Inversion method of multi-hole geostress field based on PSO-LSSVM

Abstract

The present invention discloses a multi-borehole geostress field inversion method based on PSO-LSSVM, comprising the following steps: 1) establishing a three-dimensional geostress calculation model using FLAC3D; 2) establishing a FLAC3D prediction model based on PSO-LSSVM through PSO-LSSVM training; 3) verifying the FLAC3D prediction model; and 4) obtaining the geostress field distribution of the entire region based on the FLAC3D prediction model. The present invention attempts inversion by combining different multiple boreholes each time, repeatedly testing which boreholes have a greater impact on the accuracy of the inversion results, effectively avoiding the problem of certain local borehole areas directly participating in the global inversion, resulting in large overall errors. By further dividing these boreholes into local areas for separate inversion, this method makes the inversion results closer to the global optimal solution, while avoiding the problem of local optimal solutions, thereby improving the fitting accuracy and computational efficiency of the multi-borehole inversion.

Description

Technical Field

[0001] The present invention relates to the field of rock mass engineering technology, and in particular to a multi-hole ground stress field inversion method based on PSO-LSSVM. Background Art

[0002] Excavation unloading is the primary factor contributing to surrounding rock failure during rock mass construction. This is primarily caused by excavation within a rock mass with an initial geostress field. Therefore, initial geostress is the fundamental cause of excavation unloading. The reliability of the initial geostress field and the rationality of rock mass parameter selection directly impact the reliability and safety of engineering design and construction.

[0003] The formation of geostress is primarily related to various internal and external geological dynamics, and is considered an unstable stress field. Its spatial distribution is extremely uneven and constantly changes over time. Geostress fields are difficult to express in functional form, so field measurements of geostress are the most direct way to understand a region's geostress field.

[0004] Currently, in-situ geostress measurements are primarily obtained through in-situ testing methods such as hydraulic fracturing, stress relief, and the Kaiser effect. However, these methods suffer from high testing costs and significant site limitations. Consequently, in-situ stress testing can only be performed at a small number of key locations within the project area. However, systematic analysis of the geostress field in complex tectonic zones based on a small number of measured geostress points is not feasible. Therefore, inversion of the geostress field for the entire region, combined with measured geostress data, is essential research.

[0005] Currently, commonly used geostress field inversion methods include stress (displacement) function analysis, boundary load inversion analysis, multiple linear regression analysis, and neural network algorithms. However, these methods still have many limitations in practice. For example, stress (displacement) function analysis is more suitable for uniform strata with simple geological structures; boundary load inversion analysis and multiple linear regression analysis have high errors in multi-borehole inversion; and neural network algorithms have very powerful information processing capabilities, but they may fall into local minima during the back propagation of errors. If the back propagation algorithm is trained too many times, it may also lead to slow convergence and unsatisfactory results.

[0006] In conclusion, the fitting accuracy and computational efficiency of current analysis methods, especially for multi-borehole inversion, still need further study. Summary of the Invention

[0007] In order to overcome the shortcomings of the above-mentioned technology, the purpose of the present invention is to provide a multi-borehole inversion method for geostress field based on PSO-LSSVM to improve the fitting accuracy and computational efficiency of multi-borehole inversion.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A multi-hole geostress field inversion method based on PSO-LSSVM includes the following steps:

[0010] 1) Establish a three-dimensional geostress calculation model using FLAC3D software;

[0011] 2) Establish a FLAC3D prediction model based on PSO-LSSVM through PSO-LSSVM training;

[0012] 3) FLAC3D prediction model validation;

[0013] 4) Obtain the distribution of the geostress field in the entire region based on the FLAC3D prediction model.

[0014] Preferably, the step 2) specifically includes:

[0015] 2.1) Obtaining original in-situ stress calculation values ​​at M measuring points in N0 measuring holes in the target area using the three-dimensional in-situ stress calculation model;

[0016] 2.2) Obtaining initial displacement boundary conditions based on the above-mentioned original calculated in-situ stress values ​​and measured in-situ stress values, and constructing K groups of displacement boundary condition data samples based on the initial displacement boundary conditions;

[0017] 2.3) Substitute K sets of displacement boundary condition data samples into the FLAC3D calculation model for inversion calculation and recalculate to obtain the initial in-situ stress calculation values ​​at M measuring points in N measuring holes under each displacement boundary condition data sample; where N = N0-1;

[0018] 2.4) The K-3 groups of K initial geostress calculation values ​​are taken as the training input values ​​of the PSO-LSSVM. The corresponding groups of displacement boundary condition data samples are used as output values. The three-dimensional geostress calculation model is optimized to obtain the FLAC3D prediction model based on the PSO-LSSVM.

[0019] Preferably, in step 2), the original in-situ stress calculation value and the initial in-situ stress calculation value are obtained by interpolation using a three-dimensional in-situ stress calculation model.

[0020] Preferably, in step 2.2), K groups of displacement boundary condition data samples are constructed through experience or orthogonal design.

[0021] Preferably, the step 2.4) is specifically as follows: the initial calculated ground stress values ​​at M measuring points in N measuring holes are used as training input values ​​of PSO-LSSVM, and the displacement boundary condition data are used as output values, and trained through SVM; then, the radial basis function is used as the kernel function of LS-SVM, the kernel parameters and penalty factors are optimized through PSO, and the obtained kernel parameters and penalty factor values ​​are used as LS-SVM model parameters to establish a prediction model based on PSO-LSSVM.

[0022] Preferably, the step 3) specifically includes:

[0023] 3.1) The FLAC3D prediction model based on PSO-LSSVM is tested using three sets of initial calculated in-situ stress values ​​that were not used in model training. If the test fails, the relevant parameters are adjusted until the test passes. If the test passes, the measured in-situ stress values ​​at M measuring points in N measuring holes are input into the PSO-LSSVM prediction model for calculation to obtain the corresponding gravity correction coefficient and displacement boundary conditions for the in-situ stress field.

[0024] 3.2) Based on the PSO-LSSVM prediction model, the calculated ground stress values ​​at M measuring points in N measuring holes are obtained, and compared with the corresponding measured ground stress values ​​at M measuring points in N measuring holes to obtain the relative error;

[0025] 3.3) If the relative error meets the requirements, proceed to step 4);

[0026] If the relative error does not meet the requirements, return to step 2), reselect N measuring hole positions, and obtain the initial ground stress calculation values ​​at M measuring points in N measuring holes, and repeat steps 2) to 3) until the relative error meets the requirements, and then go to step 4).

[0027] Preferably, in step 3.3), if the relative errors of N+1 options of N hole measurement positions cannot meet the requirements, reselect N hole measurement positions, set N=N-1, repeat steps 2)-3), and so on, until the relative error meets the requirements, and enter step 4). If the relative error still does not meet the requirements, and N-1=1, end the loop, divide the area to be measured into N0 small areas, and analyze each small area separately.

[0028] Preferably, the step 4) is specifically as follows:

[0029] Stress inversion analysis is performed on the N0-N measuring holes using the prediction model, and the relative errors of the inversion results of the N0-N measuring holes are calculated again. If the relative error meets the requirements, the prediction model is the final established FLAC3D prediction model based on PSO-LSSVM, and the distribution of the ground stress field in the entire area is obtained according to the FLAC3D prediction model based on PSO-LSSVM.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The multi-hole geostress field inversion method based on PSO-LSSVM proposed in this paper has significantly improved the calculation accuracy compared with conventional segmented single-hole inversion or multi-hole inversion, and provides a more refined geostress field inversion idea.

[0032] Currently, multivariate linear regression and BP neural networks are the most common methods for analyzing geostress fields. The present invention utilizes the PSO-LSSVM model to address linear and nonlinear multivariate calibration, solving the multivariate calibration problem relatively quickly and overcoming the shortcomings of SVM. PSO is used to optimize the relevant parameters in the LS-SVM, offering advantages such as being less susceptible to local minima and a simpler process. Furthermore, compared to BP neural networks, it can provide more reliable and improved generalization performance under the same training conditions. For this reason, PSO-LSSVM has been successfully applied in many fields requiring high-precision function approximation and pattern recognition. Furthermore, the accuracy of multi-well inversion is generally low, while single-well inversion suffers from the problem of local optimal solutions. The present invention combines multiple wells for inversion attempts, repeatedly testing which wells have the greatest impact on the accuracy of the inversion results. This effectively avoids the problem of large overall errors caused by the direct participation of certain well regions in the global inversion. By further dividing these wells into local regions for separate inversion, the inversion results are closer to the global optimal solution while avoiding the problem of local optimal solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flowchart of the implementation of a multi-hole geostress field inversion method based on PSO-LSSVM in the present invention. DETAILED DESCRIPTION

[0034] In order to better explain the present invention, the main contents of the present invention are further illustrated below in conjunction with specific examples, but the contents of the present invention are not limited to the following examples.

[0035] The present invention mainly relies on the idea of ​​"inversion forward calculation" to perform nonlinear inversion of the geostress field. Before performing the inversion calculation of the geostress field, it is necessary to know that the measured geostress values ​​at M measuring points in N0 measuring holes have been obtained on site.

[0036] (1) The causes of the initial geostress field of the rock mass are many, and are related to many factors such as its own properties, geological structure, temperature, groundwater, etc. From the experience of actual projects, it can be seen that the rock mass's own gravity and geological structure are the main factors affecting the initial geostress. Usually, the self-weight stress σ is used. g , the compressive stress σ of the X-direction horizontal compressive tectonic movement X , the compressive stress σ of the Y-direction horizontal compressive tectonic movement Y , shear stress τ of XY shear deformation tectonic movement XY , shear stress τ of YX-direction shear deformation tectonic movement YX , shear stress τ of ZX vertical shear deformation tectonic movement ZX , shear stress τ of ZY vertical shear deformation tectonic movement ZY The Z direction represents the vertical direction, the X and Y directions represent the horizontal directions, and the X, Y, and Z directions are orthogonal to each other. Using the linear superposition principle under the elastic working state, the initial ground stress of each point should be the result of the superposition of the above 7 calculation conditions. The initial ground stress expression of any point is:

[0037] σ=L1σ g +L2σ X +L3σ Y +L4τ XY +L5τ YX +L6τ ZX +L7τ ZY +e

[0038] Where, σ g , σ X , σ Y , τ XY , τ YX , τ ZX , τ ZY Represent the above 7 stress factors respectively; e represents error; L i (i=1, 2, 3, 4, 5, 6, 7) are the regression coefficients.

[0039] A three-dimensional geostress calculation model was established using FLAC3D software. The FLAC3D calculation model was used to calculate the seven working conditions mentioned above. The original geostress calculation values ​​at M measuring points in N0 measuring holes under the seven working conditions were obtained by interpolation.

[0040] Establish a multiple regression equation between the original calculated values ​​of the in-situ stress at M measuring points in N0 measuring holes and the measured values ​​of the in-situ stress at M measuring points in N0 measuring holes, and calculate the regression coefficient L i .

[0041] The regression coefficient L obtained by the above calculation i, the initial displacement boundary conditions A under 7 working conditions can be obtained. Based on the initial displacement boundary conditions A under these 7 working conditions, K groups of displacement boundary condition data samples are constructed through experience or orthogonal design. Each set of data consists of the initial displacement boundary conditions A of 7 working conditions, including: self-weight working condition and six working conditions: XX displacement, YY displacement, XY displacement, YX displacement, YZ displacement, and XZ displacement. Substitute the K groups of displacement boundary condition data into the FLAC3D calculation model for inversion calculation, and then use the FLAC3D calculation model to obtain the initial ground stress calculation values ​​at M measuring points in N measuring holes under each displacement boundary condition data through interpolation, that is, K initial ground stress calculation values ​​are obtained at each measuring point. It should be noted that N = N0-1 when performing the initial calculation.

[0042] (2) The initial calculated ground stress values ​​at M measuring points in N measuring holes under the K-3 groups of displacement boundary condition data are used as the training input values ​​of the particle swarm optimization least squares support vector machine (PSO-LSSVM), and the constructed displacement boundary condition data samples are used as the output values. Then, the radial basis function is used as the kernel function of the least squares support vector machine (LS-SVM), and the kernel parameter σ and the penalty factor C are optimized by the particle swarm optimization algorithm (PSO). The obtained kernel parameter σ and penalty factor C are used as the LS-SVM model parameters to establish a prediction model based on PSO-LSSVM.

[0043] (3) After the training and prediction are completed, the prediction model is tested using the other three data sets in the K data sets that did not participate in the model training as the test data sets. That is, the initial calculated values ​​of the ground stress at the M measuring points in the N measuring holes in the three data sets are used as the input of the prediction model, and the obtained output boundary conditions are compared with the initial displacement boundary conditions A obtained previously. The accuracy rate is used as the evaluation index of the prediction model to evaluate and analyze the prediction model, where the accuracy rate is defined as: for a given test sample, the ratio of the number of correctly predicted samples to the total number of samples. If the accuracy rate is low, the relevant parameters of the PSO-LSSVM model (such as inertia weight, acceleration constants c1 and c2, maximum particle velocity, etc.) are readjusted until the accuracy requirements are met. According to the actual measured values ​​of the ground stress at the M measuring points in the N measuring holes, all are input into the trained prediction model for calculation, and then the corresponding gravity correction coefficient and displacement boundary condition B of the ground stress field will be obtained.

[0044] (4) According to the obtained gravity correction coefficient and displacement boundary condition B, an inversion calculation is performed in the prediction model based on PSO-LSSVM to obtain the calculated ground stress values ​​at M measuring points in N measuring holes. The calculated ground stress values ​​at M measuring points in N measuring holes are compared with the corresponding measured ground stress values ​​at M measuring points in N measuring holes to obtain the relative error.

[0045] (5) Make error judgment.

[0046] If the error between the calculated ground stress values ​​at M measuring points in N measuring holes and the measured ground stress values ​​does not meet the actual requirements, then return to step (1) and update the measuring hole positions in the FLAC3D model, that is, reselect N measuring holes from the N0 measuring holes, which are different from the N measuring holes calculated previously, and calculate the initial ground stress values ​​at M measuring points in the reselected N measuring holes based on the K groups of displacement boundary condition data, train again based on the PSO-LSSVM model, and obtain the calculated ground stress values ​​at M measuring points in N measuring holes based on the PSO-LSSVM prediction model.

[0047] If, after N+1 cycles, the calculated values ​​of the in-situ stress at M measuring points in N measuring holes still do not meet the error requirements with the measured values, then the initial in-situ stress values ​​at M measuring points in N=N-1 measuring holes are reselected, and the process returns to step (1). The same K groups of displacement boundary condition data samples as before are used to calculate the calculated values ​​of the in-situ stress at M measuring points in the reselected N measuring holes through the same FLAC3D calculation model, and the next step is performed according to the same process until the error requirements are met.

[0048] If the error still does not meet the requirements, but it is judged that N-1=1, it is unnecessary to repeat the above steps. At this time, it is only necessary to divide the entire area into N+1 blocks and analyze them separately. The error will definitely meet the requirements. Of course, it is very rare to reach the step of N-1=1. Generally, when the analysis of the initial ground stress values ​​at M measuring points at 2 measuring holes is performed, the error will meet the actual requirements.

[0049] If the errors between the calculated values ​​of the in-situ stress at the M measuring points in the N measuring holes and the measured values ​​meet the requirements, proceed to step (6).

[0050] (6) Determine the error between the calculated ground stress values ​​and the measured ground stress values ​​at M measuring points in N0-N measuring holes: If the error between the calculated ground stress values ​​and the measured ground stress values ​​at M measuring points in N0-N measuring holes does not meet the actual requirements, then further refine the ground stress inversion analysis of the N0-N measuring holes that do not participate in the above inversion process, select a certain range of areas around the N0-N measuring holes for more refined grid division, and perform separate inversion to obtain the ground stress values ​​around the N0-N measuring holes; if the error between the calculated ground stress values ​​and the measured ground stress values ​​at M measuring points in N0-N measuring holes meets the requirements, then the model at this time is the final established FLAC3D prediction model based on PSO-LSSVM, based on which the ground stress field distribution of the entire region is obtained.

[0051] Other parts not described belong to the prior art.

Claims

1. A multi-hole geostress field inversion method based on PSO-LSSVM, characterized by: The following steps are involved: 1) Establish a three-dimensional geostress calculation model using FLAC3D software; 2) Establish a FLAC3D prediction model based on PSO-LSSVM through PSO-LSSVM training; 2.1) Obtaining original in-situ stress calculation values ​​at M measuring points in N0 measuring holes in the target area using the three-dimensional in-situ stress calculation model; 2.2) Obtaining initial displacement boundary conditions based on the above-mentioned original calculated in-situ stress values ​​and measured in-situ stress values, and constructing K groups of displacement boundary condition data samples based on the initial displacement boundary conditions; 2.3) Substitute K sets of displacement boundary condition data samples into the FLAC3D calculation model for inversion calculation and recalculate to obtain the initial in-situ stress calculation values ​​at M measuring points in N measuring holes under each displacement boundary condition data sample; where N = N0-1; 2.4) Take K-3 groups of K initial in-situ stress calculation values ​​as the training input values ​​of PSO-LSSVM, use the corresponding groups of displacement boundary condition data samples as output values, optimize the 3D in-situ stress calculation model, and obtain the FLAC3D prediction model based on PSO-LSSVM; 3) FLAC3D prediction model validation; 3.1) The FLAC3D prediction model based on PSO-LSSVM is tested using three sets of initial calculated in-situ stress values ​​that were not used in model training. If the test fails, the relevant parameters are adjusted until the test passes. If the test passes, the measured in-situ stress values ​​at M measuring points in N measuring holes are input into the PSO-LSSVM prediction model for calculation to obtain the corresponding gravity correction coefficient and displacement boundary conditions for the in-situ stress field. 3.2) Based on the PSO-LSSVM prediction model, the calculated ground stress values ​​at M measuring points in N measuring holes are obtained, and compared with the corresponding measured ground stress values ​​at M measuring points in N measuring holes to obtain the relative error; 3.3) If the relative error meets the requirements, proceed to step 4); If the relative error does not meet the requirement, return to step 2), reselect N measuring hole positions, and obtain the initial ground stress calculation values ​​at M measuring points in the N measuring holes, and repeat step 2)-step 3) until the relative error meets the requirement, and enter step 4); if the relative error cannot meet the requirement in N+1 options of selecting N measuring hole positions, reselect N measuring hole positions, set N=N-1, repeat step 2)-step 3), and so on, until the relative error meets the requirement, and enter step 4); if the relative error still does not meet the requirement, and N-1=1, end the loop, divide the area to be measured into N0 small areas, and analyze each small area separately; 4) Obtaining the distribution of the geostress field in the entire region based on the FLAC3D prediction model; Stress inversion analysis is performed on the N0-N measuring holes using the prediction model, and the relative errors of the inversion results of the N0-N measuring holes are calculated again. If the relative error meets the requirements, the prediction model is the final established FLAC3D prediction model based on PSO-LSSVM, and the distribution of the ground stress field in the entire area is obtained according to the FLAC3D prediction model based on PSO-LSSVM.

2. The multi-hole geostress field inversion method based on PSO-LSSVM according to claim 1 is characterized by: In step 2), the original in-situ stress calculation value and the initial in-situ stress calculation value are obtained by interpolation using a three-dimensional in-situ stress calculation model.

3. The multi-hole geostress field inversion method based on PSO-LSSVM according to claim 1 is characterized in that: In step 2.2), K groups of displacement boundary condition data samples are constructed through experience or orthogonal design.

4. The multi-hole geostress field inversion method based on PSO-LSSVM according to claim 1 is characterized by: The specific steps in step 2.4) are as follows: the initial calculated ground stress values ​​at M measuring points in N measuring holes are used as the training input values ​​of PSO-LSSVM, and the displacement boundary condition data are used as the output values, and trained through SVM; then, the radial basis function is used as the kernel function of LS-SVM, the kernel parameters and penalty factors are optimized through PSO, and the obtained kernel parameters and penalty factors are used as LS-SVM model parameters to establish a prediction model based on PSO-LSSVM.

Images