefore,it is a new challenge for the current seismic signal processing methods,but giving us a new opportunity for innovating an effective noise suppression method.How to improve the SNR of seismic data is the premise for visualizing the underground geological structure and exploring resources.However,traditional denoising methods are limited by the prior hypothesis and parameter tuning,which cannot meet the requirements of multi-type noise suppression.Therefore,there is an urgent need to develop a new,effective,and intelligent denoising method that is adapted for complex seismic data.Inspired by the technological revolution brought by artificial intelligence,advanced technologies such as deep learning should be introduced to seismic data processing.With the help of deep learning methods,it is expected to reduce the dependency of artificial experience and improve the reliability of Big Data-based analysis,so as to enhance the capacity for processing complex seismic data.Targeting the complex low SNR seismic data,this paper investigates various deep learning algorithms to improve the learning ability of neural networks and makes them match seismic data denoising.These researches will help to address the technical barriers between image learning and seismic data learning.Therefore,by focusing on the three elements of deep learning which are “data,algorithm,and computing capability”,our studies include noise modeling for intelligent data enhancement,developing a novel transformation learning strategy,and proposing a novel network architecture.Therefore,based on the massive data supplement,the neural networks can extract the physical features of seismic data and construct the multi-domain feature mappings for predicting noise.The above three works contribute to realizing the high-quality recovery of weak signals in the low SNR seismic data.Firstly,seismic data have the properties of high confidentiality,high-cost and regional changes,making the lack of training samples which challenges the application of deep learning.Also,seismic training data need manual labeling,aggravating the problem of training data lacking.In order to address the“data hunger”,the improved VAE(variational auto-encoder)is proposed to generate a massive number of simulated noise data for data enhancement.(1)The improved VAE uses the encoding network to represent the input as many hidden codes which are composed of two characteristic parameters.(2)The improved VAE forces the hidden codes approximate to the Mixed Gaussian Distribution instead of the Single Gaussian Distribution in the original VAE,ensuring the proposed “intelligent noise sources” have a similar random generation capability as the real sources.Therefore,a new KL(Kullback-Leibler)distance is derived to measure the similarity between the above two distributions;(3)By combining the KL distance and MSE(Mean-Squared Error)loss,a novel double-loss function is proposed for training a more optimized noise model,which can improve the coding accuracy and the simulated accuracy at the same time.In conclusion,the improved VAE noise model does not need manual labeling and does not have any confidentiality restrictions,which can be extended to noise modeling in different regions and with different characteristics,so as to address the data hunger thoroughly.Moreover,the noise data simulated by the improved VAE can be applied to the real seismic data processing,presenting the generalization and practical significance.Second,although the DAS is a transformative detective technology for unconventional seismic exploration,the DAS data contain high-intensity and multi-type optical noise that affects data interpretation and must be attenuated.Considering the various types of noise directly affect the accuracy of denoising mapping by the network,a novel transformation learning strategy-SSTCNN(Synchrosqueezing Transform Convolutional Neural Network)is proposed for DAS data denoising.(1)Benefiting the traditional signal processing method,synchrosqueezing transformation(SST)is used to obtain the time-frequency representations of the DAS data which are the input of the network;(2)Using the combination of convolution,normalization,and non-linear activation layers to extract the features,facilitate the flow of information and calculates the mapping mask from noisy data to noise in the time-frequency domain;(3)Applying the mapping mask to the input to obtain the time-frequency denoising results.Then,the final denoising results are obtained by inverse SST.In conclusion,the proposed transformation learning strategy can provide more unified and centralized prior information to the network,which has three advantages: avoiding overly redundant calculations,simplifying the training process of optimization,and reducing the need for massive training data to some extent.Compared with traditional denoising methods and general deep learning algorithms,the experiment results show that SSTCNN can effectively improve the SNR of the complex noisy data.Also,SSTCNN can help to improve the intelligent information service of the relevant field and provide technical support for oil and gas exploration.Finally,in order to extract more self features from the seismic data and solve the problem of constructing the signal training set without no pure signals.A multi-channel SVDDCNN(Singular Value Decomposition Denoising Convolutional Neural Network)is proposed to deeply mine the multi-mode features of DAS data,so as to further realize a better denoising mapping by the neural network.(1)The DAS data are decomposed by singular spectrum decomposition(SVD)into three subspace data of different modes.These decomposed data are used as the training data and input to the three-channel network;(2)A convolutional neural network with a single skip connection for residual learning is used to grasp the noise features from the noisy input data,training a model for“taking away” noise;(3)The network outputs three denoising subspace data,then adding them to obtain the rejected noise data.Overall,the multi-channel structure is to enlarge the network width and achieve the efficient supply of data.Meanwhile,the residual learning is for deepening the network,enhancing the efficiency of information propagation,and improving the accuracy of non-linear mapping.In addition,the high SNR subspace data are used as the signal training data in an innovative way,directly solving the lack of pure signal data and providing a new strategy for the construction of the signal training sets.The experiment results show that the proposed network can effectively improve the processing capability of the convolutional neural network for complex seismic data,and achieves a high-level signal recovery from low SNR data.