Microseismic P-wave first arrival picking method based on UNet++ combined with Clique Block
The MSPicking-Net network is built through UNet++ and Clique Block, which solves the problem of inefficient pickup from the first to the traditional micro-earthquake, and realizes efficient and fine low signal-to-noise ratio signal pickup, improves the accuracy and efficiency of micro-earthquake monitoring, and is energy-safe Provide a theoretical basis.
Patent Information
- Application Number
- CN202210181956.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-25
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2042-02-25
AI Technical Summary
Traditional micro-earthquake initial pickup technology is inefficient in massive data processing, and it is difficult to accurately pick up micro-earthquake P-wave initial pickup signals with low signal-to-noise ratio, affecting the accuracy of source positioning and crack prediction.
UNet++ combined with Clique Block to build the MSPicking-Net network, generate analog signals through the forward evolution method of finite difference fluctuation equation, and determine hyperparameters in combination with cross-validation to realize deep fine signal-to-noise feature decimation and signal-to-noise separation, improving the picking accuracy of low signal-to-noise ratio signals and efficiency.
It realizes efficient and fine low signal-to-noise ratio micro-seismic signal first-to-end pickup, enhances signal-to-noise discernibility, improves the accuracy of micro-seismic initial pickup, and provides a foundation for intelligent processing for micro-seismic monitoring technology to support energy Safety.
Smart Images
Figure CN114637045B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of earthquake monitoring, and in particular relates to a method for picking up the first arrival of microseismic P waves based on UNet++ combined with Clique Block. Background Art
[0002] Microseismic monitoring technology is a geophysical technology that monitors underground conditions by observing and analyzing tiny earthquake events caused by fracturing. It is of great significance to the stable and high production of oilfield development. The energy of effective microseismic signals is weak, the signal-to-noise ratio is low, and it is even completely submerged in the noise. Therefore, how to accurately and quickly pick up the first arrival of the low signal-to-noise ratio microseismic signal P wave is of great significance for earthquake source location, crack prediction, and earthquake source rupture mechanism analysis.
[0003] Although traditional picking methods (such as STA / LTA, AIC criteria, and correlation detection) have achieved certain results, their efficiency is difficult to meet the needs of first-arrival picking of microseismic signals in the era of big data. Deep learning represented by convolutional neural networks has the characteristics of deep layers and large capacity, and its advantages in high-precision and high-efficiency processing of massive data are becoming increasingly prominent.
[0004] In order to improve the accuracy and efficiency of microseismic P-wave first-arrival picking, this patent proposes a microseismic P-wave first-arrival picking method combining UNet++ and CliqueBlock. Summary of the invention
[0005] In view of the deficiency that the existing microseismic first-arrival picking technology is difficult to efficiently and accurately pick up a large number of microseismic P-wave first arrivals, the present invention uses UNet++ and Clique Block to construct a new network MSPicking-Net to further enhance the signal-to-noise discernibility of low signal-to-noise ratio microseismic signals and improve the accuracy of microseismic first-arrival picking; To achieve the above technical effects, the technical solution adopted by the present invention is:
[0006] A method for picking the first arrival of microseismic P waves based on UNet++ and Clique Block, comprising the following steps:
[0007] S1: Conduct downhole or surface observations in the fracturing area and collect hydraulic fracturing microseismic monitoring data; the monitored microseismic monitoring data has a low signal-to-noise ratio and weak effective signal energy;
[0008] S2: Use the finite difference wave equation forward modeling method to generate forward simulation signals under different main frequencies and different formation velocity models;
[0009] The finite difference wave equation forward modeling method adopts the staggered grid finite difference forward modeling method described in the paper "Sheng Guanqun. Research on microseismic signal detection and source location method [D]. China University of Petroleum (East China), 2016."
[0010] S3: setting different signal-to-noise ratios for the forward modeling signal obtained in S2, constructing an original data set together with the actual monitoring profile, performing first arrival picking on the original data set, selecting the signal waveforms at the first arrival and non-first arrival locations, and calibrating them respectively;
[0011] S4: Add Clique Block to the UNet++ network to build the MSPicking-Net network. The number of CliqueBlock blocks and the hyperparameter settings of MSPicking-Net are determined by cross-validation method.
[0012] Preferably, the cross validation method uses LOOCV or K-fold Cross Validation method for validation;
[0013] In Clique Block, except for the input nodes, any two layers in the same block are bidirectionally connected; in a block, each layer is both the input and output of other layers; in Clique Block, each layer is the output of the previous layer, and the input of the next layer is the output of any other layer; the propagation of a clique with 5 layers is shown in Table 1 below:
[0014]
[0015] Table 1:
[0016] In the first stage, the input layer (X 0 ) All layers in this block are initialized with unidirectional connections, each updated layer is connected to update the next layer; starting from the second stage, the layers begin to be updated alternately, all layers except the top layer to be updated are connected as bottom layers, and their corresponding parameters are also connected together; therefore, the output of the i-th layer (i ≥ 1) In the k-th layer (k ≥ 2) loop, it can be expressed as:
[0017]
[0018] Among them, * indicates the use of hyperparameter W for convolution operation, g represents the nonlinear activation function; W ij Maintain reuse at different stages;
[0019] S5: Training set preparation and network training: Randomly extract the first arrival position and non-first arrival position signals of the samples obtained in S3 to obtain the training set and complete the MSPicking-Net training;
[0020] S6: Output the first arrival position: Input the signal to be tested into the network, use Softmax to predict the "binary classification" probability of each sampling point of the signal to be tested belonging to the first arrival category and the non-first arrival category, and finally take the point corresponding to the probability peak in the first arrival category as the first arrival, and use the accuracy and recall rate curve to evaluate the algorithm; the Softmax formula is as follows:
[0021]
[0022] The loss function is given by the probability distribution p of the true value i (x) and the network predicted probability distribution q i The cross entropy C(p,q) between (x) is calculated as:
[0023]
[0024] Among them, x represents each point of the last layer output f(x) of the full convolutional network;
[0025] i represents the category: 1 represents the first arrival, 2 represents the non-first arrival;
[0026] q i (x) represents the predicted probability of x belonging to different categories.
[0027] Preferably, in the above-mentioned method for picking up the first arrival of microseismic P waves based on UNet++ and Clique Block, the different main frequencies and different formation velocity models in step S2 include the Mamousi model, multi-layer flat layer and inclined layer velocity models; the forward simulation signals generated by different velocity models are different, which are used to ensure sufficient abundance of samples.
[0028] Furthermore, the UNet++ network uses the UNet++ network technology in the paper "Zhou ZW, Siddiquee MMR, Tajbakhsh N, et al. 2018. Unet++: A nested u-net architecture for medical image segmentation"; the Clique Block uses the technology in the paper "Yang, Y., et al. "Convolutional Neural Networks with Alternately Updated Clique." 2018 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR) IEEE, 2018."
[0029] The beneficial effects of the present invention are:
[0030] 1. Aiming at the disadvantage of low efficiency of traditional microseismic first-arrival picking in processing massive microseismic signals, this patent uses deep learning to guide the first-arrival picking of microseismic signals, thus achieving efficient first-arrival picking of massive microseismic signals;
[0031] 2. Aiming at the low signal-to-noise ratio environment of microseismic signals and the defects of Unet++ such as shallow layers and difficulty in capturing fine microseismic signal-to-noise features, a new network MSPicking-Net is constructed by combining UNet++ network and Clique Block, which realizes deep and fine microseismic signal-to-noise feature extraction, strengthens the signal-to-noise separation of microseismic low signal-to-noise ratio signals, and realizes faster and more precise low signal-to-noise ratio microseismic first-arrival picking, further enhances the signal-to-noise distinguishability of low signal-to-noise ratio microseismic signals, and improves the accuracy of microseismic first-arrival picking.
[0032] 3. This invention points out an important application direction for microseismic monitoring technology, promotes the application and promotion of this technology, and lays the research foundation for microseismic intelligent processing algorithms by applying artificial intelligence technology to the traditional microseismic monitoring field. It can also provide an important theoretical basis for serving the country's energy strategic security. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a schematic diagram of the workflow of the present invention;
[0034] Figure 2 It is a schematic diagram of the MSPicking-Net network structure in the present invention;
[0035] Figure 3 It is a diagram of a signal to be detected with a signal-to-noise ratio of SNR=-10dB in the first arrival picking result of microseismic signals based on MSPicking-Net of the present invention;
[0036] Figure 4 It is the probability prediction curve of the first arrival position of the effective signal of the U-Net++ network in the first arrival picking result of the microseismic signal based on MSPicking-Net of the present invention;
[0037] Figure 5 It is a probability prediction curve of the first arrival position of effective signal of MSPicking-Net network in the first arrival picking result of microseismic signal based on MSPicking-Net in the present invention; DETAILED DESCRIPTION
[0038] Embodiment 1:
[0039] like Figure 1 As shown, a method for picking the first arrival of microseismic P waves based on UNet++ and Clique Block includes the following steps:
[0040] S1: Conduct downhole or surface observations in the fracturing area and collect hydraulic fracturing microseismic monitoring data; the monitored microseismic monitoring data has a low signal-to-noise ratio and weak effective signal energy;
[0041] S2: Use the finite difference wave equation forward modeling method to generate forward simulation signals under different main frequencies and different formation velocity models;
[0042] The finite difference wave equation forward modeling method adopts the staggered grid finite difference forward modeling method described in the paper "Sheng Guanqun. Research on microseismic signal detection and source location method [D]. China University of Petroleum (East China), 2016."
[0043] S3: setting different signal-to-noise ratios for the forward modeling signal obtained in S2, constructing an original data set together with the actual monitoring profile, performing first arrival picking on the original data set, selecting the signal waveforms at the first arrival and non-first arrival locations, and calibrating them respectively;
[0044] S4: Figure 2 As shown in Figure 1, Clique Block is added to the UNet++ network to construct the MSPicking-Net network. The number of Clique Blocks and the hyperparameter settings of MSPicking-Net are determined by cross-validation. The MSPicking-Net network structure is shown in Figure 1. Figure 2 As shown;
[0045] Preferably, the cross-validation method uses the LOOCV method for validation;
[0046] In Clique Block, except for the input nodes, any two layers in the same block are bidirectionally connected; in a block, each layer is both the input and output of other layers; in Clique Block, each layer is the output of the previous layer, and the input of the next layer is the output of any other layer; the propagation of a clique with 5 layers is shown in Table 1 below:
[0047]
[0048] Table 1:
[0049] In the first stage, the input layer (X 0 ) All layers in this block are initialized with unidirectional connections, each updated layer is connected to update the next layer; starting from the second stage, the layers begin to be updated alternately, all layers except the top layer to be updated are connected as bottom layers, and their corresponding parameters are also connected together; therefore, the output of the i-th layer (i ≥ 1) In the k-th layer (k ≥ 2) loop, it can be expressed as:
[0050]
[0051] Among them, * indicates the use of hyperparameter W for convolution operation, g represents the nonlinear activation function; W ij Maintain reuse at different stages;
[0052] S5: Training set preparation and network training: Randomly extract the first arrival position and non-first arrival position signals of the samples obtained in S3 to obtain the training set and complete the MSPicking-Net training;
[0053] S6: Output the first arrival position: Input the signal to be tested into the network, use Softmax to predict the "binary classification" probability of each sampling point of the signal to be tested belonging to the first arrival category and the non-first arrival category, and finally take the point corresponding to the probability peak in the first arrival category as the first arrival, and use the accuracy and recall rate curve to evaluate the algorithm; the Softmax formula is as follows:
[0054]
[0055] The loss function is given by the probability distribution p of the true value i (x) and the network predicted probability distribution q i The cross entropy C(p,q) between (x) is calculated as:
[0056]
[0057] Among them, x represents each point of the last layer output f(x) of the full convolutional network;
[0058] i represents the category: 1 represents the first arrival, 2 represents the non-first arrival;
[0059] q i (x) represents the predicted probability of x belonging to different categories.
[0060] Preferably, in the above-mentioned method for picking up the first arrival of microseismic P waves based on UNet++ and Clique Block, the different main frequencies and different formation velocity models in step S2 include the Mamousi model, multi-layer flat layer and inclined layer velocity models; the forward simulation signals generated by different velocity models are different, which are used to ensure sufficient abundance of samples.
[0061] Embodiment 2:
[0062] like Figure 3 to Figure 5 The figure shows the schematic diagram of the first arrival picking results of microseismic signals based on MSPicking-Net in practical applications;
[0063] Figure 3The forward simulation signal with -10dB Gaussian noise added has a first arrival time of 1330ms and a main frequency of 25Hz. Figure 4 It is the probability prediction curve of the first arrival position of the effective signal of the U-Net++ network; Figure 5 is the probability prediction curve of the first arrival position of the effective signal of the MSPicking-Net network; Figure 4 and Figure 5 It can be seen that the first arrival time of the microseismic signal picked up by the UNet++ network is 1480ms, while the first arrival time of the microseismic signal picked up by the method proposed in this patent is 1330ms, which is completely consistent with the first arrival time of the simulated signal; therefore, compared with the UNet++ network, the initial picking results of the microseismic signal of the low signal-to-noise ratio scheme of this patent are more reliable, which once again proves that this patent actually achieves the technical effect of optimizing the low signal-to-noise ratio characteristics of the microseismic signal and the defects of the UNet++ network. The MSPicking-Net network proposed in this patent has better performance in actual use.
[0064] The above-mentioned embodiments are only preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The embodiments and features in the embodiments of the present application can be arbitrarily combined with each other without conflict. The protection scope of the present invention shall be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.
Claims
1. A method for picking microseismic P-wave first arrivals based on UNet++ and Clique Block, characterized by: The steps include: S1: Conduct downhole or surface observations in the fracturing area and collect hydraulic fracturing microseismic monitoring data; S2: Use the finite difference wave equation forward modeling method to generate forward simulation signals under different main frequencies and different formation velocity models; S3: setting different signal-to-noise ratios for the forward modeling signal obtained in S2, constructing an original data set together with the actual monitoring profile, performing first arrival picking on the original data set, selecting the signal waveforms at the first arrival and non-first arrival locations, and calibrating them respectively; S4: Add Clique Block to the UNet++ network to build the MSPicking-Net network. The number of CliqueBlock blocks and the hyperparameter settings of MSPicking-Net are determined by cross-validation method. S5: Training set preparation and network training: Randomly extract the first arrival position and non-first arrival position signals of the samples obtained in S3 to obtain the training set and complete the MSPicking-Net training; S6: Output the first arrival position: Input the signal to be tested into the network, use Softmax to predict the "binary" probability of each sampling point of the signal to be tested belonging to the first arrival category and the non-first arrival category, and then output the point corresponding to the probability peak in the first arrival category as the first arrival position; finally, use the accuracy and recall rate curve to evaluate the algorithm; the Softmax formula is as follows: The loss function is given by the probability distribution p of the true value i (x) and the network predicted probability distribution q i The cross entropy C(p,q) between (x) is calculated as: Among them, x represents each point of the last layer output f(x) of the full convolutional network; i represents the category: 1 represents the first arrival, 2 represents the non-first arrival; q i (x) represents the predicted probability of x belonging to different categories.
2. The method for picking microseismic P wave first arrival based on UNet++ and Clique Block according to claim 1, characterized in that: In step S2, different main frequency and different formation velocity models include Mamousi model, multi-layer flat layer and inclined layer velocity models; the forward simulation signals generated by different velocity models are different.