Microseismic signal first arrival picking method based on improved WGAN-GP and picking-net

Abstract

The application provides a microseismic signal first arrival picking method based on improved WGAN-GP and Picking-Net, step one, input clean microseismic signal samples into the improved WGAN-GP, add random Gaussian noise signal conditions, generate a large number of training sample sets through a generative adversarial model, and expand a small sample data set; step two, use the improved WGAN-GP to complete training of the expanded training sample as input to train the Picking-Net; step three, input the microseismic signal to be first arrival picked into the trained Picking-Net, and output the picked first arrival signal.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of microseismic monitoring, and particularly relates to a microseismic signal first arrival picking method based on an improved WGAN-GP (Wasserstein Generative Adversarial Network based on the principle of bulldozer distance) and a Picking-Net. BACKGROUND

[0002] In recent years, with the development of oil and gas reservoir exploration technology, in the aspect of oil and gas exploitation and development, microseismic monitoring technology has gradually become a widely used new method. The microseismic monitoring technology uses a geophysical technology to monitor and analyze the formation rupture state by receiving and observing microseismic events generated by downhole fracturing using a geophone in a well or on the ground, and has important significance for oil and gas reservoir development. In actual microseismic monitoring, accurately and in real time identifying and picking up the microseismic first arrival signal is an important prerequisite, and is the basis for subsequent research work such as source positioning, fracture prediction and fracturing analysis.

[0003] The small sample problem and the low signal-to-noise ratio problem of microseismic signal first arrival picking have always been difficult problems in microseismic signal first arrival picking. The small sample problem of microseismic data makes the training efficiency of the deep learning network not high, and the low signal-to-noise ratio problem of the microseismic signal leads to the inaccuracy of the deep learning network in picking up the first arrival signal. In order to improve the efficiency and accuracy of microseismic signal first arrival picking, we propose a new method of microseismic signal first arrival picking based on an improved WGAN-GP and Picking-Net to complete the first arrival picking of microseismic signals. A signal first arrival picking method combining U-Net and Temporal Ensembling is recorded in Chinese patent document CN111695413 A. The patent mainly introduces a network structure for expanding a small amount of microseismic signals and integrating predicted effective signals. However, the patent needs to preprocess and label the data, and the steps are relatively cumbersome, and the existing feature extraction method is not improved. SUMMARY

[0004] The application provides a new method for microseismic signal first arrival picking based on improved WGAN-GP (generative adversarial network based on bulldozer distance principle) and Picking-Net, solves the problems of low efficiency and precision of microseismic signal first arrival picking, and thus completes the first arrival picking of microseismic signals. The main innovations are as follows: (1) the improved WGAN-GP replaces the full connection layer in the generator and discriminator of the WGAN-GP with a convolution layer, so that the network can expand the input microseismic data sample set before inputting the data, and improve the training efficiency of the network model; (2) the Picking-Net replaces the convolution block with a multi-scale convolution block based on UNet++, which can extract more detailed features, and adds a Transformer module at the last layer of the encoder to deepen the extraction of first arrival signal features, thereby improving the picking precision of microseismic first arrival signals; (3) the joint improved WGAN-GP and Picking-Net are used to complete the first arrival picking of microseismic signals, which can improve the efficiency and precision of first arrival picking.

[0005] To solve the above technical problems, the technical scheme adopted by the application is as follows: a method for first arrival picking of microseismic signals based on improved WGAN-GP and Picking-Net, comprising the following steps:

[0006] S1, input clean microseismic signal samples into the improved WGAN-GP (generative adversarial network based on bulldozer distance principle), add random Gaussian noise signals, generate a large number of training sample sets through a generative adversarial model, and expand the small sample data set;

[0007] S2, use the improved WGAN-GP to complete the training of the expanded training sample as input to the Picking-Net;

[0008] S3, input the microseismic signal to be first arrival picked into the trained Picking-Net, and output the picked first arrival signal.

[0009] In the preferred scheme, step S1 further comprises the following steps:

[0010] The target function of the improved WGAN-GP for generating samples is:

[0011]

[0012] Where x ~ p data (x) is the real data distribution, x ~ p g (x) is the generated data distribution, x ~ p x (x) is between the real data distribution x ~ p data (x) and the generated data distribution x ~ pg (x) is the sampling distribution in (x), L is the loss function, λ is a hyperparameter, the value of the application is 10, D(x) is the discriminator probability distribution, is the regularization.

[0013] The sampling x~p x (x) is constrained by adding a penalty x (x) is constrained by the gradient of the generated sample through regularization, x~p x (x) is the data distribution range between the real data distribution and the generated data distribution.

[0014] By sampling once on the real data distribution p data (x) and the generated data distribution p g (x) respectively, and then randomly sampling once on the two data distribution sets, x~p x (x) is as follows:

[0015] x~p x (x) = ε(x~p data (x)) + (1-ε)(x~p g (x));

[0016] Where x~p data (x) is the real data distribution, x~p g (x) is the generated data distribution, x~p x (x) is the sampling distribution between the real data distribution x~p data (x) and the generated data distribution x~p g (x), and the random number ε takes a value from the uniform distribution of [0,1].

[0017] In the preferred scheme, step S2 further comprises the following steps:

[0018] Replace the VGG Block (see Figure 2 ) originally used in UNet++ with a new multi-scale convolution block New Block (see Figure 3 ); The expanded training sample is input into Picking-Net for training.

[0019] In the preferred scheme, step S3 further comprises the following steps:

[0020] The Picking-Net structure method is:

[0021] (1) First, the initial signal features are extracted through four layers of multi-scale convolution blocks New Block in the encoder part, and down-sampling is performed layer by layer;

[0022] (2) The hidden feature map extracted by the encoder part is linearly transformed and input into the Transformer (see Figure 4 ) structure, and the feature map size is reorganized back to the decoder after passing through three layers of Transformer;

[0023] (3) In the decoder part, four layers of multi-scale convolution blocks New Block are passed through, and up-sampling is performed layer by layer;

[0024] (4) The intermediate skip connection part performs skip connection between the encoder up-sampling layer by layer and the previous layer, and also performs skip connection with the decoder structure, and finally obtains the initial feature of Picking-Net network training;

[0025] (5) The initial feature output by Picking-Net is input into the softmax layer to obtain the initial picking result map, and the Picking-Net loss function is defined as follows:

[0026]

[0027] Where, Y b is the real data, is the predicted data, N is the batch size, b [1, 2,..., N], L is the loss function, and the Picking-Net loss function is composed of cross entropy: and Dice coefficient: . BRIEF DESCRIPTION OF DRAWINGS

[0028] The application will be further described below in combination with the drawings and examples.

[0029] Figure 1 is the Picking-Net structure diagram of the application;

[0030] Figure 2 is the VGG Block convolution block structure diagram in UNet++ of the application;

[0031] Figure 3 is the New Block convolution block structure diagram in Picking-Net of the application;

[0032] Figure 4 is the Transformer structure diagram of the application;

[0033] Figure 5 is the flow chart of initial picking of microseismic signal of the application;

[0034] Figure 6 is the microseismic data material diagram before picking of the application;

[0035] Figure 7 is the Picking-Net picking the first arrival point of the microseismic signal after picking. DETAILED DESCRIPTION

[0036] As Figures 1-5 In the specific implementation, the method for picking the first arrival of the microseismic signal by improving the WGAN-GP and the Picking-Net includes the following steps:

[0037] Step one, input the clean microseismic signal sample into the improved WGAN-GP based on the bulldozer distance principle with a gradient penalty term, add a random Gaussian noise signal condition, generate a large number of training sample sets through the generation of an adversarial model, and expand the small sample data set. The principle of improving the WGAN-GP is as follows:

[0038] The objective function of the improved WGAN-GP for generating samples is defined as follows:

[0039]

[0040] In formula 1, x ~ p data (x) is a real data distribution, x ~ p g (x) is a generated data distribution, x ~ p x (x) is a sampling distribution between the real data distribution x ~ p data (x) and the generated data distribution x ~ p g (x), L is a loss function, λ is a hyperparameter, the value of which is 10 in this patent, D(x) is a discriminator probability distribution, is a regularization; in formula 1, the Lipschit function is constrained by adding the penalty sampling x ~ p x (x), the gradient of the generated sample is constrained by regularization, x ~ p x (x), the data distribution of x ~ p daya (x) is a distribution between the real data distribution and the generated data distribution, and the implementation method of the data is to sample once from the real data distribution p g and the generated data distribution p x , and then randomly sample again on the two data distribution sets, x ~ p x (x) is as follows:

[0041] x ~ p data (x) = ε(x ~ p g (x)) + (1-ε)(x ~ p data (x)) (2)

[0042] In formula 2, x ~ p g (x) is a real data distribution, x ~ p x(x) is between the real data distribution x ~ p data (x) and the generated data distribution x ~ p g (x) in the sampling distribution, the random number ε takes the value from the uniform distribution of [0, 1];

[0043] Step two, using the improved WGAN-GP to complete the training sample as input to Picking-Net, Picking-Net network structure diagram, as shown in Figure 1 .

[0044] Picking-Net structure diagram will replace the original VGG Block used in UNet++ with a new multi-scale convolution block New Block, as shown in Figures 2-3 .

[0045] The expanded training sample is input into Picking-Net for training, and the Picking-Net structure method is as follows:

[0046] (1) First, the initial signal feature is extracted through four layers of multi-scale convolution block New Block in the encoder part, and down-sampling is performed layer by layer;

[0047] (2) After that, the hidden feature map extracted by the encoder part is linearly transformed and input into the Transformer structure shown in Figure 4 After three layers of Transformer, the feature map size is reorganized back to the decoder;

[0048] (3) In the decoder part, four layers of multi-scale convolution block New Block are used, and up-sampling is performed layer by layer;

[0049] (4) The intermediate skip connection part connects the encoder layer by layer up-sampling and the previous layer, and also connects the decoder structure, and finally obtains the Picking-Net network training initial feature;

[0050] (5) The Picking-Net finally outputs the initial feature to the softmax layer to obtain the initial picking result map, and the Picking-Net loss function is defined as follows:

[0051]

[0052] In formula 3, Y b is the real data, is the predicted data, N is the batch size, b ∈ [1, 2, …, N], L is the loss function, and the Picking-Net loss function is composed of cross entropy: And Dice coefficient: ;

[0053] Step three, input the microseismic signal to be picked into the trained Picking-Net, and output the picked first arrival signal.

[0054] As Figures 6-7 shown, the comparison between the microseismic first arrival point data before and after picking confirms the effectiveness of the present application.

[0055] The above embodiments are only preferred technical solutions of the present application, and should not be regarded as limitations of the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, and include equivalent replacement solutions of the technical features recited in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present application.

Claims

1. A method for first arrival picking of microseismic signals based on improved WGAN-GP and Picking-Net, characterized by: Includes the following steps: S1. Noise-free microseismic signal samples are input into the improved WGAN-GP, and random Gaussian noise signal conditions are added. Through generative adversarial model, a large number of training sample sets are generated to expand the small sample dataset. S2. Replace the VGG Block originally used in UNet++ with a new multi-scale convolutional block (New Block); input the augmented training samples into Picking-Net for training. S3. Input the microseismic signal to be picked up into the pre-trained Picking-Net and output the picked up first arrival signal. The method for constructing the Picking-Net structure in step S2 is as follows: (1) First, the encoder part performs initial signal feature extraction through four layers of multi-scale convolutional blocks (New Blocks), and then downsamples layer by layer; (2) The hidden feature map extracted by the encoder is then linearly transformed and input into the Transformer structure. After passing through three Transformer layers, the feature map size is reassembled and returned to the decoder. (3) The decoder part passes through four layers of multi-scale convolutional blocks (New Blocks) and is upsampled layer by layer; (4) The middle skip connection part skips the encoder layer by layer upsampling and the previous layer, and then skips the decoder structure, finally obtaining the initial features of the Picking-Net network training. (5) Input the first-arrival features of the final output of Picking-Net into the softmax layer to obtain the first-arrival picking result map. The loss function of Picking-Net is defined as follows: ； in, These are real data. This is the predicted data, and N is the batch size. L is the loss function. The Picking-Net loss function is composed of cross-entropy: And Dice coefficient: composition.

2. The microseismic signal first arrival picking method based on improved WGAN-GP and Picking-Net as described in claim 1, characterized in that: Step S1 also includes the following steps: The improved objective function of WGAN-GP for generating samples is: ； in It is the actual data distribution. It generates data distribution. It is between the actual data distribution and the distribution of generated data The sampling distribution in the figure, where L is the loss function, This is a hyperparameter with a value of 10, and D(x) is the discriminator probability distribution. It is regularization; By adding penalty sampling To constrain the Lipschit function, gradient constraints are applied to the generated samples through regularization. The data distribution ranges between the actual data distribution and the generated data distribution; By real data distribution and the distribution of generated data Perform one sample from each of the two data distribution sets, and then perform another random sample from both sets. The formula for determining the value is as follows: ； in It is the actual data distribution. It generates data distribution. It is between the actual data distribution and the distribution of generated data The sampling distribution and random numbers in The values ​​follow a uniform distribution in the range [0,1].

Images