Method for real-time inversion of micro-seismic data of support resistance based on conditional image translation technology

Abstract

The invention discloses a method for real-time inversion of micro-seismic data of support resistance based on a conditional image translation technology, and belongs to the field of inversion of internal monitoring data of rock stratum movement. The technical problem to be solved is that an existing microseismic data acquisition mode is low in precision, high in cost and difficult to popularize. According to the technical scheme, the method comprises the steps that support resistance and microseismic data are collected with mining circulation as a unit; establishing a working face relative coordinate system, converting a micro-seismic event coordinate, constructing a support resistance and micro-seismic spatial-temporal characteristic matrix, and aligning the support resistance and micro-seismic spatial-temporal characteristic matrix; regarding the feature matrix as an image, and constructing and training a conditional image translation model; and applying the model to mining circulation, and inputting real-time support resistance data to invert micro-seismic data. The method is mainly used for economically and conveniently acquiring micro-seismic data and providing conditions for popularizing micro-seismic monitoring and realizing accurate rock stratum control on all working surfaces.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of inverting internal monitoring data of rock movement, and particularly relates to a support resistance real-time inversion microseismic data method based on conditional image translation technology. BACKGROUND

[0002] Microseismic monitoring is a core technology for obtaining information on internal fracture and instability of rock movement, and plays an irreplaceable role in preventing roof disasters during coal mining and achieving precise rock control. At present, microseismic data is mainly collected by laying a sensor network and calculated by using a seismic source positioning algorithm. However, this technical path has significant limitations in actual mine applications. First, the accuracy of the solved microseismic events is long-term suboptimal, which is a bottleneck problem restricted by multiple factors. Specifically, the underground monitoring environment is complex and variable, the rock mass rupture forms are various, the sound wave signal will be distorted in the propagation of non-uniform medium, and the sensor receiving performance, network layout rationality and the calculation accuracy of the positioning algorithm itself are limited, which together cause large errors in spatial position and energy size of the finally obtained microseismic data, and it is difficult to meet the needs of high-precision rock movement analysis and disaster warning. Second, the economic cost and operation burden of the microseismic monitoring system are too high. A complete set of microseismic monitoring equipment is expensive, and a large amount of manpower is needed for long-term equipment maintenance, signal reception and data processing during mining. This high threshold directly leads to the fact that the microseismic monitoring technology cannot be applied in all coal mining faces, and most coal mines only deploy it in mines with significant impact tendency. This makes the rich rock movement information contained in the microseismic data unable to be normalized and systematically used to guide daily roof management and other rock control practices, thereby limiting its potential in improving the overall safety of coal mining. For a long time, the field has been committed to seeking breakthroughs within the existing technical framework, such as optimizing sensor layout schemes or improving seismic source positioning algorithms, but often only partial improvements can be achieved, and it is difficult to fundamentally solve the two interrelated problems of precision and cost. Seeking a new data acquisition approach that can break away from the dependence on expensive hardware and complex algorithms has been a major technical difficulty. SUMMARY

[0003] To solve the above technical problems, the present application proposes a support resistance real-time inversion microseismic data method based on conditional image translation technology, which significantly reduces the dependence on expensive special monitoring equipment, reduces the manpower investment brought by equipment installation and maintenance and signal processing, and greatly reduces the overall cost and application threshold of microseismic monitoring. This provides feasibility for the popularization of microseismic monitoring technology in all working faces, thereby providing continuous and reliable data support for more accurate rock movement analysis and roof disaster prevention and control.

[0004] To achieve the above objectives, this invention provides a method for real-time inversion of microseismic data based on conditional image translation technology for support resistance, comprising: S1. Collect support resistance data and microseismic data from historical mining cycles in real time, and determine the time range for which microseismic data needs to be inverted; S2. Based on the support resistance data and microseismic data, establish a relative coordinate system for the working surface, transform the microseismic events to the relative coordinate system, and construct the spatiotemporal feature matrix of support resistance and the spatiotemporal feature matrix of microseismic events respectively. Align the two to establish a model training set. S3. Treat the spatiotemporal feature matrix of the support resistance and the spatiotemporal feature matrix of the microseismic event as images, construct a conditional image translation model, train it using the training set data, and determine the optimal model based on the evaluation index. S4. Apply the optimal model to the ongoing mining cycle, input the real-time support resistance spatiotemporal feature matrix, output the generated microseismic spatiotemporal feature matrix, and obtain the inverted microseismic data through inverse normalization processing.

[0005] Optional, collecting historical cycle support resistance data and microseismic data includes: Collect the cycle number, start and end time, and location of historical cycles. Collect complete support resistance data, with a measuring line set every 8 to 12 supports. Only the data of the measuring lines set in that cycle are collected. The collected cycle measuring line data includes complete support resistance data of all measuring lines in that cycle as well as support resistance data of the two cycles before and after that cycle. Collect complete microseismic data, only collecting microseismic event data above the coal seam floor. The microseismic events have complete time, three-dimensional spatial coordinates, and energy characteristics.

[0006] Optionally, in S2, establishing the relative coordinate system of the working surface includes: A gridded model of the entire working face area was established, with support numbers and coal cutting cycles as coordinate units. Only supports set as survey lines were selected. The origin was set at the position of support number 1 and mining cycle 1 in the coal seam floor. The working face advance direction was the positive x-axis, the direction of support number increase was the positive y-axis, and the vertical upward direction was the positive z-axis. A relative coordinate system was established, with the unit of the relative coordinate system being meters. The three-dimensional spatial coordinates of microseismic events were transformed to the relative coordinate system. The z-coordinate of the microseismic event was converted to the difference between the original elevation and the elevation of the origin of the relative coordinate system. The x and y coordinates were transformed using a planar rotation transformation.

[0007] Optionally, in S2, constructing the spatiotemporal characteristic matrix of the stent resistance includes: A spatiotemporal feature matrix of stent resistance is constructed on a cycle-by-cycle basis. For each grid, the end-cycle resistance values ​​of the 25 grids centered on that grid are obtained to form the spatiotemporal feature matrix of that grid. If the end-cycle resistance value of the grid is missing or abnormal, it is filled by the average value of other grids in the spatiotemporal feature matrix of the grid. All spatiotemporal feature matrices of the grid under that cycle are stacked to construct the spatiotemporal feature matrix of stent resistance.

[0008] Optionally, in S2, constructing the microseismic spatiotemporal feature matrix includes: A microseismic spatiotemporal feature matrix is ​​constructed in cycles. The matrix consists of all microseismic events within the time range of the microseismic data to be inverted in the cycle. Each row of the matrix contains five features of a microseismic event, including time, three-dimensional spatial coordinates, and energy features. A uniform number of rows is set according to the cycle containing the most microseismic events. For microseismic spatiotemporal feature matrices with less than a uniform number of rows, they are filled with 1s after normalization. The filled microseismic spatiotemporal feature matrix is ​​then converted into a three-dimensional shape of 5 rows and 5 columns multiplied by the uniform number of rows divided by 5.

[0009] Optionally, in S2, aligning the support resistance and microseismic spatiotemporal characteristic matrices includes: The microseismic spatiotemporal feature matrix is ​​processed by converting the time characteristics of microseismic events into the difference between the cycle end time and the original time characteristics, while keeping the energy characteristics unchanged. The coordinates of the cycle center grid in the relative coordinate system are calculated, where the x-coordinate is obtained by subtracting 0.5 from the cycle number and multiplying by the cutting depth of the coal mining machine, the y-coordinate is obtained by subtracting 1 from the survey line number, dividing by 2, adding 0.5, and multiplying by the support width, and the z-coordinate is 0. The coordinate values ​​of the cycle center grid are subtracted from the three-dimensional spatial coordinate characteristics of the microseismic events respectively. The transformed feature values ​​of all microseismic events in the training set data are statistically analyzed, and the maximum and minimum values ​​of each feature are used as normalization parameters to normalize the microseismic data. Microseismic spatiotemporal feature matrices with insufficient rows are padded with 1s to obtain aligned training set data.

[0010] Optionally, in S3, building a conditional image translation model and training it using the training set data includes: The spatiotemporal feature matrices of support resistance and microseismic spatiotemporal feature matrices are treated as images. An image translation model based on a conditional generative adversarial network architecture is constructed. The generator takes the spatiotemporal feature matrix of support resistance as input and generates the spatiotemporal feature matrix of microseismic motion. The discriminator distinguishes the generated spatiotemporal feature matrix of microseismic motion from the real spatiotemporal feature matrix of microseismic motion. The model is trained using training set data. The model parameters are optimized through adversarial training process, and the error index between the generated data and the real data is monitored to determine the optimal training rounds.

[0011] Optionally, in S4, applying the optimal model to the cycle being mined includes: A spatiotemporal feature matrix of support resistance during the mining cycle is constructed. The spatiotemporal feature matrix of support resistance is input into the optimal model, and the generated spatiotemporal feature matrix of microseismic activity is output. The generated spatiotemporal feature matrix of microseismic activity is inversely normalized using the normalization parameters used during model training to obtain inverted microseismic data.

[0012] Technical Advantages of this Invention: This invention discloses a method for real-time inversion of microseismic data based on support resistance technology. By combining support resistance data with conditional image translation technology, it opens up a completely new path for acquiring microseismic data. This method effectively overcomes the accuracy limitations of traditional microseismic monitoring technologies caused by factors such as complex environments, heterogeneous media, and equipment layout during signal reception and algorithm localization, and can directly generate microseismic data that meets application requirements. Simultaneously, this method significantly reduces reliance on expensive dedicated monitoring equipment, reduces manpower investment in equipment installation, maintenance, and signal processing, and greatly lowers the overall cost and application threshold of microseismic monitoring. This provides feasibility for the widespread application of microseismic monitoring technology in all working faces, thereby providing continuous and reliable data support for more accurate rock movement analysis and roof disaster prevention. Attached Figure Description

[0013] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a method for real-time inversion of microseismic data based on conditional image translation technology for support resistance according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the coal and rock occurrence characteristics of the working face in an embodiment of the present invention. Figure 3 This is a diagram showing the mining area and relative coordinate system of the data collected in this embodiment of the invention; Figure 4 A top view of the distribution of microseismic events was collected for embodiments of the present invention; Figure 5 A front view of the distribution of microseismic events was collected for embodiments of the present invention; Figure 6 This is a schematic diagram illustrating the construction of the spatiotemporal feature matrix of the support resistance in an embodiment of the present invention; Figure 7 A heat map of the distribution of support resistance data was collected for an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the construction process of the microseismic spatiotemporal feature matrix according to an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating the changes in data distribution during the training process of the GAN in this embodiment of the invention; Figure 10 This is a schematic diagram of the pix2pix model training process according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the generator model structure of the GAN in an embodiment of the present invention; Figure 12 This is a schematic diagram of the improved generator model structure according to an embodiment of the present invention; Figure 13 This is a schematic diagram of the discriminator's discrimination error change curve during the model training process of an embodiment of the present invention; Figure 14 This is a schematic diagram of the generator error variation curve during model training in an embodiment of the present invention. Detailed Implementation

[0014] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0016] like Figure 1 As shown, this embodiment provides a method for real-time inversion of microseismic data based on conditional image translation technology for support resistance, including: S1. Collect support resistance data and microseismic data from historical mining cycles in real time, and determine the time range for which microseismic data needs to be inverted; S2. Based on the support resistance data and microseismic data, establish a relative coordinate system for the working surface, transform the microseismic events to the relative coordinate system, and construct the spatiotemporal feature matrix of support resistance and the spatiotemporal feature matrix of microseismic events respectively. Align the two to establish a model training set. S3. Treat the spatiotemporal feature matrix of the support resistance and the spatiotemporal feature matrix of the microseismic event as images, construct a conditional image translation model, train it using the training set data, and determine the optimal model based on the evaluation index. S4. Apply the optimal model to the ongoing mining cycle, input the real-time support resistance spatiotemporal feature matrix, output the generated microseismic spatiotemporal feature matrix, and obtain the inverted microseismic data through inverse normalization processing.

[0017] Further, the collection of historical cycle support resistance data and microseismic data includes: Collect the cycle number, start and end time, and location of historical cycles. Collect complete support resistance data, with a measuring line set every 8 to 12 supports. Only the data of the measuring lines set in that cycle are collected. The collected cycle measuring line data includes complete support resistance data of all measuring lines in that cycle as well as support resistance data of the two cycles before and after that cycle. Collect complete microseismic data, only collecting microseismic event data above the coal seam floor. The microseismic events have complete time, three-dimensional spatial coordinates, and energy characteristics.

[0018] Specifically, the implementation process of this embodiment includes: S1. The real-time inversion process is based on mining cycles. The time range of microseismic data to be inverted is determined, historical cycles are selected for the production of training sets, and the support resistance and microseismic data of historical cycles are collected. Step S1 determines the time range of the microseismic data to be inverted, using the following method: By determining the time range for which microseismic data needs to be retrieved, all microseismic events related to the resistance distribution characteristics of the support system in this mining cycle are screened out, while microseismic events unrelated to these characteristics are minimized as much as possible. The end point of the time range is the end time of the mining cycle, and the start point is determined by subtracting the duration from the end point. The duration is determined based on experience and needs. Step S1 selects the historical loop used to create the training set, using the following method: The historical cycles used to create the training set must be located in the same working face or a working face with similar mining conditions as the predicted cycles. The criteria for judging similar mining conditions are: the working face of the predicted cycle is located in the same mining area and the same coal seam is being mined; the coal seam occurrence conditions, including coal seam thickness, dip angle, burial depth, and roof and floor lithology, are not significantly different; the four adjacent relationships of the working faces are not significantly different; and the mining technology adopted is not significantly different. Step S1 collects historical cycle support resistance and microseismic data, using the following method: Collect the following three data items from the historical cycle: ①The cycle number, start and end times, and position of the historical cycle; ② Complete support resistance data. A measurement line is set up every few supports (8-12 supports are optimal), and data is collected only for the measurement line set up within that cycle. The collected cyclic measurement line data should include not only complete support resistance data for all measurement lines within that cycle, but also support resistance data for the two cycles before and after that cycle, to ensure that a cyclic support resistance spatiotemporal feature matrix can be constructed; ③ Complete microseismic data (collecting only microseismic events above the coal seam floor). Identify all microseismic events within the inversion timeframe; these events must retain complete temporal, three-dimensional spatial coordinates, and energy characteristics.

[0019] This embodiment selects the 8102 working face of a mine in the Datong mining area as the research object. The required duration for microseismic data retrieval was determined to be 5 days, and a total of 630 historical cycles meeting the criteria were collected. These 630 cycles were divided into two parts at an 8:2 ratio: one part was used for model training, and the other part was considered as the current mining cycle for verifying the application effect. The mining conditions and data collection status of the 8102 working face are as follows: The 8102 working face mines the 3-5# coal seams at a depth of approximately 450m. The coal seam thickness is 16.22m, the immediate roof thickness is 3.61m, and the existing roof thickness is 15.56m. A single-strike longwall retreat fully mechanized low-level top-coal caving mining method is employed. The average strike length is 1516.5m, the dip length is 251m, and the mining height is 3.9m. Coal is caved using a combination of one-cut-one-caving sequence and intervals, and the roof is managed using the natural caving method. The coal and rock occurrence characteristics of the working face are as follows: Figure 2 As shown.

[0020] The 8102 working face is flanked by the goaf areas of the 8101 and 8103 working faces, respectively, with each section having a coal pillar width of 6 meters, making it an isolated working face. Mining of the 8101 working face ceased in October 2010, and mining of the 8103 working face ceased in March 2015, both experiencing significant mining pressure during their respective mining periods. Considering that the 8102 working face, operating as an isolated working face, would be subject to even stronger mining pressure due to the combined effects of the advance support pressure from mining and the bidirectional high support pressure from the overhanging roof of the adjacent working face goaf, mining commenced in 2020, five years after the completion of mining of the 8103 working face and after the overlying strata had stabilized.

[0021] The 8102 working face uses a total of 147 supports. The central supports are 139 ZF21000 / 27.5 / 42D four-column shield-type low-position top-coal caving supports; the transition supports are 7 ZFG21000 / 29 / 42D four-link four-column shield-type top-coal caving transition supports; and the end supports are a set of ZTZ30000 / 27.5 / 42D top-coal caving end supports. During mining, support resistance monitoring was conducted on the working face. Fourteen observation lines were evenly distributed across the face, with each line spaced approximately 10 supports apart. The 8102 working face uses the Tianma electro-hydraulic control system for support resistance monitoring. This system can accurately read the resistance value at the end of each support cycle, providing precise support resistance data for on-site monitoring and scientific research. The historical mining cycle for which data was collected spans from October 31, 2020 to April 13, 2021. The mining area and the layout of the support resistance measuring lines are as follows: Figure 3 As shown.

[0022] For microseismic data, complete data was collected from October 15, 2020 to May 26, 2021. The mining area of ​​the working face covered the historical cycle area of ​​the collected data, and all 630 cycles of microseismic data requiring inversion were fully collected. A total of 6,648 microseismic events were collected. Since only microseismic events above the overlying strata (i.e., the coal seam floor) needed to be analyzed, and the coal seam floor elevation was 791, 4,525 microseismic events were screened.

[0023] Furthermore, in S2, establishing the relative coordinate system of the working surface includes: A gridded model of the entire working face area was established, with support numbers and coal cutting cycles as coordinate units. Only supports set as survey lines were selected. The origin was set at the position of support number 1 and mining cycle 1 in the coal seam floor. The working face advance direction was the positive x-axis, the direction of support number increase was the positive y-axis, and the vertical upward direction was the positive z-axis. A relative coordinate system was established, with the unit of the relative coordinate system being meters. The three-dimensional spatial coordinates of microseismic events were transformed to the relative coordinate system. The z-coordinate of the microseismic event was converted to the difference between the original elevation and the elevation of the origin of the relative coordinate system. The x and y coordinates were transformed using a planar rotation transformation.

[0024] Specifically, the implementation process of this embodiment includes: S2. Establish a gridded model of the entire working face area with support number and coal cutting cycle as coordinate units. Based on the gridded model, establish a relative coordinate system for the working face, transform the microseismic events to the relative coordinate system, construct the support resistance and microseismic spatiotemporal characteristic matrices respectively, align the two and establish a model training set.

[0025] Step S2 establishes a relative coordinate system for the working surface and transforms microseismic events to this relative coordinate system. The method is as follows: A gridded model of the entire working face area was established, using support numbers and coal cutting cycles as coordinate units. Support numbers were only selected as survey lines. The origin was set at the position of support number 1 and mining cycle 1 in the coal seam floor, and the working face advance direction was... The positive axis direction, the direction of the bracket number increase is The positive direction of the axis, the vertically upward direction is Establish a relative coordinate system along the positive axis. In this relative coordinate system, all units are meters (m), consistent with the actual unit of measurement, length. Transform the three-dimensional spatial coordinates of microseismic events to a relative coordinate system. Microseismic events The coordinates are converted to the difference between the original elevation and the elevation relative to the origin of the coordinate system. The coordinates are transformed using a planar rotation transformation formula, as follows: ; In the formula, This represents the coordinate values ​​of a microseismic event in a relative coordinate system. This represents the coordinate values ​​of a microseismic event in the absolute coordinate system. These represent the coordinates of the origin in the absolute coordinate system. It represents the angle of clockwise rotation of the relative coordinate system relative to the absolute coordinate system.

[0026] Furthermore, in S2, the construction of the spatiotemporal characteristic matrix of the stent resistance includes: A spatiotemporal feature matrix of stent resistance is constructed on a cycle-by-cycle basis. For each grid, the end-cycle resistance values ​​of the 25 grids centered on that grid are obtained to form the spatiotemporal feature matrix of that grid. If the end-cycle resistance value of the grid is missing or abnormal, it is filled by the average value of other grids in the spatiotemporal feature matrix of the grid. All spatiotemporal feature matrices of the grid under that cycle are stacked to construct the spatiotemporal feature matrix of stent resistance.

[0027] Furthermore, in S2, the construction of the microseismic spatiotemporal feature matrix includes: A microseismic spatiotemporal feature matrix is ​​constructed in cycles. The matrix consists of all microseismic events within the time range of the microseismic data to be inverted in the cycle. Each row of the matrix contains five features of a microseismic event, including time, three-dimensional spatial coordinates, and energy features. A uniform number of rows is set according to the cycle containing the most microseismic events. For microseismic spatiotemporal feature matrices with less than a uniform number of rows, they are filled with 1s after normalization. The filled microseismic spatiotemporal feature matrix is ​​then converted into a three-dimensional shape of 5 rows and 5 columns multiplied by the uniform number of rows divided by 5.

[0028] Furthermore, in S2, aligning the support resistance and microseismic spatiotemporal characteristic matrices includes: The microseismic spatiotemporal feature matrix is ​​processed by converting the time characteristics of microseismic events into the difference between the cycle end time and the original time characteristics, while keeping the energy characteristics unchanged. The coordinates of the cycle center grid in the relative coordinate system are calculated, where the x-coordinate is obtained by subtracting 0.5 from the cycle number and multiplying by the cutting depth of the coal mining machine, the y-coordinate is obtained by subtracting 1 from the survey line number, dividing by 2, adding 0.5, and multiplying by the support width, and the z-coordinate is 0. The coordinate values ​​of the cycle center grid are subtracted from the three-dimensional spatial coordinate characteristics of the microseismic events respectively. The transformed feature values ​​of all microseismic events in the training set data are statistically analyzed, and the maximum and minimum values ​​of each feature are used as normalization parameters to normalize the microseismic data. Microseismic spatiotemporal feature matrices with insufficient rows are padded with 1s to obtain aligned training set data.

[0029] Specifically, the implementation process of this embodiment includes: The transformed feature values ​​of all microseismic events in the training set data were statistically analyzed, and the maximum and minimum values ​​of the five features were obtained as normalization parameters. The microseismic data were then normalized, and microseismic spatiotemporal feature matrices with insufficient rows were padded with 1s to obtain the training set data.

[0030] A gridded model of the entire working face area was established, using support numbers and coal cutting cycles as coordinate units. Support numbers were only selected as survey lines. A relative coordinate system was established based on the gridded model, with the origin and coordinate system defined as follows: See axis distribution Figure 3 As shown, the selected origin coordinates in the figure are (551153.4489, 4432420.5453, 791), rotated clockwise by 197° relative to the absolute coordinate system. Using the described method, the microseismic events are transformed to a relative coordinate system, and a top-down view of the microseismic event distribution is shown below. Figure 4 See the front view of the microseismic event distribution. Figure 5 .

[0031] The spatiotemporal feature matrix is ​​constructed using cycles as units. The construction process of the spatiotemporal feature matrix of stent resistance is as follows: the final resistance of the stent cycle corresponding to the grid is taken as the grid feature value. The 25 grid feature values ​​centered on the study grid constitute the spatiotemporal feature matrix of stent resistance for the study grid (if a grid feature value is missing, it is filled by the average value of other grids in the spatiotemporal feature matrix). Figure 6 As shown in the figure. The spatiotemporal characteristic matrices of the support resistance from all cyclic grids are stacked to form a cyclic support resistance spatiotemporal characteristic matrix. All grid eigenvalues ​​constitute a large matrix that reflects the distribution of the collected support resistance data, and a heatmap is plotted as shown. Figure 7 As shown.

[0032] The process of constructing the spatiotemporal characteristic matrix of microseismic events is as follows: Figure 8 As shown: By statistically analyzing the microseismic data contained in all loops, a uniform row count of 160 was determined. Each row represents one of the five feature values ​​of the microseismic data, and a microseismic spatiotemporal feature matrix was constructed sequentially according to time. Insufficient rows were padded with 1s after normalization. The microseismic spatiotemporal feature matrix was then reshaped to obtain the following shape: The final matrix.

[0033] The microseismic and support resistance data are aligned by processing the microseismic spatiotemporal feature matrix. The collected data includes a file containing the cycle number, start and end times, and mining location corresponding to each mining cycle. The temporal characteristics of microseismic events in the microseismic spatiotemporal feature matrix are transformed into the difference between the microseismic event time and the corresponding cycle end time in the file, while the energy characteristics remain unchanged. Using the cycle number and mining location information in the file, the position of the center grid of each cycle in the relative coordinate system is calculated. The coordinates are obtained by multiplying (cycle number - 0.5) by the cutting depth of the coal mining machine. The coordinates are obtained by multiplying ((number of survey lines - 1) / 2 + 0.5) by the width of the support frame. The coordinate is 0. Subtract the coordinate values ​​of the cyclic center grid in the relative coordinate system from the three-dimensional spatial coordinate features of the microseismic event to obtain the aligned microseismic spatiotemporal feature matrix.

[0034] The transformed feature values ​​of all microseismic events in the training set were statistically analyzed, and the maximum and minimum values ​​of the five features were obtained. These maximum and minimum values ​​are the normalization parameters, as shown in Table 1. The microseismic data were normalized using the normalization parameters, with rows less than 160 rows padded with 1s. 504 cycle samples were randomly selected from 630 cycles to obtain the training set.

[0035] Table 1

[0036] Furthermore, in S3, building a conditional image translation model and training it using the training set data includes: The spatiotemporal feature matrices of support resistance and microseismic spatiotemporal feature matrices are treated as images. An image translation model based on a conditional generative adversarial network architecture is constructed. The generator takes the spatiotemporal feature matrix of support resistance as input and generates the spatiotemporal feature matrix of microseismic motion. The discriminator distinguishes the generated spatiotemporal feature matrix of microseismic motion from the real spatiotemporal feature matrix of microseismic motion. The model is trained using training set data. The model parameters are optimized through adversarial training process, and the error index between the generated data and the real data is monitored to determine the optimal training rounds.

[0037] Specifically, the implementation process of this embodiment includes: By treating the support resistance and the spatiotemporal feature matrix of microseismic events as images, a conditional image translation model is constructed that can effectively capture the correlation between support resistance and microseismic data. The spatiotemporal feature matrix of support resistance is set as the input condition to generate the corresponding spatiotemporal feature matrix of microseismic events.

[0038] The model is trained using the training set data. The evaluation metric for the unnormalized data is monitored. The evaluation metric is the MAE value of the generated matrix and the true matrix. The optimal training epochs are then determined.

[0039] This paper improves upon the pix2pix model, which is based on Conditional Generative Adversarial Networks (CGAN), to construct the desired conditional image translation model. The pix2pix model is a GAN-based image translation model. GANs allow for unsupervised output of the desired image simply by specifying the objective and automatically learning a loss function that satisfies that objective. The pix2pix model introduces Conditional Information Networks (CGAN) into the GAN framework, adding precise control and guidance to the image generation process. This ensures that the generated image is not only accurate at the individual pixel level but also consistent and reasonable in its global structure and configuration.

[0040] In the development of CGAN-based image translation, the pix2pix model represents a groundbreaking research achievement. Employing a structured loss function during training, the pix2pix model successfully applied CGAN to supervised image translation tasks for the first time, achieving various image transformations, including semantic / label-to-real-image conversion, grayscale-to-color image conversion, and aeronautical chart-to-map conversion. The pix2pix model frees image translation from the limitations of traditional, task-specific algorithms, allowing for solutions within a general framework.

[0041] The constructed conditional image translation model will be introduced from three aspects: Generative Adversarial Network (GAN), pix2pix model, and improved model.

[0042] (1) Generative Adversarial Network (GAN): GANs are inspired by game theory concepts from zero-sum games. They are designed to learn the distribution of generated data x. The input noise variable is defined. The process of determining the prior distribution and then mapping it to the data space is represented as follows: , where G is a parameter The differentiable function representing the multilayer perceptron is called the generator. A second multilayer perceptron is then defined. It outputs a single scalar. This indicates that x comes from data rather than... The probability of this is called the discriminant. The model D is trained to maximize the probability of assigning the correct label to the training samples and samples from G. Simultaneously, G is trained to minimize... In other words, D and G pass through the objective function. Let's play the following zero-sum game.

[0043] ; When GANs are trained using batch stochastic gradient descent, the number of training steps for the discriminator k is a hyperparameter. The pseudocode for the model training process is shown in Table 2. Table 2

[0044] During the training process described above, the GAN synchronously updates the discriminator, thus enabling it to distinguish the distribution of generated data. and generator to generate distribution ,like Figure 9 As shown. In Figure 9 In the diagram, the lower horizontal line represents the z-domain, which is represented by a uniform distribution. The upper horizontal line represents the x-domain, and the upward arrow indicates the mapping. How to apply a non-uniform distribution to the transformed sample G in High-density areas contract while low-density areas expand. Figure 9 The training of GANs is shown in four stages. In stage (a), the adversarial model is close to convergence. and D is a near-accurate classifier; in stage (b), within the inner loop of the algorithm, D is trained to distinguish data samples, converging to... Stage (c) involves updating G, where the gradient of D guides G(z) towards regions more likely to be classified as generating data; Stage (d) occurs after multiple training iterations, where G and D reach a Nash equilibrium point, at which point neither can improve further because... The discriminator cannot distinguish between the two distributions at this time. .

[0045] (2) pix2pix model: The above describes a generative adversarial model without user constraints, which learns the mapping from a random noise vector z to the output image y. However, the image-to-image translation problem requires the input image to constrain the generated result. CGAN learns the mapping from the input image x and the random noise vector z to y. It can achieve supervised image generation based on input images. The pix2pix model uses a structured loss function during training and is the first model to successfully apply CGAN to a supervised image translation task.

[0046] Taking the task of generating a complete photo from edge images as an example, the training process of the pix2pix model is as follows: Figure 10 As shown, the generator G in the figure is trained to produce outputs that are indistinguishable from real images by the adversarially trained discriminator D. The discriminator D then tries to detect fake images created by the generator G as well as possible. Unlike unconditional GANs, both the generator and discriminator in the pix2pix model observe the input edge image.

[0047] The objective function of CGAN is modified from that of GAN, and its objective function is as follows: .

[0048] Previous research has found that combining GANs with traditional losses such as L2 distance is beneficial. Therefore, the discriminator in the pix2pix model remains unchanged from the GAN, while the generator's task is not only to fool the discriminator but also to approximate the true output in an L2 sense. The final pix2pix model chose L1 distance because experiments showed that using L1 distance reduces ambiguity.

[0049] ; Therefore, the final objective function is: .

[0050] When designing generator architectures, many previous solutions used, for example... Figure 11 The encoder-decoder structure is shown. In such a network, the input is progressively downsampled through a series of layers until a bottleneck layer, and then the process is reversed. Such a network requires all information to flow through all layers, including the bottleneck layer. For many image transformation problems, there is a significant amount of shared low-order information between the input and output, so it is desirable to pass this information directly within the network. To provide the generator with a way to bypass the bottleneck layer and directly pass this information, the pix2pix model adds skip connections, following the overall structure of the "U-Net," as shown below. Figure 11 As shown. Specifically, skip connections are added between each layer i and layer ni, where n is the total number of layers. Each skip connection simply concatenates all channels of layer i with channels of layer ni.

[0051] Existing research has shown that L1 loss can produce a blurring effect in image generation. While these losses do not improve high-frequency sharpness, they can still accurately capture low-frequency information in many cases. Therefore, the discriminator network does not need a completely new framework to judge the correctness of low-frequency information; the L1 loss in the objective function is sufficient. This allows the discriminator to model only high-frequency information. To this end, the pix2pix model divides the image into multiple... The discriminator runs through convolutions on each patch, and the final output of D is obtained by averaging the discrimination results of all patches. Such a discriminator effectively models the image as a Markov random field, assuming that pixels spaced more than one patch diameter are independent—a common assumption in texture and style models. Therefore, the discriminator in a pix2pix model can be understood as a form of texture or style loss.

[0052] (3) Improved model: The microseismic prediction model based on support drag generally adopts a pix2pix model structure. Considering the differences in input and output image sizes, the structures of the generator and discriminator were adjusted. For the microseismic prediction problem based on support drag, the model input is the support drag feature matrix, with the following shape: The output is a microseismic feature matrix with the following shape: If we consider it as an image, then the image's length and width are small enough that the discriminator cannot divide it into smaller patches, and the generator cannot perform further downsampling. Therefore, the model's discriminator directly judges the generated image without further patching. The generator, on the other hand, adopts a strategy of upsampling followed by downsampling, and its model structure is as follows: Figure 12 As shown.

[0053] Using cycles as sample units, the spatiotemporal feature matrix of support resistance as input conditions, and the spatiotemporal feature matrix of microseismic events as output, the model was trained using training set data. During training, both the discrimination error of the discriminator and the generation error of the generator were monitored, as follows: Regarding the discriminator's discrimination error: the discriminator needs to be able to classify real data as 1 and generated data as 0. Therefore, the monitored real data discrimination error is the MAE value of the discriminator output with 1, and the MAE value of the generated data with 0. Its variation curve with training rounds is shown below. Figure 13 As shown, the generator's discrimination error reaches its minimum at the 2525th training epoch, with an average error of 0.00099, indicating a strong discrimination capability. In fact, by the 300th training epoch, the discrimination error had decreased from 0.08323 to 0.00642, demonstrating that the adversarial process of the model is primarily manifested in the first 300 epochs, with no significant trend change after that. Therefore, the discrimination error variation curve for the first 300 epochs was plotted, yielding... Figure 13 .from Figure 13 As can be seen, the discrimination error initially increases with each training epoch, then gradually decreases after the 18th epoch. This is because in the first 18 epochs, the images generated by the generator become increasingly similar to real images, while the discriminator's learning ability fails to keep pace. After the 18th epoch, the discriminator develops stronger discrimination capabilities; although the generator produces increasingly realistic images, the discriminator's judgments become more accurate. Throughout the training process, the discrimination error of the generated data is consistently higher than that of the real data, indicating that the discriminator has a stronger ability to distinguish real data. When the discrimination error tends to remain constant, the discrimination errors of the real data and the generated data are nearly equal, meaning the discriminator has the same ability to distinguish between the two types of data.

[0054] For generator generation error: the monitored content is the difference in MAE between the un-inverse-normalized generated data and the real data. Samples not selected for the training set from the 630 cycle samples are considered as being mined cycles to verify the application's effectiveness. During training, after each training cycle, not only is the training set error monitored, but also these validation samples (test set) are monitored. The curves showing the change of the two types of data generation errors with training cycles are as follows: Figure 14 As shown, from Figure 14 As can be seen, the errors in both the training and test sets significantly decreased with increasing training rounds, achieving the desired training effect. This demonstrates a close correlation between the constructed support resistance input features and the microseismic output features, proving the feasibility of using support resistance to invert microseismic data. During training, the errors in the test and training sets remained relatively similar, indicating minimal overfitting and good generalization performance. The test set error reached its minimum after the 2775th training round, representing the optimal training result for the support resistance-based microseismic data inversion model. At this point, the training set error was 0.00261, and the test set error was 0.00239.

[0055] Furthermore, in S4, applying the optimal model to the ongoing mining cycle includes: A spatiotemporal feature matrix of support resistance during the mining cycle is constructed. The spatiotemporal feature matrix of support resistance is input into the optimal model, and the generated spatiotemporal feature matrix of microseismic activity is output. The generated spatiotemporal feature matrix of microseismic activity is inversely normalized using the normalization parameters used during model training to obtain inverted microseismic data.

[0056] Specifically, the implementation process of this embodiment includes: The spatiotemporal characteristic matrix of the support resistance during the mining cycle is constructed and input into the model. Using the normalization parameters from model training, the output microseismic spatiotemporal characteristic matrix is ​​inversely normalized to obtain the required real-time inversion microseismic data.

[0057] When applying the model with the optimal training rounds to an ongoing mining cycle, the spatiotemporal characteristic matrix of the support resistance for that cycle is first constructed. After being input into the model, the spatiotemporal characteristic matrix of microseismic events is output. Using the normalization parameters shown in Table 1, the microseismic events are inversely normalized to obtain the final microseismic data to be inverted.

[0058] The loop samples not included in the training set from 630 loops were taken as the loops currently being mined, and the results were evaluated under the optimal training rounds: the prediction time error was a minimum of 0.67 h and a maximum of 0.95 h; the x-coordinate error was a minimum of 4.23 m and a maximum of 5.46 m; the y-coordinate error was a minimum of 1.71 m and a maximum of 2.68 m; the z-coordinate error was a minimum of 1.51 m and a maximum of 1.92 m; and the energy error was a minimum of 150.33 J and a maximum of 192.71 J. The errors of all five features are within acceptable ranges. This invention can achieve low-cost acquisition of satisfactory microseismic data.

[0059] This invention discloses a method for real-time inversion of microseismic data based on support resistance technology. By combining support resistance data with conditional image translation technology, it opens up a new path for acquiring microseismic data. This method effectively overcomes the accuracy limitations of traditional microseismic monitoring technologies caused by factors such as complex environments, heterogeneous media, and equipment layout during signal reception and algorithm localization, and can directly generate microseismic data that meets application requirements. Simultaneously, this method significantly reduces reliance on expensive dedicated monitoring equipment, reduces manpower investment in equipment installation, maintenance, and signal processing, and substantially lowers the overall cost and application threshold of microseismic monitoring. This provides feasibility for the widespread application of microseismic monitoring technology in all working faces, thereby providing continuous and reliable data support for more accurate rock movement analysis and roof disaster prevention.

[0060] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for real-time inversion of microseismic data based on conditional image translation technology, characterized in that, The method comprises the following steps: S1, collecting support resistance data and microseismic data of historical mining cycles in real time, and determining the time range of microseismic data inversion; S2, establishing a working face relative coordinate system based on the support resistance data and the microseismic data, converting the microseismic events to the relative coordinate system, and respectively constructing a support resistance space-time feature matrix and a microseismic space-time feature matrix, aligning the two to establish a model training set; S3, regarding the support resistance space-time feature matrix and the microseismic space-time feature matrix as images, constructing a conditional image translation model, training using the training set data, and determining the optimal model based on evaluation indexes; S4, applying the optimal model to the mining cycle being mined, inputting the real-time support resistance space-time feature matrix, outputting the generated microseismic space-time feature matrix, and obtaining the inverted microseismic data through inverse normalization processing.

2. The method of real-time inversion of microseismic data using bracing resistance based on conditional image translation technique according to claim 1, wherein, The collection of support resistance data and microseismic data of historical cycles comprises: The cycle number, start and end time and location of the historical cycle are collected, complete support resistance data is collected, wherein every 8 to 12 supports are set with a measuring line, only the data of the measuring line in the cycle is collected, and the cycle measuring line data collected includes complete support resistance data of all measuring lines in the cycle and support resistance data of two cycles before and after the cycle, complete microseismic data is collected, only microseismic event data above the coal seam floor is collected, and the microseismic event saves complete time, three-dimensional space coordinates and energy characteristics.

3. The method of claim 1, wherein, In S2, the establishment of the working face relative coordinate system comprises: A full working face area gridding model is established with support number and coal cutting cycle as coordinate units, wherein the support number only selects the support set as the measuring line, the position with the support number 1 in the coal seam floor and the mining cycle 1 as the origin, the working face advancing direction as the positive direction of the x axis, the support number increasing direction as the positive direction of the y axis, and the vertical upward direction as the positive direction of the z axis to establish the relative coordinate system, the units of the relative coordinate system are all meters, the three-dimensional space coordinates of the microseismic event are converted to the relative coordinate system, wherein the z coordinate of the microseismic event is converted to the difference between the original elevation and the elevation of the origin of the relative coordinate system, and the x and y coordinates are converted by plane rotation transformation.

4. The method of claim 1, wherein, In S2, the construction of the support resistance space-time feature matrix comprises: The support resistance space-time feature matrix is constructed in units of cycles, for each grid, the cycle-end resistance values of 25 grids centered on the grid are obtained to form the grid space-time feature matrix of the grid, if the cycle-end resistance value of the grid is missing or abnormal, the average value of other grids in the grid space-time feature matrix is used for filling, and all grid space-time feature matrices in the cycle are stacked to construct the support resistance space-time feature matrix.

5. The method of claim 1, wherein, In S2, the construction of the microseismic space-time feature matrix comprises: The microseismic space-time feature matrix is constructed in cycles, which is composed of all microseismic events in the time range of the cycle to be inverted for microseismic data. The data of each row of the matrix is five features of a microseismic event, including time, three-dimensional spatial coordinates and energy features. According to the cycle containing the most number of microseismic events, a uniform number of rows is set. For the microseismic space-time feature matrix that is less than the uniform number of rows, 1 is filled after normalization. The filled microseismic space-time feature matrix is converted into a three-dimensional shape of 5 rows and 5 columns multiplied by the uniform number of rows divided by 5.

6. The method of real-time inversion of microseismic data using bracing resistance based on conditional image translation technique as claimed in claim 1, wherein, In S2, aligning the support resistance and the microseismic space-time feature matrix includes: Processing the microseismic space-time feature matrix, converting the time feature of the microseismic event into the difference between the cycle end time and the original time feature, keeping the energy feature unchanged, calculating the coordinates of the cycle center grid in the relative coordinate system, wherein the x coordinate is obtained by multiplying the cycle number minus 0.5 by the depth of the coal mining machine, the y coordinate is obtained by multiplying the support width by 0.5 after dividing the measuring line number minus 1 by 2 and adding 0.5, and the z coordinate is 0. Subtract the coordinate values of the cycle center grid from the three-dimensional spatial coordinate features of the microseismic event respectively. The maximum and minimum values of each feature are obtained as normalization parameters by counting the converted feature values of all microseismic events in the training set data. The microseismic data is normalized, and the microseismic space-time feature matrix that is less than the uniform number of rows is filled with 1 to obtain the aligned training set data.

7. The method of real-time inversion of microseismic data using bracing resistance based on conditional image translation technique as claimed in claim 1, wherein, In S3, constructing a conditional image translation model and training it using the training set data includes: The support resistance space-time feature matrix and the microseismic space-time feature matrix are regarded as images, and an image translation model based on the conditional generative adversarial network architecture is constructed, wherein the generator takes the support resistance space-time feature matrix as the input condition to generate the microseismic space-time feature matrix, the discriminator distinguishes between the generated microseismic space-time feature matrix and the real microseismic space-time feature matrix, the training set data is used to train the model, the model parameters are optimized through the adversarial training process, and the error index of the generated data and the real data is monitored to determine the optimal training epoch.

8. The method of real-time inversion of microseismic data using bracing resistance based on conditional image translation technique as claimed in claim 1, wherein, In S4, applying the optimal model to the cycle being mined includes: The support resistance space-time feature matrix of the cycle being mined is constructed, the support resistance space-time feature matrix is input into the optimal model, the generated microseismic space-time feature matrix is output, and the inverse normalization processing is performed on the generated microseismic space-time feature matrix using the normalization parameters used during model training to obtain the inverted microseismic data.

Images