CN114218982A - Micro-seismic record denoising method based on improved WGAN network and CBDNet - Google Patents

Micro-seismic record denoising method based on improved WGAN network and CBDNet Download PDF

Info

Publication number
CN114218982A
CN114218982A CN202111425452.0A CN202111425452A CN114218982A CN 114218982 A CN114218982 A CN 114218982A CN 202111425452 A CN202111425452 A CN 202111425452A CN 114218982 A CN114218982 A CN 114218982A
Authority
CN
China
Prior art keywords
cbdnet
sample
improved
micro
network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111425452.0A
Other languages
Chinese (zh)
Inventor
盛冠群
马凯
余梅
郑悦林
汤婧
张静蓝
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Three Gorges University CTGU
Original Assignee
China Three Gorges University CTGU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Three Gorges University CTGU filed Critical China Three Gorges University CTGU
Priority to CN202111425452.0A priority Critical patent/CN114218982A/en
Publication of CN114218982A publication Critical patent/CN114218982A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The microseism record denoising method based on the improved WGAN network and the CBDNet comprises the following steps: collecting micro-seismic data; generating forward simulation signals under different dominant frequencies and different stratum velocity models by using a finite difference wave equation forward method; setting different signal-to-noise ratios for the obtained forward simulation signal, constructing a noise sample, and then manufacturing an original training set by taking the noise sample and a clean sample as label categories; the prepared original training is utilized to resist and generate a large number of samples by the improved WGAN network, and the sample expansion of the training set is completed; training the CBDNet by using the extended training set sample; and inputting the seismic data to be denoised into the trained CBDNet, and outputting the denoised micro-seismic data. Compared with the traditional micro-seismic denoising scheme, the micro-seismic recording denoising method based on the improved WGAN network and the CBDNet greatly improves the efficiency and the precision of the micro-seismic section denoising.

Description

Micro-seismic record denoising method based on improved WGAN network and CBDNet
Technical Field
The invention relates to a microseism monitoring technology, in particular to a microseism record denoising method based on an improved WGAN network and CBDNet.
Background
The micro-seismic monitoring technology is a geophysical technology for monitoring the underground state by observing and analyzing micro seismic events generated by fracturing, and has important significance for stable and high yield of oil field development. The effective signal energy of the microseism data is weak, the signal-to-noise ratio is low, and even the microseism data is completely submerged in noise. Although the conventional seismic data processing methods are numerous, if the conventional seismic data processing methods are directly applied to microseism data, a satisfactory effect cannot be obtained, and the quality and the accuracy of microseism monitoring are directly influenced. Therefore, finding an appropriate method to identify weak valid signals in the microseismic data is the key to microseismic data processing and interpretation.
The microseism seismic level is weak, the signal-to-noise ratio is low, and how to rapidly and efficiently realize microseism record denoising is of great significance to microseism seismic source positioning, crack prediction and seismic source fracture mechanism analysis. The field micro-seismic monitoring time is long, the monitoring data volume is large, the efficiency of processing the micro-seismic monitoring data can be greatly reduced by manually processing the mass micro-seismic data, and the real-time requirement of processing the mass micro-seismic data cannot be met. On the other hand, the induced microseism signals are influenced by various factors such as the existence of noise sources in the surrounding environment, the distance of signal receivers and the like, and the signals are weak, wide in frequency band and seriously interfered by background noise. Therefore, how to improve the signal-to-noise ratio of the microseism monitoring data and carry out noise suppression are the prerequisites for accurately carrying out microseism data processing.
Disclosure of Invention
In order to efficiently and accurately realize the denoising of the microseism record, the invention provides a microseism record denoising method based on an improved WGAN network and a CBDNet, and compared with the traditional microseism denoising scheme, the method greatly improves the efficiency and the precision of the denoising of the microseism section.
The technical scheme adopted by the invention is as follows:
the microseism record denoising method based on the improved WGAN network and the CBDNet comprises the following steps:
the method comprises the following steps: collecting micro-seismic data;
step two: generating forward simulation signals under different dominant frequencies and different stratum velocity models by using a finite difference wave equation forward method;
step three: setting different signal-to-noise ratios for the forward analog signals obtained in the step two, constructing a noise sample, and then making the noise sample and the clean sample into an original training set by taking the noise sample and the clean sample as label categories;
step four: utilizing the original training made in the third step to generate a large number of samples by using the improved WGAN network to resist, and completing the sample expansion of the training set;
step five: training the CBDNet by using the training set sample expanded in the step four;
step six: and inputting the seismic data to be denoised into the CBDNet trained in the fifth step, and outputting the denoised micro-seismic data.
The fourth step comprises the following steps:
s4.1, deleting an active function layer of the last upsampling layer in the generator, wherein the active function in the generator uses Leaky Relu, and the expression of the active function is as in formula (4):
Figure BDA0003378186810000021
in the formula (4), a is a number close to 0.
And S4.2, removing all normalization layers in the discriminator, replacing the last Sigmoid layer with a 1x1 convolution layer, integrating the characteristics of different channels, and improving the fidelity of signals, wherein an activation function used in the network is Leaky Relu.
In the fifth step:
CBDNet network includes CNNEAnd CNND，CNNEFor obtaining a profile of noisy input y
Figure BDA0003378186810000022
Wherein: wERepresents CNNEThe network parameter of (1); CNNDIs input by y and
Figure BDA0003378186810000023
collectively, the output result of CBDNet is
Figure BDA0003378186810000024
Wherein: wDRepresents CNNDThe network parameter of (1).
In order to further improve the denoising capability of the CBDNet, the following activation function L is adopted:
L＝Lrec+λasymmLasymm+λTVLTV，
wherein:
Lrecin order to reconstruct the loss,
Figure BDA0003378186810000025
LTVis a full-variation regularization operator,
Figure BDA0003378186810000026
asymmetric loss function LasymmExpressed as:
Figure BDA0003378186810000027
the invention discloses a micro-seismic record denoising method based on an improved WGAN network and a CBDNet, which has the following technical effects:
1) the invention aims at the characteristics of low signal-to-noise ratio and large data volume of microseism monitoring signals in the shale gas exploration and development process, and establishes an intelligent denoising method for microseism monitoring under the guidance of deep learning by using the advantages of deep learning in massive data processing capacity and recognition and closely surrounding the scientific problem of noise suppression of effective microseism signals.
2) The improved WGAN network is used for expanding the microseism sample, so that the problem that the deep neural network is difficult to train under the condition of small sample microseism data is solved.
3) And (3) denoising the microseism record by using the depth CBDNet so as to greatly improve the precision and efficiency of denoising the microseism data. The invention combines deep learning and micro-seismic data denoising, and solves the bottleneck that the processing efficiency and precision of the traditional micro-seismic data are difficult to be simultaneously improved.
4) The invention adopts WGAN combined CBDNet to realize the denoising of the microseism record, and can generate two benefits: one is that an important application direction is pointed out for the hydraulic fracture micro-seismic monitoring technology, and the application and popularization of the technology are promoted; and secondly, the microseism data are accurately and quickly denoised, a microseism section with higher signal-to-noise ratio is provided for subsequent microseism positioning and first arrival picking, and compared with the traditional microseism denoising scheme, the method and the device can greatly improve the efficiency and the precision of the microseism section denoising.
Drawings
FIG. 1 is a flow chart of the present invention.
Fig. 2(a) is a first level feature diagram for improving the upsampling of the WGAN generator.
Fig. 2(b) is a third level of characterization for improving the upsampling of the WGAN generator.
Fig. 2(c) is a diagram of the generated signal during training.
Fig. 2(d) is a diagram of the signal generated by the improved WGAN.
Fig. 2(e) is a signal diagram generated by the original WGAN.
Fig. 3(a) is a diagram of the signals generated by the improved WGAN.
Fig. 3(b) is a graph of the transformed spectrum of the improved WGAN-generated signal s.
Figure 4 is a diagram of an improved WGAN generator network architecture.
Fig. 5 is a diagram of an improved WGAN arbiter network architecture.
FIG. 6(a) is a diagram of an actual microseismic recording in a certain area of Shandong.
FIG. 6(b) is a cross-sectional view of denoising microseismic data according to the present invention.
Detailed Description
As shown in fig. 1, the method for denoising the microseism record based on the improved WGAN network and the CBDNet includes the following steps:
step 1: and carrying out underground or ground observation on the fracturing area, and acquiring the hydraulic fracturing micro-seismic monitoring data. The micro-seismic monitoring data has lower signal-to-noise ratio and weak effective signal energy.
Step 2: generating forward simulation signals under different dominant frequencies and different stratum velocity models by using a finite difference wave equation forward method;
the finite difference wave equation forward modeling method is disclosed in a literature [1] contained crown group, microseismic signal detection and seismic source positioning method research [ D ], China university of Petroleum (east, 2016 ]) and is a staggered grid finite difference forward modeling method.
The models with different main frequencies and different formation velocities comprise a Mamousi model, a multilayer flat layer model, an inclined layer velocity model and the like.
The forward analog signals generated by different velocity models are different and are used for ensuring sufficient abundance of the samples.
And step 3: setting different signal-to-noise ratios for the forward analog signals obtained in the step 1, constructing a noise sample together with the actual monitoring profile, and then making an original training set by taking the noise sample and the clean sample as label categories.
The signal-to-noise ratio is given by the following equation:
Figure BDA0003378186810000041
σsand σnStandard deviation of the effective signal and additive noise, respectively.
The clean sample is a forward analog signal with no noise added.
And 4, step 4: and 3, utilizing the original training made in the third step to generate a large number of samples by using the improved WGAN network to resist, and completing the sample expansion of the training set.
The WGAN network principle is as follows:
the WGAN is a measure of the quality of the generated sample set, which is defined as follows:
Figure BDA0003378186810000042
in the formula (1), W (P)r，Pg) Is a probability distribution Pr，PgIs a distance ofIt is the lower bound, Π (P) of the norm mean of the difference between two random variables x, y of the same dimensionr，Pg) Represents PrAnd PgAll possible joint probability distributions, wherein: gamma (x, y) represents Π (P)r，Pg) Each joint probability distribution in the set also represents all the paths that can be taken to move x to y.
To solve the problems of slow convergence rate and insufficient parameter diversity in WGAN training [2 ]]Gulrajani I，Ahmed F，Arjovsky M，et al.Improved Training of Wasserstein GANs[J]2017, restraining the Lipschit function by adding a regular term to perform gradient punishment, and randomly sampling a pair of true and false samples xrAnd xgAnd a random number e; wherein: x is the number ofrCompliance PrDistribution, xgCompliance PgDistribution, ∈ obeys [0, 1]]Uniformly distributed between them, then in the real sample xrAnd generating a sample xgIntermediate random interpolation sampling
Figure BDA0003378186810000043
The value is shown in formula (2):
Figure BDA0003378186810000044
obtaining a new discriminator loss value is shown in equation (3):
Figure BDA0003378186810000045
in the formula (3), the reaction mixture is,
Figure BDA0003378186810000046
in order to generate the mathematical expectation of the sample,
Figure BDA0003378186810000047
a mathematical expectation that is an interpolated sample;
d (x) is the data to be authenticated input by the authenticator, and α is a penalty factor with a value of 10.
Figure BDA0003378186810000051
A penalty term is indicated.
Combining the characteristics of the sampled signal, the present invention makes the following improvements on the original WGAN in order to get a more appropriate signal-generating sample:
firstly, in order to ensure the diversity of the generated signal and not to cut off the negative amplitude of the generated signal, the invention deletes the active function layer of the last upsampling layer in the generator, and the active function in the generator uses Leaky Relu, and the expression is shown in formula (4):
Figure BDA0003378186810000052
in the formula (4), a is a very small number close to 0, and the formula is taken that a is 0.1,
xirepresenting the input feature vector, i.e. the feature map obtained after convolution. The generator network composition is shown in figure 4.
Secondly, because the model adds a gradient penalty to the sample, a normalization layer in the discriminator is completely removed, a last Sigmoid layer is replaced by a convolution layer of 1x1, the characteristics of different channels are integrated, the fidelity of the signal is improved, an activation function used in the network is Leaky Relu, and the network composition of the discriminator is shown in FIG. 5.
And 5: and training the CBDNet by using the training set samples expanded in the step four.
CBDNet network includes CNNEAnd CNND，CNNEFor obtaining a profile of noisy input y
Figure BDA0003378186810000053
Wherein: wERepresents CNNEThe network parameter of (1); CNNDIs input by y and
Figure BDA0003378186810000054
jointly constituting the input of CBDNetThe result is
Figure BDA0003378186810000055
Wherein: wDRepresents CNNDThe network parameter of (1).
In order to further improve the denoising capability of CBDNet, the invention adopts the following activation function L:
L＝Lrec+λasymmLasymm+λTVLTV，
wherein:
Lrecin order to reconstruct the loss,
Figure BDA0003378186810000056
wherein:
Figure BDA0003378186810000057
the output clean signal is x, which is a noisy signal.
LTVIs a full-variation regularization operator,
Figure BDA0003378186810000058
wherein:
Figure BDA0003378186810000059
the gradients in the vertical and horizontal directions are indicated separately,
Figure BDA00033781868100000510
a profile representing the noise level, i.e. the noisy input y, estimated by the noise estimation network. Asymmetric loss function LasymmExpressed as:
Figure BDA00033781868100000511
wherein: if it is not
Figure BDA0003378186810000061
If not, then,
Figure BDA0003378186810000062
alpha is a penalty factor.
Step 6: and inputting the seismic data to be denoised into the CBDNet trained in the fifth step, and outputting the denoised micro-seismic data.
As shown in fig. 2(a) to 2(e), as the generator and the discriminator continue to play a game, the features of the generated signal become finer and the amplitude-frequency characteristics become clearer, as shown in fig. 2(a), 2(b), and 2 (c). For the original WGAN and the improved WGAN generated samples, it is observed that the original WGAN does not learn the characteristics of the signal as shown in fig. 2(d), and a gradient penalty is performed by adding a regularization term to constrain the Lipschit function, so that the problem of parameter centralization is solved, and the quality of the generated signal is further improved as shown in fig. 2 (e). The distribution rule of noise and effective signals can be clearly seen by performing s-transformation on the generated signals, and the characteristics of the effective signals generated by the improved WGAN are obvious as shown in fig. 3(a) and 3 (b).
In order to ensure the diversity of the generated signals and ensure that the negative amplitude of the generated signals is not truncated, the invention deletes the activation function layer of the last upsampling layer in the generator, and the activation function in the generator uses Leaky Relu. As shown in fig. 4, a diagram of a WGAN generator network is improved.
Because the model adds gradient punishment to the sample, the normalization layer in the discriminator is completely removed, the Sigmoid layer of the last layer is replaced by the convolution layer of 1x1, the characteristics of different channels are integrated, and the fidelity of the signal is improved. Fig. 5 shows a diagram of an improved WGAN arbiter network.
FIG. 6(a) is a diagram of an actual microseismic recording in a certain area of Shandong. It can be clearly seen that the recorded signal-to-noise ratio is low, and the effective signal is difficult to be identified, and the recording signal-to-noise ratio needs to be denoised to enhance the recorded signal-to-noise ratio. FIG. 6(b) is a cross-sectional view of the micro-seismic data denoising of the present invention, showing the result of applying the present invention to denoise the record. It can be seen from fig. 6(a) and 6(b) that the method for denoising the microseism record based on the improved WGAN network and the CBDNet can greatly improve the signal-to-noise ratio of the actually monitored microseism section, and has a good denoising effect.

Claims (3)

1. The microseism record denoising method based on the improved WGAN network and the CBDNet is characterized by comprising the following steps of:
the method comprises the following steps: collecting micro-seismic data;
step two: generating forward simulation signals under different dominant frequencies and different stratum velocity models by using a finite difference wave equation forward method;
step three: setting different signal-to-noise ratios for the forward analog signals obtained in the step two, constructing a noise sample, and then making the noise sample and the clean sample into an original training set by taking the noise sample and the clean sample as label categories;
step four: utilizing the original training made in the third step to generate a large number of samples by using the improved WGAN network to resist, and completing the sample expansion of the training set;
step five: training the CBDNet by using the training set sample expanded in the step four;
step six: and inputting the seismic data to be denoised into the CBDNet trained in the fifth step, and outputting the denoised micro-seismic data.
2. The method of claim 1 for denoising microseismic records based on an improved WGAN network and CBDNet, wherein: the fourth step comprises the following steps:
s4.1, deleting an active function layer of the last upsampling layer in the generator, wherein the active function in the generator uses Leaky Relu, and the expression of the active function is as in formula (4):
Figure FDA0003378186800000011
in the formula (4), a is a number close to 0;
and S4.2, removing all normalization layers in the discriminator, replacing the last Sigmoid layer with a 1x1 convolution layer, integrating the characteristics of different channels, and improving the fidelity of signals, wherein an activation function used in the network is Leaky Relu.
3. The method of claim 1 for denoising microseismic records based on an improved WGAN network and CBDNet, wherein: in the fifth step:
CBDNet network includes CNNEAnd CNND，CNNEFor obtaining a profile of noisy input y
Figure FDA0003378186800000012
Wherein: wERepresents CNNEThe network parameter of (1); CNNDIs input by y and
Figure FDA0003378186800000013
collectively, the output result of CBDNet is
Figure FDA0003378186800000014
Wherein: wDRepresents CNNDThe network parameter of (1);
in order to further improve the denoising capability of the CBDNet, the following activation function L is adopted:
L＝Lrec+λasymmLasymm+λTVLTV，
wherein:
Lrecin order to reconstruct the loss,
Figure FDA0003378186800000021
LTVis a full-variation regularization operator,
Figure FDA0003378186800000022
asymmetric loss function LasymmExpressed as:
Figure FDA0003378186800000023
CN202111425452.0A 2021-11-26 2021-11-26 Micro-seismic record denoising method based on improved WGAN network and CBDNet Pending CN114218982A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111425452.0A CN114218982A (en) 2021-11-26 2021-11-26 Micro-seismic record denoising method based on improved WGAN network and CBDNet

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111425452.0A CN114218982A (en) 2021-11-26 2021-11-26 Micro-seismic record denoising method based on improved WGAN network and CBDNet

Publications (1)

Publication Number Publication Date
CN114218982A true CN114218982A (en) 2022-03-22

Family

ID=80698625

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111425452.0A Pending CN114218982A (en) 2021-11-26 2021-11-26 Micro-seismic record denoising method based on improved WGAN network and CBDNet

Country Status (1)

Country Link
CN (1) CN114218982A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115310488A (en) * 2022-08-16 2022-11-08 哈尔滨工业大学 Seismic oscillation recording filtering method based on generating type antagonistic neural network
CN116681618A (en) * 2023-06-13 2023-09-01 强联智创（北京）科技有限公司 Image denoising method, electronic device and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115310488A (en) * 2022-08-16 2022-11-08 哈尔滨工业大学 Seismic oscillation recording filtering method based on generating type antagonistic neural network
CN116681618A (en) * 2023-06-13 2023-09-01 强联智创（北京）科技有限公司 Image denoising method, electronic device and storage medium

Similar Documents

Publication Publication Date Title
CN108931814B (en) Multi-attribute fusion based bedrock fracture prediction method
CN114218982A (en) Micro-seismic record denoising method based on improved WGAN network and CBDNet
CA3122509A1 (en) Machine learning-augmented geophysical inversion
CN107505654A (en) Full waveform inversion method based on earthquake record integration
CN110018517B (en) Multi-scale ground micro-seismic reverse-time interference positioning method
CN111045077B (en) Full waveform inversion method of land seismic data
CN110058305A (en) A kind of DAS seismic data noise-reduction method based on convolutional neural networks
WO2012139082A1 (en) Event selection in the image domain
CN109613610B (en) Automatic picking method for microseism signal arrival time difference
CN112444841B (en) Thin-layer-containing lithology earthquake prediction method based on scale-division multi-input convolution network
CN110632662A (en) Algorithm for automatically identifying microseism signals by using DCNN-inclusion network
RU2412454C2 (en) Method to process seismic data using discrete wavelet transform
CN113077386A (en) Seismic data high-resolution processing method based on dictionary learning and sparse representation
CN106990449B (en) multi-dimensional data-based continental facies sequence stratigraphic framework construction characterization method
CN110118994B (en) Continental facies hydrocarbon source rock quantitative prediction method based on seismic inversion and machine learning
CN115826039B (en) Time slice classification model training method and system and application method and system
CN107679614B (en) Particle swarm optimization-based real-time sound wave time difference extraction method
CN106291676A (en) A kind of geological data reconstructing method based on matching pursuit algorithm
CN106291675A (en) A kind of geological data reconstructing method based on base tracer technique
CN111025397B (en) Method for obtaining depth domain velocity model by combining seismic data reflected wave and scattered wave
CN113589365B (en) Reservoir pinch-out line description method based on time-frequency domain information
CN110873895A (en) Variable grid micro-seismic reverse-time interference positioning method
CN115453621B (en) Longitudinal and transverse wave decoupling separation false image removing method based on first-order speed-stress equation
CN112462428B (en) Multi-component seismic data migration imaging method and system
CN113835123B (en) Seismic acquisition parameter analysis method based on geological target pre-stack migration imaging

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination