A well-drilled DAS noise suppression method based on subspace projection attention network

Abstract

This invention relates to a wellbore DAS noise suppression method based on a subspace projection attention network, belonging to the field of geophysical technology. It includes the establishment of the subspace projection attention network, the construction of a training set, the training of the subspace projection attention network, and DAS data denoising processing. The subspace projection attention network proposed in this invention uses projection for adaptive denoising because projection can naturally preserve the advantages of the local structure of the input signal, especially in heavily polluted or low-texture areas, which matches the characteristics of the acquired DAS data and is suitable for suppressing complex DAS noise. Denoising results from synthetic data and actual data both show that, compared with bandpass filters and feedforward denoising convolutional neural networks (DnCNN), the method of this invention can achieve more effective reduction of complex DAS noise and effective signal energy recovery, and the denoising process is simple and efficient.

Description

Technical Field

[0001] This invention belongs to the field of geophysical technology, and particularly relates to a well-drilled DAS noise suppression method based on subspace projection attention network. Background Technology

[0002] Seismic exploration is a crucial method for exploring oil, natural gas, and solid resources. Generally, seismic exploration includes three main stages: data acquisition, data processing, and profile comparison and interpretation. However, traditional geophones suffer from high costs, large channel spacing, and limited receiving bandwidth, making it impossible to obtain high-resolution exploration data to meticulously describe seismic wavefield and other information. In recent years, distributed fiber optic acoustic sensing (DAS) technology has received considerable attention, considered a new technology that can replace traditional geophones to obtain seismic data with higher sampling density. When seismic waves encounter optical fibers, they undergo slight strain, causing a change in the phase of the scattered light signal. Seismic data is then obtained through phase demodulation. Compared to traditional geophones, DAS is low-cost, heat-resistant, pressure-resistant, and corrosion-resistant, enabling long-distance continuous information acquisition and real-time measurement of multi-directional information (amplitude, frequency, and phase). Despite its many advantages, the actual seismic data acquired by DAS is often contaminated by complex background noise from the surrounding environment and low coupling between optical fibers, exhibiting characteristics of weak signals and strong interference. Besides random noise, DAS noise also includes coupling noise, background anomaly interference, fading noise, horizontal noise, and checkerboard noise, all of which adversely affect the interpretation and imaging of seismic data. How to effectively suppress various complex noises to improve the signal-to-noise ratio while clearly and completely recovering the weak effective signal has become an urgent problem to be solved in current DAS data processing.

[0003] To suppress various noises in seismic records, experts both domestically and internationally have proposed numerous solutions, such as bandpass filters, Wiener filters, FX deconvolution, Empirical Mode Decomposition (EMD), Variational Mode Decomposition (VMD), singular spectrum analysis, and fitting inversion methods. While these methods have achieved good results in some specific problems, they still have limitations. For example, bandpass filters are widely used in actual oilfield production, but they struggle to separate effective signals and noise in shared frequency bands; sparse representation methods such as curvelet transform and wavelet transform heavily rely on finding the optimal threshold for effectively separating signals and noise, but obtaining prior information to establish the optimal value is a challenge; EMD performs poorly in preserving signal amplitude when processing seismic data with high noise content and low signal-to-noise ratio; fitting inversion methods that can remove coupling noise require extensive noise analysis to obtain five parameters related to coupling noise: first fracture time, first fracture amplitude, noise period, attenuation, and average wavelet. The selection of parameters for various methods requires significant time and effort, and the processing results for DAS data from different regions often vary, failing to meet the demands for efficient and intelligent data processing.

[0004] In recent years, deep learning technology has been widely applied in the field of seismic exploration, such as seismic fault detection, microseismic event detection, first-shock acquisition, and seismic data denoising. Particularly in seismic data denoising, deep learning-based denoising methods have demonstrated superior performance compared to traditional methods by learning noise distributions from large-scale pairwise datasets. Leveraging the availability of big data, deep learning technology can extract multivariate, nonlinear, and nonparametric deep-level features from the data. Furthermore, a significant advantage is its ability to learn automatically without prior knowledge. For example, the feedforward denoising convolutional neural network (DnCNN) combines residual learning and batch normalization to achieve end-to-end denoising, establishing a mapping from noisy data to clean data and suppressing random and low-frequency noise in DAS data. While deep learning-based seismic data denoising methods have made considerable progress, they struggle to effectively suppress various complex background noise types in DAS data under challenging scenarios such as severe noise contamination or weak signals. Moreover, the recovered effective signal often falls short in terms of frequency and amplitude. Summary of the Invention

[0005] This invention provides a well-drilled DAS noise suppression method based on a subspace projection attention network to address the problem that in challenging scenarios such as severe noise pollution or weak signals, it is difficult to effectively suppress various types of complex background noise in DAS data, while the recovered effective signal is unsatisfactory in terms of frequency and amplitude.

[0006] The technical solution adopted by this invention includes the following steps:

[0007] (I) Establishment of Subspace Projection Attention Network

[0008] (1) Construction of the feature extraction part

[0009] After the signal input Y undergoes one residual convolutional block operation, it is fed into the downsampling block and the stacked residual convolutional blocks in the skip connections. The data sent to the downsampling block undergoes feature extraction through max pooling, outputting a feature map. The max pooling operation is performed by a 4×4 convolution with a stride of 2. Each feature map output from a downsampling block reduces the dimensionality of the data by half, removes redundant information, and becomes a higher-level feature map than the original input. This higher-level feature map is then passed to the next layer for subsequent higher-level feature extraction. The data after feature extraction from the stacked residual convolutional blocks becomes the low-level feature map F1. N N represents the layer number, which can be 1, 2, 3, or 4. The layers are sent to the SAM (Subspace Attention) module, which is also in a skip connection, to prepare for feature fusion. The bottom layer is connected by a residual convolutional block, and the highest-ranking feature map is extracted and shaped into F by the residual convolutional block. 5 'Sent to the feature fusion section;'

[0010] (2) Construction of the feature fusion part

[0011] High-level feature map F from the bottom 5 First, F2 is obtained by upsampling through 2×2 deconvolution. 4 The size is doubled, and then fed into the subspace attention module SAM, where the high-level feature map F2... 4 Compared with the previously fed low-level feature map F1 4 The feature map output after fusion is denoted as F. 4 The fused feature map F 4 Then compare it with the high-level feature map F from the lower layer. 5 'Connect along the channel dimension, then output as F through a residual convolution block.' 4 'Send it to the next layer, and the feature fusion parts of the following layers follow the same principle;'

[0012] Because it is necessary to better utilize local and global information in the input, feature F1 N F2 N ∈R H×W×C These components originate from different layers but are merged within the same layer. H and W represent height and width, and C represents the number of channels. They need to be the same size.

[0013] According to F1 N F2 N Estimate M basis vectors b1, b2, ..., b M Each b i ∈R H×W These are all basis vectors of the signal subspace. This process can be abstractly represented as:

[0014] f:(R HW×C ,R HW×C →R HW×M (1)

[0015] The process of generating the basis can be simply represented as:

[0016] B = f(F1) N F2 N (2)

[0017] Where B = [b1, b2, ..., b M ] is the signal subspace composed of basis vectors.

[0018] In the process of generating the signal subspace composed of basis vectors, it is necessary to first convert F1... N F2 N Connecting along the channel dimension yields F∈R H×W×2C F only represents the intermediate process of this operation step and has no practical meaning. Then it is fed into the residual convolution module with M output channels, and its output is reshaped into HW×M to generate an M-dimensional signal subspace, where the parameters are updated end-to-end during training.

[0019] In equation (2), the signal subspace B∈R is composed of the basis vectors obtained through f. H×W×M The low-level surface features F1 are transformed by orthogonal linear projection. N Projecting onto B, let P be the orthogonal projection matrix of the signal subspace. P can be obtained from B by the following operation:

[0020] P = B(B T B) -1 BT, (3)

[0021] Among them, the normalization term (B) T B) -1 This is necessary because the basis generation process does not guarantee that the basis vectors are orthogonal to each other;

[0022] Finally, in the high-level feature map F2 N Guided by F1, the low-level feature map is F1 N The projection is reconstructed into F in the signal subspace. N :

[0023] F N =PF1 N (4)

[0024] Compared to traditional UNet-like architectures that directly fuse low-level and high-level features, SA-UNet projects the low-level feature map using a subspace attention module to obtain the F-value before feature fusion. NThen it is fused with the high-level feature map obtained from upsampling;

[0025] (3) The output F of the last residual convolution block 1 The noise reduction result is obtained by subtracting the initial noisy input signal y from the 3×3 convolutional layer shaping, and finally outputting the denoised input signal.

[0026] (II) Construction of the training set

[0027] The dataset used to train the subspace projection attention network is called the training set. The training set includes a clean signal dataset s and a noisy signal dataset y, where y = s + n, and n is noise.

[0028] (1) Pure signal dataset

[0029] A large number of DAS clean signals were simulated using forward modeling, with the following specific parameter settings: 3-6 planar layers, each with a depth of 300-500m, a wave velocity of 1500-3300m / s, and a planar layer density of 1970-2245kg / m³. 3 The signal type is Ricker wavelet, the main frequency is 50-90Hz, the well-source distance is 100-200m, the sampling time interval is 1m, the recording fiber length is 2000-4000m, and the sampling frequency is 2500Hz. A 64×64 slider is used to extract K valid signal blocks from the clean record as a clean signal dataset. Based on the block size, K is greater than or equal to 20000.

[0030] (2) Noisy signal dataset

[0031] To construct a noisy signal dataset, various types of noise data samples were extracted from actual acquired DAS seismic signal records and superimposed on the clean signal to ensure that the proposed method can suppress various types of noise. In the end, K noisy signal data blocks of size 64×64 were obtained. The corresponding noisy signal dataset was obtained by randomly combining and adding the clean signal samples and the actual DAS noise, and the labels were consistent with the clean signal dataset, corresponding one-to-one, as a training sample pair.

[0032] (III) Training of Subspace Projection Attention Network

[0033] The actual noisy DAS seismic record should be a linear superposition of the clean signal s and the DAS noise n, as shown in the following equation:

[0034] y = s + n, (5)

[0035] Where s represents the desired clean signal, n represents complex DAS noise with varying phase and amplitude, including random noise, coupling noise, background anomaly interference, fading noise, horizontal noise, and checkerboard noise, and y represents the noisy DAS seismic record actually acquired in the field. The subspace projection attention network can be abstracted as a denoiser, which can establish an end-to-end mapping relationship from noisy signal data to clean signal data through continuous training, thereby achieving the removal of various noises at once and retaining only the desired effective part, as shown in the following formula:

[0036]

[0037] in Let y be the output signal data after denoising, y be the noisy signal data, SAUN represent the mapping relationship established by the denoising network, and θ abstractly represent the various parameters in the network. In order to optimize the mapping relationship, the network parameters θ are continuously optimized, thereby improving the denoising effect. To obtain a signal infinitely close to the pure signal s, a pair of records is randomly selected from the training dataset. The noisy signal data is used as the input to the subspace projection attention network, and the corresponding pure signal data is used as the expected output of the network. In order to suppress noise as much as possible without compromising the pure signal, the denoiser is trained by minimizing the following loss function:

[0038]

[0039] In the formula, s and y represent clean signal data and noisy signal data, respectively, and SAUN(y; θ) is the output of the denoiser with input y. At this time, s is the label of the denoiser, that is, the expected output of the denoiser when its input is y. Through repeated training and optimization with a large number of training samples, the parameters of the subspace projection attention network suitable for complex noise reduction in actual DAS are finally obtained. This is referred to as the obtained, fully trained model;

[0040] (iv) DAS Data Denoising Processing

[0041] The DAS data is denoised using a trained subspace projection attention network. The actual acquired seismic exploration DAS data is input into the subspace projection attention network, and the output of the network is the denoised DAS data.

[0042] In step (a) of the subspace projection attention network established in this invention, the construction of the feature fusion part (2) includes the construction of residual convolutional blocks, which includes:

[0043] The residual convolutional block is a basic component of the subspace projection attention network. It is built based on the residual network concept and includes at least one 3×3 convolutional layer and one LeakyReLU linear unit with leakage correction. The residual connection contains one 1×1 convolutional layer.

[0044] In step (a) of the establishment of the subspace projection attention network described in this invention, the construction of the feature fusion part (2) includes the following:

[0045] Use F1 N F2 represents the low-level feature map obtained at each layer. N Represents high-level feature maps, F N This represents the fused feature map output by the subspace attention module, where N represents the number of layers (1, 2, 3, 4), and F1. N and F2 N First, the features are connected along the channel dimension and fed into a residual convolutional module with M-dimensional output for shape reshaping. A small convolutional network generates the subspace basis, and finally, orthogonal linear projection is used to transform the features into F1 vectors. N Projecting the data into the generated subspace is the process of upsampling the high-level feature map F2. N Guided low-level feature map F1 N Self-attention projection.

[0046] The advantages of this invention are:

[0047] Previous methods have not been effective in processing DAS well exploration data, failing to effectively reduce various types of background noise in DAS seismic data, and inevitably resulting in loss of the recovered effective signal. This invention constructs a subspace projection attention network, whose basic structure employs a recursive residual design. A subspace attention module (SAM) is added to the skip connections, allowing low-level features to be projected under the guidance of high-level features before feature fusion, improving network performance without significantly increasing network learning complexity. A residual convolution module is constructed based on the recursive residual concept to address gradient explosion and vanishing gradient problems that may occur with increasing network depth. The LeakyReLU linear unit with leakage correction is used as the activation function of each convolutional layer, enabling more effective gradient descent and backpropagation, further optimizing network learning complexity. The subspace attention module, designed based on the projection mechanism and self-attention concept, naturally introduces global structural information into the denoising process through subspace basis generation and projection operations before fusing seismic signal data features at two different scales (low and high). This helps retain more detailed information, thereby recovering a cleaner signal with better quality. In summary, the subspace projection attention network trains a network capable of separating signal and noise by learning a set of reconstruction bases in the feature space. Noise denoising is achieved by projecting the input signal onto the selected signal subspace base. The DAS seismic data denoising method based on the subspace projection attention network effectively suppresses various complex background noises while minimizing effective signal energy attenuation. It achieves excellent and complete recovery of both shallow signals affected by strong noise and deep signals with weak energy, making it significant for DAS seismic signal processing and subsequent interpretation. Attached Figure Description

[0048] Figure 1 This is the overall architecture diagram of the subspace projection attention network of this invention;

[0049] Figure 2 This is a basic structural diagram of the residual convolution module of the present invention;

[0050] Figure 3 This is a basic structural diagram of the Subspace Attention Module (SAM) of this invention;

[0051] Figure 4(a) shows a portion of the clean signals in the training dataset, consisting of M blocks, each 64×64 in size;

[0052] Figure 4(b) shows a portion of the noise signal in the training dataset, consisting of M blocks, each 64×64 in size. The noisy signal obtained by randomly superimposing the clean signal and the noise signal must have the same label as the clean signal, and correspond one-to-one.

[0053] Figure 5(a) shows the clean DAS signal synthesized through forward modeling;

[0054] Figure 5(b) shows the actual data noise;

[0055] Figure 5(c) shows the synthesized noisy data with a signal-to-noise ratio of -3.09 dB;

[0056] Figure 6 This is a comparison chart of the denoising results of the synthesized recordings; where:

[0057] (a) is synthetic noisy data. It includes horizontal noise, fading noise, checkerboard noise, optical noise, and random noise;

[0058] (b) shows the result after bandpass filter processing. Horizontal noise and checkerboard noise remain significantly, the recovered signal frequency is reduced, and the signal-to-noise ratio is 7.73 dB.

[0059] (c) shows the result of DnCNN processing. Various background noises are suppressed to some extent, but the recovered signal amplitude is somewhat attenuated, and the signal-to-noise ratio is 11.98dB;

[0060] (d) shows the results of the iterative multi-scale attention network processing, with almost no residual background noise, and the recovered signal frequency and amplitude are well maintained, with a signal-to-noise ratio of 20.03dB.

[0061] (e) is the actual noise added to the synthetic record;

[0062] (f) is the noise separated by the bandpass filter, with obvious signal residue;

[0063] (g) represents the noise separated by DnCNN, which has a lot of residual noise and significant loss of effective signal energy.

[0064] (h) is the noise separated by the denoising network, with almost no signal residue. The separated noise is basically the same as the noise added in the synthesis record.

[0065] Figure 7 This is a comparison of the FK spectra of the results after adding noise and denoising using three different methods, where:

[0066] (a) FK spectrum of background noise added by DAS, (b) FK spectrum of noise separated by bandpass filter, with significant removal of out-of-band signal, (c) FK spectrum of noise separated by DnCNN, the matrix shows that the denoising process also removes the effective signal, (d) FK spectrum of noise separated by subspace projection attention network, with basically no effect on the effective signal in terms of frequency.

[0067] Figure 8 The image shows a comparison of the waveform of the 220th channel of the synthesized record. Among the three methods, the denoising result of this invention is basically the same as the pure signal waveform, and the signal amplitude is also the closest.

[0068] Figure 9 These are comparison images of actual denoising results, where:

[0069] (a) Actual noisy recording: The effective DAS signal is contaminated with a large amount of background noise, including horizontal noise, fading noise, optical noise and random noise;

[0070] (b) The denoising results of the actual recording were processed by a bandpass filter. Horizontal noise was clearly residual, while other noises were suppressed, and the frequency of the recovered signal was reduced.

[0071] (c) DnCNN processes the denoising results of the actual recording. Various background noises are suppressed to a certain extent, but the amplitude of the recovered signal is reduced.

[0072] (d) The denoising network of the present invention processes the actual recorded denoising results, and the background noise is almost completely reduced, while the frequency and amplitude of the recovered signal are not significantly attenuated.

[0073] Figure 10 These are actual recordings comparing noise removal process, where:

[0074] (a) The bandpass filter processes the noise separated from the actual recording, but the horizontal noise is not completely separated;

[0075] (b) DnCNN processes noise in the actual recording separation, and there is a relatively obvious signal residue;

[0076] (c) The noise reduction network of the present invention processes the noise separated from the actual recording, and the background noise is separated with basically no effective signal residue;

[0077] Figure 11 These are magnified comparison images of actual noise-removed areas;

[0078] (a) A magnified view of the actual DAS recording;

[0079] (b) The bandpass filter process locally amplifies the noise separated from the actual recording, and the blur indicates that the noise removal is mostly low frequency;

[0080] (c) DnCNN processes the noise in the actual recording separation, which is locally amplified, and most of the signal remains.

[0081] (d) The noise reduction network of the present invention processes the local amplification of noise in the actual recorded separation. Detailed Implementation

[0082] Includes the following steps:

[0083] (I) Establishment of Subspace Projection Attention Network

[0084] (1) Construction of the feature extraction part

[0085] After the signal input Y undergoes one residual convolutional block operation, it is fed into the downsampling block and the stacked residual convolutional blocks in the skip connections. The data sent to the downsampling block undergoes feature extraction through max pooling, outputting a feature map. The max pooling operation is performed by a 4×4 convolution with a stride of 2. Each feature map output from a downsampling block reduces the dimensionality of the data by half, removes redundant information, and becomes a higher-level feature map than the original input. This higher-level feature map is then passed to the next layer for subsequent higher-level feature extraction. The data after feature extraction from the stacked residual convolutional blocks becomes the low-level feature map F1. N N represents the layer number, which can be 1, 2, 3, or 4. The layers are sent to the SAM (Subspace Attention) module, which is also in a skip connection, to prepare for feature fusion. The bottom layer is connected by a residual convolutional block, and the highest-ranking feature map is extracted and shaped into F by the residual convolutional block. 5 'Sent to the feature fusion section;'

[0086] (2) Construction of the feature fusion part

[0087] High-level feature map F from the bottom 5 First, F2 is obtained by upsampling through 2×2 deconvolution. 4 The size is doubled, and then fed into the subspace attention module SAM, where the high-level feature map F2... 4 Compared with the previously fed low-level feature map F1 4 The feature map output after fusion is denoted as F. 4 The fused feature map F 4 Then compare it with the high-level feature map F from the lower layer. 5 'Connect along the channel dimension, then output as F through a residual convolution block.' 4 'Send it to the next layer, and the feature fusion parts of the following layers follow the same principle;'

[0088] ① Construct residual convolutional blocks

[0089] The residual convolutional block is a basic component of the subspace projection attention network. It is built based on the residual network concept and includes at least one 3×3 convolutional layer and one LeakyReLU linear unit with leakage correction. The residual connection contains one 1×1 convolutional layer.

[0090] ② Constructing the Subspace Attention Module (SAM)

[0091] When the input size is large, the required number of max-pooling layers will permanently lose some of the original information of the input, including the spatial relationships between the target and its surroundings. When the input size is small, only a small amount of local information can be seen, and the global information contained is insufficient. Therefore, the Subspace Attention Module (SAM) in the network is placed in each skip connection; using F1 NF2 represents the low-level feature map obtained at each layer. N Represents high-level feature maps, F N This represents the fused feature map output by the subspace attention module, where N represents the number of layers (1, 2, 3, 4), and F1. N and F2 N First, it is connected along the channel dimension and then fed into a residual convolution module with M-dimensional output for shape reshaping. Figure 3 Phase Two), in Figure 3 The third stage generates the subspace basis through a small convolutional network, and finally uses orthogonal linear projection to transform the features into F1. N Projecting the data into the generated subspace is the process of upsampling the high-level feature map F2. N Guided low-level feature map F1 N Self-attention projection.

[0092] Because it is necessary to better utilize local and global information in the input, feature F1 N F2 N ∈R H×W×C These components originate from different layers but are merged within the same layer. H and W represent height and width, and C represents the number of channels. They need to be the same size.

[0093] According to F1 N F2 N Estimate M basis vectors b1, b2, ..., b M Each b i ∈R H×W These are all basis vectors of the signal subspace. This process can be abstractly represented as:

[0094] f:(R HW×C ,R HW×C →R HW×M (1) The process of generating the basis can be simply represented as:

[0095] B = f(F1) N F2 N (2)

[0096] Where B = [b1, b2, ..., b M ] is the signal subspace composed of basis vectors.

[0097] In the process of generating the signal subspace composed of basis vectors, it is necessary to first convert F1... N F2 N Connecting along the channel dimension yields F∈R H×W×2CF only represents the intermediate process of this operation step and has no practical meaning. Then it is fed into the residual convolution module with M output channels, and its output is reshaped into HW×M to generate an M-dimensional signal subspace, where the parameters are updated end-to-end during training.

[0098] In equation (2), the signal subspace B∈R is composed of the basis vectors obtained through f. H×W×M The low-level surface features F1 are transformed by orthogonal linear projection. N Projecting onto B, let P be the orthogonal projection matrix of the signal subspace. P can be obtained from B by the following operation:

[0099] P = B(BTB) -1 BT, (3)

[0100] Among them, the normalization term (BTB) -1 This is necessary because the basis generation process does not guarantee that the basis vectors are orthogonal to each other;

[0101] Finally, in the high-level feature map F2 N Guided by F1, the low-level feature map is F1 N The projection is reconstructed into F in the signal subspace. N :

[0102] F N =PF1 N (4)

[0103] Compared to traditional UNet-like architectures that directly fuse low-level and high-level features, SA-UNet projects the low-level feature map using a subspace attention module to obtain the F-value before feature fusion. N Then it is fused with the high-level feature map obtained from upsampling;

[0104] (3) The output F of the last residual convolution block 1 The noise reduction result is obtained by subtracting the initial noisy input signal y from the 3×3 convolutional layer shaping, and finally outputting the denoised input signal.

[0105] (II) Construction of the training set

[0106] The dataset used to train the subspace projection attention network is called the training set. The training set includes a clean signal dataset s and a noisy signal dataset y, where y = s + n, and n is noise.

[0107] (1) Pure signal dataset

[0108] A large number of DAS clean signals were simulated using forward modeling, with the following specific parameter settings: 3-6 planar layers, each with a depth of 300-500m, a wave velocity of 1500-3300m / s, and a planar layer density of 1970-2245kg / m³. 3 The signal type is a Ricker wavelet, the main frequency is 50-90Hz, the well-source distance is 100-200m, the sampling time interval is 1m, the recording fiber length is 2000-4000m, and the sampling frequency is 2500Hz. A 64×64 slider is used to extract K valid signal blocks from the clean recording as a clean signal dataset. Based on the block size, K is greater than or equal to 20000. Figure 4a As shown;

[0109] (2) Noisy signal dataset

[0110] To construct a noisy signal dataset, various types of noise data samples were extracted from actual acquired DAS seismic signal records and superimposed on the clean signal. Figure 4b To ensure that the proposed method can suppress various types of noise, K noisy signal data blocks of size 64×64 are finally obtained. The corresponding noisy signal dataset is obtained by randomly combining and adding the clean signal sample and the actual DAS noise, and is labeled with the clean signal dataset, corresponding one-to-one, as a training sample pair.

[0111] (III) Training of Subspace Projection Attention Network

[0112] The actual noisy DAS seismic record should be a linear superposition of the clean signal s and the DAS noise n, as shown in the following equation:

[0113] y = s + n, (5)

[0114] Where s represents the desired clean signal, n represents complex DAS noise with varying phase and amplitude, including random noise, coupling noise, background anomaly interference, fading noise, horizontal noise, and checkerboard noise, and y represents the noisy DAS seismic record actually acquired in the field. The subspace projection attention network can be abstracted as a denoiser, which can establish an end-to-end mapping relationship from noisy signal data to clean signal data through continuous training, thereby achieving the removal of various noises at once and retaining only the desired effective part, as shown in the following formula:

[0115]

[0116] in Let y be the output signal data after denoising, y be the noisy signal data, SAUN represent the mapping relationship established by the denoising network, and θ abstractly represent the various parameters in the network. In order to optimize the mapping relationship, the network parameters θ are continuously optimized, thereby improving the denoising effect. To obtain a signal infinitely close to the pure signal s, a pair of records is randomly selected from the training dataset. The noisy signal data is used as the input to the subspace projection attention network, and the corresponding pure signal data is used as the expected output of the network. In order to suppress noise as much as possible without compromising the pure signal, the denoiser is trained by minimizing the following loss function:

[0117]

[0118] In the formula, s and y represent clean signal data and noisy signal data, respectively, and SAUN(y; θ) is the output of the denoiser with input y. At this time, s is the label of the denoiser, that is, the expected output of the denoiser when its input is y. Through repeated training and optimization with a large number of training samples, the parameters of the subspace projection attention network suitable for complex noise reduction in actual DAS are finally obtained. This is referred to as the obtained, fully trained model;

[0119] (iv) DAS Data Denoising Processing

[0120] The DAS data is denoised using a trained subspace projection attention network. The actual acquired seismic exploration DAS data is input into the subspace projection attention network, and the output of the network is the denoised DAS data.

[0121] To demonstrate the performance of the subspace projection attention network in this invention, we need to verify it using synthetic and actual acquired DAS data. Actual DAS data can be obtained directly, while synthetic records require modeling and simulation. The test datasets used in the experiment were all generated through modeling. First, we generated a clean signal as shown in Figure 5(a) by forward modeling a four-layer two-dimensional geological model. This forward model contains four planar layers, each with a depth of 300-500m. The P-wave velocities from the top to the bottom layers are 1500, 2200, 2800, and 3300 m / s, respectively, and the medium densities are 1970, 2086, 2200, and 2245 kg / m³, respectively. 3 The signal type is a Ricker wavelet with a dominant frequency of 60Hz; the well-source distance is 200m, the recording fiber length is 2048m, and the sampling frequency is 2500Hz. Background noise extracted from the actual DAS data is shown in Figure 5(b), including various types of noise such as horizontal noise, fading noise, optical noise, and checkerboard noise. The clean signal in Figure 5(a) and the actual noise in Figure 5(b) are added together to obtain the synthesized noisy data in Figure 5(c).

[0122] The effects of the present invention will be further illustrated below through experimental examples.

[0123] Experimental Example 1 Synthesis Record

[0124] To verify the effectiveness of the denoising network of this invention in processing DAS seismic data, we first denoised the synthetic noisy DAS seismic records and compared the results with those of a bandpass filter and a feedforward denoising convolutional neural network (DnCNN). Figure 6 (a) shows noisy DAS seismic data, which clearly contains various types of noise, including horizontal noise, fading noise, optical noise, checkerboard noise, and random noise. The effective signal is submerged, resulting in a low signal-to-noise ratio. Denoising was performed using a bandpass filter, DnCNN, and the denoising network of this invention, respectively. The results are shown below. Figure 6 As shown in (b)-(d). Figure 6 (e) represents the noise in the noisy record. The differences after denoising by the three methods are as follows: Figure 6 As shown in (f)-(h), the noise reduction results from the bandpass filter... Figure 6 (b) and the difference Figure 6 (f) It can be seen that the bandpass filter can suppress random noise, fading noise, and optical noise, but some details are lost, and it has no effect on horizontal noise and checkerboard noise. From Figure 6 As can be seen from the denoising results of DnCNN in (c) and the difference plot in 6(g), DnCNN, after being trained the same number of times as the method proposed in this paper, has a suppressive effect on various background noises in DAS seismic data. However, most of them remain, and the effective signal is severely attenuated. To achieve the same effect as the method proposed in this paper, Figure 6 (d) Further training is needed. In comparison, from Figure 6 (d) Noise reduction results of the method of the present invention and Figure 6 The difference plot (h) clearly shows that the new method can suppress all background noise to the greatest extent while completely preserving the effective signal. The uplink and downlink waves show almost no amplitude attenuation, and even extremely weak effective signals are preserved. This verifies the effective signal extraction capability and background noise suppression capability of the method of this invention.

[0125] To further illustrate the new method's ability to extract effective signals, Figure 7 Figures (a)-(d) show the FK spectra of the differences between the added noise, the bandpass filter, DnCNN, and the method of this invention, respectively. From the FK spectra, it can be seen that for the bandpass filter, although most of the low-frequency noise is suppressed, the noise removal within the passband is incomplete, and the effective signal outside the passband is significantly attenuated (Figure (b)). For DnCNN, the denoising effect is better than the bandpass filter, but it also removes the effective signal in each frequency band. Figure 7(c) The presence of a valid signal can be clearly observed within the rectangular box, indicating a weak ability to preserve the amplitude of the valid signal. In contrast, the noise removed by the method of this invention contains virtually no valid signal, proving that the recovered signal is closest to the clean data in frequency. Furthermore, to visually compare the amplitude preservation capabilities of different methods for the valid signal, Figure 8 The waveform comparison of channel 220 is presented, showing the results of clean data and denoising using three methods. Figure 8 This indicates that the signal waveform after bandpass filtering is not entirely consistent with the pure signal. DnCNN shows a significant reduction in amplitude at the signal peak, and the denoising result of the method in this invention best fits the pure data. The comparison of FK spectra and single-channel waveforms proves that the signal recovered by the method in this invention is closest to the pure signal in both frequency and amplitude.

[0126] Table 1 lists the signal-to-noise ratio (SNR) and mean squared error (MSE) of the denoising results from the three methods. The SNR recorded after processing with the method of this invention is 20.03 dB, which is much higher than that of the bandpass filter and DnCNN, with SNRs of 7.73 dB and 11.98 dB, respectively. At the same time, the MSE is an order of magnitude smaller than theirs.

[0127] Table 1 shows the signal-to-noise ratio (SNR) and mean squared error (MSE) of the three methods for denoising synthetic records.

[0128] method Noisy Recording bandpass filter DnCNN Method of the present invention SNR (dB) -3.09 7.73 11.98 20.03 MSE 8.68e-02 7.20E-03 2.70E-03 5.31E-04

[0129] Experiment Example 2 Actual Recording

[0130] To further verify the effectiveness of the method of the present invention, we applied it to... Figure 9 (a) shows the actual DAS data denoising process. The data used in this paper comes from the Tarim Basin in China. We can see that the DAS data is severely contaminated by noise, has poor quality, and contains strong random noise, coupling noise, fading noise, background anomaly interference, and horizontal noise. The incident downlink and reflected uplink waves, as well as other effective signals, are almost completely covered by noise and are severely discontinuous, making them impossible to fully and effectively identify. Figure 9 (b)-(d) show the processing results of the bandpass filter, DnCNN, and the denoising network of this invention, respectively. Figure 10 (a)-(c) represent the noise separated by the three methods, respectively. Figure 9 (b)-(c) and Figure 10 (a)-(b) show that both the bandpass filter and DnCNN have a certain degree of noise suppression capability, but the bandpass filter fails to suppress horizontal noise and reduces the frequency of the effective signal. DnCNN, on the other hand, retains most of the noise after processing, and the effective signal amplitude is severely attenuated. In contrast, in Figure 9 (d) and Figure 10In (c), after processing by the new method, all kinds of noise were effectively suppressed, and weak energy signals that were submerged by strong noise were also well recovered. It can be seen that the noise reduction and signal recovery capabilities of the method of the present invention are superior to the other two methods. Figure 11 for Figure 10 (a)-(b) Local magnification of the records, Figure 11 The noise appears blurry in (b), indicating that the bandpass filter removes mostly low-frequency noise. Figure 11 (c) still contains a large number of residual signals.

Claims

1. A well-dwell DAS noise suppression method based on subspace projection attention network, characterized in that, Includes the following steps: (a) Establishment of a subspace projection attention network; (1) Construction of the feature extraction part; After the signal input Y undergoes one residual convolutional block operation, it is fed into the downsampling block and the stacked residual convolutional blocks in the skip connections. The data sent to the downsampling block undergoes feature extraction through max pooling, outputting a feature map. The max pooling operation is performed by a 4×4 convolution with a stride of 2. Each feature map output from a downsampling block reduces the dimensionality of the data by half, removes redundant information, and becomes a higher-level feature map than the original input. This higher-level feature map is then passed to the next layer for subsequent higher-level feature extraction. The data after feature extraction from the stacked residual convolutional blocks becomes the low-level feature map F1. N N represents the layer number, taken as 1, 2, 3, or 4, and is sent to the SAM subspace attention module, which is also in the skip connection, for feature fusion. The bottom layer is connected through a residual convolutional block, and the highest-level feature map is extracted and shaped into F by the residual convolutional block. 5’ Send to the feature fusion section; (2) Construction of the feature fusion component; F 5’ F2 is obtained by first upsampling through 2×2 deconvolution. 4 The size is doubled, and then fed into the subspace attention module SAM, where the high-level feature map F2... 4 Compared with the previously fed low-level feature map F1 4 The feature map output after fusion is denoted as F. 4 The fused feature map F 4 Then with F 5’ The concatenation is performed along the channel dimension, and then the output is F after passing through a residual convolution block. 4’ It is sent to the next layer, and the feature fusion parts of the following layers are also processed in the same way; use N Represents a high-level feature map, feature F1 N F2 N Since they come from different layers, they need to be the same size. H and W represent height and width, and C represents the number of channels. According to F1 N F2 N Estimate M basis vectors Each These are all basis vectors of the signal subspace. This process can be abstractly represented as: （1） The process of generating the basis can be simply represented as: （2） in The signal subspace is composed of basis vectors; In the process of generating the signal subspace composed of basis vectors, it is necessary to first convert F1... N F2 N Connect along the channel dimension to obtain F only represents the intermediate process of this operation step and has no practical meaning. Then it is fed into the input with... In the residual convolution module of each output channel, its output is reshaped into... To generate A three-dimensional signal subspace, where parameters are updated end-to-end during training; In equation (2), through The signal subspace composed of the obtained basis vectors The low-level feature map F1 is transformed by orthogonal linear projection. N Projected to Above, set Let be the orthogonal projection matrix of the signal subspace. Depend on The following calculations were performed to obtain: （3） It is a normalization term; Finally, in the high-level feature map F2 N Guided by F1, the low-level feature map is F1 N The projection is reconstructed in the signal subspace as N : （4） in The subspace attention module outputs the fused feature map. Compared to traditional UNet-like architectures that directly fuse low-level and high-level features, SA-UNet projects the low-level feature map through the subspace attention module to obtain F before feature fusion. N Then it is fused with the high-level feature map obtained by upsampling; (3) The output F of the last residual convolution block 1 The noise reduction result is obtained by subtracting the initial noisy input signal y from the 3×3 convolutional layer, and finally outputting the result. ; (ii) Construction of the training set; The dataset used to train the subspace projection attention network is called the training set. The training set includes a clean signal dataset s and a noisy signal dataset y, where y = s + n, and n is noise. (1) Pure signal dataset; A large number of DAS clean signals were simulated using forward modeling, with the following specific parameter settings: 3-6 planar layers, each with a depth of 300-500m, a wave velocity of 1500-3300m / s, and a planar layer density of 1970-2245kg / m³. 3 The signal type is Ricker wavelet, the main frequency is 50-90Hz, the well-source distance is 100-200m, the sampling time interval is 1m, the recording fiber length is 2000-4000m, and the sampling frequency is 2500Hz. A 64×64 slider is used to extract K valid signal blocks from the clean record as a clean signal dataset. Based on the block size, K is greater than or equal to 20000. (2) Noisy signal dataset; To construct a noisy signal dataset, various types of noise data samples were extracted from actual acquired DAS seismic signal records and superimposed on the clean signal to ensure that the proposed method can suppress various types of noise. In the end, K noisy signal data blocks of size 64×64 were obtained. The corresponding noisy signal dataset was obtained by randomly combining and adding the clean signal samples and the actual DAS noise, and the labels were consistent with the clean signal dataset, corresponding one-to-one, as a training sample pair. (III) Training of the subspace projection attention network; Actual noisy DAS seismic records are linear superpositions of the clean signal s and DAS noise n, as shown in the following equation: （5） in To obtain the desired pure signal, Complex DAS noise, with varying phase and amplitude characteristics, includes random noise, coupling noise, background anomaly interference, fading noise, horizontal noise, and checkerboard noise. These are noisy DAS seismic records actually collected in the field. The subspace projection attention network is abstracted as a denoiser. Through continuous training, an end-to-end mapping relationship is established from noisy signal data to clean signal data, thereby achieving the removal of various noises in one go and retaining only the effective part that is desired, as shown in the following formula: （6） in The output signal data is after processing by the denoiser, where y is the noisy signal data and S is the output signal data. This represents the mapping relationship between the denoisers established by the denoising network. Abstract representations of various parameters in the network are used to continuously optimize network parameters in order to improve mapping relationships. Thus Given a pure signal s that is infinitely close to the target signal s, a pair of records is randomly selected from the training dataset. The noisy signal data is used as the input to the subspace projection attention network, and the corresponding pure signal data is used as the expected output of the network. In order to suppress noise without compromising the pure signal, the denoiser is trained by minimizing the following loss function: （7） In the formula , These represent clean signal data and noisy signal data, respectively. The input is The output of the denoiser, at this time It is the label of the noise denoiser, that is, when its input is At that time, the expected output of the denoiser is optimized through repeated training with a large number of training samples, and finally, the parameters of the subspace projection attention network suitable for complex noise reduction in actual DAS are obtained. This is called the obtained trained model; (iv) DAS data noise reduction processing; The DAS data is denoised using a trained subspace projection attention network. The actual acquired seismic exploration DAS data is input into the subspace projection attention network, and the output of the network is the denoised DAS data.

2. The well-dwell DAS noise suppression method based on subspace projection attention network according to claim 1, characterized in that, In step (a) of establishing the subspace projection attention network, the construction of the feature fusion part (2) includes the construction of residual convolutional blocks, which includes: The residual convolutional block is a fundamental component of the subspace projection attention network. It is built based on the residual network concept and includes at least one 3×3 convolutional layer and one LeakyReLU linear unit with leakage correction. The residual connection contains one 1×1 convolutional layer.

3. The well-dwell DAS noise suppression method based on subspace projection attention network according to claim 1, characterized in that, In step (a) of establishing the subspace projection attention network, the construction of the feature fusion part (2) includes the following: Low-level feature maps N and high-level feature maps N First, the features are connected along the channel dimension and fed into a residual convolutional module with M-dimensional output for shape reshaping. A small convolutional network generates the subspace basis, and finally, the low-level feature map is transformed by orthogonal linear projection. N Projecting the data into the generated subspace is the process of upsampling the high-level feature map F2. N Guided low-level feature map F1 N Self-attention projection.

Images