CN111523258B - First arrival picking method and system of effective microseismic signals based on MS-Net network - Google Patents

Abstract

The invention relates to the technical field of microseismic data processing, in particular to a method and system for first-arrival picking of effective microseismic signals based on MS-Net network. The method includes generating an original data set; data set calibration; inputting the data set into the constructed MS-Net network for training to obtain the optimal network model parameters, specifically including performing effective calibration on the part of the samples that are calibrated Supervised training, unsupervised training is performed on the other part of samples that have not been calibrated; the probability distribution of the data set is calculated point by point; the system includes a data set production module, a data set training module and an output module; the embodiment of the present invention uses MS ‑Net network combined with semi-supervised method, weighted sum of unsupervised loss and supervised loss to construct total loss, optimize network model parameters by minimizing total loss, and finally achieve accurate prediction and identification of effective signal first arrival point; reduce training set Mark the number of labels to improve the quality of the training set and detection accuracy.

Description

First arrival picking method and system of effective microseismic signals based on MS-Net network

technical field

The invention relates to the technical field of microseismic data processing, in particular to a method and system for first-arrival picking of effective microseismic signals based on an MS-Net network.

Background technique

Microseismic detection methods play an important role in engineering construction and geological disaster prevention and control. At the same time, microseismic signals have the characteristics of weak signal energy and are easily disturbed by background noise. As a result, the first arrival of microseismic signals cannot be accurately picked up, resulting in microseismic events. Therefore, the effective microseismic signal detection method is one of the key points in the field of microseismic data processing.

Traditional signal detection techniques include spectral analysis of signals through fast Fourier transform, time-frequency conversion of wavelet, curvelet and shearlet transform to achieve the purpose of removing noise and retaining effective signals. However, if the traditional methods are directly applied to microseismic data, they often cannot obtain satisfactory results, which will directly affect the quality and accuracy of microseismic monitoring. Signal monitoring based on deep learning has gradually attracted people's attention in recent years. The main reason is that it has the characteristics of many parameters and large capacity, which makes its network have a strong processing ability for massive data; MS-Net's new network model It is composed of Denseblock (Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q.Weinberger.2017) blocks added to the UNet++ network to deepen the network structure. Through the layer-hopping and pruning structure in the UNet++ network, the main and subtle signals are extracted. While optimizing the features, it avoids the problems of feature stacking and overfitting. By adding the Denseblock block, it makes up for the problem that the deep feature recognition is not obvious due to the small number of UNet++ network layers, so that deep and shallow feature construction can be accurately obtained. The MS-Net network improves the refined extraction of signal features to a certain extent.

The disadvantage of the existing technology is that it needs to manually mark the label data set and input it into the network for intensive training and learning. The quality of the training set is not high, it takes a long time and the accuracy rate is low.

Contents of the invention

In order to overcome the shortcomings of the existing technology, the embodiment of the present invention provides a method and system for first-arrival picking of effective microseismic signals based on the MS-Net network, which reduces the number of labels in the training set and improves the quality of the training set and detection accuracy.

On the one hand, the embodiment of the present invention provides a method for first-arrival picking-up of effective micro-seismic signals based on MS-Net network, comprising the following steps:

S1, generating the original data set; specifically including the use of finite difference forward modeling to generate a large number of analog signals with a main frequency range of 20-1000 Hz under different models and actual data to form the original data set;

S2, data set calibration; specifically includes first-arrival picking of some samples in the original data set, selecting the signal waveforms at the first-arrival and non-first-arrival positions of each signal sampling point, and performing calibration respectively, while the other part of the samples are not calibrated;

S3, input the data set into the constructed MS-Net network for training to obtain the optimal network model parameters; specifically including supervised training for the part of the samples that have been calibrated, and the other part that has not been calibrated samples for unsupervised training;

S4. Calculate the probability distribution of the data set point by point; specifically, use the softmax function to output the probability point by point to obtain the binary classification probability of all points, and select the peak probability of the first arrival category as the first arrival point.

On the other hand, the embodiment of the present invention provides a first-arrival pickup system for effective micro-seismic signals based on the MS-Net network, including:

The data set production module generates the original data set; specifically includes the use of finite difference forward modeling to generate a large number of analog signals with a main frequency range of 20-1000 Hz and actual data to form the original data set under different models; data set calibration; specifically includes the original data Collect a part of the samples for first-arrival picking, select the signal waveforms at the first-arrival and non-first-arrival positions of each signal sampling point and calibrate them separately, and the other part of the samples will not be calibrated;

The data set training module is to input the data set into the constructed MS-Net network for training to obtain the optimal network model parameters; specifically, supervised training is carried out for the part of the samples that have been calibrated, and all samples that have not been calibrated The other part of the sample is used for unsupervised training;

The output module calculates the probability distribution of the data set point by point; specifically, it includes using the softmax function to output the probability point by point to obtain the binary classification probability of all points, and select the peak probability of the first arrival category as the first arrival point.

The embodiment of the present invention provides a first-arrival picking method and system for effective microseismic signals based on the MS-Net network. Through the MS-Net network combined with the semi-supervised method Temporal Ensembling, the unsupervised loss and the supervised loss are weighted and summed to construct the total loss , by minimizing the total loss and optimizing the parameters of the network model, the accurate prediction and identification of the first arrival point of the effective signal is finally realized; the number of labels in the training set is reduced, and the quality of the training set and the detection accuracy are improved.

Description of drawings

In order to illustrate the technical solution of the present invention more clearly, the accompanying drawings that need to be used in the technical description of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. Ordinary technicians can also obtain other drawings based on these drawings without paying creative labor.

Fig. 1 is a kind of schematic flow chart of the microseismic effective signal first-arrival picking method based on MS-Net network of the embodiment of the present invention;

Fig. 2 is the schematic diagram of the semi-supervised method of the embodiment of the present invention in conjunction with MS-Net network training process;

Fig. 3 is the first arrival position probability prediction graph of MS-Net network effective signal of the embodiment of the present invention;

Fig. 4 is a kind of microseismic effective signal first-arrival picking-up system schematic diagram based on MS-Net network of the embodiment of the present invention;

Reference signs:

Dataset Production Module-1 Dataset Training Module-2 Output Module-3

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

Fig. 1 is a kind of micro-seismic effective signal based on MS-Net network and picks up the schematic flow chart of the first arrival method of the embodiment of the present invention; As shown in Fig. 1, comprise the following steps:

S1, generating the original data set; specifically including the use of finite difference forward modeling to generate a large number of analog signals with a main frequency range of 20-1000 Hz under different models and actual data to form the original data set;

S2, data set calibration; specifically includes first-arrival picking of some samples in the original data set, selecting the signal waveforms at the first-arrival and non-first-arrival positions of each signal sampling point, and performing calibration respectively, while the other part of the samples are not calibrated;

S3, input the data set into the constructed MS-Net network for training to obtain the optimal network model parameters; specifically including supervised training for the part of the samples that have been calibrated, and the other part that has not been calibrated samples for unsupervised training;

S4. Calculate the probability distribution of the data set point by point; specifically, use the softmax function to output the probability point by point to obtain the binary classification probability of all points, and select the peak probability of the first arrival category as the first arrival point.

Specifically, Fig. 3 is a graph showing the probability prediction curve of the first arrival position of the effective signal of the MS-Net network according to the embodiment of the present invention; The signal waveforms at the arrival and non-first arrivals are calibrated separately, and the other part of the samples are not calibrated; at the same time, these two parts of samples are input into the network training model combined with the MS-Net network and the semi-supervised method. Training and unsupervised training to obtain the optimal network model parameters, that is, the data set effective signal first arrival probability prediction curve; calculate the probability distribution of the data set point by point, use the softmax function to output the probability point by point, and obtain the binary value of each point Classification probability, only select the probability peak value of the first arrival category as the first arrival point, the formula is:

Among them, q _i (x) indicates that x belongs to the predicted probability distribution of different categories, and x indicates each point of the output f(x) of the last layer of the fully convolutional network; when the predicted probability distribution q _i (x) is 0, it means If it is not the first arrival point, the value of i is 2; when the probability value q _i (x) is 1, it represents the first arrival point, and the value of i is 1; k(x) represents the category, k=1, and 2 represents the respective Represents two types of first-arrival points and non-first-arrival points.

Fig. 2 is a schematic diagram of the semi-supervised method combined with the MS-Net network training process of the embodiment of the present invention; as shown in Fig. 2, in the step S3, the optimal network model parameters specifically include: unsupervised loss function and supervised loss The weighted summation of functions constructs a total loss function to minimize the total loss function; the value of the minimized total loss function is less than 0.1.

Specifically, in the semi-supervised method combined with the MS-Net network training process, the total loss function is:

where C is the number of different categories, and B is the mini-batch index set; the evaluation results of the two branches are divided into two different stages: first, the training set is classified without updating the weights, and then the network is different under the same input. Augmentation and dropout training of , using the just-obtained predictions as targets for the unsupervised loss component.

中(其中a是集合动量项)，/>

是训练过程中多轮xi的预测值集成的结果。/>

包含来自先前训练时期的网络集合输出的加权平均值，但最近的时期的权重大于远处的时期。为了产生训练以v为目标，我们需要通过除以因子(1-at)来校正V中的启动偏差，得到

在此过程中，我们可将第一个训练时期的无监督权重函数W(t)指定为零。At the end of each training, the network output v _j is accumulated to the output by updating the extracted feature vector V _j ←aV _j +(1-a)v _j

(where a is the collective momentum term), />

It is the result of integrating the predicted values of xi in multiple rounds during the training process. />

Contains a weighted average of network ensemble outputs from previous training epochs, but with more recent epochs weighted more than distant epochs. To generate training targeting v, we need to correct for the priming bias in V by dividing by a factor (1-at), obtaining

During this process, we can assign the unsupervised weight function W(t) to be zero for the first training epoch.

通过平方差函数形成无监督loss，当前输出V_j与标记样本通过交叉熵函数形成有监督loss，两次的loss值通过加权求和构成网络模型的总loss值，通过最小化总loss值，进而得到最优网络模型。The integration result of the current output V _j and the predicted value of multiple rounds of xi

The unsupervised loss is formed through the square difference function. The current output V _j and the marked sample form a supervised loss through the cross entropy function. The two loss values are weighted and summed to form the total loss value of the network model. By minimizing the total loss value, and then Get the optimal network model.

The embodiment of the present invention provides a first-arrival picking method for micro-seismic effective signals based on MS-Net network. Through MS-Net network combined with semi-supervised method Temporal Ensembling, unsupervised loss and supervised loss are weighted and summed to construct the total loss. Minimize the total loss, optimize the parameters of the network model, and finally realize the accurate prediction and identification of the first arrival point of the effective signal; reduce the number of labels in the training set, and improve the quality of the training set and the detection accuracy.

Based on the above embodiments, Fig. 4 is a schematic structural diagram of a microseismic effective signal first-arrival picking system based on the MS-Net network according to an embodiment of the present invention; as shown in Fig. 4 , it includes:

Data set production module 1, generating original data sets; specifically including the use of finite difference forward modeling to generate a large number of analog signals with a main frequency range of 20-1000 Hz under different models and actual data to form the original data set; data set calibration; specifically including the original data set Part of the samples in the data set are first-arrival picked, and the signal waveforms at the first-arrival and non-first-arrival of each signal sampling point are selected and calibrated separately, and the other part of the samples are not calibrated;

Data set training module 2, which inputs the data set into the constructed MS-Net network for training to obtain optimal network model parameters; specifically includes supervised training for the part of samples that are calibrated, and uncalibrated samples The other part of samples is subjected to unsupervised training;

The output module 3 calculates the probability distribution of the data set point by point; specifically, it includes using the softmax function to output the probability point by point to obtain the binary classification probability of all points, and select the peak probability of the first arrival category as the first arrival point.

The embodiment of the present invention provides a microseismic effective signal first-arrival picking system based on the MS-Net network to implement the above method, through the MS-Net network combined with the semi-supervised method Temporal Ensembling, the unsupervised loss and the supervised loss are weighted and summed to construct a total loss, by minimizing the total loss and optimizing the parameters of the network model, the accurate prediction and identification of the first arrival point of the effective signal is finally realized; the number of labels in the training set is reduced, and the quality of the training set and the detection accuracy are improved.

The device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without any creative effort.

Through the above description of the implementations, those skilled in the art can clearly understand that each implementation can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic discs, optical discs, etc., including several instructions to make a computer device (which may be a personal computer, server, or network device, etc.) execute the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (2)

1. The method for picking up the first arrival of the microseism effective signal based on the MS-Net network is characterized by comprising the following steps:

s1, generating an original data set; the method specifically comprises the steps of generating a large number of analog signals with main frequency range of 20-1000 Hz and actual data to form an original data set under different models by utilizing finite difference forward modeling;

s2, calibrating a data set; picking up a part of samples in the original data set at first arrival, selecting signal waveforms of first arrival and non-first arrival of each signal sampling point, and calibrating the signal waveforms respectively, wherein the other part of samples are not calibrated;

s3, inputting the data set into a constructed MS-Net network for training to obtain optimal network model parameters; the method specifically comprises the steps of performing supervised training on the part of the samples subjected to calibration, and performing unsupervised training on the other part of the samples not subjected to calibration; in step S3, inputting the data set into the constructed MS-Net network for training, where obtaining the optimal network model parameters specifically includes: combining an unsupervised loss function and a supervised loss function through an MS-Net network with a semi-supervised method Temporal Ensembling, and carrying out weighted summation on the unsupervised loss function and the supervised loss function to construct a total loss function, so as to obtain a minimum total loss function; the minimum total loss function value is less than 0.1;

s4, calculating probability distribution of the data set point by point; the method specifically comprises the steps of outputting probabilities point by utilizing a softmax function, obtaining two kinds of probabilities of all points, and selecting probability peaks of first arrival categories as first arrival points.

2. A microseism effective signal first arrival pickup system based on an MS-Net network, comprising:

the data set making module is used for generating an original data set; the method specifically comprises the steps of generating a large number of analog signals with main frequency range of 20-1000 Hz and actual data to form an original data set under different models by utilizing finite difference forward modeling; calibrating a data set; picking up a part of samples in the original data set at first arrival, selecting signal waveforms of first arrival and non-first arrival of each signal sampling point, and calibrating the signal waveforms respectively, wherein the other part of samples are not calibrated;

the data set training module inputs the data set into the constructed MS-Net network for training to obtain optimal network parameters; the method specifically comprises the steps of performing supervised training on the part of the samples subjected to calibration, and performing unsupervised training on the other part of the samples not subjected to calibration; the data set training module inputs the data set into a constructed MS-Net network for training, and the obtaining of the optimal network model parameters specifically comprises the following steps: combining an unsupervised loss function and a supervised loss function through an MS-Net network with a semi-supervised method Temporal Ensembling, and carrying out weighted summation on the unsupervised loss function and the supervised loss function to construct a total loss function, so as to obtain a minimum total loss function; the minimum total loss function value is less than 0.1;

the output module calculates the probability distribution of the data set point by point; the method specifically comprises the steps of outputting probabilities point by utilizing a softmax function, obtaining two kinds of probabilities of all points, and selecting probability peaks of first arrival categories as first arrival points.

Images