Early warning decision-making method and system for flood inrush disaster in submarine tunnels across faults

Abstract

The present invention belongs to the technical field related to tunnels. In order to solve the problems of inaccurate water inrush identification and lack of interpretability in high-risk scenarios where existing submarine tunnels cross fault zones, a submarine tunnel cross-fault water inrush disaster warning decision-making method and system are proposed, and a joint modeling path of attention mechanism and Bayesian inference is constructed to realize the full-process response control of input features from weighting, prediction to interpretation; an interpretable model is used for driving force analysis, thereby providing transparent and reliable risk assessment results; and the key driving forces in multi-source heterogeneous data are combined to realize real-time identification and warning of tunnel water inrush risks through numerical simulation of water inrush disaster situations, effectively improving the accuracy of water inrush risk identification and the perfection of the warning mechanism during tunnel construction, and having important engineering application value.

Description

Technical Field

[0001] The present invention belongs to the technical field related to tunnel excavation, and in particular relates to a method and system for early warning decision-making of water inrush disasters in submarine tunnels crossing faults. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid advancement of underground engineering construction, tunnel projects face an increasing number of complex geological challenges. This is especially true for undersea tunnels crossing fault structures. Water inrush disasters have become one of the most serious safety hazards during construction. During undersea tunnel construction, water inrush events often occur suddenly and without obvious warning, often due to the coupling effects of high water pressure, complex rock structures, and seawater infiltration. These events are characterized by suddenness, destructiveness, and short response times.

[0004] Fault-revealed water inrush is a particularly typical form of disaster in submarine tunnels. When tunnel excavation reveals a fault surface, significant differences in lithology, porosity, and water content on either side of the fault disrupt the previously closed underground aquifer system, potentially rapidly forming a continuous hydraulic channel. This can lead to massive water inrush at the tunnel face and even mudstone influx, posing a serious threat to the tunnel structure and personnel safety. Significant characteristics of fault activation and water inrush caused by tunnel excavation disturbance include concentrated water inrush at the tunnel face, discontinuous strata on both sides after excavation, and loose and fragmented rock mass in the fault zone. These catastrophic events are often accompanied by strong nonlinear seepage and rock stress redistribution, making them difficult to identify in advance using traditional static assessment methods.

[0005] Currently, assessments of tunnel water inrush risk mostly rely on geological exploration data and construction experience, employing statistical or empirical models such as the Q system and multivariate regression to assess risk. However, these methods struggle to accurately capture the complex geological-hydrological coupling mechanisms within submarine fault zones. In particular, when considering key driving forces such as fault activation, hydraulic fracturing, and changes in rock permeability, they often suffer from inaccurate identification and delayed response, leading to ineffective early warning of water inrush hazards.

[0006] In recent years, the application of machine learning in tunnel engineering has gradually expanded. However, mainstream models generally suffer from a "black box" nature, lacking a clear explanation of their internal logic and unable to effectively explain to engineers why and why a water inrush occurs. This has become a significant factor limiting its application in tunnel construction, which emphasizes safety and controllability.

[0007] Therefore, in the identification of water inrush in tunnels, especially in high-risk scenarios such as submarine tunnels crossing fault zones, how to identify the key driving forces of water inrush from multi-source heterogeneous data and achieve accuracy and interpretability in water inrush identification is a problem that needs to be solved at present. Summary of the Invention

[0008] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a method and system for early warning and decision-making of water inrush disasters across faults in submarine tunnels, thereby achieving the interpretability of tunnel water inrush identification and improving the accuracy of tunnel water inrush identification.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] In a first aspect, the present invention provides a method for early warning and decision-making of flood disasters in submarine tunnels across faults, comprising:

[0011] Acquire multi-source heterogeneous data from tunnels;

[0012] Based on the multi-source heterogeneous data of the tunnel, a water inrush prediction result is obtained using a trained water inrush prediction model; wherein, in the water inrush prediction model, the multi-source heterogeneous data of the tunnel is dynamically weighted using an attention mechanism to reflect the relative driving force contribution of each heterogeneous data to the water inrush risk prediction output, and a weighted feature vector is obtained, and the weighted feature vector is processed using a Bayesian neural network to obtain the water inrush prediction result;

[0013] Based on the feature weights and water inrush prediction results obtained by the attention mechanism, an interpretable model is used to analyze the driving force and obtain the key driving forces in multi-source heterogeneous data;

[0014] Based on the multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data, a numerical simulation of the tunnel water inrush is performed to obtain the numerically simulated water inrush disaster situation;

[0015] Based on the numerical simulation of water inrush disasters and water inrush prediction results, real-time identification and early warning of tunnel water inrush risks are carried out.

[0016] In a second aspect, the present invention provides a submarine tunnel cross-fault water inrush disaster early warning decision-making system, comprising:

[0017] The data collection module is configured to: obtain tunnel multi-source heterogeneous data;

[0018] The driving force analysis module is configured to: obtain a water inrush prediction result based on the multi-source heterogeneous data of the tunnel using a trained water inrush prediction model; wherein, in the water inrush prediction model, the multi-source heterogeneous data of the tunnel is dynamically weighted using an attention mechanism to reflect the relative driving force contribution of each heterogeneous data to the water inrush risk prediction output, and obtain a weighted feature vector; and the weighted feature vector is processed using a Bayesian neural network to obtain the water inrush prediction result;

[0019] The driving force screening module is configured to: perform driving force analysis using an interpretable model based on the feature weights obtained by the attention mechanism and the water inrush prediction results, and obtain the key driving forces in multi-source heterogeneous data;

[0020] The numerical simulation module is configured to: perform numerical simulation on the tunnel water inrush condition based on the multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data, and obtain the numerically simulated water inrush disaster condition;

[0021] The monitoring and early warning module is configured to: identify and warn of tunnel water inrush risks in real time based on the numerical simulation of water inrush disasters and water inrush prediction results.

[0022] One or more of the above technical solutions have the following beneficial effects:

[0023] In this invention, a joint modeling path combining attention mechanism and Bayesian inference is constructed to achieve full-process response control of input features from weighting, prediction to interpretation. This breaks through the technical limitations of traditional models, which are black-box and poorly interpretable, and ensures that the risk prediction results are physically understandable and engineering adaptable. An interpretable model is used to deeply identify and quantitatively analyze the driving forces of water inrush disasters during tunnel construction. The model can clearly demonstrate the contribution of each driving force to the water inrush risk, thereby providing transparent and reliable risk assessment results. This interpretability not only addresses the shortcomings of existing "black-box" models, but also enables engineers and construction managers to understand and trust the model's output, improving the accuracy and operability of risk assessment.

[0024] In the present invention, by combining multi-source heterogeneous data of tunnels and the key driving forces in multi-source heterogeneous data, through numerical simulation of water inrush disasters, real-time identification and early warning of tunnel water inrush risks are comprehensively realized, effectively improving the accuracy of identifying water inrush risks during tunnel construction and the perfection of the early warning mechanism, which has important engineering application value.

[0025] In this invention, an innovative method combining data collection, interpretable model analysis and real-time monitoring and early warning is combined to provide a comprehensive means of water inrush risk assessment and management for tunnel cross-fault construction, significantly improving the safety and construction efficiency of the project.

[0026] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0028] Figure 1 This is a flow chart of risk prediction and explainability analysis in Example 1 of the present invention;

[0029] Figure 2Schematic diagram of an interpretable model in Example 1 of the present invention;

[0030] Figure 3 Schematic diagram of a tunnel cross-fault water inrush disaster identification system in Example 1 of the present invention. DETAILED DESCRIPTION

[0031] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0032] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0033] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0034] Example 1

[0035] This embodiment discloses a method for early warning and decision-making of flood inrush disasters in submarine tunnels across faults, including:

[0036] Step 1: Obtain tunnel multi-source heterogeneous data.

[0037] Through a variety of advanced hydrological and geological detection equipment, data related to sudden water disasters are collected in real time to fully understand the complex hydrogeological information around the tunnel.

[0038] The collection equipment includes: groundwater level sensor: used to monitor groundwater level changes and identify water level fluctuation trends;

[0039] Pore ​​water pressure sensor: records fluctuations in pore water pressure in real time, capturing potential changes in water flow paths;

[0040] Induced polarization sensor: collects resistivity and polarizability data to identify aquifers, weak interlayers, and water-saturated fractured rock.

[0041] Geological radar (GPR): collects reflected signal data to identify fracture distribution, rock layer thickness, and spatial distribution of fracture zones;

[0042] Drilling equipment: Obtain detailed formation samples, including the rock's porosity, water content, lithology, and mineral composition.

[0043] In practical applications, sensors collect data in real time. For example, in a tunnel construction area, pore water pressure sensors recorded dramatic water pressure fluctuations ahead of the construction site. Geological radar detected an area with strong reflection signals approximately 600 meters ahead of the tunnel. Simultaneously, induced polarization revealed low resistivity and high polarizability in this area, leading to the initial identification of a water-saturated fault fracture zone.

[0044] In addition, the collected multi-source heterogeneous data is preprocessed to ensure data consistency, including:

[0045] In terms of outlier and missing value processing, the K-nearest neighbor interpolation method (KNN) is used to fill missing values ​​to ensure data continuity. In order to eliminate outliers, the data is first evaluated through box plot analysis and Z-value detection, and abnormal data points are removed to ensure that the data set used is of high quality. The standard for Z-value detection is: when the Z value of a sample is greater than 3 or less than -3, it is considered an outlier and is eliminated. Through these methods, the noise in the data is eliminated to ensure the stability of the subsequent model. Next, the Z-score standardization formula is used to normalize the data. Z-score standardization converts each feature into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate the influence of different feature dimensions on model training. The specific formula is as follows:

[0046]

[0047] in, is the characteristic mean, is the standard deviation, X represents the original eigenvalue, represents the normalized eigenvalue.

[0048] Since the data collected come from different sources, such as the unit of pore water pressure is kPa, while the polarizability is a dimensionless value, the data are processed uniformly to ensure the consistency of different data dimensions.

[0049] Step 2: Based on the multi-source heterogeneous data of the tunnel, the trained water inrush prediction model is used to obtain the water inrush prediction result; wherein, in the water inrush prediction model, the multi-source heterogeneous data of the tunnel is dynamically weighted using the attention mechanism to reflect the relative driving force contribution of each heterogeneous data to the water inrush risk prediction output, and a weighted feature vector is obtained, and the weighted feature vector is processed using a Bayesian neural network to obtain the water inrush prediction result.

[0050] like Figure 1 As shown, in this embodiment, a wide range of sample data of submarine tunnels are collected to construct a representative submarine tunnel cross-fault water inrush disaster sample library to support subsequent risk identification and modeling analysis.

[0051] Specifically, the sample library, based on measured data, compiles geological exploration reports, advanced forecasting data, and historical water inrush archives within the tunnel construction area to form a multidimensional structured dataset encompassing geological, hydrological, engineering, and spatial attributes. This dataset includes geological parameters reflecting surrounding rock properties, such as stratum lithology, rock mass integrity coefficient, and rock formation dip; hydrological parameters characterizing the response of groundwater systems, such as groundwater level, pore water pressure, and overlying seawater depth; engineering parameters characterizing engineering disturbances, such as tunnel depth, excavation dimensions, surrounding rock pressure, and excavation depth; and spatial information, such as the distribution coordinates of soluble rocks and areas of adverse geology, to describe the relationship between fault structures and water inrush pathways.

[0052] Furthermore, to improve the accuracy of identifying disaster precursors, the sample library incorporates electrical detection data such as polarizability, apparent resistivity, and time-domain attenuation index obtained through induced polarization. These parameters can effectively reflect the water content, pore structure, and degree of fragmentation of the rock mass, and are valuable for identifying water-rich fault zones and weak interlayers. They are particularly suitable for revealing potential water inrush hazards ahead of tunnel faces. To characterize the historical evolution of risks, the sample also incorporates the occurrence time, risk level, and response section of typical water inrush events, supporting the inductive extraction of risk occurrence patterns. The establishment of this sample library provides a structured, quantifiable, and key data foundation for model training, interpretable analysis, and monitoring system construction.

[0053] To ensure the quality of sample data and the effectiveness of the model, the collected data is first preprocessed. Preprocessing includes handling outliers and missing values, normalizing the data, and screening features.

[0054] In terms of outlier and missing value processing, the K-nearest neighbor interpolation method is used to fill missing values ​​to ensure the data integrity of each sample.

[0055] In the feature screening stage, recursive feature elimination (RFE) and analysis of variance (ANOVA) were combined to screen out key features highly correlated with the risk of water inrush disasters from a large number of features, including groundwater level, pore water pressure, resistivity, polarizability, and porosity.

[0056] Specifically, RFE recursively eliminates the least important features and determines the optimal feature set based on model performance evaluation. ANOVA is used to test the correlation between each feature and the target variable and identify those that significantly influence water inrush risk. In this example, groundwater level, rock formation stability, pore water pressure, fault fracture zone width, and tidal seawater depth were ultimately selected as key features.

[0057] Through feature screening and data preprocessing, it was determined that drastic changes in groundwater levels, significant decreases in resistivity, and rapid increases in pore water pressure are important influencing factors of water inrush disasters.

[0058] To accurately assess the risk of water inrush and identify key controlling factors in undersea tunnels crossing fault zones, this example constructs a water inrush prediction model that integrates a Bayesian neural network and an attention mechanism. This model focuses on practical scenarios such as fault fracture zones with complex geological conditions and severe disturbances in multi-source heterogeneous data. The water inrush prediction model incorporates an uncertainty modeling framework, using a Bayesian mechanism to jointly represent input disturbances and predicted fluctuations. It also incorporates an attention mechanism to dynamically capture the weight changes of key features in different spatiotemporal states, thereby enhancing the model's stability and engineering adaptability under abnormally fluctuating data.

[0059] Furthermore, to establish a linkage between risk prediction results and subsequent engineering responses, this model constructs a closed-loop control path of "prediction-interpretation-feedback." The water inrush probability results and driving factor ranking weights output by the water inrush prediction model serve not only as risk assessment indicators but are also directly used for reverse optimization of monitoring point configuration, numerical simulation parameter field modeling, and early warning trigger mechanism setting. This linkage design effectively overcomes the problems of traditional models' isolated outputs and difficulty in applying to actual tunneling scenarios. It realizes an integrated response system from risk prediction, interpretation and analysis to monitoring and feedback control, significantly enhancing the engineering deployment value and decision-making support capabilities of the intelligent water inrush disaster identification system.

[0060] The water inrush prediction model receives standardized, preprocessed, multi-source feature input data. These include geological parameters such as rock formation occurrence and rock mass integrity; hydrological parameters such as groundwater level and pore water pressure; engineering parameters such as excavation disturbance range and surrounding rock pressure; and electrical parameters such as polarizability, apparent resistivity, and time-domain attenuation index derived from induced polarization inversion. These feature inputs are organized into a unified feature input vector. The vector is then passed through an embedded attention mechanism module, which dynamically weights each feature based on its importance to the model's predictions and outputs a weighted feature vector. This process aims to dynamically adjust the weights of each feature based on its contribution to the prediction, ensuring that the water inrush prediction model focuses on the most influential features. The feature weights generated by the attention mechanism reflect the relative driving force of each variable in water inrush risk prediction, enhancing the water inrush prediction model's responsiveness to key drivers and providing a quantifiable basis for subsequent interpretation modules.

[0061] The feature vectors after attention weighting are input into a Bayesian neural network for risk estimation. During the training process, the weight parameters of each layer of the Bayesian neural network are represented as probability distributions, and their posterior distributions are optimized through methods such as variational inference.

[0062] During the inference phase, the water inrush prediction model outputs two core results: the probability P of a water inrush disaster occurring at the target section, a continuous value between 0 and 1; and the uncertainty indicator corresponding to the water inrush prediction value, such as the confidence interval or the range of variation. This structure enhances the model's robustness to data interference and sample inhomogeneity, making it suitable for complex conditions such as submarine fault zones with complex information and overlapping driving factors.

[0063] The output of the water inrush prediction model includes: 1. The probability of water inrush risk (P value); 2. The uncertainty of the water inrush prediction value; and 3. The feature weight set generated by the attention mechanism. These results serve as direct input to the interpretability model, providing key data support for the subsequent calculation of the driving force contribution of each variable and the identification of key risk factors.

[0064] To further enhance the water inrush prediction model's ability to identify water inrush disasters under complex geological conditions in tunnels crossing fault zones, this implementation introduces structural prior constraints to construct a "structure-aware attention mechanism," building on the existing attention mechanism. This mechanism focuses on and prioritizes modeling of geologically significant areas, such as fault fracture zones and aquifer-rich strata, thereby enhancing the model's interpretability and targeted response in real-world engineering scenarios. By embedding geological structural information as a spatial prior into the attention mechanism's score calculation process, this method encourages the water inrush prediction model to prioritize structurally sensitive areas with a high likelihood of water inrush during feature weighting.

[0065] Specifically, the water inrush prediction model architecture is as follows:

[0066] 1. Data input: For multi-source heterogeneous data such as geological, hydrological, engineering, and electrical properties, such as groundwater level, rock stability, pore water pressure, etc., missing value filling, feature screening, and normalization processing are performed to obtain input feature vectors in a unified format. .in, Indicates the Standardized feature data; n Represents the total number of features.

[0067] 2. Attention mechanism layer: The traditional attention mechanism calculates the importance score of each feature through linear transformation and bias term. The score calculation formula is as follows:

[0068]

[0069] in, is the attention weight matrix; is a bias term used to adjust the linear transformation result.

[0070] However, in complex structural environments, the contribution of features to water inrush risk depends not only on their numerical properties but also on their spatial structural semantics. Therefore, in order to guide the water inrush prediction model to focus on structural areas in the attention calculation, this embodiment adds a structural prior guidance term to the above-mentioned scoring function to form an improved scoring function:

[0071]

[0072] in, Indicates the The structural prior weight corresponding to each feature is used to characterize whether the feature is located in a structurally sensitive area. Its value comes from engineering geological prior knowledge such as fault structure maps and water-rich layer detection data. In practical applications, it can be taken as 0 or 1, or calculated as a continuous value through multi-source information fusion to reflect the degree of structural influence, where 0 indicates structural correlation and 1 indicates non-structural correlation. It is the structural guidance strength coefficient, which is used to adjust the weight of the prior guidance in the total score. The value can be tuned by cross-validation.

[0073] The above score After normalization by the softmax function, the normalized attention weights of each feature are obtained :

[0074]

[0075] in, Indicates the i Normalized attention weights of features; For the i The score of each feature; To normalize the sum of all features.

[0076] Normalized weights Indicates the The importance of each feature in water inrush risk prediction can be used to analyze the original feature vector Perform weighted combination to generate a feature representation vector that integrates structure perception , so the weighted eigenvector is expressed as:

[0077]

[0078] 3. Bayesian neural network layer: weighted feature vector processed by structure-aware attention mechanism It has fully integrated the geological structure prior constraints and data-driven dynamic characteristics, and has significantly improved the semantic sensitivity and spatial directionality of the water inrush prediction model input while retaining the information of key driving factors.

[0079] In order to further probabilistically model the water inrush risk and quantify its prediction uncertainty, this embodiment inputs the weighted features into a Bayesian neural network model based on variational inference optimization to carry out water inrush probability prediction and risk confidence analysis.

[0080] In this embodiment, the Bayesian neural network adopts a distributed parameter modeling method, that is, the weight parameters in the neural network are regarded as random variables, and the uncertainty of the model parameters is expressed by approximate inference of their posterior probability distribution. Assume that the weighted input is , which enters the BNN structure as the input layer feature vector. The weight parameters of each layer in the network are It is assumed to obey a Gaussian distribution, that is:

[0081]

[0082] in, For the The mean vector of layer weights, Its standard deviation is used to express parameter uncertainty. The output calculation formula of each layer in BNN is:

[0083]

[0084] in, represents the output of the previous layer, is the bias term of this layer, It represents the activation function such as ReLU, tanh, etc. The activation result of the final output layer is the water inrush risk probability P, which ranges from [0,1] and is used to indicate the possibility of water inrush disaster occurring in the target section.

[0085] In the process of solving the weighted posterior distribution, since the true posterior is difficult to express explicitly, a family of parameterized approximate distributions is introduced using the variational inference method. Instead of the true posterior Where D is the training data, this embodiment introduces the variational inference method, by defining the approximate distribution to replace the posterior distribution, W represents the Bayesian neural network weight parameter.

[0086] The goal of variational inference is to minimize the approximate distribution and the true posterior distribution The KL divergence between:

[0087]

[0088] Combined with Bayes' rule, the posterior distribution can be expressed as: .

[0089] After further derivation, the optimization objective is equivalent to maximizing the variational lower bound (Evidence Lower Bound, ELBO):

[0090]

[0091] in, is the fit of the Bayesian neural network to the training data, that is, the expected value of the log-likelihood; is the KL divergence between the approximate distribution and the prior distribution, which represents the constraint on the weight complexity.

[0092] To achieve optimization, assume Gaussian distribution , optimize the mean by gradient descent and variance In order to efficiently calculate the gradient, the reparameterization technique is used to express the weight parameter W as:

[0093]

[0094]

[0095] Based on this, the gradient can be obtained by sampling To estimate, variational inference has higher efficiency and stability in optimizing the weight distribution of Bayesian neural networks.

[0096] The final output of the water inrush prediction model includes: (1) the water inrush risk probability value P, which is used to quantify the possibility of water inrush disasters occurring in the current section; (2) the corresponding uncertainty indicators, such as the confidence interval width or prediction variance, which reflect the robustness of the risk output; (3) the feature weight set output by the structure-aware attention mechanism , which is used to identify the main driving factors and their contribution intensity to water inrush.

[0097] Step 3: Based on the feature weights and water inrush prediction results obtained by the attention mechanism, an interpretable model is used to perform driving force analysis to obtain the key driving forces in multi-source heterogeneous data.

[0098] like Figure 2 As shown in the figure, to improve the interpretability of water inrush risk prediction results during undersea tunnel crossings through fault zones and thereby provide a basis for the deployment of forward tunnel face monitoring systems and the adjustment of warning thresholds, a driving force identification method that integrates an attention mechanism and SHAP value theory was proposed. This method analyzes the multivariate outputs generated by a front-end Bayesian neural network model. While retaining the model's predictive performance, it incorporates game theory to achieve traceable identification of characteristic-level driving factors.

[0099] This embodiment constructs an interpretable analysis framework for analyzing the output results of the Bayesian neural network. This embodiment introduces the SHAP value analysis method to provide a multi-dimensional interpretation of the output of the risk prediction model and establish a full-path sensitivity mapping from "feature input" to "water inrush probability output". This method can quantify the marginal contribution of a single feature in a single sample prediction one by one, thereby achieving variable sorting and driving force identification. For physical characteristics obtained by induced polarization detection in submarine fault zones (such as polarizability, apparent resistivity, and time-domain attenuation value), their SHAP values ​​are often significantly higher than other conventional engineering parameters, indicating that the response intensity and prediction contribution of such variables in the model are highly coupled.

[0100] By integrating the output of the attention mechanism with the results of SHAP value analysis, a "Driving Force Importance Report" can be generated, clearly presenting the ranking of the main factors contributing to water inrush risk. This report typically includes one to five dominant driving forces. These driving forces are highly sensitive, spatially concentrated, and have good on-site correlation. This report can directly guide sensor deployment and monitoring frequency adjustment strategies ahead of the subsea tunnel face, reducing redundant data collection costs and improving the accuracy and efficiency of risk warnings.

[0101] Specifically, the input of the model is first a feature set that has been processed by multi-source heterogeneity, denoted as , including geophysical parameters such as apparent resistivity, polarizability, time domain attenuation value, and engineering background parameters such as grouting pressure, cycle footage, etc. After receiving this set of inputs, the neural network outputs three key contents: first, the probability value of the predicted water inrush risk, reflecting the possibility of water inrush in the future at the current tunnel face position; second, the confidence index of the predicted result , which represents the stability and uncertainty range of the model prediction; the third is the feature response weight set generated by the attention mechanism , which is used to measure the contribution of each input feature to the network activation function.

[0102] Based on the above three types of output, in order to further characterize the marginal effect of each feature variable on the final decision of the model, the SHAP (Shapley Additive exPlanations) value calculation framework is introduced. This method uses the theory of Shapley value in game theory and, according to the setting of "features as cooperative players", simulates the changes in the model prediction output before and after the removal and addition of feature combinations, thereby quantifying the marginal contribution of a single variable. In mathematical expression, let represents the set of all input features, For a specific feature, its SHAP value The calculation formula is:

[0103]

[0104] in, Indicates that it does not contain features Any subset of features, Indicates that only a subset of features is used The predicted value of water inrush probability output after inputting the water inrush prediction model, Indicates that in the collection Introducing features based on The predicted value is recalculated and quantified By traversing and calculating all subset combinations, we can finally get the average marginal contribution of each feature. , forming a set of SHAP values.

[0105] Considering that the fault zone where the submarine tunnel passes often exhibits significant spatial heterogeneity and geological structural differences, this embodiment further introduces a spatial domain driving force attribution mechanism. The specific approach is to divide the entire input area into multiple physical space sub-areas. , each sub-region corresponds to a set of spatial position related feature sets SHAP value calculation is performed independently in each spatial unit, and a weight parameter is set according to the physical weight of the area in the water inrush induction mechanism. , then The spatially weighted SHAP value of each feature in the cross-regional overall model is expressed as follows:

[0106]

[0107] in, Indicates that it does not contain features Any feature subset of Indicates that only a subset of features is used The predicted value of water inrush probability output after inputting the water inrush prediction model, Indicates that in the collection Introducing features based on The predicted value is recalculated and quantified The marginal improvement.

[0108] This weighted processing fully retains the local contribution differences of each partition characteristic in the model, improves the interpretation granularity and physical spatial consistency of the water inrush driving factors, and makes the model interpretation closer to the real geological induction mechanism.

[0109] In order to enhance the stability of the interpretation and the consistency of the physical meaning of the features, this embodiment fuses the above SHAP value with the output of the attention mechanism in the neural network, that is, constructs a joint feature importance expression function , expressed as:

[0110]

[0111] in, Indicates the The fusion contribution of the features, Quantify its marginal driving force value, The contribution index after fusion processing not only has mathematical explanations, but also reflects the response strength of the feature in actual geophysical exploration activation, thus avoiding biased interpretation caused by a single indicator.

[0112] This embodiment further considers the uncertainty factors in the prediction results of the water inrush prediction model and introduces a confidence adjustment mechanism in the interpretability analysis process. Assume that the prediction result of the water inrush risk probability is , and the corresponding confidence interval width is , the confidence information is used as the weighted item in the driving force fusion index. After adjustment, The final contribution calculation formula of each feature is updated as follows:

[0113]

[0114] in, is a confidence-based weight adjustment function. When the uncertainty of the water inrush prediction model is high, If the value is larger, the driving force contribution automatically decreases to reduce the interpretation bias caused by unstable predictions.

[0115] Optionally, in the spatial domain driving force attribution mechanism, the regional average confidence can also be used as the partition weight coefficient The adjustment items are used to further enhance the robustness of the regional attribution analysis results.

[0116] Finally, this embodiment can output a feature contribution The module creates a "Primary Driving Force Ranking Report," which lists the top five dominant features with significant explanatory power in the prediction results and provides physical explanations, such as groundwater level changes, decreased fault resistivity, and increased polarizability. It also generates contribution visualization charts and spatial distribution maps, providing a scientific basis for optimizing sensor deployment strategies and monitoring cycles at the tunnel face ahead. This module also serves as a feedback signal, linking the front-end neural network training module with the risk response control module to implement a closed-loop intelligent control mechanism of "risk prediction—driving force interpretation—response adjustment," significantly improving the accurate identification of fault water inrush risks and engineering controllability during undersea tunnel construction.

[0117] Step 4: Based on the multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data, a numerical simulation of the tunnel water inrush is performed to obtain the numerically simulated water inrush disaster situation.

[0118] Based on the collected multi-source heterogeneous data, including geological, hydrological, and electrical data, this embodiment utilizes interpretable machine learning algorithms, such as SHAP value analysis, to identify the key drivers of water inrush. Based on this, a multiphysics numerical model with an embedded driver factor weighting mechanism is constructed. By embedding the weighted control logic of the driving factors within the physical parameter settings, boundary condition configuration, and coupling paths of the multiphysics numerical model, the multiphysics numerical model achieves sensitive response and adaptive evolution to high-risk areas, thereby improving the simulation accuracy of water inrush path prediction, seepage pressure response, and abnormal electrical property evolution. Ultimately, by comparing the output of the multiphysics numerical model with field-measured data, a closed-loop early warning mechanism of "interpretation and analysis - model simulation - risk identification - field verification" is formed, significantly enhancing the system's field adaptability and engineering decision-making value.

[0119] In order to realize the universal adaptability of the method under multiple types of water inrush main controlling factors, this embodiment constructs a universal mapping function mechanism based on "driving factor-model response". , through a family of functions Mapping to various physical input parameters and boundary conditions of multiphysics numerical models, such as permeability , water head , conductivity , elastic modulus This mechanism allows different types of driving factors to participate in the regulation of multi-physics numerical models and forms targeted simulation responses based on their spatial distribution and weight influence, with good universal adaptability and modular scalability.

[0120] The system embeds driving factor weights to implement a targeted parameter assignment mechanism through the following three levels:

[0121] (1) Driving factors are mapped to boundary conditions and input parameter formulas.

[0122] For any high-weight driving factor , its boundary response or input parameter mapping in the multi-physics numerical model can be expressed as:

[0123]

[0124] in, Indicates the Multi-physics numerical model input parameters or boundary control conditions such as head boundary, permeability, elastic modulus, conductivity, etc. Indicates the Driving factors at spatial points The numerical value of :For boundary conditions The form of the designed mapping function; Represents a set of function control parameters, such as normalization coefficient, threshold, adjustment coefficient, etc.

[0125] (2) Fault structural properties guide the adjustment of permeability and mechanical parameters.

[0126] Fault width It can be regarded as a structural attribute variable that controls the evolution of the water inrush path. The system determines whether it has the main driving contribution in a specific engineering section based on SHAP analysis. If it ranks in the top bit, then include it in the driving factor set , and mapped to the permeability and elastic modulus The adjustment function is used to enhance the model's resolution in the fault zone area.

[0127] Fault width Mapping to the permeability adjustment coefficient k and the elastic modulus adjustment coefficient E:

[0128]

[0129]

[0130] in, represents the standard fault width; is the adjustment coefficient; Represents the base permeability and elastic modulus.

[0131] This mechanism enables the model to automatically load the “high permeability-low strength” attribute in the water-rich area of ​​the fault to respond to the evolution of the water inrush path.

[0132] (3) Verification of consistency between multi-physics numerical model output and interpretation model.

[0133] To test the effectiveness of the spatial response of driving factors during the simulation process, the system introduces the explanation-response consistency coefficient (ERC), which is defined as a consistency index between two spatial distributions, analogous to the similarity coefficient in information retrieval or spatial overlap in geology. Its calculation form is as follows:

[0134]

[0135] in, The numerical simulation is shown in The response strength of each unit, Indicates the explanation weight value of the SHAP value driving strength mapping in this unit, is the total number of units. The ERC indicator is used to quantify the degree of consistency between the spatial projection of the explanatory variables and the model response.

[0136] In an actual engineering case, such as a section of an undersea tunnel crossing a fault, the cumulative contribution of the model's top five SHAP variables reached 86.4%, with polarization alone accounting for 21.7%. A parameter-weighted model constructed on this basis improved the accuracy of predicting high-risk water inrush areas from 78.9% to 91.6%, with a spatial overlap rate of 85.2% and an average ERC value of 0.84, significantly outperforming traditional empirical methods (ERC ≈ 0.61).

[0137] Step 5: Based on the numerical simulation of water inrush disasters and water inrush prediction results, real-time identification and early warning of tunnel water inrush risks are carried out.

[0138] In this embodiment, the top five driving forces and numerical simulations of water inrush disasters are selected to construct an on-site monitoring system for the undersea tunnel. Based on the risk contribution ranking results of the driving force variables, combined with the evolution of the water inrush path, the distribution of high-risk intervals, and the change characteristics of the deformation boundary in the numerical simulation, the on-site monitoring layout points and collection indicators are determined, completing the full-process perception and dynamic modeling of the water inrush evolution process.

[0139] The monitoring design follows the principle of "model response sensitivity first + simulation result verification and reinforcement", and the top five main driving forces are selected as the core monitoring objects. Specifically, the system monitoring parameters include rock deformation rate, groundwater level changes, pore water pressure, fault fracture zone permeability indicators, and electrical parameters such as polarizability, apparent resistivity and time domain attenuation index. The electrical anomaly areas inverted by the induced polarization technology will be used to optimize the electrode array layout in the monitoring system, assist in identifying the water-rich weak zone and hidden seepage channels in front of the tunnel face, and improve the spatial directionality and foresight of the monitoring data.

[0140] Monitoring points are prioritized in fault-revealed sections, simulated areas showing high-pressure water inrush channels, and areas with a high incidence of historical disasters. The system continuously collects key data, including changes in ground stress, water pressure response, and electrical field changes, and automatically transmits this data to the backend platform. After a pre-defined data cleaning and standardization process, it is used as input in real time into the trained water inrush prediction model for risk assessment.

[0141] The water inrush prediction model's assessment results include a water inrush probability value and its credible interval. Based on the weight distribution output by the attention mechanism and the interpretable SHAP analysis results, a real-time trend chart of driving force variations is generated. Comparing the predicted risk value with the project's preset threshold automatically triggers a water inrush warning program and simultaneously generates a warning report, highlighting the main risk factors and their response areas. Through this monitoring-analysis-response linkage mechanism, the system achieves real-time, intelligent, and interpretable early warning control of water inrush risks in subsea tunnels crossing fault zones.

[0142] To effectively respond to and precisely intervene in cross-fault flooding disasters in undersea tunnels, this embodiment establishes a multi-level early warning trigger mechanism and response strategy based on risk prediction results. This model, trained using real-time monitoring data, calculates flooding risk and generates a flooding probability value and its credible interval. The system pre-sets thresholds for flooding risk assessment, categorizing the risk into three levels based on the predicted probability P. When P is less than the warning value P1, the system operates in a routine monitoring state. When P is between P1 and the high-risk threshold P2, the system triggers a level 1 warning, prompting attention to changes in risk. When P exceeds P2, the system initiates a level 2 warning response and recommends initiating a construction intervention plan.

[0143] The warning level is determined not only by the current water inrush probability but also by the dynamic trends in the contribution of the driving forces. Based on the SHAP value analysis results, the system plots a weight variation curve for each driving force variable. If any of the top five driving forces undergoes a sudden change within a short period of time, such as a jump in weight exceeding a set threshold, this is considered a signal of a "precursorial evolution sensitive period," triggering early warning measures. This approach, by integrating feature-level dynamic information with probabilistic risk indicators, enhances warning sensitivity and proactive response.

[0144] Once triggered, the system generates a water inrush warning report, including the current risk level, a ranking of the primary driving factors, and trends in corresponding collected indicators. It also sends an alert to construction management and automatically retrieves and recommends emergency plans based on historical case studies, providing decision support for on-site construction scheduling. This multi-parameter joint identification and multi-level response mechanism ensures timely warning and effective control of water inrush disasters in complex structural areas of submarine tunnels.

[0145] This example uses an interpretable model to deeply identify and quantify the driving forces of water inrush during tunnel construction. This model clearly demonstrates the contribution of each driving force to water inrush risk, providing transparent and reliable risk assessment results. This interpretability not only addresses the shortcomings of existing "black box" models but also enables engineers and construction managers to understand and trust the model's output, improving the accuracy and operability of risk assessments.

[0146] This embodiment features real-time monitoring and analysis capabilities, enabling continuous monitoring of key parameters during tunnel construction. When high-risk conditions are detected, timely warnings are issued, providing construction personnel with ample time to react and implement appropriate countermeasures. This real-time and early warning mechanism significantly enhances construction site safety and effectively mitigates potential safety hazards associated with sudden flooding.

[0147] This method is applicable to tunnel projects under a variety of complex geological conditions, especially those across faults. It has broad application prospects. It integrates data collection, analysis, and early warning, and can flexibly adjust parameters to suit different construction environments. Its scalability and adaptability make this method suitable not only for submarine tunnel construction but also for other underground projects with similar risk characteristics.

[0148] This embodiment provides construction parties with a scientific and quantitative basis for decision-making by clearly identifying and quantifying driving forces. This helps optimize construction plans and implement effective disaster reduction measures in advance, significantly reducing the likelihood of sudden flooding disasters, lowering engineering risks and costs during construction, and improving the overall economic efficiency and safety of the project.

[0149] Example 2

[0150] like Figure 3 As shown, the purpose of this embodiment is to provide a method for early warning and decision-making of flood disasters in submarine tunnels across faults, including:

[0151] The data collection module is configured to: obtain tunnel multi-source heterogeneous data;

[0152] The driving force analysis module is configured to: obtain a water inrush prediction result based on the multi-source heterogeneous data of the tunnel using a trained water inrush prediction model; wherein, in the water inrush prediction model, the multi-source heterogeneous data of the tunnel is dynamically weighted using an attention mechanism to reflect the relative driving force contribution of each heterogeneous data to the water inrush risk prediction output, and obtain a weighted feature vector; and the weighted feature vector is processed using a Bayesian neural network to obtain the water inrush prediction result;

[0153] The driving force screening module is configured to: perform driving force analysis using an interpretable model based on the feature weights obtained by the attention mechanism and the water inrush prediction results, and obtain the key driving forces in multi-source heterogeneous data;

[0154] The numerical simulation module is configured to: perform numerical simulation on the tunnel water inrush condition based on the multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data, and obtain the numerically simulated water inrush disaster condition;

[0155] The monitoring and early warning module is configured to: identify and warn of tunnel water inrush risks in real time based on the numerical simulation of water inrush disasters and water inrush prediction results.

[0156] This embodiment also includes a data preprocessing module and a model training module. The data preprocessing module is configured to process outliers and missing values ​​and normalize data. The model training module is configured to optimize the Bayesian neural network weight distribution using variational inference and optimize the mean and variance of the weights using gradient descent.

[0157] In this embodiment, based on multidimensional geological and hydrological data, the system predicts the probability of occurrence of different geological bodies. For example, a water inrush risk in a certain area is predicted to be 90%, primarily due to a rapid rise in groundwater levels and abnormal fluctuations in pore water pressure. Combining various characteristic data, the system can accurately predict the spatial distribution and location of water inrush hazards, such as a fault fracture zone approximately 500 meters ahead of the tunnel, and also provides the scale (length, depth, and width) of the affected area. This embodiment uses a three-dimensional geological model to display the distribution and prediction results of water inrush hazards, including risk probability, geological body size, and spatial location, helping construction teams intuitively understand risk information.

[0158] Based on the prediction results, this embodiment automatically generates construction recommendations to guide the construction team in taking appropriate measures. For example, when the system identifies a high risk of water inrush (risk probability ≥ 80%), it recommends reducing the tunnel excavation speed, increasing the surrounding rock support strength, and starting the drainage pump to lower the water level. If a weak interlayer containing water is identified, grouting reinforcement is recommended to improve the stability of the formation. The system supports real-time data input and can dynamically update the model prediction results based on the latest on-site monitoring data. If the groundwater level continues to rise or the pore water pressure fluctuates significantly, the system will immediately issue an audible and visual alarm and update the risk report at the same time, providing the construction team with real-time adjustment recommendations.

[0159] This embodiment, in particular, enabled timely drainage and support measures during peak tidal periods, effectively reducing construction risks. Furthermore, the system's 3D visualization helped the construction team clearly understand the distribution characteristics of the disaster, significantly improving construction efficiency and safety. The use of this system not only optimized resource allocation but also reduced economic losses caused by water inrush.

[0160] In further embodiments, there is also provided:

[0161] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed by the processor. When the computer instructions are executed by the processor, the method described in Example 1 is performed. For the sake of brevity, no further details are given here.

[0162] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0163] The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0164] Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0165] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for early warning and decision-making of flood inrush disasters in submarine tunnels across faults, characterized by: include: Acquire multi-source heterogeneous data from tunnels; Based on the multi-source heterogeneous data of the tunnel, a water inrush prediction result is obtained using a trained water inrush prediction model; wherein, in the water inrush prediction model, the multi-source heterogeneous data of the tunnel is dynamically weighted using an attention mechanism to reflect the relative driving force contribution of each heterogeneous data to the water inrush risk prediction output, and a weighted feature vector is obtained, and the weighted feature vector is processed using a Bayesian neural network to obtain the water inrush prediction result; Based on the feature weights and water inrush prediction results obtained by the attention mechanism, an interpretable model is used to analyze the driving force and obtain the key driving forces in multi-source heterogeneous data; Based on multi-source heterogeneous tunnel data and the key driving forces within that data, a numerical simulation of water inrush in the tunnel is conducted to obtain the simulated water inrush disaster scenario. Based on the simulated water inrush disaster scenario and water inrush prediction results, real-time identification and early warning of tunnel water inrush risks are performed. Among them, based on the feature weights and water inrush prediction results obtained by the attention mechanism, an interpretable model is used to analyze the driving force, and the key driving forces in multi-source heterogeneous data are obtained, specifically: The spatial domain driving force attribution mechanism is introduced to divide the fault zone area of ​​the submarine tunnel into multiple physical space sub-areas; For the water inrush prediction results corresponding to the characteristic variables at the spatial location corresponding to each physical space sub-area, the SHAP value of each characteristic variable is calculated using an interpretable model to obtain the spatially weighted SHAP value of each characteristic variable in the fault zone area across the submarine tunnel. The fusion contribution of each feature variable is obtained based on the spatially weighted SHAP value of each feature variable in the cross-fault zone area of ​​the submarine tunnel and the corresponding feature weight obtained based on the attention mechanism; The key driving forces in multi-source heterogeneous data are determined based on the fusion contribution of each characteristic variable and the confidence of the water inrush prediction results.

2. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 1, characterized in that: Before the multi-source heterogeneous data are input into the trained water inrush prediction model, the multi-source heterogeneous data are also preprocessed, and the preprocessing includes: processing of outliers and missing values ​​and normalization of data; the multi-source heterogeneous data include geological parameters, hydrological parameters, engineering parameters and electrical parameters.

3. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 1, characterized in that: In dynamically weighting the multi-source heterogeneous data of the tunnel using the attention mechanism, a structural prior guidance item is added to characterize whether the feature is located in a structurally sensitive area to obtain a weighted feature vector.

4. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 1, characterized in that: In the optimization training of the water inrush prediction model, variational inference is used to optimize the weight distribution of the Bayesian neural network, and the mean and variance of the weight parameters are optimized by the gradient descent method.

5. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 1, characterized in that: Based on the multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data, a numerical simulation of the tunnel water inrush is conducted, and the numerical simulation of the water inrush disaster is obtained, specifically: Construct a multi-physics numerical model of the fault-crossing area of ​​an undersea tunnel; Mapping key driving forces in multi-source heterogeneous data to boundary control conditions or input parameters of multi-physics numerical models; Based on the multi-physics field numerical model that introduces the key driving forces in multi-source heterogeneous data, the tunnel water inrush situation is numerically simulated to obtain the numerically simulated water inrush disaster situation.

6. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 1, characterized in that: The water inrush prediction results are obtained based on the water inrush prediction model, and early warning is carried out by combining the key driving forces in multi-source heterogeneous data. Specifically: Based on the preset water inrush risk classification threshold and the water inrush prediction probability obtained by the water inrush prediction model, the water inrush risk level is determined, and corresponding warnings are issued according to the determined water inrush risk level; If any key driving force variable in the multi-source heterogeneous data shows a sudden change, an early warning will be issued.

7. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 5, characterized in that: The numerical simulation of tunnel water inrush conditions based on multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data also includes: mapping the fault width to the permeability and elastic modulus adjustment coefficient, so that the multi-physics field numerical model responds to the evolution of the water inrush path.

8. The method for early warning and decision-making of flood inrush disaster in a submarine tunnel across a fault as claimed in claim 1, characterized in that: Before the multi-source heterogeneous data are input into the trained water inrush prediction model, the method also includes: extracting key features from the multi-source heterogeneous data by combining recursive feature elimination method and variance analysis.

9. The early warning and decision-making system for flood disasters across faults in submarine tunnels is characterized by: include: The data collection module is configured to: obtain tunnel multi-source heterogeneous data; The driving force analysis module is configured to: obtain a water inrush prediction result based on the multi-source heterogeneous data of the tunnel using a trained water inrush prediction model; wherein, in the water inrush prediction model, the multi-source heterogeneous data of the tunnel is dynamically weighted using an attention mechanism to reflect the relative driving force contribution of each heterogeneous data to the water inrush risk prediction output, and obtain a weighted feature vector; and the weighted feature vector is processed using a Bayesian neural network to obtain the water inrush prediction result; The driving force screening module is configured to: Based on the feature weights and water inrush prediction results obtained by the attention mechanism, use an interpretable model to perform driving force analysis to identify key driving forces in multi-source heterogeneous data. Specifically, it introduces a spatial domain driving force attribution mechanism to divide the fault zone area of ​​the submarine tunnel into multiple physical spatial sub-regions; For the water inrush prediction results corresponding to the characteristic variables at the spatial location corresponding to each physical space sub-area, the SHAP value of each characteristic variable is calculated using an interpretable model to obtain the spatially weighted SHAP value of each characteristic variable in the fault zone area across the submarine tunnel. The fusion contribution of each feature variable is obtained based on the spatially weighted SHAP value of each feature variable in the cross-fault zone area of ​​the submarine tunnel and the corresponding feature weight obtained based on the attention mechanism; According to the fusion contribution of each characteristic variable and the confidence of the water inrush prediction results, the key driving force in multi-source heterogeneous data is determined; The numerical simulation module is configured to: perform numerical simulation on the tunnel water inrush condition based on the multi-source heterogeneous data of the tunnel and the key driving forces in the multi-source heterogeneous data, and obtain the numerically simulated water inrush disaster condition; The monitoring and early warning module is configured to: identify and warn of tunnel water inrush risks in real time based on the numerical simulation of water inrush disasters and water inrush prediction results.

Images