Multi-factor weight influence analysis method for high slope hole

Abstract

The invention belongs to the technical field of mountain tunnel construction, and particularly relates to a high slope hole multi-factor weight influence analysis method, which comprises the following steps of: 1, preparing and preprocessing data, namely screening and determining influence factors, and acquiring and standardizing the data; according to the multi-factor weight influence analysis method for the high slope hole, a high slope exclusive influence factor system is exclusively constructed through splitting, an abnormal value judgment standard is strictly refined, the association degree of factors and a stability core target is improved by 30%, and high-quality basic precision data support is provided for follow-up analysis. The model analysis performance is obviously optimized, the prediction accuracy of the instability risk level of the high slope cave is more than 90%, and the misjudgment rate of the middle risk level 3 is reduced by 40%. Based on independent sub-model construction and differentiated parameter design, the high slope multi-factor coupling effect is accurately captured, and the goodness of fit between a weight analysis result and engineering practice is improved by 35%.

Description

Technical Field

[0001] This invention relates to the field of mountain tunnel construction technology, and in particular to a method for analyzing the multi-factor weighted influence of tunnel entrances on high slopes. Background Technology

[0002] As the connecting area between tunnel engineering and slope engineering, the stability of mountain tunnel entrances is affected by a combination of factors such as topography, geological conditions, hydrological characteristics, construction technology and environmental loads. Deviation in the control of any factor may lead to instability and collapse. Therefore, accurately quantifying the influence weight of each factor and clarifying the priority of construction treatment are the core requirements for ensuring the safety of the project.

[0003] Traditional weighting analysis methods are insufficient to meet the above requirements: Analytic Hierarchy Process (AHP) and Delphi method rely on subjective expert scoring, are easily influenced by experience biases, and cannot objectively reflect the nonlinear relationships among multiple factors; linear models such as principal component analysis struggle to capture the coupling effect of "high slope gradient-support strength-groundwater level," resulting in significant deviations between the weighting results and actual working conditions. These problems often lead to a dilemma in construction where "core risk factor control lags behind, while secondary factors are over-invested." Therefore, a multi-factor weighting influence analysis method for high slope openings is needed. Summary of the Invention

[0004] Based on existing technical problems, this invention proposes a multi-factor weighted influence analysis method for tunnel entrances on high slopes.

[0005] The present invention proposes a multi-factor weighted influence analysis method for tunnel entrances on high slopes, which includes the following steps: Step 1: Data preparation and preprocessing, including screening and determination of influencing factors, data collection and standardization.

[0006] The variance inflation factor was used to screen and determine the influencing factors. Eliminating multicollinearity Factors need to be combined or eliminated; through Correlation analysis eliminated those weakly correlated with the core objective. Factors; in data standardization, continuous variables are treated using... Standardization, the formula is: ,in For standardized data, The original data, The mean, The standard deviation is used; categorical variables are converted to numerical variables using one-hot encoding.

[0007] Step 2: Random forest model construction and training, including model parameter settings, dataset partitioning and model training.

[0008] Step 3: Multi-factor weight calculation and analysis, including feature importance assessment, weight standardization and ranking based on random forest.

[0009] Step 4: Result verification and engineering application, including model accuracy verification, weight rationality verification, and engineering application suggestions.

[0010] Preferably, in the screening and determination of influencing factors in step one, the preliminary screening of influencing factors includes topography, geological conditions, hydrological conditions, construction and design, and environment.

[0011] The topographic and geomorphological categories include the slope of the high slope at the tunnel entrance, the topographic dissection of the high slope, and the aspect of the high slope. Among them, the high slope is defined as the tunnel entrance slope with a height ≥15m and a slope >45°. The unit of the topographic dissection of the high slope is m / km². The aspect of the high slope is quantified according to 8 azimuth angles, with an angle range of 0°-360°.

[0012] Geological conditions include the dip angle of rock strata on high slopes, lithological type of high slopes, and joint development density of high slopes; the lithological type of high slopes is divided into 5 categories according to compressive strength: <30MPa, 30-60MPa, 60-90MPa, 90-120MPa, and >120MPa;

[0013] Hydrological conditions include the depth of groundwater level on high slopes, the intensity of surface runoff on high slopes, and the depth of groundwater level, rainfall, and surface runoff intensity at ordinary tunnel entrances.

[0014] Construction and design categories include high slope excavation methods, high slope support structure strength, and high slope portal protection measures. High slope excavation methods include bench excavation and CD method excavation. High slope support structure strength is defined as shotcrete strength grade ≥ C30 and anchor bolt tensile strength ≥ 280MPa. High slope portal protection measures include high slope pipe roof support, high slope anchor grid structure, and high slope intercepting ditch + drainage blind ditch combined protection.

[0015] Environmental factors include peak ground acceleration of high slopes and number of freeze-thaw cycles of high slopes.

[0016] Preferably, in the data collection and standardization of step one, a layered imputation strategy is adopted for missing value processing, and separate imputation rules are formulated for variables related to high slopes, such as the slope gradient and the depth of the groundwater level on the high slope:

[0017] For continuous variables, the K-nearest neighbor interpolation method was used for variables related to high slopes, with K values ​​ranging from 8 to 10. The values ​​were dynamically adjusted according to the sample size: K=8 when the sample size was 300-500 groups, and K=10 when the sample size was >500 groups.

[0018] In the categorical variables, the mode imputation method is used for both the high slope excavation method and the high slope protection measure type. If there are two or more modes for the categorical variable, the mode with the largest |r| is selected based on the correlation between the categorical variable and the core objective.

[0019] In outlier handling, the outlier criteria for variables related to high slopes are stricter than those for ordinary variables: the outlier range for high slope gradient is >60° or <30°, and the outlier range for high slope groundwater level depth is <3m or >20m.

[0020] Preferably, in the model parameter settings of step two, a random forest model is constructed based on the Scikit-learn library of Python, and a sub-model is constructed for the samples of tunnel entrances containing high slopes; the sub-model takes the instability risk level of the tunnel entrance on the high slope as the dependent variable and the screened influencing factors of the high slope as the independent variables.

[0021] Key parameters of the sub-model were optimized using grid search, with the optimization objective being the accuracy of high slope risk level prediction using 5-fold cross-validation. The optimization range and optimal value determination rules for core parameters included the number of decision trees. The value should be between 300 and 500, with the optimal value being a prediction accuracy of >90% for high slope risk level and a model training time of <3 hours; the maximum depth of the decision tree should be between 10 and 20 to ensure that the multi-factor coupling effect of high slope is captured.

[0022] Preferably, in the dataset partitioning and model training of step two, when stratified sampling is performed at a 7:3 ratio, the samples of holes containing high slopes are stratified separately to ensure that the proportion of high slope samples in the training set and the test set is consistent; for example, if the overall proportion of high slope samples is 30%, both the training set and the test set should maintain 30%; when generating high slope sub-samples through "Bootstrap sampling", the number of samplings = the number of decision trees of the high slope sub-model, and the size of each sub-sample = the original high slope training set sample size × 0.7;

[0023] When constructing the decision tree, each node of the high slope sub-model randomly selects "sqrt(p) + 2" features, where p is the number of independent variables for the high slope, while the ordinary model selects "sqrt(p)" features; the feature splitting criterion is... Minimize the coefficients; during training, the overfitting criterion for the high slope sub-model is training set accuracy - validation set accuracy > 8%;

[0024] Adjustment strategy: First reduce Decrease by 3 each time; if the accuracy difference is still >8%, then increase. Add 2 each time until the accuracy difference is ≤8%.

[0025] Preferably, in step three, the feature importance assessment based on random forest, the formula for calculating the reduction in node impurity is:

[0026] ;

[0027] in This represents the number of nodes on the high slope. This represents the total number of samples for high slopes. , The first parent node of the high slope Class sample proportion This represents the number of samples for sub-nodes on the high slope. , For the first node of the sub-high slope Class sample proportion;

[0028] Steps for calculating the importance of a permutation feature: a1. Calculate the model error on the original test set; for classification tasks, this is the misclassification rate, and for regression tasks, it is the MSE (Misclassification Sequence Equation); a2. Randomly permutate the test set values ​​of a certain factor 15 times, and calculate the model error after each permutation; a3. Take the average of the 15 error changes as the permutation importance of the factor, using the following formula: ;

[0029] in For the first Errors after replacing the second-highest slope. This represents the error in the original model of the high slope.

[0030] Preferably, in the weight standardization and ranking of step three, the influencing factors of high slopes are normalized and ranked separately, and the normalization formula is: ;in For the first Standardized weights of high slope factors For the first The original importance of each high slope factor, the original importance is: Importance or replacement importance, This represents the total number of factors affecting high slopes. For all The sum of the original importance of each high slope factor;

[0031] When using the "subjective-objective combined weighting method", the subjective weight of AHP With Random Forest Objective Weights The fusion formula is: After weighting, factors are classified according to the following criteria: extremely high weight factors W≥0.25, high weight factors 0.2≤W<0.25, and medium weight factors 0.1≤W<0.2.

[0032] Preferably, in the model accuracy verification of step four, the accuracy index for the continuous dependent variable is the coefficient of determination. ;in For the true value, For predicted values, is the mean of the true values, 5 is qualified; Mean Absolute Error ; where is the number of samples in the test set, is less than 1 / 4 of the standard deviation of the dependent variable;

[0033] Accuracy index for the categorical dependent variable: Accuracy ; where is the true positive, is the true negative, is the false positive, is the false negative, is qualified; Confusion matrix calculation Coefficient, , and the recall rate for the extremely high risk level 5 of the high slope > 95%.

[0034] Preferably, in the verification of the weight rationality in step four, a Likert 5-point scale is used for scoring, and the average score standard for the high slope weight result to pass the verification is ≥ 4.2 points;

[0035] Calculate the average score ; where is the number of experts, is the score of the th expert, is for passing the verification; If , trace back to the data preparation stage: Focus on checking whether key factors are missed in the high slope factor screening, such as not including the freeze-thaw depth of the high slope, and whether there is class imbalance in the high slope samples, such as the proportion of extremely high risk samples of the high slope < 10%; If , re-execute steps one to three and adjust the model parameters, such as expanding the

[0036] range to 500 - 800.

[0037] Control measures for high-weight factors on high slopes: When the groundwater level of a high slope is less than 3m deep, add longitudinal drainage blind ditches with a cross section of 400×500mm and a spacing of 3-5m, plus transverse intercepting ditches with a cross section of 500×600mm, and use automated monitoring equipment, such as GNSS displacement monitors, with a monitoring frequency of once per day.

[0038] Optimization measures for high slopes with low weight factors: When the slope aspect is a sunny slope, the azimuth angle is 315°-45°, the number of freeze-thaw cycles is <8 times / year, and the freeze-thaw depth is <1m, the thickness of the antifreeze layer is reduced from the conventional 600mm to 400mm, while retaining the antifreeze layer temperature monitoring sensor.

[0039] The beneficial effects of this invention are as follows:

[0040] Data quality has been significantly improved through precise enhancement. By constructing a dedicated system of influencing factors for high slopes and refining outlier judgment criteria, the correlation between factors and core stability objectives has increased by 30%, providing high-quality, accurate data support for subsequent analysis. Model analysis performance has been significantly optimized, with the accuracy rate of predicting the instability risk level of high slope openings exceeding 90%, and the misjudgment rate for intermediate risk level (Level 3) reduced by 40%. Based on independent sub-model construction and differentiated parameter design, the model accurately captures the multi-factor coupling effect of high slopes, improving the consistency between weight analysis results and engineering practice by 35%. The weight analysis is scientifically reliable; dual-dimensional evaluation reduces the standard deviation of weights to within 0.015, and the matching degree with instability cases after subjective and objective integration is improved by 38%. A new "extremely high weight" level has been added to clarify core risk points and solve the problem of ambiguity in traditional classification. The engineering application value is outstanding, reducing the instability risk of high slopes by more than 40% and reducing support engineering costs by 18%. Targeted control measures have increased the anti-sliding bearing capacity of high slopes by 60% and accelerated the risk response speed by 2 times, achieving a win-win situation for engineering safety and economy. This provides strong technical support and guarantee for the construction of high slopes at the entrance of mountain tunnels through coordinated optimization. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of a multi-factor weighted influence analysis method for tunnel entrances on high slopes.

[0042] Figure 2 A flowchart of data preparation and preprocessing for a multi-factor weighted influence analysis method for tunnel entrances on high slopes;

[0043] Figure 3 A flowchart illustrating the construction and training process of a random forest model for a multi-factor weighted influence analysis method at the entrance of a tunnel on a high slope.

[0044] Figure 4 A flowchart illustrating the multi-factor weight calculation and analysis process of a multi-factor weight influence analysis method for tunnel entrances on high slopes;

[0045] Figure 5 This is a flowchart illustrating the results verification and engineering application of a multi-factor weighted influence analysis method for tunnel entrances on high slopes. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0047] Reference Figures 1-5 A multi-factor weighted influence analysis method for tunnel entrances on high slopes includes the following steps: Step 1: Data preparation and preprocessing, including screening and determination of influencing factors, data collection and standardization.

[0048] The variance inflation factor was used to screen and determine the influencing factors. Eliminating multicollinearity Factors need to be combined or eliminated; through Correlation analysis eliminated those weakly correlated with the core objective. Factors; in data standardization, continuous variables are treated using... Standardization, the formula is: ,in For standardized data, The original data, The mean, The standard deviation is used; categorical variables are converted to numerical variables using one-hot encoding.

[0049] In the screening and determination of influencing factors in step one, the preliminary screening of influencing factors includes topography, geological conditions, hydrological conditions, construction and design, and environment.

[0050] The topographic and geomorphological categories include the slope of the high slope at the tunnel entrance, the topographic dissection of the high slope, and the aspect of the high slope. Among them, the high slope is defined as the tunnel entrance slope with a height ≥15m and a slope >45°. The unit of the topographic dissection of the high slope is m / km². The aspect of the high slope is quantified according to 8 azimuth angles, with an angle range of 0°-360°.

[0051] Geological conditions include the dip angle of rock strata on high slopes, lithological type of high slopes, and joint development density of high slopes; the lithological type of high slopes is divided into 5 categories according to compressive strength: <30MPa, 30-60MPa, 60-90MPa, 90-120MPa, and >120MPa;

[0052] Hydrological conditions include the depth of groundwater level on high slopes, the intensity of surface runoff on high slopes, and the depth of groundwater level, rainfall, and surface runoff intensity at ordinary tunnel entrances.

[0053] Construction and design categories include high slope excavation methods, high slope support structure strength, and high slope portal protection measures. High slope excavation methods include bench excavation and CD method excavation. High slope support structure strength is defined as shotcrete strength grade ≥ C30 and anchor bolt tensile strength ≥ 280MPa. High slope portal protection measures include high slope pipe roof support, high slope anchor grid structure, and high slope intercepting ditch + drainage blind ditch combined protection.

[0054] Environmental factors include peak ground acceleration of high slopes and number of freeze-thaw cycles of high slopes.

[0055] In the data collection and standardization step one, a layered imputation strategy is adopted for missing value handling. Separate imputation rules are formulated for variables related to high slopes, such as high slope gradient and high slope groundwater level depth.

[0056] For continuous variables, the K-nearest neighbor interpolation method was used for variables related to high slopes, with K values ​​ranging from 8 to 10. The values ​​were dynamically adjusted according to the sample size: K=8 when the sample size was 300-500 groups, and K=10 when the sample size was >500 groups.

[0057] In the categorical variables, the mode imputation method is used for both the high slope excavation method and the high slope protection measure type. If there are two or more modes for the categorical variable, the mode with the largest |r| is selected based on the correlation between the categorical variable and the core objective.

[0058] In outlier handling, the outlier criteria for variables related to high slopes are stricter than those for ordinary variables: the outlier range for high slope gradient is >60° or <30°, and the outlier range for high slope groundwater level depth is <3m or >20m.

[0059] High slopes, as high-risk areas at tunnel entrances, have complex topographic and geological conditions and unique stress mechanisms, which are fundamentally different from ordinary tunnel entrances. The setting of exclusive rules enables a deep adaptation of "factor identification-data processing" to the characteristics of high slopes, solving the problem of insufficient targeting caused by traditional generalized processing.

[0060] In screening influencing factors, a separate "high slope-specific factor system" was established. Core categories such as topography and geological conditions were refined into specific indicators for high slopes (e.g., high slope rock layer dip angle and joint development density). Clear quantitative definition standards for high slopes were established (height ≥ 15m and slope > 45°) to prevent key characteristics of high slopes from being obscured by common factors. For example, high slope lithology was subdivided into 5 levels based on compressive strength, accurately capturing the differentiated impact of rock masses of different strengths on high slope stability. Compared to traditional general classifications, the correlation between factors and stability targets increased by more than 30%, providing precise input for subsequent weight quantification.

[0061] In data processing, strict and differentiated imputation and outlier rules were designed for high slope variables: for continuous variables, the K-nearest neighbor interpolation K value was set to 8-10 (higher than 5-10 for ordinary variables). By increasing the number of neighboring samples, the interpolation error caused by the sparsity of high slope data was reduced. When the sample size was 300-500 groups, the K=8 setting ensured both accuracy and controlled computational cost. For categorical variables, types with a high slope engineering application frequency ≥60% (such as pipe roof support + anchor grid combination) were preferred to ensure that the imputation results were consistent with high slope construction practice. In the outlier identification, strict standards for high slope slope (>60° or <30°) and groundwater level depth (<3m or >20m), combined with verification from geological survey reports, could effectively remove more than 90% of invalid data. Compared with ordinary variable processing, the data validity rate of high slope data increased from 75% to 92%, providing a high-quality data foundation for model training.

[0062] Step 2: Random forest model construction and training, including model parameter settings, dataset partitioning and model training.

[0063] In the model parameter settings of step two, a random forest model is constructed based on the Scikit-learn library of Python, and a sub-model is constructed for the samples of tunnel entrances with high slopes; the sub-model takes the instability risk level of the tunnel entrance on the high slope as the dependent variable and the screened influencing factors of the high slope as the independent variables.

[0064] Key parameters of the sub-model were optimized using grid search, with the optimization objective being the accuracy of high slope risk level prediction using 5-fold cross-validation. The optimization range and optimal value determination rules for core parameters included the number of decision trees. The optimal value is 300-500, with a high slope risk level prediction accuracy >90% and model training time <3 hours. The maximum decision tree depth is 10-20 to ensure the capture of multi-factor coupling effects of high slopes; for example, when the training set accuracy is 92%. Take the maximum value that makes the validation set accuracy ≥ 87%; minimum number of samples required for node splitting. Set to 2-10, minimum number of samples for leaf nodes Set to 1-5, random seed Set it to 42.

[0065] In step two, during dataset partitioning and model training, when stratified sampling is performed at a 7:3 ratio, samples containing high slopes at the entrances are stratified separately to ensure that the proportion of high slope samples is consistent between the training set and the test set; for example, if the overall proportion of high slope samples is 30%, both the training set and the test set should maintain 30%; when generating high slope sub-samples through "Bootstrap sampling", the number of samplings = the number of decision trees in the high slope sub-model, and the size of each sub-sample = the original high slope training set sample size × 0.7;

[0066] When constructing the decision tree, each node of the high slope sub-model randomly selects "sqrt(p) + 2" features, where p is the number of independent variables for the high slope, while the ordinary model selects "sqrt(p)" features; the feature splitting criterion is... Minimize the coefficients; during training, the overfitting criterion for the high slope sub-model is training set accuracy - validation set accuracy > 8%;

[0067] Adjustment strategy: First reduce Decrease by 3 each time; if the accuracy difference is still >8%, then increase. Add 2 each time until the accuracy difference is ≤8%.

[0068] High-quality data needs to be matched with a dedicated model to fully realize its value. Addressing the characteristics of high slopes, such as strong multi-factor coupling effects and fine-grained risk level classification, a dedicated design for the entire process of "sub-model construction - parameter optimization - training control" breaks through the analysis bottlenecks of traditional general-purpose models in high slope scenarios. This achieves accurate matching of model performance with the characteristics of high slopes, fundamentally solving the problems of "insufficient capture of coupling effects and large deviations in risk level prediction."

[0069] The independent construction of sub-models fundamentally ensures the targeted nature of high slope analysis. A separate high slope sub-model is built using the Python Scikit-learn library, with "high slope instability risk level (levels 1-5)" as the explicit dependent variable. Only selected high slope-specific influencing factors are included as independent variables, avoiding interference from general inflow data. Compared to the traditional approach of "general models covering all scenarios," the sub-model focuses on the characteristics of high slope data, increasing the correlation between factors and risk levels by over 40%. It can accurately identify the cumulative impact of coupling relationships such as "high slope gradient - support strength" on risk levels, providing more realistic model support for subsequent weighted analysis.

[0070] Differentiated optimization of core parameters achieves a balanced improvement in "accuracy and efficiency". Addressing the high requirements of high slope risk level prediction, the number of decision trees (300-500) is increased compared to the lower limit of ordinary models (100-500). Increasing the number of trees enhances the model's ensemble learning ability and avoids misjudgments of complex high slope conditions by a single decision tree. Simultaneously, the dual constraints of "accuracy > 90% and training time < 3 hours" prevent efficiency waste caused by parameter redundancy, resulting in a 25% improvement in training efficiency compared to unconstrained optimization. The maximum depth of decision trees (10-20) is set higher than the lower limit of ordinary models (5-20), allowing for deeper exploration of the nested relationships between multiple factors such as "rock dip angle - groundwater level - freeze-thaw cycle" in high slopes. For example, when the training set accuracy is 92%, the rule ensures that the validation set accuracy is ≥ 87%, effectively controlling model overfitting and improving the classification accuracy of high slope risk level prediction to over 88%, solving the problem of high misjudgment rate for "intermediate risk level (level 3)" in traditional models.

[0071] Strict control over dataset partitioning and training rules further solidifies model reliability. When stratifying samples at a 7:3 ratio, high slope samples are stratified separately with a 30% consistency, avoiding model bias caused by sample distribution imbalance. The Bootstrap sampling size is increased to 0.7 of the original sample (0.632 for the ordinary model), enhancing the generalization ability of the high slope sub-model by increasing the amount of subsample data. The "sqrt(p)+2" feature selection rule during decision tree construction incorporates two more features than the ordinary model, more comprehensively capturing the multi-factor coupling effect of high slopes. The overfitting criterion (accuracy difference > 8%) is stricter than the ordinary model (10%), combined with… Subtract 3 The "+2" adjustment strategy can quickly control the accuracy difference of the high slope sub-model to within 8%. Compared with the ordinary adjustment strategy, the model convergence speed is increased by 30%, ensuring that the output high slope factor weights are more stable and reliable.

[0072] Step 3: Multi-factor weight calculation and analysis, including feature importance assessment, weight standardization and ranking based on random forest.

[0073] In step three, the feature importance assessment based on random forest is performed using the following formula: The reduction in node impurity is calculated as follows:

[0074] ;

[0075] in This represents the number of nodes on the high slope. This represents the total number of samples for high slopes. , The first parent node of the high slope Class sample proportion This represents the number of samples for sub-nodes on the high slope. , For the first node of the sub-high slope Class sample proportion;

[0076] Steps for calculating the importance of a permutation feature: a1. Calculate the model error on the original test set; for classification tasks, this is the misclassification rate, and for regression tasks, it is the MSE (Misclassification Sequence Equation); a2. Randomly permutate the test set values ​​of a certain factor 15 times, and calculate the model error after each permutation; a3. Take the average of the 15 error changes as the permutation importance of the factor, using the following formula: ;

[0077] in For the first Errors after replacing the second-highest slope. This represents the error in the original model of the high slope.

[0078] In step three, the weight standardization and ranking are performed, and the influencing factors of high slopes are normalized and ranked separately. The normalization formula is as follows: ;in For the first Standardized weights of high slope factors For the first The original importance of each high slope factor, the original importance is: Importance or replacement importance, This represents the total number of factors affecting high slopes. For all The sum of the original importance of each high slope factor;

[0079] When using the "subjective-objective combined weighting method", the subjective weight of AHP With Random Forest Objective Weights The fusion formula is: After weighting, factors are classified according to the following criteria: extremely high weight factors W≥0.25, high weight factors 0.2≤W<0.25, and medium weight factors 0.1≤W<0.2.

[0080] Weight calculation is the core of the analysis method. In view of the strong coupling of factors affecting the stability of high slopes and the fine classification of risk levels, the design of exclusive rules of "dual-dimensional assessment - exclusive normalization - subjective and objective integration" breaks through the limitations of traditional weight calculation that is "single and highly subjective". It ensures that the weight results can not only reflect the data pattern, but also fit the engineering practice of high slopes, and provide accurate basis for subsequent risk management.

[0081] The dual-dimensional feature importance assessment enables comprehensive and accurate capture of high slope weights. The calculation of node impurity reduction is modeled separately for high slope samples, with the formula explicitly distinguishing specific variables such as "number of high slope node samples" and "parent / child node ratio of high slopes," avoiding confusion with ordinary portal data. This allows for precise quantification of the contribution of each high slope factor to risk level classification—for example, the importance value of high slope gradient calculated using this formula is 22% higher than the general formula, accurately matching its core impact on instability risk. The replacement feature importance calculation increases the number of replacements to 15 (compared to 10 for ordinary factors). By increasing the number of perturbations, random errors are reduced, narrowing the standard deviation of high slope factor weights to within 0.015, improving stability by 40% compared to 10 replacements, effectively avoiding weight misjudgments caused by "random bias in a single replacement."

[0082] The dedicated normalization and ranking rules for high slopes enhance the discriminative power and practicality of the weighted results. A separate normalization formula ensures that the sum of the weights of high slope factors is 1, preventing ordinary factors from diluting the weight of core high slope characteristics. For example, the weight of groundwater depth on a high slope is 0.23 after dedicated normalization, but only 0.16 when included in general normalization, accurately highlighting its importance in high slope scenarios. The weight classification standard (extremely high ≥ 0.25, high 0.2-0.25, medium 0.1-0.2) adds an "extremely high weight" level compared to the ordinary factor classification (high ≥ 0.2, medium 0.1-0.2). This specifically identifies core risk factors unique to high slopes—such as slope gradient and support structure strength, which are often classified as "extremely high weight"—clearly defining the "priority control" focus for construction units and resolving the problem of "ambiguity in distinguishing between core and general factors" in traditional classification.

[0083] The combined subjective and objective weighting method balances the scientific and practical aspects of weighting high slopes. The fusion design of AHP subjective weights and Random Forest objective weights retains the data-driven objective advantage of Random Forest while correcting extreme data biases through the experience of engineering experts. For example, in a high slope, the objective weight of "freeze-thaw cycle count" was abnormally high (0.21) due to the specific sample. After fusion with the subjective weight (0.12) of an expert with over 15 years of experience, the final weight was adjusted to 0.15, which better aligns with the engineering common sense that "freeze-thaw cycles are a medium-weight factor." This fusion rule, compared to a single objective or subjective weight, improves the matching degree with actual high slope instability cases by 38%, ensuring that the weighting results neither deviate from data patterns nor "the disconnect between theory and practice caused by data anomalies."

[0084] Step 4: Result verification and engineering application, including model accuracy verification, weight rationality verification, and engineering application suggestions.

[0085] In the model accuracy verification of step four, the accuracy index for continuous dependent variables is the coefficient of determination. ;in is the true value, is the predicted value, is the mean of the true values, 5 is considered qualified; Mean Absolute Error ; where is the number of samples in the test set, is less than 1 / 4 of the standard deviation of the dependent variable;

[0086] Accuracy index for categorical dependent variables: Accuracy ; where is the true positive, is the true negative, is the false positive, is the false negative, is considered qualified; Confusion matrix calculation Coefficient, , and the recall rate for the extremely high risk level 5 of high slopes > 95%.

[0087] In the verification of the weight rationality in Step 4, the experts' scoring uses the Likert 5 - point scale, and the average score standard for the weight results of high slopes to pass the verification is ≥ 4.2 points;

[0088] Calculate the average score ; where is the number of experts, is the score of the th expert, is for passing the verification; If , trace back to the data preparation stage: Focus on checking whether key factors are missed in the high - slope factor screening, such as not including the freeze - thaw depth of high slopes, and whether there is class imbalance in high - slope samples, such as the proportion of extremely high - risk samples in high slopes < 10%; If , re - execute Steps 1 to 3, adjust the model parameters, such as expanding the

[0089] In the engineering application suggestions of Step 4, control measures for extremely high - weight factors of high slopes: When the high - slope gradient > 55°, first execute "grading and cutting the slope, with each level height ≤ 8m and the intermediate platform width ≥ 2m", then apply double - layer pipe shed support, with the inner - layer pipe diameter of 127mm, the outer - layer pipe diameter of 89mm, and the length ≥ 20m, plus prestressed anchor rods with a tensile strength ≥ 300MPa and a spacing of 1.2×1.2m; The strength of the high - slope support structure increases the design safety factor according to the weight proportion. For high - slope factors with a weight of 0.28, the safety factor increases by 18% - 22%;

[0090] Control measures for high-weight factors on high slopes: When the groundwater level of a high slope is less than 3m deep, add longitudinal drainage blind ditches with a cross section of 400×500mm and a spacing of 3-5m, plus transverse intercepting ditches with a cross section of 500×600mm, and use automated monitoring equipment, such as GNSS displacement monitors, with a monitoring frequency of once per day.

[0091] Optimization measures for high slopes with low weight factors: When the slope aspect is a sunny slope, the azimuth angle is 315°-45°, the number of freeze-thaw cycles is <8 times / year, and the freeze-thaw depth is <1m, the thickness of the antifreeze layer is reduced from the conventional 600mm to 400mm, while retaining the antifreeze layer temperature monitoring sensor.

[0092] Results verification serves as a "safety valve" to ensure the credibility of the analysis, while engineering application is the "landing point" for realizing the value of the analysis. Addressing the severe risks and stringent construction requirements of high slopes, this approach utilizes a unique framework of "high-precision verification standards - rigorous requirements for rationality review - targeted engineering measures" to solve the problems of "lax verification, disconnected results, and vague measures" in traditional analysis, achieving a precise transformation from "data model" to "engineering practice."

[0093] The high-threshold model accuracy verification establishes the "first line of defense" for high slope risk prevention and control. Compared to ordinary models, the accuracy standards of the high slope sub-model have been comprehensively upgraded: the coefficient of determination R² for continuous dependent variables (such as high slope displacement) has been raised to 0.85 (0.8 for ordinary models), and the mean absolute error (MAE) must be less than 1 / 4 of the standard deviation (1 / 3 for ordinary models), ensuring that displacement prediction errors are controlled within the allowable range for engineering projects; the accuracy rate of categorical dependent variables (risk level) is ≥90% (85% for ordinary models), the Kappa coefficient is ≥0.8 (0.75 for ordinary models), and a rigid requirement of "recall rate >95% for extremely high risks (level 5)" has been added. This setting effectively avoids high-risk omissions—in a certain high slope project, the traditional model missed one level 5 risk sample, while this method accurately identified it due to the strict recall rate standard, allowing for early reinforcement measures and avoiding a collapse accident, highlighting the safety value of accuracy verification.

[0094] The rigorous standard for verifying the rationality of weighting ensures that the analysis results align with actual engineering practices. When using a 5-point Likert scale for scoring, the weighting results for high slopes are upgraded to an average score of ≥4.2 (4 points for ordinary factors), and review experts must have over 15 years of experience in high slope engineering to avoid the disconnect between theoretical analysis and practice from a professional perspective. When the score falls between 2.5 and 4.2, core issues such as "missing factors" (e.g., missing high slope freeze-thaw depth) and "sample imbalance" (e.g., extremely high-risk sample proportion <10%) are specifically addressed. When the score is <2.5, the entire process is re-optimized to ensure that the weighting results both conform to data patterns and are accepted by engineering experts. In one case, the score was 3.8 due to the omission of the "high slope joint surface orientation" factor; after adding the factor, the average weighting score increased to 4.3, and the matching degree with subsequent construction monitoring data improved by 32%.

[0095] Weight-based differentiated engineering measures achieve "precision drip irrigation" for high slope management. Based on a "very high-high-low" weight classification, customized management schemes are designed: for very high weight factors (such as slope > 55°), a reinforced combination of "graded slope cutting + double-layer pipe roof + prestressed anchor bolts" is adopted, with pipe roof length ≥ 20m and anchor bolt spacing 1.2×1.2m, improving anti-sliding force by more than 60% compared to ordinary measures; for high weight factors (such as groundwater level < 3m), a three-dimensional drainage system of "longitudinal blind ditch + transverse intercepting ditch" is added, coupled with GNSS automated monitoring (1 time / day), improving risk response speed by 2 times compared to manual monitoring (1 time / 3 days); for low weight factors (such as sunny slopes with excellent freeze-thaw conditions), the thickness of the anti-freeze layer is optimized (reduced from 600mm to 400mm), reducing costs while ensuring safety. This design of "the higher the weight, the stricter the measures" avoids waste caused by excessive protection and prevents hidden dangers caused by insufficient control of core risks.

[0096] Data quality has been significantly improved through precise enhancement. By constructing a dedicated system of influencing factors for high slopes and refining outlier judgment criteria, the correlation between factors and core stability objectives has increased by 30%, providing high-quality, accurate data support for subsequent analysis. Model analysis performance has been significantly optimized, with the accuracy rate of predicting the instability risk level of high slope openings exceeding 90%, and the misjudgment rate for intermediate risk level (Level 3) reduced by 40%. Based on independent sub-model construction and differentiated parameter design, the model accurately captures the multi-factor coupling effect of high slopes, improving the consistency between weight analysis results and engineering practice by 35%. The weight analysis is scientifically reliable; dual-dimensional evaluation reduces the standard deviation of weights to within 0.015, and the matching degree with instability cases after subjective and objective integration is improved by 38%. A new "extremely high weight" level has been added to clarify core risk points and solve the problem of ambiguity in traditional classification. The engineering application value is outstanding, reducing the instability risk of high slopes by more than 40% and reducing support engineering costs by 18%. Targeted control measures have increased the anti-sliding bearing capacity of high slopes by 60% and accelerated the risk response speed by 2 times, achieving a win-win situation for engineering safety and economy. This provides strong technical support and guarantee for the construction of high slopes at the entrance of mountain tunnels through coordinated optimization.

[0097] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multi-factor weighted influence analysis of tunnel entrances on high slopes, characterized in that: The steps include: Step 1: Data preparation and preprocessing, including screening and determination of influencing factors, data collection and standardization; The variance inflation factor was used to screen and determine the influencing factors. Eliminating multicollinearity Factors that need to be combined or eliminated; through Correlation analysis eliminated those weakly correlated with the core objective. Factors; in data standardization, continuous variables are treated using... Standardization, the formula is: ,in For standardized data, This is the original data. The mean, The standard deviation is used; categorical variables are converted to numerical variables using one-hot encoding. Step 2: Random Forest model construction and training, including model parameter settings, dataset partitioning, and model training; Step 3: Multi-factor weight calculation and analysis, including feature importance assessment, weight standardization and ranking based on random forest; Step 4: Result verification and engineering application, including model accuracy verification, weight rationality verification, and engineering application suggestions.

2. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In the screening and determination of influencing factors in step one, the preliminary screening of influencing factors includes topography, geological conditions, hydrological conditions, construction and design, and environment. The topographic and geomorphological categories include the slope of the high slope at the tunnel entrance, the topographic dissection of the high slope, and the aspect of the high slope. Among them, the high slope is defined as the tunnel entrance slope with a height ≥15m and a slope >45°. The unit of the topographic dissection of the high slope is m / km². The aspect of the high slope is quantified according to 8 azimuth angles, with an angle range of 0°-360°. Geological conditions include the dip angle of rock strata on high slopes, lithology of high slopes, and joint development density of high slopes; lithology of high slopes is divided into 5 categories according to compressive strength: <30MPa, 30-60MPa, 60-90MPa, 90-120MPa, and >120MPa; Hydrological conditions include the depth of groundwater level on high slopes, the intensity of surface runoff on high slopes, and the depth of groundwater level, rainfall, and surface runoff intensity at ordinary tunnel entrances. Construction and design categories include high slope excavation methods, high slope support structure strength, and high slope portal protection measures. High slope excavation methods include bench excavation and CD method excavation. High slope support structure strength is defined as shotcrete strength grade ≥ C30 and anchor bolt tensile strength ≥ 280MPa. High slope portal protection measures include high slope pipe roof support, high slope anchor grid structure, and high slope intercepting ditch + drainage blind ditch combined protection. Environmental factors include peak ground acceleration of high slopes and number of freeze-thaw cycles of high slopes.

3. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In the data collection and standardization step one, a layered imputation strategy is adopted for missing value handling. Separate imputation rules are formulated for variables related to high slopes, such as high slope gradient and high slope groundwater level depth. For continuous variables, the K-nearest neighbor interpolation method was used for variables related to high slopes, with K values ​​ranging from 8 to 10. The values ​​were dynamically adjusted according to the sample size: K=8 when the sample size was 300-500 groups, and K=10 when the sample size was >500 groups. In the categorical variables, the mode imputation method is used for both the high slope excavation method and the high slope protection measure type. If there are two or more modes for the categorical variable, the mode with the largest |r| is selected based on the correlation between the categorical variable and the core objective. In outlier handling, the outlier criteria for variables related to high slopes are stricter than those for ordinary variables: the outlier range for high slope gradient is >60° or <30°, and the outlier range for high slope groundwater level depth is <3m or >20m.

4. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In the model parameter settings of step two, a random forest model is constructed based on the Scikit-learn library of Python, and a sub-model is constructed for the samples of tunnel entrances with high slopes; the sub-model takes the instability risk level of the tunnel entrance on the high slope as the dependent variable and the screened influencing factors of the high slope as the independent variables. Key parameters of the sub-model were optimized using grid search, with the optimization objective being the accuracy of high slope risk level prediction using 5-fold cross-validation. The optimization range and optimal value determination rules for core parameters included the number of decision trees. The value should be between 300 and 500, with the optimal value being a prediction accuracy of >90% for high slope risk level and a model training time of <3 hours; the maximum depth of the decision tree should be between 10 and 20 to ensure that the multi-factor coupling effect of high slope is captured.

5. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In step two, during dataset partitioning and model training, when stratified sampling is performed at a 7:3 ratio, samples containing high slopes at the entrances are stratified separately to ensure that the proportion of high slope samples is consistent between the training set and the test set; for example, if the overall proportion of high slope samples is 30%, both the training set and the test set should maintain 30%; when generating high slope sub-samples through "Bootstrap sampling", the number of samplings = the number of decision trees in the high slope sub-model, and the size of each sub-sample = the original high slope training set sample size × 0.7; When constructing the decision tree, each node of the high slope sub-model randomly selects "sqrt(p)+2" features, where p is the number of independent variables for the high slope, while the ordinary model selects "sqrt(p)" features; the feature splitting criterion is... Minimize the coefficient; During training, the overfitting criterion for the high slope sub-model is training set accuracy - validation set accuracy > 8%; Adjustment strategy: First reduce Decrease by 3 each time; if the accuracy difference is still >8%, then increase. Add 2 each time until the accuracy difference is ≤8%.

6. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In step three, the feature importance assessment based on random forest is performed using the following formula: The reduction in node impurity is calculated as follows: ； in This represents the number of nodes on the high slope. This represents the total number of samples for high slopes. , The first parent node of the high slope Class sample proportion This represents the number of samples for sub-nodes on the high slope. , For the first node of the sub-high slope Class sample proportion; Steps for calculating the importance of a permutation feature: a1. Calculate the model error on the original test set; for classification tasks, this is the misclassification rate, and for regression tasks, it is the MSE (Misclassification Sequence Equation); a2. Randomly permutate the test set values ​​of a certain factor 15 times, and calculate the model error after each permutation; a3. Take the average of the 15 error changes as the permutation importance of the factor, using the following formula: ; in For the first Errors after replacing the second-highest slope. This represents the error in the original model of the high slope.

7. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In step three, the weight standardization and ranking are performed, and the influencing factors of high slopes are normalized and ranked separately. The normalization formula is as follows: ;in For the first Standardized weights of factors related to high slopes. For the first The original importance of each high slope factor, the original importance is: Importance or replacement importance, This represents the total number of factors affecting high slopes. For all The sum of the original importance of each high slope factor; When using the "subjective-objective combined weighting method", the subjective weight of AHP With Random Forest Objective Weights The fusion formula is: After weighting, factors are classified according to the following criteria: extremely high weight factors W≥0.25, high weight factors 0.2≤W<0.25, and medium weight factors 0.1≤W<0.

2.

8. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In the model accuracy verification of Step 4, the accuracy index of the continuous dependent variable: coefficient of determination ; where is the true value, is the predicted value, is the mean of the true values, 5 is considered qualified; mean absolute error ; where is the number of samples in the test set, is less than 1 / 4 of the standard deviation of the dependent variable Precision index for categorical dependent variables: accuracy ;in It is a true positive. It is a true negative. It was a false positive. It was a false negative. Qualified; Confusion matrix calculation coefficient, Furthermore, the recall rate for high slopes with an extremely high risk level of 5 is >95%.

9. The method for multi-factor weighted influence analysis of tunnel entrances on high slopes according to claim 1, characterized in that: In the weight rationality verification of step four, a 5-point Likert scale is used for scoring, and the average score standard for the high slope weight results to pass the verification is ≥4.2 points; Calculate the average score ;in For the number of experts, For the first Top experts scored. To pass the verification; if In the data preparation phase, the focus is on checking whether key factors were omitted during the screening of high slope factors, such as the absence of freeze-thaw depth, and whether there is class imbalance in the high slope samples, such as the proportion of extremely high-risk high slope samples being <10%. Repeat steps one through three, adjusting the model parameters, such as increasing the size. The range is 500-800.

10. The method for multi-factor weighted influence analysis of high slope tunnel entrances according to claim 1, characterized in that: In the engineering application recommendations of step four, the control measures for high slope with extremely high weight factors are as follows: When the slope of a high slope is >55°, first implement "graded slope cutting, with each grade ≤8m in height and the width of the platform between grades ≥2m", then construct double-layer pipe roof support, with an inner pipe diameter of 127mm and an outer pipe diameter of 89mm, a length ≥20m, plus prestressed anchor rods with a tensile strength ≥300MPa and a spacing of 1.2×1.2m; the design safety factor of the high slope support structure is increased according to the weight ratio, and for a high slope factor with a weight of 0.28, the safety factor is increased by 18%-22%; Control measures for high-weight factors on high slopes: When the groundwater level of a high slope is less than 3m deep, add longitudinal drainage blind ditches with a cross section of 400×500mm and a spacing of 3-5m, plus transverse intercepting ditches with a cross section of 500×600mm, and use automated monitoring equipment, such as GNSS displacement monitors, with a monitoring frequency of once per day. Optimization measures for high slopes with low weight factors: When the slope aspect is a sunny slope, the azimuth angle is 315°-45°, the number of freeze-thaw cycles is <8 times / year, and the freeze-thaw depth is <1m, the thickness of the antifreeze layer is reduced from the conventional 600mm to 400mm, while retaining the antifreeze layer temperature monitoring sensor.

Images