CN-121189076-B - Real-time reconstruction method and system for multiple physical fields of dam based on fusion of finite element and AI proxy model

CN121189076BCN 121189076 BCN121189076 BCN 121189076BCN-121189076-B

Abstract

The invention discloses a real-time reconstruction method and a real-time reconstruction system for multiple physical fields of a dam based on fusion of finite elements and an AI proxy model. Belongs to the technical field of hydraulic engineering safety monitoring. The system comprises a finite element simulation data generation module, an AI proxy model training module, a sensor interface module and a physical field real-time reconstruction module. Firstly, constructing a multi-physical-field simulation model and generating multi-working-condition simulation data by combining dam geometry and geological parameters through a finite element technology. And then training an AI proxy model by adopting a physically guided graph operator or a multi-scale physical attention mechanism, and realizing efficient mapping of the complex physical field. The system is accessed into the multi-source sensor data in real time, and drives the proxy model after dynamic time warping and normalization processing, so that the efficient reconstruction and anomaly detection of the physical field of the dam are realized. The model thawing and physical constraint countermeasure training is automatically triggered through error analysis, and the self-adaptive capacity and reconstruction accuracy of the model to the sudden working condition are improved.

Inventors

CHENG HENG
WAN TAO
LEI ZHENGQI
WANG PENGNAN
LIU HAIBO
Bian Wenpeng
MAO YANPIAN
LIU YI
ZHAO YUNTIAN
QI ZHIYONG
ZHOU QIUJING
YUAN JIN

Assignees

中国水利水电科学研究院
中国长江电力股份有限公司

Dates

Publication Date: 20260512
Application Date: 20250910

Claims (8)

1. The dam multi-physical-field real-time reconstruction system based on the fusion of the finite element and the AI proxy model is characterized by comprising a finite element simulation data generation module, an AI proxy model training module, a sensor interface module and a physical-field real-time reconstruction module; The finite element simulation data generation module is used for constructing a finite element grid model based on dam geometric parameters and geological parameters, setting boundary conditions and generating a multi-physical field simulation data set; The AI proxy model training module is used for splitting the simulation data set into a training set, a verification set and a test set, constructing an AI proxy model by adopting a physically guided graph calculation sub-network, training, and learning a finite element input-output mapping relation; The sensor interface module is used for collecting actual measurement data of the sensor in real time, preprocessing the data and synchronizing multiple sources; The physical field real-time reconstruction module is used for driving the AI proxy model to reconstruct the physical field in real time by using measured data and triggering a model correction mechanism through error analysis; The finite element simulation data generation module is used for constructing a finite element grid model comprising a dam body and a dam foundation based on the geometric parameters and the geological parameters of the dam, adding initial conditions, boundary conditions and loads, and performing simulation calculation to generate a multi-physical-field simulation data set; the method specifically comprises the following steps: the geometric model construction unit is used for constructing a three-dimensional geometric model according to dam design data and geological parameters; the mesh subdivision unit is used for carrying out fine mesh subdivision on the geometric model and carrying out mesh encryption on key parts of the dam body; The boundary condition setting unit is used for adding initial conditions, boundary conditions and loads to construct a dam simulation model; The simulation calculation unit is used for inputting material parameters and strength parameters, performing finite element calculation and generating simulation data sets under different working conditions, wherein each working condition is not less than 100 groups of data; The AI proxy model training module splits the multi-physical field simulation data set into a training set, a verification set and a test set, builds and trains an AI proxy model by adopting PhysGeO-transducer mixed network, learns finite element input-output mapping relation, The method specifically comprises the following steps: The model architecture construction unit adopts PhysGeO-transform mixed architecture, wherein a graph structure encoder extracts spatial characteristics through a graph convolution operator, a multi-scale physical attention module captures a time sequence change rule, an input layer contains water level, temperature and deformation time sequence parameters, and an output layer generates a three-dimensional physical field cloud graph and a key part time path curve; the data set splitting unit splits the simulation data set into a training set, a verification set and a test set in a ratio of 8:1:1; the loss calculation unit is used for calculating the deviation between the model predicted value and the finite element simulation value by adopting a mean square error loss function; and the precision evaluation unit adopts average absolute error and normalized average absolute error to evaluate the precision of the model, and when NMAE times continuously exceed 5%, the model correction mechanism is triggered, and the network weight is finely adjusted by using measured data.
2. The real-time reconstruction system of a dam multiphysics based on the fusion of finite elements and an AI proxy model of claim 1, wherein the sensor interface module collects actual measurement data of various sensors in real time and processes the input monitoring data, the sensor interface module comprises a protocol conversion sub-module and a data synchronization sub-module, and the system specifically comprises: The multi-source sensor access unit is used for integrating heterogeneous sensor networks arranged in the dam foundation, the dam abutment and the dam foundation curtain area; The protocol conversion and integration unit adopts modularized hardware and a protocol converter to solve the problem of sensor communication isomerism; the data preprocessing unit is used for performing triple cleaning and standardization on the original sensor data; and the multisource data synchronization unit adopts a dynamic time warping algorithm (DTW) to align sensor data time stamps of different sampling frequencies.
3. The real-time reconstruction system of multiple physical fields of a dam based on fusion of finite elements and an AI proxy model of claim 1, wherein the real-time reconstruction module of the physical field drives the AI proxy model to reconstruct the physical field in real time by using measured data, and a model correction mechanism is triggered by error analysis; the method specifically comprises the following steps: the model deployment unit deploys the trained AI proxy model to the real-time computing server; The data receiving unit is used for receiving the standardized monitoring data output by the sensor interface module in real time; the physical field reconstruction unit drives an AI model based on the monitoring data and rapidly reconstructs a stress field, a temperature field and a seepage field into a plurality of physical fields; and the result output unit is used for outputting a physical field cloud picture and a key part time course curve.
4. A real-time reconstruction method of a dam multiphysics based on fusion of finite elements and an AI proxy model, wherein the method is applicable to the system as set forth in any one of claims 1 to 3, and the method comprises: Step 1, constructing a multi-physical-field finite element model comprising a dam body and a dam foundation based on dam design data and geological parameters, carrying out grid encryption treatment on key parts of the dam foundation to form gradient grid distribution, adding static/dynamic load boundary conditions and displacement constraint conditions, carrying out finite element simulation calculation, and generating a simulation data set comprising at least 12 working conditions of normal operation, historical highest water level, earthquake working conditions and extreme air temperature; Step 2, splitting the simulation data set into a training set, a verification set and a test set according to the proportion of 8:1:1, constructing an AI proxy model by adopting a physically guided graph calculation sub-network, and training; step 3, acquiring environmental data and dam monitoring data in real time by utilizing a sensor network of which the dams are arranged in key areas of the dam body and the dam foundation, wherein the sensor network comprises a water level gauge, a thermometer, an osmometer, a strain gauge and a positive and negative vertical line; And 4, driving the AI proxy model to reconstruct the physical field in real time by using measured data, and when the normalized average absolute error exceeds 5% three times continuously, automatically triggering a dynamic error correction mechanism based on physical guide countermeasure training by the system, and finely adjusting the network weight of the model so as to improve the prediction accuracy.
5. The method for reconstructing the dam multiphysics based on the fusion of the finite element and the AI proxy model in real time is characterized in that the graph-computing sub-network in the step 2 initializes weights through Chebyshev polynomials, and the multiscale physical attention module reorganizes features in a Fourier domain to enhance the perception capability of the dam heel stress gradient.
6. The method for reconstructing the dam multi-physical-field real-time based on the fusion of the finite element and the AI proxy model is characterized in that the multi-scale physical attention module comprises a frequency domain decomposition layer, a physical query weighting layer and a characteristic reconstruction layer, wherein the frequency domain decomposition layer is used for extracting and separating physical field characteristics through fast Fourier transformation, the physical query weighting layer takes a water level change rate as a weight distribution coefficient, the characteristic reconstruction layer is used for carrying out inverse transformation reconstruction on characteristic vectors, and the sensing and distinguishing capacity of the model on physical field anomalies of local key parts of a dam heel is remarkably improved.
7. The method for real-time reconstruction of multiple physical fields of a dam based on fusion of finite elements and an AI proxy model as set forth in claim 4, wherein the error correction mechanism comprises a layered defrosting strategy, and further comprises a physical guidance countermeasure training mechanism, specifically, when a normalized average absolute error exceeds a defined threshold and conventional fine tuning is invalid, an automatic defrosting model multi-scale physical attention module performs network weight optimization by generating a countermeasure training sample conforming to physical constraints, and finally rapid adaptation and high-precision reconstruction of physical field changes under sudden working conditions are realized.
8. The method for reconstructing the dam multiphysics based on the fusion of finite elements and the AI proxy model in real time is characterized in that multisource monitoring data fusion is used for carrying out alignment processing on a sensor asynchronous sampling time sequence by using a dynamic time-warping sliding window matching algorithm, so that time scales of generated characteristic sequences are consistent, influences caused by different dimensions and sensor types are eliminated through Z-score normalization, and high-quality and low-noise real-time data are provided for the input of a subsequent AI proxy model.

Description

Real-time reconstruction method and system for multiple physical fields of dam based on fusion of finite element and AI proxy model Technical Field The invention belongs to the technical field of hydraulic engineering safety monitoring, and particularly relates to a real-time reconstruction method and system for multiple physical fields of a dam based on fusion of finite elements and an AI proxy model. Background The dam is used as an important water conservancy infrastructure, and physical field monitoring of the dam faces the following technical bottlenecks: 1. The finite element analysis efficiency is low, millions of grids are required to be built for dam fine modeling by the traditional finite element method, single simulation takes more than 24 hours, linear elastic FEM (FEM) calculates distortion in singular points (such as a dam heel) and further increases time consumption by depending on a nonlinear model; 2. The existing sensor arrangement is limited by cost, so that the coverage of a dam monitoring area is insufficient, and particularly, the monitoring of the internal physical field of the dam body is difficult; 3. the data heterogeneity problem is that the sampling frequency of the sensor data is inconsistent, the output signal form and the communication protocol are also inconsistent, and the heterogeneous data are difficult to be effectively fused by the traditional monitoring method; 4. The internal field quantity is not measurable, key parameters such as a dam foundation seepage field cannot be directly measured through a traditional sensor, an estimation error is large depending on an empirical formula, and an observation instrument buried in the dam body or the dam foundation cannot be replaced or is extremely high in replacement cost after being damaged, so that thorough perception of a physical field in the dam is lacked. 5. The traditional CNN-LSTM combined architecture has obvious limitations, mainly has the advantages that space-time characteristics are fused in a simple cascade mode, physical characteristic differences of a stress field (space leading) and a seepage field (time sequence leading) are not considered, a loss function only pays attention to data fitting, physical conservation law constraint is ignored, so that extrapolation errors of extreme working conditions are increased, global retraining is needed for model updating, and abnormal conditions such as local material degradation of a dam are difficult to adapt to. Disclosure of Invention Aiming at the problems, the invention provides a dam physical field simulation method with high precision and real-time performance, which realizes the rapid reconstruction of the whole physical field of the dam by deeply fusing finite element simulation and artificial intelligence, thereby solving the problems of low efficiency, incomplete monitoring coverage and the like of the traditional finite element simulation. The system comprises a finite element simulation data generation module, an AI proxy model training module, a sensor interface module and a physical field real-time reconstruction module; The finite element simulation data generation module is used for constructing a finite element grid model based on dam geometric parameters and geological parameters, setting boundary conditions and generating a multi-physical field simulation data set; The AI proxy model training module is used for splitting the simulation data set into a training set, a verification set and a test set, constructing an AI proxy model by adopting a physically guided graph calculation sub-network, training, and learning a finite element input-output mapping relation; The sensor interface module is used for collecting actual measurement data of the sensor in real time, preprocessing the data and synchronizing multiple sources; The physical field real-time reconstruction module is used for driving the AI proxy model to reconstruct the physical field in real time by using measured data and triggering a model correction mechanism through error analysis; In one scheme, the finite element simulation data generation module constructs a finite element grid model comprising a dam body and a dam foundation based on geometric parameters and geological parameters of the dam, adds initial conditions, boundary conditions and loads, carries out simulation calculation and generates a multi-physical-field simulation data set; the method specifically comprises the following steps: the geometric model construction unit is used for constructing a three-dimensional geometric model according to dam design data and geological parameters; the mesh subdivision unit is used for carrying out fine mesh subdivision on the geometric model and carrying out mesh encryption on key parts of the dam body; The boundary condition setting unit is used for adding initial conditions, boundary conditions and loads to construct a dam simulation model; And the simulation calculation unit is used fo