A method for constructing a digital twin model of a high earth-rockfill dam

Abstract

The present invention provides a method for constructing a digital twin model of a high earth-rockfill dam, which belongs to the technical field of health monitoring of high earth-rockfill dams. The method includes constructing a measuring point monitoring model based on sensor monitoring data; constructing a mapping model, and constructing a dynamic monitoring loss function of the mapping model that adopts a domain adaptive method for adversarial training; inputting the predicted values ​​of the deformation monitoring effect quantity and the seepage pressure monitoring quantity of the measuring point monitoring model for the seepage process and deformation process of the high earth-rockfill dam into the mapping model to obtain a digital twin model, and constructing an extreme working condition digital twin model based on the digital twin model; using knowledge graph reasoning combined with the dam slope stability coefficient to realize the interaction of the extreme working condition digital twin model with the physical model. Based on deep learning and finite element method, the present invention proposes a deformation-seepage digital twin framework of high earth-rockfill dam for the first time, in order to achieve a more economical, efficient and reliable monitoring solution, and provide technical support for the future structural health monitoring of high earth-rockfill dams.

Description

Technical Field

[0001] The present invention belongs to the technical field of health monitoring of high earth-rockfill dams, and in particular relates to a method for constructing a digital twin model of a high earth-rockfill dam. Background Art

[0002] High earth-rockfill dams are widely used in water conservancy projects worldwide due to their low cost, strong adaptability, and simple construction. In recent years, a number of high earth-rockfill dam projects have been completed, such as Gushui, Shuangjiangkou, Rumei, and Lianghekou. Their safety is crucial to the safety of people and property. However, the complex geological environment and high stress conditions make structural health monitoring of high earth-rockfill dams challenging. Currently, point-based monitoring models are the mainstream approach for high earth-rockfill dam health monitoring. Statistical models (such as HST) or shallow machine learning methods are often used, but these methods struggle to capture complex multi-feature relationships and suffer from insufficient generalization. Deep learning models (such as LSTM, GRU, and GCN) have been applied in dam safety monitoring due to their powerful nonlinear modeling capabilities. However, these models generally perform poorly under extreme operating conditions. Transfer learning can improve model prediction accuracy and generalization through knowledge transfer, providing support for the reliability of deep learning models under unknown extreme operating conditions. Furthermore, digital twin technology holds promise for comprehensive monitoring of high earth-rockfill dams, but currently faces challenges in constructing physical digital models, integrating sensor data, and ensuring the flow of data between the digital and physical worlds. In summary, existing monitoring models for measuring points struggle to meet these requirements, and the application of digital twin technology remains in its infancy. Summary of the Invention

[0003] In response to the above-mentioned deficiencies in the existing technology, the present invention provides a method for constructing a digital twin model of a high earth-rockfill dam, aiming to solve the technical difficulties in digital representation of the physical model of a high earth-rockfill dam and fusion of monitoring information, focusing on breaking through the deficiencies of existing methods in terms of real-time performance, accuracy, intelligent reasoning and safety warning.

[0004] In order to achieve the above objectives, the technical solution adopted by the present invention is: a method for constructing a digital twin model of a high earth-rockfill dam, comprising the following steps:

[0005] S1. Build a monitoring model based on sensor monitoring data;

[0006] S2. Using a long short-term memory neural network (LSTM) framework with measurement point connections, we learn the information transmission relationship between the finite element node seepage pressure or deformation value corresponding to the sensor measurement point, the actual sensor monitoring value, and the seepage pressure or deformation value of the surrounding nodes, respectively, to build a mapping model and a dynamic monitoring loss function for the mapping model.

[0007] S3. Input the predicted values ​​of the deformation monitoring effect and seepage pressure monitoring quantity of the high earth-rockfill dam's seepage and deformation processes from the measurement point monitoring model into the mapping model. Use domain adaptation for adversarial training, combine it with the measurement point monitoring model to obtain a digital twin model, and build a digital twin interaction model for extreme working conditions based on the digital twin model.

[0008] S4. Construct a high earth-rockfill dam inference dataset based on the knowledge graph, use the high earth-rockfill dam inference dataset to train the extreme working condition digital twin interaction model, and use the dam body global deformation monitoring effect and seepage pressure monitoring prediction value output by the extreme working condition digital twin interaction model as input. Use the trained extreme working condition digital twin interaction model output to obtain the cross-section treatment measures for the physical project, and complete the construction of the high earth-rockfill dam digital twin model.

[0009] The beneficial effects of the present invention are: based on deep learning and finite element methods, the present invention proposes for the first time a deformation-seepage digital twin framework for high earth-rockfill dams, in order to achieve a more economical, efficient and reliable monitoring solution, and provide technical support for future structural health monitoring of high earth-rockfill dams.

[0010] Furthermore, the S1 includes the following steps:

[0011] Construct a graph convolutional feature extractor: Extract environmental influencing factors and use them as input variables. Use GCN to adopt a feature extractor-decoder architecture to construct a graph convolutional feature extractor.

[0012] Percolation-deformation feature integration: Use the graph convolution feature extractor to construct the percolation feature extractor and the deformation feature extractor respectively, extract the environment-percolation influence features and the environment-deformation influence features, and then perform splicing to obtain the percolation-deformation coupling features. Among them, the seepage-deformation coupling characteristics Use a shared trainable feature decoder to simultaneously output seepage monitoring prediction values and deformation monitoring prediction values

[0013]

[0014] Where σ represents a nonlinear activation function, W represents a trainable weight matrix, and b represents a bias;

[0015] Physical feature constraints under extreme working conditions: Supervised learning of the hydraulic load characteristics and soil-rock creep characteristics of the earth-rock dam deformation feature extractor is performed. The loss function of the multi-teacher distillation in the soil-rock deformation feature extractor is expressed as follows:

[0016] L MYD =αL task (l S ,t)+(1-α)LKD (p(f s ,T),E(p(f T H ,T),p(f T θ,T)))

[0017] L KD =DF(p(f S ,T),p(f T ,T))

[0018] L=αL task +(1-α)L KD

[0019] p(f,T)=solfmax(f / T)

[0020] Among them, L MYD represents the loss function of the soil-rock deformation feature extractor, α represents the loss balance parameter, and L task represents the loss of regression prediction of the student network of the earth-rock dam deformation feature extractor, l S represents the student network deformation prediction value, t represents the deformation measured value, L KD represents the gap between the teacher network and the student network, p represents the distillation loss calculation method, and f s represents the characteristics of the student network, T represents the temperature parameter, E() represents the integrated output function of the teacher network, and f T H represents the hydraulic load characteristics of the teacher network, f T θ represents the soil creep characteristics of the teacher network, DF() represents the distance metric function, and f T represents the teacher network features, f represents the extracted deformation features, and L represents the calculation method of the knowledge distillation loss function;

[0021] Time-varying effect model update: By parameter-sharing and fine-tuning the feature decoder trained on the long-term monitoring data source task on the new sensor monitoring dataset, the construction of the measurement point monitoring model is completed. The measurement point monitoring model outputs the predicted values ​​of the deformation monitoring effect and the seepage pressure monitoring.

[0022] The technical effects of the above further solution are: achieving more accurate prediction of the dam structure condition under extreme water level conditions and achieving continuous iterative updates to improve the prediction performance of time-varying parameter problems caused by particle breakage.

[0023] Furthermore, the expression of the graph convolutional feature extractor is as follows:

[0024]

[0025] Among them, H0 represents the environmental features input to the graph convolutional feature extractor, and represents the set of hydraulic load influence factors, temperature influence factors and time-effect influence factors at time t, W (l) and W 0 Both represent trainable weight matrices, H (l+1) represents the environmental impact factor feature matrix of the l-th layer node, σ represents the nonlinear activation function, represents the normalized degree matrix, Represents the adjacency matrix of the impact factor plus self-connection, H (l+1) and H (l) They represent the environmental impact factor feature matrices of the l+1th layer nodes and the lth layer nodes respectively, A represents the impact factor adjacency matrix, and I represents the unit matrix.

[0026] The technical effect brought about by the above further solution is: using a graph convolutional feature extractor to extract the complex correlation between numerous environmental quantity influencing factors.

[0027] Furthermore, the expression of the seepage-deformation coupling characteristic is as follows:

[0028]

[0029] in, represents the seepage-deformation coupling characteristics, Indicates the environmental-seepage influence characteristics, Represents the environment-deformation influence characteristics.

[0030] The technical effect brought about by the above further solution is: through the seepage-deformation coupling feature, the mutual coupling modeling of the seepage process and the deformation process is realized.

[0031] Furthermore, the expression of the mapping model is as follows:

[0032]

[0033] θ δ =(W,b)

[0034] in, represents the seepage pressure value of the nodes around the finite element node, δ Monitoring point Indicates the measurement value of the monitoring point, Model LSTM () represents a model built based on the long short-term memory neural network LSTM, represents the calculated value of the finite element node corresponding to the sensor measurement point, ⊙ represents the element multiplication, W, b and θ δ All represent trainable parameters, and [] represents feature fusion.

[0035] The technical effect brought about by the above further solution is: the present invention uses a mapping model with measurement point connections to achieve a breakthrough transition from point monitoring to full-area engineering monitoring, and improves prediction accuracy.

[0036] Furthermore, the expression of the dynamic monitoring loss function of the mapping model is as follows:

[0037]

[0038] L Digital twin =λ1L accuracy +λ2L space

[0039]

[0040] Among them, θ Digital twin represents the parameters obtained by domain adaptive adversarial training of the mapping model, L Digital twin represents the loss function of the mapping model for domain adaptive adversarial training, θ represents the trainable parameters of the mapping model, λ1 and λ2 represent weight parameters, which are used to balance the effects of accuracy loss and spatial variability loss, and L accuracy Represents the accuracy loss function, L space represents the spatial variability loss function, N represents the number of finite element nodes, l represents the loss calculation method, i represents the cyclic calculation, and represents the spatial gradient between the predicted value of the mapping model and the numerical simulation result, α represents the loss balance parameter, Indicates the loss when the input is the sensor measurement point, Indicates the loss when the input is the sensor measurement point corresponding to the finite element node, represents the loss of all finite element nodes.

[0041] The technical effect brought about by the above further solution is: the present invention realizes the dynamic deep fusion of numerical simulation calculation results and sensor measurement results through the loss function of the mapping model.

[0042] Furthermore, the expression of the extreme working condition digital twin model is as follows:

[0043]

[0044] in, Indicates the parameters after fine-tuning the extreme working condition data parameters, θ δ represents the trainable parameters for fine-tuning parameters under extreme working conditions, L() represents the loss calculation method, represents the predicted value of the digital twin model under extreme working conditions, represents the calculated value of the finite element model under extreme working conditions, represents the seepage prediction value of the monitoring model at the measuring point, represents the deformation prediction value of the monitoring model at the measuring point, θ Digital twin represents the parameters obtained by domain adaptive adversarial training under the time series monitoring data of the mapping model, f Digital twin Represents a digital twin model.

[0045] The beneficial effect of the above further solution is that the present invention optimizes the parameter fine-tuning strategy under extreme working conditions to achieve accurate extrapolation of the digital twin model to the deformation and seepage conditions under extreme working conditions.

[0046] Furthermore, the expression of the digital twin interaction model is as follows:

[0047]

[0048] J(θ)=|F s -f|

[0049] Among them, P() represents the failure mode of the digital twin model, S represents the treatment method obtained by reasoning in the high soil and rock failure knowledge graph, represents the predicted value of the digital twin model under extreme working conditions, σ represents the activation function, and W s and b s represents the trainable parameters, θ t+1 represents the optimized parameters, θ t Parameter W representing the digital twin interaction model s and b s Set, η represents the learning rate, J(θ t ) represents the loss function, F s It represents the real-time calculated dam slope stability safety factor, and f represents the recommended value of the dam slope stability safety factor in relevant specifications.

[0050] The beneficial effect of the above further scheme is: the present invention realizes the guidance of the digital twin model on the physical engineering treatment measures in the case of structural damage through the digital twin interaction model enhanced by the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flow chart of the method of the present invention.

[0052] Figure 2 This is the layout diagram of the deformation and seepage measurement points for monitoring section 210 of PB Dam.

[0053] Figure 3A schematic diagram of the process line for monitoring the measured values ​​at the measuring points.

[0054] Figure 4 This is a table of mesh division and partitioning of the finite element model.

[0055] Figure 5 Schematic diagram comparing the deformation prediction accuracy of the proposed model, the finite element model, and the ablation model.

[0056] Figure 6 Schematic diagram comparing the seepage prediction accuracy of the proposed model and the finite element model.

[0057] Figure 7 Schematic diagram comparing the prediction accuracy of the measurement point monitoring model and the baseline model. DETAILED DESCRIPTION

[0058] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0059] Example

[0060] This invention aims to address the technical challenges of digitally representing the physical model of high earth-rockfill dams and integrating monitoring information, focusing on overcoming the shortcomings of existing methods in terms of real-time performance, accuracy, intelligent reasoning, and safety early warning. First, existing digital models have poor adaptability to environmental factors (such as water level fluctuations, temperature effects, and time-dependent effects), making it difficult to accurately describe the seepage and deformation states of the dam under complex operating conditions. To address this issue, this invention proposes a digital twin framework for high earth-rockfill dam structures based on deep learning, transfer learning, and physical constraints. This framework enables the model to adapt to environmental changes, improving the real-time reflection and prediction accuracy of the dam's state. Second, the integration of monitoring data with finite element simulation data has flaws, resulting in deviations in the model's mapping of the dam's true state. In traditional methods, the node relationships in finite element calculations are difficult to accurately match with the monitoring data, affecting the accuracy of dam safety assessments. This invention utilizes a long short-term memory neural network (LSTM) with measurement point connections and a dynamic monitoring loss function, combined with the finite element method to establish a mapping model. This optimizes the integration of monitoring data and numerical calculations, improving data matching accuracy and computational efficiency. Furthermore, existing technologies lack the intelligent reasoning and feedback capabilities between digital twin models and physical dams, making it difficult to effectively utilize monitoring data to guide dam maintenance and optimize operation. This invention combines slope stability analysis with knowledge graph reasoning technology to establish an intelligent feedback mechanism, enabling the digital twin model to infer the safety status of the dam structure based on real-time monitoring data, providing a scientific basis for project management and decision-making. Finally, under extreme conditions (such as heavy rain), the dam stability prediction and safety warning capabilities are limited, and traditional monitoring systems struggle to provide reliable warning information. This invention employs a transfer learning approach to enable the model to update time-varying effects and extrapolate predictions for extreme conditions, thereby enhancing the dam's safety warning capabilities and providing technical support for disaster prevention and mitigation. The main objective of this invention is to construct a digital twin framework for high earth-rockfill dams based on deep learning, finite element methods, and knowledge graph reasoning, enabling high-precision real-time perception and prediction of the dam's seepage and deformation states. Furthermore, deep learning methods incorporating physical constraints improve the matching between monitoring data and numerical calculations, ensuring that the digital model accurately maps to the actual dam. By integrating slope stability analysis with intelligent reasoning mechanisms, reliable assessments and early warnings are provided for dam structural safety, enhancing disaster prevention and mitigation capabilities. Furthermore, the model's enhanced time-varying update capabilities enable it to adapt to long-term operating conditions and provide extrapolated predictions under extreme conditions, providing intelligent support for dam operation and maintenance.

[0061] like Figure 1 As shown, the present invention provides a method for constructing a digital twin model of a high earth-rockfill dam, and its implementation method is as follows:

[0062] S1. Construct a monitoring model based on sensor monitoring data. The implementation method is as follows:

[0063] Construct a graph convolutional feature extractor: Extract environmental influencing factors and use them as input variables. Use GCN to adopt a feature extractor-decoder architecture to construct a graph convolutional feature extractor.

[0064] Percolation-deformation feature integration: Use the graph convolution feature extractor to construct the percolation feature extractor and the deformation feature extractor respectively, extract the environment-percolation influence features and the environment-deformation influence features, and then perform splicing to obtain the percolation-deformation coupling features. Among them, the seepage-deformation coupling characteristics A shared trainable feature decoder is used to simultaneously output the prediction value of seepage monitoring and deformation monitoring prediction values

[0065] Physical feature constraints of extreme working conditions: supervised learning of hydraulic load characteristics and soil-rock creep characteristics of earth-rock dam deformation feature extractor;

[0066] Time-varying effect model update: By parameter-sharing and fine-tuning the feature decoder trained on the long-term monitoring data source task on the new sensor monitoring dataset, the construction of the measurement point monitoring model is completed. The measurement point monitoring model outputs the predicted values ​​of the deformation monitoring effect and the seepage pressure monitoring.

[0067] In this embodiment, a graph convolutional feature extractor is constructed. First, environmental impact factors are extracted as input variables based on the HST model, and a feature extractor-decoder architecture is adopted based on GCN.

[0068]

[0069] Among them, H 0 represents the environmental features input to the graph convolutional feature extractor, and represents the set of hydraulic load influence factors, temperature influence factors and time-effect influence factors at time t, W (l) and W 0 Both represent trainable weight matrices, H (l+1) represents the environmental impact factor feature matrix of the l-th layer node, σ represents the nonlinear activation function, represents the normalized degree matrix, Represents the adjacency matrix of the impact factor plus self-connection, H (l+1) and H (l)Denote the environmental impact factor feature matrices of the l+1th and lth layer nodes, respectively. A denotes the impact factor adjacency matrix, and I denotes the identity matrix. The essence of this formula is to normalize and aggregate the information of the impact factor adjacent nodes, so that the impact factor node characteristics of each layer depend not only on itself but also on the information of its neighboring impact factor nodes. The output of formula (1) is the input of formula (2).

[0070] In this embodiment, the percolation-deformation feature integration: the graph convolution feature extractor is used to construct the percolation feature extractor and the deformation feature extractor respectively, and the environment-percolation impact feature is extracted. and environmental-deformation influence characteristics Then the feature fusion is achieved by splicing:

[0071]

[0072] in, represents the seepage-deformation coupling characteristics, Indicates the environmental-seepage influence characteristics, Represents the environment-deformation influence characteristics.

[0073] Seepage-deformation coupling characteristics A shared trainable feature decoder is used to simultaneously output the seepage monitoring prediction value and deformation monitoring prediction values

[0074]

[0075] Among them, σ represents a nonlinear activation function, W represents a trainable weight matrix, and b represents a bias.

[0076] In this embodiment, supervised learning is performed on the hydraulic load characteristics and soil-rock creep characteristics of the earth-rock dam deformation feature extractor at the same time. The loss function of the multi-teacher distillation process can be expressed as:

[0077] L MYD =αL task (l S ,t)+(1-α)L KD (p(f s ,T),E(p(f T H ,T),p(f T θ ,T))) (5)

[0078] L KD =DF(p(f S ,T),p(f T ,T)) (6)

[0079] L=αL task +(1-α)L KD (7)

[0080] p(f,T)=solfmax(f / T) (8)

[0081] Among them, L MYD represents the loss function of the soil-rock deformation feature extractor, α represents the loss balance parameter, and L task represents the loss of regression prediction of the student network of the earth-rock dam deformation feature extractor, l S represents the student network deformation prediction value, t represents the deformation measured value, L KD represents the gap between the teacher network and the student network, p represents the distillation loss calculation method, and f s represents the characteristics of the student network, T represents the temperature parameter, E() represents the integrated output function of the teacher network, and f T H represents the hydraulic load characteristics of the teacher network, f T θ represents the soil creep characteristics of the teacher network, DF() represents the distance metric function, and f T represents the teacher network features, f represents the extracted deformation features, and L represents the calculation method of the knowledge distillation loss function.

[0082] In this embodiment, the time-varying effect model is updated by performing parameter sharing and fine-tuning on the feature decoder trained on the long-term monitoring data source task on a new recent monitoring dataset to improve the model's prediction performance.

[0083] θ shared (W decoder ,b decoder )=θ pretrain (W decoder ,b decoder ) (9)

[0084]

[0085] Among them, θ pretrain represents the training parameters of the feature decoder, θ shared and θ update Denote the shared and updated parameters respectively, f Cell boby (·) represents the monitoring model of the measuring point, L Recent monitoring data Represents the loss function for updating the most recent monitoring data task, W decoder and b decoder represents the trainable parameters, θ pretrain represents the obtained parameters, θ represents the set of trainable parameters, They represent the set of hydraulic load influence factors, temperature influence factors and time-effect influence factors at time t respectively.

[0086] S2. Using a long short-term memory neural network (LSTM) framework with measurement point connections, we learn the information transmission relationship between the finite element node seepage pressure or deformation value corresponding to the sensor measurement point, the actual sensor monitoring value, and the seepage pressure or deformation value of the surrounding nodes, respectively, to build a mapping model and a dynamic monitoring loss function for the mapping model.

[0087] In this embodiment, an LSTM framework with measurement point connections is used to learn the finite element node seepage pressure or deformation value corresponding to the sensor measurement point. Actual monitoring value of the sensor δ Monitoring point The seepage pressure or deformation value of the nodes around it The information transmission relationship between them is used to obtain the mapping model:

[0088]

[0089] θ δ =(W,b) (12)

[0090] in, represents the seepage pressure value of the nodes around the finite element node, δ Monitoring point Indicates the measurement value of the monitoring point, Model LSTM () represents the mapping model based on the long short-term memory neural network LSTM, represents the calculated value of the finite element node corresponding to the sensor measurement point, ⊙ represents the element multiplication, W, b and θ δ All represent trainable parameters, and [] represents feature fusion.

[0091] In this embodiment, a dynamic monitoring loss function is constructed. During the model training process, a dynamic monitoring loss function is used to adjust the mapping model's learning of the association between the finite element mesh nodes and the actual sensor monitoring. The loss function is divided into the accuracy loss function L accuracy and spatial variability loss function L space Two parts:

[0092]

[0093] L Digital twin =λ1L accuracy +λ2L space (14)

[0094]

[0095] Among them, θ Digital twinrepresents the parameters obtained by domain adaptive adversarial training of the mapping model, L Digital twin represents the loss function of the mapping model for domain adaptive adversarial training, θ represents the trainable parameters of the mapping model, λ1 and λ2 represent weight parameters, which are used to balance the effects of accuracy loss and spatial variability loss, and L accuracy Represents the accuracy loss function, L space represents the spatial variability loss function, N represents the number of finite element nodes, l represents the loss calculation method, i represents the cyclic calculation, and represents the spatial gradient between the predicted value of the mapping model and the numerical simulation result, α represents the loss balance parameter, Indicates the loss when the input is the sensor measurement point, Indicates the loss when the input is the sensor measurement point corresponding to the finite element node, represents the loss of all finite element nodes.

[0096] S3. Input the predicted values ​​of the deformation monitoring effect and seepage pressure monitoring quantity of the high earth-rockfill dam's seepage and deformation processes from the measurement point monitoring model into the mapping model. Use domain adaptation for adversarial training, combine it with the measurement point monitoring model to obtain a digital twin model, and build a digital twin interaction model for extreme working conditions based on the digital twin model.

[0097] In this embodiment, a digital twin model is constructed. The predicted values ​​of the seepage process and deformation process of the high earth-rockfill dam are input into the measuring point mapping model through the measuring point monitoring model to obtain the digital twin model:

[0098]

[0099] In this embodiment, a digital twin model for extreme working conditions is constructed. Extreme working conditions such as high water levels are simulated to improve the generalization of the digital twin model:

[0100]

[0101] in, Indicates the parameters after fine-tuning the extreme working condition data parameters, θ δ represents the trainable parameters for fine-tuning parameters under extreme working conditions, L() represents the loss calculation method, represents the predicted value of the digital twin model under extreme working conditions, represents the calculated value of the finite element model under extreme working conditions, represents the seepage prediction value of the monitoring model at the measuring point, represents the deformation prediction value of the monitoring model at the measuring point, θ Digital twin represents the parameters obtained by domain adaptive adversarial training under the time series monitoring data of the mapping model, fDigital twin Represents a digital twin model.

[0102] S4. Construct a high earth-rockfill dam inference dataset based on the knowledge graph, use the high earth-rockfill dam inference dataset to train the extreme working condition digital twin interaction model, and use the dam body global deformation monitoring effect and seepage pressure monitoring prediction value output by the extreme working condition digital twin interaction model as input. Use the trained extreme working condition digital twin interaction model output to obtain the cross-section treatment measures for the physical project, and complete the construction of the high earth-rockfill dam digital twin model.

[0103] In this embodiment, knowledge graph reasoning is combined with the dam slope stability coefficient to realize the interaction between the digital twin model and the physical model:

[0104]

[0105] J(θ)=|F s -f| (21)

[0106] Among them, P() represents the failure mode of the digital twin model, S represents the treatment method obtained by reasoning in the high soil and rock failure knowledge graph, represents the predicted value of the digital twin model under extreme working conditions, σ represents the activation function, and W s and b s represents the trainable parameters, θ t+1 represents the optimized parameters, θ t Parameter W representing the digital twin interaction model s and b s Set, η represents the learning rate, J(θ t ) represents the loss function, F s It represents the real-time calculated dam slope stability safety factor, and f represents the recommended value of the dam slope stability safety factor in relevant specifications.

[0107] The present invention will be further described below.

[0108] The PB Hydropower Station is a large-scale hydropower project primarily focused on power generation, with additional benefits in flood control, sediment retention, and other comprehensive utilization. The dam is a gravel-filled earth core rockfill dam with a crest elevation of 856m, a maximum dam height of 186m, a normal water level of 850m, and a dead water level of 790m. Figure 2As shown in the figure, the typical section of PB Dam 0+210 was studied in particular. A set of horizontal displacement gauges using tension wires were arranged at the downstream filtration layer, transition layer and rockfill at elevations of 808m, 758m and 731m. These gauges are numbered from CH1 to CH16, with a total of 16 measurement points, including deformation measurement points and seepage measurement points. Currently, there are three seepage pressure monitoring points on the downstream dam slope of this section, numbered P22, P29 and P32. The process line of the measured reservoir water level and the values ​​of some deformation and seepage monitoring points are shown in the figure below. Figure 3 As shown, Figure 3 Where "Date" represents time. A dataset was created using deformation and seepage monitoring data from the 0+210 section of the PB Dam from May 21, 2012, to May 23, 2024, as well as finite element simulation data. The data division ratios for training, validation, and testing are shown in Table 1, which also provides a dataset overview. To more accurately evaluate the performance of the proposed model, the following nine baseline models were selected for comparison: a classical statistical HST (water pressure-temperature-time dependence model), an FEM (finite element model), five baseline deep learning models (long short-term memory neural network (LSTM), GRU gated recurrent unit (GRU) model, GCN graph convolutional neural network (GCN) model, Transformer model, and TGCN spatiotemporal graph convolutional model), a shallow machine learning model (SVM support vector machine (SVM) model), and an ablation model. The ablation model uses only the LSTM framework as the mapping model.

[0109] Table 1

[0110]

[0111] The digital twin model of the typical section 0+210 of the PB Dam is established using the present invention. The specific steps are as follows:

[0112] (1) Finite element model establishment. In the finite element model, the actual conditions such as the dam zoning layout and the thick foundation cover are taken into account. In addition, based on the results of the field drilling test, the finite element model divides the dam into two areas, the upper and lower areas, with the dead water level of 790m as the boundary. The simulation range is twice the dam height upstream, downstream and below the dam foundation. The finite element model consists of 6572 units and 13348 nodes. The finite element mesh and material division are as follows: Figure 4As shown in the figure (main rockfill area: the main body and main load-bearing structure of the dam; core wall: the central anti-seepage body of the earth-rock dam composed of materials such as clay; secondary rockfill area: protects the stability of the main rockfill area and the downstream dam slope; cover layer: loose accumulations and sediments of various origins covering the bedrock; bedrock: intact new mineral rock beneath the surface weathering layer of the continental crust). The PB Dam is divided into the upstream levee area, transition layer, filter layer, gravel core wall, downstream secondary levee area, and downstream main levee area. These design parameter values ​​are shown in Tables 2 and 3, and the design permeability coefficient of each zone of the PB Dam is shown in Table 4. Table 2 shows the parameters of the PB Dam's seven-parameter creep model, Table 3 shows the parameters of the Duncan-Zhang EB constitutive model of the PB Dam, and Table 4 shows the permeability coefficient of the PB Dam.

[0113] Table 2

[0114]

[0115] Table 3

[0116]

[0117] Table 4

[0118]

[0119]

[0120] (2) Establishment of the monitoring model of measuring points. The monitoring model of measuring points is constructed, and the input is the hydraulic load factor (H, H) based on the HST (water pressure-temperature-time model). 2 ,H 3 ,H 4 ), temperature factor and the aging factor (θ, ln(θ), e -θ ), output the deformation monitoring effect size of CH1-CH16 and the predicted value of the osmotic pressure monitoring size of P10, P22, P29 and P32. Training was performed using Pycharm IDE (python integrated development environment) and Pytorch2.5.1 (open source deep learning framework) based on python3.10 under the Windows 10 system environment. The server configuration is as follows: CPU is Intel Core i7-13700, GPU is Nvidia RTX 4070SUPER. The learning rate is set to 0.001, the batch size is 64, the optimizer selects Adam optimizer to update the network parameters, the number of iterations is set to 500, the model hidden layer is 64, the graph convolution layer is 2, and the multilayer perceptron layer is 2. Compared with the baseline model, Figure 7 As shown, Figure 7In the figure, MAE stands for mean absolute error, RMSE stands for root mean square error, and MAPE stands for mean absolute error.

[0121] (3) Establishment of mapping model. The input of the mapping model is the deformation monitoring effect quantity numbered CH1-CH16 and the seepage monitoring quantity numbered P10, P22, P29 and P32 and their corresponding finite element node calculation values. The output is the predicted value of all finite element nodes. During the domain adaptive adversarial training process, the loss function is divided into two parts: spatial variability loss and precision loss (the loss of prediction of actual monitoring quantity, the loss of prediction of node calculation value corresponding to the layout measurement point, and the loss of prediction of all finite element node calculation value) to dynamically adjust the comprehensive learning of actual monitoring status and finite element calculation status. Training was performed using Pycharm IDE (python integrated development environment) and Pytorch2.5.1 (open source deep learning framework) based on python3.10 in the Windows 10 system environment. The server configuration is as follows: CPU is Intel Core i7-13700 and GPU is Nvidia RTX 4070SUPER. The learning rate is set to 0.001, the batch size is 64, the Adam optimizer is selected to update the network parameters, the number of iterations is set to 500, and the model hidden layer is 64.

[0122] (4) Construct a digital twin model for extreme working conditions. The deformation monitoring effect of CH1-CH16 and the predicted values ​​of the seepage monitoring of P10, P22, P29 and P32 output by the monitoring point model are input into the measuring point mapping model to output the deformation and seepage process of the cross section. Figure 5 and Figure 6 shown.

[0123] (5) Construct a digital twin interaction model. Based on the knowledge graph, a high earth-rock dam (digital twin status - high earth-rock dam failure - treatment measures) inference data set is constructed. The dam body global deformation monitoring effect and seepage pressure monitoring prediction values ​​output by the digital twin model are used as inputs, and the output section is used for the physical engineering treatment measures (such as adding drainage facilities, correcting the dam slope angle, water level scheduling measures, etc.).

[0124] This example applies a digital twin framework to a high earth-rockfill dam project. This method effectively realizes a digital twin of the deformation and seepage coupling process of a high earth-rockfill dam, with an average error of only 3.17% between the predicted deformation and actual monitoring data. Compared with traditional numerical simulation methods, this method significantly improves simulation accuracy, with deformation simulation accuracy increasing by 64.42%. Furthermore, the proposed dynamic monitoring loss method effectively improves the prediction accuracy of the fusion of the finite element model and actual monitoring data, increasing deformation prediction accuracy by 32.55%. Compared with the baseline model, the proposed monitoring point prediction model achieves the best prediction accuracy and better generalization capability. Results show that compared with the ablation model, the proposed deformation-seepage coupling prediction method can effectively improve the prediction accuracy of the monitoring point monitoring model by 19.44%. This method can more accurately and real-timely reflect the deformation and seepage processes of high earth-rockfill dams.

[0125] In summary, current deformation monitoring of high earth-rockfill dams primarily relies on measurement point models based on statistical methods or shallow machine learning. These models are limited in capturing the relationship between complex environmental variables and dam deformation characteristics, as well as in modeling temporal characteristics, making it difficult to fully reflect the operational status of high earth-rockfill dams. Furthermore, while deep learning models possess powerful nonlinear modeling capabilities, they generally generalize poorly to unknown extreme conditions, making them inadequate for real-time monitoring of high earth-rockfill dams. This paper proposes a digital twin framework for high earth-rockfill dams to address the limitations of existing monitoring methods. This framework combines graph convolutional neural networks, transfer learning, LSTM, and finite element methods to establish a measurement point model for integrated seepage and deformation monitoring, a measurement point mapping model, an interaction model, and an extrapolation prediction model for extreme conditions. This enables real-time perception, prediction, and early warning of the deformation and seepage status of high earth-rockfill dams. This framework dynamically updates the monitoring model, enhancing its extrapolation capabilities for unknown conditions. It also uses slope stability analysis to further assess structural safety and intelligently generate treatment measures. Compared with traditional numerical simulation methods, the proposed method improves simulation accuracy by integrating actual sensor monitoring information, avoids extensive numerical calculations, and exhibits real-time performance, improving timeliness. Compared with models based on sensor measurement points, its prediction accuracy and generalization capabilities are superior. This method achieves a breakthrough shift from traditional "point monitoring" of high earth-rockfill dam structures to "global monitoring." It can more accurately and in real time reflect the surface deformation and seepage processes of high earth-rockfill dams. It can also use mechanistic methods such as slope stability analysis to assess structural safety in real time, providing a cost-effective and efficient intelligent monitoring solution for practical projects and has broad application prospects.

Claims

1. A method for constructing a digital twin model of a high earth-rockfill dam, characterized in that: The following steps are involved: S1. Build a monitoring model based on sensor monitoring data; S2. Using a long short-term memory neural network (LSTM) framework with measurement point connections, we learn the information transmission relationship between the finite element node pressure or deformation value corresponding to the sensor measurement point, the actual sensor monitoring value, and the pressure or deformation value of the surrounding nodes, respectively, to build a mapping model and a dynamic monitoring loss function for the mapping model. The expression of the dynamic monitoring loss function of the mapping model is as follows: in, represents the parameters obtained by domain adaptive adversarial training of the mapping model, Represents the loss function of the mapping model for domain adaptive adversarial training, represents the trainable parameters of the mapping model, and Both represent weight parameters, which are used to balance the impact of accuracy loss and spatial variability loss. represents the accuracy loss function, represents the spatial variability loss function, represents the number of finite element nodes, Indicates the loss calculation method, i Indicates a loop calculation, and represents the spatial gradient between the predicted value of the mapping model and the numerical simulation result, represents the loss balance parameter, Indicates the loss when the input is the sensor measurement point, Indicates the loss when the input is the sensor measurement point corresponding to the finite element node, represents the loss of all finite element nodes; S3. Input the predicted values ​​of the deformation monitoring effect and seepage pressure monitoring quantity of the high earth-rockfill dam's seepage and deformation processes from the measurement point monitoring model into the mapping model. Use domain adaptation for adversarial training, combine it with the measurement point monitoring model to obtain a digital twin model, and build a digital twin interaction model for extreme working conditions based on the digital twin model. S4. Construct a high earth-rockfill dam inference dataset based on the knowledge graph, and use the high earth-rockfill dam inference dataset to train the extreme working condition digital twin interaction model. The dam body global deformation monitoring effect quantity and seepage pressure monitoring quantity prediction value output by the extreme working condition digital twin interaction model are used as input. The trained extreme working condition digital twin interaction model output is used to obtain the treatment measures for the cross-section physical project, and the construction of the high earth-rockfill dam digital twin model is completed.

2. The method for constructing a digital twin model of a high earth-rockfill dam according to claim 1, characterized in that: Said S1 comprises the following steps: Construct a graph convolutional feature extractor: Extract environmental influencing factors and use them as input variables. Use GCN to adopt a feature extractor-decoder architecture to construct a graph convolutional feature extractor. Infiltration-deformation feature integration: Use the graph convolution feature extractor to construct the infiltration feature extractor and the deformation feature extractor respectively, extract the environment-infiltration influence features and the environment-deformation influence features, and splice them to obtain the infiltration-deformation coupling features , where the seepage-deformation coupling characteristic Use a shared trainable feature decoder to simultaneously output seepage monitoring prediction values and deformation monitoring prediction values : in, represents a nonlinear activation function, represents the trainable weight matrix, Indicates bias; Physical feature constraints under extreme working conditions: Supervised learning of the hydraulic load characteristics and soil-rock creep characteristics of the earth-rock dam deformation feature extractor is performed. The loss function of the multi-teacher distillation in the soil-rock deformation feature extractor is expressed as follows: in, represents the loss function of the soil-rock deformation feature extractor, represents the loss balance parameter, represents the loss of regression prediction of the student network of the earth-rock dam deformation feature extractor, represents the student network deformation prediction value, represents the measured deformation value, represents the gap between the teacher network and the student network, Indicates the calculation method of distillation loss, represents the student network characteristics, represents the temperature parameter, represents the integrated output function of the teacher network, represents the hydraulic load characteristics of the teacher network, represents the soil-rock creep characteristics of the teacher network, represents the distance metric function, represents the teacher network characteristics, represents the extracted deformation features, Represents the calculation method of knowledge distillation loss function; Time-varying effect model update: By parameter-sharing and fine-tuning the feature decoder trained on the long-term monitoring data source task on the new sensor monitoring dataset, the construction of the measurement point monitoring model is completed. The measurement point monitoring model outputs the predicted values ​​of the deformation monitoring effect and the seepage pressure monitoring.

3. The method for constructing a digital twin model of a high earth-rockfill dam according to claim 2, characterized in that: The expression of the graph convolutional feature extractor is as follows: in, represents the environmental features input to the graph convolutional feature extractor, 、 and Indicates separation time t The combination of hydraulic load influence factor, temperature influence factor and time influence factor, and Both represent trainable weight matrices, represents a nonlinear activation function, represents the normalized degree matrix, Represents the adjacency matrix of the impact factor plus self-connection, and Respectively represent Layer nodes and The environmental impact factor characteristic matrix of the layer node, represents the impact factor adjacency matrix, Represents the identity matrix.

4. The method for constructing a digital twin model of a high earth-rockfill dam according to claim 2, characterized in that: The expression of the seepage-deformation coupling characteristic is as follows: in, represents the seepage-deformation coupling characteristics, Indicates the environmental-seepage influence characteristics, Represents the environment-deformation influence characteristics.

5. The method for constructing a digital twin model of a high earth-rockfill dam according to claim 1, characterized in that: The expression of the mapping model is as follows: in, represents the seepage pressure value of the nodes around the finite element node, Indicates the measurement value of the monitoring point, Represents a model built based on the long short-term memory neural network LSTM, Indicates the calculated value of the finite element node corresponding to the sensor measurement point, represents element-wise multiplication, 、 and All represent trainable parameters, and [] represents feature fusion.

6. The method for constructing a digital twin model of a high earth-rockfill dam according to claim 1, characterized in that: The expression of the extreme working condition digital twin model is as follows: in, Indicates the parameters after fine-tuning the extreme working condition data parameters. represents the trainable parameters for fine-tuning parameters under extreme working conditions, Indicates the loss calculation method, represents the predicted value of the digital twin model under extreme working conditions, represents the calculated value of the finite element model under extreme working conditions, represents the seepage prediction value of the monitoring model at the measuring point, represents the deformation prediction value of the monitoring model of the measuring point, represents the parameters obtained by domain adaptive adversarial training under the time series monitoring data of the mapping model, Represents a digital twin model.

7. The method for constructing a digital twin model of a high earth-rockfill dam according to claim 1, characterized in that: The expression of the digital twin interaction model is as follows: in, represents the failure mode of the digital twin model, represents the treatment method obtained by reasoning in the high soil and rock failure knowledge graph, represents the predicted value of the digital twin model under extreme working conditions, represents the activation function, and represents the trainable parameters, represents the optimized parameters, Parameters representing the digital twin interaction model and gather, represents the learning rate, represents the loss function, represents the real-time calculated dam slope stability safety factor, Indicates the recommended value of the dam slope stability safety factor in relevant specifications.

Images