A method for quantitative evaluation of tunnel rock mass integrity

Abstract

The present invention relates to the technical field of tunnel engineering, and in particular to a method for quantitatively evaluating the integrity of a tunnel rock mass, comprising the following steps: S1, acquiring engineering data; S2, selecting and / or recalculating the engineering data to obtain sample data; S3, using the sample data and an XGBoost model to train a machine learning model for predicting the integrity of surrounding rocks; S4, calculating the SHAP value of each indicator in the sample data, performing interpretability analysis on the machine learning model, and determining the weight of each indicator; S5, using the weight of each indicator obtained in step S4, using a multivariate instability index analysis method to obtain a quantitative evaluation model for the integrity of a tunnel rock mass; S6, using the quantitative evaluation model for the integrity of a tunnel rock mass to evaluate the integrity of actual engineering data. By improving the utilization rate and rationality of drilling parameters, the accuracy of the final quantitative indicators is guaranteed, and the usability in the technical field of tunnel engineering is also guaranteed.

Description

Technical Field

[0001] The invention relates to the technical field of tunnel engineering, and in particular to a method for quantitatively evaluating the integrity of a tunnel rock mass. Background Art

[0002] During tunnel excavation, the quality evaluation of the integrity of the unexcavated rock mass in front of the face is an important process to determine the subsequent tunnel excavation method and support measures and ensure the safety of tunnel construction. The research and development trend of tunnel rock mass quality evaluation at home and abroad has also shown a trend from qualitative to quantitative. The mainstream quantitative evaluation methods of tunnel rock mass quality, such as Q value method, RMR method, BQ and modified [BQ] method, RQD method and GSI method, all regard the integrity of the rock mass as an important reference indicator, but they are more dependent on in-situ tests and indoor tests to obtain, which greatly increases the difficulty of engineering application. For the above reasons, more research is currently turning to the use of on-site advanced geological prediction results to seek reasonable methods for rapid and accurate quantitative evaluation of the integrity of the tunnel rock mass. Compared with conventional geophysical exploration methods (geophysical radar, TSP, infrared water exploration, etc.), digital drilling can most intuitively reflect the real geological information in front of the face (including the contained integrity information) by directly drilling into the surrounding rock in front of the face and returning a large number of drilling parameters. Therefore, it has become an important research object in this regard.

[0003] In terms of quantitative evaluation of rock mass integrity based on drilling data, there are currently two main technical solutions: (1) Mathematical statistics (data fitting) method, which is a common method for establishing quantitative evaluation relationships between indicators. In this regard, Professor Li Ning used the drilling method to experimentally study the response characteristics of drilling energy to rock mass discontinuity, and proposed an empirical method for determining rock quality index (RQD) and drilling energy changes based on discontinuity frequency; Professor Yue Zhongqi developed a digital drilling process monitoring (DPM) system for engineering rock mass quality evaluation, which improved and enhanced the existing engineering rock mass quality evaluation effect including integrity; Cao Ruilang built a new digital drilling monitoring system for geological drilling rigs and proposed a new quantitative indicator DPI for expressing rock mass integrity. (2) Machine learning method. Since digital drilling parameters are numerous and show nonlinear correlation characteristics, it is difficult for conventional mathematical statistics to achieve an ideal fitting effect. Therefore, some scholars have also used machine learning models that have performed well in the field of data analysis in recent years to conduct prediction research on drilling data. For example, Fang Yuwei used a neural network model to identify complete rock masses, incomplete rock masses and weak interlayers in limestone areas; Liu Shiyang used a convolutional neural network model to invert and predict broken zones prone to landslides before tunnel excavation.

[0004] The technical problems existing in the above research are as follows:

[0005] (1) The quantitative relationship between drilling parameters and mainstream rock mass integrity evaluation indicators is obtained through mathematical statistics (data fitting), or new quantitative evaluation indicators are proposed. In order to ensure the fitting effect, this method often makes certain idealized assumptions during application, ignoring the complexity of the on-site drilling operation environment and process. The selection of drilling indicators and the determination of weights are greatly affected by qualitative and subjective factors, and the evaluation effect is not specific enough (focusing on intact rock mass and incomplete rock mass);

[0006] (2) The method of machine learning is used to predict the integrity of the rock mass based on drilling data. The disadvantage is that it relies on a large amount of integrity data that matches the drilling data. Secondly, machine learning itself is a "black box" model that cannot directly provide quantitative indicators or formulas. The interpretability of the results and engineering practicality are poor.

[0007] (3) Both of the above methods have a common disadvantage, that is, the quantitative evaluation object is mostly a single tunnel section, while the tunnel construction object is a continuous rock mass. On the one hand, the single section evaluation undoubtedly greatly increases the randomness, and on the other hand, it has limited guiding significance for tunnel construction. Summary of the invention

[0008] The purpose of the present invention is to overcome the problem that the evaluation method based on machine learning in the prior art cannot directly give quantitative indicators or formulas, resulting in poor interpretability of the results and poor engineering practicality, and to provide a method for quantitative evaluation of the integrity of tunnel rock mass.

[0009] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0010] A method for quantitatively evaluating the integrity of a tunnel rock mass comprises the following steps:

[0011] S1, obtain engineering data;

[0012] S2, performing fine selection and / or secondary calculation on the engineering data to obtain sample data;

[0013] S3, using sample data and XGBoost model, train a machine learning model for predicting the integrity of surrounding rock;

[0014] S4, calculating the SHAP value of each indicator in the sample data, performing an interpretability analysis on the machine learning model for predicting the integrity of the surrounding rock, and determining the weight of each indicator;

[0015] S5, using the weight of each indicator obtained in step S4, using a multivariate instability index analysis method to obtain a quantitative evaluation model for the integrity of the tunnel rock mass;

[0016] S6, using the tunnel rock mass integrity quantitative evaluation model to evaluate the rock mass integrity of actual engineering data.

[0017] Preferably, the engineering data includes four primary indicators, namely propulsion speed, propulsion force, torque and rotation speed.

[0018] Furthermore, the step S2 of refining and / or recalculating the engineering data includes one or more of the following steps:

[0019] S21, outlier filtering;

[0020] S22, segment the engineering data according to the same section distance, calculate the mean and variance of the four first-level indicators in each segmented section, and obtain eight second-level indicators;

[0021] S23, performing oversampling processing on the secondary index;

[0022] S23, performing feature correlation analysis on the primary indicators or the secondary indicators, and eliminating features with correlations higher than a first threshold; and / or, performing feature importance analysis on the primary indicators or the secondary indicators, and eliminating indicators with feature importance lower than a second threshold.

[0023] In view of the characteristics of the original drilling sampling data, the present invention provides data denoising and equidistant segmentation steps in a preferred embodiment, segments the engineering data according to the same section distance, calculates the mean and variance of the four primary indicators in each segmented section, avoids the contingency caused by the evaluation of a single tunnel section, deeply explores the data rules that potentially reflect the integrity of the rock mass, and provides a high-quality data set.

[0024] Preferably, the first threshold is 0.8.

[0025] Preferably, the second threshold is 0.1.

[0026] Furthermore, the mathematical model D of the multivariate instability index analysis method in step S5 t as follows:

[0027]

[0028] Among them, n is the number of influencing factors, d1~d n is the instability index factor, w1~w n is the instability index weight.

[0029] Furthermore, the weight of each indicator is calculated according to the following formula:

[0030]

[0031] Among them, SHAP i Represents the SHAP value of the i-th indicator.

[0032] Preferably, in step S2, the sample data includes five indicators, namely, propulsion speed mean DRM, propulsion force mean TPM, torque mean TM, rotation speed mean RM and propulsion speed variance DRV.

[0033] The present invention performs indicator (i.e., feature) screening in a preferred embodiment and removes the thrust variance TPV, torque variance TV, and rotation speed variance RV, which are strongly correlated or have low importance, among the eight secondary indicators, so that the performance evaluation indicators of the machine learning model used to predict the integrity of the surrounding rock are improved by 3% to 5%.

[0034] Furthermore, the calculation formula IRFI of the tunnel rock integrity quantitative evaluation model in step S5 is as follows:

[0035]

[0036] Among them, W1 is the weight of the indicator DRM, W2 is the weight of the indicator TPM, W3 is the weight of the indicator TM, W4 is the weight of the indicator RM, and W5 is the weight of the indicator DRV.

[0037] Furthermore, the evaluation results of the tunnel rock mass integrity quantitative evaluation model in step S6 include:

[0038] When DRM ≥ 120, the evaluation result is mud filling;

[0039] When DRM<120 and the IRFI value range is (0,31), the evaluation result is relatively complete;

[0040] When DRM<120 and the IRFI value range is [31,55), the evaluation result is relatively fragmented;

[0041] When DRM<120 and IRFI≥55, the evaluation result is broken.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention selects and / or recalculates engineering data to obtain sample data, and then uses the sample data and the XGBoost model to train a machine learning model for predicting the integrity of the surrounding rock. The machine learning model for predicting the integrity of the surrounding rock is then subjected to an interpretability analysis to determine drilling indicators closely related to the integrity and the weights corresponding to these indicators. A quantitative evaluation model for the integrity of the tunnel rock mass is established using a multivariate instability index analysis method, and the integrity of the tunnel rock mass is quantitatively interpreted based on digital drilling data. On the one hand, the instability index and exponential power of the multivariate instability index analysis method are quantified and confirmed using machine learning and its interpretability as tools, thereby improving the utilization rate and rationality of the drilling parameters and ensuring the accuracy of the final quantitative indicators. On the other hand, it ensures its usability in the field of tunnel engineering technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of a method for quantitative evaluation of tunnel rock mass integrity.

[0045] Figure 2 Schematic diagram of surrounding rock with different degrees of completeness.

[0046] Figure 3 The original data for advance drilling.

[0047] Figure 4 This is a schematic diagram of the proportion of each label in the secondary indicator data set.

[0048] Figure 5 It is the result of the feature correlation analysis of the secondary indicator dataset.

[0049] Figure 6 This is the preliminary feature importance analysis result of the secondary indicator dataset.

[0050] Figure 7 A bar chart comparing the performance of the XGBoost classification model before and after feature filtering.

[0051] Figure 8 Schematic diagram of model interpretability analysis based on SHAP value.

[0052] Fig. 9 is the sample distribution and SHAP value of DRM eigenvalue.

[0053] Fig.10 It is a scatter plot based on IRFI.

[0054] Fig.11 This is a visualization image of the original drilling data in Example 2. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0056] Example 1

[0057] This invention innovatively combines mathematical statistics with machine learning, and with a series of effective drilling data preprocessing measures, to achieve high-quality quantitative evaluation of the integrity of the tunnel rock mass:

[0058] (1) Using the multivariate instability index analysis method as a mathematical model, a comprehensive index for quantitative interpretation of tunnel rock mass integrity is constructed to facilitate engineering applications;

[0059] (2) Based on the characteristics of the original drilling sampling data, through data noise reduction and equidistant segmentation, we deeply explore the data patterns that potentially reflect the integrity of the rock mass and provide high-quality data sets;

[0060] (3) The instability index and exponential power of the multivariate instability index analysis method are quantified and confirmed using machine learning and its interpretability as tools to improve the utilization and rationality of drilling parameters and ensure the accuracy of the final quantitative indicators.

[0061] This embodiment provides a method for calculating comprehensive indicators of tunnel rock mass integrity quantitative interpretation based on digital drilling, which can achieve quantitative evaluation of rock mass integrity. Figure 1 As shown, the following steps are included:

[0062] S1, obtain engineering data;

[0063] S2, performing fine selection and / or secondary calculation on the engineering data to obtain sample data;

[0064] S3, using sample data and XGBoost model, train a machine learning model for predicting the integrity of surrounding rock;

[0065] S4, calculating the SHAP value of each indicator in the sample data, performing an interpretability analysis on the machine learning model for predicting the integrity of the surrounding rock, and determining the weight of each indicator;

[0066] S5, using the weight of each indicator obtained in step S4, using a multivariate instability index analysis method to obtain a quantitative evaluation model for the integrity of the tunnel rock mass;

[0067] S6, using the tunnel rock mass integrity quantitative evaluation model to evaluate the rock mass integrity of actual engineering data.

[0068] Specifically, in step S1, the engineering data is the original drilling data collected from a tunnel under construction in Guangxi. The engineering geological conditions of the tunnel site are complex. It mainly passes through relatively broken to broken weathered sandstone strata, and there are mud-filled caves and cracks. It is very easy to have adverse geological disasters such as mud burst and landslides on the face. Figure 2 shown.

[0069] To ensure the safety of tunnel construction, the C6-2 multifunctional crawler drilling rig produced by Casagrande was used to conduct advance drilling operations and to make geological forecasts based on drilling data and on-site conditions. The original drilling data collected was as follows: Figure 3 As shown, the four primary indicators are propulsion speed, propulsion force, torque and rotation speed. The present invention collects more than 10,000 original drilling data including the left and right sections of the tunnel, and the tunnel length covered by the data is about 150 meters in total.

[0070] In step S2, the engineering data is refined and / or recalculated to obtain sample data. This step is a step of preprocessing the data, and specifically includes one or more of the following steps:

[0071] (1) Outlier Filtering

[0072] During the drilling process, due to the operating environment, operator operation and other reasons, a large amount of first-level indicator data collected inevitably contains abnormal data and some special drilling status data, which is specifically manifested as one or several first-level indicator collection values ​​being 0. Therefore, before using drilling data to train machine learning models, outlier filtering is required. The filtering rule is shown in the following formula (1):

[0073] T=x 推进速度 ·x 推进力 ·x 扭矩 ·x 旋转速度 (1)

[0074] In formula (1), x 推进速度 、x 推进力 、x 扭矩 and x rotation speed represent the data values ​​of the four primary indicators of a certain sampling point respectively. If T=0, the value of the sampling point is judged to be an abnormal value and filtered. Otherwise, if T≠0, it is judged to be a normal value and retained.

[0075] (2) Data equidistant segmentation and feature engineering

[0076] In order to solve the problem mentioned above that the cross section is used as the evaluation unit in the field of advanced drilling data interpretation, after the data outlier filtering is completed, the present invention performs equidistant segmentation processing on the original drilling data according to the spacing d=0.5m, and calculates the mean and variance of the four primary indicators in each segment according to formula (2) to formula (3), that is, performs feature engineering processing to form new secondary indicators, which are specifically shown as follows:

[0077]

[0078]

[0079] In the formula, n is the number of samples, x is n is the value of the nth sample, x i is the value of the i-th sample, is the mean of n samples, is the variance of n samples.

[0080] (3) Analysis and processing of secondary indicator data sets

[0081] After data segmentation and feature engineering, the secondary indicators formed include the mean driving speed (DRM), the mean propulsion force (TPM), the mean torque (TM), the mean rotation speed (RM), the propulsion speed variance (DRV), the propulsion force variance (TPV), the torque variance (TV), and the rotation speed variance (RV). The secondary indicator data set has a total of 295 samples, and the proportion of each completion level is as follows: Figure 4 shown.

[0082] Depend on Figure 4 It can be seen that the dataset has a certain degree of sample imbalance. In order to ensure the accuracy and effectiveness of the subsequent machine learning model, the dataset needs to be oversampled. The SMOTE algorithm is an oversampling technology used to synthesize minority samples, which can be regarded as an improved strategy for the random oversampling algorithm. Traditional random oversampling uses a simple sample duplication strategy to increase the number of minority samples, which is prone to overfitting and reduces the generalization ability of the classifier. In order to overcome this problem, the basic principle of the SMOTE algorithm is to analyze minority samples and synthesize new samples based on the minority samples and add them to the dataset.

[0083] The SMOTE oversampling process is shown in formula (4):

[0084] x new =x i +rand(0,1)|x i -x j | (4)

[0085] In formula (4): x i represents the minority class sample, and calculates the difference between it and all other minority class samples x i The Euclidean distance between them is used to obtain the k nearest neighbors. Usually k is taken as 5; x j represents a sample randomly selected from the k nearest neighbors; rand(0,1) represents a random number between 0 and 1. For the nearest neighbor sample x j and minority class samples x i , synthesize a new sample x at a random point on their connecting line new .

[0086] The present invention utilizes the SMOTE algorithm, an oversampling technology for synthesizing minority samples, in a preferred embodiment, to analyze minority samples and synthesize new samples based on the minority samples and add them to the data set, thereby solving the sample imbalance problem and avoiding the problem of overfitting and reduced classifier generalization ability caused by increasing the number of minority class samples by simply copying samples.

[0087] The processed data set is subjected to Pearson feature correlation analysis according to formula (5) to explore the possibility of reducing data dimensions before machine model training.

[0088]

[0089] In formula (5): a i and b i is the data value, and is the average value.

[0090] The value of Pearson's coefficient ρ is between -1 and +1. Values ​​close to +1 indicate strong positive correlation, values ​​close to -1 indicate strong negative correlation, and values ​​close to 0 indicate no correlation. Figure 5 Correlation heat map shown.

[0091] Depend on Figure 5 It can be seen that the highest positive correlation is 0.46 between TPV and DRV, and the highest negative correlation is -0.68 between DRV and TPM. There is no feature with a high correlation (absolute value>0.8), so no feature screening is performed based on the correlation analysis. In addition to simply screening features from the numerical perspective of the features through correlation analysis, the importance of features should also be considered from the perspective of the model. Therefore, the feature_importances function of the XGBoost model is used to preliminarily visualize the importance of features. The results are as follows: Figure 6 shown.

[0092] pass Figure 6It can be seen that among the eight secondary indicators, the first five DRM, TPM, TM, RM and DRV feature importances are all over 0.1, which are relatively important features, and the last three TPV, TV and RV feature importances are all less than 0.1, which are unimportant features. Based on this, TPV, TV and RV are removed. Finally, the first five are retained to form the data set for machine learning model training and learning.

[0093] After step S2 preprocesses the engineering data, step S3 is executed to train a machine learning model for predicting the integrity of the surrounding rock using the sample data and the XGBoost model.

[0094] XGBoost stands for Extreme Gradient Boosting, which can be translated as the extreme gradient boosting algorithm. Different from traditional algorithms, it is a highly efficient machine learning algorithm developed from the traditional machine learning classification and regression tree algorithm (CART). As a representative algorithm of boosting in the integrated algorithm, XGBoost builds multiple weak evaluators, namely CART, on the data one by one, and gradually accumulates and summarizes the modeling results of multiple weak evaluators through multiple iterations to obtain better regression or classification performance than a single model. This superposition strategy with a single decision tree as a weak evaluator can be expressed in the form of addition, as shown in formula (6):

[0095]

[0096] In formula (6): (or ) represents the prediction result of the entire model on this sample, K represents the total number of weak evaluators, and f k represents the kth decision tree, x i Represents the eigenvector corresponding to sample i.

[0097] XGBoost introduces model complexity to measure the computational efficiency of the algorithm. Therefore, the objective function consists of two parts: the traditional loss function and the model complexity. The formula is shown in formula (7):

[0098]

[0099] Ω(f k )=γT+λ||ω|| 2 / 2 (8)

[0100] In formula (7), Obj represents the objective function of the model, n represents the total amount of data imported into the k-th tree, and the first term represents the traditional loss function, which measures the true label y i and predicted values The second term represents the complexity of the model, which is represented by a certain transformation Ω of the tree model. This change represents a formula for measuring the complexity of the tree model from the structure of the tree, which is expanded as shown in formula (8). In formula (8), γ and λ represent the coefficients of the model complexity, and T represents the number of tree nodes in the model decision tree.

[0101] In order to solve the objective function, Taylor expansion can be used to calculate equation (7), and the result is shown in equation (9):

[0102]

[0103] In formula (9), γ and λ represent the coefficients of model complexity, T represents the number of tree nodes in the model decision tree, and g i 、h i Respectively represent samples x i The first and second derivatives of , j represents the index of each leaf node, ω j represents the sample weight on the jth leaf node, I j Represents the sample subset of the jth leaf node.

[0104] Bring the tree structure into the loss function, that is, j Taking the derivative and setting the derivative function equal to zero, the minimum value of the objective function Obj can be obtained. min .Obj min It is an important indicator to measure the quality of the model. The smaller its minimum value is, the better the performance of the model is. The calculation formula is shown in formula (10) to formula (11):

[0105]

[0106]

[0107] The present invention uses machine learning as a tool to obtain the rationality of the index and corresponding exponential power of the multivariate instability index analysis method. One of the most important measurement indicators is whether the performance of the machine learning model is good enough. Only by performing interpretable analysis on high-performance models and then obtaining important drilling indicators and weights can it be sufficiently reasonable. To this end, it is necessary to first evaluate the performance of the model.

[0108] In order to comprehensively evaluate the model performance, the selected model evaluation indicators include accuracy (ACC), precision (PRE), recall (REC), and harmonic mean score (F1). The calculation formulas (12) to (15) of all the above evaluation indicators are shown.

[0109]

[0110]

[0111]

[0112]

[0113] Where: TP is true positive, which indicates the number of samples that are actually positive and correctly predicted as negative by the classifier; FN is false negative, which indicates the number of samples that are actually negative but are incorrectly predicted as positive; FP is false positive, which indicates the number of samples that are actually negative but are incorrectly predicted as positive; TN is true negative, which indicates the number of samples that are actually negative and correctly predicted as negative.

[0114] Finally, the performance of the XGBoost classification model constructed by the present invention is as follows Figure 7 shown. Figure 7 The left side of each group of bar charts shows the model performance evaluation results without feature screening (8 features), and the right side shows the model performance evaluation results after feature screening (5 features). It can be found that the four model performance evaluation indicators have all been improved by 3% to 5%, which verifies the effectiveness of feature engineering through feature importance in the present invention.

[0115] Therefore, based on the characteristics of the original drilling sampling data, the present invention, on the one hand, deeply explores the data patterns that potentially reflect the integrity of the rock mass through data denoising and equidistant segmentation to provide a high-quality data set; on the other hand, through indicator (i.e., feature) screening, removes strongly correlated indicators or indicators with low importance, so that the performance evaluation indicators of the machine learning model used to predict the integrity of the surrounding rock are improved by 3% to 5%.

[0116] After the machine learning model for predicting the integrity of the surrounding rock is trained, step S4 is executed to perform interpretability analysis on the model.

[0117] At present, many machine learning models are "black box" algorithms. Although they can make very good predictions, they cannot explain how they make predictions. Many data scientists find it difficult to know why the algorithm gets such prediction results, which is actually very fatal. Therefore, the interpretability of machine learning, or XAI (Explainable Artificial Intelligence), has always been the focus and difficulty in this field.

[0118] Feature importance analysis is the most commonly used interpretability method in tree models and ensemble algorithms based on tree model theory. By outputting feature importance data, we can intuitively know which features are the most important for the model, which are key features, and which features have little impact on the model prediction results. Based on this, feature engineering can be performed, or a reference can be provided for the prediction of subsequent new sample data (try to ensure the authenticity of important feature data). In XGBoost, importance_type=gain is often used, that is, the importance of each feature is measured by calculating the splitting gain of each feature. According to formula (11), the gain calculation definition of the split point is as shown in formula (16):

[0119]

[0120] In the formula, L and R represent the tree model splitting to the left and splitting to the right respectively.

[0121] However, this method can only obtain the feature importance of the entire data set, that is, a rough global interpretability, and this importance is usually based on the training set. On the one hand, when the model is overfitted, the importance of the feature is misleading, and on the other hand, it is easily affected by high cardinality features, which often makes numerical variables rank high. For classification problems, if you want to further understand the importance of each indicator corresponding to each type of label, as well as the predictive interpretability of local samples, you still need to use other interpretability theories and methods.

[0122] The SHAP model is based on the SHAP value (Shapley Value) for model interpretability, a concept that comes from game theory. This model is currently a relatively versatile and novel method for model interpretability, which can be used for both previous global explanations and local explanations. SHAP is a method for post-model interpretation. Its core idea is to calculate the marginal contribution of features to the model output, and then explain the "black box model" from both the global and local levels. SHAP constructs an additive interpretation model, in which all features are considered "contributors." For each predicted sample, the model generates a predicted value, and the SHAP value is the value assigned to each feature in the sample.

[0123] Assume that the model benchmark score (usually the mean of the target variable for all samples) is y base , the i-th sample is x i , the kth feature of the i-th sample is x i,k , the Shap value of this feature is f(x i,k ), then the model is good for sample xi The predicted value of is:

[0124] y i =y base +f(x i,1 )+f(x i,1 )+...+f(x i,k ) (17)

[0125] When f(x i,k )>0, indicating that the feature has a positive effect on the prediction of the target value, otherwise the feature has an opposite effect on the target prediction value. Therefore, the SHAP value not only gives the magnitude of the feature's influence, but also reflects the positive and negative influence of the feature in each sample. Based on the calculated SHAP value of each sample, the model can be partially and fully interpreted.

[0126] In step S4 of this implementation, the importance of each feature to the integrity of the four surrounding rocks is analyzed by calculating the SHAP value. The results are as follows: Figure 8 As shown, Figure 8 The horizontal axis is the calculated SHAP value, and the vertical axis is the order of features from top to bottom according to the SHAP value. It can be seen that, except for mud filling, the five important features have made different contributions to the correct prediction of rock masses of various integrity levels.

[0127] In order to further analyze the interpretability of DRM prediction for mud filling, the sample distribution and SHAP value of DRM eigenvalues ​​are plotted under the premise of mud filling label. Fig. 9 As shown. Fig. 9 It can be seen that the DRM feature value for the prediction of mud filling basically takes 120 as the critical value: DRM<120, the feature makes a negative contribution to the prediction of mud filling, that is, the prediction is wrong; DRM>120, the feature makes a positive contribution to the prediction of mud filling, that is, the prediction is correct.

[0128] In step S4, the SHAP value of each indicator in the sample data is calculated. After the interpretability analysis is performed on the machine learning model for predicting the integrity of the surrounding rock, the present invention needs to obtain a quantitative evaluation model for the integrity of the tunnel rock mass by a multivariate instability index analysis method. Therefore, it is necessary to determine the weight of each indicator. The specific process is as follows.

[0129] In the prior art, the multivariate instability index analysis method is to obtain the instability index (D) under a relative relationship based on quantitative statistics. t ), a comprehensive quantitative indicator used to describe the severity of an event, D t The larger the value, the more severe the event is.

[0130] The mathematical model of the standard multivariate instability index analysis method is shown in the following formula (18):

[0131]

[0132] In formula (18), n is the number of influencing factors, d n is the instability index factor, W n is the instability index weight.

[0133] When applying digital drilling data to quantitatively evaluate the integrity of tunnel rock mass, comprehensive consideration of modeling complexity and evaluation accuracy means that it is necessary to clarify the drilling indicators that are closely related to the integrity and the weights corresponding to these indicators.

[0134] According to formula (18) and Figure 8 The present invention proposes an "Interval Rock Fracture Index (IRFI)", and the instability index is DRM, RM, TPM, TM, DRV. Each index power is calculated according to the following formula (19):

[0135]

[0136] In the formula, SHAP i Represents the SHAP value of the i-th feature (i.e. indicator) of the machine learning model.

[0137] In step S5, the weight W of each indicator obtained above is used i (i.e., exponential power), and the multivariate instability index analysis method is used to obtain a quantitative evaluation model for the integrity of the tunnel rock mass; finally, the IRFI quantitative formula determined in this embodiment is shown in formula (20):

[0138] IRFI=DRM 0.27 ×TPM 0.08 ×TM 0.08 ×RM 0.09 ×DRV 0.49 (20)

[0139] In order to verify the effectiveness of formula (20) for quantitatively evaluating the integrity of tunnel rock mass, the IRFI scatter plot of the data set is drawn. The results are shown in Fig.10 shown. Fig.10 In the figure, the horizontal axis is the number of the sample data, that is, the serial number. The sample data marked with light blue are relatively complete rock masses, the sample data marked with dark blue are relatively broken rock masses, the sample data marked with purple are broken rock masses, and the sample data marked with brown are mud-filled. The vertical axis is the IRFI value calculated for each sample data.

[0140] Depend on Fig.10It can be seen that the IRFI value can distinguish relatively intact, relatively broken, and broken rock masses with good results, and the comprehensive data coverage rate can reach about 90%. However, it cannot distinguish mud fillings, which is caused by the particularity of the mud filling data type, which can be seen from Figure 8 As can be seen in the figure, unlike the other three types of data, the data type of mud filling is basically only related to DRM (interval drilling rate mean), while IRFI is a comprehensive indicator. However, due to its particularity, we can distinguish it by adding discrimination conditions, and the final quantitative interpretation scheme is shown in Table 1 below.

[0141] Table 1 Quantitative interpretation table and accuracy based on model interpretability and IRFI index

[0142]

[0143] Example 2

[0144] In order to verify the evaluation effect of the quantitative evaluation index of rock mass integrity based on the advance drilling data proposed in Table 2, the advance drilling data of ZK109+240 at the left tunnel face of the tunnel were selected for verification. Specifically, 20 m from 21 to 41 m of hole No. 2 were selected (because it is representative and has many types of surrounding rock integrity). The original drilling data image is shown in the figure below. Fig.11 shown.

[0145] The evaluation results of professional technicians in the advanced geological forecast of this section are that the rock mass at 21-28m is relatively intact to relatively broken, and there is suspected mud-filled weak interlayer near 25m, and the rock mass at 29-41m is relatively broken to broken, and there is suspected mud-filled weak interlayer at 38-40m. The IRFI value judgment results are shown in Table 2 below:

[0146] Table 2 Quantitative discrimination results based on IRFI values

[0147] depth DRM TPM TM RM DRV IRFI Quantitative interpretation 21.5 18.2 101.2 48.43 135.05 45.35 41.91 More broken 22 15.55 101.35 48.95 130.76 43.92 39.48 More broken 22.5 16.27 100.98 54.62 127.63 35.47 36.3 More broken 23 23.54 101.54 63.15 148.6 488.42 144.9 broken 23.5 20.79 101.57 56.94 138.43 201.8 90.34 broken 24 26.21 101.63 59.6 136.14 53.02 50.74 More broken 24.5 26.21 100.94 60.53 145.41 627.73 167.27 broken 25 115.35 101.13 64.1 138.6 5998.92 737.77 Clay filling 25.5 23.85 101.31 64.88 137.07 21.85 30.56 More complete 26 21.87 101.51 63.19 148.62 2.49 11.28 More complete 26.5 22.17 101.21 59.82 140.73 12.97 24.75 More complete 27 20.95 101.41 59.14 146.86 22.16 30.61 More complete 27.5 23.84 101.59 58.16 145.51 28.47 36.84 More broken 28 24.19 101.66 56.05 135.27 62.43 53.41 broken 28.5 16.99 101.56 57.91 142.6 51.43 44.57 More broken 29 20.48 101.58 51.94 134.86 53.4 47.07 More broken 29.5 17.02 101.71 55.13 143.97 59.8 47.79 More broken 30 15.29 101.81 53.65 135.82 50.06 42.32 More broken 30.5 19.05 101.63 54.28 143.1 88.73 59.43 broken 31 22.29 101.7 51.03 135.04 47.14 45.31 More broken 31.5 19.79 101.96 52.17 139.12 49.24 45.01 More broken 32 23.59 102 57.61 147.46 36.75 41.56 More broken 32.5 18.1 101.73 55.14 141.09 35.82 37.93 More broken 33 19.92 102 55.36 147.97 72.42 54.83 More broken 33.5 26.03 102.24 54.72 136.01 75.04 59.45 broken 34 21.29 102.4 53.61 138.04 43.38 43.27 More broken 34.5 16.11 101.97 52.72 140.58 44.7 40.72 More broken 35 17.24 101.84 54.67 148.6 45.95 42.36 More broken 35.5 18.95 101.91 55.38 145.48 69.33 52.89 More broken 36 33.54 102.1 58.6 141.02 544.33 166.23 broken 36.5 32.35 101.88 66.28 148.88 573.5 171.27 broken 37 15.51 101.72 57.45 144.16 63.08 47.98 More broken 37.5 11.64 101.78 59.8 151.37 34.61 33.54 More broken 38 10.64 102.08 58.21 144.2 35.06 32.73 More broken 38.5 11.98 101.97 60.15 143.25 197.39 77.63 broken 39 217.77 101.84 63.72 142.3 1087.18 386.76 Clay filling 39.5 187.77 101.79 69.83 147.03 2893.91 600.63 Clay filling 40 138.19 101.59 67.81 145.84 10707.95 1032.73 Clay filling 40.5 16.46 101.51 61.78 135.48 750.77 160.09 broken 41 15.78 101.86 65.17 136.3 583.14 140.9 broken

[0148] It can be seen from Table 2 that the quantitative evaluation index of rock integrity in the IRFI interval proposed in the present invention can provide detailed and relatively accurate rock integrity evaluation results based on the advanced drilling data, and the prediction accuracy meets the actual engineering requirements.

[0149] The above description is only a preferred embodiment of the present invention and is 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 method for quantitatively evaluating the integrity of a tunnel rock mass, characterized in that: The following steps are involved: S1, obtain engineering data; S2, performing fine selection and / or secondary calculation on the engineering data to obtain sample data; the sample data includes five indicators, namely, propulsion speed mean DRM, propulsion force mean TPM, torque mean TM, rotation speed mean RM and propulsion speed variance DRV; S3, using sample data and XGBoost model, train a machine learning model for predicting the integrity of surrounding rock; S4, calculating the SHAP value of each indicator in the sample data, performing an interpretability analysis on the machine learning model for predicting the integrity of the surrounding rock, and determining the weight of each indicator; S5, using the weight of each indicator obtained in step S4, using the multivariate instability index analysis method to obtain a quantitative evaluation model for the integrity of the tunnel rock mass; the mathematical model D of the multivariate instability index analysis method t as follows: Among them, n is the number of influencing factors, d1~d n is the instability index factor, W1~W n is the instability index weight; The weight of each indicator is calculated according to the following formula: Among them, SHAP i Represents the SHAP value of the i-th indicator; The calculation formula IRFI of the tunnel rock mass integrity quantitative evaluation model is as follows: Among them, W1 is the weight of the indicator DRM, W2 is the weight of the indicator TPM, W3 is the weight of the indicator TM, W4 is the weight of the indicator RM, and W5 is the weight of the indicator DRV; S6, using the tunnel rock mass integrity quantitative evaluation model to evaluate the rock mass integrity of actual engineering data.

2. A method for quantitatively evaluating the integrity of a tunnel rock mass according to claim 1, characterized in that: The engineering data includes four primary indicators, namely propulsion speed, propulsion force, torque and rotation speed.

3. A method for quantitatively evaluating the integrity of a tunnel rock mass as claimed in claim 2, characterized in that: The step S2 includes performing refinement and / or secondary calculation on the engineering data, including one or more of the following steps: S21, outlier filtering; S22, segment the engineering data according to the same section distance, calculate the mean and variance of the four first-level indicators in each segmented section, and obtain eight second-level indicators; S23, performing oversampling processing on the secondary index; S23, performing feature correlation analysis on the primary indicators or the secondary indicators, and eliminating features with correlations higher than a first threshold; and / or, performing feature importance analysis on the primary indicators or the secondary indicators, and eliminating indicators with feature importance lower than a second threshold.

4. A method for quantitatively evaluating the integrity of a tunnel rock mass as claimed in claim 3, characterized in that: The first threshold is 0.

8.

5. A method for quantitatively evaluating the integrity of a tunnel rock mass as claimed in claim 3, characterized in that: The second threshold is 0.

1.

6. A method for quantitatively evaluating the integrity of a tunnel rock mass according to claim 1, characterized in that: The evaluation results of the tunnel rock mass integrity quantitative evaluation model in step S6 include: When DRM ≥ 120, the evaluation result is mud filling; When DRM<120 and the IRFI value range is (0,31), the evaluation result is relatively complete; When DRM<120 and the IRFI value range is [31,55), the evaluation result is relatively fragmented; When DRM<120 and IRFI≥55, the evaluation result is broken.

Images