Microseismic Profile Denoising Method Based on Deep Learning

Abstract

The present application provides a denoising method for microseismic profiles based on deep learning. In step one, a batch of noise-free profiles and noise signals are input into the generator of a conditional generative adversarial network (cGAN), and constraint conditions are given to the generator, thereby obtaining noisy profiles for training. The discriminator receives real profiles and profiles generated from the generator, and constraint conditions are given to the discriminator. Through the mutual game between the discriminator and the generator, more realistic profiles are generated, thereby achieving the purpose of expanding the dataset. In step two, X^0,0 is input into the main network, and the noise prediction X^0,4 of the network is output. The main network is mainly structured by UNet++ and incorporates a Clique block structure. In step three, residual learning R(•) is performed on the noise prediction X^0,4 and the network input X^0,0 to obtain the denoised profile. In step four, the microseismic profile to be denoised is passed through this network, and the denoised profile is output, solving the problem of removing random noise from the signal during microseismic monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of microseismic monitoring, and in particular to a method for denoising microseismic profiles based on deep learning. Background Art

[0002] Microseismic monitoring technology is a new method developed in recent years mainly for the exploration of unconventional oil and gas reservoirs, which can obtain relatively accurate and timely signal information. Due to the particularity of the signal acquisition method, during the process of collecting signals, the signals will contain a large amount of noise, especially random noise is particularly obvious. The noise causes the effective signals to be masked, resulting in a very low signal-to-noise ratio of the microseismic signals. Therefore, the demand for microseismic denoising technology is becoming stronger and stronger. Microseismic signal profile denoising is to use denoising technology to remove as much noise as possible within a profile of microseismic signals and leave the actually detected signals to ensure that the signals have great reference value for researchers.

[0003] The signal-to-noise ratio of the signals obtained by traditional denoising methods is relatively low, and the difference between the obtained profile and the original profile is not large, so the effect of traditional denoising methods is not good. By using some methods in deep learning and performing denoising processing through fixed patterns, it is more likely to obtain a higher peak signal-to-noise ratio and structural similarity, and there will also be a certain degree of improvement in the profile quality. Chinese Patent Document CN104977618A records a method for evaluating shale gas reservoirs and finding sweet spots, which records a pre-stack denoising scheme. However, the pre-stack denoising scheme is relatively diverse, and how to select a reasonable denoising method for surface wave interference, acoustic wave interference, linear coherent noise, high-energy interference noise, and three-dimensional pre-stack random noise is not specifically disclosed in the literature. Summary of the Invention

[0004] The present invention solves the problem of removing random noise in signals during microseismic monitoring. The technical solution adopted is: a method for denoising microseismic profiles based on deep learning, specifically including the following steps:

[0005] S1. Input a batch of noise-free profiles and noise signals into the generator of the conditional generative adversarial network, given the constraint conditions of the generator, and obtain the noisy profiles for training;

[0006] The discriminator receives the real profiles and the profiles generated from the generator, sets the constraint conditions of the discriminator, and uses the mutual game between the discriminator and the generator to generate more real profiles, realizing the expansion of the dataset;

[0007] S2. Input the profile X after expansion 0,0 into the main network, and output the noise prediction X 0,4 ;

[0008] S3. Input the noise prediction X0,4 The expanded profile X of the main network input 0,0 Perform residual learning Obtain an improved main network;

[0009]

[0010] is the denoised profile, X 0，0 is the expanded profile, X 0，4 is the noise prediction, Θ is the network parameter;

[0011] S4. Input the microseismic profile to be denoised into the improved main network, and output the denoised profile.

[0012] In the preferred solution, step S1 further includes the following steps:

[0013] The microseismic profile data acquisition model is: y = X + n;

[0014] where x is the clean data, n is the noise, and y is the noisy data;

[0015] Establish the relationship between x and y of the neural network model based on deep learning:

[0016] x = Net(y; Θ);

[0017] where Net is the neural network model and Θ is the network parameter;

[0018] The neural network model is used to output the residual parameters:

[0019]

[0020] where is the residual learning. Optimize the parameters according to the output of the residual learning to obtain the optimal hyperparameters, that is, obtain the optimal neural network model.

[0021] In the preferred solution, step S1 further includes the following steps:

[0022] The generator receives a random noise z and generates a noisy profile G(z|y) through this noise; the discriminator outputs the probability D(x|y) that x is the real input profile by comparing the output G(z|y) of the generator with the input profile x - if the output is 1, it means the probability that the profile x is a real profile is 100%; if the output is 0, it means the profile x is generated by the generator; the objective function of the conditional generative adversarial network cGAN is:

[0023]

[0024] where p data is the distribution of the corresponding data;

[0025] The cross-section after cGAN expansion is as follows:

[0026] X 0,0 = cGAN(x).

[0027] In the preferred solution, step S2 further includes the following steps:

[0028] The main network is based on the structure of UNet++ and integrates the Clique block structure;

[0029] Replace the downsampling in the convolution and encoding parts of the main network with the multi-feature layer and fusion layer in MFF-CNN:

[0030] Multi-feature layer: Use convolution blocks of three sizes, 3×3, 5×5, and 7×7, to extract and fuse the input features;

[0031] The normalization method is replaced from batch normalization to batch renormalization to adapt to normalization in the small-sample environment; the activation function is replaced from the rectified linear unit ReLU to the parametric rectified linear unit PReLU to prevent overfitting of the model;

[0032] Finally, the extracted features are obtained using the structure of the residual learning block.

[0033] In the preferred solution, the calculation method of the main structure of the main network is:

[0034]

[0035] Where MSC is the multi-scale feature extraction module, u(·) is the upsampling layer, and the bilinear interpolation method is used; [·] is the concatenation layer.

[0036] In the preferred solution, the calculation method of the volume feature extraction is:

[0037] X 0,1 = Concat(X 0，0 , u(X 1，0 ))

[0038] X 1,0 = MSC(X 0，0 ).

[0039] In the preferred solution, in the Clique Block, except for the input node, any two layers in the same module are bidirectionally connected, that is, each layer is both the input and the output of other layers;

[0040] In the first stage, the input layer initializes all the layers in the block through unidirectional connections, and the connected layers are updated one by one to update the next layer; starting from the second stage, the layers are updated alternately. All layers except the top layer to be updated are connected as the bottom layer, and the corresponding parameters of each layer are also connected together; that is, the output of the i-th layer in the k-th loop is as follows:

[0041]

[0042] where * represents a convolution operation with weight W, g represents an activation function, k represents the number of stages, k ≥ 2, and i represents the layer number, i ≥ 1;

[0043] Let x 4，0 The feature obtained after passing through the Cllique Block is denoted as Then we get:

[0044]

[0045] where CB represents the Clique Block structure.

[0046] In a preferred solution, upsampling replaces the deconvolution method of UNet++ with bilinear interpolation to restore the image, and the finally obtained noise prediction is X 0，4 .

[0047] A microseismic profile denoising method based on deep learning provided by the present invention has the following beneficial effects: Compared with the prior art method of analyzing and then removing different noises in the pre-stack denoising scheme, the present invention adopts an artificial intelligence method to denoise according to the difference in noise data characteristics. By using the method of combining cGAN with UNet++ and Clique Block in deep learning to denoise the microseismic profile, there are two advantages: First, cGAN can input certain conditions into the generator, and the generator automatically generates a profile that meets the requirements, and then through the discriminator for mutual game, to automatically generate a microseismic profile as close as possible to the real profile for network training; Second, using UNet++ for signal-to-noise separation and using Clique Block for feature reuse. Compared with the previous denoising methods, the present invention can greatly improve the learning ability of the network, and thus improve the ability of microseismic signal profile denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below with reference to the drawings and embodiments.

[0049] Figure 1 is an example diagram of the M-C block structure of the present invention.

[0050] Figure 2 is the workflow diagram of the present invention.

[0051] Figure 3 is the CGAN structure diagram of the present invention.

[0052] Figure 4 is the multi-scale feature extraction structure diagram of the present invention.

[0053] Figure 5 is the Clique Block structure diagram of the present invention.

[0054] Figure 6 is the overall network structure diagram of the present invention.

[0055] Figure 7 Noisy microseismic signal diagram.

[0056] Figure 8 is the processing result diagram of DenseNet.

[0057] Figure 9 is the intermediate feature diagram after the second convolution of DenseNet.

[0058] Figure 10 is the processing result diagram of DnCNN.

[0059] Figure 11 is the intermediate feature diagram after the second convolution of DnCNN.

[0060] Figure 12 is the processing result diagram of this network.

[0061] Figure 13 is the intermediate feature diagram after the second convolution of this network. Detailed implementation manner

[0062] Such as Figures 1 to 5 in, the seismic data acquisition model is:

[0063] y = x + n (1);

[0064] where x is the clean data, n is the noise, and y is the noisy data. The basic idea of the deep learning neural network is to establish the following relationship between x and y:

[0065] x = Net(y; Θ) ( 2) ;

[0066] where Net is the proposed network and Θ is the network parameter. In actual applications, the residual part is used as the output:

[0067]

[0068] where It is residual learning. Parameter optimization is performed based on the output of residual learning to obtain the optimal hyperparameters, that is, the optimal network is obtained.

[0069] As Figures 6 to 13 shown in

[0070] The CliUNet microseismic profile denoising method includes the following steps:

[0071] Step 1: Input a batch of noise-free profiles and noise signals into the generator of cGAN, and given the constraint conditions of the generator, so as to obtain the noisy profiles for training; the discriminator receives the real profiles and the profiles generated from the generator, and given the constraint conditions of the discriminator, and uses the mutual game between the discriminator and the generator to generate more real profiles, so as to achieve the purpose of expanding the dataset.

[0072]

[0073] where p data is the distribution of the corresponding data.

[0074] The profiles after expanding the dataset by cGAN are:

[0075] X 0，0 = CGAN(x)(5);

[0076] Step 2: Input X 0,0 into the main network, and output the noise prediction X 0,4 of the network. The main network takes the structure of UNet++ as the main body and integrates the Clique block structure. The specific steps are as follows:

[0077] 1. Replace the downsampling of the convolution and encoding parts with the multi-feature layer and fusion layer in MFF-CNN:

[0078] Multi-feature layer: Convolution blocks of three sizes, 3×3, 5×5, and 7×7, are used to extract and fuse the input features; the normalization method is replaced from batch normalization (BN. Ioffe and Szegedy, 2015) to batch renormalization (BRN. Ioffe, 2017) to adapt to the normalization in the small-sample environment; the activation function is replaced from the ReLU activation function to the PReLU activation function to prevent overfitting of the model; finally, the structure of the residual learning block is used to obtain the extracted features.

[0079] The main structure of the network - the M-C block is obtained from Equation (4):

[0080]

[0081] where MSC is the multi-scale feature extraction module, u(·) is the upsampling layer, and the bilinear interpolation method is used; [·] is the concatenation layer. Specifically, in Figure 1

[0082] X 0，1 = Concat(X 0，0 , u(X 1，0 ))

[0083] X 1，0 = MSC(X 0，0 )

[0084] The intermediate feature after encoding is denoted as X 4，0 .

[0085] 2. Use the Clique Block for feature reuse to improve the feature recognition ability of the model:

[0086] The Clique Block structure: In the Clique Block, except for the input node, any two layers in the same block are bidirectionally connected - each layer is both the input and the output of other layers. In the first stage, the input layer initializes all layers in the block through a unidirectional connection, and the connected layers will be updated one by one to update the next layer; starting from the second stage, the layers start to be updated alternately. All layers except the top layer to be updated are connected as the bottom layer, and their corresponding parameters are also connected together; therefore, the output of the i-th layer (i≥1) in the k-th layer (k≥2) loop can be expressed as:

[0087]

[0088] In Equation (1), * is the convolution operation with weight W, g is the activation function, k represents the number of stages, and i represents the number of layers. Denote the feature obtained by passing X 4,0 through the Clique Block as then we can get:

[0089]

[0090] where CB represents the Clique Block structure.

[0091] 3. The upsampling replaces the deconvolution method used in UNet++ with the bilinear interpolation method to restore the image, and the finally obtained noise prediction is X 0，4 . ​

[0092] Step 3: Perform residual learning on the noise prediction X 0,4 and the network input X 0,0 to obtain a denoised profile (with network parameters Θ): (The network parameters are Θ):

[0093]

[0094] Step 4: Pass the microseismic profile to be denoised through this network to output the denoised profile.

[0095] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations on the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A microseismic profile denoising method based on deep learning, characterized in that It includes the following steps: S1. Input a batch of noise-free profiles and noise signals into the generator of the conditional generative adversarial network. Given the constraints of the generator, obtain the noisy profiles for training; The discriminator receives real profiles and the profiles generated from the generator. Set the constraints of the discriminator, and use the mutual game between the discriminator and the generator to generate more real profiles, realizing the expansion of the dataset; S2. Input the expanded profile X 0,0 into the main network and output the noise prediction X 0,4 ; S3. Perform noise prediction X 0,4 and the expanded profile X of the main network input 0,0 to perform residual learning to obtain an improved main network; is the denoised profile, X 0,0 is the expanded profile, X 0,4 is the noise prediction, Θ is the network parameter; S4. Input the microseismic profile to be denoised into the improved main network, and output the denoised profile.

2. The method for denoising microseismic profiles based on deep learning according to claim 1, wherein: The following steps are also included in step S1: The microseismic profile data acquisition model is: y = x + n; where x is the clean data, n is the noise, and y is the noisy data; Establish the relational expression between x and y of the neural network model based on deep learning: x = Net(y; Θ); where Net is the neural network model and Θ is the network parameter; The neural network model is used to output the residual parameter: Among them is residual learning. Parameter optimization is carried out according to the output of residual learning to obtain the optimal hyperparameters, that is, to obtain the optimal neural network model.

3. The method for denoising microseismic profiles based on deep learning according to claim 1, wherein: The following steps are also included in step S1: The generator receives a random noise z and generates a noisy profile G(z|y) through this noise; the discriminator outputs the probability D(x|y) that x is the real input profile by comparing the generator output G(z|y) with the input profile x - if the output is 1, it means the probability that the profile x is a real profile is 100%; if the output is 0, it means the profile x is the profile generated by the generator; the objective function of the conditional generative adversarial network cGAN is: where p data is the distribution of the corresponding data; The profile after expansion by cGAN is:

4. The microseismic profile denoising method based on deep learning according to claim 1, characterized in that: The following steps are also included in step S2: The main network takes the structure of UNet++ as the main body and integrates the Clique block structure; Replace the downsampling of the convolution and encoding parts in the main network with the multi-feature layer and fusion layer in MFF-CNN: Multi-feature layer: Use convolution blocks of three sizes, 3×3, 5×5, and 7×7, to extract and fuse the input features; The normalization method is replaced from batch normalization to batch renormalization to adapt to the normalization in the small-sample environment; the activation function is replaced from the rectified linear unit ReLU to the parametric rectified linear unit PReLU to prevent overfitting of the model; Finally, use the structure of the residual learning block to obtain the extracted features.

5. The method for denoising microseismic profiles based on deep learning according to claim 4, wherein: The calculation method of the main structure of the main network is: where MSC is the multi-scale feature extraction module, u(·) is the upsampling layer, and the bilinear interpolation method is used; [·] is the concatenation layer.

6. The method for denoising microseismic profiles based on deep learning according to claim 5, wherein: The specific feature extraction calculation method is: X 0，1 = Concat(X 0，0 , u(X 1，0 ) X 1,0 = MSC(X 0,0 )。 7. The method for denoising microseismic profiles based on deep learning according to claim 5, characterized in that: In the Clique Block, except for the input node, any two layers in the same module are bidirectionally connected, that is, each layer is both the input of other layers and the output of other layers; In the first stage, the input layer initializes all layers in the block through a unidirectional connection, and the connected layers updated each time will update the next layer; Starting from the second stage, the layers start to be updated alternately. All layers except the top layer to be updated are connected as the bottom layer, and the corresponding parameters of each layer are also connected together; that is, the output of the i-th layer in the k-th loop is as follows: In the formula, * is the convolution operation with weight W, g is the activation function, k represents the number of stages, k ≥ 2, i represents the number of layers, i ≥ 1; Take X 4,0 The features obtained after passing through the Clique Block are denoted as Then we get: where CB represents the Clique Block structure.

8. The method for denoising microseismic profiles based on deep learning according to claim 5, wherein: Upsampling replaces the deconvolution method of UNet++ with bilinear interpolation method to restore the image, and the finally obtained noise prediction is X 0,4 。

Images