A method and system for predicting tunnel water inflow intervals based on physical residual adaptive constraints and seepage mechanism-guided conformal calibration.

Abstract

This invention relates to the field of intelligent prediction and hydrogeological risk prevention and control in tunnel engineering. Specifically, it discloses a method and system for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanisms. This system integrates multi-source engineering data and seepage analysis mechanisms to construct a full-process intelligent prediction system comprising "spatiotemporal decoupling characteristic engineering – physical residual adaptive constraints – geological grouping seepage mechanism-guided conformal calibration – PDP sensitivity analysis and dual risk assessment." It is applicable to high-precision interval prediction and construction safety management of water inflow in mountain tunnels, urban tunnels, and deeply buried underground engineering projects. This invention solves the problems of poor physical interpretability of pure data-driven models, inability to adaptively adjust physical constraint strength, insufficient reliability of prediction intervals, and a single risk assessment dimension in traditional tunnel water inflow prediction, significantly improving the accuracy, generalization ability, physical rationality, and engineering practicality of water inflow prediction.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction and hydrogeological risk prevention and control technology for tunnel engineering. Specifically, it discloses a method and system for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism. Background Technology

[0002] In tunnel construction, water inflow is a key technical indicator for evaluating the water-bearing capacity of surrounding rock, preventing sudden water and mud inrush disasters, determining the scale of drainage systems, and guiding construction safety management. To predict water inflow, existing technologies have disclosed machine learning prediction methods. However, the physical constraints applied by these methods generally employ fixed-weight regularization, failing to adaptively adjust the constraint strength according to changes in geological conditions. This results in excessively strong constraints in highly heterogeneous strata, suppressing the model's expressive ability, and insufficient constraints in homogeneous strata, leading to physical anomalies. Under the same geological conditions, phenomena such as increased water inflow with increased grouting ring thickness and decreased water inflow with increased hydraulic head height, which violate the basic laws of seepage mechanics, severely limit the widespread application of these models in engineering projects.

[0003] Existing methods for predicting water inflow primarily rely on single-point numerical outputs, lacking the ability to quantify uncertainty and provide reliable interval predictions, thus failing to support risk-based risk management. Conventional models only provide a single predicted water inflow value, neglecting uncertainties arising from data dispersion, model errors, and operational variability. Some quantile prediction methods generate intervals that are not effectively calibrated, resulting in significant deviations between actual coverage and the set confidence levels. Intervals that are too wide lose their engineering guidance value, while intervals that are too narrow fail to provide early warning. While standard conformal prediction does not rely on distribution assumptions, its inconsistency measurement is based solely on prediction error, ignoring the degree of deviation between the predicted results and physical laws. This leads to narrow prediction intervals even on physically inconsistent samples, resulting in false reliability due to "narrow but inaccurate" intervals. Current technologies cannot achieve highly reliable interval predictions and uncertainty assessments, failing to meet the practical needs of drainage system design, construction plan development, and risk level classification.

[0004] Most predictive models are black-box models, lacking interpretability, visualization, and engineering applicability, and their risk assessment dimensions are too narrow. Existing methods only classify risk levels based on the width of the prediction interval, failing to incorporate the degree of deviation between the predicted trend and physical laws into the risk assessment criteria. When the model's predicted trend contradicts physical laws but the interval is narrow, traditional methods may misjudge it as low risk, leading to engineering decision-making errors. The models are only applicable to working conditions similar to the training samples, and their generalization ability under different geological conditions and cross-sectional forms is limited. They cannot provide technical support for dynamic design, construction parameter optimization, and early risk prediction, and overall remain in a passive prediction state, making it difficult to achieve proactive risk prevention and control. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a method and system for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanisms. By constructing a spatiotemporally decoupled feature engineering system, introducing a physical residual adaptive constraint mechanism and a segmented multi-analytical solution constraint strategy, and combining this with conformal calibration guided by geological grouping seepage mechanisms, highly reliable interval prediction is achieved. Furthermore, a dual risk cross-judgment is implemented, thereby improving prediction accuracy, physical rationality, and engineering practicality, providing technical support for water inflow risk prevention and control in tunnel engineering.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanisms, comprising the following steps: Acquire multi-source tunnel engineering data; Missing value filling, outlier removal and standardization processing are performed on multi-source tunnel engineering data; Perform spatiotemporal decoupling feature engineering to determine time delay parameters by conducting Granger causality tests on the time series of rainfall and water inflow. We constructed time-delay weighted cumulative rainfall characteristics and attenuation coefficient weighted effective rainfall characteristics, constructed dimensionless derivative characteristics of seepage mechanics and geological heterogeneity factors, and used Spearman correlation coefficient for feature screening and correlation analysis. In the CatBoost gradient boosting tree framework, a joint loss function consisting of a data fitting loss term and a physical residual penalty term is constructed. A constraint weight adjustment factor that is adaptively adjusted according to the formation heterogeneity is introduced. A partitioned and segmented multi-analytical solution constraint strategy is adopted. The model training and hyperparameter optimization are completed by combining grid search and early stopping strategies. Based on the ternary water inflow prediction value output by the trained optimal model, a basic prediction interval is formed. The validation set samples are cross-grouped according to the hydrogeological dimension. Within each geological group, the seepage consistency score of the fused physical residual is calculated as a non-consistency measure of conformal calibration to achieve prediction interval calibration of target coverage probability. Based on the prediction interval width and physical consistency index, dual risk cross-judgment is performed and risk levels are divided to form intelligent construction decision suggestions.

[0007] As a further technical solution, the time-delay weighted cumulative rainfall characteristics The construction method is In the formula τ max The significant lag days determined by the Granger causality test, weighted RF t-i The rainfall on day ti; the attenuation coefficient weighted effective rainfall characteristic. λ is the attenuation coefficient, and i and j are both non-negative integers.

[0008] As a further technical solution, the dimensionless derived features of seepage mechanics include thickness / relative diameter, hydraulic head / relative diameter, and fault / relative diameter; wherein thickness / relative diameter is the ratio of grouting ring thickness to tunnel diameter, hydraulic head / relative diameter is the ratio of hydraulic head height to tunnel diameter, and fault / relative diameter is the ratio of fault fracture zone length to tunnel diameter, which can eliminate the influence of dimensional differences and size; the geological heterogeneity factor is the coefficient of variation of permeability coefficient. In the formula σ k Let μ be the standard deviation of the permeability coefficient along the tunnel mileage direction. K This is the mean permeability coefficient, used to describe the degree of formation heterogeneity.

[0009] As a further technical solution, the joint loss function is: In the formula For data fitting loss, To predict water inflow using a neural network model, To measure the actual water inflow on-site, This is a penalty term for physical residuals. Includes the geological and hydrological parameters (permeability coefficient, hydraulic head, tunnel geometry, etc.) required for the analytical solution of water inflow. The constraint weight adjustment factor In the formula This is the upper limit of the weight. It is the attenuation constant. The critical threshold for heterogeneity is when the coefficient of variation of the permeability is... As λ increases, λ decreases to achieve an adaptive match between the strength of physical constraints and the complexity of the formation. Homogeneous formations strengthen physical constraints to ensure physical consistency, while heterogeneous formations relax constraints to retain data-driven flexibility.

[0010] As a further technical solution, the physical residual penalty term It includes a three-stage physical consistency penalty: a penalty for the monotonically increasing permeability coefficient and inflow rate, a penalty for the monotonically decreasing grouting ring thickness and inflow rate, and a penalty for the monotonically increasing head height and inflow rate. In each gradient improvement iteration, the violation ratio of the current model to the above monotonicity conditions is calculated, and the violation ratio is weighted and included in the physical residual term, forcing the model output trend to conform to the basic laws of seepage mechanics.

[0011] As a further technical solution, the segmented and multi-analytical solution constraint strategy is as follows: in the homogeneous and intact section, the Dupuit steady-flow analytical solution is used as the constraint benchmark; in the fault fracture zone section, the Goodman inflow formula is used as the constraint benchmark; and in the karst development section, the mirror method superposition solution is used as the constraint benchmark. Physical constraints corresponding to the analytical solution are applied independently to each section to ensure the geographical relevance and engineering applicability of the constraints.

[0012] As a further technical solution, the calculation method of the seepage consistency score (SCS) is as follows: In the formula This refers to the seepage consistency score of the xth sample; The predicted water inflow is given by the 0.5 quantile for the x-th sample. Let x be the measured inflow rate of the x-th sample. The physical consistency deviation score for the x-th sample is set to 1 when the predicted inflow trend is inconsistent with the analytical solution trend of seepage, and 0 when they are consistent; α and β are preset weighting coefficients; within each geological group, the seepage consistency scores of the n validation samples in that group are sorted from smallest to largest, and the x-th score is selected. The interval expansion corresponding to each seepage consistency score is used as the calibration offset for that group. The offset is added to the lower and upper limits of the basic prediction interval to obtain the calibrated prediction interval, thus realizing the conformal interval calibration of the group.

[0013] As a further technical solution, the hydrogeological dimension cross-grouping is based on three dimensions: surrounding rock grade, permeability coefficient range, and water abundance. Each verification sample belongs to a unique group, and each group independently calculates the calibration offset to ensure the fairness and reliability of the prediction range under different geological conditions.

[0014] As a further technical solution, the dual risk cross-determination is based on the prediction interval width. Physical consistency index The cross matrix: where , The 0.9 and 0.1 quantiles of the predicted inflow are used as follows: when the PW is narrow and the PCI is high, it is considered low risk with high certainty and construction can proceed directly; when the PW is wide but the PCI is high, it is considered medium risk, requiring enhanced monitoring but the predicted trend is reliable; when the PW is narrow but the PCI is low, it is considered high risk, although the predicted range is narrow, it contradicts physical laws and needs to be re-verified; when the PW is wide and the PCI is low, it is considered extremely high risk, and construction should be suspended for supplementary investigation.

[0015] Secondly, the present invention also provides a tunnel water inflow interval prediction system based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism, comprising: The data preprocessing module is configured to acquire multi-source tunnel engineering data and perform missing value filling, outlier removal, and standardization on the multi-source tunnel engineering data. The spatiotemporal decoupling feature engineering module is configured to perform spatiotemporal decoupling feature engineering, perform Granger causality tests on the time series of rainfall and water inflow to determine time lag parameters, construct time lag weighted cumulative rainfall features and attenuation coefficient weighted effective rainfall features, construct seepage mechanics dimensionless derived features and geological heterogeneity factors, and perform feature screening and correlation analysis through Spearman correlation coefficient. The physical residual adaptive constraint model training module is configured to construct a joint loss function consisting of a data fitting loss term and a physical residual penalty term within the CatBoost gradient boosting tree framework. It introduces a constraint weight adjustment factor λ that is adaptively adjusted according to the formation heterogeneity, adopts a partitioned and segmented multi-analytical solution constraint strategy, and combines grid search and early stopping strategies to complete model training and hyperparameter optimization. The geological grouping seepage mechanism guides the conformal calibration module; it is configured to form a basic prediction interval based on the ternary inflow prediction value output by the trained optimal model, cross-group the validation set samples according to the hydrogeological dimension, calculate the seepage consistency score of the fused physical residual within each geological group as the non-consistency measure of conformal calibration, and realize the prediction interval calibration of the target coverage probability. The PDP sensitivity analysis and dual risk assessment module is configured to support the rendering of correlation heatmaps, partial dependency graphs, and prediction interval strip plots, and to perform parameter sensitivity back-calculation optimization. The dual risk assessment and decision-making module automatically classifies four risk levels based on the cross matrix of the prediction interval width and the physical consistency index PCI, and outputs construction safety suggestions, drainage configuration references, and risk warning conclusions, ultimately forming an intelligent prediction system that can be used for dynamic decision-making in tunnel engineering construction.

[0016] The beneficial effects of this invention are as follows: This invention achieves standardized processing of multi-source engineering data and systematic construction of spatiotemporal decoupling features, fundamentally solving the problems of single feature dimensions and lack of time delay effects; it deeply integrates physical residual adaptive constraints with a segmented and multi-analytical solution strategy to solve the core problems of purely data-driven models being physically unreasonable and constraint strengths being unable to be adaptively adjusted; it constructs a conformal calibration mechanism guided by a geological grouping seepage mechanism, fundamentally improving the ability to predict high-reliability intervals with a 90% confidence level and quantify uncertainty; it establishes a dual-risk cross-judgment mechanism, significantly improving interpretability, engineering implementation capability, and decision-making safety; and it enhances the ability to reverse-engineer construction parameters and proactively control and prevent decision-making, comprehensively transforming the traditional passive prediction model.

[0017] This invention acquires multi-source tunnel engineering data, performs missing value filling, outlier removal, and data standardization, ensuring that data of different dimensions and magnitudes have unified input conditions. Based on this, Granger causality tests are used to determine the time lag parameter of rainfall on water inflow, constructing time-lag weighted cumulative rainfall features and attenuation coefficient weighted effective rainfall features to effectively capture the hysteretic response law between rainfall and water inflow. Dimensionless derived features of seepage mechanics are constructed to eliminate the scale effect caused by tunnel cross-sectional dimensions. The permeability coefficient variation coefficient is introduced as a geological heterogeneity factor, providing a quantitative basis for subsequent adaptive constraints. Spearman correlation analysis is used to identify redundant features and verify correlation directions, automatically removing redundant features, ultimately forming a standardized, systematic, and physically driven high-quality input feature set.

[0018] This invention constructs a joint loss function based on the CatBoost algorithm, organically unifying the data fitting loss and the physical residual penalty term to achieve collaborative training of data-driven and physical priors. An adaptive constraint weight adjustment factor is introduced to dynamically adjust the constraint strength according to the heterogeneity of the formation: in homogeneous formations, physical constraints are strengthened to ensure that the model output strictly conforms to the laws of seepage mechanics; in heterogeneous formations, constraints are moderately relaxed to retain the flexibility of data-driven training. Simultaneously, a segmented, multi-analytical solution constraint strategy is adopted. The Dupuit stable seepage analytical solution is used in homogeneous and intact sections; the Goodman inflow formula is used in fault fracture zones; and the mirror method superposition solution is used in karst development sections. Constraints are applied independently to each section to ensure the geographical relevance and engineering applicability of the physical constraints. The model uses a grid search and early stopping strategy for automatic optimization, automatically selecting the optimal model based on minimizing the joint loss of the validation set, avoiding overfitting, and ensuring the predictive stability of the model under unknown conditions. Attached Figure Description

[0019] Figure 1 A flowchart of a method and system for predicting tunnel water inflow intervals based on physical residual adaptive constraints and seepage mechanism-guided conformal calibration; Figure 2 A schematic diagram of the joint loss function structure and adaptive weight adjustment mechanism for physical residual adaptive constraint training; Figure 3 A schematic diagram of the conformal calibration process guided by the multi-analytical solution constraint strategy and geological grouping seepage mechanism for segmented and segmented regions; Figure 4 This is a schematic diagram of the sensitivity analysis of parameter PDP and the dual-risk cross-determination matrix. Detailed Implementation

[0020] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless otherwise expressly indicated by the invention, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. As described in the background section, existing technologies have problems. This invention proposes a method and system for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanisms. This prediction method and system are applicable to high-precision interval prediction, uncertainty quantification, and construction-period risk management of water inflow under multi-source geological and design parameters in mountain tunnels, urban tunnels, and deep underground engineering projects. The system includes data preprocessing, spatiotemporal decoupling feature engineering, physical residual adaptive constraint model training, conformal calibration guided by geological grouping and seepage mechanisms, prediction visualization, PDP sensitivity analysis, and dual risk assessment. Data preprocessing performs missing value filling, outlier removal, and standardization on multi-source tunnel engineering data. The spatiotemporal decoupling feature engineering module constructs rainfall-water inflow time-delay coupling features through Granger causality tests, combines dimensionless derivative features of seepage mechanics with geological heterogeneity factors, completes feature optimization, and performs correlation analysis using Spearman correlation coefficients. The residual adaptive constraint model training module, based on the CatBoost framework, constructs a joint loss function consisting of data fitting loss and physical residual penalty. It introduces a constraint weight adjustment factor adaptively adjusted according to formation heterogeneity and employs a segmented, multi-analytical solution constraint strategy, combined with grid search and early stopping strategies to achieve model training and hyperparameter optimization. The geological grouping seepage mechanism-guided conformal calibration module constructs 0.1, 0.5, and 0.9 ternary regression models to form basic inflow prediction intervals. These are cross-grouped according to hydrogeological dimensions, and the seepage consistency score (SCS) integrating physical residuals is used as a measure of inconsistency in conformal calibration, achieving a highly reliable prediction interval with a 90% target coverage probability. The PDP sensitivity analysis and dual risk assessment module displays parameter sensitivity patterns based on partial dependency graphs, rendering prediction interval band maps and correlation heatmaps. Risk levels are classified based on the cross-judgment matrix of prediction interval width and physical consistency index, forming an intuitive engineering risk assessment result.

[0022] Example 1 This embodiment proposes a method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanisms. This method is based on the implementation process of an intelligent interval prediction system for tunnel water inflow under complex hydrogeological conditions. With physical residual adaptive constraints and conformal calibration guided by seepage mechanisms as its core, it deeply integrates multi-source engineering data and seepage analytical solution mechanisms. Through data preprocessing, spatiotemporal decoupling feature engineering, physical residual adaptive constraint training, geological grouping conformal interval calibration, PDP sensitivity analysis, and dual risk assessment, it achieves high accuracy, high reliability, and intelligence in predicting tunnel water inflow intervals. The specific steps include: Acquire multi-source tunnel engineering data; Missing value filling, outlier removal and standardization processing are performed on multi-source tunnel engineering data; Perform spatiotemporal decoupling feature engineering to determine time delay parameters by conducting Granger causality tests on the time series of rainfall and water inflow. We constructed time-delay weighted cumulative rainfall characteristics and attenuation coefficient weighted effective rainfall characteristics, constructed dimensionless derivative characteristics of seepage mechanics and geological heterogeneity factors, and used Spearman correlation coefficient for feature screening and correlation analysis. In the CatBoost gradient boosting tree framework, a joint loss function consisting of a data fitting loss term and a physical residual penalty term is constructed. A constraint weight adjustment factor λ, which is adaptively adjusted according to the formation heterogeneity, is introduced. A partitioned and segmented multi-analytical solution constraint strategy is adopted, and grid search and early stopping strategies are combined to complete model training and hyperparameter optimization. Based on the ternary water inflow prediction value output by the trained optimal model, a basic prediction interval is formed. The validation set samples are cross-grouped according to the hydrogeological dimension. Within each geological group, the seepage consistency score of the fused physical residual is calculated as a non-consistency measure of conformal calibration, so as to achieve prediction interval calibration with a 90% target coverage probability. Based on the prediction interval width and physical consistency index, a dual risk cross-judgment is performed and risk levels are divided to form intelligent construction decision suggestions.

[0023] As a further technical solution, multi-source engineering data is collected based on the tunnel construction project. Data sources include geological survey reports, borehole data, construction progress records, hydrological monitoring logs, and grouting design parameters. Specifically, these include geological parameters (permeability coefficient, surrounding rock grade, water abundance, groundwater type, fault fracture zone length), construction parameters (construction method, number of tunnel sections, number of tunnel faces, construction cross-sectional area, cumulative advance, daily advance, tunnel depth, still water level), environmental parameters (rainfall, infiltration coefficient), and grouting parameters (grouting ring thickness, grouting ring permeability coefficient). Multiple interpolation is used to fill in missing values, preserving the statistical distribution characteristics of the original data. IQR (Independent Quality Reduction) is used to identify and remove outliers; that is, when a data point is less than a certain value... or greater than The time markers are marked as anomalies. The removed data is then Z-score standardized to eliminate the influence of different units and provide a unified input condition for subsequent feature engineering and model training.

[0024] Specifically, in the spatiotemporal decoupling characteristic engineering phase, a vector autoregression model is first used to perform a Granger causality test on the rainfall time series and the water inflow time series to determine the maximum lag days τmax for a significant causal relationship (p < 0.05). Here, p is the statistical significance probability value of the Granger causality test; p < 0.05 indicates that the probability of no causal relationship between the rainfall time series and the water inflow time series is less than 5%, meaning the causal relationship has statistical significance. τmax is then verified using actual engineering data. max The time-lag weighted cumulative rainfall characteristics were constructed based on a 5-day period. In the formula τ max Significant lag days determined by Granger causality test; weights RFt-i represents the rainfall on day ti, where i and j are non-negative integers, i is the index of the number of lag days, and j is the index of the normalized summation of the denominator with the same traversal range as i; the weights are distributed in decreasing order of time, with rainfall closer to the current time having a larger weight; simultaneously, a weighted effective rainfall feature based on attenuation coefficients is constructed. λ is the attenuation coefficient, taken as 0.3. Secondly, dimensionless derived features of seepage mechanics are constructed: grout ring thickness / tunnel diameter, hydraulic head height / tunnel diameter, and fault fracture zone length / tunnel diameter, to eliminate the scale effect caused by different tunnel cross-sectional dimensions, enabling the model to have good generalization ability among tunnels with different cross-sectional forms. Thirdly, the coefficient of variation of the permeability coefficient is calculated. As a geological heterogeneity factor, σ in the formula k Let μ be the standard deviation of the permeability coefficient along the tunnel mileage direction. K The mean permeability coefficient is used to describe the degree of formation heterogeneity; it quantitatively describes the non-uniformity of permeability coefficient along the tunnel mileage direction, providing input for subsequent adaptive constraint weight adjustment. Finally, the Spearman correlation coefficient between each feature and the water inflow is calculated, and highly redundant features are removed with a threshold of |ρ|≥0.95 (where ρ is the Spearman rank correlation coefficient, with a value range of [...]). [1, 1], the closer |ρ| is to 1, the stronger the monotonic correlation between the two features, and |ρ| ≥ 0.95 indicates that the statistical information of the two features is highly overlapping. Plot the correlation heat map to verify whether the correlation direction between each feature and the inflow rate conforms to the laws of seepage mechanics, and ensure the physical rationality and statistical validity of the feature set.

[0025] Specifically, such as Figure 2 As shown, during the training phase of the physical residual adaptive constraint model, a joint loss function is constructed for each training sample based on the CatBoost gradient boosting tree framework. In the formula The data fitting loss is calculated using the mean squared error as the standard, and it continuously reduces the deviation between the model-predicted inflow and the measured inflow. To predict water inflow using a neural network model, To measure the actual water inflow on-site, This is a physical residual penalty term, which sets monotonicity constraints on the permeability coefficient, grouting ring thickness, hydraulic head, and inflow rate based on seepage mechanics. Includes the geological and hydrological parameters (permeability coefficient, hydraulic head, tunnel geometry, etc.) required for the analytical solution of water inflow; among which is the data fitting loss. Mean squared error loss is used to ensure that the model accurately fits the measured data.

[0026] Physical residual penalty The three-stage monotonicity penalty includes a monotonically increasing penalty for permeability coefficient and inflow rate, a monotonically decreasing penalty for grouting ring thickness and inflow rate, and a monotonically increasing penalty for head height and inflow rate. This covers the three basic monotonicity conditions of seepage mechanics: increasing permeability coefficient equals increasing inflow rate, increasing grouting ring thickness equals decreasing inflow rate, and increasing head height equals increasing inflow rate. In each gradient improvement iteration, the violation ratio of the current model on the training set of the above three monotonicity conditions is calculated, and the violation ratio is weighted and included in the physical residual term, forcing the model output trend to strictly conform to the laws of seepage mechanics.

[0027] Adaptive constraint weights that vary with the coefficient of variation of the permeability coefficient Based on the dynamic adjustment of formation heterogeneity, the formula is as follows: This is the upper limit of the weight. It is the attenuation constant. This is the critical threshold for heterogeneity; when the permeability coefficient varies with the coefficient of variation. As λ increases, λ decreases to achieve an adaptive match between physical constraint strength and formation complexity. Homogeneous formations strengthen physical constraints to ensure physical consistency, while heterogeneous formations relax constraints to retain data-driven flexibility. Specifically, when... When the value is relatively small and the layer is homogeneous, λ approaches its maximum value, indicating strong physical constraints that ensure the model strictly conforms to the trend of the analytical solution; when... When the value is large and the layer is heterogeneous, λ approaches 0, the constraints are relaxed, and space is reserved for data-driven learning of complex patterns.

[0028] Specifically, such as Figure 3As shown, a segmented, multi-analytical-solution constraint strategy is adopted: the entire tunnel is divided into homogeneous and intact sections, fault-bounded sections, and karst-developed sections based on geological conditions and mileage. Physical constraints corresponding to the analytical solutions are independently applied to each section to ensure the geographical relevance of the constraints. Specifically, the Dupuit steady-state seepage analytical solution is used as the constraint benchmark in the homogeneous and intact sections; the Goodman inflow formula is used as the constraint benchmark in the fault-bounded sections; and the mirror-image superposition solution is used as the constraint benchmark in the karst-developed sections. Physical constraints corresponding to the analytical solutions are independently applied to each section to ensure the geographical relevance and engineering applicability of the constraints. Simultaneously, a grid search + early-stop strategy is set to optimize the model. Grid search is used to optimize the key hyperparameters of CatBoost. The number of early-stop rounds is set to 50. Training stops when the joint loss on the validation set no longer decreases after 50 consecutive rounds. The optimal model structure and parameter combination are automatically selected based on minimizing the joint loss on the validation set.

[0029] Specifically, in the conformal calibration stage guided by the geological grouping seepage mechanism, the predicted inflow values ​​based on the trained optimal model are output with 0.1, 0.5, and 0.9 ternary regression values, forming the basic prediction intervals. The 0.1 quantile represents the lower conservative limit, the 0.5 quantile represents the median, and the 0.9 quantile represents the upper risk limit. The validation set samples are cross-grouped in three dimensions according to the surrounding rock grade, permeability coefficient range, and water abundance, forming multiple non-overlapping geological groups. Each validation sample belongs to only one group. Within each geological group, a seepage consistency score (SCS) is calculated for each sample, where the seepage consistency score of the x-th sample is... In the formula This refers to the seepage consistency score of the xth sample; The predicted water inflow is given by the 0.5 quantile for the x-th sample. Let x be the measured inflow rate of the x-th sample. The physical consistency deviation score for the x-th sample is set to 1 when the predicted inflow trend is inconsistent with the analytical solution trend of seepage, and 0 when they are consistent. α and β are preset weighting coefficients. Within each geological group, the SCS values ​​of the n validation samples in that group are sorted in ascending order, and the x-th sample is selected. The interval expansion corresponding to each SCS value is used as the calibration offset for that group. ,Will The offsets are added to the lower and upper limits of the base prediction interval, respectively, to correct the prediction interval and obtain the calibrated prediction interval. This achieves conformal interval calibration for grouping. In this embodiment, α=0.7 and β=0.3, where α is the dominant weight reflecting the contribution of prediction error to interval calibration, and β is the auxiliary weight introducing the corrective effect of physical consistency deviation, thus quantifying both prediction error and physical consistency deviation. The SCS values ​​of all validation samples within a group are sorted from smallest to largest, and the [number]th [sample] is selected. The interval expansion corresponding to each SCS value is used as the calibration offset for that group. ,Will Add the lower and upper limits of the base prediction interval respectively to obtain the calibrated prediction interval. The coverage probability of the predicted intervals was verified to reach the target of 90% on an independent test set. This group calibration strategy ensures that the predicted intervals have both coverage fairness and physical rationality under geological conditions with different surrounding rock grades, permeability, and water abundance.

[0030] Specifically, such as Figure 4 The image shows the optimal prediction model after training, which includes PDP parameter sensitivity analysis and dual risk cross-determination stage. The specific process of PDP parameter sensitivity analysis is as follows: Two core parameters, grouting ring thickness (t) and grouting ring permeability coefficient (k), are selected. Partial dependency plots are used to analyze their marginal impact on the predicted inflow rate. A t-PDP curve is plotted to verify the negative correlation between increased grouting ring thickness and decreased inflow rate, while a k-PDP curve is plotted to verify the positive correlation between increased grouting ring permeability coefficient and increased inflow rate. This verifies the parameter variation patterns and confirms that the model output conforms to basic understanding of seepage mechanics. Based on the partial dependency plots, construction parameters are optimized and inversely extrapolated. When t-PDP shows high sensitivity to grouting ring thickness, increasing the grouting ring thickness can effectively reduce inflow rate and narrow the range; when k-PDP shows high sensitivity to grouting ring permeability coefficient, decreasing the grouting ring permeability coefficient can significantly improve the water-blocking effect.

[0031] The specific process of dual-risk cross-judgment is as follows: calculate the interval width PW and the physical consistency index PCI, and establish dual-risk cross-judgment based on the predicted interval width. Physical consistency index The two-dimensional cross-determination matrix: where , The predicted inflow rates are defined as 0.9 and 0.1 quantiles, respectively. The median width of all sample intervals (PW) is used as the narrow-to-wide boundary, and 0.8 is used as the high-to-low boundary (PCI). Based on the cross-determination matrix of PW and PCI, four risk levels are defined, and corresponding solutions are derived. When PW is narrow and PCI is high, it is considered low risk, allowing normal tunneling, and the drainage system is configured according to the design drainage capacity. When PW is wide but PCI is high, it is considered medium risk, requiring increased inflow monitoring frequency and sufficient drainage capacity margin. When PW is narrow but PCI is low, it is considered high risk, requiring re-verification of input parameter accuracy and cross-validation with analytical solutions. When PW is wide and PCI is low, it is considered extremely high risk, requiring suspension of construction, supplementary hydrogeological investigation, and re-evaluation of the grouting plan. Finally, the analysis results are integrated to output a visual report and final construction decision recommendations. This dual-risk assessment mechanism effectively compensates for the shortcomings of traditional single-dimensional risk assessment, which cannot identify "physically contradictory high risks," by simultaneously introducing two dimensions: uncertainty (PW) and physical plausibility (PCI). This provides a more reliable and comprehensive basis for decision-making regarding tunnel construction safety.

[0032] This embodiment achieves standardized processing of multi-source engineering data and systematic construction of spatiotemporal decoupling features, fundamentally solving the problems of single feature dimensions and missing time-lag effects. It unifies and standardizes multi-source survey, design, hydrological, and construction data, including surrounding rock grade, surrounding rock permeability coefficient, hydraulic head height, grout ring thickness, tunnel diameter, grout permeability coefficient, and rainfall, completing missing value filling, outlier removal, and data standardization processing to ensure unified input conditions for data of different dimensions and magnitudes. Based on this, the Granger causality test is used to determine the time-lag parameter of rainfall on water inflow, constructing time-lag weighted cumulative rainfall features and attenuation coefficient weighted effective rainfall features to effectively capture the hysteresis response law between rainfall and water inflow. Dimensionless derived features of seepage mechanics, such as grout ring thickness / relative diameter, hydraulic head height / relative diameter, and fault fracture zone length / relative diameter, are constructed to eliminate the scale effect caused by tunnel cross-sectional dimensions. The permeability coefficient variation coefficient is introduced as a geological heterogeneity factor, providing a quantitative basis for subsequent adaptive constraints. Spearman correlation analysis is used to identify redundant features and verify the correlation direction. Redundant features with |ρ|≥0.95 are automatically removed, ultimately forming a standardized, systematic, and physically driven high-quality input feature set.

[0033] This embodiment deeply integrates adaptive physical residual constraints with a segmented, multi-analytical solution strategy, addressing the core challenges of purely data-driven models exhibiting physical inconsistencies and unadaptive constraint strength adjustment. Based on the CatBoost algorithm, this invention constructs a joint loss function, organically unifying data fitting loss and physical residual penalty terms to achieve collaborative training between data-driven and physical priors. It innovatively introduces an adaptive constraint weight adjustment factor, dynamically adjusting constraint strength according to formation heterogeneity: strengthening physical constraints in homogeneous formations to ensure the model output strictly conforms to seepage mechanics laws, and moderately relaxing constraints in heterogeneous formations to retain the flexibility of data-driven approaches. The physical residual penalty term is a three-stage monotonic penalty, covering three fundamental monotonic conditions of seepage mechanics: increasing permeability coefficient (increasing inflow), increasing grouting ring thickness (decreasing inflow), and increasing hydraulic head (increasing inflow). The violation proportion is calculated and weighted in each gradient boosting iteration. Meanwhile, a segmented, multi-analytical solution constraint strategy is adopted. The Dupuit steady-flow analytical solution is used in homogeneous and intact sections, the Goodman inflow formula is used in fault-fractured zones, and the mirror method superposition solution is used in karst-developed sections. Constraints are applied independently to each section to ensure the geographical relevance and engineering applicability of the physical constraints. The model employs grid search and early stopping strategies for automatic optimization, automatically selecting the optimal model based on minimizing the joint loss on the validation set to avoid overfitting and ensure the model's predictive stability under unknown conditions.

[0034] This embodiment constructs a conformal calibration mechanism guided by a geological grouping seepage mechanism, fundamentally improving the ability to predict high-reliability intervals with a 90% confidence level and quantify uncertainty. This invention trains 0.1, 0.5, and 0.9 ternary regression models, outputting a conservative lower limit, most likely central value, and risk upper limit for inflow, respectively, forming an initial prediction interval. Based on this, it innovatively proposes a strategy of cross-grouping validation set samples according to three hydrogeological dimensions: surrounding rock grade, permeability coefficient interval, and water abundance. Each sample belongs to a unique group, and each group is calibrated independently. Within each geological group, the seepage consistency score (SCS) fused with physical residuals is used as a measure of inconsistency in conformal calibration. By sorting the SCS of each group and determining the calibration offset with a 90% coverage probability, the final prediction interval stably reaches the target coverage probability of 90% on the test set without requiring any prior assumptions about the data distribution. Compared to general conformal calibration methods, the SCS metric of this invention integrates both prediction error and physical consistency information, which can effectively solve the problem of false reliability of "narrow but inaccurate intervals" and ensure that the prediction intervals under different geological conditions have both coverage and physical rationality.

[0035] This embodiment establishes a dual-risk cross-judgment mechanism and complete parameter sensitivity analysis capabilities, significantly improving interpretability, engineering feasibility, and decision-making safety. This invention achieves parameter sensitivity analysis through a partial dependency graph. By observing the rising and falling trends of the PDP curve, it intuitively demonstrates the marginal impact of key parameters such as surrounding rock permeability coefficient, hydraulic head, grouting ring thickness, and grouting ring permeability coefficient on water inflow, verifying the physical consistency of the model. Simultaneously, it provides quantitative basis for optimizing construction parameters, adjusting grouting schemes, and designing drainage systems. It innovatively proposes a dual-risk cross-judgment matrix based on the prediction interval width (PW) and the physical consistency index (PCI): when PW is narrow and PCI is high, it is low risk, and construction can proceed directly; when PW is wide but PCI is high, it is medium risk, requiring enhanced monitoring, but the prediction trend is reliable; when PW is narrow but PCI is low, it is high risk, although the interval is narrow, it contradicts physical laws and requires re-verification; when PW is wide and PCI is low, it is extremely high risk, and construction should be suspended for supplementary investigation. This dual-judgment mechanism solves the fundamental defect of traditional single-dimensional risk judgment in failing to identify "physically contradictory high risks." Meanwhile, the entire prediction process is visualized through Spearman correlation heatmaps and prediction interval band plots, making it easier for engineers to quickly understand the basis of model predictions and risk situations.

[0036] This embodiment enhances the ability to reverse-engineer construction parameter sensitivity and proactively control and make decisions, comprehensively transforming the traditional passive prediction model. Based on the results of PDP sensitivity analysis, this invention can deduce the key construction parameter adjustment directions that narrow the predicted water inflow range: when the grouting ring thickness is highly sensitive, increasing the grouting ring thickness can effectively reduce the water inflow and narrow the prediction range; when the grouting ring permeability coefficient is highly sensitive, decreasing the grouting ring permeability coefficient can significantly improve the water-blocking effect. This reverse-engineering mechanism realizes a closed loop from "predicting water inflow" to "optimizing construction parameters to reduce water inflow," upgrading the prediction system from passive forecasting to proactive control, providing technical support for dynamic construction decision-making.

[0037] Example 2 This embodiment provides a tunnel water inflow interval prediction system based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism. It includes five core modules: data preprocessing module, spatiotemporal decoupling feature engineering module, physical residual adaptive constraint model training module, geological grouping seepage mechanism guided conformal calibration module, and dual risk judgment and decision module. Each module is linked in sequence to realize the intelligent process from multi-source data input to high-precision interval prediction of water inflow, uncertainty quantification guided by seepage mechanism, and dual risk decision output. The data preprocessing module is configured to acquire multi-source tunnel engineering data and perform missing value filling, outlier removal, and standardization on the multi-source tunnel engineering data. The spatiotemporal decoupling feature engineering module is configured to perform spatiotemporal decoupling feature engineering, perform Granger causality tests on the time series of rainfall and water inflow to determine time lag parameters, construct time lag weighted cumulative rainfall features and attenuation coefficient weighted effective rainfall features, construct seepage mechanics dimensionless derived features and geological heterogeneity factors, and perform feature screening and correlation analysis through Spearman correlation coefficient. The physical residual adaptive constraint model training module is configured to construct a joint loss function consisting of a data fitting loss term and a physical residual penalty term within the CatBoost gradient boosting tree framework. It introduces a constraint weight adjustment factor λ that is adaptively adjusted according to the formation heterogeneity, adopts a partitioned and segmented multi-analytical solution constraint strategy, and combines grid search and early stopping strategies to complete model training and hyperparameter optimization. The geological grouping seepage mechanism guides the conformal calibration module; it is configured to form a basic prediction interval based on the ternary inflow prediction value output by the trained optimal model, cross-group the validation set samples according to the hydrogeological dimension, calculate the seepage consistency score of the fused physical residual within each geological group as the non-consistency measure of conformal calibration, and realize the prediction interval calibration of the target coverage probability. The PDP sensitivity analysis and dual risk assessment module is configured to support the rendering of correlation heatmaps, partial dependency graphs, and prediction interval strip plots, and to perform parameter sensitivity back-calculation optimization. The dual risk assessment and decision-making module automatically classifies four risk levels based on the cross matrix of the prediction interval width and the physical consistency index PCI, and outputs construction safety suggestions, drainage configuration references, and risk warning conclusions, ultimately forming an intelligent prediction system that can be used for dynamic decision-making in tunnel engineering construction.

[0038] As a further technical solution, the data preprocessing module supports the import of multi-source data from tunnel geology, hydrology, design, and construction. It employs multiple interpolation and IQR methods for missing value filling and outlier removal, and eliminates dimensional differences through Z-score standardization. The spatiotemporal decoupling feature engineering module supports automatic determination of time-delay parameters using Granger causality tests, automatic calculation of time-delay weighted cumulative rainfall and attenuation coefficient weighted effective rainfall features, automatic generation of dimensionless derived features, and geological heterogeneity factors. The system includes calculation of Spearman correlation coefficient and automatic removal of redundant features with |ρ|≥0.95; the physical residual adaptive constraint model training module supports partitioned and segmented multi-analytical solution constraint embedding, dynamic calculation of adaptive constraint weight factor λ, construction of joint loss function, grid search and early stopping strategy execution; the geological grouping seepage mechanism-guided conformal calibration module supports three-dimensional cross-grouping of hydrogeology, automatic calculation of seepage consistency score (SCS), determination of group calibration offset and adjustment of interval boundaries, ensuring that the interval achieves 90% target coverage probability on the test set.

[0039] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0040] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A tunnel water inflow interval prediction method based on physical residual adaptive constraint and seepage mechanism guided conformal calibration, characterized in that, Includes the following steps: Acquire multi-source tunnel engineering data; Missing value filling, outlier removal and standardization processing are performed on multi-source tunnel engineering data; Spatiotemporal decoupling feature engineering was performed to determine time lag parameters by Granger causality tests on the time series of rainfall and water inflow. Time lag weighted cumulative rainfall features and attenuation coefficient weighted effective rainfall features were constructed. Dimensionless seepage mechanics derived features and geological heterogeneity factors were constructed. Feature selection and correlation analysis were performed using Spearman correlation coefficient. In the CatBoost gradient boosting tree framework, a joint loss function consisting of a data fitting loss term and a physical residual penalty term is constructed. A constraint weight adjustment factor that is adaptively adjusted according to the formation heterogeneity is introduced. A partitioned and segmented multi-analytical solution constraint strategy is adopted. The model training and hyperparameter optimization are completed by combining grid search and early stopping strategies. Based on the ternary water inflow prediction value output by the trained optimal model, a basic prediction interval is formed. The validation set samples are cross-grouped according to the hydrogeological dimension. Within each geological group, the seepage consistency score of the fused physical residual is calculated as a non-consistency measure of conformal calibration, so as to achieve the prediction interval calibration of the target coverage probability. Based on the prediction interval width and physical consistency index, a dual risk cross-judgment is performed and risk levels are divided to form intelligent construction decision suggestions.

2. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism as described in claim 1, characterized in that, The time-lag weighted cumulative rainfall feature is constructed in a manner of , where τ max is a significant lag number determined by Granger causality test, the weight , RF t-i is the rainfall on the t-i day; the decay coefficient weighted effective rainfall feature , λ is a decay coefficient; where i, j are both non-negative integers.

3. The method according to claim 1, characterized in that, The dimensionless derived features of seepage mechanics include thickness / relative diameter, hydraulic head / relative diameter, and fault / relative diameter; where thickness / relative diameter is the ratio of grouting ring thickness to tunnel diameter, hydraulic head / relative diameter is the ratio of hydraulic head height to tunnel diameter, and fault / relative diameter is the ratio of fault fracture zone length to tunnel diameter, which can eliminate the influence of dimensional differences and size; the geological heterogeneity factor is the coefficient of variation of permeability coefficient. In the formula σ k Let μ be the standard deviation of the permeability coefficient along the tunnel mileage direction. K This represents the average permeability coefficient.

4. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism as described in claim 1, characterized in that, The joint loss function In the formula For data fitting loss, To predict water inflow using a neural network model, To measure the actual water inflow on-site, This is a penalty term for physical residuals. Includes the stratigraphic and hydrological parameters required for the analytical solution of water inflow; The constraint weight adjustment factor In the formula This is the upper limit of the weight. It is the attenuation constant. The threshold value for heterogeneity is λ. When the coefficient of variation of the permeability increases, λ decreases to achieve an adaptive match between the strength of physical constraints and the complexity of the formation. Homogeneous formations strengthen physical constraints to ensure physical consistency, while heterogeneous formations relax constraints to retain data-driven flexibility.

5. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism according to claim 4, characterized in that, The physical residual penalty term is a three-stage physical consistency penalty, including a penalty for the monotonically increasing permeability coefficient and inflow rate, a penalty for the monotonically decreasing grouting ring thickness and inflow rate, and a penalty for the monotonically increasing head height and inflow rate. In each gradient improvement iteration, the violation ratio of the current model to the above monotonicity conditions is calculated, and the violation ratio is weighted and included in the physical residual term to force the model output trend to conform to the basic laws of seepage mechanics.

6. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism according to claim 1, characterized in that, The segmented and multi-analytical solution constraint strategy is as follows: in homogeneous and intact sections, the Dupuit steady-flow analytical solution is used as the constraint benchmark; in fault fracture zone sections, the Goodman inflow formula is used as the constraint benchmark; and in karst development sections, the mirror method superposition solution is used as the constraint benchmark. Physical constraints corresponding to the analytical solution are applied independently to each section to ensure the geographical relevance and engineering applicability of the constraints.

7. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism according to claim 1, characterized in that, The calculation method for the seepage consistency score is as follows: In the formula This refers to the seepage consistency score of the xth sample; The predicted water inflow is given by the 0.5 quantile for the x-th sample. Let x be the measured inflow rate of the x-th sample. The physical consistency deviation score for the x-th sample is set to 1 when the predicted inflow trend is inconsistent with the analytical solution trend of seepage, and 0 when they are consistent; α and β are preset weighting coefficients; within each geological group, the seepage consistency scores of the n validation samples in that group are sorted from smallest to largest, and the x-th score is selected. The interval expansion corresponding to each seepage consistency score is used as the calibration offset for that group. The offset is added to the lower and upper limits of the basic prediction interval to obtain the calibrated prediction interval, thus realizing the conformal interval calibration of the group.

8. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism according to claim 1, characterized in that, The cross-grouping of hydrogeological dimensions includes three dimensions: surrounding rock grade, permeability coefficient range, and water abundance. Each verification sample belongs to a unique group, and each group independently calculates the calibration offset to ensure the fairness and reliability of the prediction range under different geological conditions.

9. The method for predicting tunnel water inflow intervals based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism according to claim 1, characterized in that, The dual-risk cross-judgment is based on the cross matrix of the prediction interval width (PW) and the physical consistency index (PCI): when PW is narrow and PCI is high, it is judged as low risk, with high certainty and construction can proceed directly; when PW is wide but PCI is high, it is judged as medium risk, requiring enhanced monitoring but the predicted trend is reliable; when PW is narrow but PCI is low, it is judged as high risk, with a narrow prediction interval but contradicting physical laws, requiring re-verification; when PW is wide and PCI is low, it is judged as extremely high risk, and construction should be suspended for supplementary exploration.

10. A tunnel water inflow interval prediction system based on physical residual adaptive constraints and conformal calibration guided by seepage mechanism, characterized in that, include: Data preprocessing module; It is configured to acquire multi-source tunnel engineering data and perform missing value filling, outlier removal and standardization processing on the multi-source tunnel engineering data; Spatiotemporal decoupling feature engineering module; Configured to perform spatiotemporal decoupling feature engineering, Granger causality test is performed on the time series of rainfall and water inflow to determine time lag parameters, time lag weighted cumulative rainfall features and attenuation coefficient weighted effective rainfall features are constructed, seepage mechanics dimensionless derived features and geological heterogeneity factor are constructed, and feature screening and correlation analysis are performed through Spearman correlation coefficient; The physical residual adaptive constraint model training module is configured to construct a joint loss function consisting of a data fitting loss term and a physical residual penalty term within the CatBoost gradient boosting tree framework. It introduces a constraint weight adjustment factor λ that is adaptively adjusted according to the formation heterogeneity, adopts a partitioned and segmented multi-analytical solution constraint strategy, and combines grid search and early stopping strategies to complete model training and hyperparameter optimization. The geological grouping seepage mechanism guides the conformal calibration module; it is configured to form a basic prediction interval based on the ternary inflow prediction value output by the trained optimal model, cross-group the validation set samples according to the hydrogeological dimension, calculate the seepage consistency score of the fused physical residual within each geological group as the non-consistency measure of conformal calibration, and realize the prediction interval calibration of the target coverage probability. PDP Sensitivity Analysis and Dual Risk Assessment Module; It is configured to support the rendering of correlation heatmaps, partial dependency graphs, and prediction interval strip plots, and to perform parameter sensitivity back-calculation optimization; the dual risk judgment and decision-making module automatically divides the risk into four levels based on the cross matrix of the prediction interval width and the physical consistency index PCI, and outputs construction safety suggestions, drainage configuration references and risk warning conclusions, ultimately forming an intelligent prediction system that can be used for dynamic decision-making in tunnel engineering construction.

Images