A dike breach collapse potential intelligent deduction and inundation probability visualization system, method and medium

Abstract

The present application belongs to the technical field of embankment breach rescue, and discloses an embankment breach collapse potential intelligent deduction and submerged probability visualization system, method and medium. The engineering intervention deduction engine is constructed based on a physical guidance neural operator, takes a two-dimensional multi-physical reconstruction field containing pile body positions, material distribution, flow velocity field and terrain as input, and directly outputs a water flow velocity-water depth physical field time course of a future period. The online self-evolution engine triggers model fine-tuning by comparing the deviation of the deduction result and real-time monitoring data. The uncertainty quantification visualization interface dynamically displays the deduction result in the form of a probabilistic submerged heat map. The present application initiates a deduction architecture combining physical constraints and neural operators, solves the technical problems of low calculation efficiency of traditional numerical simulation and physical inconsistency of pure data-driven methods, realizes the unity of minute-level deduction speed and compliance with physical laws, and provides real-time and reliable scientific decision support for embankment breach rescue.

Description

Technical Field

[0001] This invention relates to the field of dike breach emergency repair technology, and in particular to a system, method and medium for intelligent prediction of dike breach and visualization of inundation probability. Background Technology

[0002] my country experiences frequent floods, and emergency response decisions regarding dike breaches heavily rely on expert experience and traditional numerical simulation methods, presenting significant technical bottlenecks. While traditional hydrodynamic simulations have clear mechanisms, they consume tens of hours of computation, failing to meet minute-level decision-making needs. Furthermore, they struggle to integrate real-time monitoring data with engineering intervention measures and are prone to convergence issues under complex boundary conditions. Purely data-driven machine learning methods, although computationally fast, are prone to violating physical laws, exhibiting poor generalization ability, and potentially producing absurd predictions such as "water flows uphill" when data is sparse. The main shortcomings of existing technologies are: difficulty in balancing computational efficiency and extrapolation accuracy; insufficient integration of multi-source data with physical mechanisms; a lack of dynamic response simulation capabilities for engineering interventions; and a lack of uncertainty quantification in prediction results.

[0003] This invention achieves an organic integration of physical laws and data-driven approaches through an engineering intervention inference engine using physically guided neural operators; it ensures the model dynamically adapts to the disaster relief process through an online self-evolutionary engine; and it provides risk perception capabilities through uncertainty quantification and visualization. While maintaining physical consistency, it increases the inference speed by two orders of magnitude, effectively solving key challenges of timeliness, accuracy, and practicality in traditional methods. Summary of the Invention

[0004] The purpose of this invention is to provide a system, method, and medium for intelligent prediction of dike breach and visualization of inundation probability, in order to solve the problems existing in the prior art and realize the functions of intelligent prediction of dike breach and visualization of dike inundation probability under high-speed water flow environment and engineering intervention.

[0005] The technical solution of the present invention is as follows: A system for intelligent prediction of dike breach and visualization of inundation probability, comprising:

[0006] The engineering intervention simulation engine is based on the construction of a physically guided neural operator. It takes a two-dimensional multi-physics reconstruction field as input and directly outputs the time history of the water velocity-depth physical field and the breach shape, which includes the water velocity field and the water depth field, for a specified future time period.

[0007] The online self-evolution engine automatically compares the predictions of the engineering intervention simulation engine with the actual monitoring data. When the deviation between the two exceeds a threshold, the fine-tuning system of the online self-evolution engine is triggered.

[0008] The uncertainty quantification visualization interface displays the time history of the water flow velocity-water depth physical field and the breach morphology output by the engineering intervention deduction engine on the real scene map in the form of a probabilistic inundation heat map, wherein the color saturation represents the inundation probability.

[0009] The two-dimensional multi-physical reconstruction field includes the number, diameter, sinking position and depth of the plugging pile, the size and weight of the plugging material, the PIV water flow velocity field, and the CAD point cloud terrain field.

[0010] The physically guided neural operator is deduced by a multi-dimensional neural operator convolution layer, and the loss function includes a data-driven loss and a physical constraint loss.

[0011] The engineering intervention deduction engine is trained in the training stage using different two-dimensional multi-physical reconstruction fields for numerical simulation and the time history of the water flow velocity-water depth physical field and the breach morphology obtained by simulation to learn the nonlinear mapping relationship between the two-dimensional multi-physical reconstruction field and the time history of the water flow velocity-water depth physical field and the breach morphology. The physically guided neural operator function expression is as follows,

[0012] ;

[0013] is the to-be-trained solver, is the projection operator, is the activation function of the tth layer, is the linear operator of the tth layer, used to provide linear transformation, is the integral kernel operator of the tth layer, is the bias tensor of the tth layer, is the activation function of the first layer, is the last layer of the multi-dimensional Fourier convolution layer; is the first layer of the lifting operator;

[0014] , is the initial hidden feature function, is the domain of , is the range of ; is the input function, represents , wherein is the domain, is the range; is the integral kernel operator of the ith layer, embedding a kernel function, and the expression is as follows:

[0015] ;

[0016] ;

[0017] in, As an integral operator, it provides nonlocal information interaction, allowing information transmission between distant points; for The Borel measure, where, It is a space of arbitrary dimensions; It is a nonlinear kernel function; The hidden characteristic function of the i-th integral kernel operator, Let i be the cumulative space of the first i arbitrary-dimensional spaces; For any function, Let be the dimension of the value space corresponding to the velocity field of the i-th sample;

[0018] The total loss function consists of data-driven loss and physical constraint loss, and its formula is as follows:

[0019] ;

[0020] For data-driven loss, Weights for data-driven loss; For physical constraint loss, The weights for the physical constraint loss;

[0021] The data-driven loss is calculated from the root mean square error, and the formula is as follows.

[0022] ;

[0023] For sample size, The predicted velocity field value of the i-th sample, The observed velocity field value of the i-th sample, The predicted water depth field value of the i-th sample The water depth field observation value of the i-th sample;

[0024] The physical constraint loss consists of multiple loss values, including the continuity equation constraint of the two-dimensional shallow water equation, the momentum equation constraint of the two-dimensional shallow water equation, the pile-flow interaction constraint, the improved material deposition constraint, and the improved sediment transport equation constraint, the formulas of which are as follows.

[0025] Physical constraint loss ;

[0026] Continuity equation constraints : ;

[0027] Momentum equation constraints :

[0028] ;

[0029] Pile-water interaction constraint : ;

[0030] Improved material deposition constraint :

[0031] ;

[0032] Improved sediment transport equation constraint :

[0033] ;

[0034] where g is the gravitational acceleration, z i is the bed elevation field of the i-th sample, S vis is the viscous source term, S fri is the friction source term, S ine is the sediment inertia source term, is the flow area ratio of the i-th sample, is the pile obstruction area of the i-th sample, is the grid cell area, is the flow velocity correction coefficient of the i-th sample, is the effective diffusion coefficient of the i-th sample, is the flow velocity correction adjustment coefficient, is the initial diffusion coefficient, is the diffusion adjustment coefficient, is the sediment mass, is the saturated sediment concentration, is the sediment concentration field, is the material settling characteristic velocity, is the critical starting flow velocity of the i-th sample, is the sediment diffusion coefficient of the i-th sample, is the erosion rate coefficient, is the deposition rate coefficient; is the i-th water depth field, is the sample number, is the divergence symbol, is the vector Hadamard product.

[0035] The implementation of the online self-evolution engine is to calculate the deviation of the last time water flow velocity-water depth physical field calculated by the engineering intervention deduction engine and the two-dimensional multi-physical reconstruction field updated by monitoring after deploying the embankment breach collapse potential intelligent deduction and inundation probability visualization system, the deviation includes flow velocity deviation, water depth size deviation, the calculation formula is as follows,

[0036] Flow velocity deviation:

[0037] Depth deviation:

[0038] Deviation:

[0039] Trigger condition:

[0040] wherein, is the flow velocity vector of the i-th grid point at the last time step, is the flow velocity vector of the i-th grid point after monitoring and updating, is the water depth of the i-th grid point at the last time step, is the water depth of the i-th grid point after monitoring and updating, is the weight corresponding to the flow velocity deviation, is the weight corresponding to the water depth size deviation, is the deviation threshold value, is the flow velocity threshold value, is the water depth threshold value; the fine-tuning system is triggered when any condition in the trigger condition formula is met.

[0041] The specific implementation of the fine-tuning system is that the two-dimensional multi-physical reconstruction field after monitoring and updating is taken as a training sample, the weights, biases and kernel function hyperparameters of all layers except the last layer of the multi-dimensional Fourier convolution layer are frozen, and only the weights, biases and kernel function hyperparameters of the last layer of the multi-dimensional Fourier convolution layer are adjusted; the residual error of the training history is recorded during the adjustment iteration process, and an early stopping mechanism is introduced to avoid overfitting during training.

[0042] The specific implementation of the uncertainty quantification visualization interface is that based on the water flow velocity-water depth physical field time history and breach morphology output by the engineering intervention deduction engine, the reliability analysis method is used to calculate the inundation probability of the location of the real scene map at each future time point, and the real scene map is rendered with a transparent gradient color layer.

[0043] The reliability analysis method obtains the water flow velocity field and the water depth field according to the water flow velocity-water depth physical field time history output by the engineering intervention deduction engine; calculates the water depth gradient feature, the kinetic energy density feature, the potential energy feature, the flow field divergence feature and the cumulative time effect feature according to the water flow velocity field and the water depth field, and the calculation formula is as follows,

[0044] Water depth gradient feature:

[0045] Kinetic energy density feature:

[0046] Potential energy feature:

[0047] Flow field divergence feature:

[0048] Cumulative time effect feature:

[0049] wherein, h threshoold is the water depth gradient safety threshold, K ed is the kinetic energy density coefficient, E v,threshoold is the kinetic energy density safety threshold, P is the potential energy coefficient, E g,threshoold is the potential energy safety threshold, t measure is the cumulative time, v ed is the water flow velocity field; the risk value is calculated based on the five features, and the calculation is as follows,

[0050] Risk value:

[0051] wherein, sigmoid is the sigmoid function, w i is the weight corresponding to the i-th feature; the improved diffusion reaction equation is constructed based on the risk value, and the formula is as follows,

[0052] Improved diffusion reaction equation:

[0053] wherein, P is the probability value distribution, D ed is the terrain adaptive diffusion coefficient, λ is the probability decay rate, P0 is the initial probability condition, and 0 is the zero vector; the probability value distribution is obtained by solving the improved diffusion reaction equation, and the color saturation is adjusted according to the probability value of each plane position at each moment.

[0054] A dike breach collapse potential intelligent deduction and inundation probability visualization method is realized by using a dike breach collapse potential intelligent deduction and inundation probability visualization system, including the following steps: inputting a two-dimensional multi-physical reconstruction field fused by the number, diameter, sinking position and depth of the pile body, the size and weight of the plugging material, the PIV water flow velocity field, and the CAD point cloud terrain field into the pre-trained engineering intervention deduction engine to quickly deduce the water flow velocity-depth physical field time course and breach shape containing the water flow velocity field and water depth field within a specified period in the future; through the uncertainty quantification visualization interface, the deduction result is dynamically displayed in the form of a probabilistic inundation heat map; during the system operation, the online self-evolution engine is used to continuously fine-tune the engineering intervention deduction engine by using the real-time updated monitoring data.

[0055] The coupling relationship between the flow field and the water depth field is strictly maintained during the deduction process, and the coupling is ensured through the following ways: when training the engineering intervention deduction engine, the loss function thereof simultaneously contains the continuity equation constraint of the two-dimensional shallow water equation, the momentum equation constraint of the two-dimensional shallow water equation, the pile-water flow interaction constraint, the improved material deposition constraint, and the improved sediment transport equation constraint.

[0056] The pre-training method of the engineering intervention deduction engine comprises the following steps: based on different two-dimensional multi-physical reconstruction fields, simulating the water flow velocity-water depth physical field time course and the breach shape, constructing the two-dimensional multi-physical reconstruction field and the corresponding water flow velocity-water depth physical field time course as training samples; and based on the physical constraint loss, training the physically guided neural operator to learn the nonlinear mapping relationship between the two-dimensional multi-physical reconstruction field and the water flow velocity-water depth physical field time course and the breach shape.

[0057] A computer readable storage medium, having stored thereon a computer program, when the computer program is executed by a processor, realizes the dike breach collapse potential intelligent deduction and inundation probability visualization method.

[0058] The beneficial effects of the present application are as follows: compared with the prior art, the present application first creates an intelligent deduction architecture based on the fusion of a physically guided neural operator and an online self-evolution mechanism, realizes the deep coupling of physical laws and data-driven, develops a multi-physical field coupling constraint system considering the engineering intervention of pile-water-sediment material, and establishes a complete closed-loop system from two-dimensional multi-physical reconstruction monitoring data to dike inundation probability visualization. The deduction speed is improved by 100 times compared with traditional numerical simulation, reaching a minute-level response; under the premise of ensuring physical consistency, the prediction accuracy is significantly improved; the system is adapted to the dynamic process of rescue through real-time self-evolution; and the system provides probabilistic risk assessment, providing a more comprehensive scientific basis for rescue decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The training flowchart of the present application;

[0060] Figure 2 The internal structure diagram of the engineering intervention deduction engine of the present application;

[0061] Figure 3 The uncertainty quantification visualization workflow diagram of the present application;

[0062] Figure 4 The water flow velocity-water depth physical field time course and the breach shape and the inundation probability distribution diagram of the present application; (a) is the water flow velocity-water depth physical field time course, and (b) is the breach shape and the inundation probability distribution diagram;

[0063] Figure 5Figures (a), (b), (c) and (d) are respectively the average time course of the water flow velocity field, the average time course of the water depth field, the time course of the submergence probability and the time course of the submergence area of the application.

[0064] Figure 6 Fig. 4 is a frequency diagram of the submergence probability of the application. DETAILED DESCRIPTION

[0065] For the convenience of the technical solutions of the application, some concepts involved in the application are first illustrated below, and the technical solutions of the application are clearly and completely described in combination with the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are part of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0066] EMBODIMENT

[0067] The application is further described in detail below with the example of the dike breach rescue plugging work of Xinhua Village in Huarong County, Yueyang City, Hunan Province on July 10, 2016. The water level of the dike breach of Xinhua Village reaches 35.15 meters, which is 0.15 meters higher than the guaranteed water level. The piping flow increases from 1 m³ / s to 10 m³ / s within half an hour, which finally leads to the cracking and breach of the dike.

[0068] REFERENCE Figure 1 In the embodiment, an intelligent inference and visualization system for dike breach collapse is provided, which comprises an engineering intervention inference engine based on a neural operator constructed based on physical guidance, configured to directly output a water flow velocity-depth physical field time course containing a water flow velocity field and a water depth field and a breach shape in a specified period in the future with a two-dimensional multi-physical reconstruction field as input; an online self-evolution engine, which automatically compares and infers prediction and subsequent actual monitoring data when the deviation between the two exceeds a threshold value, and triggers a fine-tuning system of the online self-evolution engine; an uncertainty quantification visualization interface, which displays the water flow velocity-depth physical field time course and the breach shape output by the engineering intervention inference engine in the form of a probabilistic submergence heat map on a real scene map, wherein the color saturation represents the submergence probability.

[0069] REFERENCE Figure 2In this embodiment, the physics-guided neural operator is derived by a multidimensional neural operator convolutional layer, and the loss function includes data-driven loss and physical constraint loss. The engineering intervention derivation engine is trained during the training phase using numerical simulations based on different two-dimensional multiphysics reconstruction fields, as well as the simulated water velocity-depth physical field time histories and breach morphologies, thereby learning the nonlinear mapping relationship between the two-dimensional multiphysics reconstruction field and the water velocity-depth physical field time histories and breach morphologies. The functionalized expression of the physics-guided neural operator is as follows:

[0070]

[0071] in, For the solution operator to be trained, For projection operators, Let be the activation function of the t-th layer. For the linear operator at level t, used to provide linear transformations, For the integral kernel of the t-th layer, Let be the bias tensor of the t-th layer. The activation function for layer 1. The last layer is a multidimensional Fourier convolutional layer; the first layer's lifting operator is... This maps the input function to a high-dimensional feature space. The initial hidden feature function is... Let be the activation function of the i-th layer. For the i-th layer, there is a linear operator used to provide linear transformations. Let be the bias function of the i-th layer, and construct a function similar to the input function and the hidden feature function. Let i be the integral kernel operator of the i-th layer, which embeds a kernel function, expressed as follows:

[0072]

[0073]

[0074] in, As an integral operator, it provides nonlocal information exchange, allowing information to be transmitted between distant points. for Borel measure, where In the formula It is a nonlinear kernel function.

[0075] The physical constraint loss consists of multiple loss values, including the continuity equation constraint of the two-dimensional shallow water equation, the momentum equation constraint of the two-dimensional shallow water equation, the pile-flow interaction constraint, the improved material deposition constraint, and the improved sediment transport equation constraint. The formulas are as follows.

[0076] Physical constraints: ;

[0077] Continuity equation constraints: ;

[0078] Momentum equation constraints:

[0079]

[0080] Pile-water flow interaction constraints:

[0081] Improved material deposition constraints:

[0082] ;

[0083] Improved sediment transport equation constraints:

[0084] ;

[0085] Where g is the acceleration due to gravity, z i Let S be the riverbed elevation field for the i-th sample. vis S is a viscous source term. fri For the friction source term, S ine For sediment inertial source term, Let be the circulation area ratio of the i-th sample. Let be the obstruction area of ​​the pile for the i-th sample. For the area of ​​the grid cell, Let be the flow velocity correction coefficient for the i-th sample. Let be the effective diffusion coefficient of the i-th sample. For flow rate correction adjustment coefficient, The initial diffusion coefficient is . The diffusion regulation coefficient is... For sediment quality, The saturated sand content, For the sediment concentration field, The characteristic velocity of material settling. Let be the critical starting flow velocity for the i-th sample. Let be the sediment diffusion coefficient of the i-th sample. The erosion rate coefficient is... is the deposition rate coefficient.

[0086] In this embodiment, the online self-evolutionary engine is implemented by automatically calculating the deviation between the water flow velocity-depth physical field at the last moment calculated by the engineering intervention simulation engine and the updated two-dimensional multiphysics reconstruction field after system deployment. The deviation includes flow velocity deviation and water depth deviation, and the calculation formula is as follows.

[0087] Flow rate deviation:

[0088] Depth deviation:

[0089] Deviation:

[0090] Trigger condition:

[0091] wherein, is the flow velocity vector of the i-th grid point at the last time step, is the flow velocity vector of the i-th grid point after monitoring and updating, is the water depth of the i-th grid point at the last time step, is the water depth of the i-th grid point after monitoring and updating, is the weight corresponding to the flow velocity deviation, is the weight corresponding to the water depth size deviation, is the deviation threshold value, is the flow velocity threshold value, is the water depth threshold value; the fine-tuning system is triggered when any condition in the trigger condition formula is met.

[0092] In the embodiment, the specific implementation of the fine-tuning system is that the two-dimensional multi-physical reconstruction field after monitoring and updating is taken as a training sample, the weights, biases and kernel function hyperparameters of all layers except the last multi-dimensional Fourier convolution layer are frozen, and only the weights, biases and kernel function hyperparameters of the last multi-dimensional Fourier convolution layer are adjusted; the residual error of the training history is recorded during the adjustment iteration process, and an early stopping mechanism is introduced to avoid overfitting of the training.

[0093] Referring to Figure 3 and Figure 4 In the embodiment, the specific implementation of the uncertainty quantification visualization interface is that, based on the water flow velocity-water depth physical field time history and breach morphology output by the engineering intervention deduction engine, the reliability analysis method is used to calculate the inundation probability of the location of the real scene map at each future time point, and the real scene map is rendered with a transparent gradient color layer.

[0094] In the embodiment, the reliability analysis method, according to the water flow velocity field and water depth field obtained from the water flow velocity-water depth physical field time history output by the engineering intervention deduction engine, respectively calculates the water depth gradient feature, kinetic energy density feature, potential energy feature, flow field divergence feature and cumulative time effect feature, and the calculation formula is as follows,

[0095] Water depth gradient feature:

[0096] Kinetic energy density feature:

[0097] Potential energy characteristics:

[0098] Flow field divergence characteristics:

[0099] Cumulative time effect characteristics:

[0100] Among them, h threshoold K is the safety threshold for the water depth gradient. ed E is the kinetic energy density coefficient. v,threshoold Where P is the kinetic energy density safety threshold, and E is the potential energy coefficient. g,threshoold For the potential energy safety threshold, t measure For cumulative time, v ed This represents the water flow velocity field. The risk value is calculated based on five characteristics, as follows:

[0101] Risk value:

[0102] Where sigmoid is the sigmoid function, w i Let be the weight corresponding to the i-th feature. An improved diffusion reaction equation is constructed based on the risk value, as follows:

[0103] Improved diffusion reaction equation:

[0104] Where P is the flooding probability distribution, and D ed Let λ be the terrain-adaptive diffusion coefficient, λ be the probability decay rate, P0 be the initial probability condition, and 0 be the zero vector. The flooding probability distribution is obtained by solving the improved diffusion reaction equation, and the color saturation is adjusted according to the probability value of each planar location at each time step.

[0105] See Figure 5 and Figure 6 In this embodiment, the average time history of water flow velocity field, the average time history of water depth field, the breach morphology, and the flooding probability time history and flooding area time history plot calculated by the reliability analysis method are output by the engineering intervention simulation engine.

[0106] Compared to existing technologies, this embodiment pioneers an intelligent inference architecture that integrates a physics-guided neural operator with an online self-evolutionary mechanism, achieving deep coupling between physical laws and data-driven approaches. It develops a multi-physics field coupling constraint system considering engineering interventions involving piles, water flow, and sedimentary materials; and establishes a complete closed-loop system from two-dimensional multi-physics reconstructed monitoring data to a visualized probability of dike inundation. The inference speed is 100 times faster than traditional numerical simulations, achieving minute-level response; prediction accuracy is significantly improved while ensuring physical consistency; it adapts to the dynamic process of emergency response through real-time self-evolution; and it provides probabilistic risk assessment, offering a more comprehensive scientific basis for emergency response decisions.

Claims

1. A system for intelligent prediction of dike breach and flood probability visualization, characterized in that, include: The engineering intervention simulation engine is based on the construction of a physically guided neural operator. It takes a two-dimensional multi-physics reconstruction field as input and directly outputs the time history of the water velocity-depth physical field and the breach shape, which includes the water velocity field and the water depth field, for a specified future time period. The online self-evolution engine automatically compares the predictions of the engineering intervention simulation engine with the actual monitoring data. When the deviation between the two exceeds a threshold, the fine-tuning system of the online self-evolution engine is triggered. The uncertainty quantification visualization interface displays the water flow velocity-depth physical field time history and breach morphology output by the engineering intervention simulation engine on the real-world map in the form of a probabilistic inundation heat map, where color saturation represents the probability of inundation occurring. The online self-evolutionary engine is implemented by deploying an intelligent simulation and inundation probability visualization system for dike breaches. Then, the deviation between the water flow velocity-depth physical field simulated by the engineering intervention simulation engine at the final moment and the updated two-dimensional multi-physics reconstruction field is calculated. This deviation includes velocity deviation and depth deviation, calculated as follows: Flow rate deviation: ; Water depth deviation: ; deviation: Triggering conditions: ; in, To deduce the velocity vector at the i-th grid point at the final moment, To monitor the velocity vector of the i-th grid point after the update, To deduce the water depth at the i-th grid point at the final moment, To monitor the water depth at the i-th grid point after the update, The weight corresponding to the flow velocity deviation. The weights corresponding to the deviations in water depth are: The deviation threshold, For flow rate threshold, The water depth threshold is used; the fine-tuning system is triggered if any condition in the triggering condition formula is met.

2. The intelligent simulation and flood probability visualization system for dike breach and collapse as described in claim 1, characterized in that, The two-dimensional multiphysics reconstruction field includes the number, diameter, pile location and depth of the sealing piles, size and weight of the sealing material, PIV water velocity field, and CAD point cloud terrain field.

3. The intelligent simulation and flood probability visualization system for dike breach and collapse as described in claim 1, characterized in that, The physical-guided neural operator is derived by a multidimensional neural operator convolutional layer, and the loss function includes data-driven loss and physical constraint loss. The engineering intervention simulation engine is trained during the training phase using numerical simulations based on different two-dimensional multiphysics reconstruction fields, as well as the simulated water velocity-depth physical field time histories and breach morphologies. It learns the nonlinear mapping relationship between the two-dimensional multiphysics reconstruction fields and the water velocity-depth physical field time histories and breach morphologies. The physical-guided neural operator functional expression is as follows. ； For the solution operator to be trained, For projection operators, Let be the activation function of the t-th layer. For the linear operator at level t, used to provide linear transformations, For the integral kernel of the t-th layer, Let be the bias tensor of the t-th layer. The activation function for layer 1. This is the last layer of a multidimensional Fourier convolutional layer; This is the first-level lifting operator; , The initial hidden feature function is... for The domain, for The range of values; For the input function, Represented as ,in For the domain, The range; For the integral kernel operator of the i-th layer, embed a kernel function, expressed as follows: ； ； in, As an integral operator, it provides nonlocal information interaction, allowing information transmission between distant points; for The Borel measure, where, For any dimensional space; It is a nonlinear kernel function; The hidden characteristic function of the i-th integral kernel operator, Let i be the cumulative space of the first i arbitrary-dimensional spaces; For any function, Let be the dimension of the value space corresponding to the velocity field of the i-th sample; The total loss function consists of data-driven loss and physical constraint loss, and its formula is as follows: ； For data-driven loss, Weights for data-driven loss; For physical constraint loss, The weights for the physical constraint loss; The data-driven loss is calculated from the root mean square error, and the formula is as follows. ； For sample size, The predicted velocity field value of the i-th sample, The observed velocity field value of the i-th sample, The predicted water depth field value of the i-th sample The water depth field observation value of the i-th sample; The physical constraint loss consists of multiple loss values, including the continuity equation constraint of the two-dimensional shallow water equation, the momentum equation constraint of the two-dimensional shallow water equation, the pile-flow interaction constraint, the improved material deposition constraint, and the improved sediment transport equation constraint, the formulas of which are as follows. Physical constraint loss ; Continuity equation constraints ; Momentum equation constraints : ； Pile-flow interaction constraint : ; Improved material deposition constraints : ； Improved sediment transport equation constraints : ； Where g is the acceleration due to gravity, z i Let S be the riverbed elevation field for the i-th sample. vis S is a viscous source term. fri For the friction source term, S ine For sediment inertial source term, Let be the circulation area ratio of the i-th sample. Let be the obstruction area of ​​the pile for the i-th sample. For the area of ​​the grid cell, Let be the flow velocity correction coefficient for the i-th sample. Let be the effective diffusion coefficient of the i-th sample. For flow rate correction adjustment coefficient, The initial diffusion coefficient is . The diffusion regulation coefficient is... For sediment quality, The saturated sand content, For the sediment concentration field, The characteristic velocity of material settling. Let be the critical starting flow velocity for the i-th sample. Let be the sediment diffusion coefficient of the i-th sample. The erosion rate coefficient is... This is the deposition rate coefficient; For sample size, For the divergence sign, It is the Hadamard product of vectors.

4. The intelligent simulation and flood probability visualization system for dike breach and collapse as described in claim 3, characterized in that, The specific implementation of the fine-tuning system is as follows: based on the monitored and updated two-dimensional multi-physics reconstruction field as training samples, the weights, biases, and kernel function hyperparameters of all layers except the last multi-dimensional Fourier convolutional layer are frozen, and only the weights, biases, and kernel function hyperparameters of the last multi-dimensional Fourier convolutional layer are adjusted. Adjust the residuals recorded during the iteration process and introduce an early stopping mechanism to avoid overfitting.

5. The intelligent simulation and flood probability visualization system for dike breach and collapse as described in claim 1, characterized in that, The specific implementation of the uncertainty quantification visualization interface is as follows: based on the water flow velocity-depth physical field time history and breach morphology output by the engineering intervention simulation engine, the reliability analysis method is used to calculate the inundation probability of the real-scene map at each future time point, and the real-scene map is rendered with a transparency gradient color layer.

6. The intelligent simulation and flood probability visualization system for dike breach and collapse as described in claim 5, characterized in that, The reliability analysis method obtains the water velocity field and water depth field respectively based on the time history of the water velocity-depth physical field output by the engineering intervention simulation engine; and calculates the water depth gradient characteristics, kinetic energy density characteristics, potential energy characteristics, flow field divergence characteristics, and cumulative time effect characteristics based on the water velocity field and water depth field, with the following calculation formulas. Water depth gradient characteristics: ; Kinetic energy density characteristics: ; Potential energy characteristics: ; Flow field divergence characteristics: ; Cumulative time effect characteristics: ; Among them, h threshoold K is the safety threshold for the water depth gradient. ed E is the kinetic energy density coefficient. v,threshoold As the kinetic energy density safety threshold, C PE E is the potential energy coefficient. g,threshoold For the potential energy safety threshold, t measure For cumulative time, v ed The water flow velocity field is used; the risk value is calculated based on five characteristics, as follows. Risk value: ; Where sigmoid is the sigmoid function, w i Let be the weight corresponding to the i-th feature; an improved diffusion reaction equation is constructed based on the risk value, as follows. Improved diffusion reaction equation: ; Where P is the flooding probability distribution, and D ed λ is the terrain-adaptive diffusion coefficient, λ is the probability decay rate, P0 is the initial probability condition, and 0 is the zero vector. The flooding probability distribution is obtained by solving the improved diffusion reaction equation, and the color saturation is adjusted according to the probability value of each plane position at each time.

7. A method for intelligent prediction of dike breach and visualization of inundation probability, characterized in that, The system, described in any one of claims 1-6, employs an intelligent simulation and inundation probability visualization system for dike breaches. The system comprises the following steps: inputting a two-dimensional multi-physics reconstruction field, fused from the number, diameter, location, and depth of the sealing piles, the size and weight of the sealing material, the PIV water velocity field, and the CAD point cloud topographic field, into a pre-trained engineering intervention simulation engine. This rapidly simulates the time history of the water velocity-depth physical field and the breach morphology within a specified future time period, including both the water velocity and depth fields. The simulation results are dynamically displayed as a probabilistic inundation heatmap through an uncertainty quantification visualization interface. During system operation, the engineering intervention simulation engine is continuously fine-tuned using real-time updated monitoring data via an online self-evolutionary engine.

8. The method for intelligent prediction of dike breach and visualization of inundation probability according to claim 7, characterized in that, The pre-training method of the engineering intervention simulation engine includes: numerical simulation based on different two-dimensional multi-physics reconstruction fields, simulating the time history of the water velocity-depth physical field and the breach shape, and using the two-dimensional multi-physics reconstruction field and the corresponding water velocity-depth physical field time history and breach shape as training samples; training the physically guided neural operator based on physical constraint loss, so that it learns the nonlinear mapping relationship between the two-dimensional multi-physics reconstruction field and the water velocity-depth physical field time history and breach shape.

9. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the method for intelligent simulation of dike breach and flood probability visualization as described in claim 7 or 8.

Images