A DAS data multi-scale noise reduction method based on a globally-informed discriminator GAN

Abstract

The present invention relates to a method for multi-scale noise reduction of DAS data based on a global information discriminant GAN, belonging to the method for denoising DAS data in borehole seismic exploration. It includes the acquisition of actual DAS records in boreholes, the construction of a multi-scale global information GAN model, the construction of a training set for the multi-scale global information GAN, the training of the multi-scale global information GAN model, and the processing of on-site data of actual seismic data. The advantages are that a targeted global information discriminant generative adversarial network denoising model is designed according to the multi-scale characteristics of DAS data in borehole seismic exploration and the idea of adversarial training, integrating multi-scale feature extraction, adversarial training and global information fusion, making the optimization target space more inclined towards the discriminative network, effectively reducing various types of noise in seismic exploration DAS data, improving the signal-to-noise ratio of exploration DAS records, being beneficial to subsequent imaging, inversion and interpretation work, and having strong practicality for oil and gas resource exploration.

Description

Technical Field

[0001] The present invention belongs to the method for denoising DAS data in borehole seismic exploration, and particularly refers to a denoising method based on a multi-scale global information generative adversarial model for improving the signal-to-noise ratio of seismic DAS data under complex noise interference. Background Art

[0002] With the continuous increase in the demand for oil and gas resources and the continuous improvement of exploration and development levels, the geological environment faced by oil drilling is becoming increasingly complex, and the research on seismic exploration is gradually moving towards deepening, refinement, and intelligentization. As an important means of modern seismic exploration, the vertical seismic profile (VSP) must develop in the direction of "light and dense" with the improvement of the complexity of the wellbore environment and the continuous increase in accuracy requirements. Therefore, traditional electronic geophones can no longer meet the requirements of modern seismic exploration VSP data acquisition. The distributed acoustic sensing (DAS) technology measures the optical phase change in Rayleigh backscattered light by using the phase change of the scattered light signal caused by the optical fiber, and can convert any standard optical cable into a distributed sensor. In recent years, DAS sensors have gradually been applied to the field of seismic exploration due to their advantages such as low cost, high sensitivity, high precision, large array, and repeatable acquisition.

[0003] Affected by the surrounding complex environment and downhole operation equipment, such as the difficulty in achieving good coupling between the optical cable and the receiving surface, in terms of the technology itself, DAS uses the weak scattered light signal in the optical fiber to record seismic waves, and the received effective signal often shows as a "weak signal", and the signal-to-noise ratio decreases significantly with the increase of the propagation distance. Therefore, how to suppress the complex noise interference with strong energy, random distribution, and wide existence in DAS-VSP data and improve the signal-to-noise ratio of the obtained DAS data is a core issue related to the effective utilization of DAS data and the sustainable development of DAS technology.

[0004] The DAS-VSP data contains not only random noise, but also superimposed noise formed by various causes such as background noise, fading noise, horizontal noise, and checkerboard noise. Many prior-based noise reduction methods (such as band-pass filtering, F-X deconvolution, wavelet transform, curve transform, shear wave transform, empirical mode decomposition (EMD), and time-frequency peak filtering (TFPF), etc.) have limited suppression effects on DAS noise when the nature of the noise is unknown. In recent years, with the development of deep learning theory and the improvement of hardware computing power, deep learning neural networks can obtain more accurate intrinsic feature representations of the original data through multi-layer neural network models with super strong feature learning capabilities, showing their absolute advantages in signal processing. Among many network architectures, the generative adversarial idea proposed by Goodfellow et al. innovatively uses an adversarial training mechanism to train two neural networks, improving application efficiency and being widely used in many fields such as image generation, super-resolution reconstruction, and data augmentation.

[0005] Current seismic exploration data denoising methods have limited ability to reduce DAS noise with complex causes, few priors, and various types, resulting in low utilization rate of effective information in low signal-to-noise ratio DAS-VSP data. Summary of the Invention

[0006] The present invention provides a multi-scale noise reduction method for DAS data based on global information discriminant GAN to solve the problems of complex and unknown characteristics, reduction of various types of noise, and incomplete and inaccurate retention of global effective signal energy in current seismic exploration DAS-VSP data.

[0007] The technical solution adopted by the present invention includes the following steps:

[0008] 1) Acquisition of actual records of downhole DAS

[0009] 2) Construction of a multi-scale global information GAN model

[0010] The multi-scale global information GAN model for seismic exploration DAS data processing consists of a generator G DAS and a discriminator D DAS The two parts. G DAS uses an encoder-decoder structure, receives DAS data containing various types of noise and performs encoding processing, and is responsible for outputting a noise-free pure DAS signal; the discriminator D DAS through the limitation of the loss function, plays a game with the generation network G DAS to perform alternating iterative training to balance their respective networks;

[0011] 3) Construction of a training set for multi-scale global information GAN

[0012] The dataset for training the multi-scale global information GAN model consists of two parts: a pure DAS signal set and a DAS noise set. Since there are many types of DAS noise, the DAS noise set can be composed of a series of DAS noise subsets. During the training process, either a single noise subset can be used, or several noise subsets can be combined and used;

[0013] 4) Training of the multi-scale global information GAN model

[0014] 5) Field data processing of actual seismic data

[0015] Using the G of the multi-scale global information GAN after training DAS Perform denoising processing on the DAS data actually collected in the field, that is, use the noisy DAS actual record collected in the field as the input signal and send it into G DAS , and the output data of this network is the DAS seismic data after denoising the predicted record.

[0016] In step 1) of the present invention, the vertical seismic profile technology VSP is used. A distributed fiber optic acoustic sensing DAS device is arranged along the smooth wellbore in a vertical well, and a seismic source shot point is set on the ground to generate seismic waves. The data collected after the shot point excitation constitutes the actual record of DAS in the well.

[0017] Step 2) of the present invention includes:

[0018] (1) Construction of the generator G DAS

[0019] The generator G DAS is designed based on U-net. This encoder-decoder structure network is divided into a contracting path and an expanding path. The contracting path consists of a series of downsampling operation modules such as convolutional layers, batch normalization, activation functions, and max pooling to obtain relevant information of the input DAS data through feature extraction; the expanding path consists of a series of upsampling operation modules such as transposed convolutional layers, convolutional layers, batch normalization, and activation functions, mainly for the precise positioning of DAS information. Among them, multiple pooling layers help the network to identify multi-scale features of DAS data, which is beneficial for better distinguishing DAS noise from effective signals; the upsampling result is tuned with part of the output of the contracting path, which can promote the information flow between the large-scale features from upsampling and the same-scale features from the left contracting path, is beneficial for the multi-scale feature fusion between the original DAS data and the denoised data, and better retains the effective signals;

[0020] (2) Construction of the discriminator D DAS

[0021] The discriminator D DAS consists of an input module, an output module, and several M DASModule composition, where the input module consists of one layer of Conv and one layer of LeakyRelu; the output module consists of an Adaptive AvgPool layer, one layer of Conv, one layer of LeakyRelu, and one layer of fully connected layer FC; M DAS The M module consists of one layer of Conv, one layer of batch normalization layer BN, and one layer of LeakyRelu. The number of convolutional kernels in each M DAS module doubles in turn. At the same time, D DAS All convolutional layers of use strided convolution to replace spatial pooling, expand the receptive field, downsample the results of convolutional layer feature extraction to capture the abstract information of the input signal. By continuously extracting features from the input signal, the discriminator finally determines the specific category of the input data.

[0022] Step 3) in the present invention includes:

[0023] (1) Pure DAS signal set X

[0024] The formation structure information is obtained by analyzing the actual DAS seismic data. The Ricker wave closest to the main frequency of the effective signal is used. See Equation 1. The propagation of seismic waves in the formation is simulated by the wave equation and the finite difference method to obtain the corresponding forward modeling data of seismic exploration DAS.

[0025]

[0026] where A is the amplitude, f0 is the main frequency of the wavelet, the frequency range is 20Hz - 80Hz, t0 is the wavelet delay time. According to the DAS system layout parameters, and during the network training process, the sliding window parameters and the initial number of DAS records are expanded according to the hardware environment, network convergence speed, denoising efficiency, and denoising effect to ensure the completeness and generalization requirements of the training set.

[0027] (2) DAS noise set N

[0028] The DAS noise set N for the training of the global information discrimination generative adversarial network is mainly composed of six types of typical noises obtained from actual surveys, including the horizontal noise, checkerboard noise, wellbore environment background noise, and fading noise that can be obtained before the shot point excitation, and also including the coupling noise and optical long-period noise generated after the shot point excitation. The size of each type of noise data is the same as that of the pure DAS data block, and the quantity is slightly less.

[0029] (3) Noisy training set Y

[0030] From the pure DAS signal set X = {x i |i = 1, 2, … p} and the DAS noise set N = {n jTake a signal data x from each of {j = 1, 2, …, q} i (i ∈ {1, 2, …, p}) and a noise data n j (j ∈ {1, 2, …, q}), and construct noisy data y with different noise levels through Equation (2) i,j,m ;

[0031] y i,j,m = x i + m · n j (2)

[0032] where m is a noise level adjustment factor, m ∈ N(0, 5);

[0033] The noisy data set Y = {y i,j,m |i = 1, 2, …, p; j = 1, 2, …, q; m ∈ (0, 5]} and the paired data {x i , y i,j,m} can be used for the model training of the global information discrimination generative adversarial network.

[0034] In step 4) of the present invention, in order to enable the established multi-scale global information GAN network to have the ability to obtain high signal-to-noise ratio, high resolution, and high fidelity DAS signals from noisy input signals, and to more comprehensively retain the detailed information of DAS signals, a certain number of paired data {x i , y i,j,m} are used to train the multi-scale global information GAN network, implicitly realizing the end-to-end mapping between noisy signal input and denoised signal output;

[0035] (1) Discriminative loss and D DAS Network parameter update

[0036] Let the predicted signal output by the generative network G DAS be That is:

[0037]

[0038] G DAS (·) represents the generative network mapping, y i,j,m is the noisy signal constructed for the training set, and the discriminative loss Draw on the idea of Wasserstein distance. The Wasserstein distance measures the minimum value of the average distance required to move data from distribution P to distribution Q, that is, measures the minimum distance between the joint distributions of pure data and data to be discriminated to describe the difference between pure DAS signals and predicted DAS signals, avoiding the vanishing gradient when the discriminator saturates, aiming to improve G DASThe discrimination ability is as shown in Equation (5):

[0039]

[0040] where E[·] represents the data distribution of their respective functions, and G DAS (·), D DAS (·) represent the mapping of the generator network and the discriminator network respectively. Since this binary classifier has more superior smoothing characteristics than before, based on this, the parameters of each module of D DAS are updated, which can better distinguish the distribution differences between pure DAS data and predicted DAS data from a global perspective, thus providing more reliable training indicators for network model optimization.

[0041] (2) Generator loss and G DAS network parameter update

[0042] Generator loss includes content loss L MSE and improved adversarial loss L ADV , where L MSE can measure the mean square error of the prediction result itself, which is beneficial to output a predicted DAS signal with a higher signal-to-noise ratio; the adversarial loss L ADV includes both the Wasserstein distance between the distribution of the predicted DAS signal and the pure DAS signal, and adds a gradient penalty operator to improve the training stability, expands the optimization space of the discriminator, and significantly improves the ability of the multi-scale global information GAN to describe the structural features of the DAS signal. The specific form is:

[0043]

[0044] where λ0 is the penalty factor, ||·|| F represents the F-norm, λ1 and λ2 are the weights of the content loss and the adversarial loss respectively, and

[0045]

[0046] (3) Overall optimization objective and network alternating iteration

[0047] The overall optimization objective of the multi-scale global information GAN can be expressed as:

[0048]

[0049] During the network training process, the parameters of each module of G DAS are updated to minimize the objective function, and the parameters of each module of D DAS are updated to maximize the objective function. The two networks alternate and iterate, which not only improves G DASThe prediction ability for DAS signals also improves the discrimination ability of D DAS until the two reach dynamic equilibrium. At this time, the training of the multi-scale global information GAN model approaches completion, and the obtained G DAS can be used for noise reduction of DAS data.

[0050] The global information discriminative generative adversarial network constructed in the present invention utilizes a U-net encoding and decoding structure to extract and fuse DAS signal features at multiple scales for adversarial training. By defining the overall network optimization objective of the global information discriminative GAN for reducing complex multi-type noises in DAS data through the Wasserstein distance, content loss, and adversarial loss, etc., and training the global information discriminative generative adversarial network with a rich training set of DAS data and noises to obtain a robust representation of DAS signal features. It realizes the reconstruction of DAS signals from noisy data, improves the signal-to-noise ratio of seismic data, and restores the global information of field actual seismic exploration DAS data, improves the reliability of seismic exploration DAS data, is conducive to accurately inferring formation lithology and structure, and enables the DAS technology to be better applied to the exploration of formation fine structure and unconventional oil and gas resources.

[0051] Effectively reducing complex multiple types of noises in the data collected by the high-precision measurement system DAS is an important processing link for obtaining high-quality DAS seismic data, and is crucial for improving the rationality of geological formation interpretation and the efficiency of oil and gas exploitation. The solution proposed in the present invention can effectively suppress the various types of seismic exploration DAS noises with diverse characteristics and different origins, restore the information of the effective wave reflection wave with weak energy in the seismic exploration DAS data, is more conducive to the accurate extraction of subsequent information such as reflection amplitude, velocity, and frequency, effectively improves the accuracy of geological formation analysis, and is conducive to accurately estimating the oil and gas reserves and distribution range.

[0052] The advantages of the present invention are that a targeted global information discriminative generative adversarial network denoising model is designed according to the multi-scale characteristics and adversarial training ideas of borehole seismic exploration DAS data. The ideas of multi-scale feature extraction, adversarial training, and global information fusion are incorporated into the design schemes of multiple links such as training set construction, network structure design, and loss function construction, so that the optimization target space is more inclined towards the discriminative network, which not only expands the optimization space of the discriminative network but also can use global features to identify and extract effective DAS signals. The present invention can improve the problems of insufficient optimization objectives, unstable gradients, poor integrity of signal retention, and insufficient accuracy of existing seismic exploration data denoising algorithms, effectively reduce multiple types of noises in seismic exploration DAS data, improve the signal-to-noise ratio of exploration DAS records, is conducive to subsequent imaging, inversion, and interpretation work, and has strong practicability for oil and gas resource exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1This is the structural diagram of the multi-scale global information GAN model of the present invention, where the generator G DAS receives DAS data containing multiple types of noise and performs encoding and decoding processing, and is responsible for outputting a pure DAS signal without noise; the discriminator D DAS through the limitation of the loss function, plays a game with the generator network G DAS for learning, and the two are alternately iteratively trained to balance their respective networks;

[0054] Figure 2 This is the generator G DAS structural diagram based on the U-Net network of the present invention, which is divided into an encoder and a decoder. The encoder is the contraction path, and the signal extracts features of each scale layer by layer through a downsampling module composed of a convolutional layer, a normalization layer, a linear unit, etc.; the decoder is the expansion path, and uses an upsampling module with different parameters to fuse the features between scales to achieve signal-to-noise separation and dimension recovery in the data. The network parameter settings of each module are shown in Table 1, where the convolution kernel size is set to 3×3;

[0055] Figure 3 This is the discriminator D DAS structural diagram of the multi-scale global information GAN of the present invention; the discriminator D DAS consists of an input module, an output module, and several M DAS modules;

[0056] Figure 4 This is the DAS-VSP data map actually obtained in seismic exploration. The acquisition process is as follows: DAS receivers are arranged in the well along the direction perpendicular to the ground. Excitation is selected at a position 200 meters from the wellhead on the ground. According to the requirements of the observed formation depth and resolution, the decoding distance width of the receiver is 1m, 6144 data each time, and the sampling frequency is 2500Hz; the DAS receivers receive reflected waves, direct waves, downward waves, and various types of noise caused by various factors from underground, forming a multi-channel DAS-VSP seismic exploration record, which can reflect the reflection and propagation of the wave field between formations;

[0057] Figure 5 This is a set of formation velocity model maps used to generate DAS data sets. By solving the elastic wave equation by the finite difference method, the propagation process of seismic waves in the actual well can be simulated, and a pure DAS data can be generated. In this example, a total of 10 sets of models are established. The triangles in the figure represent the seismic sources (shot points), and the thick vertical black lines represent the DAS sensors;

[0058] Figure 6 This is a map of 6 types of DAS noise data segments in the noise training set N. Among them, from left to right in the upper row are horizontal noise, fading noise, and wellbore environment background noise, and the noise types in the lower row are checkerboard noise, optical long-period noise, and coupling noise;

[0059] Figure 7 It is a paired data graph for cross - validating the performance of the multi - scale global information GAN denoising network. The left part is a pure DAS data that can be used as the ground truth of the denoising result. Its generation process is the same as that of the data training set, but with different parameters. The right part is the noisy paired data, and the noise in this data does not appear in the training set either;

[0060] Figure 8 It is to use the multi - scale global information GAN denoising network for Figure 7 processing the noisy data;

[0061] Figure 9 It is Figure 7 the difference between the noisy data and Figure 8 the denoising result, reflecting the noise components separated by the multi - scale global information GAN denoising network;

[0062] Figure 10 It is to use the multi - scale global information GAN denoising network for Figure 4 processing the actual DAS - VSP data to obtain the denoising result, where the parameters of each module in the network model are exactly the same as those in the forward - modeling data experiment;

[0063] Figure 11 It is the denoising result graph of processing Figure 4 the actual data using a band - pass filter. To effectively reduce various types of noise, the pass - band is set to 20Hz - 80Hz;

[0064] Figure 12 It is the denoising result of processing the actual Figure 4 data using the original GAN network, where the parameters are selected according to the main frequency range of the reflection signal and the optimal conditions. Specific implementation method

[0065] It includes the following steps:

[0066] 1), Obtaining the actual record of down - hole DAS

[0067] Using the vertical seismic profile technology (VSP), a distributed fiber optic acoustic sensor (DAS) device is arranged along the smooth wellbore in a vertical well with a depth of about 5000 meters. The decoding distance is 1 meter. A seismic source (shot point) is set on the ground to generate seismic waves, with an offset of 200 meters. The receiver accepts the time from 10ms to 4000ms. The data collected after the shot point excitation constitutes the actual record of down - hole DAS;

[0068] 2), Constructing the multi - scale global information GAN model

[0069] The multi-scale global information GAN model for seismic exploration DAS data processing consists of a generator G DAS (Generator Network) and Discriminator D DAS (Identification Network) consists of two parts, G DAS Receives DAS data containing multiple types of noise and performs encoding processing, and is responsible for outputting a noise-free pure DAS signal; the discriminator D DAS Through the restriction of loss function, and the generation network G DAS The two networks are trained alternately and iteratively to achieve a balance. In the network design part, according to the characteristics of DAS data and the design purpose of the present invention, the generator G DAS Using the encoder-decoder structure, merge the extended discriminator D DAS The data block size and depth are used to fuse and identify the multi-scale features of the signal, so that the trained G DAS Better maintain the spatial structure and global information of DAS data to achieve global and multi-scale denoising;

[0070] (1) Generator G DAS Construction

[0071] Generator G DAS Based on the U-net design, the codec structure network is divided into a contraction path and an expansion path. The contraction path consists of a series of downsampling operation modules such as convolution layer, batch normalization, activation function and maximum pooling, and obtains relevant information of the input DAS data through feature extraction; the expansion path consists of a series of upsampling operation modules such as deconvolution layer, convolution layer, batch normalization and activation function, which is mainly used for the precise positioning of DAS information. Among them, multiple pooling layers help the network to recognize the multi-scale features of DAS data, which is conducive to better distinguishing DAS noise from effective signals; the upsampling results are tuned with part of the output of the contraction path, which can promote the information flow between the large-scale features from the upsampling and the same-scale features from the left contraction path, which is conducive to the multi-scale feature fusion between the original DAS data and the denoised data, and better retains the effective signal.

[0072] (2) Discriminator D DAS Construction

[0073] Discriminator D DAS It consists of input module, output module and several M DAS The input module consists of a Conv layer and a LeakyRelu layer; the output module consists of an Adaptive AvgPool layer, a Conv layer, a LeakyRelu layer, and a fully connected layer FC; DASThe module consists of a layer of Conv, a layer of batch normalization layer BN, and a layer of LeakyRelu. To give full play to the global discrimination role of the discriminator and expand the receptive field of the discriminator, each M DAS The number of convolution kernels in the module doubles in turn. At the same time, D DAS All convolutional layers of use strided convolution to replace spatial pooling, and downsample the results of convolutional layer feature extraction to capture the abstract information of the input signal. By continuously extracting features from the input signal, the discriminator finally determines the specific category of the input data;

[0074] 3) Construction of the training set for the multi-scale global information GAN

[0075] The data set used to train the multi-scale global information GAN model includes two parts: a pure DAS signal set and a DAS noise set. Since there are many types of DAS noises, the DAS noise set can be composed of a series of DAS noise subsets. During the training process, either a single noise subset can be used, or several noise subsets can be superimposed and used;

[0076] (1) Pure DAS signal set X

[0077] By analyzing the actual DAS seismic data to obtain the formation structure information, using the Ricker wavelet (see Equation 1) closest to the main frequency of the effective signal, and simulating the propagation of seismic waves in the formation by the wave equation and the finite difference method, the corresponding forward DAS data for seismic exploration can be obtained;

[0078]

[0079] Among them, A is the amplitude, f0 is the main frequency of the wavelet, the frequency range is 20Hz - 80Hz, t0 is the wavelet delay time. According to the DAS system layout parameters, such as formation characteristics, the layout well depth of the DAS system, the distance between the seismic source and the wellhead, etc., more than twenty pure DAS signal sets X are constructed, each record has no less than a thousand channels, and each channel has at least 1 second of data volume (in the case of a sampling frequency of 2500Hz). Use a sliding window to intercept no less than ten thousand DAS data blocks from the data set, and during the network training process, expand the sliding window parameters and the initial DAS record quantity according to the hardware environment, network convergence speed, denoising efficiency, and denoising effect to ensure the completeness and generalization requirements of the training set;

[0080] (2) DAS noise set N

[0081] The DAS noise set N for global information discriminant generative adversarial network training consists of six types of typical noises mainly obtained from actual surveys, including horizontal noise, checkerboard noise, wellbore environmental background noise, and fading noise that can be obtained before shot excitation, and also including coupling noise and optical long-period noise generated after shot excitation. The size of each type of noise data is the same as that of the pure DAS data block, but the quantity is slightly less. In addition to being intercepted from actual records, some noise data can be obtained by simulating the wave equation according to the noise type, spatio-temporal characteristics, and generation mechanism.

[0082] (3) Noisy training set Y

[0083] From the pure DAS signal set X = {x i | i = 1, 2, … p} and the DAS noise set N = {n j | j = 1, 2, … q}, one signal data x i (i ∈ {1, 2, … p}) and one noise data n j (j ∈ {1, 2, … q}) are taken respectively, and the noisy data y i,j,m with different noise levels is constructed through Equation (2);

[0084] y i,j,m = x i + m·n j (2)

[0085] where m is the noise level adjustment factor, m ∈ N(0, 5];

[0086] The noisy data set Y = {y i,j,m | i = 1, 2, …, p; j = 1, 2, …, q; m ∈ (0, 5]} and the paired data {x i , y i,j,m} generated by Equation (2) can be used for the model training of the global information discriminant generative adversarial network;

[0087] 4) Multi-scale global information GAN model training

[0088] To enable the established multi-scale global information GAN network to have the ability to obtain high signal-to-noise ratio, high-resolution, and high-fidelity DAS signals from noisy input signals and more comprehensively retain the detailed information of DAS signals, a certain number of paired data {x i , y i,j,m} are used to train the multi-scale global information GAN network, implicitly realizing the end-to-end mapping between the noisy signal y i,j,m and the pure DAS signal x i ;

[0089] This design aims to improve the optimization space for the overall GAN network training. From a global perspective, it recovers the DAS signal affected by complex noise and proposes to establish a content loss L based on MSE MSE and an adversarial loss L DAS based on G DAS and D ADV to train the multi-scale global information GAN. This idea can better play the discrimination role of the discriminator and improve the limitation that the optimization target space is concentrated in the generation network;

[0090] (1) Discrimination loss and the update of the network parameters of D DAS Let the predicted signal output by the generation network G

[0091] be DAS That is

[0092]

[0093] G DAS (·) represents the mapping of the generation network, y i,j,m is the noisy signal constructed for the training set. The discrimination loss draws on the idea of the Wasserstein distance. The Wasserstein distance measures the minimum value of the average distance required to move the data from distribution P to distribution Q, that is, it measures the minimum distance of the joint distribution between the pure data and the data to be discriminated to describe the difference degree between the pure DAS signal and the predicted DAS signal, avoiding the gradient disappearance when the discriminator is saturated, aiming to improve the discrimination ability of G DAS as shown in Equation (5):

[0094]

[0095] where E[·] represents the data distribution of their respective functions, G DAS (·), D DAS (·) represent the mappings of the generation network and the discrimination network respectively. Since this binary classifier has more superior smoothing characteristics than before, based on this, the parameters of each module of D DAS are updated, and it can better distinguish the distribution differences between the pure DAS data and the predicted DAS data from a global perspective. Based on this, the parameters of each module of D DAS are updated, and it can better distinguish the distribution differences between the pure DAS data and the predicted DAS data from a global perspective, thus providing a more reliable training index for the optimization of the network model;

[0096] (2) Generation loss and the update of the network parameters of G DAS Generation loss

[0097] Generation loss including content loss L MSE and improved adversarial loss L ADV , where L MSE can measure the mean square error of the prediction result itself, which is beneficial to output a prediction DAS signal with a higher signal-to-noise ratio; the adversarial loss L ADV includes both the Wasserstein distance between the predicted DAS signal and the pure DAS signal distribution, and adds a gradient penalty operator to improve the training stability, expand the optimization space of the discriminator, and significantly improve the ability of the multi-scale global information GAN to describe the structural features of the DAS signal. The specific form is:

[0098]

[0099] where λ0 is the penalty factor, ||·|| F represents the F-norm, λ1 and λ2 are the weights of the content loss and the adversarial loss respectively, and

[0100]

[0101] (3) Overall optimization objective and network alternating iteration

[0102] The overall optimization objective of the multi-scale global information GAN can be expressed as:

[0103]

[0104] During the network training process, the parameters of each module of G DAS are updated to minimize the objective function, and the parameters of each module of D DAS are updated to maximize the objective function. The two networks alternate and iterate, which not only improves the prediction ability of G DAS for the DAS signal but also improves the discrimination ability of D DAS until the two reach a dynamic balance. At this time, the training of the multi-scale global information GAN model approaches completion, and the obtained G DAS can be used for noise reduction of DAS data;

[0105] 5) Field data processing of actual seismic data

[0106] Using G of the multi-scale global information GAN after training to DAS denoise the DAS data collected in the field, that is, taking the noisy DAS actual record collected in the field as the input signal and sending it into G DAS , and the output data of this network is the DAS seismic data after denoising the predicted record.

[0107] The following further illustrates the present invention through specific application examples.

[0108] The actual seismic exploration DAS-VSP record is a 2432*6144 DAS data collected in the Tahe area of the Tarim Basin, Xinjiang in 2018, as attached Figure 4 As shown, this seismic exploration record has 2432 channels, the sampling time is 0.0004 seconds, and each channel has 6144 data points. This seismic exploration DAS record contains a large number of background random noises, fading noises, coupling noises, long-period noises and other types of noises with different causes, resulting in the reflection event information being interfered, covered and submerged by the noises, and its amplitude, position and continuity are all damaged, affecting the reliability of subsequent information processing and imaging interpretation.

[0109] (1) Construct a multi-scale global information GAN model

[0110] According to factors such as hardware processing speed and memory environment, in this example, the network model based on the global information discriminant generative adversarial network is constructed as Figure 1 built, where the generator G DAS is designed based on U-net. This encoder-decoder network is divided into a contracting path and an expanding path, as Figure 2 shown. The contracting path consists of a series of downsampling operation modules such as convolutional layer (Conv), batch normalization layer (BN), activation function (Relu) and max pooling layer (Maxpool). During the downsampling process, the number of convolutional kernels doubles in turn, that is, 64, 128, 256, 512 and 1024 respectively; the expanding path consists of a series of upsampling operation modules such as deconvolutional layer (DeConv), convolutional layer (Conv), batch normalization layer (BN) and activation function (Tanh). During the upsampling process, the number of convolutional kernels shrinks in turn, that is, 512, 256, 128 and 64 respectively. Downsampling is used for multi-scale feature extraction to obtain more information, upsampling is used for precise positioning of DAS information, and to restore abstract features. The intermediate copy connection realizes multi-scale feature fusion of input and output. Finally, add a layer of Conv(1×1) and Tanh activation function for signal output. Table 1 gives the relevant parameters used in this example.

[0111] Table 1. Generator G DAS Network parameter settings

[0112] <![CDATA[Encoder G enc (Contraction path)]]> <![CDATA[Decoder G dec (Expansion 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

[0113] Discriminator D DAS consists of an input module, an output module and several M DAS modules, and the network structure is as Figure 3。The specific parameters used in this example are as follows: The input module consists of a layer of Conv(3×3) and a layer of LeakyRelu. The output module consists of an Adaptive AvgPool layer, a layer of Conv(1×1), a layer of LeakyRelu, and a fully connected layer FC; M DAS The M module consists of a layer of Conv(3×3), a layer of BN layer, and a layer of LeakyRelu. To give full play to the global discrimination of D DAS and expand its receptive field, in this example, the number of M DAS modules is 5, and the number of convolutional kernels in each M DAS module doubles in turn, namely 64, 128, 256, 512, and 1024 respectively. At the same time, all convolutional layers of D DAS use strided convolution to replace spatial pooling to expand the receptive field, downsample the results of convolutional layer feature extraction to capture the abstract information of the input signal. By continuously extracting features from the input DAS denoised signal, D DAS finally makes a determination on the specific category of the input data.

[0114] (2) Establish a multi-scale global information GAN data training set

[0115] Pure DAS signal set X: According to the DAS system layout parameters, such as formation characteristics, the layout well depth of the DAS system, the distance between the seismic source and the wellhead, etc., construct 10 pure DAS signal sets X, each record has no less than 1000 channels, and each channel has at least 1 second of data volume (in the case of a sampling frequency of 2500 Hz). Table 2 shows the selection range of the DAS system layout parameters in this example.

[0116] Table 2. DAS system layout and forward simulation formation parameters

[0117]

[0118] DAS noise set N:

[0119] The DAS noise set N for the training of the global information discrimination generative adversarial network is mainly composed of six types of typical noises obtained from actual surveys, including horizontal noise, checkerboard noise, wellbore environment background noise, and fading noise that can be obtained before the shot point excitation, and also including coupling noise and optical long-period noise generated after the shot point excitation. Four types of noises such as horizontal noise and fading noise can be taken from the data before the first arrival or obtained by simulating the wave equation based on the statistical characteristics and generation mechanisms of such noises in terms of frequency, time, space, etc. In this example, they are all taken from the noise data obtained before the first arrival; the coupling noise and optical long-period noise generated along with the signal are taken from the actual DAS records or obtained jointly by combining artificial synthesis methods in this example. In this example, 5 images of each type of noise are selected, and the size is the same as that of the data in the pure DAS signal set X. Figure 6 Figure 2 shows the data segments of six typical DAS noises in the selected downhole DAS noise set.

[0120] Noisy training set Y:

[0121] From the pure DAS signal set X = {x i | i = 1, 2, … p} and the DAS noise set N = {n j | j = 1, 2, … q}, a signal data x i (i ∈ {1, 2, … p}) and a noise data n j (j ∈ {1, 2, … q}) are respectively taken, and the noisy data set Y = {y i,j,m | i = 1, 2, …, p; j = 1, 2, …, q; m ∈ (0, 5]} is obtained through the noise level adjustment factor m. In this example, m ∈ N[1, 5], and x i from the pure DAS signal set and the corresponding y i,j,m from the noisy data set form paired data {x i , y i,j,m} for the model training of the global information discrimination generative adversarial network.

[0122] According to factors such as the hardware processing speed and software and memory environment requirements, in this example, a sliding window of size 128×128 is used to segment and normalize the above paired data {x i , y i,j,m}. After checking, a total of 20926 pairs of paired data of size 128×128 are actually used for the model training of the global information discrimination generative adversarial network in this example.

[0123] (3) Multi-scale global information GAN model training

[0124] Randomly assign initial values to the parameters of the multi-scale global information GAN model. Take every 32 groups out of 20,926 pairs of matching data as a batch and feed them into the model for training batch by batch. During the training process in this example, use formulas (4) and (5) to calculate the loss functions and Use the Adam optimizer and continuously correct the parameters of each layer of the network using the forward propagation and backpropagation algorithms to alternately update D DAS and G DAS , until the overall optimization objective of the multi-scale global information GAN in formula (9) tends to dynamic balance, indicating that the training of the multi-scale global information GAN model is completed. The obtained G DAS is the DAS denoising network we need. In this example, to balance the adversarial property and convergence speed, take the penalty factors λ0 = 10, λ1 = 1, and λ2 = 0.01, and the number of iterations is 200 times.

[0125] DAS Data Noise Reduction Based on Multi-Scale Global Information GAN

[0126] After the training of the multi-scale global information GAN is completed, perform cross-validation on its noise reduction performance. In this example, imitate the generation process of the training paired data, and adopt formation parameters that are completely different from the generation process of the training model. Newly generate 6 pairs of data with different sizes. Input the noisy records in the paired data into the trained G DAS network, and compare the output DAS data noise reduction results with the pure DAS data in the paired data to analyze the noise reduction performance of the multi-scale global information GAN. Figure 7 One pair of records is given, including a pure DAS data that can be used as the ground truth of the denoising result, and a paired noisy data that serves as the input to the multi-scale global information GAN denoising network. The denoising result of the multi-scale global information GAN in this example is as Figure 8 shown. It can be seen that the similarity with the Figure 7 pure DAS data is relatively high, and it can restore the positions, energies, and continuities of the main reflection axes. The weak converted waves and interlayer reflection energies that are completely submerged in various complex noises in the deep part are also restored to a certain extent. Figure 9 is Figure 7 the difference between the noisy data and the Figure 8 denoising result, reflecting the noise components separated by the multi-scale global information GAN denoising network.

[0127] To quantitatively measure the denoising performance of the multi-scale denoising network based on the global information discriminant GAN, use equation (10) to calculate the signal-to-noise ratio (SNR) of each DAS record,

[0128]

[0129] where u represents the pure DAS data, is the mean value of u, represents the DAS data after denoising. L and M respectively represent the number of channels and the number of sampling points per channel in the two-dimensional DAS record. Table 3 gives the SNR of 6 simulated records used for cross-validation before and after processing with the multi-scale global information GAN denoising network. It can be seen that the average SNR is increased by about 20 dB, which verifies the effectiveness of the multi-scale global information GAN denoising network.

[0130] Table 3 Comparison of SNR of denoising for 6 cross-validation DAS records

[0131]

[0132] Process the actual seismic data using the multi-scale global information GAN denoising network completed by the above training and validation Figure 4 The denoising result is as Figure 10 shown. The multi-scale global information GAN denoising network can thoroughly eliminate the pre-first-arrival noises such as fading noise, horizontal noise, and checkerboard noise. At the same time, it can also better suppress the optical noise and reveal most of the weak reflection signals covered under the optical long-period noise. The coupling noise at the right edge of the record is not completely eliminated. This type of noise has a considerable degree of similarity with the reflection signal characteristics. Therefore, the multi-scale global information GAN denoising network also has weak ability to distinguish the coupling noise. In addition to the method proposed in the present invention, we also processed the Figure 4 record using the most commonly used band-pass filtering method in geophysical exploration signals, as Figure 11 shown. Considering the frequency band range of the DAS signal, the selected passband is 20 Hz - 80 Hz. The band-pass filter can effectively remove the random noise and long-period optical noise, but it performs poorly in suppressing horizontal noise, checkerboard noise, and some fading noises, and even greatly attenuates the deep weak reflection signals.

[0133] In addition, Figure 12 the method of the original GAN is used to process the noise of the actual seismic exploration DAS record. Although we used a generative adversarial network with a depth similar to that of the multi-scale global information GAN network of the present invention and performed almost the same number of iterative trainings, the results show that although most of the effective signal spectra can be restored, there are still obvious residual horizontal noise and fading noise, and there is still obvious noise interference in the overall restoration of the reflection signal. By comparison, our multi-scale global information GAN denoising network performs particularly well in the global signal restoration ability and noise suppression when processing DAS data.

[0134] In summary, the multi-component noise removal method for seismic exploration DAS data based on the multi-scale global information GAN is effective and practical. It uses the U-net encoder-decoder structure and the multi-scale network architecture to extract the features of DAS data, and gradually realizes the feature extraction and noise reduction of DAS data through the adversarial training of the generator and the discriminator. It can not only remove the interference such as horizontal noise and random background noise in the actual records, but also restore the weak reflection signals covered by the optical long-period noise, protect the energy of various signals, and restore the continuity of the reflection axis. This is not only conducive to improving the signal-to-noise ratio of the actual seismic exploration DAS records, but also can guide accurate velocity inversion and imaging, which is beneficial to subsequent processing, interpretation, and accurately exploring the underground geological structure.

Claims

1. A multi-scale noise reduction method for DAS data based on a global information discriminant GAN, characterized in that Including the following steps: 1) Acquisition of the actual DAS records in the well 2) Construction of the multi-scale global information GAN model The multi-scale global information GAN model for seismic exploration DAS data processing consists of a generator G DAS and a discriminator D DAS The generator G DAS uses an encoder-decoder structure, receives DAS data containing multiple types of noise and performs encoding processing, and is responsible for outputting a pure DAS signal without noise; the discriminator D DAS through the limitation of the loss function, plays a game with the generator network G DAS for adversarial learning, and the two are alternately iteratively trained to balance their respective networks; 3) Construction of the training set for the multi-scale global information GAN The data set for training the multi-scale global information GAN model consists of two parts: a pure DAS signal set and a DAS noise set. Since there are many types of DAS noise, the DAS noise set can be composed of a series of DAS noise subsets. During the training process, either a single noise subset can be used, or several noise subsets can be superimposed and used. Among them: (1) Pure DAS signal set X By analyzing the actual DAS seismic data to obtain the formation structure information, using the Ricker wavelet closest to the main frequency of the effective signal, as shown in Equation 1, simulating the propagation of seismic waves in the formation by the wave equation and the finite difference method, so as to obtain the corresponding forward modeling data of seismic exploration DAS; Where A is the amplitude, f0 is the main frequency of the wavelet, the frequency range is 20Hz - 80Hz, t0 is the wavelet delay time. According to the layout parameters of the DAS system, and during the network training process, the sliding window parameters and the initial number of DAS records are expanded according to the hardware environment, network convergence speed, denoising efficiency and denoising effect to ensure the completeness and generalization requirements of the training set; (2) DAS noise set N The DAS noise set N for training the global information discriminative generative adversarial network is mainly composed of six types of typical noises obtained from actual surveys, including horizontal noise, checkerboard noise, wellbore environment background noise and fading noise that can be obtained before the shot point excitation, and also including coupling noise and optical long-period noise generated after the shot point excitation. The size of each type of noise data is the same as that of the pure DAS data block, and the quantity is slightly less; (3) Noisy training set Y From the pure DAS signal set X = {x i | i = 1, 2, … p} and the DAS noise set N = {n j | j = 1, 2, … q}, a signal data x i (i ∈ {1, 2, … p}) and a noise data n j (j ∈ {1, 2, … q}) are respectively taken, and the noisy data y with different noise levels is constructed through Equation (2) i,j,m ; y i,j,m = x i + m·n j (2) Where m is the noise level adjustment factor, m ∈ N(0, 5); The noisy dataset Y = {y i,j,m | i = 1, 2, …, p; j = 1, 2, …, q; m ∈ (0, 5]} and the paired data {x i , y i,j,m} can be used for the model training of the global information discriminant generative adversarial network; 4) Training of the multi-scale global information GAN model 5) On-site data processing of actual seismic data Using G of the multi-scale global information GAN after training is completed DAS Perform noise reduction processing on the DAS data actually collected in the field, that is, take the noisy DAS actual record collected in the field as the input signal and send it into G DAS The output data of this network is the DAS seismic data after noise reduction of the predicted record.

2. The multi-scale noise reduction method for DAS data based on global information discriminant GAN according to claim 1, wherein: In step 1), the vertical seismic profile technology VSP is used. A distributed fiber optic acoustic sensor DAS device is arranged along the smooth wellbore in a vertical well, and a seismic source shot point is set on the surface to excite seismic waves. The data collected after the shot point excitation constitutes the actual DAS records in the well.

3. A method for multi-scale noise reduction of DAS data based on a globally informed discriminative GAN according to claim 1, characterized in that: In step 2), it includes: (1) Generator G DAS Construction Generator G DAS Based on the U-net design, the encoder-decoder structure network is divided into a contracting path and an expanding path. The contracting path consists of a series of downsampling operation modules such as convolutional layers, batch normalization, activation functions, and max pooling, which are used to obtain relevant information of the input DAS data through feature extraction. The expanding path consists of a series of upsampling operation modules such as transposed convolutional layers, convolutional layers, batch normalization, and activation functions, which are mainly used for the precise positioning of DAS information. Among them, multiple pooling layers help the network to perform multi-scale recognition of DAS data features, which is beneficial to better distinguish DAS noise from effective signals. The upsampling results are tuned with partial outputs of the contracting path, which can promote the information flow between the large-scale features from upsampling and the same-scale features from the left contracting path, which is beneficial to the multi-scale feature fusion between the original DAS data and the denoised data, and better retains the effective signals. (2) Discriminator D DAS Construction Discriminator D DAS consists of an input module, an output module, and several M DAS modules. Among them, the input module consists of a layer of Conv and a layer of Leaky Rectified Linear Unit (LeakyRelu); the output module consists of an Adaptive Average Pooling (Adaptive AvgPool) layer, a layer of Conv, a layer of LeakyRelu, and a fully connected layer (FC); the M DAS module consists of a layer of Conv, a layer of Batch Normalization (BN), and a layer of LeakyRelu. The number of convolutional kernels in each M DAS module doubles sequentially. At the same time, all convolutional layers of D DAS use strided convolution to replace spatial pooling to expand the receptive field, downsample the results of convolutional layer feature extraction to capture the abstract information of the input signal, and by continuously extracting features from the input signal, the discriminator finally determines the specific category of the input data.

4. A multi-scale noise reduction method for DAS data based on a globally informed discriminator GAN according to claim 1, characterized in that: In step 4), to enable the established multi-scale global information GAN network to have the ability to obtain DAS signals with high signal-to-noise ratio, high resolution, and high fidelity from noisy input signals, and to more comprehensively retain the detailed information of DAS signals, a certain number of paired data {x i , y i,j,m} are used to train the multi-scale global information GAN network, implicitly realizing the end-to-end mapping between the input of noisy signals and the output of denoised signals; (1) Discriminative loss and D DAS Network parameter update Let the prediction signal output by the generation network G DAS be That is: G DAS (·) represents the generation network mapping, y i,j,m is the noisy signal constructed for the training set, and the discriminative loss Drawing on the idea of the Wasserstein distance, the Wasserstein distance measures the minimum value of the average distance required to move data from distribution P to distribution Q, that is, it measures the minimum distance between the joint distributions of the clean data and the data to be discriminated to describe the difference between the clean DAS signal and the predicted DAS signal, avoiding the vanishing gradient when the discriminator saturates, aiming to improve the discriminative ability of G DAS as shown in Equation (5): where E[·] represents the data distribution of their respective functions, G DAS (·), D DAS (·) represent the mapping of the generator network and the discriminator network respectively. Since the binary classifier has more superior smoothing characteristics than before, based on this, the parameters of each module of D DAS are updated, which can better distinguish the distribution differences between pure DAS data and predicted DAS data from a global perspective, thus providing more reliable training indicators for the optimization of the network model; (2) Generate loss and G DAS Update network parameters Generated loss including content loss L MSE and improved adversarial loss L ADV , where L MSE can measure the mean square error of the prediction result itself, which is beneficial to output a predicted DAS signal with a higher signal-to-noise ratio; the adversarial loss L ADV includes both the Wasserstein distance between the predicted DAS signal and the distribution of the pure DAS signal, and adds a gradient penalty operator to improve the training stability, expands the optimization space of the discriminator, and significantly improves the ability of the multi-scale global information GAN to describe the structural features of the DAS signal. The specific form is as follows: where λ0 is the penalty factor, and ||·|| F represents the F-norm, λ1 and λ2 are the weights of the content loss and adversarial loss respectively, and (3) Overall optimization objective and network alternating iteration The overall optimization objective of the multi-scale global information GAN can be expressed as: During network training, G DAS updates the parameters of each module to minimize the objective function, and D DAS updates the parameters of each module to maximize the objective function. The two networks alternate and iterate, which not only improves G DAS 's prediction ability for DAS signals but also improves D DAS 's discrimination ability until the two reach a dynamic balance. At this time, the training of the multi-scale global information GAN model approaches completion, and the obtained G DAS can be used for noise reduction of DAS data.

Images