Microseismic event intelligent positioning method and system

Abstract

The invention discloses an intelligent positioning method and system for a microseism event. Intelligent positioning of the microseism event is realized by combining a convolutional neural network and probability density function mapping. Comprising the following steps: according to a work area speed model, generating four-dimensional training data representing detector coordinates and seismic phase arrival time and a three-dimensional label representing a seismic source position through Poisson disk sampling and a Gaussian probability density function; designing a convolutional neural network structure, optimizing network weight through back propagation, and constructing a micro-seismic event intelligent positioning model; and inputting actual data into the positioning model to obtain probability density distribution of the seismic source in a three-dimensional space, and finally outputting a high-precision seismic source positioning result through peak value extraction. According to the method, the observation system is creatively integrated into network input, and the strong nonlinear feature extraction capability of the convolutional neural network is combined, so that the generalization capability of the positioning model to different observation systems and the robustness to detector coordinate deviation and arrival time error are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of earth exploration and information technology, and in particular, to a method and system for intelligent microseismic event location. Background Art

[0002] In the field of microseismic monitoring, accurately locating the source position is crucial for microseismic monitoring applications in small and medium-sized work areas such as mine dynamic disaster warning, carbon dioxide storage safety monitoring, unconventional oil and gas field hydraulic fracture evaluation, coalbed methane extraction monitoring, and geothermal reservoir development. Accurate, stable, and reliable microseismic location is a necessary step to improve the monitoring effect. Traditional microseismic location methods mainly include: source location methods based on travel time, grid search methods, energy superposition scanning methods, etc. However, each technology has disadvantages in terms of use: The travel time inversion methods (such as the Geiger method, double-difference location method, etc.) are sensitive to both the error of the underground velocity model and the error of the arrival time. Most work areas are based on a simplified layered velocity model, and the stability of travel time-based algorithms is relatively low, and the location accuracy drops sharply as the geological complexity increases; The resolution of the source location by the grid search method highly depends on the scale of the grid divided for the work area in the early stage, and the high-resolution grid traversal takes a long time, resulting in poor real-time performance of this method; The energy superposition scanning method has high stability, but it is difficult to distinguish dense microseismic events at low resolution, and a large amount of computing power is urgently needed at high resolution, making it difficult to achieve a balance.

[0003] Some machine learning methods attempt to improve the stability, reliability, and automation level of the algorithm through deep neural networks. However, most of these methods rely on fixed geophone coordinates and arrival time information of seismic phases, and the model has a high dependence on the observation system, making it difficult to meet the requirements of multiple deployments of different observation systems in time-lapse monitoring, and the generalization ability is poor. In the actual monitoring process, the real-time monitoring requirements for mine dynamic disaster warning require a millisecond-level response, and conventional methods are difficult to be practical due to high computational loads; while carbon dioxide storage monitoring, hydraulic fracturing monitoring, and geothermal development monitoring have relatively high requirements for the resolution of the results. These factors lead to the general technical contradiction that "high resolution and high efficiency cannot be achieved simultaneously" in the current technology. To address these problems, a new microseismic location method needs to be proposed. By combining the powerful feature extraction ability of the deep convolutional network, integrating the high generalization, strong robustness, and high stability of deep learning methods, and leveraging the high computational efficiency of neural networks, it provides technical support with both real-time performance and credibility for decision-making such as oil and gas field development optimization and geological disaster risk assessment. Summary of the Invention

[0004] The object of the present invention is to provide a method and system for intelligent microseismic event location, which realizes intelligent microseismic event location through the combination of a convolutional neural network and probability density function mapping, can overcome the main disadvantages mentioned in the above background technology, has a certain generalization ability, can adapt to different observation systems in different stages of time-lapse monitoring, and has good adaptability to the uncertainties caused by coordinate deviation and arrival time deviation.

[0005] To achieve the above object, the present invention provides a method and system for intelligent microseismic event location, including the following steps:

[0006] Step 1: Establish a relative coordinate system based on the data of the work area, determine the surface monitoring range of the monitoring work area, convert the positions of the reference point, geophones, and wellheads into geodetic coordinates and altitudes, and establish a relative coordinate system;

[0007] Step 2: Establish a velocity model for the work area, and establish a layered velocity model based on the actual geological and geophysical data of the work area;

[0008] Step 3: Design a simulated observation system, and establish an observation system in the monitoring area according to the principles of uniformity and randomness;

[0009] Step 4: Estimate the seismogenic range and establish a training set of the source location. Referring to the actual engineering situation, estimate the horizontal range and depth range where microseismic events may occur within the work area, establish the spatial distribution interval of the sources used in the training set, and generate a large number of sources within the seismogenic range at a certain interval;

[0010] Step 5: Calculate the travel times, and calculate the P-wave travel times and S-wave travel times from each source location to each geophone location;

[0011] Step 6: Generate training data and labels;

[0012] Step 7: Train the convolutional neural network model. After initializing the network model, input the training data into the network model for calculation, compare the output result with the corresponding label to obtain the error, and then backpropagate the error to train the weights in the network. Record the error each time during training. Stop the network training when the error tends to a certain value and remains stable for several consecutive trainings;

[0013] Step 8: Input the data to be located into the trained convolutional neural network model for realizing source location to obtain the predicted information of the output;

[0014] Step 9: Extract the peak value from the output result, and the peak value is the microseismic location result.

[0015] Preferably, in step 1, the relative coordinate system has the east direction as the X-axis, the north direction as the Y-axis, the opposite direction of gravity as the Z-axis, and the reference point as the coordinate origin. The reference point is selected as a certain lower altitude point below the southwest corner of the work area and the monitored target layer. The X-direction coordinate increases gradually eastward, the Y-direction coordinate increases gradually northward, and the Z-direction coordinate increases gradually upward. All coordinates within the monitoring work area are positive numbers.

[0016] Preferably, the velocity model in step 2 is established based on the previous geological and geophysical data of the work area, including the relative coordinates of the top interface of each layer of stratum, the P-wave velocity of each layer of stratum, and the S-wave velocity of each layer of stratum.

[0017] Preferably, in step 3, Poisson disk random two-dimensional sampling is used to obtain the positions of simulated geophones with uniform distribution and random distribution. This method requires specifying the range of the coordinates of the generated points, the number of generated points, and the minimum distance between the generated points. First, a grid of candidate points is generated within the specified range. After randomly shuffling the order, the candidate points are selected one by one. Then, it is checked whether there are selected points within a certain adjacent range of the grid that are less than the minimum distance. If there is no conflict, the candidate point is retained, and finally, the specified number of sampling points is obtained. These sampling points are used as the XY coordinates of the simulated geophones, and a Z coordinate is randomly assigned to each geophone position within a certain depth range from the ground surface, and finally, the positions of the simulated geophones are obtained.

[0018] Preferably, the source positions in the training set generated in step 4 start from the southwest lower corner of the estimated earthquake occurrence range, and the source coordinates are moved at a fixed interval to generate source positions. First, the X coordinate is moved according to the specified interval. When the X direction reaches the end, the Y coordinate is moved. When the Y direction reaches the end, the Z coordinate is moved, and so on, thus covering a cubic area. The estimated earthquake occurrence range extends a certain area outward from the possible source range to avoid a small number of isolated events falling outside the estimated range.

[0019] Preferably, the number of channels represented by the second dimension of the training data in step 6 is always 5, which respectively represent: the X coordinate of the geophone, the Y coordinate of the geophone, the Z coordinate of the geophone, the arrival time of the P wave on the geophone, and the arrival time of the S wave on the geophone. The number of channels represented by the second dimension of the label is always 3, which respectively represent: the X coordinate of the source, the Y coordinate of the source, and the Z coordinate of the source.

[0020] Preferably, the training data and labels in step 6 represent the position and the arrival time of seismic waves with a Gaussian probability density function. Generate a Gaussian probability density function with a specified number of points, and map this number of points to the length of the work area in this direction or the intercepted time length of the seismic record. The value of the relative coordinate or the arrival time of the seismic wave determines the peak position of the bell-shaped curve. Increasing the number of points of the Gaussian probability density function helps to improve the resolution of the result, but at the cost that it is more difficult for the network to be trained to convergence, requiring more training data and a longer number of training epochs, and slightly reducing the prediction speed of the network. The width of the bell-shaped curve represents the uncertainty of the coordinate or the error of the arrival time. A wider bell-shaped curve is likely to improve the stability of the result, but is not conducive to the reliability of the result. The resolution and real-time performance, as well as the stability and reliability, can be flexibly balanced according to actual needs.

[0021] Preferably, the convolutional neural network in step 7 consists of a downsampling encoder, an upsampling decoder, and an output convolutional layer. The downsampling encoder consists of 8 convolutional layers and 8 max-pooling layers. The convolutional layers used in the downsampling encoder part use 3×3 convolutional kernels to extract input features, keep the output size the same as the input size through padding, and expand the number of channels, then activate through the ReLU activation function, and then perform downsampling using a 2×2 max-pooling kernel. The upsampling decoder part consists of 8 convolutional layers and 8 upsampling layers. The convolutional layers used in the upsampling decoder part use 1×3 convolutional kernels to process the channel data, keep the output size the same as the input size through padding, and reduce the number of channels at the same time, then activate through the ReLU activation function, and then perform upsampling to restore the number of data points using a 1×2 upsampling kernel. The output convolutional layer consists of 4 convolutional layers. The first 3 convolutional layers all use 1×3 convolutional kernels to process the channel data, keep the output size the same as the input size through padding, and then activate through the ReLU activation function; the final output convolutional layer uses 1×3 convolutional kernels to process the channel data, keep the output size the same as the input size through padding, and then activate through the Sigmoid activation function to normalize the output range to a unified value range (from 0 to 1).

[0022] Preferably, the network used in step 8 is the network trained in step 7. The generalization ability of this training result for different observation systems is strong. For a specific work area, only need to train the network once to convergence and then it can be continuously used subsequently, without the need for repeated training.

[0023] The present invention adopts a microseismic positioning method based on a convolutional neural network and probability density function mapping, and has the following beneficial effects compared with the prior art:

[0024] In the training data of the present invention, the geophone positions are converted into Gaussian probability density function representations. The biggest difference from the previous microseismic positioning methods based on convolutional neural networks is that the geophone position information is also used as an input parameter, which is beneficial to adapting to different observation systems at different stages of time-lapse monitoring.

[0025] The present invention represents the travel times of different seismic phases from different source positions to different geophones using Gaussian probability density functions, without the need to use the original waveform records, being unaffected by radiation patterns and noise interference, and effectively improving the anti-interference ability of the positioning model.

[0026] The present invention utilizes the high computational efficiency of convolutional neural networks to promote the real-time performance of the microseismic positioning method, and at the same time utilizes its non-linear feature extraction ability to improve the stability and reliability of the results.

[0027] In summary, the present invention innovatively combines convolutional neural networks and probability density function mapping to propose a microseismic event intelligent positioning method and system, achieving high-efficiency intelligent positioning of microseismic events. For a work area, only one training is required to be applicable to the positioning of microseismic events under different observation systems at different times. Its technical effects cover multiple dimensions such as cost savings, accuracy improvement, and process simplification, providing a new solution for microseismic positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0029] Figure 1 is the flowchart of the implementation case of the present invention;

[0030] Figure 2 is the relationship diagram between the simulated observation system and the source positions of the training set involved in the specific implementation case;

[0031] Figure 3 is the source label and output result diagram involved in the specific implementation case. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following will clearly and completely describe the specific technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the implementation descriptions of the present invention and cannot reflect all the embodiments. Based on the embodiments of the present invention, all other embodiments implemented by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0033] With reference to Figure 1 the process of[[ID=]], a micro-seismic event intelligent positioning method and system disclosed by the present invention realizes intelligent positioning of micro-seismic events through the combination of a convolutional neural network and probability density function mapping. The implementation steps of a specific implementation case are as follows:

[0034] Step 1: Organize the work area scope and establish a relative coordinate system. The implementation case reflects the technical solution of the present invention based on a numerical simulation model. The scope of the established simulation numerical model is 5000m×5000m×2500m.

[0035] Step 2: The simulation numerical model adopts a layered velocity model, and specifies the thickness, P-wave velocity, and S-wave velocity of each layer.

[0036] Step 3: Design a simulation observation system, and use the Poisson disk two-dimensional random sampling method to generate the positions of simulated geophones evenly distributed and randomly distributed within the work area. Set the minimum spacing to 135, and the range of generated geophones is X = [200, 4800], Y = [200, 4800]. The distance from the geophone to the work area boundary is slightly greater than the minimum geophone spacing, and 200 geophones are generated. Then, a random Z coordinate is generated for each geophone within a certain depth. The simulated geophone positions are as shown by the triangular markers in Figure 2 .

[0037] Step 4: Estimate the seismogenic range and establish the training set of seismic source positions. Seismic sources are generated at intervals of 10 within the range of X = [1700, 3000], Y = [1900, 3000], Z = [300, 1500]. The range is as shown by the square boxes in Figure 2 .

[0038] Step 5: Use the ray tracing method accelerated by GPU parallel computing to calculate the P-wave travel time and S-wave travel time from each seismic source point to each geophone.

[0039] Step 6: Extract a certain number of seismic sources from all seismic sources, randomly select several geophones from 200 surface geophones, and generate training data using the geophone positions and the arrival times of seismic waves on the geophones. Randomly select a certain number of geophones, fill the arrival times on these geophones with 0 as simulated bad channels; for each seismic source, randomly specify a time length within a certain range and add it to the arrival times of all geophones to simulate that the seismogenic moment of micro-seismic monitoring is unknown. Generate training labels using the selected seismic source positions, as shown in the upper part of Figure 3 . The length of the training data and labels is 1024 points. The number of seismic sources included in the training set is 12800.

[0040] Step 7: Perform range normalization on the training data and labels, normalizing them to the range from 0 to 1. Measure the error between the network output result and the labels using the binary cross-entropy function. Use the Adam optimizer to train the network with an initial learning rate of 0.001. After 100 epochs of training, the network reaches a convergence state.

[0041] Step 8: Use the test data to input into the network to test the network.

[0042] Step 9: Extract the peak value of the final positioning result.

[0043] The seismic source for testing is located at the position of X = 2480, Y = 2200, Z = 1020. After the specific implementation steps of a microseismic positioning method based on convolutional neural network and probability density function mapping disclosed by the present invention, the obtained seismic source positioning result is X = 2474, Y = 2210, Z = 1037. Evaluate the error of this method in this test. The horizontal direction error is about 11m, the vertical direction error is about 17m, and the total error is about 20m; the errors in each direction are respectively: about 0.12% in the X direction, about 0.20% in the Y direction, about 0.16% in the horizontal direction, about 0.68% in the Z direction, and the total error is about 0.23%. Since the errors are all less than 1%, it can be considered that this method can achieve good positioning accuracy.

Claims

1. A method and system for intelligent microseismic event location, characterized in that, It includes the following steps: Step 1: Establish a relative coordinate system based on the data of the work area, determine the surface monitoring range of the monitoring work area, convert the positions of the reference points, geophones, and wellheads into geodetic coordinates and altitudes, and establish a relative coordinate system; Step 2: Establish a velocity model for the work area, and establish a layered velocity model based on the actual geological and geophysical data of the work area; Step 3: Design a simulated observation system, and establish an observation system in the monitoring area according to the principles of uniformity and randomness; Step 4: Estimate the seismogenic range to establish a training set of the source positions. Referring to the actual engineering situation, estimate the horizontal range and depth range of possible microseismic events within the work area, establish the spatial distribution interval of the sources used in the training set, and generate a large number of sources within the seismogenic range at a certain interval; Step 5: Calculate the travel times, and calculate the P-wave travel times and S-wave travel times from each source position to each geophone position; Step 6: Generate training data and labels; Step 7: Train the convolutional neural network model. After initializing the network model, input the training data into the network model for calculation, compare the output result with the corresponding label to obtain the error, and then backpropagate the error to train the weights in the network. Record the error each time of training. When the error tends to a certain value and remains stable for several consecutive trainings, stop the network training; Step 8: Input the data to be located into the trained convolutional neural network model for source location to obtain the predicted information of the output; Step 9: Extract the peak value from the output result, and the peak value is the microseismic location result.

2. A microseismic event intelligent positioning method and system according to claim 1, which realizes microseismic event intelligent positioning by jointly using a convolutional neural network and probability density function mapping, and is characterized in that The source positions of the training set described in Step 4 are obtained by Poisson disk random two-dimensional sampling to get the positions of simulated geophones with uniform distribution and random distribution; it is necessary to specify the range of the coordinates of the generated points, the number of generated points, and the minimum distance between the generated points. Generate a candidate point grid within the specified range, randomly shuffle the order, and then start selecting candidate points one by one. Then check whether there are selected points with a distance less than the minimum distance in the grids within a certain adjacent range. If there is no conflict, retain the candidate point. Finally, obtain the specified number of sampling points; use these sampling points as the X and Y coordinates of the simulated geophones, and randomly assign a Z coordinate to each geophone position within a certain depth range from the ground. Finally, obtain the positions of the simulated geophones.

3. The intelligent microseismic event location method and system according to claim 1, characterized in that The training data and labels described in step 6 represent the coordinates of the position and the arrival time of the seismic wave using the Gaussian probability distribution function; generate a specified number of Gaussian probability density functions, and map this number of points to the length of the work area in this direction or the intercepted time length of the seismic record. The relative coordinate value or the arrival time of the seismic wave determines the peak position of the bell-shaped curve; the training data is a four-dimensional array, where the first dimension represents the index of the microseismic event, the second dimension represents the data channel, the third dimension represents the geophone index, and the fourth dimension represents the number of data points; the label is a three-dimensional array, where the first dimension represents the index of the source, the second dimension represents the data channel, and the third dimension represents the number of data points; the number of channels represented by the second dimension of the training data is always 5, which respectively represent: the X coordinate of the geophone, the Y coordinate of the geophone, the Z coordinate of the geophone, the arrival time of the P wave on the geophone, and the arrival time of the S wave on the geophone; the number of channels represented by the second dimension of the label is always 3, which respectively represent: the X coordinate of the source, the Y coordinate of the source, and the Z coordinate of the source.

4. The intelligent microseismic event location method and system according to claim 1, characterized in that The convolutional neural network model described in step 7 should have the following structure: it consists of a downsampling encoder, an upsampling decoder, and an output convolutional layer; the downsampling encoder consists of 8 convolutional layers and 8 max pooling layers; the convolutional layers used in the downsampling encoder part use 3×3 convolutional kernels to extract input features, and through padding, the output size is kept the same as the input size, and the number of channels is expanded, then it is activated by the ReLU activation function, and then downsampled using a 2×2 max pooling kernel. The upsampling decoder part consists of 8 convolutional layers and 8 upsampling layers; the convolutional layers used in the upsampling decoder part use 1×3 convolutional kernels to process the channel data, and through padding, the output size is kept the same as the input size, and at the same time the number of channels is reduced, then it is activated by the ReLU activation function, and then upsampled using a 1×2 upsampling kernel to restore the number of data points; the output convolutional layer consists of 4 convolutional layers. The first 3 convolutional layers all use 1×3 convolutional kernels to process the channel data, and through padding, the output size is kept the same as the input size, and then it is activated by the ReLU activation function; the final output convolutional layer uses a 1×3 convolutional kernel to process the channel data, and through padding, the output size is kept the same as the input size, and then it is activated by the Sigmoid activation function.

5. A microseismic event intelligent positioning method and system according to claim 1, characterized in that, The convolutional neural network described in step 7 realizes the mapping from the probability density function of the geophone coordinates and the arrival time of the seismic phase to the probability density function of the source position, and the network output is the source coordinates represented by the Gaussian probability density function.

6. A microseismic event intelligent positioning method and system according to claim 1, characterized in that, The network prediction and positioning information described in step 8 only need to complete the process of training the network from step 1 to step 7 once. In the case where the monitoring work area remains unchanged, the trained network can adapt to multiple observation systems. Especially in the time-lapse monitoring environment, the positioning of microseismic events can be realized without repeated training.

Images