Tunnel surrounding rock strength advanced prediction method based on machine learning

Abstract

The invention provides a tunnel surrounding rock strength advanced prediction method based on machine learning, which comprises the following steps: carrying out element test and compression and tensile test on various weathered rocks, and establishing an experimental database; generating intensity parameters and unit parameters through statistical sampling, performing numerical simulation, and constructing a simulation database; according to experiment and simulation data, respectively training element contents and test results, and simulating a mapping model between input and output; constructing a target function by taking the minimum model error as a target, and establishing a parameter matching mechanism by using an optimization algorithm in combination with element data of the rock to be measured to obtain optimal calculation parameters; and finally, parameters are tested and predicted according to field elements, surrounding rock modeling and deformation analysis are carried out, and support is provided for construction risk judgment.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel engineering, and in particular relates to a method for advanced prediction of tunnel surrounding rock strength based on machine learning. Background Art

[0002] In rock engineering practice, accurate prediction of rock properties is key to ensuring construction safety and reliability. Poor rock mass, especially in complex weathered geological environments, often leads to engineering accidents and affects the stability of mining and tunneling. Traditional rock classification methods, such as standard penetration tests, wave velocity ratio tests, and indoor uniaxial compressive strength (UCS) tests, although providing preliminary rock mass information, are questionable in terms of their reliability due to their reliance on a single evaluation criterion and lack of generalization capabilities. Therefore, it is particularly important to develop new methods that can accurately predict rock mass properties in real time.

[0003] X-ray fluorescence (XRF) technology, as an efficient tool for analyzing chemical element composition, plays an important role in geotechnical engineering optimization design. However, there is still a lack of effective tools to process large amounts of rock chemical element variation information to quantitatively characterize rock strength changes and drive actual engineering construction.

[0004] With the advancement of computing technology, numerical methods have become an important tool for solving geotechnical engineering problems and assisting decision-making. However, establishing precise correspondences between multiple microscopic parameter combinations and the actual rock mechanical response has been a challenge. In recent years, the application of machine learning in geotechnical engineering has continued to expand, particularly in predicting engineering stability based on key parameter learning, demonstrating significant potential and providing a potential solution to these problems. Summary of the Invention

[0005] The purpose of the embodiments of the present invention is to provide a method for advanced prediction of tunnel surrounding rock strength based on machine learning to achieve real-time risk assessment of rock engineering, solving the problems of insufficient data accuracy and real-time performance in existing technologies and the lack of accurate description of rock material damage and fragmentation.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is a method for advanced prediction of tunnel surrounding rock strength based on machine learning, which is carried out in the following steps:

[0007] S1. Conduct XRF and UCS / BTS tests on various types of weathered rocks; establish a test database based on the test results;

[0008] S2. randomly generating Mohr-Coulomb intensity parameters and cohesive unit parameters using a statistical sampling method, performing MC-FDEM simulation model calculations based on the Mohr-Coulomb intensity parameters and cohesive unit parameters, and establishing a numerical database based on the input parameters and corresponding simulation output results;

[0009] S3. Based on the experimental database and the numerical database, the relationship model Net-1 between XRF element test results and UCS / BTS results and the relationship model Net-2 between MC-FDEM simulation model input parameters and numerical results were established respectively;

[0010] S4. Establish an objective function based on the goal of minimizing the error between the output results of the model established in S3; establish a numerical parameter allocation mechanism using a machine learning method based on the XRF test results of the rock mass to be tested; and output the calculation parameters of the optimal MC-FDEM simulation model of the target shelter;

[0011] S5. Predict numerical parameters based on on-site XRF test results and perform MC-FDEM simulation model calculations on the surrounding rock. Analyze surrounding rock deformation based on the numerical model results and conduct corresponding construction risk assessments.

[0012] Furthermore, the specific process of establishing the relationship model Net-1 between the XRF element test results and the UCS / BTS results in S3 is as follows:

[0013] For a group of rocks with different degrees of weathering, the corresponding XRF data, compressive strength, and tensile strength were tested in the laboratory. The relationship function between the XRF test and the UCS / BTS results was established based on the following formula:

[0014]

[0015] Among them, f a is the activation function in the form of Sigmoid, ω ij is the weight of the node, θ i is the threshold, n is the number of input features, j is the index variable, y k is the prediction result, x k For input data.

[0016] Furthermore, the specific process of establishing the relationship model Net-2 between the MC-FDEM simulation numerical simulation input parameters and numerical results in S3 is as follows:

[0017] A completely random method was used to generate the solid element parameters and cohesion element parameters required for the MC-FDEM simulation model. These parameters were then input into the MC-FDEM simulation model to perform compression and tensile strength calculations. The calculation results were retained, and the relationship function between the randomly generated numerical parameters and the compression, tensile strength, and failure mode in the MC-FDEM simulation was established based on the following formula:

[0018]

[0019] Among them, f a is the activation function in the form of Sigmoid, ω ij is the weight of the node, θ i is the threshold, n is the number of input features, j is the index variable, y k is the prediction result, x k For input data.

[0020] Furthermore, the objective function in S4 is specifically:

[0021]

[0022] Where m represents the number of key test points on the selected stress-strain curve, F1, F2, and F3 are the UCS, BTS error, and failure shape error objective functions, T1, T2, D1, and D2 represent the UCS, BTS test results, and numerical model test results, respectively. N represents the number of XRF test elements, M represents the number of numerical calculation parameters, and x represents the number of XRF test elements. M Indicates numerical calculation parameters, x N represents the XRF test element, i represents the index variable, L represents the number of categories of the optimization model, P FP Indicates the probability of the corresponding target destruction mode.

[0023] Furthermore, the key test points include a loading starting point, a peak failure point, and a final failure point.

[0024] Furthermore, the values ​​of T1 and T2 are the outputs of the Net-1 model by inputting the XRF test data; the values ​​of D1 and D2 and the failure probability P FP It is the output of the Net-1 model with the optimal MC-FDEM numerical calculation parameters input.

[0025] Furthermore, the XRF test data is obtained through on-site testing.

[0026] Furthermore, the optimal MC-FDEM numerical parameter determination process is as follows:

[0027] S401. Generate random parameter combinations within a given range. Each parameter combination forms an individual in the population. All individuals form a parent population P0 with a population size of N1.

[0028] S402, based on conventional selection, crossover and mutation operations, a progeny population Q0 is formed, with a population size of N1;

[0029] S403, the parent population P generated by the Xth generation X and Q X Combined to form R X , the population size is 2N1;

[0030] S404, perform non-dominated sorting to generate non-dominated set F i , and determine the individual crowding attribute, if the size of F1 is smaller than N1, then move to the next parent population P X+1 Add F2 until it is added to F m The number of populations after the experiment is not less than N1. If the population size is equal to N1, the experiment stops. If the population size is greater than N1, the F species must be eliminated through the crowding attribute. m The inferior individuals, the remaining N1 individuals serve as the parent population of the next generation;

[0031] S405, based on conventional selection, crossover and mutation operations to form the offspring population Q X+1 , the population size is N1;

[0032] S406. If the number of iterations reaches the maximum genetic generation, output P i+1 All individuals in are taken as Pareto solutions as the optimal MC-FDEM numerical parameters, otherwise, a sub-population is generated and steps S403-S405 are repeated.

[0033] Furthermore, the individual congestion attribute in S404 is specifically:

[0034]

[0035] Where i represents the index variable i-th objective function, C j is the current non-dominated solution ζ j The degree of congestion; f i (ζ j-1 ), f i (ζ j+1 ) is ζ j-1 ,ζ j+1 The corresponding objective function value, f i max , f i min are the upper and lower thresholds of the current objective function solution, ζ j-1 ,ζ j+1are located at ζ in the sorted sequence j Before and after solutions.

[0036] Compared with existing technologies, the present invention offers the following advantages: It automates the entire process from non-destructive elemental testing to numerical simulation parameter assignment. It rapidly acquires high-quality data support without destroying rock samples, significantly improving modeling efficiency and adaptability. It can quickly and accurately map XRF test results to the mechanical parameters required by FDEM, overcoming the time-consuming and labor-intensive limitations of traditional testing methods. This invention not only significantly shortens the parameter acquisition cycle and improves the accuracy of the numerical model, but also provides a more reliable technical approach for solving complex geotechnical engineering problems, thereby enhancing the efficiency and accuracy of numerical model parameter acquisition. The present invention effectively integrates the advantages of multiple information sources, combining XRF's high-resolution component identification with FDEM's ability to describe discontinuous failure behaviors such as rock fracture and crack propagation, achieving efficient coupling between the material's microscopic composition and its macroscopic mechanical response. By integrating multiple sources of information, the present invention enables the model to not only reflect the true composition of geological materials but also accurately predict their failure behavior under different stress conditions, providing a crucial basis for engineering decision-making. It also provides a more proactive safeguard for engineering safety, facilitating the early identification and response to potential geological hazards, ensuring a safe and stable construction process. The present invention demonstrates excellent robustness and adaptability in the face of changing and complex geological environments. Even in field environments lacking conventional testing conditions, it can provide reliable strength parameter estimation and risk identification capabilities, making it particularly suitable for geotechnical engineering projects under various complex geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 Flowchart of the present invention.

[0039] Figure 2 This is a schematic diagram of a granite mine according to an embodiment of the present invention.

[0040] Figure 3 1. This is a graph showing the elemental test results of granite according to an embodiment of the present invention; (a) is the content of simple elements; (b) is the content of oxides.

[0041] Figure 4This is the NRBO-NSGA-NSGA-II workflow of an embodiment of the present invention.

[0042] Figure 5 Graphs showing granite strength test results according to an embodiment of the present invention; (a) shows the uniaxial compression stress-strain result, (b) shows the uniaxial compression failure morphology, (c) shows the Brazilian splitting stress-strain result, and (d) shows the Brazilian splitting failure morphology.

[0043] Figure 6 Schematic diagram of the numerical model strength database of an embodiment of the present invention; wherein, (a) is the BTS numerical database, and (b) is the UCS numerical database.

[0044] Figure 7 Schematic diagram of the destruction morphology database according to an embodiment of the present invention; wherein (a) is the UCS numerical database, and (b) is the BTS numerical database.

[0045] Figure 8 Schematic diagram of the prediction results of moderately weathered cylindrical granite according to an embodiment of the present invention; (a) is the XRF test result of moderately weathered rock, (b) is the numerical calculation result of cylindrical FDEM, and (c) is the comparison of the experimental and numerical results of the stress-strain curve.

[0046] Figure 9 Schematic diagram of the prediction results of slightly weathered cubic granite according to an embodiment of the present invention; (a) is the XRF test result of slightly weathered rock, (b) is the numerical calculation result of the cube FDEM, (c) is the morphology of the granite sample, and (d) is the comparison of the experimental and numerical results of the stress-strain curve.

[0047] Figure 10 This is a schematic diagram of the surrounding rock deformation prediction under three weathering conditions according to the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] like Figure 1This embodiment provides a machine learning-based method for advanced prediction of tunnel surrounding rock strength. Specifically, it is a real-time tunnel excavation risk assessment method that integrates XRF testing, machine learning, and the MC-FDEM method. Through machine learning, non-destructive test results obtained using XRF testing are rapidly used to provide the optimal numerical parameters required for MC-FDEM calculations, thereby reconstructing the geomechanical model for potential risk assessment and prediction. This embodiment combines the rapid and precise quantitative advantages of XRF with the significant advantages of FDEM in describing rock material damage and fragmentation, achieving a leap from compositional testing to rock mass strength assessment, providing a new solution for real-time risk assessment in rock engineering.

[0050] In some specific embodiments, the machine learning-based method for advanced prediction of tunnel surrounding rock strength first establishes a database between X-ray fluorescence (XRF) analysis results and rock strength through indoor experiments to reveal the relationship between rock chemical composition and mechanical properties. Subsequently, the generalized discrete element method (FDEM) simulation technology is used to construct a database of associations between MC-FDEM input parameters and model-predicted strength. Next, a machine learning method is used to process and analyze the relationship between XRF element characteristics, actual rock strength, and MC-FDEM parameters to establish a parameter prediction model. Finally, based on on-site XRF test data, the aforementioned model is used to predict the numerical simulation parameters of the rock mass ahead to achieve surrounding rock risk assessment. The details are as follows:

[0051] S1. Test and verify the correlation between XRF test results and UCS (uniaxial compressive strength) / BTS (indirect tensile strength) results of various types of weathered rocks through indoor experiments, and establish a test database based on the test data;

[0052] S2, batch calculation of MC-FDEM models using randomly generated Mohr-Coulomb parameters and Cohesive unit parameters to establish a numerical database;

[0053] Specifically, the S2 process is as follows: During the MC-FDEM numerical simulation, statistical sampling methods (such as Monte Carlo or Latin Hypercube sampling) are used to randomly generate multiple sets of Mohr-Coulomb strength parameters (such as internal friction angle, cohesion, and tensile strength) and cohesive element parameters (such as peak stress, fracture energy, and stiffness) based on the heterogeneous and anisotropic characteristics of the rock mass. Based on these parameter combinations, MC-FDEM simulation tasks are executed in batches. Ultimately, all input parameters and corresponding simulation output results are unified into a structured numerical database.

[0054] S3. Based on the experimental and numerical databases, a machine learning method (NRBO-BPNN) was used to establish the correspondence between XRF element test results and UCS / BTS test results, and the relationship between FDEM model input parameters and numerical results;

[0055] This embodiment uses the BPNN model to establish a functional relationship and make predictions by learning a set of data xy: k When the prediction result y k It can be determined according to the following formula:

[0056]

[0057] Where, f a is the activation function in the form of Sigmoid, ω ij is the weight of the node, θ i is the threshold, n is the number of input features, and j is the index variable. These parameters are calculated by comparing the actual data y with the predicted data y k The error is automatically adjusted and determined, and the value effect of the weight parameter can be further enhanced by introducing the NRBO enhancement operator.

[0058] First, for a group of rocks with different degrees of weathering, the corresponding XRF data, compressive strength (UCS), and tensile strength (BTS) were tested in the laboratory. Based on formula (1), the relationship function between XRF test and rock compressive and tensile strength, namely Net-1, was established.

[0059] Next, a completely randomized approach was used to generate over 300 sets of solid element and cohesion element parameters required for the MC-FDEM model. These were then input into the MC-FDEM software to perform compression and tensile strength calculations, and the results were saved. Similarly, a relationship function, Net-2, was established based on Equation (1) to establish the relationship between the randomly generated numerical parameters and the compression, tensile strength, and failure modes in the MC-FDEM simulation.

[0060] This completes the establishment of the relationship model between XRF testing and rock strength (Net-1) and the relationship model between input parameters and FDEM strength (Net-2), which is convenient for subsequent calls in S4.

[0061] S4. An objective function is established based on the minimum error of the UCS / BTS results of the Net-1 and Net-2 models. Then, a numerical parameter allocation mechanism is established using a machine learning method (optimization algorithm NSGA-II) based on the XRF test results of the rock mass to be tested. Specifically,

[0062] The objective function consists of three parts: compressive strength error (UCS), tensile strength error (BTS), and failure form error (FP). Whether it is an indoor test or MC-FDEM numerical simulation, the UCS and BTS results are always a stress-strain curve. Figure 5 As shown in (a) to (d), taking the key points on the stress-strain curve: the loading starting point, the peak failure point, and the final failure point can approximately replace the entire curve, so that the errors of these key points can be calculated to calculate the compressive strength error (UCS) and the tensile strength error (BTS). Since there are only two cases of similarity and dissimilarity in the failure morphology, it is only necessary to use 0 to represent similarity, that is, no error, and 1 to represent dissimilarity, that is, a large error. Based on the above principle, the objective function is established as follows:

[0063]

[0064] Where m represents the number of key test points on the selected stress-strain curve. The key test points in this embodiment include: loading starting point, peak failure point, and final failure point. F1, F2, and F3 are the objective functions of UCS, BTS error, and failure shape error. T1, T2, D1, and D2 represent the UCS, BTS test results, and numerical model test results, respectively. N represents the number of XRF test elements. M represents the number of numerical calculation parameters. M Indicates numerical calculation parameters, x N represents the XRF test element, i represents the index variable. L represents the number of categories of the optimization model, P FP Represents the probability of the corresponding target failure mode. This function can ensure that when the failure mode does not meet the requirements, the input value is continuously modified during the iterative optimization process so that P FP Close to 1, the final destruction result is consistent with expectations.

[0065] According to the XRF test data, the values ​​of T1 and T2 are calculated by inputting them into the Net-1 function. According to the optimal MC-FDEM numerical calculation parameters, the values ​​of D1 and D2 as well as the failure probability P are calculated by inputting them into the Net-2 function. FP XRF test results can be obtained through field testing, and the optimal MC-FDEM numerical parameters need to be determined through the optimization algorithm NSGA-II, so that the three objective functions mentioned above are minimized at the same time, and finally the numerical calculation parameters are output. The implementation steps of the NSGA-II algorithm are as follows:

[0066] (1) Generate random parameter combinations within a given range. Each parameter combination forms an individual in the population. All individuals form a parent population P0 with a population size of N1.

[0067] (2) Based on conventional selection (e.g., roulette wheel selection, tournament selection, elite selection, etc.), crossover and mutation operations form a descendant population Q0 with a population size of N1;

[0068] (3) The parent population P generated by the Xth generation X and Q X Combined to form R X , the population size is 2N1;

[0069] (4) Perform non-dominated sorting to generate non-dominated set F i , and calculate the individual crowding attribute (as shown in formula (3)). If the size of F1 is smaller than N1, then move to the next parent population P X+1 Add F2 to the m The number of populations after the experiment is not less than N1. If the population size is equal to N1, the experiment stops. If the population size is greater than N1, the F species must be eliminated through the crowding attribute. m The inferior individuals, the remaining N1 individuals serve as the parent population of the next generation;

[0070] (5) Based on conventional selection, crossover and mutation operations, the offspring population Q is formed X+1 , the population size is N1;

[0071] (6) If the number of iterations reaches the maximum genetic generation, output P i+1 All individuals in are taken as Pareto solutions as the optimal MC-FDEM numerical parameters, otherwise, a sub-population is generated and steps (3)-(5) are repeated.

[0072] The crowding distance operator is introduced to evaluate the local clustering of solutions, and the individual crowding attribute can be expressed as:

[0073]

[0074] Where i represents the index variable i-th objective function, C j is the current non-dominated solution ζ j The degree of congestion; f i (ζ j-1 ), f i (ζ j+1 ) is ζ j-1 ,ζ j+1 The corresponding objective function value, f i max , f i min are the upper and lower thresholds of the current objective function solution, ζ j-1 ,ζ j+1 are located at ζ in the sorted sequence j Before and after solutions.

[0075] In summary, the model framework process established in this embodiment is as follows Figure 4 shown.

[0076] S5. Predict numerical parameters based on on-site XRF test results and perform MC-FDEM simulation model calculations on the surrounding rock. Analyze surrounding rock deformation based on the numerical model results and conduct corresponding construction risk assessments.

[0077] Example 1

[0078] S1. Rock sample collection and numbering

[0079] First, collect and number the mine rock samples at different depths. Figure 2 The database was established using the granite mine in Baishi Town, Shandong Province, in northern China as an example. This region has a warm temperate monsoon climate, with an average annual temperature ranging from -3.4 to 31.6°C. Its hilly topography and rich mineral resources make it an ideal location for geochronological and geochemical studies. The mine studied is a yellow granite mine. According to common weathering grade classification standards, the rocks are classified along the slope surface as Type I (unweathered), Type II (slightly weathered), Type III (moderately weathered), and Type IV (strongly weathered). Type III granite was sampled from the slope surface at depths ranging from 20 to 100 meters. Type II granite was sampled from a surface depth of 150 to 200 meters, and unweathered rock was sampled from a depth of approximately 500 meters. Cylindrical BTS test specimens measured 50 mm × 25 mm, and UCS test specimens measured 50 mm × 100 mm. Three specimens were tested in each group for repetitive testing. These specimens were numbered 1-1 to 1-3 (BTS) for strongly weathered granite, 1-4 to 1-6 (UCS) for moderately weathered granite, 2-1 to 2-3 (BTS) for moderately weathered granite, 2-4 to 2-6 (UCS) for slightly weathered granite, and 3-1 to 3-3 (BTS) for slightly weathered granite, and 3-4 to 3-6 (UCS) for slightly weathered granite. Cube specimens measured 100 mm × 100 mm and were named A-1 to A-3 (fresh granite), B-1 to B-3 (slightly weathered granite), C-1 to C-3 (moderately weathered granite), and D-1 to D-3 (strongly weathered granite) according to their weathering intensity.

[0080] S2. Rock sample XRF test

[0081] At present, the main methods for determining the degree of rock weathering include standard penetration test, wave velocity ratio test and indoor uniaxial compressive strength (UCS) test. First, a standard penetration test is carried out to obtain the anti-penetration parameters of the rock mass. Subsequently, the wave velocity ratio test is used to measure the wave velocity changes inside the rock mass to evaluate its integrity and degree of weathering. Next, rock samples are collected and UCS tests are carried out under laboratory conditions to obtain their compressive strength data. In addition, in order to analyze the elemental composition characteristics of the rock, a portable XRF instrument is used for testing. Before testing, it is necessary to ensure that the sample surface is smooth and clean to reduce measurement errors. Place the sample on the XRF test platform, set appropriate test parameters, and perform multiple measurements to improve data stability. The test results are as follows: Figure 3 (a) and (b) show the changes in weathered granite. For weathered granite, the focus is on changes in key elements, including O, Si, Al, Na, K, and Ca. Because K is highly mobile and Na is relatively prone to residue or enrichment, the changes in Si, Al, K, and Ca content are calculated during data analysis to serve as a criterion for determining the degree of granite weathering.

[0082] S3, Rock Sample UCS / BTS Test

[0083] The UCS and BTS tests use cylindrical specimens with standard sizes of 50mm×100mm and 50mm×25mm, respectively. Each set of tests contains three identical specimens. First, the dimensions of the specimens are measured to ensure that they meet the standard requirements, and the surface is checked for obvious cracks or defects. The test is carried out using an RMT-301 digitally controlled electric servo testing machine. In the UCS test, the specimen is placed vertically on the loading platform, the loading rate is adjusted to the strain control mode that meets the experimental standard, and the stress-strain data of the entire loading process is recorded until the specimen is damaged. In the BTS test, the specimen is placed horizontally, and a uniform radial tensile force is applied using a loading plate to ensure that the fracture mode meets the tensile failure characteristics, and the force-displacement data is recorded, such as Figure 5 (a) to (d) show this. During the test, key parameters such as loading rate, peak load, and stress-strain curves were simultaneously recorded to analyze the mechanical properties of rocks with varying degrees of weathering. After all tests were completed, the fracture surfaces of the specimens were observed and recorded, and the influence of the degree of rock weathering on the UCS and BTS results was analyzed.

[0084] S4. Establishment of rock sample numerical database

[0085] The numerical result database needs to cover all possible rock stress-strain relationships in the form of an envelope to provide reliable combined parameter data. Including 5 MC model parameters of tetrahedral elements and 6 parameters of cohesive elements, a total of 11 random variables are generated, and 100 sets of parameters are generated by random sampling. All randomly generated parameters are stored in a dictionary format and input into the compression (UCS) and tension (BTS) numerical calculation models constructed by MC-FDEM through Python scripts. In the UCS / BTS numerical model, the loading conditions are set as: upper and lower loading rates of 0.05m / s, and the contact friction coefficient between the loading plate and the specimen is set to 0.1. After the calculation is completed, the Python script is used to batch extract stress-strain data, displacement, cohesive element damage, and crack morphology (FP) results and store them in the numerical database, such as Figure 6 (a) and (b) are shown. Subsequently, the specific process of establishing the numerical parameter allocation mechanism for XRF test results is as follows: using random parameters as input and the numerical result database as output, a BPNN model is established to form a callable function Net-2, and the NRBO algorithm is used to optimize its hyperparameters to improve the generalization ability of the model. The detailed steps are given in S4 and used Figure 4 Detailed integration was carried out.

[0086] After 141 sets of numerical calculations were completed, Python scripts were used to organize all stress-strain data and extract key failure morphology parameters, such as Figure 7 As shown in (a) and (b). The numbering rule is set as follows: UCS-U-XX represents the displacement result of the XXth case failure of the UCS model, and BTS-P-XX represents the distribution of the plastic zone of the XXth case failure of the BTS model. The loading limit is set to 0.79mm for the UCS model and 0.45mm for the BTS model. Under specific parameter combinations, different failure modes may occur, including plastic yield but no failure (such as case-76), crushing failure (such as case-22), compression and shear failure of the BTS model end (such as case-81), and single shear plane failure (such as case-45). Based on this, the failure morphology objective function F3 is added to enable further constraints on the corresponding failure morphology parameters, so that the compression (UCS) and tension (BTS) test experiments are completely consistent with the numerical results established by the MC-FDEM method in addition to the stress-strain. The failure morphology is also completely consistent, which significantly improves the rationality of the machine learning model to optimize the numerical calculation parameters.

[0087] S5. Rock sample verification

[0088] Taking the moderately weathered granite sample numbered 2-5 as an example, Figure 8The prediction results of medium weathering cylinder are shown. Based on the model established by S4, the XRF element test results of the target rock material are used as input ( Figure 8 (a)), the numerical parameters of MC-FDEM were obtained as output (the calculation results are shown in Figure 8 (b)), in the UCS test and BTS test, the stress-strain curves are similar to the results obtained from the experimental values ​​(as shown in Figure 8 (c)). Since the compressive and tensile strengths of the numerical / experimental database are based on the results of cylindrical samples, the predictions for cylindrical samples are very accurate, with the error in the prediction of the compressive strength of moderately weathered rock being only 5.47%.

[0089] However, since the database is constructed based on cylindrical samples, its generalization performance needs to be further tested. Taking the cubic sample B-2 as an example, the prediction performance of the method of this embodiment under different samples is verified. Figure 9 (a) to (d) respectively represent the element test results of the cubic specimen B-2, the comparison between the MC-FDEM test results and the experimental results, and the comparison of the stress-strain curves. From the XRF results, the test results of the B-2 sample are based on the difference between the medium-weathered and slightly-weathered cylindrical samples, which is consistent with the compression strength test results. The peak strength of the cube is higher than the peak strength of the cylindrical medium-weathered granite (106MPa) and lower than that of the slightly-weathered granite (193MPa). This shows that the method of establishing rock strength based on XRF test results in this embodiment is effective. From the failure mode, B-2 shows an obvious "dumbbell-shaped" failure mode with obvious oblique shear cracks. The MC-FDEM prediction of the peak strain is more conservative than the experimental value. This is because as the rock sample increases, there are more internal cracks, and the corresponding crack compaction stage lasts longer during the compression test. Therefore, the relative displacement prediction will be relatively conservative.

[0090] S6. Engineering Example Verification

[0091] like Figure 2 As shown in Figure 2, an actual granite mine in a certain area is selected as the research object, and a mine numerical model is established based on the geological conditions of the area, as shown in Figure 2. Figure 10 During the modeling process, the influence of the stratum deadweight and lateral in-situ stress was not considered. A vertical uniform stress load of 60 MPa was applied to the top of the model. Fixed constraints were used at the bottom and lateral boundaries (i.e., no displacement degrees of freedom). The model geometry was set according to the on-site engineering parameters.

[0092] In this embodiment, three types of granite geological conditions with different degrees of weathering that may be encountered during mine excavation are simulated: slightly weathered, moderately weathered, and strongly weathered granite. XRF elemental composition testing is performed on the rocks exposed during excavation at the construction site to obtain information on the main element composition of the rock samples. Subsequently, the rock mechanics parameter prediction method established by the present invention is used as input and imported into the FDEM numerical modeling platform to carry out numerical simulation analysis of the surrounding rock deformation and failure behavior. The simulation results are shown in Figure 2. Figure 10 As shown. In the case where the surrounding rock is slightly weathered granite, the overall structural deformation is small, and no obvious cracks or damage appear; under the same load conditions, the medium-weathered granite surrounding rock shows a large displacement response, accompanied by local crack initiation and expansion, but has not yet led to overall instability; in particular, under the conditions of strongly weathered granite, the surrounding rock undergoes obvious collapse and damage, indicating that its bearing capacity has significantly decreased. In summary, this embodiment obtains rock element information through on-site rapid XRF testing methods, and combines the parameter prediction model to construct an FDEM numerical model. It can effectively evaluate the stability of mine tunnels under conditions of rock masses with different weathering degrees, identify potential construction risks in advance, and has good engineering application prospects and promotion value.

[0093] This implementation allows for rapid, non-destructive, and efficient on-site acquisition of rock mass elemental XRF test results to assess construction safety, thereby preventing tunnel collapses in weathered granite. Furthermore, as the database continues to grow using on-site XRF test data and mine displacement monitoring results, the corresponding prediction results will become more accurate.

[0094] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A method for advanced prediction of tunnel surrounding rock strength based on machine learning, characterized in that: Follow these steps: S1. Conduct XRF and UCS / BTS tests on various types of weathered rocks; establish a test database based on the test results; S2. randomly generating Mohr-Coulomb intensity parameters and cohesive unit parameters using a statistical sampling method, performing MC-FDEM simulation model calculations based on the Mohr-Coulomb intensity parameters and cohesive unit parameters, and establishing a numerical database based on the input parameters and corresponding simulation output results; S3. Based on the experimental database and the numerical database, the relationship model Net-1 between XRF element test results and UCS / BTS results and the relationship model Net-2 between MC-FDEM simulation model input parameters and numerical results were established respectively; S4. Establish an objective function with the goal of minimizing the error between the output results of the model established in S3; and establish a numerical parameter allocation mechanism using a machine learning method based on the XRF test results of the rock mass to be tested; Output the calculation parameters of the optimal MC-FDEM simulation model of the target shelter; S5. Predict numerical parameters based on the on-site XRF test results and perform MC-FDEM simulation model calculations of the surrounding rock. Analyze the surrounding rock deformation based on the numerical model results.

2. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 1, characterized in that: The specific process of establishing the relationship model Net-1 between the XRF element test results and the UCS / BTS results in S3 is as follows: For a group of rocks with different degrees of weathering, the corresponding XRF data, compressive strength, and tensile strength were tested in the laboratory. The relationship function between the XRF test and the UCS / BTS results was established based on the following formula: Among them, f a is the activation function in the form of Sigmoid, ω ij is the weight of the node, θ i is the threshold, n is the number of input features, j is the index variable, y k is the prediction result, x k For input data.

3. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 1, characterized in that: The specific process of establishing the relationship model Net-2 between the input parameters and numerical results of the MC-FDEM simulation in S3 is as follows: A completely random method was used to generate the solid element parameters and cohesion element parameters required for the MC-FDEM simulation model. These parameters were then input into the MC-FDEM simulation model to perform compression and tensile strength calculations. The calculation results were retained, and the relationship function between the randomly generated numerical parameters and the compression, tensile strength, and failure mode in the MC-FDEM simulation was established based on the following formula: Among them, f a is the activation function in the form of Sigmoid, ω ij is the weight of the node, θ i is the threshold, n is the number of input features, j is the index variable, y k is the prediction result, x k For input data.

4. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 1, characterized in that: The objective function of S4 is specifically: Where m represents the number of key test points on the selected stress-strain curve, F1, F2, and F3 are the UCS, BTS error, and failure shape error objective functions, T1, T2, D1, and D2 represent the UCS, BTS test results, and numerical model test results, respectively. N represents the number of XRF test elements, M represents the number of numerical calculation parameters, and x represents the number of XRF test elements. M Indicates numerical calculation parameters, x N represents the XRF test element, i represents the index variable, L represents the number of categories of the optimization model, P FP Indicates the probability of the corresponding target destruction mode.

5. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 4 is characterized in that: The key test points include the loading starting point, the peak failure point, and the final failure point.

6. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 4, characterized in that: The values ​​of T1 and T2 are the output of the Net-1 model by inputting the XRF test data; the values ​​of D1 and D2 and the probability of the failure mode P FP It is the output of the Net-1 model with the optimal MC-FDEM numerical calculation parameters input.

7. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 6, characterized in that: The XRF test data is obtained through field testing.

8. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 6, characterized in that: The optimal MC-FDEM numerical parameter determination process is as follows: S401. Generate random parameter combinations within a given range. Each parameter combination forms an individual in the population. All individuals form a parent population P0 with a population size of N1. S402, based on conventional selection, crossover and mutation operations, a progeny population Q0 is formed, with a population size of N1; S403, the parent population P generated by the Xth generation X and the offspring population Q X Combined to form R X , the population size is 2N1; S404, perform non-dominated sorting to generate non-dominated set F i , and determine the individual crowding attribute, if the size of F1 is smaller than N1, then move to the next parent population P X+1 Add F2 until it is added to F m The number of the population is not less than N1, F1, F2...F m ∈F i If the population size is equal to N1, stop. If the population size is greater than N1, it is necessary to eliminate F by the crowding attribute. m The inferior individuals, the remaining N1 individuals serve as the parent population of the next generation; S405, based on conventional selection, crossover and mutation operations to form the offspring population Q X+1 , the population size is N1; S406. If the number of iterations reaches the maximum genetic generation, output P i+1 All individuals in are taken as Pareto solutions as the optimal MC-FDEM numerical parameters, otherwise, a sub-population is generated and steps S403-S405 are repeated.

9. The method for advanced prediction of tunnel surrounding rock strength based on machine learning according to claim 8, characterized in that: The individual congestion attribute in S404 is specifically: Where i represents the index variable i-th objective function, C j is the current non-dominated solution ζ j The degree of congestion; f i (ζ j-1 ), f i (ζ j+1 ) is ζ j-1 ,ζ j+1 The corresponding objective function value, f i max , f i min are the upper and lower thresholds of the current objective function solution, ζ j-1 ,ζ j+1 are located at ζ in the sorted sequence j Before and after solutions.

Images