CN111507155A - U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device - Google Patents

U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device Download PDF

Info

Publication number
CN111507155A
CN111507155A CN202010051627.5A CN202010051627A CN111507155A CN 111507155 A CN111507155 A CN 111507155A CN 202010051627 A CN202010051627 A CN 202010051627A CN 111507155 A CN111507155 A CN 111507155A
Authority
CN
China
Prior art keywords
data
supervised
loss
net
training
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
CN202010051627.5A
Other languages
Chinese (zh)
Inventor
盛冠群
郭小龙
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.)
Yangtze University
Original Assignee
Yangtze 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 Yangtze University filed Critical Yangtze University
Priority to CN202010051627.5A priority Critical patent/CN111507155A/en
Publication of CN111507155A publication Critical patent/CN111507155A/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
    • G06F2218/12—Classification; Matching
    • 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. for interpretation or for event detection
    • G01V1/288—Event detection in seismic signals, e.g. microseismics
    • 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
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00—Details of seismic processing or analysis
    • G01V2210/30—Noise handling
    • G01V2210/32—Noise reduction
    • G01V2210/324—Filtering
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00—Details of seismic processing or analysis
    • G01V2210/40—Transforming data representation
    • G01V2210/48—Other transforms
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00—Details of seismic processing or analysis
    • G01V2210/60—Analysis
    • G01V2210/67—Wave propagation modeling

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Remote Sensing (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Environmental & Geological Engineering (AREA)
  • Business, Economics & Management (AREA)
  • Emergency Management (AREA)
  • Signal Processing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Acoustics & Sound (AREA)
  • Geology (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • Geophysics (AREA)
  • Geophysics And Detection Of Objects (AREA)

Abstract

The invention discloses a U-Net + + and UDA combined microseism effective signal first arrival pickup method and a device, wherein the method comprises the following steps: acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set; making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets; putting the labeled data set and the unlabeled data set into a U-Net + + network together to perform Gaussian noise-based supervised training and data enhancement-based unsupervised training respectively, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model; and carrying out microseism effective signal first arrival pickup through the trained network model. The invention can reduce the labor cost and improve the accuracy of integrated prediction.

Description

U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device
Technical Field
The invention belongs to the field of microseism signal detection, and particularly relates to a method and a device for realizing first arrival pickup of microseism effective signals by combining a UDA (Universal data acquisition) network and a U-Net + +.
Background
Micro-earthquakes are small earthquakes that are often unavoidable during deep mining in underground mines. The microseism monitoring technology is used as an important way for shale gas exploitation, plays an increasingly important role in the exploration process, and has important significance for automatic processing of massive microseism data if automatic identification and automatic first arrival pickup of microseism effective signals can be realized.
There has been a great deal of research into the first arrival pickup of microseismic effect signals by Convolutional Neural Networks (CNNs). A Nested-U-Net (UNet + +) network adds a layer jump structure on the basis of the U-Net network, so that different dimensional characteristics can be effectively extracted, but the number of layers of the UNet + + network is shallow, so that the identification precision is still to be improved. The traditional microseism effective signal detection needs a large amount of manually marked samples, and is long in time consumption and low in accuracy. Therefore, how to accurately detect the continuously input signals under the semi-supervised condition is very important, and the invention focuses on researching how to quickly and accurately realize the first arrival picking of the micro-seismic effective signals of the small samples under the strong noise background.
Disclosure of Invention
On the basis of U-Net + +, various and real noise data are generated by combining Unsupervised Data Augmentation (UDA) and the model keeps consistency on the noise, so that high-precision first-arrival picking of micro-seismic signals of small samples can be performed, loss can be optimized, and the problems that calculation of a large amount of sample data is required and picking precision is low in traditional micro-seismic signal detection are solved.
In a first aspect of the invention, a U-Net + + and UDA combined microseism effective signal first arrival pickup method is provided, and the method comprises the following steps:
s1, acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set;
s2, making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets;
s3, putting the labeled data set and the unlabeled data set into a U-Net + + network to respectively perform supervised training based on Gaussian noise and unsupervised training based on data enhancement, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model;
and S4, carrying out microseism effective signal first arrival pickup through the trained network model.
Preferably, the step S1 specifically includes:
drawing a profile of original microseism signal data by using matlab, marking sampling points of signals, extracting effective signal channels, removing ineffective channels, and realizing separation of signals and noise by low-pass or high-pass filtering;
performing time-frequency analysis on the micro-seismic signals by using S transformation to obtain a micro-seismic signal data set; the formula of the S transformation is as follows:
Figure BDA0002371373270000021
wherein
Figure BDA0002371373270000022
Is a gaussian window function, t represents time and f represents frequency.
Preferably, in step S3, the process of the supervised training based on gaussian noise is:
dividing a labeled data set into two batches of label signals, wherein one batch is kept unchanged, and the other batch is interfered by adding Gaussian noise; putting the two batches of labeled signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
for data that remains unchanged, a conditional probability distribution p is calculatedθ(y | x) calculating conditional probability distribution for data interfered by Gaussian noise
Figure BDA0002371373270000024
The cross entropy losses of two batches of features at different levels are calculated in an integrated mode, a dice coefficient is introduced into the cross entropy losses and serves as a loss value of a supervised part, and a loss formula of the supervised part is as follows:
Figure BDA0002371373270000023
where x denotes input, y and y*Respectively, true and predicted values, b refers to the b-th signal, N is the batch size, theta refers to the model parameter, L refers to the labeled sample,
Figure BDA00023713732700000310
refers to noise.
Preferably, in step S3, the process of the unsupervised training based on data enhancement is:
dividing the non-tag data set into two batches of non-tag signals, performing data enhancement on one batch, and keeping the other batch unchanged; putting the two batches of non-label signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
the signal obtained after up-sampling is processed as follows: for data that remains unchanged, a conditional probability distribution of y given an input x is calculated
Figure BDA0002371373270000031
For enhanced data
Figure BDA0002371373270000032
Calculating a given input
Figure BDA0002371373270000033
Conditional probability distribution of time y
Figure BDA0002371373270000034
Integrating and calculating consistency losses of two batches of different layers, taking the consistency losses as an unsupervised loss value, wherein the calculation formula is as follows:
Figure BDA0002371373270000035
wherein
Figure BDA0002371373270000036
U refers to the non-tagged data,
Figure BDA0002371373270000037
is a data enhancement conversion of the data to be enhanced,
Figure BDA0002371373270000038
is a fixed copy of the current parameter theta.
Preferably, in step S3, the final loss value is formed by balancing the supervised loss and the unsupervised loss using a weighting factor, and the expression is:
final _ loss is cross _ entry _ loss + λ consistency _ loss λ as a weight factor;
for final output final _ loss values of the network, if the final _ loss values are smaller than or equal to a preset threshold value, the model convergence is good, and training data of the model are reserved; and if the final _ loss value is larger than the preset threshold value, advancing the updated parameters of the feature results obtained from the last layer by adopting a BP neural network algorithm until the model converges.
Preferably, in step S3, in the semi-supervised model training, a random gradient descent is used as an optimizer training sample, θ is derived by a loss function of each sample to obtain a corresponding gradient, so as to update θ, and a training formula of the semi-supervised model training is as follows:
Figure BDA0002371373270000039
Figure BDA0002371373270000041
wherein theta is an iteration parameter, y is a function output quantity, h (theta) is a solved fitting function, j is a parameter number, and theta' is an updated parameter along the negative direction of theta gradient;
in a second aspect of the present invention, a U-Net + + and UDA combined microseismic effective signal first arrival pickup apparatus is provided, the apparatus comprising:
a dataset acquisition module: acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set;
a dataset tagging module: making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets;
a semi-supervised training module: putting the labeled data set and the non-labeled data set into a U-Net + + network model together to perform supervised training based on Gaussian noise and unsupervised training based on data enhancement respectively, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model;
first arrival picking up module: and carrying out microseism effective signal first arrival pickup through the trained network model.
Preferably, the semi-supervised model training module specifically includes:
the supervised training unit:
the device is used for dividing a labeled data set into two batches of label signals, wherein one batch of label signals is kept unchanged, and the other batch of label signals is interfered by adding Gaussian noise; putting the two batches of labeled signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
for data that remains unchanged, a conditional probability distribution p is calculatedθ(y | x) calculating conditional probability distribution for data interfered by Gaussian noise
Figure BDA0002371373270000044
The cross entropy losses of two batches of features at different levels are calculated in an integrated mode, a dice coefficient is introduced into the cross entropy losses and serves as a loss value of a supervised part, and a loss formula of the supervised part is as follows:
Figure BDA0002371373270000042
where x denotes input, y and y*Respectively, true and predicted values, b refers to the b-th signal, N is the batch size, theta refers to the model parameter, L refers to the labeled sample,
Figure BDA0002371373270000043
refers to noise;
an unsupervised training unit:
the system is used for dividing a non-tag data set into two batches of non-tag signals, one batch is subjected to data enhancement, and the other batch is kept unchanged; putting the two batches of non-label signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
the signal obtained after up-sampling is processed as follows: for data that remains unchanged, a conditional probability distribution of y given an input x is calculated
Figure BDA0002371373270000051
For enhanced data
Figure BDA0002371373270000052
Calculating a given input
Figure BDA0002371373270000053
Conditional probability distribution of time y
Figure BDA0002371373270000054
Integrating and calculating consistency losses of two batches of different layers, taking the consistency losses as an unsupervised loss value, wherein the calculation formula is as follows:
Figure BDA0002371373270000055
wherein
Figure BDA0002371373270000056
U means noneThe data of the tag is transmitted to the mobile terminal,
Figure BDA0002371373270000057
is a data enhancement conversion of the data to be enhanced,
Figure BDA0002371373270000058
is a fixed copy of the current parameter θ;
final loss value calculation unit:
calculating a final loss value for the supervised and unsupervised losses balanced using the weighting factors, expressed as:
final _ loss is cross _ entry _ loss + λ consistency _ loss λ as a weighting factor.
Preferably, in the semi-supervised model training module, a random gradient descent is used as an optimizer training sample, a loss function of each sample is used to derive θ to obtain a corresponding gradient, so as to update a parameter θ, and a training formula is as follows through a parameter adjusting optimization network:
Figure BDA0002371373270000059
Figure BDA00023713732700000510
where θ is the iteration parameter, y is the function output, h (θ) is the found fitting function, j is the number of parameters, and θ' is the resulting parameter updated in the negative direction of the θ gradient.
The invention combines the UDA and the U-Net + + network to realize the first arrival pickup of the microseism effective signal and has the following two advantages:
firstly, theoretical advantages are as follows: the invention combines the UDA and the U-Net + + network to generate a large amount of real data to be transmitted to the network for training through the unique capacity expansion mode of the UDA, and simultaneously, the accuracy of integrated prediction can be improved, thereby effectively solving the problem.
Secondly, the practical advantages are as follows: existing microseism detection technology needs to rely on a large amount of manual marking sample training, and this process is consuming time and power, uses this patent to handle the microseism data set and can practice thrift a large amount of human costs to a certain extent, and the more efficient picks up microseism first arrival signal point simultaneously.
Drawings
In order to more clearly illustrate the technical solution of the present invention, the drawings needed to be used in the technical description of the present invention will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
FIG. 1 is a schematic flow chart of a U-Net + + and UDA combined microseism effective signal first-break pickup method provided by the invention;
FIG. 2 is a schematic diagram of a U-Net + + network model used in the present invention;
FIG. 3 is a diagram illustrating a residual block structure;
FIG. 4 is a diagram of signals under test;
FIG. 5 is the first arrival pickup of the microseismic signal of FIG. 4.
Detailed Description
In order to make the objects, features and advantages of the present invention more obvious and understandable, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the embodiments described below are only a part of the embodiments of the present invention, but not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, the present invention provides a U-Net + + and UDA combined microseismic effective signal first arrival pickup method, including:
s1, acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set;
drawing a profile of original microseism signal data by using matlab, marking sampling points of signals, extracting effective signal channels, removing ineffective channels, and realizing separation of signals and noise by low-pass or high-pass filtering;
performing time-frequency analysis on the micro-seismic signals by using S transformation to obtain a micro-seismic signal data set; the formula of the S transformation is as follows:
Figure BDA0002371373270000071
wherein
Figure BDA0002371373270000072
Is a gaussian window function, t represents time and f represents frequency.
S2, making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets;
specifically, effective signal channels are found in a signal profile generated by matlab, the first arrival positions of the effective signals are marked manually, and a labeled data set is manufactured.
S3, putting the labeled data set and the unlabeled data set into a U-Net + + network to respectively perform supervised training based on Gaussian noise and unsupervised training based on data enhancement, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model;
specifically, a labeled data set and an unlabeled data set are converted into two-dimensional signals in a zero-adding and completion mode, the two-dimensional signals are put into a U-Net + + network together for integrated prediction, the whole network training process is divided into two stages of supervised training and unsupervised training, and the final _ loss of the two stages is taken as the overall loss value as the standard for measuring the quality of the model to carry out semi-supervised model training on the network.
Referring to fig. 2, fig. 2 is a schematic diagram of a U-Net + + network model structure used in the present invention, compared with a U-Net network, the added skip layer connection of the U-Net + + network can integrate deep and shallow features of different layers, and the newly added deep supervision structure reduces numerous and complex parameters of the network, and can improve the efficiency and the extraction accuracy of the network.
The U-Net + + network model starts to execute layer jump connection when i is 0, and the operation formula is as follows
Figure RE-GDA0002556815810000081
Where H (-) represents a convolution and an activation function, u (-) represents an upsampled layer, and [ - ] represents the connection layer;
introducing a residual block between the down-sampling and the up-sampling of the U-Net + + network, said residual block having a structure as shown in fig. 3, clipping the signal after each convolutional layer using a relu function, wherein the relu function is:
f(x)＝max(0,x) (3)
where x is the model output value.
Further, the process of the supervised training based on gaussian noise is as follows:
dividing a labeled data set into two batches of label signals, wherein one batch is kept unchanged, and the other batch is interfered by adding Gaussian noise; putting the two batches of labeled signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output; two batches of label signals are subjected to down-sampling coding of a U-Net + + network, robustness of small disturbances of input signals can be improved, overfitting is reduced, the operation amount of the label signals is reduced, the size of a receptive field is increased, parameters lost in down-sampling are recovered through up-sampling decoding operation, feature extraction is carried out, and classification probability is output through a softmax function.
For data that remains unchanged, a conditional probability distribution p is calculatedθ(y | x) calculating conditional probability distribution for data interfered by Gaussian noise
Figure BDA0002371373270000083
The cross entropy losses of two batches of features at different levels are calculated in an integrated mode, a dice coefficient is introduced into the cross entropy losses, the cross entropy losses introduced into the dice coefficient are used as loss values of a supervised part, and a loss formula of the supervised part is as follows:
Figure BDA0002371373270000082
where x denotes input, y and y*Respectively, true and predicted values, b refers to the b-th signal, N is the batch size, theta refers to the model parameter, L refers to the labeled sample,
Figure BDA0002371373270000084
refers to noise.
Further, the process of the unsupervised training based on data enhancement is as follows:
dividing the non-tag data set into two batches of non-tag signals, performing data enhancement on one batch, and keeping the other batch unchanged; putting the two batches of non-label signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
the signal obtained after up-sampling is processed as follows: for data that remains unchanged, a conditional probability distribution of y given an input x is calculated
Figure BDA0002371373270000091
For enhanced data
Figure BDA0002371373270000092
Calculating a given input
Figure BDA0002371373270000093
Conditional probability distribution of time y
Figure BDA0002371373270000094
Integrating and calculating consistency losses of two batches of different layers, taking the consistency losses as unsupervised loss values, and adopting a calculation formula as follows:
Figure BDA0002371373270000095
wherein
Figure BDA0002371373270000096
U refers to the non-tagged data,
Figure BDA0002371373270000097
is a data enhancement conversion of the data to be enhanced,
Figure BDA0002371373270000098
is a fixed copy of the current parameter theta.
Further, the final loss value is obtained by balancing the supervised loss and the unsupervised loss by using a weight factor, and the expression is as follows:
final_loss＝cross_entropy_loss+λconsistency_loss (6)
λ is a weighting factor;
for final output final _ loss values of the network, if the final _ loss values are smaller than or equal to a preset threshold value, the model convergence is good, and training data of the model are reserved; and if the final _ loss value is larger than the preset threshold value, advancing the updated parameters of the feature results obtained from the last layer by adopting a BP neural network algorithm until the model converges.
By minimizing divergence measures of both distributions
Figure BDA00023713732700000910
Not only can the noise immunity of the model be improved, but also the label information can be transferred from the label data to the label-free data, thereby smoothing the model.
The semi-supervised model training is formed by the supervised training and the unsupervised training, random Gradient Descent (SGD) is used as an optimizer training sample in the semi-supervised model training, the loss function of each sample is used for deriving theta to obtain a corresponding Gradient so as to update theta, and a training formula is as follows through a parameter adjusting optimization network:
Figure BDA0002371373270000099
Figure BDA0002371373270000101
wherein theta is an iteration parameter, y is a function output quantity, h (theta) is a solved fitting function, j is a parameter number, and theta' is an updated parameter along the negative direction of theta gradient;
and S4, carrying out microseism effective signal first arrival pickup through the trained network model.
Referring to fig. 4, fig. 4 is a signal to be measured, the signal to be measured of fig. 4 is input into the model trained in step S3, and the first arrival picking result of the micro seismic signal is shown in fig. 5.
Corresponding to the embodiment of the method, the invention provides a U-Net + + and UDA combined microseismic effective signal first arrival pickup device, which comprises:
a dataset acquisition module: acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set;
a dataset tagging module: making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets;
a semi-supervised training module: putting the labeled data set and the non-labeled data set into a U-Net + + network model together to perform supervised training based on Gaussian noise and unsupervised training based on data enhancement respectively, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model;
first arrival picking up module: and carrying out microseism effective signal first arrival pickup through the trained network model.
Further, the semi-supervised model training module specifically comprises:
the supervised training unit:
the device is used for dividing a labeled data set into two batches of label signals, wherein one batch of label signals is kept unchanged, and the other batch of label signals is interfered by adding Gaussian noise; putting the two batches of labeled signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
for data that remains unchanged, a conditional probability distribution p is calculatedθ(y | x) calculating conditional probability distribution for data interfered by Gaussian noise
Figure BDA0002371373270000102
The cross entropy losses of two batches of features at different levels are calculated in an integrated mode, a dice coefficient is introduced into the cross entropy losses and serves as a loss value of a supervised part, and a loss formula of the supervised part is as follows:
Figure BDA0002371373270000111
where x denotes input, y and y*Respectively, true and predicted values, b refers to the b-th signal, N is the batch size, theta refers to the model parameter, L refers to the labeled sample,
Figure BDA00023713732700001112
refers to noise;
an unsupervised training unit:
the system is used for dividing a non-tag data set into two batches of non-tag signals, one batch is subjected to data enhancement, and the other batch is kept unchanged; putting the two batches of non-label signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
the signal obtained after up-sampling is processed as follows: for data that remains unchanged, a conditional probability distribution of y given an input x is calculated
Figure BDA0002371373270000112
For enhanced data
Figure BDA0002371373270000113
Calculating a given input
Figure BDA0002371373270000114
Conditional probability distribution of time y
Figure BDA0002371373270000115
Integrating and calculating consistency losses of two batches of different layers, taking the consistency losses as an unsupervised loss value, wherein the calculation formula is as follows:
Figure BDA0002371373270000116
wherein
Figure BDA0002371373270000117
U refers to the non-tagged data,
Figure BDA0002371373270000118
is a data enhancement conversion of the data to be enhanced,
Figure BDA0002371373270000119
is a fixed copy of the current parameter θ;
final loss value calculation unit:
for calculating the final loss value, the expression is:
final _ loss is cross _ entry _ loss + λ consistency _ loss λ as a weighting factor.
Further, in the semi-supervised model training module, random gradient descent is used as an optimizer training sample, a loss function of each sample is used for deriving theta to obtain a corresponding gradient so as to update a parameter theta, and a training formula is as follows by tuning a parameter optimization network:
Figure BDA00023713732700001110
Figure BDA00023713732700001111
where θ is the iteration parameter, y is the function output, h (θ) is the found fitting function, j is the number of parameters, and θ' is the resulting parameter updated in the negative direction of the θ gradient.
The above apparatus embodiments and method embodiments are in one-to-one correspondence, and reference may be made to the method embodiments for a brief point of the apparatus embodiments.
The embodiments in the present specification are described in a progressive manner, each embodiment focuses on differences from other embodiments, similar parts between the embodiments are referred to, and parts not described in the specification are all the prior art or common general knowledge.
Those of skill would further appreciate that the various illustrative components and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in random access memory, read only memory, electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
Although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and the modifications or the substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (9)

1. A U-Net + + and UDA combined microseismic effect signal first arrival pickup method, the method comprising:
s1, acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set;
s2, making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets;
s3, putting the labeled data set and the unlabeled data set into a U-Net + + network to respectively perform Gaussian noise-based supervised training and data enhancement-based unsupervised training, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model;
and S4, carrying out microseism effective signal first arrival pickup through the trained network model.
2. The U-Net + + and UDA combined microseismic effect signal first arrival pickup method according to claim 1, wherein the step S1 specifically comprises:
drawing a profile of original microseism signal data by using matlab, marking sampling points of signals, extracting effective signal channels, removing ineffective channels, and realizing separation of signals and noise by low-pass or high-pass filtering;
performing time-frequency analysis on the micro-seismic signals by using S transformation to obtain a micro-seismic signal data set; the formula of the S transformation is as follows:
Figure FDA0002371373260000011
wherein
Figure FDA0002371373260000012
Is a gaussian window function, t represents time and f represents frequency.
3. The U-Net + + and UDA combined microseismic effect signal first arrival picking method of claim 1 wherein the supervised training based on gaussian noise in step S3 comprises:
dividing the tagged data set into two batches of tag signals, wherein one batch is kept unchanged, and the other batch is interfered by adding Gaussian noise; putting the two batches of labeled signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
for data that remains unchanged, a conditional probability distribution p is calculatedθ(y | x) for the data interfered by the Gaussian noise, calculating to obtain a conditional probability distribution pθ(y*|x,∈)；
The cross entropy losses of two batches of features at different levels are calculated in an integrated mode, a dice coefficient is introduced into the cross entropy losses and serves as a loss value of a supervised part, and a loss formula of the supervised part is as follows:
Figure FDA0002371373260000021
where x denotes input, y and y*Respectively, true and predicted values, b refers to the b-th signal, N is the batch size, θ refers to the model parameter, L refers to the labeled sample, ∈ refers to noise.
4. The U-Net + + and UDA joint microseismic valid signal first arrival picking method according to claim 3 wherein in step S3, the process of unsupervised training based on data enhancement is as follows:
dividing the non-tag data set into two batches of non-tag signals, performing data enhancement on one batch, and keeping the other batch unchanged; putting the two batches of non-label signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
the signal obtained after up-sampling is processed as follows: for data that remains unchanged, a conditional probability distribution of y given an input x is calculated
Figure FDA0002371373260000022
For enhanced data
Figure FDA0002371373260000023
Calculating a given input
Figure FDA0002371373260000024
Conditional probability distribution of time y
Figure FDA0002371373260000025
Integrating and calculating consistency loss of two batches of different layers, taking the consistency loss as a loss value of an unsupervised part, and adopting a calculation formula as follows:
Figure FDA0002371373260000026
wherein
Figure FDA0002371373260000027
U refers to the non-tagged data,
Figure FDA0002371373260000028
is a data enhancement conversion of the data to be enhanced,
Figure FDA0002371373260000029
is a fixed copy of the current parameter theta.
5. The U-Net + + and UDA joint microseismic effect signal first arrival picking method of claim 5 wherein in step S3, the final loss value is obtained by balancing supervised and unsupervised losses using weighting factors, and is expressed as:
final_loss＝cross_entropy_loss+λconsistency_loss
λ is a weighting factor;
for final output final _ loss values of the network, if the final _ loss values are smaller than or equal to a preset threshold value, the model convergence is good, and training data of the model are reserved; and if the final _ loss value is larger than the preset threshold value, advancing the updated parameters of the feature results obtained from the last layer by adopting a BP neural network algorithm until the model converges.
6. The U-Net + + and UDA joint microseism significant signal first arrival picking method according to claim 1, wherein in step S3, random gradient descent is used as an optimizer training sample in the semi-supervised model training, a loss function of each sample is used to derive θ to obtain a corresponding gradient, and then a parameter θ is updated, and the training formula is as follows:
Figure FDA0002371373260000031
Figure FDA0002371373260000032
wherein theta is an iteration parameter, y is a function output quantity, h (theta) is a solved fitting function, j is the number of parameters, and theta' is an updated parameter along the negative direction of theta gradient;
7. a U-Net + + and UDA combined microseismic effect signal first arrival pickup apparatus, comprising:
a dataset acquisition module: acquiring original micro-seismic signal data and preprocessing the original micro-seismic signal data to obtain a micro-seismic signal data set;
a dataset tagging module: making a labeled data set according to part of data of the micro seismic signal data set, and making the rest labeled data sets;
a semi-supervised training module: putting the labeled data set and the non-labeled data set into a U-Net + + network model together to perform supervised training based on Gaussian noise and unsupervised training based on data enhancement respectively, integrating the predicted loss value of the supervised part and the predicted loss value of the unsupervised part into a final loss value, and performing semi-supervised model training by taking the final loss value as a standard for measuring the quality of the whole network model;
first arrival picking up module: and carrying out microseism effective signal first arrival pickup through the trained network model.
8. The U-Net + + and UDA combined microseismic active signal first arrival pickup apparatus as claimed in claim 7, wherein the semi-supervised model training module specifically comprises:
the supervised training unit:
the device is used for dividing a labeled data set into two batches of label signals, wherein one batch is kept unchanged, and the other batch is interfered by adding Gaussian noise; putting the two batches of labeled signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
for data that remains unchanged, a conditional probability distribution p is calculatedθ(y | x) for the data interfered by the Gaussian noise, calculating to obtain a conditional probability distribution pθ(y*|x,∈)；
The cross entropy losses of two batches of features at different levels are calculated in an integrated mode, a dice coefficient is introduced into the cross entropy losses and serves as a loss value of a supervised part, and a loss formula of the supervised part is as follows:
Figure FDA0002371373260000041
where x denotes input, y and y*Respectively indicating a true value and a predicted value, b indicating a b-th signal, N being the size of a batch, theta indicating a model parameter, L indicating a sample with a label, and ∈ indicating noise;
an unsupervised training unit:
the system is used for dividing a non-tag data set into two batches of non-tag signals, one batch is subjected to data enhancement, and the other batch is kept unchanged; putting the two batches of non-label signals into a U-Net + + network for convolution and pooling operation to obtain end-to-end output;
the signal obtained after up-sampling is processed as follows: for data that remains unchanged, a conditional probability distribution of y given an input x is calculated
Figure FDA0002371373260000042
For enhanced numbersAccording to
Figure FDA0002371373260000043
Calculating a given input
Figure FDA0002371373260000044
Conditional probability distribution of time y
Figure FDA0002371373260000045
Integrating and calculating consistency loss of two batches of different layers, taking the consistency loss as a loss value of an unsupervised part, and adopting a calculation formula as follows:
Figure FDA0002371373260000046
wherein
Figure FDA0002371373260000047
U refers to the non-tagged data,
Figure FDA0002371373260000048
is a data enhancement conversion of the data to be enhanced,
Figure FDA0002371373260000049
is a fixed copy of the current parameter θ;
final loss value calculation unit:
for calculating a final loss value using the weighting factors to balance the supervised and unsupervised losses, the expression is:
final_loss＝cross_entropy_loss+λconsistency_loss
λ is a weighting factor.
9. The U-Net + + and UDA combined microseismic valid signal first arrival pickup apparatus according to claim 8, wherein the semi-supervised model training module uses stochastic gradient descent as optimizer training samples, and updates θ by deriving θ according to the loss function of each sample, and the training formula is as follows by tuning the parameter optimization network:
Figure FDA0002371373260000051
Figure FDA0002371373260000052
where θ is the iteration parameter, y is the output of the function, h (θ) is the fitting function found, j is the number of parameters, and θ' is the resulting parameter updated in the negative direction of the θ gradient.
CN202010051627.5A 2020-01-17 2020-01-17 U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device Pending CN111507155A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010051627.5A CN111507155A (en) 2020-01-17 2020-01-17 U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010051627.5A CN111507155A (en) 2020-01-17 2020-01-17 U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device

Publications (1)

Publication Number Publication Date
CN111507155A true CN111507155A (en) 2020-08-07

Family

ID=71863877

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010051627.5A Pending CN111507155A (en) 2020-01-17 2020-01-17 U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device

Country Status (1)

Country Link
CN (1) CN111507155A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112229952A (en) * 2020-09-30 2021-01-15 重庆科技学院 Method for detecting toxic and harmful gas in chemical industrial park
CN112926631A (en) * 2021-02-01 2021-06-08 大箴(杭州)科技有限公司 Financial text classification method and device and computer equipment
CN113255807A (en) * 2021-06-03 2021-08-13 北京的卢深视科技有限公司 Face analysis model training method, electronic device and storage medium
CN113627073A (en) * 2021-07-01 2021-11-09 武汉大学 Underwater vehicle flow field result prediction method based on improved Unet + + network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106405640A (en) * 2016-08-26 2017-02-15 中国矿业大学(北京) Automatic microseismic signal arrival time picking method based on depth belief neural network
CN110619264A (en) * 2019-07-30 2019-12-27 长江大学 UNet + + based microseism effective signal identification method and device

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106405640A (en) * 2016-08-26 2017-02-15 中国矿业大学(北京) Automatic microseismic signal arrival time picking method based on depth belief neural network
CN110619264A (en) * 2019-07-30 2019-12-27 长江大学 UNet + + based microseism effective signal identification method and device

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
QIZHE XIE等: "UNSUPERVISED DATA AUGMENTATION FOR CONSISTENCY TRAINING" *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112229952A (en) * 2020-09-30 2021-01-15 重庆科技学院 Method for detecting toxic and harmful gas in chemical industrial park
CN112229952B (en) * 2020-09-30 2022-11-15 重庆科技学院 Method for monitoring toxic and harmful gases in chemical industrial park
CN112926631A (en) * 2021-02-01 2021-06-08 大箴(杭州)科技有限公司 Financial text classification method and device and computer equipment
CN113255807A (en) * 2021-06-03 2021-08-13 北京的卢深视科技有限公司 Face analysis model training method, electronic device and storage medium
CN113255807B (en) * 2021-06-03 2022-03-25 北京的卢深视科技有限公司 Face analysis model training method, electronic device and storage medium
CN113627073A (en) * 2021-07-01 2021-11-09 武汉大学 Underwater vehicle flow field result prediction method based on improved Unet + + network
CN113627073B (en) * 2021-07-01 2023-09-19 武汉大学 Underwater vehicle flow field result prediction method based on improved Unet++ network

Similar Documents

Publication Publication Date Title
CN111507155A (en) U-Net + + and UDA combined microseism effective signal first-arrival pickup method and device
CN109993100B (en) Method for realizing facial expression recognition based on deep feature clustering
CN111243674B (en) Base sequence identification method, device and storage medium
CN113642698B (en) Geophysical well logging intelligent interpretation method, system and storage medium
CN112761628B (en) Shale gas yield determination method and device based on long-term and short-term memory neural network
CN112233675B (en) Voice wake-up method and system based on separated convolutional neural network
CN111695413A (en) Signal first arrival pickup method and device combining U-Net and Temporal encoding
CN114239749B (en) Modulation identification method based on residual shrinkage and two-way long-short-term memory network
CN114492521A (en) Intelligent lithology while drilling identification method and system based on acoustic vibration signals
CN111882057A (en) Multi-stage deep learning method oriented to space-time sequence data modeling and application
CN112101418A (en) Method, system, medium and equipment for identifying breast tumor type
CN115795369B (en) Coal-rock interface identification method and device
CN116386720A (en) Single cell transcription factor prediction method based on deep learning and attention mechanism
CN114706128A (en) Microseism data noise suppression method and system, storage medium and seismic information processing equipment
CN113850185A (en) Multi-classification method, device, terminal and storage medium for underground acoustic emission source
CN116226623B (en) Mark layer division method and device based on SegNet segmentation model and computer equipment
CN113487519B (en) Image rain removing method based on artificial intelligence
CN117633658B (en) Rock reservoir lithology identification method and system
CN117647837B (en) First arrival pickup method, system and computer equipment for seismic data
CN116304996B (en) Method and device for extracting body trend of stone relic monitoring data
CN116469398A (en) Motor fault diagnosis method based on binary mode and PBA-ANN
CN117669795A (en) Deep learning-based tight gas reservoir yield prediction method and system
CN116204760A (en) Drilling strain data anomaly extraction method based on GRU network
CN117711505A (en) Enzyme kinetic parameter prediction method, device, equipment and medium
CN116959483A (en) Music feature extraction method, device, equipment and medium

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20200807

RJ01 Rejection of invention patent application after publication