CN114706128A - Microseism data noise suppression method and system, storage medium and seismic information processing equipment - Google Patents

Microseism data noise suppression method and system, storage medium and seismic information processing equipment Download PDF

Info

Publication number
CN114706128A
CN114706128A CN202210278729.XA CN202210278729A CN114706128A CN 114706128 A CN114706128 A CN 114706128A CN 202210278729 A CN202210278729 A CN 202210278729A CN 114706128 A CN114706128 A CN 114706128A
Authority
CN
China
Prior art keywords
model
noise
micro
data
seismic
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
CN202210278729.XA
Other languages
Chinese (zh)
Inventor
董宏丽
商柔
王闯
管闯
宋金波
申雨轩
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.)
Northeast Petroleum University
Original Assignee
Northeast Petroleum University
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 Northeast Petroleum University filed Critical Northeast Petroleum University
Priority to CN202210278729.XA priority Critical patent/CN114706128A/en
Publication of CN114706128A publication Critical patent/CN114706128A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. analysis, for interpretation, for correction
    • G01V1/36—Effecting static or dynamic corrections on records, e.g. correcting spread; Correlating seismic signals; Eliminating effects of unwanted energy
    • G01V1/364—Seismic filtering

Abstract

A method and a system for suppressing the noise of micro-seismic data, a storage medium and a seismic information processing device belong to the micro-seismic data processing technology and aim to solve the problems that the existing method can not eliminate all noise and can also lose effective signals. The technical points are as follows: collecting microseism data of a target block by using a detector; dividing the acquired micro-seismic data into a training set, a verification set and a test set according to a proportion, inputting the training set data into a time attention mechanism mutual information generation countermeasure network model for learning and outputting a micro-seismic noise suppression pre-training model; inputting verification set data into a trained model, and verifying model parameters by using signal-to-noise ratio to obtain an optimal micro-seismic noise pressure system model hyper-parameter combination; testing the optimal hyper-parameter combination by using the test set data and obtaining a final micro-seismic noise pressure modeling model; and inputting the field micro-seismic data into a final micro-seismic noise pressure control model to obtain de-noised micro-seismic data, and realizing micro-seismic data noise control.

Description

Microseism data noise suppression method and system, storage medium and seismic information processing equipment
Technical Field
The invention relates to a method and a system for suppressing microseism data noise, a storage medium and seismic information processing equipment, and belongs to the technical field of microseism data processing.
Background
In recent years, the potential of conventional oil and gas exploitation in China is gradually reduced, but the oil and gas demand caused by rapid development of national industry is continuously improved, and the external dependence of oil in China is as high as 73.5% by 2020. The unconventional oil gas in China can be used for detecting about 1075 hundred million tons of reserves, the development potential is huge, and the development of unconventional oil gas exploration and development is a necessary way for ensuring the energy safety in China. The microseism monitoring technology is a key technology for evaluating the hydraulic fracturing effect and acquiring parameters such as the azimuth, the length and the height of a crack, and is an important means for guaranteeing the efficient development of unconventional oil and gas.
In the practical engineering, the ground micro-seismic detector not only collects a large amount of random noise but also contains strong interference caused by complex ground noise sources (such as 50Hz industrial alternating current interference, drilling interference and the like) in the data acquisition process, so that a micro-seismic effective signal is easily submerged, and the identification of a micro-seismic effective event and the accuracy and reliability of seismic source positioning are seriously influenced. Therefore, noise suppression is a core element of microseismic data preprocessing. However, in establishing a noise suppression model, we find that the conventional noise reduction method usually takes subjective expert experience guidance as a main guide when distinguishing effective signals from noise, that is, the noise signal is considered to exist only in the high-frequency part of the whole signal, and then the high-frequency noise is eliminated by using low-pass filtering. However, noise signals contained in the micro-seismic signals have coupling, and the subjective identification of the noise frequency and the corresponding method for elimination cannot eliminate all the noise and also lose effective signals. In addition, the effective event characteristic distribution of the microseisms presents discretization and only exists in local positions of the time series. The microseism signals collected by different detectors also present different characteristics due to geological changes, so that loss of effective events is easily caused in the noise suppression process of microseism data. Therefore, finding new ways to suppress noise in microseismic data without losing significant events has become paramount in unconventional oil and gas exploration development.
Disclosure of Invention
In order to solve the technical problems mentioned in the background art: the traditional denoising method generally takes subjective expert experience guidance as a main guide when distinguishing effective signals and noise, namely, the noise signals are only determined to exist in the high-frequency part of the whole signals, and then the low-pass filtering is utilized to eliminate the high-frequency noise. However, noise signals contained in the micro-seismic signals have coupling, and the noise frequency is subjectively determined and eliminated by adopting a corresponding method, so that not only can all the noise be eliminated, but also effective signals can be lost, and therefore the micro-seismic data noise suppression method and system, the storage medium and the seismic information processing equipment are provided, wherein the micro-seismic data noise suppression method and system are based on the time attention mechanism mutual information generation countermeasure network.
The technical scheme adopted by the invention for solving the technical problems is as follows: a microseism data noise suppression method for generating a countermeasure network based on time attention mechanism mutual information is characterized by comprising the following steps:
step 1: collecting the micro-seismic data of the target block by using a detector to obtain the micro-seismic data X with noise and the non-noise micro-seismic data
Figure BDA0003557213690000021
Step 2: dividing the microseism data acquired in the step 1 into a training set, a verification set and a test set according to the ratio of 8:1: 1; the training set is represented as
Figure BDA0003557213690000022
Wherein xkThe K training trace set samples are K training samples in total; the verification set is represented as
Figure BDA0003557213690000023
Wherein xaThe verification trace set sample is the a-th verification trace set sample, and the A-th verification trace set sample is the A-th verification trace set sample; the test set is represented as
Figure BDA0003557213690000024
Wherein xbB test trace set samples are the B test trace set samples; the sampling frequency of the training set, the verification set and the test set is f;
and step 3: extraction of the samples collected in step 1
Figure BDA0003557213690000025
Structural semantic feature set of
Figure BDA0003557213690000026
Wherein, cd∈RH×1The d-th micro-seismic structure semantic vector is H;
and 4, step 4: sampling a random noise distribution (such as Gaussian distribution) with a sampling frequency f, and acquiring a random noise vector set
Figure BDA0003557213690000027
Wherein z ism∈RN×1The m-th random noise vector is composed of N sampling points;
and 5: subjecting the product obtained in step 3
Figure BDA0003557213690000028
With the product obtained in step 4
Figure BDA0003557213690000029
Dimension combination is carried out to obtain feature vectors
Figure BDA00035572136900000210
Wherein v isu∈R(H+N)×1Is the u-th merged feature vector;
step 6: the method comprises the following steps of utilizing a countermeasure network generated based on time attention mechanism mutual information to learn training set data and output a trained microseism noise suppression pre-training model, wherein the specific path is as follows:
the microseism noise pressing model adopted in the step comprises a generation model, a discrimination model and a separation model;
the generating model is composed of a time attention module and a generating module, wherein the time attention module is specifically defined as follows:
Figure BDA00035572136900000211
wherein the input to the temporal attention module is v obtained via step 5u,j(ii) a The output of the temporal attention module is the weight αu；σtRepresents the tanh activation function;
Figure BDA00035572136900000212
a jth channel representing a convolution kernel having a size of Z' x 1;
Figure BDA00035572136900000213
is a bias term;
the generation module consists of a Dense layer, a Batch Normalization layer, a ReLU layer, a Deconv layer and a Tanh layer; the discrimination model is composed of a time attention module and a discrimination module, wherein the time attention module is specifically defined as follows:
Figure BDA00035572136900000214
wherein the input to the temporal attention module is obtained via step 2
Figure BDA00035572136900000215
The output of the temporal attention module is the weight βk；σsRepresents the softmax activation function;
Figure BDA00035572136900000216
a jth channel representing a convolution kernel having a convolution kernel size of F' × 1;
Figure BDA00035572136900000217
is a bias term;
the discrimination module consists of a Dense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Softmax layer;
the separation model is composed of a time attention module and a separation module, wherein the time attention module is defined as:
Figure BDA0003557213690000031
wherein the input to the temporal attention module is G (z, c) obtained via the generative modelh,j(ii) a The output of the temporal attention module is the weight γh；σsigRepresenting a sigmoid activation function;
Figure BDA0003557213690000032
a jth channel representing a convolution kernel having a size of Y' × 1;
Figure BDA0003557213690000033
is a bias term;
the separation module consists of a Dense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Sigmoid layer;
subjecting the product obtained in step 5
Figure BDA0003557213690000034
Inputting the data into a generating model to obtain a generated noise-free micro seismic data set; the generated noise-free micro-seismic data set and the data obtained in the step 2
Figure BDA0003557213690000035
Respectively inputting the data into a discrimination model, training the discrimination model and reversely transmitting the gradient to a generation model; inputting the generated noise-free micro-seismic data set into a separation model to extract structural semantic features, calculating mutual information between the generated noise-free micro-seismic data and the extracted structural semantic features, training the separation model, and reversely transmitting the gradient to the generation model; finally, when the maximum iteration times are reached, a pre-training model for suppressing the microseism noise can be obtained;
and 7: verification set data to be obtained via step 2
Figure BDA0003557213690000036
Inputting the data into the noise suppression pre-training model obtained in the step 6, verifying the performance of the trained noise suppression pre-training model by utilizing signal-to-noise ratio, and obtaining the optimal hyper-parameter combination of the micro-seismic noise suppression model;
and 8: utilizing test set data obtained via step 2
Figure BDA0003557213690000037
Testing the optimal hyper-parameter combination of the microseism noise pressing model obtained in the step 7, and obtaining a final microseism data noise pressing model;
and step 9: and (4) inputting the field micro-seismic data into the final micro-seismic noise suppression model obtained in the step (8) to obtain the de-noised micro-seismic data, so as to suppress the noise of the micro-seismic data.
Further, the generative model training process in step 6 is performed according to the following path:
2.1 the generative model is composed of a time attention module and a generation module, wherein the time attention module is defined as:
Figure BDA0003557213690000038
where the input to the temporal attention module is v obtained via step 5u,j(ii) a The output of the temporal attention module is the weight αu；σtRepresents the tanh activation function;
Figure BDA0003557213690000039
a jth channel representing a convolution kernel having a size of Z' x 1;
Figure BDA00035572136900000310
is the bias term.
2.2 v to be obtained via step 5u,jAnd alpha obtained via equation (1)uIs input into formula (2) to obtain an increaseStrong feature mapping vattSpecifically, the following are defined:
Figure BDA00035572136900000311
wherein v isattFor enhanced feature mapping of merged feature vectors, vuFor the u-th merged feature vector, αuThe temporal attention weight of the u-th merged feature vector.
2.3 v to be obtained via equation (2)attInputting the data into a generation module to obtain a generated noise-free micro-seismic data set
Figure BDA0003557213690000041
Wherein, G (z, c)h∈RS×1 is the noise-free microseism data generated at the h-th; the generating module is composed of a Dense layer, a Batch Normalization layer, a ReLU layer, a Deconv layer and a Tanh layer.
Further, the discriminant model training process in step 6 is performed according to the following path:
3.1 the discriminant model is composed of a time attention module and a discriminant module, wherein the time attention module is defined as:
Figure BDA0003557213690000042
wherein the input to the temporal attention module is obtained via step 2
Figure BDA0003557213690000043
The output of the temporal attention module is the weight βk；σsRepresents the softmax activation function;
Figure BDA0003557213690000044
a jth channel representing a convolution kernel having a convolution kernel size of F' × 1;
Figure BDA0003557213690000045
is offset byAn item.
3.2 obtaining via step 2
Figure BDA0003557213690000046
With β obtained via equation (3)kInput into formula (4) to obtain an enhanced feature map xattSpecifically, the following are defined:
Figure BDA0003557213690000047
wherein xattIs composed of
Figure BDA0003557213690000048
Enhanced feature mapping of (2);
Figure BDA0003557213690000049
for the kth training set sample, βkA temporal attention weight for the kth training set sample;
3.3 x to be obtained via equation (4)attInput into a discrimination module to obtain
Figure BDA00035572136900000410
Probability of belonging to a true sample. The discrimination module is composed of a sense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Softmax layer.
Further, the separation model training process in step 6 is performed according to the following path:
4.1 the separation model is composed of a time attention module and a separation module, wherein the time attention module is defined as:
Figure BDA00035572136900000411
wherein the input to the temporal attention module is G (z, c) obtained via the generative modelh,j(ii) a The output of the temporal attention module is the weight γh；σsigRepresenting a sigmoid activation function;
Figure BDA00035572136900000412
a jth channel representing a convolution kernel having a size of Y' × 1;
Figure BDA00035572136900000413
is the bias term.
4.2 to be obtained via feature 2.3
Figure BDA00035572136900000414
And gamma obtained via the formula (5)hInput into equation (6) to obtain an enhanced feature map qattSpecifically, the following are defined:
Figure BDA00035572136900000415
wherein q isattEnhanced feature mapping for G (z, c); g (z, c)hFor the h-th generated noiseless microseismic data, gammahTime attention weight for the h-th generated noise-free microseismic data.
4.3 q to be obtained via equation (6)attInputting into a separation module to obtain
Figure BDA00035572136900000416
Structural semantic vector of
Figure BDA00035572136900000417
Wherein
Figure BDA00035572136900000418
For the w-th extracted structural semantic feature vector, and calculate G (z, c) and cqMutual information of (2); the separation module is composed of a sense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Sigmoid layer.
Further, the model verification process in step 7 is performed according to the following path:
verification set data to be obtained by step 2
Figure BDA0003557213690000051
Inputting the data into the noise suppression pre-training model obtained in the step 6, verifying the performance of the trained noise suppression pre-training model by using the signal-to-noise ratio, and obtaining the optimal hyper-parameter combination of the micro-seismic noise suppression model.
Further, the model testing process in step 8 is performed according to the following path:
test set data to be obtained by step 2
Figure BDA0003557213690000052
And 7, testing the optimal hyper-parameter combination of the micro-seismic noise suppression model obtained in the step 7, and obtaining a final micro-seismic data noise suppression model.
Further, the noise suppression process in step 9 is performed according to the following path: and (4) inputting the field micro-seismic data into the final micro-seismic noise suppression model obtained in the step (8) to obtain the de-noised micro-seismic data, so as to suppress the noise of the micro-seismic data.
A micro-seismic data noise suppression system for generating a countermeasure network based on time attention mechanism mutual information is provided with a program module corresponding to the steps of the technical scheme, and the steps in the micro-seismic data noise suppression method for generating the countermeasure network based on the time attention mechanism mutual information are executed during running.
A computer readable storage medium having stored thereon a computer program configured to, when invoked by a processor, perform the steps of the method for noise suppressing microseismic data based on time attention mechanism mutual information generation countermeasure networks described above.
A seismic information processing apparatus comprising at least one processor, and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executable by the at least one processor to enable the at least one processor to perform a method of noise suppressing microseismic data based on time attention mechanism mutual information generation countermeasure network as described above.
The invention has the following beneficial technical effects:
the microseism noise suppression method overcomes the defects that the coupling noise cannot be completely removed and the loss of microseism effective events is easily caused by the traditional filtering method. The traditional denoising method generally takes subjective expert experience guidance as a main guide when distinguishing effective signals and noise, namely, the noise signals are only determined to exist in the high-frequency part of the whole signals, and then the low-pass filtering is utilized to eliminate the high-frequency noise. However, noise signals contained in the micro-seismic signals have coupling, and the subjective identification of the noise frequency and the corresponding method for elimination cannot eliminate all the noise and also lose effective signals. In addition, the effective event characteristic distribution of the microseisms presents discretization and only exists in local positions of the time series. The microseism signals collected by different detectors also present different characteristics due to geological changes, so that loss of effective events is easily caused in the noise suppression process of microseism data.
The method comprises the following steps: collecting microseism data of a target block by using a detector; dividing the acquired micro-seismic data into a training set, a verification set and a test set according to a certain proportion, inputting the training set data into time attention mechanism mutual information to generate a confrontation network model for learning, and outputting a micro-seismic noise suppression pre-training model; inputting verification set data into a trained model, verifying model parameters by utilizing signal-to-noise ratio, and obtaining an optimal micro-seismic noise pressure system model hyper-parameter combination; testing the optimal hyper-parameter combination by using the test set data and obtaining a final micro-seismic noise pressure modeling model; and inputting the field micro-seismic data into a final micro-seismic noise pressure control model to obtain de-noised micro-seismic data, and realizing micro-seismic data noise control.
The invention considers that the noise contained in the microseism data has the characteristics of strong coupling and large energy, namely the noise characteristic and the microseism data characteristic are coded in a complex and disordered state in a data space, so that the ideal denoising effect is difficult to achieve by the traditional method. Therefore, the method provided by the invention can be used for learning the noise distribution in the micro-seismic data by introducing the potential noise vector and maximizing the mutual information between the potential noise vector distribution and the generated micro-seismic data so as to achieve the purpose of removing the noise. Finally, the effective event characteristic distribution in the microseism monitoring data is considered to be discretized and only exists in the local position of the time sequence. Therefore, the invention adopts a time attention mechanism to amplify the characteristics of the effective events, avoids the phenomenon that the effective events are lost in the denoising process, and further obtains better denoising effect. The invention provides a generation countermeasure network based on mutual information, which solves the problem that strong coupling noise is difficult to remove in microseism noise suppression, introduces a time attention mechanism to amplify effective events in microseism data, and avoids the phenomenon that the effective events are lost in the denoising process. A set of systematic and scientific microseism data noise suppression method is provided, and the method has operability and practical value.
Therefore, the method can effectively solve the problems and amplify the microseism effective events while realizing microseism noise suppression.
The effectiveness of the method of the invention is reflected in the summary of the invention:
1. the method is built by adopting a deep convolutional neural network, has excellent nonlinear characterization capability, and improves the precision of microseism noise suppression on the network structure level.
2. The method adopts a time attention mechanism to amplify the local characteristics of the micro-seismic data time sequence effective events, can avoid the loss of effective signals in the micro-seismic data in the process of realizing noise suppression, and improves the precision of micro-seismic noise suppression on the characteristic level of the data.
3. The method adopts the calculation mutual information to strengthen the correlation between the effective characteristics of the micro-seismic data and the micro-seismic signals, can effectively suppress the coupling noise contained in the micro-seismic data, and improves the precision of suppressing the micro-seismic noise on the algorithm level.
In addition, the method belongs to an unsupervised learning method based on data driving, dependence on manual labeling is eliminated, and more manpower and material resources are reduced. On-site micro-seismic data can be directly subjected to noise suppression through a trained model, a micro-seismic data processing technology which is easier to operate is provided for on-site constructors, the on-site working efficiency is improved, and the best effect is achieved.
Drawings
FIG. 1 is a block flow diagram of a microseismic noise suppression model (method and system) according to an embodiment of the present invention;
FIG. 2 is a diagram of a temporal attention mechanism mutual information generation countermeasure network core, according to an embodiment of the invention;
FIGS. 3 and 4 are diagrams of examples of noise suppressed visualizations of microseismic data according to embodiments of the invention, FIG. 3 is a diagram of examples of noise suppressed visualizations of microseismic data (noisy microseismic noise data), FIG. 4 is a diagram of examples of noise suppressed visualizations of microseismic data (denoised microseismic noise data); FIGS. 5 and 6 are diagrams of another example of noise-suppressed visualization of microseismic data (noisy microseismic noise data);
in fig. 4 and 6, the four lines after denoising represent the in-phase axes (the in-phase axes are the lines of extreme values of the same vibration phase of each trace on the seismic record), and represent the layering of the stratum; the two groups of the two images have different directions of the same phase axis and different straight potentials, and represent different geological structures. The provided figure shows that the invention can be used for compressing a plurality of different geological structure data, and the practicability of the algorithm of the invention can be better embodied.
Detailed Description
The invention will be further described with reference to the accompanying drawings in which:
in order to make those skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.
Fig. 1 is a flow chart of microseismic noise suppression in the technical solution of the present invention, and the whole flow chart is completely implemented by Python language programming (of course, other programming languages may be used to implement the flow chart). According to the figure 1, firstly, on-site micro-seismic noise-free data and noise-containing data are collected, then the data are divided into a training set, a verification set and a test set according to a specific proportion, and the training set data are input into time attention mechanism mutual information to generate a pre-training model for learning in a countermeasure network model and outputting micro-seismic noise suppression; inputting verification set data into a trained model, verifying model parameters by utilizing signal-to-noise ratio, and obtaining an optimal micro-seismic noise pressure system model hyper-parameter combination; testing the optimal hyper-parameter combination by using the test set data, and obtaining a final micro-seismic noise pressure modeling model; and inputting the field micro-seismic data into a final micro-seismic noise pressure making model to obtain de-noised micro-seismic data, thereby realizing micro-seismic data noise suppression.
In the embodiment, the validity of the method is verified by taking field microseism data as an example.
Step 1: and acquiring the micro-seismic data of the target block by using a detector to obtain a micro-seismic data signal A with noise and a micro-seismic data signal B without noise.
Step 2: and (3) dividing the microseism data acquired in the step (1) into a training set, a verification set and a test set according to the ratio of 8:1: 1. The training set is represented as
Figure BDA0003557213690000071
Wherein x iskIs the k training set sample; the verification set is represented as
Figure BDA0003557213690000072
Wherein x isaIs the a-th verification set sample; the test set is represented as
Figure BDA0003557213690000073
Wherein x isbIs the b test set sample. And the lengths of the sampling points of the training set, the verification set and the test set are all S.
And step 3: extracting the structural semantic feature set of the signal B acquired in the step 1
Figure BDA0003557213690000074
Wherein, cd∈RH×1The length of the d micro-seismic structure semantic vector is H.
And 4, step 4: sampling a random noise distribution (e.g., gaussian distribution) and obtaining a set of random noise vectors
Figure BDA0003557213690000075
Wherein z ism∈RN×1And the m-th random noise vector is composed of N sampling points.
And 5: subjecting the product obtained in step 3
Figure BDA0003557213690000076
With the product obtained in step 4
Figure BDA0003557213690000077
Dimension combination is carried out to obtain a combined feature vector
Figure BDA0003557213690000078
Wherein v isu∈R(H+N)×1Is the u-th merged feature vector.
And 6: referring to fig. 2, a countermeasure network is generated by using time attention mechanism mutual information to learn training set data and output a trained microseism noise suppression pre-training model, and the specific path is as follows:
the microseism noise pressing model adopted in the step comprises a generation model, a discrimination model and a separation model.
The generating model is composed of a time attention module and a generating module, wherein the time attention module is specifically defined as:
Figure BDA0003557213690000081
wherein the input to the temporal attention module is v obtained via step 5u,j(ii) a The output of the temporal attention module is the weight αu；σtRepresents the tanh activation function;
Figure BDA0003557213690000082
a jth channel representing a convolution kernel having a size of Z' x 1;
Figure BDA0003557213690000083
is the bias term.
The generation module is composed of a sense layer, a Batch Normalization layer, a ReLU layer, a Deconv layer and a Tanh layer, and belongs to the field of the prior art.
The discrimination model is composed of a time attention module and a discrimination module, wherein the time attention module is specifically defined as follows:
Figure BDA0003557213690000084
wherein the input to the temporal attention module is obtained via step 2
Figure BDA0003557213690000085
The output of the temporal attention module is the weight βk；σsRepresents the softmax activation function;
Figure BDA0003557213690000086
a jth channel representing a convolution kernel having a convolution kernel size of F' × 1;
Figure BDA0003557213690000087
is the bias term.
The discrimination module is composed of a Dense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Softmax layer.
The separation model is composed of a time attention module and a separation module, wherein the time attention module is defined as follows:
Figure BDA0003557213690000088
wherein, whenThe input to the inter-attention module is G (z, c) obtained via the generative modelh,j(ii) a The output of the temporal attention module is the weight γh；σsigRepresenting a sigmoid activation function;
Figure BDA0003557213690000089
a jth channel representing a convolution kernel having a size of Y' × 1;
Figure BDA00035572136900000810
is the bias term.
The separation module is composed of a Dense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Sigmoid layer.
FIG. 2 illustrates a microseismic noise pressure modeling model: subjecting the product obtained in step 5
Figure BDA00035572136900000811
Inputting the data into a generating model to obtain a generated noise-free micro seismic data set; collecting the generated noise-free micro-seismic data and the data obtained in step 2
Figure BDA00035572136900000812
Respectively inputting the data into a discrimination model, training the discrimination model and reversely transmitting the gradient to a generation model; the generated noise-free micro-seismic data set is input into a separation model to extract structural semantic features, mutual information of the generated noise-free micro-seismic data and the extracted structural semantic features is calculated, the separation model is trained, and the gradients are reversely transmitted to the generation model, so that micro-seismic can be enhanced, the correlation between effective features of the micro-seismic data and micro-seismic signals can be enhanced through the method, and coupling noise in the micro-seismic data can be effectively suppressed.
And 7: verification set data to be obtained via step 2
Figure BDA00035572136900000813
Inputting the noise suppression pre-training model obtained in the step 6, verifying the performance of the trained noise suppression pre-training model by using the signal-to-noise ratio, and obtainingAnd (3) carrying out optimal hyper-parameter combination on the microseism noise suppression model.
And step 8: utilizing test set data obtained via step 2
Figure BDA0003557213690000091
And 7, testing the optimal hyper-parameter combination of the micro-seismic noise suppression model obtained in the step 7, and obtaining a final micro-seismic data noise suppression model.
TABLE 1 comparison of different denoising model effects
Figure BDA0003557213690000092
And step 9: and (4) inputting the field micro-seismic data into the final micro-seismic noise suppression model obtained in the step (8) to obtain the de-noised micro-seismic data, so as to suppress the noise of the micro-seismic data. The results of the noise suppression experiment for the microseismic data are shown in fig. 3 and 4 and fig. 5 and 6.
The provided figure shows that the method can be used for suppressing a plurality of different geological structure data, and the practicability of the algorithm can be better shown. The method adopts a time attention mechanism to amplify the characteristics of the effective events, avoids the phenomenon that the effective events are lost in the denoising process, further obtains a better denoising effect, and has good universality.
The method provided by the invention verifies the claimed technical effect through simulation experiments and practical application.
The algorithm (method) provided by the invention is the bottom technical kernel of the invention, and various products can be derived based on the algorithm.
The algorithm (method) provided by the invention utilizes a program language to develop a micro-seismic data noise suppression system for generating a confrontation network based on the mutual information of the time attention mechanism, the system is provided with a program module corresponding to the steps of the technical scheme, and the steps in the micro-seismic data noise suppression method for generating the confrontation network based on the mutual information of the time attention mechanism are executed during running.
A computer program of the developed system (software) is stored on a computer readable storage medium, the computer program being configured to implement the steps of the above-described method of noise-suppressing microseismic data for generating a counterpoise network based on time attention mechanism mutual information when invoked by a processor. I.e. the invention may be embodied on a carrier, to form a computer program product.
A seismic information processing apparatus comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executable by the at least one processor to enable the at least one processor to perform a method of noise suppressing microseismic material based on time attention mechanism mutual information generation countermeasure network as described above. The seismic information processing equipment is used as a terminal product applied by the invention.
Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.
The computer programs (also known as programs, software applications, or code) in the present invention include machine instructions for a programmable processor, and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.
It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present application may be executed in parallel, sequentially, or in different orders, and are within the scope of the present invention as long as the desired results of the technical solutions disclosed in the present application can be achieved.

Claims (10)

1. A microseism data noise suppression method for generating a countermeasure network based on time attention mechanism mutual information is characterized by comprising the following steps:
step 1: collecting the micro-seismic data of the target block by using a detector to obtain the micro-seismic data X with noise and the non-noise micro-seismic data
Figure FDA0003557213680000011
And 2, step: dividing the microseism data acquired in the step 1 into a training set, a verification set and a test set according to the ratio of 8:1: 1; the training set is represented as
Figure FDA0003557213680000012
Wherein xkFor the kth training trace set sample, K training samples are shared; the verification set is represented as
Figure FDA0003557213680000013
Wherein xaThe verification trace set sample is the a-th verification trace set sample, and the A-th verification trace set sample is the A-th verification trace set sample; the test set is represented as
Figure FDA0003557213680000014
Wherein x isbB test trace set samples are the B test trace set samples; the sampling frequency of the training set, the verification set and the test set is f;
and step 3: extraction of the samples collected in step 1
Figure FDA0003557213680000015
Structural semantic feature set of
Figure FDA0003557213680000016
Wherein, cd∈RH×1The d-th micro-seismic structure semantic vector is H;
and 4, step 4: sampling random noise distribution with sampling frequency f, and acquiring random noise vector set
Figure FDA0003557213680000017
Wherein z ism∈RN×1The m-th random noise vector is composed of N sampling points;
and 5: subjecting the product obtained in step 3
Figure FDA0003557213680000018
And obtained in step 4
Figure FDA0003557213680000019
Dimension combination is carried out to obtain feature vectors
Figure FDA00035572136800000110
Wherein v isu∈R(H+N)×1Merging the feature vectors into the u-th merged feature vector;
step 6: the method comprises the following steps of utilizing a time attention mechanism-based mutual information to generate a confrontation network to learn training set data and output a trained microseism noise suppression pre-training model, wherein the specific path is as follows:
the microseism noise pressing model adopted in the step comprises a generation model, a discrimination model and a separation model;
the generating model is composed of a time attention module and a generating module, wherein the time attention module is specifically defined as:
Figure FDA00035572136800000111
wherein the input to the temporal attention module is v obtained via step 5u,j(ii) a The output of the temporal attention module is the weight αu；σtRepresents the tanh activation function;
Figure FDA00035572136800000112
a jth channel representing a convolution kernel having a size of Z' x 1;
Figure FDA00035572136800000113
is a bias term;
the generation module consists of a Dense layer, a Batch Normalization layer, a ReLU layer, a Deconv layer and a Tanh layer;
the discrimination model is composed of a time attention module and a discrimination module, wherein the time attention module is specifically defined as follows:
Figure FDA00035572136800000114
wherein the input to the temporal attention module is obtained via step 2
Figure FDA00035572136800000115
The output of the temporal attention module is the weight βk；σsRepresents the softmax activation function;
Figure FDA00035572136800000116
representing the jth channel of the convolution kernel, wherein the size of the convolution kernel is F' multiplied by 1;
Figure FDA00035572136800000117
is a bias term;
the discrimination module consists of a Dense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Softmax layer;
the separation model is composed of a time attention module and a separation module, wherein the time attention module is defined as:
Figure FDA0003557213680000021
wherein the input to the temporal attention module is G (z, c) obtained via the generative modelh,j(ii) a The output of the temporal attention module is the weight γh；σsigRepresenting a sigmoid activation function;
Figure FDA0003557213680000022
a jth channel representing a convolution kernel having a size of Y' × 1;
Figure FDA0003557213680000023
is a bias term;
the separation module consists of a Dense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Sigmoid layer;
subjecting the product obtained in step 5
Figure FDA0003557213680000024
Inputting the data into a generating model to obtain a generated noise-free micro seismic data set; the generated noise-free micro-seismic data set and the data obtained in the step 2
Figure FDA0003557213680000025
Respectively inputting the data into a discrimination model, training the discrimination model and reversely transmitting the gradient to a generation model; inputting the generated noiseless micro-seismic data set into a separation model to extract structural semantic features, calculating mutual information between the generated noiseless micro-seismic data and the extracted structural semantic features, and carrying out separation modelingTraining is carried out, and the gradient is reversely transmitted to a generated model; finally, when the maximum iteration times are reached, a pre-training model for suppressing the microseism noise can be obtained;
and 7: verification set data to be obtained via step 2
Figure FDA0003557213680000026
Inputting the data into the noise suppression pre-training model obtained in the step 6, verifying the performance of the trained noise suppression pre-training model by utilizing signal-to-noise ratio, and obtaining the optimal hyper-parameter combination of the micro-seismic noise suppression model;
and 8: utilizing test set data obtained via step 2
Figure FDA0003557213680000027
Testing the optimal hyper-parameter combination of the microseism noise pressing model obtained in the step 7, and obtaining a final microseism data noise pressing model;
and step 9: and (4) inputting the field micro-seismic data into the final micro-seismic noise suppression model obtained in the step (8) to obtain the de-noised micro-seismic data, so as to suppress the noise of the micro-seismic data.
2. The method of claim 1, wherein the method comprises: the training process of the generative model in the step 6 is performed according to the following path:
2.1 the generative model is composed of a time attention module and a generation module, wherein the time attention module is defined as:
Figure FDA0003557213680000028
where the input to the temporal attention module is v obtained via step 5u,j(ii) a The output of the temporal attention module is the weight αu；σtRepresents the tanh activation function;
Figure FDA0003557213680000029
a jth channel representing a convolution kernel having a size of Z' x 1;
Figure FDA00035572136800000210
is the bias term.
2.2 v to be obtained via step 5u,jAnd alpha obtained via equation (1)uInputting the data into formula (2) to obtain an enhanced feature map vattSpecifically, the following are defined:
Figure FDA00035572136800000211
wherein v isattFor enhanced feature mapping of merged feature vectors, vuFor the u-th merged feature vector, αuThe temporal attention weight of the u-th merged feature vector.
2.3 v to be obtained via equation (2)attInputting the data into a generation module to obtain a generated noise-free micro-seismic data set
Figure FDA0003557213680000031
Wherein, G (z, c)h∈RS×1Generating noiseless microseismic data for the h-th time; the generating module is composed of a Dense layer, a Batch Normalization layer, a ReLU layer, a Deconv layer and a Tanh layer.
3. The method for suppressing the noise of the microseismic data based on the time attention mechanism mutual information generation countermeasure network as claimed in claim 1 or 2, wherein: the discriminant model training process in the step 6 is performed according to the following path:
3.1 the discriminant model is composed of a time attention module and a discriminant module, wherein the time attention module is defined as:
Figure FDA0003557213680000032
wherein the input to the temporal attention module is obtained via step 2
Figure FDA0003557213680000033
The output of the temporal attention module is the weight βk；σsRepresents the softmax activation function;
Figure FDA0003557213680000034
a jth channel representing a convolution kernel having a convolution kernel size of F' × 1;
Figure FDA0003557213680000035
is the bias term.
3.2 obtaining via step 2
Figure FDA0003557213680000036
With β obtained via equation (3)kInput into formula (4) to obtain an enhanced feature map xattSpecifically, the following are defined:
Figure FDA0003557213680000037
wherein x isattIs composed of
Figure FDA0003557213680000038
Enhanced feature mapping of (2);
Figure FDA0003557213680000039
for the kth training set sample, βkA temporal attention weight for the kth training set sample;
3.3 x to be obtained via equation (4)attInput into a discrimination module to obtain
Figure FDA00035572136800000310
Belong toProbability of a true sample. The discrimination module is composed of a sense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Softmax layer.
4. The method of claim 3, wherein the method comprises: the training process of the separation model in the step 6 is carried out according to the following path:
4.1 the separation model is composed of a time attention module and a separation module, wherein the time attention module is defined as:
Figure FDA00035572136800000311
wherein the input to the temporal attention module is G (z, c) obtained via the generative modelh,j(ii) a The output of the temporal attention module is the weight γh；σsigRepresenting a sigmoid activation function;
Figure FDA00035572136800000312
representing the j channel of the convolution kernel, wherein the size of the convolution kernel is Y' × 1;
Figure FDA00035572136800000313
is a bias term.
4.2 to be obtained via feature 2.3
Figure FDA00035572136800000314
And gamma obtained via the formula (5)hInput into equation (6) to obtain an enhanced feature map qattSpecifically, the following are defined:
Figure FDA00035572136800000315
wherein q isattEnhanced feature mapping for G (z, c); g (z, c)hNoise-free micro for h-th generationSeismic data, gammahTime attention weight for the h-th generated noise-free microseismic data.
4.3 q to be obtained via equation (6)attInputting into a separation module to obtain
Figure FDA0003557213680000041
Structural semantic vector of
Figure FDA0003557213680000042
Wherein
Figure FDA0003557213680000043
For the w-th extracted structural semantic feature vector, and calculate G (z, c) and cqMutual information of (2); the separation module is composed of a sense layer, a Batch Normalization layer, a LeakyReLU layer, a Conv layer and a Sigmoid layer.
5. The method of claim 4, wherein the method comprises: the model verification process in step 7 is performed according to the following path:
verification set data to be obtained by step 2
Figure FDA0003557213680000044
Inputting the data into the noise suppression pre-training model obtained in the step 6, verifying the performance of the trained noise suppression pre-training model by using the signal-to-noise ratio, and obtaining the optimal hyper-parameter combination of the micro-seismic noise suppression model.
6. The method of claim 5 for noise suppressing microseismic data based on time attention mechanism mutual information generation countermeasure network, wherein the method comprises the following steps: the model testing process in step 8 is performed according to the following path:
test set data to be obtained by step 2
Figure FDA0003557213680000045
And 7, testing the optimal hyper-parameter combination of the micro-seismic noise suppression model obtained in the step 7, and obtaining a final micro-seismic data noise suppression model.
7. The method of claim 6, wherein the method comprises: the noise suppression process in step 9 is performed according to the following path:
and (4) inputting the field micro-seismic data into the final micro-seismic noise suppression model obtained in the step (8) to obtain the de-noised micro-seismic data, so as to suppress the noise of the micro-seismic data.
8. A microseism data noise suppression system for generating a countermeasure network based on time attention mechanism mutual information is characterized in that: the system has program modules corresponding to the steps of any one of claims 1 to 7, and is operable to perform the steps of the method for noise suppressing microseismic data based on time attention mechanism mutual information generation countermeasure network.
9. A computer-readable storage medium characterized by: the computer readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the method of noise-suppressing microseismic data based on temporal attention mechanism mutual information generating a counterpoise network of any of claims 1-7.
10. A seismic information processing apparatus characterized by: the seismic information processing apparatus comprises at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executable by the at least one processor to enable the at least one processor to perform a method of noise suppressing a microseismic data based on time attention mechanism mutual information generation countermeasure network as claimed in any one of claims 1-7.
CN202210278729.XA 2022-03-21 2022-03-21 Microseism data noise suppression method and system, storage medium and seismic information processing equipment Pending CN114706128A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210278729.XA CN114706128A (en) 2022-03-21 2022-03-21 Microseism data noise suppression method and system, storage medium and seismic information processing equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210278729.XA CN114706128A (en) 2022-03-21 2022-03-21 Microseism data noise suppression method and system, storage medium and seismic information processing equipment

Publications (1)

Publication Number Publication Date
CN114706128A true CN114706128A (en) 2022-07-05

Family

ID=82169003

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210278729.XA Pending CN114706128A (en) 2022-03-21 2022-03-21 Microseism data noise suppression method and system, storage medium and seismic information processing equipment

Country Status (1)

Country Link
CN (1) CN114706128A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116660982A (en) * 2023-08-02 2023-08-29 东北石油大学三亚海洋油气研究院 Full waveform inversion method based on attention convolution neural network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116660982A (en) * 2023-08-02 2023-08-29 东北石油大学三亚海洋油气研究院 Full waveform inversion method based on attention convolution neural network
CN116660982B (en) * 2023-08-02 2023-09-29 东北石油大学三亚海洋油气研究院 Full waveform inversion method based on attention convolution neural network

Similar Documents

Publication Publication Date Title
CN111723329B (en) Seismic phase feature recognition waveform inversion method based on full convolution neural network
Wang et al. Blind source extraction of acoustic emission signals for rail cracks based on ensemble empirical mode decomposition and constrained independent component analysis
CN109613610B (en) Automatic picking method for microseism signal arrival time difference
CN111950503B (en) Aviation transient electromagnetic data processing method and device and computing equipment
CN108710777B (en) Diversified anomaly detection identification method based on multi-convolution self-coding neural network
CN104599292A (en) Noise-resistant moving target detection algorithm based on low rank matrix
CN105116442A (en) Lithologic oil-gas reservoir weak-reflection seismic signal reconstruction method
CN110632662A (en) Algorithm for automatically identifying microseism signals by using DCNN-inclusion network
CN113687433B (en) Bi-LSTM-based magnetotelluric signal denoising method and system
Yang et al. Denoising of distributed acoustic sensing data using supervised deep learning
CN111948711B (en) Method and system for extracting low-frequency part of seismic data by deep learning method
CN113962244A (en) Rayleigh wave seismic data noise removal method, storage medium and electronic device
CN111507155A (en) U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device
CN104143115A (en) Technological method for achieving soil water content classified identification through geological radar technology
CN114706128A (en) Microseism data noise suppression method and system, storage medium and seismic information processing equipment
RU2412454C2 (en) Method to process seismic data using discrete wavelet transform
CN112464721A (en) Automatic microseism event identification method and device
CN113822201A (en) Deep learning method for underwater object shape recognition based on flow field velocity component time course
Li et al. Magnetotelluric data denoising method combining two deep-learning-based models
CN115690772A (en) Gravity fault automatic identification method based on deep learning
CN111380680A (en) Check valve fault feature extraction method based on improved permutation entropy
CN116049637A (en) Gas-liquid two-phase flow pattern recognition method based on flow noise signals
CN115828069A (en) End-to-end magnetic anomaly signal noise reduction method based on deep learning
CN115563480A (en) Gear fault identification method for screening octave geometric modal decomposition based on kurtosis ratio coefficient
CN106842322B (en) Carbon dioxide flooding monitoring earthquake time difference correction method

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