A dam multi-physical field real-time reconstruction method and system based on finite element and AI agent model fusion

Abstract

The application discloses a kind of based on finite element and AI agent model fusion dam multi-physical field real-time reconstruction method and system.Belong to water conservancy engineering safety monitoring technical field.The system includes finite element simulation data generation module, AI agent model training module, sensor interface module and physical field real-time reconstruction module.First, by finite element technology combining dam geometry and geological parameters, construct multi-physical field simulation model and generate multi-working condition simulation data.Subsequently, physical-guided graph operator or multi-scale physical attention mechanism is used to train AI agent model, to realize the efficient mapping of complex physical field.The system accesses multi-source sensor data in real time, drives the agent model after dynamic time warping and normalization processing, realizes the efficient reconstruction and anomaly detection of dam physical field.Through error analysis, model thawing and physical constraint countermeasures training are automatically triggered, to improve the adaptive ability and reconstruction accuracy of the model to sudden working conditions.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project safety monitoring technology, and more specifically relates to a method and system for real-time reconstruction of multiphysics fields of dams based on the fusion of finite element and AI agent model. Background Technology

[0002] As an important water conservancy infrastructure, dams face the following technical bottlenecks in physical field monitoring:

[0003] 1. Low efficiency of finite element analysis: Traditional finite element methods require the construction of millions of meshes for detailed modeling of dams, with a single simulation taking more than 24 hours. Furthermore, linear elastic FEMs suffer from computational distortion at singular points (such as the dam heel), requiring the reliance on nonlinear models to further increase the time consumption. Traditional FEM mesh generation is time-consuming, and the mesh quality for complex structures is low, failing to meet the real-time assessment needs of emergency conditions such as flood seasons. In seismic response analysis, traditional finite element methods rely on simplified linear elastic models, which cannot capture the progressive process of panel crushing failure, leading to delayed disaster early warning.

[0004] 2. Insufficient monitoring coverage: Existing sensor deployment is limited by cost, resulting in insufficient coverage of the dam monitoring area, especially making it difficult to monitor the physical field inside the dam body;

[0005] 3. Data heterogeneity problem: Sensor data sampling frequencies are inconsistent, as are the output signal formats and communication protocols, making it difficult for traditional monitoring methods to effectively integrate this heterogeneous data;

[0006] 4. Internal field quantities are unmeasurable: Key parameters such as the seepage field of the dam foundation cannot be directly measured by traditional sensors. Estimation based on empirical formulas has large errors. Furthermore, the observation instruments buried inside the dam body or foundation cannot be replaced or are extremely costly to replace if damaged, resulting in a lack of thorough understanding of the physical field inside the dam.

[0007] 5. The traditional CNN-LSTM combined architecture has obvious limitations. It mainly uses simple cascade fusion of spatiotemporal features without considering the differences in physical characteristics between the stress field (spatial dominance) and the seepage field (temporal dominance); the loss function only focuses on data fitting and ignores the constraints of physical conservation laws, which leads to increased extrapolation errors under extreme conditions; the model update requires global retraining and is difficult to adapt to abnormal situations such as local material degradation in dams. Summary of the Invention

[0008] To address the aforementioned problems, this invention provides a dam physical field simulation method that combines high precision and real-time performance. This method achieves rapid reconstruction of the entire dam's physical field by deeply integrating finite element simulation with artificial intelligence, thereby solving the problems of low efficiency and incomplete monitoring coverage in traditional finite element simulation.

[0009] To achieve the above objectives, the present invention employs the following technical solution: the system includes: a finite element simulation data generation module; an AI proxy model training module; a sensor interface module; and a real-time physical field reconstruction module;

[0010] The finite element simulation data generation module is used to construct a finite element mesh model based on the geometric and geological parameters of the dam, set boundary conditions, and generate a multiphysics simulation dataset.

[0011] The AI ​​agent model training module is used to split the simulation dataset into a training set, a validation set, and a test set, and to build and train the AI ​​agent model using a physically guided graph operator network to learn the finite element input-output mapping relationship.

[0012] The sensor interface module is used to collect sensor measurement data in real time, and to preprocess and synchronize the data from multiple sources.

[0013] The real-time physical field reconstruction module is used to drive the AI ​​agent model to reconstruct the physical field in real time using measured data, and triggers the model correction mechanism through error analysis.

[0014] In one embodiment, the finite element simulation data generation module constructs a finite element mesh model containing the dam body and foundation based on the dam's geometric and geological parameters, adds initial conditions, boundary conditions, and loads, performs simulation calculations, and generates a multiphysics simulation dataset.

[0015] Specifically, it includes:

[0016] Geometric model building unit: Construct a three-dimensional geometric model based on dam design data and geological parameters;

[0017] Mesh subdivision unit: Refine the mesh of the geometric model and densify the mesh of key parts of the dam body;

[0018] Boundary condition setting unit: Add initial conditions, boundary conditions and loads to build a dam simulation model;

[0019] Simulation calculation unit: Input material parameters and strength parameters, perform finite element calculations, and generate simulation datasets under different working conditions, with no less than 100 sets of data for each working condition.

[0020] In one approach, the AI ​​agent model training module splits the multiphysics simulation dataset into training, validation, and test sets, and uses a PhysGeO-Transformer hybrid network to construct and train the AI ​​agent model, learning the finite element input-output mapping relationship.

[0021] Specifically, it includes:

[0022] Model architecture building unit: The PhysGeO-Transformer hybrid architecture is adopted, in which the graph structure encoder extracts spatial features through graph convolution operators, the multi-scale physical attention module captures the temporal change pattern, the input layer contains water level, temperature, and deformation temporal parameters, and the output layer generates a three-dimensional physical field cloud map and time history curves of key parts.

[0023] Dataset splitting unit: The simulation dataset is split into training set, validation set and test set in a ratio of 8:1:1;

[0024] Loss calculation unit: The mean square error loss function is used to calculate the deviation between the model prediction value and the finite element simulation value;

[0025] Accuracy Evaluation Unit: The model accuracy is evaluated using mean absolute error and normalized mean absolute error. When the NMAE exceeds 5% for three consecutive times, the model correction mechanism is triggered, and the network weights are fine-tuned using measured data.

[0026] In one embodiment, the sensor interface module acquires real-time measured data from various sensors and processes the input monitoring data. The sensor interface module includes a protocol conversion submodule and a data synchronization submodule. Specifically, it includes:

[0027] Multi-source sensor access unit: A heterogeneous sensor network integrated and arranged in the dam heel, dam abutment and dam foundation curtain area;

[0028] Protocol conversion and integration unit: Modular hardware + protocol converter are used to solve the problem of heterogeneity in sensor communication;

[0029] Data preprocessing unit: performs triple cleaning and standardization on the raw sensor data;

[0030] Multi-source data synchronization unit: Employs dynamic time warping (DTW) algorithm to align timestamps of sensor data with different sampling frequencies.

[0031] In one approach, the real-time physical field reconstruction module uses measured data to drive an AI agent model to reconstruct the physical field in real time, and triggers a model correction mechanism through error analysis.

[0032] Specifically, it includes:

[0033] Model Deployment Unit: Deploys the trained AI agent model to the real-time computing server;

[0034] Data receiving unit: Receives standardized monitoring data output from the sensor interface module in real time;

[0035] Physical field reconstruction unit: Based on monitoring data-driven AI models, it can quickly reconstruct multiple physical fields such as stress field, temperature field, and seepage field.

[0036] Output unit: Outputs physical field cloud map and time history curves of key parts.

[0037] On the other hand, a real-time multiphysics reconstruction method for dams based on the fusion of finite element method and AI agent model, wherein the system described in the method includes:

[0038] Step 1: Based on the dam design data and geological parameters, construct a multi-physics finite element model including the dam body and dam foundation. Perform mesh refinement on key parts such as the dam heel to form a gradient mesh distribution. Add static / dynamic load boundary conditions and displacement constraints, perform finite element simulation calculations, and generate a simulation dataset including at least 12 types of working conditions such as normal operation, historical highest water level, seismic conditions, and extreme temperatures.

[0039] Step 2: Split the simulation dataset into training, validation and test sets in a ratio of 8:1:1, and build and train the AI ​​agent model using a physically guided graph operator network.

[0040] Step 3: Utilize the sensor network deployed in key areas of the dam body and foundation, including water level gauges, thermometers, piezometers, strain gauges, and plumb lines, to collect environmental data and dam monitoring data in real time.

[0041] Step 4: Use measured data to drive the AI ​​agent model to reconstruct the physical field in real time. When the normalized mean absolute error exceeds 5% for three consecutive times, the system automatically triggers a dynamic error correction mechanism based on physical-guided adversarial training to fine-tune the network weights of the model in order to improve prediction accuracy.

[0042] In one scheme, the graph structure encoder described in step 2 initializes the weights using Chebyshev polynomials; the multi-scale physical attention module reorganizes the features in the Fourier domain to enhance the ability to perceive the stress gradient at the dam heel.

[0043] In one scheme, the multi-scale physical attention module includes: a frequency domain decomposition layer, a physical query weighting layer, and a feature reconstruction layer. The frequency domain decomposition layer extracts and separates physical field features through fast Fourier transform. The physical query weighting layer uses the water level change rate as the weight allocation coefficient. The feature reconstruction layer performs inverse transformation and reconstruction on the feature vector, which significantly improves the model's ability to perceive and distinguish physical field anomalies in local key parts such as dam heels.

[0044] In one scheme, the error correction mechanism includes a layered unfreezing strategy and a physics-guided adversarial training mechanism. Specifically, when the normalized mean absolute error exceeds a certain threshold and conventional fine-tuning is ineffective, the multi-scale physical attention module of the model is automatically unfrozen. The network weights are optimized by generating adversarial training samples that conform to physical constraints, ultimately achieving rapid adaptation and high-precision reconstruction to changes in the physical field under sudden operating conditions.

[0045] In one approach, the multi-source monitoring data fusion utilizes a dynamic time warping sliding window matching algorithm to align the asynchronous sampling time series of sensors, ensuring consistent time stamps in the generated feature sequences. Furthermore, Z-score normalization eliminates the influence of different dimensions and sensor types, providing high-quality, low-noise real-time data for subsequent AI agent model input.

[0046] Beneficial effects of this invention:

[0047] 1. This invention constructs an AI proxy model using PhysGeO-Transformer to replace traditional finite element simulation, reducing the simulation speed of dam physical fields from the traditional finite element simulation time of days (>24 hours) to the second level (<30 seconds), meeting the requirements for synchronous simulation of dam structural behavior, and promoting the evolution of simulation from "post-analysis" to "real-time intervention".

[0048] 2. The dynamic fusion layer of this invention uses the water level change rate to adjust the spatiotemporal feature weights using Query, achieving a processing capacity of ≥100 sets of monitoring data per second while ensuring accuracy (the system is deployed on an NVIDIA A100 GPU).

[0049] 3. This invention integrates finite element calculation data and sensor measurement data. The trained AI agent model can quickly generate a three-dimensional physical field cloud map, increasing the monitoring coverage of the dam body to 100% and completely solving the problem of blind spots in the monitoring of hidden areas such as the dam heel and the curtain area of ​​the dam foundation.

[0050] 4. This invention innovatively designs a physical constraint decoder, which integrates the percolation field equation into the input layer through a hard constraint mechanism. This design reduces the extrapolation error of the traditional CNN-LSTM model and solves the prediction distortion problem caused by physical mismatch.

[0051] 5. This invention employs adaptive error correction. When NMAE continuously exceeds 5%, an error correction mechanism is triggered. Through a layered unfreezing strategy and a dual-modal learning rate update strategy, only the network weights in abnormal regions are updated, avoiding the resource consumption of global retraining. This ensures both the accuracy of the model and its timeliness in emergency situations.

[0052] This invention enables rapid and accurate identification of abnormal dam conditions through rapid reconstruction and precise simulation of the dam's physical field, ensuring dam operation and maintenance safety and reducing the operation and maintenance costs throughout the dam's life cycle. Attached Figure Description

[0053] Figure 1 This is a diagram of a real-time multi-physics reconstruction method and system architecture for dams based on the fusion of finite element method and AI agent model, including a finite element simulation data generation module, an AI agent model training module, a sensor interface module, and a real-time physical field reconstruction module.

[0054] Figure 2 The finite element simulation data generation module includes four main units: geometric model construction, mesh generation, boundary condition setting, and simulation calculation, generating a multiphysics simulation dataset containing 12 types of working conditions.

[0055] Figure 3 The training process for AI proxy models is shown, illustrating the complete training process from dataset splitting to model training, accuracy evaluation, and error correction mechanisms.

[0056] Figure 4 The network architecture of PhysGeO-Transformer is shown in detail, including the network structure design and data flow of the graph encoder (graph convolution operator, Chebyshev polynomial basis initialization, mesh topology modeling), multi-scale physical attention module (frequency domain decomposition, physical query weighting, feature recombination), and physical constraint decoder (hard constraint mechanism, seepage field equation injection).

[0057] Figure 5 It is a sensor interface module, including functional units such as multi-source sensor access, protocol conversion, data preprocessing, and multi-source data synchronization;

[0058] Figure 6 The flowchart of the real-time physical field reconstruction module shows the complete workflow of the trained AI agent model on the GPU server, including deployment, real-time data reception, multi-physics field reconstruction (stress field, temperature field, seepage field, response time ≤30 seconds), result output (3D physical field cloud map, time history curve of key parts, 100% dam body coverage) and dynamic error correction mechanism.

[0059] Figure 7 The flowchart of the Dynamic Time Warping (DTW) algorithm illustrates the algorithm execution process for solving the problem of inconsistent sampling frequencies from multiple sensors. It includes detailed steps such as sampling frequency difference detection, DTW distance matrix construction, cumulative distance calculation, optimal path finding, sliding window matching algorithm, and timestamp synchronization verification. Detailed Implementation

[0060] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0061] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0062] Example 1:

[0063] A real-time multi-physics reconstruction system for dams based on the fusion of finite element method and AI agent model is constructed. The system includes: a finite element simulation data generation module; an AI agent model training module; a sensor interface module; and a real-time physical field reconstruction module.

[0064] The finite element simulation data generation module constructs a finite element mesh model containing the dam body and foundation based on the dam's geometric and geological parameters, adds initial conditions, boundary conditions, and loads, performs simulation calculations, and generates a multiphysics simulation dataset.

[0065] The finite element simulation data generation module specifically includes:

[0066] Geometric model building unit: Construct a three-dimensional geometric model based on dam design data and geological parameters;

[0067] Mesh subdivision unit: The geometric model is finely meshed, and the mesh is refined for key parts of the dam body (such as the abutment and heel);

[0068] Boundary condition setting unit: Add initial conditions, boundary conditions and loads to build a dam simulation model;

[0069] Simulation calculation unit: Input material parameters and strength parameters, perform finite element calculations, and generate simulation datasets under different working conditions, with no less than 100 sets of data for each working condition.

[0070] The specific implementation steps of the finite element simulation data generation module are as follows:

[0071] A geometric model of the dam is constructed based on the dam's design data and geological parameters.

[0072] The geometric model of the dam is finely meshed, and the mesh of important parts is densified to construct the dam mesh model;

[0073] Add boundary conditions such as water level, temperature, and constraints to construct a dam simulation model;

[0074] Input the dam and foundation material parameters, strength parameters, etc., and perform finite element simulation calculations to obtain simulation datasets of various physical fields of the dam. The simulation datasets should include normal operation, historical highest water level, seismic conditions, extreme temperature conditions, etc., and each condition should have no less than 100 sets of simulation data.

[0075] The AI ​​agent model training module splits the multiphysics simulation dataset into training, validation, and test sets. It then uses a PhysGeO-Transformer hybrid network to build and train the AI ​​agent model, learning the finite element input-output mapping relationship.

[0076] The AI ​​agent model training module specifically includes:

[0077] Model architecture building unit: The PhysGeO-Transformer hybrid architecture is adopted, in which the graph structure encoder extracts spatial features through graph convolution operators, the multi-scale physical attention module captures the temporal change pattern, the input layer contains temporal parameters such as water level, temperature, and deformation, and the output layer generates a three-dimensional physical field cloud map and time history curves of key parts.

[0078] Dataset splitting unit: The simulation dataset is split into training set, validation set and test set in a ratio of 8:1:1;

[0079] Loss calculation unit: The mean square error loss function is used to calculate the deviation between the model prediction value and the finite element simulation value;

[0080] Accuracy evaluation unit: The model accuracy is evaluated using mean absolute error (MAE) and normalized mean absolute error (NMAE). When the NMAE exceeds 5% for three consecutive times, the model correction mechanism is triggered, and the network weights are fine-tuned using measured data.

[0081] The specific implementation steps of the AI ​​agent model training module are as follows:

[0082] (1) An AI agent model is constructed using a physical guided graph operator network (PhysGeO-Transformer). The graph structure encoder extracts spatial features through graph convolution operators, and the multi-scale physical attention module captures the temporal variation patterns. The model input layer contains temporal parameters such as water level, temperature, and deformation, and the output layer generates the three-dimensional physical field of the dam body and the time history curves of key parts.

[0083] (2) The dam physical field simulation dataset was divided into a training set, a validation set, and a test set in a ratio of 8:1:1. The training set was used to train the AI ​​agent model, and the test set was used to test the accuracy of the trained AI agent model. The AdamW optimizer (β1 = 0.9, β2 = 0.999) was used. Spectral graph convolution initialization (Chebyshev polynomial basis) was used. The initial learning rate was set to 3×10.-5 Training batch size: 64; Total training epochs: 800.

[0084] (3) Incorporate the seepage field equation into the loss function:

[0085]

[0086] Where: y is the measured value, y' is the predicted value of the AI ​​agent model, is the regularization coefficient, and K is the penetration coefficient. For gradient operators;

[0087] (4) The accuracy of the trained AI agent model is evaluated using MAE and NMAE:

[0088]

[0089] Where: N is the number of samples in the test set; i is the sample number; y i Let y be the finite element simulation value of the i-th sample; i ′ represents the AI ​​agent model prediction value for the i-th sample;

[0090] (5) When NMAE > 5% for three consecutive times, the model error correction mechanism is triggered. The network weights are fine-tuned using the latest measured data (≥72 hours of continuous measured data). This mechanism adopts a layered unfreezing strategy; first, the weights of the output layer of the physical constraint decoder are unfrozen, and a fixed learning rate of 1×10 is used. -5 After 10 rounds of fine-tuning, if the NMAE is still >5% after weak fine-tuning, Physical Guided Adversarial Training (P-AT) is triggered.

[0091]

[0092] Where: θ is the model parameter, δ is the small perturbation, and f θ This is a parametric model, meaning a prediction function with θ as a parameter, L as the loss function, and P(f θ (x) represents the physical constraints;

[0093] Unfreeze the weights of the multi-scale physical attention module to generate samples that conform to physical laws, and adopt a dual-modal learning rate update strategy: η = η0(1-t / T) for normal operation, and η = η0e for emergency operation. -λt The basic learning rate η0 = 3 × 10 -5 Where λ = 0.05, and t is the number of fine-tuning rounds;

[0094] (6) Establish a multi-level prediction mechanism. Under normal operating conditions, the full model is used for prediction. Under emergency operating conditions (such as sudden rise in water level, earthquake, etc.), a simplified model branch is activated, sacrificing spatial resolution to ensure real-time performance.

[0095] The sensor interface module collects real-time measured data from various sensors and processes the input monitoring data. The sensor interface module (including a protocol conversion submodule and a data synchronization submodule) specifically includes:

[0096] Multi-source sensor access unit: A heterogeneous sensor network integrated and arranged in the dam heel, dam abutment and dam foundation curtain area;

[0097] Protocol conversion and integration unit: Modular hardware + protocol converter are used to solve the problem of heterogeneity in sensor communication;

[0098] Data preprocessing unit: performs triple cleaning and standardization on the raw sensor data;

[0099] Multi-source data synchronization unit: Employs dynamic time warping (DTW) algorithm to align timestamps of sensor data with different sampling frequencies.

[0100] The specific implementation steps of the sensor interface module are as follows:

[0101] (1) Real-time collection of dam environmental and structural parameters through sensor networks to form a dynamically updated data stream;

[0102] (2) The sensor network is integrated into a unified interface by adopting a modular hardware + protocol conversion approach, supporting water conservancy monitoring protocols such as Modbus-RTU / TCP, IEC 60870-5-104, and MQTT;

[0103] (3) Preprocess the input sensor data, including removing outliers using the 3σ principle and standardizing the data using Z-score normalization, and convert it into the data format required by the surrogate model;

[0104] (4) The dynamic time warping algorithm is used to synchronize multi-source data for each sensor.

[0105] The real-time physics field reconstruction module uses measured data to drive an AI proxy model to reconstruct the physics field in real time, and triggers a model correction mechanism through error analysis. The real-time physics field reconstruction module specifically includes:

[0106] Model Deployment Unit: Deploys the trained AI agent model to the real-time computing server;

[0107] Data receiving unit: Receives standardized monitoring data output from the sensor interface module in real time;

[0108] Physical field reconstruction unit: Based on monitoring data-driven AI models, it can quickly reconstruct multiple physical fields such as stress field, temperature field, and seepage field;

[0109] Output unit: Outputs physical field cloud map, time history curves of key parts, etc.

[0110] The specific implementation steps of the real-time physical field reconstruction module are as follows:

[0111] (1) Deploy the trained AI agent model;

[0112] (2) Receive real-time sensor measurement data input from the sensor interface module;

[0113] (3) Based on the monitoring data, the multi-physics field of the dam is quickly reconstructed, and the physical field cloud map and time history curves of key parts are output.

[0114] Example 2:

[0115] Based on the system in Example 1, a real-time multiphysics reconstruction method for dams based on the fusion of finite element method and AI proxy model is constructed. The specific implementation steps are as follows:

[0116] Step 1: Based on the dam design data and geological parameters, construct a multi-physics finite element model including the dam body and foundation. The multi-physics field includes stress field, temperature field, seepage field, etc. For key parts such as the dam heel, perform mesh refinement (mesh size ≤ 0.5m in the refined area, mesh size ≤ 2m in the unrefined area) to form a gradient mesh distribution. Add static / dynamic load boundary conditions (including water level pressure, temperature field, and seismic acceleration time history) and displacement constraints. Perform finite element simulation calculations to generate simulation datasets containing at least 12 types of working conditions, including normal operation, historical highest water level, seismic conditions, and extreme temperatures, with no less than 100 sets of data for each working condition.

[0117] Step 2: Split the simulation dataset into training, validation, and test sets in an 8:1:1 ratio. Construct and train an AI agent model using a PhysGeO-Transformer network (Physically Guided Graph Operator Network). This model includes: a graph structure encoder (using graph convolution operators...) Aggregated adjacency matrix A mesh Defined grid topology relationships), multi-scale physical attention modules (based on water level change rate) For the query, perform frequency domain eigenvalue decomposition Att(Q,K,V)=F -1 [W k ·F(x)]), physical constraint decoder (to convert the seepage equation (Injected into the input layer as a hard constraint).

[0118] The physical constraint decoder injects the seepage field equation into the loss function using the Lagrange multiplier method:

[0119]

[0120] Where: y is the measured value, y' is the predicted value of the AI ​​proxy model, λ is the regularization coefficient, and K is the penetration coefficient. This is the gradient operator.

[0121] The graph-structured encoder uses Chebyshev polynomial bases. (in Let U be the regularized graph Laplacian matrix, and λ be the graph Laplacian matrix. The eigenmatrix and eigenvalues, T j () is a j-th order Chebyshev polynomial, and diag() is a diagonal matrix. The weights are initialized to explicitly model the grid adjacency relationship. The multi-scale physical attention module reorganizes features in the Fourier domain to enhance the ability to perceive the stress gradient at the dam heel.

[0122] The multi-scale physical attention module includes:

[0123] Frequency domain decomposition layer: Separates physical field features through fast Fourier transform;

[0124] Physical Query Weighting Layer: Assign weighting coefficients;

[0125] Feature Recombination Layer: Execute F -1 The inverse transform of [Ω·F(x)], where

[0126] Step 3: Utilize the sensor network deployed in key areas of the dam body and foundation, including water level gauges, thermometers, piezometers, strain gauges, and plumb lines, to collect environmental data and dam monitoring data in real time. Use the 3σ principle to remove outliers and Z-score normalization for preprocessing. Use a sliding window matching algorithm based on dynamic time warping (DTW) to align the sampling time scales of heterogeneous sensors.

[0127] Step 4: Use measured data to drive the AI ​​agent model to reconstruct the physical field in real time. When the normalized mean absolute error (NMAE) exceeds 5% for three consecutive times, the system automatically triggers the dynamic error correction mechanism based on physics-guided adversarial training (P-AT) to fine-tune the network weights of the model to improve prediction accuracy.

[0128] The response time of the real-time reconstruction of the physical field does not exceed 30 seconds, which meets the real-time assessment requirements of emergency conditions such as flood season.

[0129] The error correction mechanism includes a layered unfreezing strategy and a dual-modality learning rate update strategy. The layered unfreezing strategy first unfreezes the weights of the constraint decoder output layer for fine-tuning. If the requirements are still not met after fine-tuning, a physics-guided adversarial training (P-AT) mechanism is triggered. (where θ is the model parameter, δ is the small perturbation, and f) θThis is a parametric model, meaning a prediction function with θ as a parameter, L as the loss function, and P(f θ (x) represents the physical constraints) Unfreeze the weights of the multi-scale physical attention module to generate adversarial examples that conform to physical laws; adopt a dual-modal learning rate update strategy: under normal conditions, η = η0(1-t / T), under emergency conditions, η = η0e -λt .

[0130] Example 3:

[0131] Reference Figure 1 A method and system for real-time reconstruction of multi-physics fields of dams based on the fusion of finite element method and AI agent model, including a finite element simulation data generation module, an AI agent model training module, a sensor interface module, and a real-time reconstruction module of physical fields.

[0132] Reference Figure 2 The finite element simulation data generation module is responsible for building finite element simulation models and generating simulation datasets.

[0133] Specifically, taking a concrete gravity dam (120m high, 500m crest length) as an example, a geometric model is first constructed based on design data and geological survey data. The dam body and foundation rock mass are divided using hexahedral elements (element size ≤ 0.5m). Hexahedral swept meshes are used to refine the mesh at the dam heel (size gradient: 0.1m~0.5m) to improve calculation accuracy. Then, initial conditions, boundary conditions, and loads are added, and material parameters are input (dam body material parameters are set as follows: elastic modulus 28GPa, Poisson's ratio 0.167, density 2400kg / m³). 3 Simulation calculations are performed on parameters such as strength, and finally a simulation dataset is generated that includes normal operation, historical highest water level, seismic conditions, etc., with no less than 100 sets of data for each condition.

[0134] Reference Figures 3-4 The AI ​​agent model training module trains an AI agent model built using a PhysGeO-Transformer hybrid network based on a simulation dataset.

[0135] Specifically, refer to Figure 3 The training process for the AI ​​agent model includes dataset splitting (training, validation, and test sets in an 8:1:1 ratio), model training, loss calculation, and accuracy evaluation. When the NMAE exceeds 5% three times consecutively, a model correction mechanism is triggered to ensure the accuracy of the physics reconstruction. (Refer to...) Figure 4 The AI ​​agent model adopts the PhysGeO-Transformer architecture, which includes three levels of modules:

[0136] Graph structure encoder: using spectral graph convolution operator Extracting spatial features; Multi-scale physical attention: using water level acceleration For dynamic weighting coefficients, frequency domain feature reconstruction F' = F -1 [Ω·F(x)], where Physically Constrained Decoder: Equations for Hard Injection Seepage Field

[0137] Reference Figure 5 The sensor interface module collects and preprocesses measured data.

[0138] Specifically, the sensor data processing flow includes: first, collecting environmental and structural parameters through a sensor network, and integrating different protocol data through modular hardware and protocol conversion interfaces; then, using the 3σ principle to remove outliers, and using Z-score normalization to standardize the data; finally, using the dynamic time warping algorithm (see...) Figure 7 This solves the problem of inconsistent sampling frequencies of multi-source data and outputs standardized data in a unified format.

[0139] Reference Figure 6 The real-time physical field reconstruction module drives the AI ​​model through measured data and outputs visualized physical field results.

[0140] Specifically, the workflow of the real-time physical field reconstruction module includes: the trained AI agent model receives the measured data output by the sensor interface module, quickly reconstructs multiple physical fields such as the dam stress field and temperature field based on the monitoring data, and outputs the physical field cloud map and time history curves of key parts of the dam in real time, providing a visual analysis basis for dam safety monitoring. In this embodiment, the real-time physical field reconstruction module adopts a dual-GPU hot standby architecture, with a fault switching time of ≤0.5 seconds; the system response time is shortened from the traditional FEM >24 hours to 28 seconds, and the response time in emergency mode is 15 seconds (average of 10 times) (the system is deployed on an NVIDIA A100 GPU server).

[0141] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0142] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time multiphysics reconstruction system for dams based on the fusion of finite element method and AI proxy model, characterized in that: The system includes: a finite element simulation data generation module; an AI agent model training module; a sensor interface module; and a real-time physics field reconstruction module. The finite element simulation data generation module is used to construct a finite element mesh model based on the geometric and geological parameters of the dam, set boundary conditions, and generate a multiphysics simulation dataset. The AI ​​agent model training module is used to split the simulation dataset into a training set, a validation set, and a test set, and to build and train the AI ​​agent model using a physically guided graph operator network to learn the finite element input-output mapping relationship. The sensor interface module is used to collect sensor measurement data in real time, and to preprocess and synchronize the data from multiple sources. The real-time physical field reconstruction module is used to drive the AI ​​agent model to reconstruct the physical field in real time using measured data, and triggers the model correction mechanism through error analysis. The finite element simulation data generation module, based on the geometric and geological parameters of the dam, constructs a finite element mesh model including the dam body and foundation, adds initial conditions, boundary conditions and loads, performs simulation calculations, and generates a multiphysics simulation dataset. Specifically, it includes: Geometric model building unit: Construct a three-dimensional geometric model based on dam design data and geological parameters; Mesh subdivision unit: Refine the mesh of the geometric model and densify the mesh of key parts of the dam body; Boundary condition setting unit: Add initial conditions, boundary conditions and loads to build a dam simulation model; Simulation calculation unit: Input material parameters and strength parameters, perform finite element calculations, and generate simulation datasets under different working conditions, with no less than 100 sets of data for each working condition; The AI ​​agent model training module splits the multiphysics simulation dataset into training, validation, and test sets, and uses a PhysGeO-Transformer hybrid network to construct and train the AI ​​agent model, learning the finite element input-output mapping relationship. Specifically, it includes: Model architecture building unit: The PhysGeO-Transformer hybrid architecture is adopted, in which the graph structure encoder extracts spatial features through graph convolution operators, the multi-scale physical attention module captures the temporal change pattern, the input layer contains water level, temperature and deformation temporal parameters, and the output layer generates three-dimensional physical field cloud map and time history curves of key parts. Dataset splitting unit: The simulation dataset is split into training set, validation set and test set in a ratio of 8:1:1; Loss calculation unit: The mean square error loss function is used to calculate the deviation between the model prediction value and the finite element simulation value; Accuracy Evaluation Unit: The model accuracy is evaluated using mean absolute error and normalized mean absolute error. When the NMAE exceeds 5% for three consecutive times, the model correction mechanism is triggered, and the network weights are fine-tuned using measured data.

2. The dam multiphysics real-time reconstruction system based on the fusion of finite element method and AI proxy model as described in claim 1, characterized in that: The sensor interface module acquires real-time measured data from various sensors and processes the input monitoring data. The sensor interface module includes a protocol conversion submodule and a data synchronization submodule; specifically, it includes: Multi-source sensor access unit: A heterogeneous sensor network integrated and arranged in the dam heel, dam abutment and dam foundation curtain area; Protocol conversion and integration unit: Modular hardware + protocol converter are used to solve the problem of heterogeneity in sensor communication; Data preprocessing unit: performs triple cleaning and standardization on the raw sensor data; Multi-source data synchronization unit: Employs dynamic time warping (DTW) algorithm to align timestamps of sensor data with different sampling frequencies.

3. The dam multiphysics real-time reconstruction system based on the fusion of finite element method and AI proxy model as described in claim 1, characterized in that: The aforementioned real-time physical field reconstruction module uses measured data to drive an AI proxy model to reconstruct the physical field in real time, and triggers a model correction mechanism through error analysis; Specifically, it includes: Model Deployment Unit: Deploys the trained AI agent model to the real-time computing server; Data receiving unit: Receives standardized monitoring data output from the sensor interface module in real time; Physical field reconstruction unit: Based on monitoring data-driven AI models, it can quickly reconstruct multiple physical fields such as stress field, temperature field, and seepage field. Output unit: Outputs physical field cloud map and time history curves of key parts.

4. A real-time multiphysics reconstruction method for dams based on the fusion of finite element method and AI proxy model, wherein the method is applicable to the system as described in any one of claims 1-3, characterized in that: The method includes: Step 1: Based on the dam design data and geological parameters, construct a multi-physics finite element model including the dam body and dam foundation. Perform mesh refinement on key parts of the dam heel to form a gradient mesh distribution. Add static / dynamic load boundary conditions and displacement constraints. Perform finite element simulation calculations to generate a simulation dataset containing at least 12 types of working conditions, including normal operation, historical highest water level, seismic conditions, and extreme temperature. Step 2: Split the simulation dataset into training, validation and test sets in a ratio of 8:1:1, and build and train the AI ​​agent model using a physically guided graph operator network. Step 3: Utilize the sensor network deployed in key areas of the dam body and foundation, including water level gauges, thermometers, piezometers, strain gauges, and plumb lines, to collect environmental data and dam monitoring data in real time. Step 4: Use measured data to drive the AI ​​agent model to reconstruct the physical field in real time. When the normalized mean absolute error exceeds 5% for three consecutive times, the system automatically triggers the dynamic error correction mechanism based on physical-guided adversarial training to fine-tune the network weights of the model in order to improve prediction accuracy.

5. The method for real-time reconstruction of multiphysics fields of a dam based on the fusion of finite element method and AI proxy model as described in claim 4, characterized in that: In step 2, the graph operator network initializes the weights using Chebyshev polynomials; the multi-scale physical attention module reorganizes the features in the Fourier domain to enhance the ability to perceive the stress gradient at the dam heel.

6. The method for real-time reconstruction of multiphysics fields of a dam based on the fusion of finite element method and AI proxy model as described in claim 5, characterized in that: The multi-scale physical attention module includes a frequency domain decomposition layer, a physical query weighting layer, and a feature reconstruction layer. The frequency domain decomposition layer extracts and separates physical field features through fast Fourier transform. The physical query weighting layer uses the water level change rate as the weight allocation coefficient. The feature reconstruction layer performs inverse transformation and reconstruction on the feature vector, which significantly improves the model's ability to perceive and distinguish physical field anomalies in key local parts of the dam heel.

7. The method for real-time reconstruction of multiphysics fields of a dam based on the fusion of finite element method and AI proxy model as described in claim 4, characterized in that: The error correction mechanism includes a layered unfreezing strategy and a physics-guided adversarial training mechanism. Specifically, when the normalized average absolute error exceeds a certain threshold and conventional fine-tuning is ineffective, the multi-scale physical attention module of the model is automatically unfrozen. The network weights are optimized by generating adversarial training samples that conform to physical constraints, ultimately achieving rapid adaptation and high-precision reconstruction of physical field changes under sudden operating conditions.

8. The method for real-time reconstruction of multiphysics fields of a dam based on the fusion of finite element method and AI agent model as described in claim 4, characterized in that: Multi-source monitoring data fusion utilizes a dynamic time warping sliding window matching algorithm to align the asynchronous sampling time series of sensors, ensuring consistent time scales in the generated feature sequences. Furthermore, Z-score normalization eliminates the influence of different dimensions and sensor types, providing high-quality, low-noise real-time data for subsequent AI proxy model input.

Images