Microseismic signal denoising method combined with wgan-gp and sadnet

Abstract

A microseismic signal denoising method combining WGAN-GP and SADNet is proposed. Microseismic signal samples are input into the WGAN-GP network, and noise signal conditions are added. A generative adversarial model is used to generate a large training sample set, expanding the small sample dataset. The expanded training samples from WGAN-GP are then used as input to train SADNet. The acquired microseismic data is input into the trained SADNet, and the denoised microseismic data is output. This invention utilizes WGAN-GP to expand the small sample dataset of microseismic signals. The variable convolution introduced by the residual space adaptive block in SADNet can adapt to spatial texture and edge changes, and the focus on the global context encoding / decoding structure can extract multi-scale information, thereby improving the denoising accuracy of complex noise in microseismic signals.

Description

Technical Field

[0001] This invention belongs to the field of microseismic signal denoising, and specifically relates to a microseismic signal denoising method that combines WGAN-GP and SADNet. Background Technology

[0002] Currently, with the gradual exploitation of shale oil and gas and unconventional oil and gas resources, microseismic monitoring technology for oil and gas extraction has long attracted attention for the rational and safe development of resources. Generally, methods such as hydraulic fracturing or conventional water injection are used to monitor underground fractures, oil and gas distribution, and pressure changes, providing an efficient and safe approach to oil and gas extraction. However, microseismic signals acquired during monitoring are easily interfered with by various external noises, resulting in a low proportion of effective signal. Without denoising, it is difficult to obtain the desired signal from these low signal-to-noise ratio (SNR) microseismic signals. From the current methods, the acquired microseismic signal samples suffer from "complex interference noise, low effective signal ratio," and when the acquired microseismic signal dataset is small, traditional denoising methods struggle to improve the SNR of the effective signal, resulting in unsatisfactory efficiency and quality. Therefore, finding a highly efficient and high-quality microseismic denoising method is currently essential.

[0003] In recent years, deep learning has shown great promise in denoising microseismic signals. However, addressing the challenges of small sample datasets and complex interference noise in effective microseismic signals, and aiming to improve the signal-to-noise ratio (SNR) of the effective microseismic signal, the inventors' team has proposed a novel method combining WGAN-GP [Ishaan Gulrajani, Faruk Ahmed et al., 2017] and SADNet [Meng Chang, Qi Li et al., 2020] to denoise microseismic signals. However, some technical challenges remain, including low denoising accuracy and unsatisfactory denoising effects for complex noise in microseismic signals. Summary of the Invention

[0004] In view of the technical problems existing in the background technology, the present invention provides a microseismic signal denoising method combining WGAN-GP and SADNet. The present invention uses WGAN-GP to expand the small sample dataset of microseismic signals, and the variable convolution introduced by the residual space adaptive block in SADNet can adapt to spatial texture and edge changes. By focusing on the global context encoding and decoding structure, multi-scale information can be extracted, thereby improving the denoising accuracy of complex noise in microseismic signals.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A method for denoising microseismic signals combining WGAN-GP and SADNet, comprising the following steps:

[0007] Step 1: Input the microseismic signal samples into the WGAN-GP network, add noise signal conditions, and generate a large number of training sample sets through a generative adversarial model to expand the small sample dataset.

[0008] The objective function for generating samples in WGAN-GP is defined as follows:

[0009]

[0010] In the formula, x ~ p data (x) is the true data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is a value between the true data distribution x ~ p data (x) and the generated data distribution x~p g The sampling distribution in (x) is λ, where λ is a hyperparameter and D(x) is the discriminator probability distribution. It is regularization; in formula (1), the penalty sampling x ~ p is added. x (x) is used to constrain the Lipschit function, and gradient constraints are applied to the generated samples through regularization;

[0011] Step 2: Use the augmented training samples obtained from WGAN-GP as input to train SADNet;

[0012] SADNet uses four multi-scale encoder-decoder structures, with the number of channels in each structure set to 32, 64, 128, and 256 respectively in this paper. The kernel size of the first and last Conv layers is set to 1×1, and the final output is set to 1 according to the input. A 2×2 filter convolution kernel is used in the sampling layer, and the kernel size of the other Conv layers is 3×3.

[0013] SADNet first extracts initial features from the input training samples using a single Conv layer. These initial features are then input into a multi-layer decoder-encoder architecture to denoise the samples. In the encoding part, ResBlock is used to extract features at different scales. Finally, it flexibly captures signal features to address the complex noise in microseismic signals. The principle is as follows:

[0014]

[0015] In the formula, w i This represents the weight at position p in the convolution kernel, p iIndicates the position in N(p); Δp i For position p i The learning offset, Δm i It is a learnable modulation variable, with a value range of [0,1].

[0016] The reconstructed features are input into the Conv layer to restore and denoise the image, yielding the output result. x is the input data sample. The sample data after denoising;

[0017] Step 3: Input the collected microseismic data to be measured into the trained SADNet, and output the denoised microseismic data.

[0018] In the preferred embodiment, in step one, x~p x The data distribution of (x) falls between the real data distribution and the generated data distribution. The data is implemented by using the real data distribution P. data and the generated data distribution P g Perform one sampling on each of the two data distribution sets, and then perform another random sampling on these two data distribution sets, x ~ p x The formula for determining the value of (x) is as follows:

[0019] x~p x (x)=ε(x~p data (x))+(1-ε)(x~p g (x)) (2)

[0020] In the formula, x ~ p data (x) is the true data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is a value between the true data distribution x ~ p data (x) and the generated data distribution x~p g The sampling distribution in (x) has random number ε that takes values ​​that follow a uniform distribution in [0,1].

[0021] In the preferred embodiment, the value of λ is 10.

[0022] This patent can achieve the following beneficial effects:

[0023] This invention employs a combination of WGAN-GP and SADNet to denoise microseismic signals. It offers three main advantages: 1. WGAN-GP can expand the sample size, and the resulting training sample set serves as the basis for SADNet, requiring only a small training set to complete the microseismic signal denoising. 2. The variable convolutions in the residual space adaptive block of SADNet can extract signal texture and edge features from noisy microseismic signals. 3. The use of a multi-layer encoder-decoder structure in SADNet can extract global contextual information. The encoding part uses ResBlocks for multi-scale feature extraction, and the context block expands the receptive field and captures multi-scale information. In the decoding part, RSABs are used to eliminate noise and reconstruct textures, removing noise from coarse to fine.

[0024] This invention, combining WGAN-GP and SADNet, achieves excellent results in denoising signals with small sample datasets and complex noise interference, significantly improving the denoising efficiency and accuracy of microseismic monitoring signals. Figure 5-8 As can be seen, DnCNN, UDnCNN, and DUDnCNN can all reduce noise. The denoising results of DnCNN and UDnCNN show that residual interference noise remains significant. Compared to the previous two networks, the new method proposed in this invention, after expanding the small sample dataset using WGAN-GP, shows that the DUDnCNN network has a higher effective signal-to-noise ratio for microseismic signals, indicating that the method in this patent has higher denoising accuracy and better performance. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0026] Figure 1 This is a flowchart illustrating the combined WGAN-GP and SADNet method of this invention.

[0027] Figure 2 This is a diagram of the SADNet structure of the present invention;

[0028] Figure 3 This is a structural diagram of the Context block of the present invention;

[0029] Figure 4 This is a structural diagram of the RASBs of the present invention;

[0030] Figure 5 This is a microseismic noise signal diagram of the present invention;

[0031] Figure 6 This is the result of DnCNN processing in this invention;

[0032] Figure 7 This is the result of UDnCNN processing in this invention;

[0033] Figure 8 This is the result of DUDnCNN processing in this invention. Detailed Implementation

[0034] This invention addresses the problem of "small sample datasets and complex interference noise" in microseismic signals. WGAN-GP [Ishaan Gulrajani, Faruk Ahmed et al., 2017] can expand the small sample datasets of microseismic signals. The variable convolution introduced by the residual space adaptive block in SADNet [Meng Chang, Qi Li et al., 2020] can adapt to spatial texture and edge changes, and focusing on the global context encoding / decoding structure can extract multi-scale information, thereby improving the denoising accuracy of complex noise in microseismic signals. Therefore, a new method combining WGAN-GP [Ishaan Gulrajani, Faruk Ahmed et al., 2017] and SADNet [Meng Chang, Qi Li et al., 2020] is proposed for denoising microseismic signals. The main improvements are as follows:

[0035] 1) We expanded the small sample dataset using WGAN-GP [Ishaan Gulrajani, Faruk Ahmed et al., 2017], and used it as the input training sample set for the SADNet network [Meng Chang, Qi Li et al., 2020].

[0036] 2) The variable convolution introduced by the residual space adaptive block in SADNet [Meng Chang, Qi Li et al., 2020] can adapt to the texture and edge changes of microseismic signals. The decoding and encoding with context blocks can extract multi-scale information of microseismic signals and address complex interference noise in the training sample set. By combining sample expansion and denoising networks, the efficiency and accuracy of denoising microseismic signals are improved.

[0037] The specific noise reduction methods are as follows:

[0038] Preferred solutions include Figures 1 to 8 As shown, a microseismic signal denoising method combining WGAN-GP and SADNet is described, with the following steps:

[0039] Step 1: Input clean microseismic signal samples into the WGAN-GP network, add noisy signal conditions, and generate a large training sample set through a generative adversarial model to expand the small sample dataset. The principle of WGAN-GP is as follows:

[0040] The objective function for generating samples in WGAN-GP is defined as follows:

[0041]

[0042] In formula (1), x ~ p data (x) is the true data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is a value between the true data distribution x ~ p data (x) and the generated data distribution x~p g In the sample distribution (x), λ is a hyperparameter, which is set to 10 in this embodiment, and D(x) is the discriminator probability distribution. It is regularization; in formula (1), the penalty sampling x ~ p is added. x (x) is used to constrain the Lipschit function, and gradient constraints are applied to the generated samples through regularization.

[0043] x~p x The data distribution of (x) falls between the real data distribution and the generated data distribution. The data is implemented by using the real data distribution P. data and the generated data distribution P g Perform one sampling on each of the two data distribution sets, and then perform another random sampling on these two data distribution sets, x ~ p x The formula for determining the value of (x) is as follows:

[0044] x~p x (x)=ε(x~p data (x))+(1-ε)(x~p g (x)) (2)

[0045] In formula (2), x ~ p data (x) is the true data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is a value between the true data distribution x ~ p data (x) and the generated data distribution x~p g The sampling distribution in (x) has random number ε that takes values ​​that follow a uniform distribution in [0,1].

[0046] Step 2: Use the augmented training samples obtained from WGAN-GP as input to train SADNet. See the SADNet network structure diagram below. Figure 2 As shown,

[0047] SADNet uses four multi-scale encoder-decoder architectures, with the number of channels in each architecture set to 32, 64, 128, and 256 respectively in this paper. The kernel size of the first and last Conv layers is set to 1×1, and the final output is set to 1 according to the input. A 2×2 filter convolution kernel is used in the sampling layer, and the kernel size of the other Conv layers is 3×3.

[0048] SADNet first extracts initial features from the input training samples using a single Conv layer. These initial features are then input into a multi-layer decoder-encoder architecture to denoise the samples. In the encoding part, ResBlock is used to extract features at different scales. Here, ResBlock uses Leaky ReLU as the activation function, and the Batch Normalization layer is removed. A context block is used instead. Figure 3 As shown in the figure, several dilated convolutions with different dilation rates are used to expand the receptive field and capture multi-scale information; the number of parameters and the computational cost of the network are reduced; and finally, the features extracted from different receptive fields are fused for the output of the decoding part.

[0049] In the decoding section, RSABs are discussed. Figure 3 As shown in the figure, variable convolution is used to learn the migration map, thereby avoiding the influence of interfering features on the output results. Furthermore, the principle behind flexibly capturing signal features in response to complex noise in microseismic signals is as follows:

[0050]

[0051] w in formula (3) i This represents the weight at position p in the convolution kernel, p i This represents the position in N(p). Δp i For position p i The learning offset, Δm i It is a learnable modulation variable with a value range of [0,1].

[0052] The reconstructed features are input into the Conv layer to restore and denoise the image, yielding the output result. x is the input data sample. The data shown is the denoised sample data. To ensure the denoising performance of SADNet, this paper uses L2Loss to train the generated sample data and L1Loss to train the real sample data.

[0053] Step 3: Input the collected microseismic data to be measured into the trained SADNet, and output the denoised microseismic data.

[0054] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for denoising microseismic signals combining WGAN-GP and SADNet, characterized in that... Includes the following steps: Step 1: Input the microseismic signal samples into the WGAN-GP network, add noise signal conditions, and generate a large number of training sample sets through a generative adversarial model to expand the small sample dataset. The objective function for generating samples in WGAN-GP is defined as follows: In the formula, x ~ p data (x) is the true data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is a value between the true data distribution x ~ p data (x) and the generated data distribution x~p g The sampling distribution in (x) is λ, where λ is a hyperparameter and D(x) is the discriminator probability distribution. It is regularization; in formula (1), the penalty sampling x ~ p is added. x (x) is used to constrain the Lipschit function, and gradient constraints are applied to the generated samples through regularization; Step 2: Use the augmented training samples obtained from WGAN-GP as input to train SADNet; SADNet uses four multi-scale encoder-decoder structures, with the number of channels in each structure set to 32, 64, 128, and 256 respectively in this paper. The kernel size of the first and last Conv layers is set to 1×1, and the final output is set to 1 according to the input. A 2×2 filter convolution kernel is used in the sampling layer, and the kernel size of the other Conv layers is 3×3. SADNet first extracts initial features from the input training samples using a single Conv layer. These initial features are then input into a multi-layer decoder-encoder architecture to denoise the samples. In the encoding part, ResBlock is used to extract features at different scales. Finally, it flexibly captures signal features to address the complex noise in microseismic signals. The principle is as follows: In the formula, w i This represents the weight at position p in the convolution kernel, p i Indicates the position in N(p); Δp i For position p i The learning offset, Δm i It is a learnable modulation variable, with a value range of [0,1]. The reconstructed features are input into the Conv layer to restore and denoise the image, yielding the output result. x is the input data sample. The sample data after denoising; Step 3: Input the collected microseismic data to be measured into the trained SADNet, and output the denoised microseismic data.

2. The microseismic signal denoising method combining WGAN-GP and SADNet according to claim 1, characterized in that: In step one, x~p x The data distribution of (x) falls between the real data distribution and the generated data distribution. The data is implemented by using the real data distribution P. data and the generated data distribution P g Perform one sampling on each of the two data distribution sets, and then perform another random sampling on these two data distribution sets, x ~ p x The formula for determining the value of (x) is as follows: x~p x (x)=ε(x~p data (x))+(1-ε)(x~p g (x)) (2) In the formula, x ~ p data (x) is the true data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is a value between the true data distribution x ~ p data (x) and the generated data distribution x~p g The sampling distribution in (x) has random number ε that takes values ​​that follow a uniform distribution in [0,1].

3. The microseismic signal denoising method combining WGAN-GP and SADNet according to claim 1, characterized in that: The value of λ is 10.

Images