CN115905805A - DAS data multi-scale noise reduction method based on global information judgment GAN - Google Patents

Abstract

The invention relates to a DAS data multi-scale noise reduction method based on global information judgment GAN, and belongs to a DAS data noise reduction method for borehole seismic exploration. The method comprises the steps of obtaining DAS actual records in a well, constructing a multi-scale global information GAN model, constructing a training set of the multi-scale global information GAN, training the multi-scale global information GAN model and processing actual seismic data on site. The method has the advantages that a targeted global information discrimination generation countermeasure network denoising model is designed according to the multi-scale features and the countermeasure training thought of the seismic exploration DAS data in the well, multi-scale feature extraction, countermeasure training and global information fusion are integrated, so that the optimized target space is inclined to a discrimination network more, various types of noise in the seismic exploration DAS data are effectively reduced, the signal to noise ratio of exploration DAS records is improved, the follow-up imaging, inversion and interpretation work is facilitated, and the method has strong practicability on oil and gas resource exploration.

Description

DAS data multi-scale noise reduction method based on global information judgment GAN

Technical Field

The invention belongs to a method for denoising borehole seismic exploration DAS data, and particularly relates to a denoising method for improving the signal-to-noise ratio of seismic DAS data under complex noise interference, wherein a countermeasure model is generated based on multi-scale global information.

Background

With the increasing demand for oil and gas resources and the increasing degree of exploration and development, the geological environment faced by petroleum drilling is more and more complex, and the research aiming at seismic exploration gradually advances towards deepening, refining and intellectualization. As the complexity of the borehole environment of a Vertical Seismic Profile (VSP) serving as an important means of modern seismic exploration is improved and the requirement on precision is continuously improved, the corresponding data acquisition technology must be developed towards the direction of light and density, so that the traditional electronic detector can not meet the requirement of the modern seismic exploration VSP data acquisition gradually. Distributed optical fiber acoustic sensing (DAS) technology measures optical phase changes in rayleigh backscattered light using fiber-induced phase changes in the scattered light signal, converting any standard optical cable into a distributed sensor. In recent years, DAS sensors have been increasingly used in the field of seismic exploration due to their advantages of low cost, high sensitivity, high accuracy, large arrays, and repeatable acquisition.

Due to the influence of surrounding complex environment and borehole operation equipment, such as difficulty in achieving good coupling between the optical cable and the receiving surface, actual DAS-VSP data obtained by domestic technology at the present stage usually contains higher noise level and complex and various noise types. In addition, the difference exists between the national DAS instrument precision index and the international advanced level; in the aspect of the technology, the DAS records seismic waves by using scattered light signals with weak energy in optical fibers, the received effective signals are often represented as 'weak signals', and the signal-to-noise ratio is remarkably reduced along with the increase of the propagation distance. Therefore, how to suppress the complex noise interference which is strong in energy, random in distribution and widely existed in the DAS-VSP data and improve the signal to noise ratio of the acquired DAS data is a core problem related to how to effectively utilize the DAS data and whether the DAS technology can be continuously developed.

DAS-VSP data not only contains random noise, but also superimposed noise formed by a plurality of causes such as background noise, fading noise, horizontal noise, chessboard noise and the like. Many a priori-based noise reduction methods (e.g., bandpass filtering, F-X deconvolution, wavelet transformation, curve transformation, shear wave transformation, empirical Mode Decomposition (EMD), and time-frequency peak filtering (TFPF)) have limited suppression of DAS noise in situations where the noise properties are unknown. In recent years, with the development of deep learning theory and the improvement of hardware computing capability, the deep learning neural network acquires more accurate intrinsic feature representation of original data by virtue of a multilayer neural network model with ultra-strong feature learning capability, and shows the absolute advantage of the deep learning neural network in signal processing. In a plurality of network architectures, the generative confrontation thought proposed by Goodfellow et al is innovatively trained on two neural networks by using a confrontation training mechanism, so that the application efficiency is improved, and the method is more applied to a plurality of fields such as image generation, super-resolution reconstruction and data enhancement.

The existing seismic exploration data denoising method has limited DAS noise reduction capability with complex cause, less prior and more types, so that the effective information utilization rate of the DAS-VSP data with low signal to noise ratio is low.

Disclosure of Invention

The invention provides a DAS data multi-scale noise reduction method based on global information judgment GAN, which is used for solving the problems of complex unknown characteristics, multi-type noise reduction, incomplete global effective signal energy retention and insufficient accuracy in the existing seismic exploration DAS-VSP data.

The technical scheme adopted by the invention is that the method comprises the following steps:

1) Acquisition of actual DAS records in a well

2) Construction of multi-scale global information GAN model

Multi-scale global information GAN model for seismic exploration DAS data processing is generated by generator G _DAS And discriminator D _DAS Two-part construction, G _DAS Using a coding and decoding structure, receiving DAS data containing various types of noise, performing coding processing, and outputting clean DAS signals without noise; discriminator D _DAS By limiting of loss functions, with generation of network G _DAS Carrying out game learning, and alternately carrying out iterative training on the game learning and the game learning to enable respective networks to reach balance;

3) Construction of training set of multi-scale global information GAN

The data set used for training the multi-scale global information GAN model comprises a pure DAS signal set and a DAS noise set, the DAS noise set can be composed of a series of DAS noise subsets due to the fact that the DAS noise is numerous in types, and a single noise subset can be used in the training process, and a plurality of noise subsets can be used in a superposition mode;

4) Multi-scale global information GAN model training

5) Actual seismic data field data processing

G utilizing multi-scale global information GAN after training is finished _DAS Carrying out denoising processing on DAS data actually acquired in the field, namely sending the DAS actual record containing noise acquired in the field into G as an input signal _DAS The output data of the network is the DAS seismic data after denoising the predicted record.

According to the invention, in the step 1), a vertical seismic profile technology VSP is used, distributed optical fiber acoustic wave sensing DAS equipment is distributed in a vertical well along a smooth well wall, seismic source shot points are arranged on the ground surface to excite seismic waves, and acquired data after shot point excitation form DAS actual records in the well.

The step 2) of the invention comprises the following steps:

(1) Generator G _DAS Construction of

Generator G _DAS Based on U-net design, the coding/decoding structure network branchThe method comprises the steps that a contraction path and an expansion path are adopted, the contraction path is composed of a series of downsampling operation modules such as a convolution layer, batch normalization, an activation function and a maximum pooling, and relevant information of input DAS data is obtained through feature extraction; the extended path is composed of a series of up-sampling operation modules such as an deconvolution layer, a convolution layer, batch normalization and an activation function, and is mainly used for accurately positioning DAS information, wherein the pooling layers are beneficial to multi-scale identification of DAS data characteristics by a network and better distinguishing DAS noise and effective signals; the up-sampling result is tuned with partial output of the contraction path, so that information flow between the up-sampled large-scale feature and the same-scale feature from the left contraction path can be promoted, multi-scale feature fusion between original DAS data and de-noising data is facilitated, and effective signals are better reserved;

(2) Discriminator D _DAS Construction of

Discriminator D _DAS Comprises an input module, an output module and a plurality of M _DAS The input module consists of a layer of Conv and a layer of modified linear unit LeakyRelu; the output module consists of an Adaptive average pooled Adaptive AvgPool layer, a Conv layer, a LeakyRelu layer and a full connection layer FC; m is a group of _DAS The module consists of a Conv layer, a batch normalization layer BN layer and a LeakyRelu layer, each M being _DAS The number of convolution kernels in the module is doubled in turn, and at the same time, D _DAS All the convolutional layers use step convolution to replace space pooling, the receptive field is enlarged, the convolutional layer feature extraction result is sampled down to capture the abstract information of the input signal, the features are continuously extracted from the input signal, and the discriminator finally judges the specific category of the input data.

The step 3) of the invention comprises the following steps:

(1) Pure DAS signal set X

Stratum structure information is obtained by analyzing actual DAS seismic data, and the propagation condition of seismic waves in a stratum is simulated by a wave equation and a finite difference method with a rake wavelet closest to the dominant frequency of an effective signal, see formula 1, so that corresponding DAS forward data for seismic exploration is obtained;

wherein A is amplitude, f ₀ The dominant frequency of the wavelet is in the frequency range of 20Hz-80Hz ₀ Laying parameters according to a DAS (data-to-analog converter) system, and expanding sliding window parameters and the number of initial DAS records according to a hardware environment, a network convergence speed, denoising efficiency and denoising effect in a network training process to ensure the completeness and generalization requirements of a training set;

(2) DAS noise set N

The DAS noise set N used for global information discrimination generation confrontation network training is mainly composed of six types of typical noises obtained by actual survey, wherein the noise set N comprises horizontal noise, chessboard noise, well surrounding environment background noise and fading noise which can be obtained before shot point excitation, and coupling noise and optical long period noise which are generated along with shot point excitation, and the data of various types of noise and the DAS data are same in size and slightly less in number;

(3) Noisy training set Y

From a clean DAS signal set X = { X = _i I =1,2, \ 8230p } and DAS noise set N = { N = { N } _j Respectively taking a signal data x from | j =1,2, \8230; q |) _i (i ∈ {1,2, \8230; p }) and a noisy data n _j (j epsilon {1,2, \8230; q }), and constructing noisy data y with different noise levels by the formula (2) _i,j,m ；

y _i,j,m ＝x _i +m·n _j (2)

Wherein m is a noise level adjustment factor, and m belongs to N (0, 5);

noisy data set Y = { Y) generated by equation (2) _i,j,m |i＝1,2,…,p；j＝1,2,…,q；m∈(0,5]And pairing data { x } _i ,y _i,j,m Can be used for global information discrimination to generate model training of the confrontation network.

In the step 4), in order to enable the established multi-scale global information GAN network to have the capability of acquiring DAS signals with high signal-to-noise ratio, high resolution and high fidelity from noisy input signals, the DA signals are more comprehensively reservedDetail information of S signal, using a certain amount of pairing data { x _i ,y _i,j,m Training a multi-scale global information GAN network, and implicitly realizing end-to-end mapping between noise-containing signal input and noise-removed signal output;

(1) Discriminating loss

And D _DAS Network parameter update

Routing generation network G _DAS The output prediction signal is

Namely:

G _DAS (. Represents generating a network map, y _i,j,m Identifying loss for noisy signals constructed for training set

By using the Wasserstein distance idea for reference, the Wasserstein distance measures the minimum value of the average distance which needs to be moved when the data is moved from the distribution P to the distribution Q, namely the minimum distance of the joint distribution between pure data and data to be identified is measured to describe the difference degree between the pure DAS signal and the predicted DAS signal, the disappearance of the gradient when the identifier is saturated is avoided, and the purpose of improving G is to _DAS The discrimination ability of (3) is represented by the formula (5):

wherein E [. C]Representing the distribution of data of respective functions therein, G _DAS (·)、D _DAS The (·) represents the generation network and the identification network mapping respectively, because the binary classifier has superior smooth characteristic compared with the prior art, the pair D is based on _DAS The parameters of each module are updated, and purities can be distinguished from the global perspectiveThe DAS data and the predicted DAS data are distributed differently, and therefore a more reliable training index is provided for network model optimization.

(2) Generating losses

And G _DAS Network parameter update->

Generating losses

Including content loss L _MSE And improved resistance to loss L _ADV Wherein L is _MSE The mean square error of the prediction result can be measured, and the DAS signal with higher signal-to-noise ratio can be output; against loss L _ADV The method not only comprises the steps of predicting Wasserstein distance between the DAS signal and the distribution of the pure DAS signal, but also increases a gradient penalty operator to improve training stability, expands optimization space of a discriminator, and obviously improves the capability of describing DAS signal structural features of multi-scale global information GAN, and the specific form is as follows:

wherein λ is ₀ As a penalty factor, | ·| luminance _F Denotes the F norm, λ ₁ 、λ ₂ Are respectively the weight of the content loss and the counter-loss, and

(3) Overall optimization objective and network alternating iteration

The overall optimization goal of the multi-scale global information GAN can be expressed as:

in the network training process, G _DAS Updating the parameters of the modules to minimize the objective function, D _DAS The parameters of each module are updated to maximize the objective function, and the two networks are alternately iterated, thereby improving G _DAS The prediction capability of DAS signals is improved _DAS Until the two reach dynamic balance, the multi-scale global information GAN model training is nearly completed, and the obtained G _DAS I.e., for noise reduction of the DAS data.

The constructed global information discrimination generation countermeasure network utilizes a U-net coding and decoding structure, multi-scale extraction and DAS signal characteristics are fused for countermeasure training, global information discrimination GAN overall network optimization targets for reducing complex and multi-type noise of DAS data are defined through Wassertein distance, content loss, countermeasure loss and the like, the global information discrimination generation countermeasure network is trained through constructing rich DAS data and a noise training set, and steady DAS signal characteristic representation is obtained. The method realizes the reconstruction of DAS signals from noisy data, improves the signal-to-noise ratio of seismic data and the global information recovery of the DAS data of field actual seismic exploration, improves the reliability of DAS data of seismic exploration, is favorable for accurately deducing the lithology and structure of the stratum, and enables the DAS technology to be better applied to exploration of fine structures of the stratum and exploration of unconventional oil and gas resources.

The method is an important processing link for obtaining high-quality DAS seismic data, and is of great importance for improving rationality of geological bottom interpretation, improving oil and gas exploitation efficiency and the like. The scheme provided by the invention can effectively suppress DAS noise of seismic exploration with various types, complex characteristics and different causes, recover the information of effective wave reflected waves with weak energy of DAS data of seismic exploration, is more favorable for accurately extracting information such as subsequent reflected amplitude, velocity and frequency, effectively improves the accuracy of geological formation analysis, and is favorable for accurately estimating the oil gas reserves and the distribution range.

The method has the advantages that a targeted global information discrimination generation countermeasure network denoising model is designed according to the multi-scale features and the countermeasure training thought of the seismic exploration DAS data in the well, and ideas such as multi-scale feature extraction, countermeasure training, global information fusion and the like are integrated into the design schemes of multiple links such as training set construction, network structure design, loss function construction and the like, so that the optimized target space is inclined to the identification network more, the optimization space of the identification network is expanded, and the DAS effective signals can be recognized and extracted by using the global features. The method can solve the problems of insufficient optimization target, unstable gradient, poor signal retention integrity, insufficient accuracy and the like of the existing data denoising algorithm in seismic exploration, effectively reduce various types of noise in the DAS data in seismic exploration, improve the signal-to-noise ratio of DAS record in exploration, is beneficial to subsequent imaging, inversion and interpretation work, and has strong practicability on oil-gas resource exploration.

Drawings

FIG. 1 is a diagram of a multi-scale global information GAN model structure of the present invention, in which a generator G _DAS Receiving DAS data containing various types of noise, performing coding and decoding processing, and outputting clean DAS signals without noise; discriminator D _DAS By limiting of loss functions, with generation of network G _DAS Carrying out game learning, and alternately carrying out iterative training to balance respective networks;

FIG. 2 is a generator G based on U-Net network of the present invention _DAS The structure diagram is divided into an encoder and a decoder, wherein the encoder is a contraction path, and signals pass through a down-sampling module consisting of a convolution layer, a normalization layer, a linear unit and the like to extract the characteristics of each scale layer by layer; the decoder, namely an extended path, fuses the characteristics among all scales by utilizing the upsampling modules with different parameters to realize the signal-noise separation and the dimensionality recovery in data, the network parameter settings of all the modules are shown in table 1, and the sizes of convolution kernels are all set to be 3 multiplied by 3;

FIG. 3 is a diagram of a multi-scale global information GAN discriminator D according to the present invention _DAS A structure diagram; discriminator D _DAS Comprises an input module, an output module and a plurality of M _DAS Module composition;

fig. 4 is a DAS-VSP data map actually acquired by seismic exploration, and the acquisition process is as follows: arranging DAS receivers in a well along a direction vertical to the ground, selecting positions 200 meters away from the well mouth on the ground for excitation, according to the requirements of the depth of an observed stratum and resolution, the decoding distance width of the receivers is 1m, 6144 data are obtained each time, and the sampling frequency is 2500Hz; the DAS receiver receives reflected waves, direct waves, downlink waves and various types of noise caused by various causes from the underground to form a plurality of DAS-VSP seismic exploration records which can reflect the reflection and propagation conditions of wave fields among stratums;

FIG. 5 is a set of stratum velocity model diagrams for generating DAS data sets, an elastic wave equation is solved through a finite difference method, the propagation process of seismic waves in an actual well can be simulated, and pure DAS data is generated, in the example, 10 sets of models are built, in the diagrams, a triangle represents a seismic source (shot point), and a vertical thick black line represents a DAS sensor;

fig. 6 is a data fragment diagram of DAS-type 6 noise data in the noise training set N, where the upper row of noise data is horizontal noise, fading noise, and background noise of the surrounding well environment from left to right, and the lower row of noise data is chessboard noise, optical long-period noise, and coupled noise;

fig. 7 is a diagram of paired data for cross-validation of performance of a multi-scale global information GAN denoising network, in which the left part is a pure DAS data that can be used as a true value (ground route) of a denoising result, the generation process is the same as the generation process of a data training set, but the parameters are different, the right part is paired data with noise, and noise in the data does not appear in the training set;

FIG. 8 is a diagram of the results of processing the noisy data of FIG. 7 using a multi-scale global information GAN denoising network;

FIG. 9 is a difference between the noisy data of FIG. 7 and the denoising result of FIG. 8, reflecting the noise components separated by the multi-scale global information GAN denoising network;

fig. 10 is a denoising result obtained by processing the actual DAS-VSP data of fig. 4 with the multi-scale global information GAN denoising network, where parameters of each module in the network model are completely consistent with the forward data experiment;

FIG. 11 is a graph of the de-noising results of the actual data of FIG. 4 processed with a band pass filter, the pass band being set to 20Hz-80Hz for effective multi-type noise reduction;

fig. 12 is a noise cancellation result of processing actual data of fig. 4 using an original GAN network, where parameters are selected according to dominant frequency range of a reflected signal and optimal conditions.

Detailed Description

Comprises the following steps:

1) DAS actual record acquisition in the well

Distributing distributed optical fiber acoustic wave sensing (DAS) equipment along a smooth well wall in a vertical well with the well depth of about 5000 meters by using a vertical seismic profiling technology (VSP), setting a seismic source (shot point) on the ground surface to excite seismic waves with the offset distance of 200 meters, and enabling a receiver to receive time of 10ms to 4000ms, wherein data acquired after shot point excitation form DAS actual records in the well;

2) Construction of multi-scale global information GAN model

Multi-scale global information GAN model for seismic exploration DAS data processing is generated by generator G _DAS (Generation network) and discriminator D _DAS (authentication network) two-part construction, G _DAS Receiving DAS data containing various types of noise, performing coding processing, and outputting clean DAS signals without noise; discriminator D _DAS By limiting of loss functions, with generation of network G _DAS And (5) carrying out game learning, and alternately carrying out iterative training to balance respective networks. A network design part, a generator G according to the characteristics of DAS data and the design purpose of the invention _DAS Merging an extension discriminator D using a codec structure _DAS The data block size and depth of the signal, and multi-scale feature fusion and identification are carried out on the signal so as to train out G _DAS The spatial structure and global information of DAS data are better kept, and the purpose of global and multi-scale noise elimination is achieved;

(1) Generator G _DAS Construction of

Generator G _DAS Based on U-net design, the productThe decoding structure network is divided into a contraction path and an expansion path, the contraction path is composed of a series of downsampling operation modules such as a convolution layer, batch normalization, an activation function and maximum pooling, and relevant information of input DAS data is obtained through feature extraction; the extended path is composed of a series of up-sampling operation modules such as an deconvolution layer, a convolution layer, batch normalization and an activation function, and is mainly used for accurately positioning DAS information, wherein the plurality of pooling layers are beneficial to multi-scale identification of DAS data characteristics by a network and better distinguishing DAS noise and effective signals; the up-sampling result is tuned with partial output of the contraction path, so that information flow between the large-scale feature from up-sampling and the same-scale feature from the left contraction path can be promoted, multi-scale feature fusion between original DAS data and de-noising data is facilitated, and effective signals are better reserved.

(2) Discriminator D _DAS Construction of

Discriminator D _DAS Comprises an input module, an output module and a plurality of M _DAS The input module consists of a layer of Conv and a layer of modified linear unit LeakyRelu; the output module consists of an Adaptive average pooled Adaptive AvgPool layer, a Conv layer, a LeakyRelu layer and a full connection layer FC; m is a group of _DAS The module consists of a Conv layer, a batch normalization layer BN layer and a LeakyRelu layer, and each M is used for fully playing the global identification function of the identifier and expanding the reception field of the identifier _DAS The number of convolution kernels in the module is doubled in turn, and at the same time, D _DAS All convolutional layers of (2) use step-by-step convolution to replace spatial pooling, and the results of convolutional layer feature extraction are downsampled to capture abstract information of an input signal. By continuously extracting features from the input signal, the discriminator finally determines the specific category of the input data;

3) Construction of training set of multi-scale global information GAN

The data set used for training the multi-scale global information GAN model comprises a pure DAS signal set and a DAS noise set, the DAS noise set can be composed of a series of DAS noise subsets due to the fact that the DAS noise is numerous in types, and a single noise subset can be used in the training process, and a plurality of noise subsets can be used in a superposition mode;

(1) Pure DAS signal set X

Stratum structure information is obtained through analyzing actual DAS seismic data, propagation conditions of seismic waves in a stratum are simulated by a wave equation and a finite difference method with a Ricker wavelet (see formula 1) closest to effective signal dominant frequency, and accordingly corresponding DAS forward data of seismic exploration are obtained;

wherein A is amplitude, f ₀ The dominant frequency of the wavelet is in the frequency range of 20Hz-80Hz ₀ Constructing more than twenty frames according to DAS layout parameters such as stratum characteristics, DAS layout well depth, DAS well-mouth distance and the like, wherein each frame is recorded with at least one thousand DAS signal sets X with at least 1 second data volume (sampling frequency is 2500 Hz), intercepting at least ten thousand DAS data blocks from the data sets by utilizing a sliding window, and expanding the sliding window parameters and the initial DAS record number according to hardware environment, network convergence speed, denoising efficiency and denoising effect in the network training process to ensure the completeness and generalization requirements of the training set;

(2) DAS noise set N

The DAS noise set N used for global information discrimination and generation of the confrontation network training is mainly composed of six types of typical noise obtained by actual survey, wherein the typical noise includes horizontal noise, chessboard noise, well surrounding environment background noise and fading noise which can be obtained before shot point excitation, coupling noise and optical long period noise which are concomitantly generated after shot point excitation, the noise data of various types and the DAS data block have the same size and slightly less quantity, and part of noise data can be obtained by wave equation simulation according to the noise type, space-time characteristics and generation mechanism except for being intercepted from actual records.

(3) Noisy training set Y

From a clean DAS signal set X = { X = _i I =1,2, \8230p } and DAS noise set N = { N = _j L j =1,2, \8230q } in each caseTaking a signal data x _i (i ∈ {1,2, \8230; p }) and a noisy data n _j (j ∈ {1,2, \8230; q }), noisy data y with different noise levels are constructed by equation (2) _i,j,m ；

y _i,j,m ＝x _i +m·n _j (2)

Wherein m is a noise level adjusting factor, m belongs to N (0, 5);

noisy data set Y = { Y) generated by equation (2) _i,j,m |i＝1,2,…,p；j＝1,2,…,q；m∈(0,5]And pairing data { x } _i ,y _i,j,m The method can be used for model training of global information discrimination generation confrontation network;

4) Multi-scale global information GAN model training

In order to enable the established multi-scale global information GAN network to have the capability of acquiring DAS signals with high signal-to-noise ratio, high resolution and high fidelity from noisy input signals and more comprehensively retain the detail information of the DAS signals, a certain amount of pairing data { x } is utilized _i ,y _i,j,m Training a multi-scale global information GAN network, and implicitly realizing a noisy signal y _i,j,m And the pure DAS signal x _i End-to-end mapping between;

the design aims to improve the optimization space of the whole GAN network training, recover the DAS signals interfered by the complex noise from the global angle, and provide the establishment of the content loss L based on the MSE _MSE And based on G _DAS 、D _DAS Against loss L _ADV The idea of training the multi-scale global information GAN can better play the identification role of the identifier and improve the limitation that the optimization target space is concentrated on generating the network;

(1) Discriminating loss

And D _DAS Network parameter update

Routing generation network G _DAS The output prediction signal is

Namely, it is

G _DAS (. Represents generating a network map, y _i,j,m Identifying loss for noisy signals constructed for training set

By using the Wasserstein distance concept, the Wasserstein distance measures the minimum value of the average distance which needs to be moved when the data is moved from the distribution P to the distribution Q, namely the minimum distance of the joint distribution of pure data and data to be identified is measured to describe the difference degree between the pure DAS signal and the predicted DAS signal, the disappearance of gradient when the identifier is saturated is avoided, and the purpose of improving G is to _DAS The discrimination ability of (2) is represented by the formula (5):

wherein E [. C]Data distribution representing respective functions therein, G _DAS (·)、D _DAS (. C) represents the generation network and the identification network respectively, because the binary classifier has better smooth characteristic than the past, based on the pair D _DAS Parameters of each module are updated, distribution differences of pure DAS data and predicted DAS data can be distinguished from the global perspective, and D is calculated based on the distribution differences _DAS Parameters of each module are updated, distribution differences of pure DAS data and predicted DAS data can be distinguished from the global perspective, and therefore reliable training indexes are provided for network model optimization;

(2) Generating losses

And G _DAS Network parameter update

Generating losses

Including content loss L _MSE And improved resistance to loss L _ADV Wherein L is _MSE The mean square error of the prediction result can be measured, and the DAS signal with higher signal-to-noise ratio can be output; against loss L _ADV The method not only comprises the steps of predicting Wasserstein distance between the DAS signal and the distribution of the pure DAS signal, but also increases a gradient penalty operator to improve training stability, expands the optimization space of a discriminator, and obviously improves the capability of describing DAS signal structural features by multi-scale global information GAN, and the specific form is as follows:

wherein λ ₀ As a penalty factor, | \ | charging _F Denotes the F norm, λ ₁ 、λ ₂ Are respectively the weight of the content loss and the counter loss, an

(3) Overall optimization objective and network alternating iteration

The overall optimization goal of the multi-scale global information GAN can be expressed as:

in the network training process, G _DAS Updating the parameters of the modules to minimize the objective function, D _DAS The parameters of each module are updated to maximize the objective function, and the two networks are alternately iterated, thereby improving G _DAS The prediction capability of DAS signals is improved at the same time _DAS Is judgedCapacity, until the two reach dynamic balance, at this time the multi-scale global information GAN model training is nearly completed, and the obtained G _DAS The method can be used for noise reduction of DAS data;

5) Actual seismic data field data processing

G utilizing multi-scale global information GAN after training _DAS Denoising DAS data actually acquired in the field, namely, noise-containing DAS actual records acquired in the field are sent to G as input signals _DAS The output data of the network is the DAS seismic data after denoising the predicted record.

The present invention is further illustrated by the following specific application examples.

The actual seismic survey DAS-VSP record is DAS data of a segment 2432 × 6144 collected in 2018 in the townrea basin tahe area of Xinjiang, and as shown in fig. 4, the seismic survey record has 2432 tracks in total, the sampling time is 0.0004 seconds, and each track has 6144 data points. The DAS record contains a large amount of noise with different types of causes, such as background random noise, fading noise, coupling noise, long period noise and the like, so that the reflected event information is covered and submerged by noise interference, the amplitude, position and continuity of the reflected event information are damaged, and the reliability of subsequent information processing and imaging interpretation is influenced.

(1) Construction of multi-scale global information GAN model

According to factors such as hardware processing speed and memory environment, a network model for generating the countermeasure network based on global information discrimination is built as shown in FIG. 1, wherein a generator G _DAS Based on the U-net design, the codec network is divided into a contraction path and an expansion path, as shown in fig. 2. The contraction path is composed of a series of downsampling operation modules such as a convolutional layer (Conv), a batch normalization layer (BN), an activation function (Relu) and a maximum pooling layer (Maxpool), and the number of convolutional kernels is doubled in sequence in the downsampling process, namely 64, 128, 256, 512 and 1024; the extended path is composed of a series of up-sampling operation modules such as a deconvolution layer (DeConv), a convolution layer (Conv), a batch normalization layer (BN) and an activation function (Tanh), and the number of convolution kernels is reduced in the up-sampling process, namely 512, 256, 128 and 64 respectively. Lower miningMulti-scale feature extraction is carried out to obtain more information, up-sampling is used for accurately positioning DAS information, abstract features are restored, input and output multi-scale feature fusion is realized through intermediate copy connection, and finally a layer of Conv (1 x 1) and Tanh activation function are added for signal output. Table 1 gives the relevant parameters used in this example.

TABLE 1 Generator G _DAS Network parameter setting

Encoder G enc (contracted route)	Decoder G dec (extended Path)
		Conv(64,3,1),BN,Relu,Max pool(2,2)	DeConv(3,1),Conv(512,3,1),BN,Relu
Conv(128,3,1),BN,Relu,Max pool(2,2)	DeConv(3,1),Conv(256,3,1),BN,Relu
		Conv(256,3,1),BN,Relu,Max pool(2,2)	DeConv(3,1),Conv(128,3,1),BN,Relu
Conv(512,3,1),BN,Relu,Max pool(2,2)	DeConv(3,1),Conv(64,3,1),BN,Relu
		Conv(1024,3,1),BN,Relu	Conv(1,1,1),Tanh

Discriminator D _DAS Comprises an input module, an output module and a plurality of M _DAS The module is formed, and the network structure is as shown in figure 3. The specific parameters used in this example are: the input module consists of a Conv (3 multiplied by 3) layer and a modified linear unit LeakyRelu, and the output module consists of an Adaptive average pooling (Adaptive AvgPool) layer, a Conv (1 multiplied by 1) layer, a modified linear unit LeakyRelu and a full connection layer FC; m is a group of _DAS The module consists of a Conv (3 × 3) layer, a BN layer and a LeakyRelu layer. To give full play to D _DAS Global discrimination and expansion of the receptive field of (1), in this case M _DAS The number of modules is 5, each M _DAS The number of convolutional cores in the module is doubled in turn, i.e., 64, 128, 256, 512, and 1024, respectively. At the same time, D _DAS All the convolutional layers use step convolution to replace space pooling, the receptive field is expanded, and the result of convolutional layer feature extraction is sampled down to capture abstract information of input signals. By continuously extracting features from the input DAS denoised signal, D _DAS A decision is ultimately made as to the particular category of input data.

(2) Establishing a multi-scale global information GAN data training set

Clean DAS signal set X: and constructing 10 pure DAS signal sets X, wherein each pure DAS signal set X is recorded at least 1000 times and has a data volume of at least 1 second (the sampling frequency is 2500 Hz) per time according to the layout parameters of the DAS system, such as the stratum characteristics, the layout well depth of the DAS system, the distance between a seismic source and a well head and the like. Table 2 shows the selected range of the DAS system layout parameters in this example.

TABLE 2 DAS System layout and forward modeling of formation parameters

DAS noise set N:

the DAS noise set N used for global information discrimination and generation of the antagonistic network training is mainly composed of six types of typical noise obtained by actual survey, wherein the typical noise comprises horizontal noise, chessboard noise, well surrounding environment background noise and fading noise which can be obtained before shot point excitation, and coupling noise and optical long-period noise which are concomitantly generated after shot point excitation. Four types of noise such as horizontal noise, fading noise and the like can be obtained from data before first arrival, and can also be obtained by wave equation simulation through the statistical characteristics and the generation mechanism of the noise such as frequency, time, space and the like, wherein the noise data obtained before first arrival are obtained in the example; the coupling noise, optical long period noise, etc. associated with the signals are taken from the actual DAS recordings in this example or obtained in combination with artificial synthesis. In this example, each noise is 5 pieces, and has the same size as the data in the clean DAS signal set X. Figure 6 shows data slices for six typical DAS noises in a selected DAS noise set in a well.

Noisy training set Y:

from a clean DAS signal set X = { X = _i I =1,2, \ 8230p } and DAS noise set N = { N = { N } _j Respectively taking a signal data x from | j =1,2, \8230; q |) _i (i ∈ {1,2, \8230; p }) and a noisy data n _j (j ∈ {1,2, \8230; q }), a noisy data set Y = { Y } is obtained by a noise level adjustment factor m _i,j,m |i＝1,2,…,p；j＝1,2,…,q；m∈(0,5]}. In this example m ∈ N [1, 5]]X from clean DAS signal set _i With corresponding y from noisy data sets _i,j,m Composition pairing data { x _i ,y _i,j,m And the global information is judged and generated to train the model of the confrontation network.

According to the requirements of hardware processing speed, software and memory environment and other factors, the embodiment uses a sliding window with the size of 128 x 128 to match the pairing data { x _i ,y _i,j,m Segmentation and normalization are carried out, and after the inspection, the matching data of 20926 pairs with the size of 128 multiplied by 128 are used for the practice of the training of the confrontation network model in the embodiment for global information discrimination.

(3) Multi-scale global information GAN model training

Randomly giving initial values of parameters of the multi-scale global information GAN model, and sending 20926 matched data in each 32 groups as one batch into the model to train. In the training process of this example, the loss functions are calculated by the formulas (4) and (5) respectively

And &>

D is alternately updated by using Adam optimizer and continuously modifying parameters of each layer of network by utilizing forward propagation and backward propagation algorithms _DAS And G _DAS Until the overall optimization target of the multi-scale global information GAN in the formula (9) tends to be dynamically balanced, marking that the training of the multi-scale global information GAN model is finished, and obtaining G _DAS I.e. the DAS denoising network we need. This example is to balance the antagonism and convergence rate, taking the penalty factor lambda ₀ ＝10、λ ₁ =1 and lambda ₂ =0.01, the number of iterations is 200. />

DAS data noise reduction based on multi-scale global information GAN

And after the multi-scale global information GAN training is finished, performing cross validation on the noise reduction performance of the multi-scale global information GAN. The example simulates the generation process of training paired data, and newly generates 6 pieces of paired data with different sizes by adopting stratum parameters completely different from the generation process of the training model. Inputting noise-containing records in paired data into trained G _DAS And the network compares the output DAS data noise reduction result with pure DAS data in the matched data, and analyzes the noise elimination performance of the multi-scale global information GAN. Fig. 7 shows 1 pair record, which contains a clean DAS data that can be used as a true value (ground route) of the denoising result, and a pair noisy data that is used as an input of the multi-scale global information GAN denoising network. In this example, as shown in fig. 8, the noise reduction result using the multi-scale global information GAN shows that the DAS data has a high similarity to the pure DAS data of fig. 7, and can recover the position, energy, and continuity of the main reflection axis, and the weak transition wave and the interlayer reflection energy, which are completely submerged in various complex noises, are also recovered to a certain extent. Fig. 9 is a difference between the noisy data of fig. 7 and the denoising result of fig. 8, which reflects the noise components separated by the multi-scale global information GAN denoising network.

In order to quantitatively measure the denoising performance of the multi-scale denoising network based on global information judgment GAN, the signal-to-noise ratio (SNR) recorded by each DAS is calculated by using the formula (10),

where u represents the clean DAS data,

is the mean value of u>

Representing the denoised DAS data, wherein L and M respectively represent the number of channels recorded by the two-dimensional DAS and the number of sampling points of each channel. Table 3 gives the SNR of 6 simulated records for cross-validation before and after processing using the multi-scale global information GAN noise cancellation network. It can be seen that the average boosted signal-to-noise ratio is about 20dB, which confirms the effectiveness of the multi-scale global information GAN denoising network.

TABLE 3 Cross-validation DAS record denoising SNR comparison

The actual seismic data in fig. 4 are processed by using the multi-scale global information GAN denoising network completed by the training and verification, and the denoising result is shown in fig. 10. The multi-scale global information GAN denoising network thoroughly eliminates the early-arrival noises such as fading noise, horizontal noise, chessboard noise and the like, simultaneously well inhibits optical noise, and exposes most weak reflection signals covered under optical long-period noise. The coupling noise at the right edge of the record is not completely eliminated, and the noise has a considerable similarity with the characteristics of the reflected signal, so that the multi-scale global information GAN denoising network has a weak capability of distinguishing the coupling noise. In addition to the method proposed by the present invention, we have processed the FIG. 4 recordings using the most commonly used band-pass filtering method for geophysical survey signals, as shown in FIG. 11. Considering the frequency band range of the DAS signals, the selected pass band is 20Hz-80Hz, the band-pass filter can effectively remove random noise and long-period optical noise, but the band-pass filter does not perform well in the aspects of inhibiting horizontal noise, chessboard noise and some fading noises, and even greatly attenuates deep weak reflection signals.

In addition, fig. 12 uses the original GAN method to process the noise recorded by the DAS, and although we use the generation countermeasure network with depth similar to that of the multi-scale global information GAN network of the present invention and perform iterative training for almost the same number of times, the result shows that although most of the effective signal spectrum can be recovered, there is still significant residual horizontal noise and fading noise, and the overall recovery of the reflected signal still has significant noise interference. By contrast, our multiscale global information GAN denoising network performs particularly well in terms of global signal recovery capability and noise suppression when processing DAS data.

In conclusion, the method for removing the multi-component noise of the seismic exploration DAS data based on the multi-scale global information GAN is effective and practical, the characteristics of the DAS data are extracted by utilizing the U-net coding and decoding structure and the multi-scale network system structure, the characteristic extraction and the noise elimination of the DAS data are gradually realized through the countertraining of the generator and the discriminator, the interference of horizontal noise, random background noise and the like in actual recording can be removed, weak reflection signals under the coverage of optical long-period noise can be recovered, the energy of various signals can be protected, and the continuity of a reflection axis can be recovered. The method is not only beneficial to improving the signal-to-noise ratio of the practical seismic exploration DAS record, but also can guide accurate speed inversion and imaging, and is beneficial to subsequent processing, explanation and accurate exploration of the underground geological structure.

Claims (5)

1. A DAS data multi-scale noise reduction method based on global information judgment GAN is characterized by comprising the following steps:

1) DAS actual record acquisition in the well

2) Construction of multi-scale global information GAN model

Multi-scale global information GAN model for seismic exploration DAS data processing is generated by generator G _DAS And discriminator D _DAS Two-part construction, G _DAS Using a coding and decoding structure to receive DAS data containing multiple types of noise and carry out coding processingOutputting a clean DAS signal without noise; discriminator D _DAS By limiting of loss functions, with generation of network G _DAS Carrying out game learning, and alternately carrying out iterative training to balance respective networks;

3) Construction of training set of multi-scale global information GAN

The data set used for training the multi-scale global information GAN model comprises a pure DAS signal set and a DAS noise set, the DAS noise set can be composed of a series of DAS noise subsets due to the fact that the DAS noise is numerous in types, and a single noise subset can be used in the training process, and a plurality of noise subsets can be used in a superposition mode;

4) Multi-scale global information GAN model training

5) Actual seismic data field data processing

G utilizing multi-scale global information GAN after training _DAS Denoising DAS data actually acquired in the field, namely, noise-containing DAS actual records acquired in the field are sent to G as input signals _DAS The output data of the network is the DAS seismic data after denoising the predicted record.

2. The DAS data multi-scale noise reduction method according to claim 1, wherein the DAS data multi-scale noise reduction method includes: in the step 1), a vertical seismic profile technology VSP is used, distributed optical fiber acoustic wave sensing DAS equipment is distributed in a vertical well along a smooth well wall, seismic source shot points are arranged on the ground surface to excite seismic waves, and acquired data after shot point excitation form DAS actual records in the well.

3. The DAS data multi-scale noise reduction method according to claim 1, wherein the DAS data multi-scale noise reduction method includes: the step 2) comprises the following steps:

(1) Generator G _DAS Construction of

Generator G _DAS Based on U-net design, the coding and decoding structure network is divided into a contraction path and an expansion path, wherein the contraction path is composed of a series of convolution layers, batch normalization, activation functions, maximum pooling and the likeThe down-sampling operation module is used for acquiring relevant information of input DAS data through feature extraction; the extended path is composed of a series of up-sampling operation modules such as an deconvolution layer, a convolution layer, batch normalization and an activation function, and is mainly used for accurately positioning DAS information, wherein the plurality of pooling layers are beneficial to multi-scale identification of DAS data characteristics by a network and better distinguishing DAS noise and effective signals; the up-sampling result is tuned with partial output of the contraction path, so that information flow between the large-scale feature from up-sampling and the same-scale feature from the left contraction path can be promoted, multi-scale feature fusion between original DAS data and de-noising data is facilitated, and effective signals are better reserved;

(2) Discriminator D _DAS Construction of

Discriminator D _DAS Comprises an input module, an output module and a plurality of M _DAS The input module consists of a layer of Conv and a layer of modified linear unit LeakyRelu; the output module consists of an adaptive average pooled adaptive AvgPool layer, a Conv layer, a LeakyRelu layer and a full connection FC layer; m _DAS The module consists of a Conv layer, a batch normalization layer BN layer and a LeakyRelu layer, each M being _DAS The number of convolution kernels in the module is doubled in turn, and at the same time, D _DAS All the convolutional layers use step convolution to replace space pooling, the receptive field is enlarged, the convolutional layer feature extraction result is sampled down to capture the abstract information of the input signal, the features are continuously extracted from the input signal, and the discriminator finally judges the specific category of the input data.

4. The DAS data multi-scale noise reduction method according to claim 1, wherein the DAS data multi-scale noise reduction method includes: the step 3) comprises the following steps:

(1) Pure DAS signal set X

Stratum structure information is obtained by analyzing actual DAS seismic data, and the propagation condition of seismic waves in a stratum is simulated by a wave equation and a finite difference method with a rake wavelet closest to the dominant frequency of an effective signal, see formula 1, so that corresponding DAS forward data for seismic exploration is obtained;

wherein A is amplitude, f ₀ Is wavelet main frequency, the frequency range is 20Hz-80Hz ₀ The wavelet delay time is a parameter laid according to the DAS, and in the network training process, the sliding window parameters and the initial DAS record number are expanded according to the hardware environment, the network convergence speed, the denoising efficiency and the denoising effect, so that the completeness and the generalization requirements of a training set are ensured;

(2) DAS noise set N

The DAS noise set N used for global information discrimination generation confrontation network training is mainly composed of six types of typical noise obtained by actual survey, wherein the typical noise includes horizontal noise, chessboard noise, well surrounding environment background noise and fading noise which can be obtained before shot point excitation, and also includes coupling noise and optical long period noise which are concomitantly generated after shot point excitation, and the data of various types of noise has the same size with the pure DAS data block and is slightly less in quantity;

(3) Noisy training set Y

From a clean DAS signal set X = { X = _i I =1,2, \ 8230p } and DAS noise set N = { N = { N } _j Respectively taking a signal data x from | j =1,2, \8230; q |) _i (i ∈ {1,2, \8230; p }) and a noisy data n _j (j epsilon {1,2, \8230; q }), and constructing noisy data y with different noise levels by the formula (2) _i,j,m ；

y _i,j,m ＝x _i +m·n _j (2)

Wherein m is a noise level adjustment factor, and m belongs to N (0, 5);

noisy data set Y = { Y) generated by equation (2) _i,j,m |i＝1,2,…,p；j＝1,2,…,q；m∈(0,5]And pairing data { x } _i ,y _i,j,m And the method can be used for global information discrimination to generate model training of the countermeasure network.

5. The method of claim 1, wherein the method comprises discriminating the DAS data multi-scale noise reduction based on global informationIs characterized in that: in the step 4), in order to enable the established multi-scale global information GAN network to have the capability of acquiring DAS signals with high signal-to-noise ratio, high resolution and high fidelity from noisy input signals, and more comprehensively retain the detail information of the DAS signals, a certain amount of pairing data { x ] is utilized _i ,y _i,j,m Training a multi-scale global information GAN network, and implicitly realizing end-to-end mapping between noise-containing signal input and noise-removed signal output;

(1) Discriminating loss

And D _DAS Network parameter update

Routing generation network G _DAS The output prediction signal is

Namely:

G _DAS (. Represents generating a network map, y _i,j,m Noise-containing signals constructed for training set, discriminating loss

By using the Wasserstein distance idea for reference, the Wasserstein distance measures the minimum value of the average distance which needs to be moved when the data is moved from the distribution P to the distribution Q, namely the minimum distance of the joint distribution between pure data and data to be identified is measured to describe the difference degree between the pure DAS signal and the predicted DAS signal, the disappearance of the gradient when the identifier is saturated is avoided, and the purpose of improving G is to _DAS The discrimination ability of (2) is represented by the formula (5): />

Wherein E [. C]Representing the respective functions thereinData distribution of G _DAS (·)、D _DAS (. C) represents the generation network and the identification network respectively, because the binary classifier has better smooth characteristic than the past, based on the pair D _DAS Parameters of each module are updated, distribution differences of pure DAS data and predicted DAS data can be distinguished from the global perspective, and therefore reliable training indexes are provided for network model optimization;

(2) Generating losses

And G _DAS Network parameter update

Generating losses

Including content loss L _MSE And improved resistance to loss L _ADV Wherein L is _MSE The prediction result can be subjected to mean square error measurement, so that the output of a predicted DAS signal with a higher signal-to-noise ratio is facilitated; against loss L _ADV The method not only comprises the steps of predicting Wasserstein distance between the DAS signal and the distribution of the pure DAS signal, but also increases a gradient penalty operator to improve training stability, expands optimization space of a discriminator, and obviously improves the capability of describing DAS signal structural features of multi-scale global information GAN, and the specific form is as follows:

wherein λ ₀ As a penalty factor, | \ | charging _F Denotes the F norm, λ ₁ 、λ ₂ Are respectively the weight of the content loss and the counter loss, an

(3) Overall optimization objective and network alternating iteration

The overall optimization goal of the multi-scale global information GAN can be expressed as:

in the network training process, G _DAS Updating the parameters of the modules to minimize the objective function, D _DAS The parameters of each module are updated to maximize the objective function, and the two networks are alternately iterated, thereby improving G _DAS The prediction capability of DAS signals is improved _DAS Until the two reach dynamic balance, the multi-scale global information GAN model training is nearly completed, and the obtained G _DAS I.e., for noise reduction of the DAS data.

Images