Rock mass crustal stress field inversion method and system based on machine learning, and storage medium

Abstract

The invention relates to the technical field of machine learning, in particular to a rock mass crustal stress field inversion method and system based on machine learning and a storage medium. The method comprises the following steps: S1, constructing a three-dimensional virtual model of a target rock mass, and obtaining geological data at each position in an inversion target area; s2, acquiring a geological type in the inversion area, and acquiring rock mass parameters of a rock mass corresponding to a sampling point; s3, a plurality of sets of test data are obtained through a hydraulic fracturing method and a drilling method and preprocessed, and learning data and detection data are obtained; s4, a rock mass stress prediction model is created, the rock mass stress prediction model comprises m sub-models, the learning data are used for training the sub-models, and the sub-models are verified based on the detection data. According to the invention, the problem of inaccurate inversion of the rock mass stress field is solved, and the inversion precision of the rock mass crustal stress field is improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and more specifically to a rock mass in-situ stress field inversion method, system and storage medium based on machine learning. Background Art

[0002] Rock stress inversion is an important research topic in the field of geotechnical engineering. It analyzes the stress response of rock mass under specific conditions and infers the stress state or mechanical parameters of rock mass. This technology plays a key role in the design, construction and stability assessment of underground engineering. With the increasing demand for underground space development, the development of rock stress inversion technology is particularly important. A similar prior art is a Chinese patent with publication number CN116861704B, which proposes a method for dynamic and rapid prediction of large deformation levels of high geostress soft rock tunnels, involving the field of tunnel large deformation prediction, including: first, constructing a three-dimensional geological model based on the acquired geological data; then performing geostress inversion analysis based on the three-dimensional geological model and the measured geostress results to determine the magnitude and direction of geostress within the tunnel axis; then conducting a point load strength test on the test rock sample to calculate the uniaxial compressive strength of the rock in the section to be predicted; then estimating the rock compressive strength based on the uniaxial compressive strength of the rock in the section to be predicted; finally, determining the large deformation level of the tunnel based on the maximum geostress in the geostress inversion analysis results and the estimated rock compressive strength, and formulating corresponding construction methods and support parameters based on the determination results; the geostress characteristics and surrounding rock properties of different prediction sections can be considered, and dynamic and rapid prediction of large deformation of high geostress soft rock tunnels can be achieved. In addition, similar prior art includes a U.S. patent with publication number US11536864B1, which discloses a method for establishing the distribution of geostress field on a slope in a canyon area, including: obtaining the duration of the fracture surface according to the length of the structural surface trace of the fracture surface and the length of the rock bridge, and then obtaining the fracture stage of the crack according to the fracture stage of the fracture surface. According to the progressive failure characteristics of the rock mass, combined with the characteristics of the fracture surface, the magnitude and direction of the maximum principal stress are obtained to establish the distribution of the geostress field. This method is simple to operate, does not require geostress testing, does not require a large amount of manpower and material resources, does not require unnecessary capital investment, and can simply and effectively obtain geostress field data. Both of the above documents solve the problem of geostress field inversion, but the accuracy is low and the problem of rock mass stress distribution cannot be accurately inverted. Summary of the invention

[0003] In order to better solve the above problems, the present invention provides a rock mass in-situ stress field inversion method based on machine learning, comprising the following steps:

[0004] Step S1: scanning and imaging the target rock mass and acquiring a first geological image, filtering and interpolating the first geological image to acquire a second geological image, and constructing a three-dimensional virtual model of the target rock mass based on the second geological image; acquiring geological data at each position in the inversion area based on the three-dimensional virtual model, wherein the geological data includes measurement data of faults, fractures and terrain trends;

[0005] Step S2: Based on the three-dimensional virtual model, the geological type in the inversion area is obtained, and a number of sampling points are selected within the area corresponding to each geological type, and the rock mass parameters corresponding to the sampling point are obtained by sampling the rock mass at each sampling point and performing in-situ compression, and the rock mass parameters are marked in the three-dimensional virtual model;

[0006] Step S3: obtaining a plurality of test data sets respectively by hydraulic fracturing and drilling according to the sub-region where each sampling point is located, preprocessing the test data, and generating learning data and detection data from the preprocessed test data, the rock mass parameters of the sampling points corresponding to the test data, and the geological data;

[0007] Step S4: Create a rock stress prediction model, the rock stress prediction model includes m sub-models, the m sub-models are created based on the geological type, the sub-models are trained using the learning data, and the sub-models are verified based on the detection data, wherein each of the sub-models corresponds to a geological type.

[0008] As a preferred technical solution of the present invention, in step S1, filtering and interpolating the first geological image to obtain the second geological image includes:

[0009] Step S11: abnormal pixel detection is performed on pixels in the first geological image, and a difference between the brightness value of each pixel and a first value is obtained. When the difference is greater than a set difference, the pixel is an abnormal pixel, and when the difference is less than or equal to the set difference, the pixel is a normal pixel, wherein the first value is the average of the maximum and minimum brightness values ​​of the pixels in the first geological image;

[0010] Step S12: With the abnormal pixel as the center, within a set range, obtain the number of other abnormal pixels, and based on the number, obtain the normal pixels on a circle with a set radius around the abnormal pixel with the abnormal pixel as the center, and average the brightness values ​​of the normal pixels on the circle to obtain an interpolation value, wherein the larger the number, the smaller the set radius, and vice versa, the larger the radius, the number of abnormal pixels in the set range is less than the total number of pixels in the set range;

[0011] Step S13: replacing the brightness value of each abnormal pixel with the interpolation value corresponding to the abnormal pixel.

[0012] As a preferred technical solution of the present invention, step S2 includes: obtaining the geological type of the rock mass in the three-dimensional virtual model, and selecting a landmark sampling area within the area corresponding to each geological type, selecting multiple sampling points at different positions in the sampling area, sampling the rock mass at each sampling point and performing in-situ compression to obtain the rock mass parameters corresponding to the sampling point, wherein the rock mass parameters include Young's modulus E, Poisson's ratio v, strain hardening modulus H, peak strain μ, and post-peak drop modulus M of the rock mass, and the geological types include at least folds, cracks, and faults.

[0013] As a preferred technical solution of the present invention, step S3 includes the following steps:

[0014] Step S31: based on the three-dimensional virtual model, the sub-region where the sampling point in step S2 is located is obtained, and N positions are uniformly selected in the sub-region as data collection positions, where the value of N is a positive integer greater than or equal to 10;

[0015] Step S32: obtaining a plurality of groups of test data by the hydraulic fracturing method and the drilling method respectively, and marking the plurality of groups of test data at the data acquisition positions in the three-dimensional virtual model;

[0016] Step S33: According to the geological data and rock mass parameters at the data collection position of the test data and the geological data and rock mass parameters at the data collection positions of other surrounding test data, obtain the size relationship between the test data and other surrounding test data, and judge whether the test data needs to be corrected based on the size relationship. When the test data satisfies the size relationship, the test data does not need to be corrected. When the test data does not satisfy the size relationship, the test data needs to be corrected, and weight the test data based on the four closest distances to the data collection position of the test data, namely, the front, back, left, and right, and use the weighted values ​​as the correction data of the test data, wherein the weights of the four test data are set by the distance from the test data, and the closer the distance, the higher the weight, and vice versa.

[0017] As a preferred technical solution of the present invention, the step S33 further includes a step S34:

[0018] Based on the three-dimensional virtual model, the test data, the geological data at the data collection location of the test data and the rock parameters of the sub-area where the test data is located are combined into training data, and any three consecutive training data belonging to the corresponding area of ​​any geological type are selected as detection data, and the remaining two training data are used as learning data.

[0019] As a preferred technical solution of the present invention, step S4 includes:

[0020] Create the rock stress prediction model, obtain m geological types in the inversion area through the geological data, and create m sub-models based on the geological types, use the learning data in the sub-area corresponding to each geological type in the three-dimensional virtual model as the learning data of the corresponding sub-model, and train the sub-model through the learning data, wherein the value of m is a positive integer greater than or equal to 3.

[0021] As a preferred technical solution of the present invention, step S4 also includes: inputting the geological data and rock parameters in the detection data into the rock stress prediction model, and obtaining the geological type to which the detection data belongs according to the position of the detection data in the three-dimensional virtual model, inputting the corresponding sub-model according to the geological type, obtaining the prediction result, comparing the prediction result with the stress value in the detection data, and obtaining a comparison difference. When the comparison difference is less than or equal to a preset difference, the prediction result is accurate. When the comparison difference is greater than the preset difference and the prediction result satisfies the size relationship between the stress value in the detection data and the stress value in the closest learning data in the sub-region, the sub-model is retrained by increasing the number of learning data of the geological type. When the comparison difference is greater than the preset difference and the prediction result does not satisfy the size relationship between the detection data and the closest learning data in the sub-region, step S2 is repeated to re-obtain the rock parameters corresponding to the geological type and retrain the sub-model.

[0022] As a preferred technical solution of the present invention, the test data is the value of the geostress field at the data collection location.

[0023] The present invention also provides a rock mass in-situ stress field inversion system based on machine learning, the system is used to implement the above method, and the system comprises:

[0024] A virtual model generation unit is used to obtain a first geological image by scanning and imaging the target rock mass, filter and interpolate the first geological image to obtain a second geological image, and construct a three-dimensional virtual model of the target rock mass based on the second geological image; and obtain geological data at each position in the inversion area based on the three-dimensional virtual model, wherein the geological data includes measurement data of faults, fractures and terrain trends;

[0025] A marking unit, used to obtain the geological type in the inversion area based on the three-dimensional virtual model, select a number of sampling points within the area corresponding to each geological type, and obtain rock mass parameters corresponding to the sampling point by sampling and in-situ compressing the rock mass at each sampling point, and mark the rock mass parameters in the three-dimensional virtual model;

[0026] A data generating unit, configured to obtain a plurality of test data respectively through hydraulic fracturing and drilling according to the sub-region where each sampling point is located, pre-process the test data, and generate learning data and detection data from the pre-processed test data, the rock mass parameters of the sampling points corresponding to the test data, and the geological data;

[0027] A model training unit is used to create a rock stress prediction model, wherein the rock stress prediction model includes m sub-models, wherein the m sub-models are created based on the geological type, the sub-models are trained using the learning data, and the sub-models are verified based on the detection data, wherein each of the sub-models corresponds to a geological type.

[0028] The present invention also provides a computer storage medium, wherein the storage medium stores program instructions, wherein when the program instructions are executed, the device where the storage medium is located is controlled to execute the above method.

[0029] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0030] The present invention obtains a first geological image of a target rock mass through a rock mass scanning technology, performs filtering and denoising on the first geological image, obtains abnormal pixels in the first geological image, obtains the number of other abnormal pixels within a set range, and performs filtering and interpolation on the abnormal pixels according to the number of abnormal pixels within the set range, obtains a second geological image, and obtains accurate geological data based on the second geological image. According to the geological type of the rock mass in the three-dimensional virtual model, a landmark sampling area is selected within the area corresponding to each geological type, and multiple sampling points are selected at different positions in the sampling area, and rock mass parameters of each sampling point are obtained, so as to obtain relatively rich and comprehensive sampling rock mass parameters of each geological type. Based on the three-dimensional virtual model, each sampling point in step S2 The sub-region where the point is located, and N data collection positions are uniformly selected in the above sub-region, and several groups of test data are obtained by hydraulic fracturing and drilling, and the above test data are marked in the above three-dimensional virtual model. Since the sampling point and the above data collection position are in the same sub-region, the rock mass parameters of the above sampling point and the above test data are combined into learning data and detection data to improve the accuracy of the rock stress prediction model. The stress magnitude relationship at two relative positions in the above three-dimensional virtual model can be judged by the stress distribution law, and the above magnitude relationship can be used to judge whether the above test data needs to be corrected. When correction is required, the four test data closest to the front, back, left and right of the above test data are weighted to obtain more accurate stress field data, laying the foundation for further improving the accuracy of the rock stress prediction model. Create multiple sub-models, and fully train the learning data corresponding to each of the above address types through the above sub-models, so as to obtain the stress prediction value corresponding to each geological type. Through the mutual cooperation of the above technical solutions, the accuracy of the inversion of the ground stress field can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of the rock mass in-situ stress field inversion method based on machine learning of the present invention;

[0032] Figure 2 This is a structural diagram of the rock mass in-situ stress field inversion system based on machine learning of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0034] The present invention provides a rock mass in-situ stress field inversion method based on machine learning, such as Figure 1 As shown, the following steps are included:

[0035] Step S1: scanning and imaging the target rock mass and acquiring a first geological image, filtering and interpolating the first geological image to acquire a second geological image, and constructing a three-dimensional virtual model of the target rock mass based on the second geological image; acquiring geological data at each position in the inversion target area based on the three-dimensional virtual model, wherein the geological data includes measurement data of faults, fractures and terrain trends;

[0036] Specifically, a first geological image of the target rock mass is obtained through rock mass scanning technology. Since noise interference is inevitably introduced during the scanning process, noise points, i.e., abnormal pixels, will appear. In order to obtain more accurate geological data, it is necessary to filter and denoise the first geological image to obtain the brightness value of each pixel in the first geological image, i.e., the difference between the grayscale value and the first value. The first value is the average of the maximum and minimum values ​​of the pixel brightness. Based on the difference, the abnormal pixels and normal pixels in the first geological image are accurately identified. In addition, in the first geological image, the structure and geological characteristics of the rock layer are expressed by the brightness value of the pixel, and due to the continuity between the pixels, therefore, within the set range, the normal pixels are obtained. The number of other abnormal pixels mentioned above is calculated, and the abnormal pixels are filtered and interpolated according to the number of abnormal pixels within the above-mentioned set range. The abnormal pixels are taken as the center of the circle, and the average value of the normal pixels on the circle with the set radius is taken as the brightness value of the abnormal pixels. The more the number of abnormal pixels is, the more accurately the brightness value of the abnormal pixels needs to be interpolated. Conversely, when the number of abnormal pixels is small, it needs to be consistent with the brightness value in the current area. Therefore, the more the number of abnormal pixels is, the larger the corresponding set radius is. Through the above-mentioned technical solution, a more accurate interpolation of the abnormal pixels can be obtained, thereby obtaining a more accurate three-dimensional virtual model, laying a foundation for further obtaining accurate geological data.

[0037] Step S2: Based on the three-dimensional virtual model, the geological type in the inversion area is obtained, and a number of sampling points are selected within the area corresponding to each geological type, and the rock mass parameters corresponding to the sampling point are obtained by sampling the rock mass at each sampling point and performing in-situ compression, and the rock mass parameters are marked in the three-dimensional virtual model;

[0038] Specifically, according to the geological type of the rock mass in the three-dimensional virtual model, a landmark sampling area is selected within the area corresponding to each geological type, that is, different areas with landmark characteristics are selected as sampling areas, for example: different positions at equal distances and positions at different distances on both sides of the fold area are used as sampling areas, and multiple sampling points are selected at different positions in the sampling area, and the rock mass parameters of each of the sampling points are obtained. Through the technical solution, relatively rich and comprehensive sampling rock mass parameters of each geological type can be obtained, laying a foundation for further obtaining accurate inversion results of the geostress field of the target rock mass.

[0039] Step S3: obtaining a plurality of test data sets respectively by hydraulic fracturing and drilling according to the sub-region where each sampling point is located, preprocessing the test data, and generating learning data and detection data from the preprocessed test data, the rock mass parameters of the sampling points corresponding to the test data, and the geological data;

[0040] Specifically, based on the sub-region where each sampling point in step S2 is located in the above-mentioned three-dimensional virtual model, N data acquisition positions are uniformly selected in the above-mentioned sub-region, and several groups of test data are obtained through hydraulic fracturing and drilling, and the above-mentioned test data are the stress at the above-mentioned data acquisition position, and the above-mentioned test data are annotated in the above-mentioned three-dimensional virtual model. Since the sampling point and the above-mentioned data acquisition position are in the same sub-region, the rock mass parameters of the above-mentioned sampling point and the above-mentioned test data are combined into learning data and detection data to improve the accuracy of the rock mass stress prediction model, and then a more accurate inversion of the ground stress field is obtained. The two relative stress distribution laws in the above-mentioned three-dimensional virtual model can be judged. The stress magnitude relationship at the position, and judge whether the above test data needs to be corrected by the above magnitude relationship. When the above test data meets the above magnitude relationship, the above test data does not need to be corrected. When the above test data does not meet the above magnitude relationship, the test data needs to be corrected, and the four test data closest to the above test data, namely, the front, back, left and right, are weighted, wherein the sum of the weights of the left and right test data and the sum of the weights of the front and back test data are both 0.5, and the closer the distance to the above test data, the higher the weight, and vice versa. Through the above technical solution, more accurate stress field data can be obtained, laying a foundation for further improving the accuracy of the rock stress prediction model.

[0041] Step S4: Create a rock stress prediction model, the rock stress prediction model includes m sub-models, the m sub-models are created based on the geological type, the sub-models are trained using the learning data, and the sub-models are verified based on the detection data, wherein each of the sub-models corresponds to a geological type.

[0042] Specifically, multiple sub-models are created, and the learning data corresponding to each of the above-mentioned address types are fully trained through the above-mentioned sub-models, so that the stress prediction value corresponding to each geological type can be obtained, thereby improving the accuracy of stress inversion, and the input data in the detection data of the corresponding sub-model is input into the above-mentioned rock stress prediction model, and the above-mentioned input data is the rock parameters and geological data in the detection data, and the prediction result, i.e., the predicted stress, is obtained, and the above-mentioned predicted stress is compared with the corresponding stress value in the detection data, and the comparison difference is obtained, and the above-mentioned rock stress prediction model is corrected according to the above-mentioned comparison difference. When the comparison difference is less than or equal to the preset difference, it is considered that the prediction result is relatively accurate and does not need to be corrected. When the comparison difference is greater than the preset difference, and the prediction result satisfies the detection data and the sub-area When the size relationship between the detection data and the nearest learning data is obtained, it is considered that the prediction result is not accurate enough, but the error is relatively small, which may be caused by the insufficient number of learning data. The accuracy of the corresponding sub-model is improved by increasing the number of learning data. When the comparison difference is greater than the preset difference, and the prediction result does not satisfy the size relationship between the detection data and the nearest learning data in the sub-area, it means that the error is relatively large and the accuracy of the sub-model is poor. It is possible that the rock parameters in the learning data are not accurate enough. By repeating the above step S2 to re-acquire the above rock parameters, re-acquire the above learning data and detection data, and re-train the above sub-model, the accuracy of the above sub-model is improved. Through the above technical solution, a rock stress model with higher accuracy can be obtained, thereby obtaining an accurate inversion rock stress field.

[0043] Furthermore, in step S1, filtering and interpolating the first geological image to obtain the second geological image includes:

[0044] Step S11: abnormal pixel detection is performed on pixels in a set area in the middle of the first geological image, and a difference between the brightness value of each pixel and a first value is obtained. When the difference is greater than the set difference, the pixel is an abnormal pixel, and when the difference is less than or equal to the set difference, the pixel is a normal pixel, wherein the first value is the average of the maximum and minimum values ​​of the pixel brightness in the first geological image;

[0045] Specifically, when scanning and imaging the target rock mass, noise will inevitably be introduced, so noise points, i.e., abnormal pixels, will appear. In order to obtain more accurate geological data, it is necessary to filter and denoise the first geological image. First, abnormal pixel detection is performed. Since when the brightness of the pixel is maximum, it is all white, i.e., the brightest, and when the brightness of the pixel is minimum, it is all black, neither of which can represent the specific image content. The pixel brightness that can represent the image content should be located between the maximum and minimum brightness values. Therefore, the brightness value of each pixel in the first geological image, i.e., the difference between the grayscale value and the first value, is obtained. The first value is the average of the maximum and minimum values ​​of the pixel brightness. When the difference is less than or equal to the set difference, i.e., close to the first value, the pixel is a normal pixel. When the difference is greater than the set difference, the pixel is an abnormal pixel. Through the above technical solution, abnormal pixels and normal pixels in the first geological image can be accurately identified.

[0046] Step S12: With the abnormal pixel as the center, within a set range, obtain the number of other abnormal pixels, and based on the number, obtain the normal pixels on a circle with a set radius around the abnormal pixel with the abnormal pixel as the center, and average the gray values ​​of the normal pixels on the circle to obtain an interpolation value, wherein the more the number is, the smaller the set radius is, and vice versa, the larger the radius is, the number of abnormal pixels within the set range is less than the total number of pixels within the set range;

[0047] Step S13: replacing the brightness value of each abnormal pixel with the interpolation value corresponding to the abnormal pixel.

[0048] Specifically, since the structure and geological characteristics of the rock layer are expressed by the brightness value of the pixel in the above-mentioned first geological image, and due to the continuity between the pixels, the number of other abnormal pixels mentioned above is obtained within the set range, wherein the larger the number is, the greater the density of the abnormal pixels within the above-mentioned set range is, and vice versa, the smaller the density is, and the abnormal pixels are filtered and interpolated according to the number of abnormal pixels within the above-mentioned set range, with the abnormal pixel as the center of the circle and the average value of the normal pixels on the circle with the set radius as the brightness value of the abnormal pixel, the larger the number of abnormal pixels is, the more accurately the brightness value of the abnormal pixel needs to be interpolated, and vice versa, when the number of abnormal pixels is small, it needs to be consistent with the brightness value in the current area, therefore, the larger the number of abnormal pixels is, the larger the corresponding set radius is, and through the above-mentioned technical solution, a more accurate interpolation of the abnormal pixel points can be obtained, thereby obtaining a more accurate three-dimensional virtual model, laying a foundation for further obtaining accurate geological data.

[0049] Furthermore, the step S2 includes: obtaining the geological type of the rock mass in the three-dimensional virtual model, and selecting a landmark sampling area within the area corresponding to each geological type, selecting multiple sampling points at different positions in the sampling area, sampling the rock mass at each sampling point and performing in-situ compression to obtain the rock mass parameters corresponding to the sampling point, wherein the rock mass parameters include Young's modulus E, Poisson's ratio v, strain hardening modulus H, peak strain μ, and post-peak drop modulus M of the rock mass, and the geological types include at least folds, cracks, and faults.

[0050] Specifically, according to the geological type of the rock mass in the three-dimensional virtual model, a landmark sampling area is selected within the area corresponding to each geological type, that is, different areas with landmark characteristics are selected as sampling areas, for example: different positions at equal distances and positions at different distances on both sides of the fold area are used as sampling areas, and multiple sampling points are selected at different positions in the sampling area, and the rock mass parameters of each of the sampling points are obtained. Through the technical solution, relatively rich and comprehensive sampling rock mass parameters of each geological type can be obtained, laying a foundation for further obtaining accurate inversion results of the geostress field of the target rock mass.

[0051] Furthermore, step S3 includes the following steps:

[0052] Step S31: based on the three-dimensional virtual model, the sub-region where the sampling point in step S2 is located is obtained, and N positions are uniformly selected in the sub-region as data collection positions, where the value of N is a positive integer greater than or equal to 10;

[0053] Step S32: obtaining a plurality of groups of test data by the hydraulic fracturing method and the drilling method respectively, and marking the plurality of groups of test data at the data acquisition positions in the three-dimensional virtual model;

[0054] Specifically, based on the sub-region where each sampling point in step S2 is located in the above-mentioned three-dimensional virtual model, N data acquisition positions are evenly selected in the above-mentioned sub-region, and several groups of test data are obtained through hydraulic fracturing and drilling. The above-mentioned test data are the stress at the above-mentioned data acquisition position, and the above-mentioned test data are marked in the above-mentioned three-dimensional virtual model. Since the sampling point and the above-mentioned data acquisition position are in the same sub-region, the rock mass parameters of the above-mentioned sampling point and the above-mentioned test data are combined into learning data and detection data to improve the accuracy of the rock mass stress prediction model, thereby obtaining a more accurate inversion of the ground stress field.

[0055] Step S33: According to the geological data and rock mass parameters at the data collection position of the test data and the geological data and rock mass parameters at the data collection positions of other surrounding test data, obtain the size relationship between the test data and other surrounding test data, and judge whether the test data needs to be corrected based on the size relationship. When the test data satisfies the size relationship, the test data does not need to be corrected. When the test data does not satisfy the size relationship, the test data needs to be corrected, and weight the test data based on the four closest distances to the data collection position of the test data, namely, the front, back, left, and right, and use the weighted values ​​as the correction data of the test data, wherein the weights of the four test data are set by the distance from the test data, and the closer the distance, the higher the weight, and vice versa.

[0056] Specifically, through the above-mentioned three-dimensional virtual model, the above-mentioned test data, geological data and rock parameters at each position can be intuitively seen. Although it is impossible to judge whether the test data is accurate alone, the stress magnitude relationship at two relative positions can be judged through the stress distribution law, and the above-mentioned magnitude relationship can be used to judge whether the above-mentioned test data needs to be corrected. When the above-mentioned test data meets the above-mentioned magnitude relationship, the above-mentioned test data does not need to be corrected. When the above-mentioned test data does not meet the above-mentioned magnitude relationship, the test data needs to be corrected, and the four test data closest to the front, back, left and right of the above-mentioned test data are weighted, among which the sum of the weights of the left and right two test data and the sum of the weights of the front and back two test data are both 0.5, and the closer the distance to the above-mentioned test data, the higher the weight, and vice versa. Through the above-mentioned technical solution, more accurate stress field data can be obtained, laying a foundation for further improving the accuracy of the rock stress prediction model.

[0057] Furthermore, the step S33 further includes a step S34:

[0058] Based on the three-dimensional virtual model, the test data, the geological data at the data collection location of the test data and the rock parameters of the sub-area where the test data is located are combined into training data, and any three consecutive training data belonging to the corresponding area of ​​any geological type are selected as detection data, and the remaining two training data are used as learning data.

[0059] Specifically, through the distribution position of the test data in the above-mentioned three-dimensional virtual model, the geological data, rock parameters and the above-mentioned test data at the position of the above-mentioned test data are combined into training data, and any three consecutive training data within the corresponding area of ​​any geological type are selected as detection data, and the remaining two training data are used as learning data. Through the above-mentioned technical scheme, more comprehensive learning data and detection data for each geological type are obtained to improve the accuracy of the rock stress prediction model.

[0060] Furthermore, the step S4 comprises:

[0061] Create the rock stress prediction model, obtain m geological types in the inversion area through the geological data, and create m sub-models based on the geological types, use the learning data in the sub-area corresponding to each geological type in the three-dimensional virtual model as the learning data of the corresponding sub-model, and train the sub-model through the learning data, wherein the value of m is a positive integer greater than or equal to 3.

[0062] Specifically, by creating a plurality of the above sub-models and fully training the learning data corresponding to each of the above address types through the above sub-models, the stress prediction value corresponding to each geological type can be obtained, thereby improving the accuracy of stress inversion.

[0063] Furthermore, step S4 also includes: inputting the geological data and rock parameters in the detection data into the rock stress prediction model, and obtaining the geological type to which the detection data belongs according to the position of the detection data in the three-dimensional virtual model, inputting the corresponding sub-model according to the geological type, obtaining a prediction result, comparing the prediction result with the stress value in the detection data, and obtaining a comparison difference. When the comparison difference is less than or equal to a preset difference, the prediction result is accurate. When the comparison difference is greater than the preset difference and the prediction result satisfies the size relationship between the stress value in the detection data and the stress value in the closest learning data in the sub-region, the sub-model is retrained by increasing the number of learning data of the geological type. When the comparison difference is greater than the preset difference and the prediction result does not satisfy the size relationship between the detection data and the closest learning data in the sub-region, step S2 is repeated to re-obtain the rock parameters corresponding to the geological type and retrain the sub-model.

[0064] Specifically, by inputting the input data in the detection data of the corresponding sub-model into the above-mentioned rock stress prediction model, the above-mentioned input data is the rock mass parameters and geological data in the detection data, obtaining the prediction result, that is, the predicted stress, and comparing the above-mentioned predicted stress with the corresponding stress value in the detection data, and obtaining the comparison difference, and correcting the above-mentioned rock stress prediction model according to the above-mentioned comparison difference. When the comparison difference is less than or equal to the preset difference, it is considered that the prediction result is relatively accurate and does not need to be corrected. When the comparison difference is greater than the preset difference, and the prediction result satisfies the size relationship between the detection data and the nearest learning data in the sub-area, it is considered that the prediction result is not accurate enough, but the error is relatively small. , which may be caused by the insufficient number of learning data. The accuracy of the corresponding sub-model is improved by increasing the number of learning data. When the comparison difference is greater than the preset difference, and the prediction result does not satisfy the size relationship between the detection data and the nearest learning data in the sub-region, it means that the error is relatively large and the accuracy of the sub-model is poor. It is possible that the rock parameters in the learning data are not accurate enough. By repeating the above step S2 to re-acquire the above rock parameters, re-acquire the above learning data and detection data, and re-train the above sub-model, the accuracy of the above sub-model is improved. Through the above technical solution, a rock stress model with higher accuracy can be obtained, thereby obtaining an accurate inversion rock stress field.

[0065] Furthermore, the test data is the value of the geostress field at the data collection location.

[0066] The present invention also provides a rock mass in-situ stress field inversion system based on machine learning, the system is used to implement the above method, such as Figure 2 As shown, the system comprises:

[0067] A virtual model generation unit is used to obtain a first geological image by scanning and imaging the target rock mass, filter and interpolate the first geological image to obtain a second geological image, and construct a three-dimensional virtual model of the target rock mass based on the second geological image; and obtain geological data at each position in the inversion area based on the three-dimensional virtual model, wherein the geological data includes measurement data of faults, fractures and terrain trends;

[0068] A marking unit, used to obtain the geological type in the inversion area based on the three-dimensional virtual model, select a number of sampling points within the area corresponding to each geological type, and obtain rock mass parameters corresponding to the sampling point by sampling and in-situ compressing the rock mass at each sampling point, and mark the rock mass parameters in the three-dimensional virtual model;

[0069] A data generating unit, configured to obtain a plurality of test data respectively through hydraulic fracturing and drilling according to the sub-region where each sampling point is located, pre-process the test data, and generate learning data and detection data from the pre-processed test data, the rock mass parameters of the sampling points corresponding to the test data, and the geological data;

[0070] A model training unit is used to create a rock stress prediction model, wherein the rock stress prediction model includes m sub-models, wherein the m sub-models are created based on the geological type, the sub-models are trained using the learning data, and the sub-models are verified based on the detection data, wherein each of the sub-models corresponds to a geological type.

[0071] The present invention also provides a computer storage medium, wherein the storage medium stores program instructions, wherein when the program instructions are executed, the device where the storage medium is located is controlled to execute the above method.

[0072] In summary, the present invention obtains a first geological image of a target rock mass through rock mass scanning technology, performs filtering and denoising on the first geological image, obtains abnormal pixels in the first geological image, obtains the number of other abnormal pixels within a set range, and performs filtering and interpolation on the abnormal pixels according to the number of abnormal pixels within the set range, obtains a second geological image, and obtains accurate geological data based on the second geological image, selects a landmark sampling area within the corresponding area of ​​each geological type according to the geological type of the rock mass in the three-dimensional virtual model, selects a plurality of sampling points at different positions in the sampling area, and obtains rock mass parameters of each sampling point, thereby obtaining relatively rich and comprehensive sampling rock mass parameters of each geological type, and based on the three-dimensional virtual model, each in step S2 The sub-region where the sampling point is located, and N data collection positions are uniformly selected in the above sub-region, and several groups of test data are obtained by hydraulic fracturing and drilling, and the above test data are marked in the above three-dimensional virtual model. Since the sampling point and the above data collection position are in the same sub-region, the rock mass parameters of the above sampling point and the above test data are combined into learning data and detection data to improve the accuracy of the rock stress prediction model. The stress size relationship at two relative positions in the above three-dimensional virtual model can be judged by the stress distribution law, and the above size relationship can be used to judge whether the above test data needs to be corrected. When correction is required, the four test data closest to the front, back, left and right of the above test data are weighted to obtain more accurate stress field data, laying the foundation for further improving the accuracy of the rock stress prediction model. Create multiple sub-models, and fully train the learning data corresponding to each of the above address types through the above sub-models, so as to obtain the stress prediction value corresponding to each geological type. Through the mutual cooperation of the above technical solutions, the accuracy of the inversion of the ground stress field can be improved.

[0073] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0074] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0075] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A rock mass in-situ stress field inversion method based on machine learning, characterized in that: The steps include: Step S1: scanning and imaging the target rock mass and acquiring a first geological image, filtering and interpolating the first geological image to acquire a second geological image, and constructing a three-dimensional virtual model of the target rock mass based on the second geological image; acquiring geological data at each position in the inversion area based on the three-dimensional virtual model, wherein the geological data includes measurement data of faults, fractures and terrain trends; Step S2: Based on the three-dimensional virtual model, the geological type in the inversion area is obtained, and a number of sampling points are selected within the area corresponding to each geological type, and the rock mass parameters corresponding to the sampling point are obtained by sampling the rock mass at each sampling point and performing in-situ compression, and the rock mass parameters are marked in the three-dimensional virtual model; Step S3: obtaining a plurality of test data sets respectively by hydraulic fracturing and drilling according to the sub-region where each sampling point is located, preprocessing the test data, and generating learning data and detection data from the preprocessed test data, the rock mass parameters of the sampling points corresponding to the test data, and the geological data; Step S4: Create a rock stress prediction model, the rock stress prediction model includes m sub-models, the m sub-models are created based on the geological type, the sub-models are trained using the learning data, and the sub-models are verified based on the detection data, wherein each of the sub-models corresponds to a geological type.

2. The method according to claim 1, characterized in that In step S1, filtering and interpolating the first geological image to obtain the second geological image includes: Step S11: abnormal pixel detection is performed on pixels in the first geological image, and a difference between the brightness value of each pixel and a first value is obtained. When the difference is greater than a set difference, the pixel is an abnormal pixel, and when the difference is less than or equal to the set difference, the pixel is a normal pixel, wherein the first value is the average of the maximum and minimum brightness values ​​of the pixels in the first geological image; Step S12: With the abnormal pixel as the center, within a set range, obtain the number of other abnormal pixels, and based on the number, obtain the normal pixels on a circle with a set radius around the abnormal pixel with the abnormal pixel as the center, and average the brightness values ​​of the normal pixels on the circle to obtain an interpolation value, wherein the larger the number, the smaller the set radius, and vice versa, the larger the radius, the number of abnormal pixels in the set range is less than the total number of pixels in the set range; Step S13: replacing the brightness value of each abnormal pixel with the interpolation value corresponding to the abnormal pixel.

3. The method according to claim 1, characterized in that The step S2 comprises: obtaining the geological type of the rock mass in the three-dimensional virtual model, selecting a landmark sampling area within the area corresponding to each geological type, selecting a plurality of sampling points at different positions in the sampling area, sampling the rock mass at each sampling point and performing in-situ compression to obtain the rock mass parameters corresponding to the sampling point, wherein the rock mass parameters include Young's modulus E, Poisson's ratio v, strain hardening modulus H, peak strain μ, and post-peak drop modulus M of the rock mass, and the geological types include at least folds, cracks, and faults.

4. The method according to claim 1, characterized in that The step S3 comprises the following steps: Step S31: based on the three-dimensional virtual model, the sub-region where the sampling point in step S2 is located is obtained, and N positions are uniformly selected in the sub-region as data collection positions, where the value of N is a positive integer greater than or equal to 10; Step S32: obtaining a plurality of groups of test data by the hydraulic fracturing method and the drilling method respectively, and marking the plurality of groups of test data at the data acquisition positions in the three-dimensional virtual model; Step S33: According to the geological data and rock mass parameters at the data collection position of the test data and the geological data and rock mass parameters at the data collection positions of other surrounding test data, obtain the size relationship between the test data and other surrounding test data, and judge whether the test data needs to be corrected based on the size relationship. When the test data satisfies the size relationship, the test data does not need to be corrected. When the test data does not satisfy the size relationship, the test data needs to be corrected, and weight the test data based on the four closest distances to the data collection position of the test data, namely, the front, back, left, and right, and use the weighted values ​​as the correction data of the test data, wherein the weights of the four test data are set by the distance from the test data, and the closer the distance, the higher the weight, and vice versa.

5. The method according to claim 4, characterized in that The step S33 further includes a step S34: Based on the three-dimensional virtual model, the test data, the geological data at the data collection location of the test data and the rock parameters of the sub-area where the test data is located are combined into training data, and any three consecutive training data within the corresponding area of ​​any geological type are selected as detection data, and the remaining two training data are used as learning data.

6. The method according to claim 1, characterized in that The step S4 comprises: Create the rock stress prediction model, obtain m geological types in the inversion area through the geological data, and create m sub-models based on the geological types, use the learning data in the sub-area corresponding to each geological type in the three-dimensional virtual model as the learning data of the corresponding sub-model, and train the sub-model through the learning data, wherein the value of m is a positive integer greater than or equal to 3.

7. The method according to claim 1, characterized in that The step S4 also includes: inputting the geological data and rock mass parameters in the detection data into the rock mass stress prediction model, and obtaining the geological type to which the detection data belongs according to the position of the detection data in the three-dimensional virtual model, inputting the corresponding sub-model according to the geological type, obtaining a prediction result, comparing the prediction result with the stress value in the detection data, and obtaining a comparison difference. When the comparison difference is less than or equal to a preset difference, the prediction result is accurate. When the comparison difference is greater than the preset difference and the prediction result satisfies the size relationship between the stress value in the detection data and the stress value in the closest learning data in the sub-region, the sub-model is retrained by increasing the number of learning data of the geological type. When the comparison difference is greater than the preset difference and the prediction result does not satisfy the size relationship between the detection data and the closest learning data in the sub-region, the step S2 is repeated to re-obtain the rock mass parameters corresponding to the geological type and retrain the sub-model.

8. The method according to claim 4, characterized in that The test data is the value of the geostress field at the data collection location.

9. A rock mass in-situ stress field inversion system based on machine learning, the system being used to implement the method according to any one of claims 1 to 8, characterized in that: The system comprises: A virtual model generation unit is used to obtain a first geological image by scanning and imaging the target rock mass, filter and interpolate the first geological image to obtain a second geological image, and construct a three-dimensional virtual model of the target rock mass based on the second geological image; and obtain geological data at each position in the inversion area based on the three-dimensional virtual model, wherein the geological data includes measurement data of faults, fractures and terrain trends; A marking unit, used to obtain the geological type in the inversion area based on the three-dimensional virtual model, select a number of sampling points within the area corresponding to each geological type, and obtain rock mass parameters corresponding to the sampling point by sampling and in-situ compressing the rock mass at each sampling point, and mark the rock mass parameters in the three-dimensional virtual model; A data generating unit, configured to obtain a plurality of test data respectively through hydraulic fracturing and drilling according to the sub-region where each sampling point is located, pre-process the test data, and generate learning data and detection data from the pre-processed test data, the rock mass parameters of the sampling points corresponding to the test data, and the geological data; A model training unit is used to create a rock stress prediction model, wherein the rock stress prediction model includes m sub-models, wherein the m sub-models are created based on the geological type, the sub-models are trained using the learning data, and the sub-models are verified based on the detection data, wherein each of the sub-models corresponds to a geological type.

10. A computer storage medium, characterized in that: The storage medium stores program instructions, wherein when the program instructions are executed, the device where the storage medium is located is controlled to execute the method according to any one of claims 1 to 8.

Images