Rock-fill dam deformation digital twinborn body construction method based on generative AI

Abstract

The invention discloses a rock-fill dam deformation digital twin construction method based on generative AI, and the core is that a conditional denoising diffusion probability model is adopted, a conditional sampling mechanism is constructed through classifier-free guidance and continuous conditional vectors, and finite element simulation data, monitoring data and operation data are efficiently fused. And further combining a residual neural network ResNet-18 and a K-means clustering method to carry out finite element data unsupervised classification, introducing combinatorial optimization, and identifying a finite element data category which is most matched with monitoring data, so that the condition-guided deformation field is generated. The framework is applied and verified on the highest two-estuary rockfill dam (303 meters) in the current built world. The result shows that the rockfill dam deformation digital twinborn body construction method based on the generative AI can efficiently reconstruct rockfill dam deformation, has high precision and real-time performance, remarkably improves the global deformation thorough sensing ability of the rockfill dam, and provides key technical support for safe operation of the rockfill dam.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of water conservancy and geotechnical engineering, and particularly relates to a rock-fill dam deformation digital twin construction method based on generative AI. BACKGROUND

[0002] Rock-fill dams are an important type of large-scale hydraulic structure, and have become one of the widely used dam types due to their outstanding economy and adaptability. With the continuous development of rock-fill dam construction technology, the dam height has fully entered the 300-meter level, and the structure scale is huge, so the safety operation risk management is crucial. Among them, dam body deformation control is one of the key issues to ensure the safety of the whole life cycle of the rock-fill dam, and is highly concerned in various stages such as design, construction, operation and maintenance. The use of various monitoring technologies for in-situ deformation monitoring is an important basis for deformation control. However, due to factors such as sensor layout density, burial depth, environmental interference and instrument aging, the deformation monitoring data often presents the phenomena of sparse spatial distribution, discontinuity, and even failure, which seriously restricts the comprehensive perception and analysis of the true deformation state of the dam body. In addition, although traditional finite element analysis can analyze the overall deformation of the dam, it relies on parameter inversion, has low calculation efficiency and is sensitive to initial conditions, and it is difficult to meet the engineering needs of dynamic perception in the operation period.

[0003] In recent years, digital twinning has provided more and more intelligent and data support for the safety management of water conservancy engineering, and with the development of artificial intelligence, especially generative artificial intelligence (Generative AI), generative models have been gradually introduced into the field of structural health monitoring and physical field reconstruction to enhance the prediction capability. Although existing researches have adopted deep learning methods for dam body deformation prediction or anomaly identification, most of the models are limited to single-point prediction of existing monitoring points, and it is difficult to realize comprehensive reconstruction of the two-dimensional or even three-dimensional deformation field of the dam body, and the robustness to noise and uncertainty of monitoring data is lacking.

[0004] In summary, how to reconstruct the deformation field of the dam body with high precision and high efficiency under the condition of sparse monitoring data, by integrating the finite element physical prior information and the engineering operation data, to improve the comprehensive perception ability of the deformation of the rock-fill dam, has become a key technical problem that needs to be broken through in the current safety monitoring field of rock-fill dams. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the present application provides a rock-fill dam deformation digital twin construction method based on generative AI, which significantly improves the accuracy, efficiency and physical consistency of deformation reconstruction by using generative diffusion modeling, category guiding mechanism and continuous condition vector joint control generation process.

[0006] According to one aspect of the present application, a rock-fill dam deformation digital twin construction method based on generative AI is provided, comprising: A finite element numerical model was constructed based on deformation monitoring data and operational data of the rockfill dam, and multiple sets of deformation field simulation samples were obtained by adopting a parameter sampling strategy. Based on the obtained deformation field simulation samples, two-dimensional deformation field data of key monitoring sections are extracted, and depth features of two-dimensional deformation field images are extracted through deep neural networks. Unsupervised clustering methods are used to classify and label the simulation samples. A conditional denoising diffusion probability model is constructed. The model is trained by using the classified two-dimensional deformation field simulation samples, the corresponding category labels, and the running data used in the simulation as conditional inputs, and outputs a controllable deformation field reconstruction model. By combining actual deformation monitoring data with simulation data of the corresponding cross sections, a combined optimization framework is constructed to search for the optimal data category. The optimal data category and the running data are then input into a controllably generated deformation field reconstruction model, which outputs the reconstructed deformation field of the rockfill dam.

[0007] As a further technical solution, after extracting the two-dimensional deformation field data of the key monitoring sections based on the obtained deformation field simulation samples, the solution also includes: The extracted two-dimensional deformation field data is uniformly interpolated and mapped to an image matrix with a fixed resolution. The deformation values ​​are then converted into single-channel grayscale images using standardization processing.

[0008] As a further technical solution, depth features of two-dimensional deformation field images are extracted using deep neural networks, including: Using a pre-trained ResNet-18 network as the backbone, the fully connected classification layers at the ends are removed, and only the convolutional layers and global average pooling layers are retained to extract the depth feature vector of each two-dimensional deformation field image.

[0009] As a further technical solution, unsupervised clustering methods are used for the classification and labeling of simulation samples, including: The depth feature vectors corresponding to all two-dimensional deformation field images are input into the K-means clustering algorithm to perform unsupervised clustering operations, classify the simulation samples, and assign a unique class label to each simulation sample.

[0010] As a further technical solution, the training of the conditional denoising diffusion probability model includes: Gaussian noise is gradually added to the simulated deformation field data of each section of the rockfill dam through a forward diffusion process to obtain a noisy image; The noisy image is input into the U-Net network and trained to predict the noise components added to the original two-dimensional deformable field image; The trained U-Net network is used to gradually remove noise during the backdiffusion process, reconstructing the original deformable field image.

[0011] As a further technical solution, the method also includes: Discrete category labels are encoded into high-dimensional vectors through a fully connected layer, and a category guidance mechanism is used to randomly mask the category labels with a set probability during the training phase. The running data is transformed into feature representations through fully connected layers and fused with the category label embeddings under the category guidance mechanism to jointly guide the decoding path of the U-Net network.

[0012] As a further technical solution, a combined optimization framework is constructed by combining actual deformation monitoring data with simulation data of the corresponding cross-sections to search for the optimal data category. The optimal data category and the running data are then input into a controllably generated deformation field reconstruction model, including: The multi-category data extracted from the finite element simulation samples are used to construct a candidate set. A combined optimization model is introduced, and several categories that are closest to the measured deformation values ​​are selected as conditions for generation guidance. The optimal data category is obtained by using a genetic algorithm to optimize the combined optimization model.

[0013] The labels of the optimal data categories are combined with the running data to form a conditional vector, which is then input into a trained conditional denoising diffusion probability model for conditional sampling to obtain generative deformation field results that match the measured data.

[0014] According to one aspect of the present invention, a system for constructing a deformable digital twin of a rockfill dam based on generative AI is provided for implementing the method, the system comprising: The first main module is used to construct a finite element numerical model based on the deformation monitoring data and operation data of the rockfill dam, and to obtain multiple sets of deformation field simulation samples by adopting a parameter sampling strategy. The second main module is used to extract two-dimensional deformation field data of key monitoring sections based on the obtained deformation field simulation samples, extract the depth features of the two-dimensional deformation field images through deep neural networks, and classify and label the simulation samples using unsupervised clustering methods. The third main module is used to construct a conditional denoising diffusion probability model. It uses the classified two-dimensional deformation field simulation samples, the corresponding category labels, and the running data used in the simulation as conditional inputs to train the model and outputs a controllable deformation field reconstruction model. The fourth main module is used to combine actual deformation monitoring data with simulation data of the corresponding cross sections to construct a combined optimization framework, search for the optimal data category, and input the searched optimal data category and the running data into the controllably generated deformation field reconstruction model to output the reconstructed rockfill dam deformation field.

[0015] According to one aspect of the present invention, a device for constructing a deformable digital twin of a rockfill dam based on generative AI is provided, comprising a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute the method for constructing a deformable digital twin of a rockfill dam based on generative AI.

[0016] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the described method for constructing a deformable digital twin of a rockfill dam based on generative AI.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention supports comprehensive perception of the deformation of rockfill dams under sparse monitoring data conditions. This invention integrates finite element simulation data, in-situ monitoring records, and operational data, overcoming the limitations of traditional models that are highly dependent on monitoring points and have weak spatial reconstruction capabilities. It can still achieve high-precision reconstruction of the two-dimensional deformation field under conditions of limited monitoring data. The reconstruction results of the dam deformation field not only match the measured data in amplitude but also have good physical consistency in deformation time and space, supporting the safety assessment of the dam.

[0018] (2) This invention supports dynamic real-time reconstruction of the deformation field of rockfill dams. Compared with the traditional parametric inversion finite element method, which takes several hours for a single simulation, the model of this invention can complete the generation of a deformation field in a few seconds after training, improving the reconstruction efficiency by more than four orders of magnitude, and has the ability to monitor in real time and respond quickly.

[0019] (3) This invention has high robustness and strong generalization ability. It adopts a generative diffusion model training mechanism, which has stable noise modeling and prediction capabilities. It also achieves deformation field reconstruction under different monitoring data inputs through a combination optimization strategy, ensuring that reliable rockfill dam deformation fields can still be generated even when the number of monitoring points is reduced. This enhances the scalability and practical value of the method in actual rockfill dam projects. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a method for constructing a deformable digital twin of a rockfill dam based on generative AI, provided in an embodiment of the present invention; Figure 2 The following is a schematic diagram illustrating the application process of DDPM in the deformation field prediction of rockfill dams in the embodiments of the present invention: (a) Example of rockfill dam model and deformation field of core wall section; (b) Deformation field prediction using DDPM; Figure 3 This is a schematic diagram of the CDDPM model structure for reconstructing the deformation field of a rockfill dam according to an embodiment of the present invention; Figure 4 The following are schematic diagrams illustrating the relevant aspects of the Lianghekou rockfill dam used in this embodiment of the invention: (a) aerial photograph of the dam; (b) BIM model; (c) construction-operation process and material zoning. Figure 5 This is a schematic diagram of the process of generating the cross-sectional deformation field by performing 200 steps of progressive noise reduction through a U-Net network when using CDDPM in an embodiment of the present invention. Figure 6 The following is a schematic diagram of the deformation field reconstruction of the core wall and section 3-3 at a typical moment in an embodiment of the present invention: (a) the iterative process of the objective function in the combinatorial optimization; (b) the deformation field reconstructed by CDDPM; Figure 7 A schematic diagram showing the time consumption statistics of each part when applied to an embodiment of the present invention; Figure 8 The following is a schematic diagram of the statistical analysis of the deformation field reconstruction results of each section under the condition of not using CFG guidance in an embodiment of the present invention: (a) the average value of section 3-3; (b) the average value of the core wall section; (c) the median of section 3-3; (d) the median of the core wall section; (e) the maximum value of section 3-3; (f) the maximum value of the core wall section. Figure 9 This is a heatmap of the SSIM index under the condition of gradually reducing the number of measurement points in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0023] This invention provides a method for constructing a deformable digital twin of a rockfill dam based on generative AI, which improves the accuracy of rockfill dam simulation. It mainly consists of four parts: dataset preparation, finite element simulation data preprocessing, CDDPM model training, and CDDPM conditional sampling. (See reference...) Figure 1 The specific steps include: Step S1: Collect finite element deformation simulation data, operational data, and deformation monitoring data of the rockfill dam profile. The finite element simulation data is obtained through multiple finite element analyses. The operational data includes the construction and operation duration T and the reservoir water level. H R Deformation monitoring data is obtained from various monitoring equipment and technologies used in rockfill dam projects.

[0024] In step S1, the acquisition of deformation monitoring data, operational data, and finite element simulation data includes: Step S1.1: Various monitoring equipment and technologies, including internal deformation monitoring equipment such as water pipe settling meters, electromagnetic settling rings, and flexible inclinometers, and external monitoring technologies such as measuring robots, GNSS monitoring, and InSAR, are used to obtain in-situ deformation monitoring data. Step S1.2: Water level change data and construction and operation duration are typically provided by the rockfill dam project management unit. Both types of data are used not only in finite element numerical simulation analysis but also in subsequent model training and condition sampling processes. Step S1.3, obtaining finite element simulation data, involves constructing a finite element model and performing calculations using a suitable constitutive model and analysis software. This step involves constructing a finite element model of the rockfill dam and applying numerical sampling methods to generate multiple sets of constitutive model parameter combinations. These combinations are then substituted into the finite element model of the rockfill dam for numerical simulation, resulting in a large amount of finite element simulation data.

[0025] Step S1.3 further includes the following steps: (1) Based on the geometric structure, material zoning, construction conditions and boundary conditions of the rockfill dam, a three-dimensional finite element model is constructed. The element type used can be a three-dimensional solid element (such as C3D8), and the finite element analysis software used includes, but is not limited to, ABAQUS, ANSYS, etc. (2) Select a constitutive model suitable for describing the deformation behavior of riprap, such as the Duncan-Zhang EB model for simulating nonlinear elastic response, the creep model for characterizing long-term effects, and the wetting deformation model for describing volume change behavior under water action. (3) Using parameter design methods such as Latin Hypercube Sampling (LHS), multiple sets of representative constitutive parameter combinations are generated and substituted into the finite element model one by one for simulation calculation to reflect the influence of material parameter uncertainty on the dam deformation field. (4) Based on the actual filling and impounding process, set up loading steps for multiple time periods to simulate the deformation response of the rockfill dam from filling to operation. The simulation process may include the dam body response at multiple typical moments (such as construction completion, initial impoundment, maximum impoundment, etc.), and extract the two-dimensional deformation results of key sections; (5) Map each set of simulation results to two-dimensional deformation field data under a regular grid and perform global normalization to construct the standardized finite element simulation sample set required for training and generating the model.

[0026] Step S2 involves preprocessing the simulation data. First, the two-dimensional deformation field data of the key monitoring sections are extracted. Then, the depth feature vectors are extracted using a ResNet-18 residual neural network. Finally, the simulation data is classified using the K-means clustering method to obtain two-dimensional deformation field simulation samples with category labels.

[0027] In step S2, the simulation data is preprocessed, mainly including unsupervised deep clustering and data classification. The main steps are as follows: Step S2.1, image format standardization. The extracted two-dimensional deformation field data includes the x and y coordinates of each node ( x , y ) and corresponding deformation value ( δ ), constituting two-dimensional deformation field data ( x , y , δ The extracted two-dimensional deformation field data is uniformly interpolated and mapped to an image matrix with a fixed resolution (e.g., 128×128), and the deformation values ​​are converted into single-channel grayscale images by standardization processing (e.g., min-max normalization or [-1, 1] interval standardization) to facilitate subsequent neural network input processing.

[0028] Step S2.2, Deep Feature Extraction. The ResNet-18 residual neural network is used to extract features from the aforementioned two-dimensional deformation field images. Specifically, using a pre-trained ResNet-18 network as the backbone, the fully connected classification layers at the ends are removed, leaving only the convolutional layers and the global average pooling layer, to extract the deep feature vector of each image, representing its deformation structure features.

[0029] Step S2.3, Unsupervised Clustering Classification. Input the depth feature vectors corresponding to all 2D deformation field samples into the K-means clustering algorithm to perform unsupervised clustering and classify the simulation samples. The number of cluster centers K can be set according to sample diversity and training requirements (e.g., K = 50). The clustering result will assign a unique class label to each simulation sample for subsequent model training.

[0030] Step S2.4, Data Consistency Check. Ensure that the samples of each category are representative and diverse in terms of spatial morphology and deformation amplitude. If necessary, balance the number of samples in each category to improve the stability and generalization ability of the model training.

[0031] Step S3: The classified two-dimensional deformation field simulation data, the corresponding category labels, and the running data used in the simulation are used as input to train the conditional denoising diffusion probability model CDDPM, thereby obtaining a controllable deformation field reconstruction model.

[0032] In this step, to make the generated deformation field closer to the real deformation, CFG (Classifier-Free Guidance) and continuous conditional vectors are introduced. Multi-source monitoring data and running data are used as generation conditions for the model, guiding the model to generate a deformation field that better matches the actual measurement. The limited metadata category labels obtained from deep clustering and the running data at the corresponding time steps are used as the conditional inputs for training.

[0033] In step S3, CDDPM (Conditional Denoising Diffusion Probabilistic Model) training is performed, and the main steps include: Step S3.1, Model Input Construction: The following three types of data will be used as model input: (1) Two-dimensional deformation field simulation image: derived from the deformation field simulation sample after standardization in step S2, and represented as a single-channel grayscale image with a fixed resolution (e.g., 128×128); (2) Category label: The cluster category number assigned to each simulation sample, using one-hot or multi-hot encoding as discrete condition input; (3) Operational data: including total construction and operation time T Current water level H R These are combined to form a continuous conditional vector, which serves as a representation of the actual running data.

[0034] Step S3.2, Model Structure Design. This method uses CDDPM as the core model for generating the deformation field of the rockfill dam. Its core idea is to gradually add noise to the original data through a forward diffusion process and train a neural network model to gradually remove noise during the back diffusion process, thereby reconstructing the original data. The model consists of two main parts: the forward diffusion process and the parameter back diffusion process.

[0035] (1) Forward diffusion process: In this process, Gaussian noise is gradually added to the simulated deformation field data of each section of the rockfill dam. This process constitutes a Markov chain, that is, the deformation field data is gradually diffused into pure Gaussian noise. Specifically, from a uniformly distributed... U (1, 2,…, T Sample one time step in ) t and from the standard normal distribution N ( 0 , I Sampling in) t Sub-Gaussian noise ε 1, ε 2,…, ε t Add them sequentially to the initial image. x 0~ q ( x 0), of which x 0 represents the original two-dimensional deformation field image of a certain cross-section of the rockfill dam, from which the noisy image is finally obtained. x t The image was then fed into the U-Net network structure, and its predictions were trained and added to... x 0 noise component ε Its denoising objective function is:

[0036] In the formula, x 0 represents the initial time ( t = 0) Clean deformation field image, x t Indicates diffusion t A version with noise added after the step. β t Let be a fixed scalar parameter representing diffusion intensity, with a value range of (0,1). I N(0, ..., represents the standard normal distribution) I The covariance matrix in ).

[0037] (2) Parametric Backdiffusion Process: The backdiffusion process is the process of gradually restoring the deformed field image from a Gaussian noise image. Specifically, this process uses a standard normal distribution... N ( 0 , IPure noise image obtained by sampling from ) x T Starting from this point. At each diffusion step... t The current noisy image x t The input is fed into the U-Net network to predict its corresponding Gaussian noise component. At the same time, from N ( 0 , I Another noise term is sampled in the sample. z The image at the next time step is calculated using the reparameterization formula. x t-1 Through multiple iterations, the final target deformation field image is generated. x 0, its process can be represented as:

[0038] In the formula, ε t This represents the noise component predicted by the neural network; the entire process is equivalent to learning the mean vector. μ ( x t , t ) and covariance matrix Σ( x t , t ),parameter α t =1- β Indicates the first t The original image proportions are preserved in the first step, which is in the first step. t The cumulative product of the steps reflects the degree to which the original information is preserved overall.

[0039] Step S3.3, Conditional Information Embedding. Discrete class labels are encoded into high-dimensional vectors through a fully connected layer. Using the CFG mechanism, class labels are randomly masked with a certain probability during the training phase. c cat ∈{0, 1} C ,Right now:

[0040] This enables the model to flexibly learn and generalize from conditional information, guiding the generative model to learn different deformation patterns.

[0041] The continuous condition vector [ T , H R The noise is transformed into a feature representation, embedded into a neural network through independent fully connected layers, and then fused with the class label embedding. Conditional CDDPM learns to predict noise under both conditional and unconditional settings. During inference, U-Net predicts noise when CFG conditions are available.ε θ ( x t , c CFG ), and when blocked, it will predict ε θ ( x t During the sampling process, these two predictions are combined in a linear combination, as shown in the following equation:

[0042] In the formula, ω It represents the condition-guided intensity, which can flexibly control the influence of CFG conditions during the generation process and balance the diversity and fidelity of the reconstructed deformation field.

[0043] In addition, continuous condition vectors are extracted from the runtime data. c cont = [ T , H R This vector is embedded into the network through a fully connected layer and directly guides the CDDPM together with the CFG conditional label, as shown in the following equation:

[0044] clock, f cat (·)and f cont (·) indicates an independent fully connected layer. λ These are hyperparameters that control the influence of continuous conditional vectors. These fused embeddings collectively guide the decoding path of U-Net, shaping the spatial pattern and amplitude of the reconstructed deformation field. Compared with label-driven tuning, this dual-conditional mechanism enhances the model's adaptability to complex and variable dam operation scenarios, while achieving high-fidelity and controllable deformation reconstruction under sparse monitoring data.

[0045] Step S3.4, Model Training Mechanism: The model training follows the basic process of DDPM, adding Gaussian noise to the real image through the forward diffusion process, and then training the network to predict the added noise; the loss function adopts mean squared error (MSE) or its improved form to measure the error between the model's predicted noise and the real noise; to enhance model performance, training techniques such as automatic mixed precision (AMP), learning rate scheduler, gradient clipping, etc. can be introduced.

[0046] Step S3.5, Model Training Output. After training is complete, a CDDPM model is obtained that can be used for subsequent conditional sampling generation; the model has the ability to generate deformation field images with reasonable spatial distribution and good structural consistency based on specified category labels and running data.

[0047] Step S4: Combining actual monitoring data with simulation data of the corresponding cross-sections, a combined optimization framework is constructed to search for the finite element simulation data category that best matches the measured deformation. The optimal category label and the actual running data are input into CDDPM for conditional sampling, ultimately obtaining the reconstructed generative deformation field.

[0048] In step S4, conditional sampling of the CDDPM model is performed based on monitoring data and actual operating data to reconstruct the deformation field.

[0049] In this step, when reconstructing the deformation field, monitoring data and operational data at the corresponding time are collected as inputs. A combined optimization method is used to select the finite metadata category that best matches the monitoring data and use it as the CFG guiding condition. The selected optimal category label and the actual operational data are input into the CDDPM model to realize the reconstruction of the deformation field during the operation of the rockfill dam.

[0050] Step S4 further includes: Step S4.1, Modeling the Combinatorial Optimization Problem. A candidate set is constructed from the multi-category data extracted from the finite element simulation samples. This candidate set is then introduced into the combinatorial optimization model, and several categories that most closely resemble the measured deformation values ​​are selected as guiding conditions for generation. This problem is formalized as a discrete combinatorial optimization problem, with the objective function being:

[0051] In the formula, x i This is a selection scalar for multi-hot encoding, with a value of 0 or 1, indicating whether to select the first hot code. i Class data; δ i,j For the first i Simulation data in the first j Deformation values ​​at each measuring point; The measured deformation value is used. By minimizing this objective function, the data combination that best matches the actual monitoring results is found.

[0052] Step S4.2, Solving with Intelligent Algorithm. The classic Genetic Algorithm is used to solve the above combinatorial optimization model. During the optimization process, the individual encoding adopts a multi-hot encoding format, which is suitable for parallel searching of the global optimal solution for multiple class combinations, and has good search capabilities and convergence characteristics.

[0053] Step S4.3, Condition Input Construction and Sampling Generation. The optimal simulation data category labels obtained through combined optimization are then combined with the actual construction and operation data condition vectors […]. T , H RThe conditional vectors are jointly constructed and input into the trained CDDPM for conditional sampling, ultimately obtaining generative deformation field results that match the measured data.

[0054] As one implementation method, this invention uses the aforementioned generative AI-based method for constructing a digital twin of a rockfill dam deformation to reconstruct the deformation field of the Lianghekou core rockfill dam during its operation.

[0055] Figure 2 (a) and (b) with Figure 3 The working principle of the proposed CDDPM is further clarified. Figure 4 (a) through (c) illustrate the relevant circumstances of the rockfill dams used in the embodiments of the present invention.

[0056] First, three types of data sources were obtained from the rockfill dam project: in-situ deformation monitoring data, construction and operation data, and finite element simulation data. In-situ deformation monitoring data came from actually deployed equipment such as water-tube settlement gauges, flexible inclinometers, and electromagnetic settlement rings. Operational data included water level changes and construction and operation duration. Finite element simulation data was generated from a three-dimensional finite element model established under the actual geometry and material zoning conditions of the rockfill dam. The model used C3D8 elements and set 87 loading steps, covering the filling and impoundment processes. Each set of simulation samples corresponded to a set of constitutive parameters and a two-dimensional deformation field for multiple time steps. The two-dimensional deformation field data of the core wall and the typical section 3-3 during the operation phase (steps 59-87) were extracted for subsequent model training.

[0057] The extracted two-dimensional deformation field data were uniformly interpolated into fixed-size images (128×128) and converted into grayscale images. Depth features of the images were extracted using a pre-trained ResNet-18 residual neural network, and then the K-means algorithm was used to perform cluster analysis on the simulation samples. Each simulation sample was assigned a category label based on its deformation morphology, resulting in 50 categories, providing structural information guidance for subsequent conditional generative models.

[0058] The model takes three types of conditional vectors as input: a two-dimensional deformation field image, data clustering category labels, and information such as water level during operation and total construction and operation time. The CDDPM architecture is a deep U-Net network that combines temporal embedding, conditional embedding, and spatial coding to support flexible control and fusion of conditions. Noise is added through forward diffusion, and the model is then trained to learn the reverse denoising process, thereby achieving condition generation.

[0059] Figure 5 The process of generating a deformable field from pure Gaussian noise using the CDDPM proposed in this invention is demonstrated. Figure 6(a) shows the convergence of the objective function during the combinatorial optimization process. The function converges stably to below 0.01 after 20 iterations, indicating that the optimization method proposed in this invention can effectively identify representative simulation data categories based on monitoring data for CFG condition guidance. Figure 6 (b) shows the deformation field results of a typical cross section of the estuary obtained using this reconstruction framework. Since the conditional DDPM model is trained based on finite element data, its generated results are highly consistent with the spatial distribution of the finite element deformation field.

[0060] Table 1 shows the relative error between the reconstruction results and the monitoring data. The reconstruction error of all measuring points on the two cross sections is basically controlled within 10%, and the overall reconstruction accuracy reaches 92.8%, which verifies the effectiveness and accuracy of this method in the high-precision deformation sensing task of rockfill dams.

[0061] Table 1. Relative error between reconstructed deformation field and monitoring data. .

[0062] Figure 7 The computation time consumption of each stage of the method of this invention under the aforementioned equipment conditions is demonstrated. Similar to traditional parameter inversion methods, the computation time is mainly concentrated in the finite element sample preparation stage. After the CDDPM model is trained, real-time reconstruction of the deformation field can be achieved, generating a single deformation field image in only about 3 seconds, which can meet the needs of real-time monitoring. In contrast, the average time for a single finite element analysis based on inversion parameters is about 6 hours, and the structural health monitoring model based on time-series monitoring data and deep learning also needs to be trained before it can make predictions. Both of these methods are difficult to achieve real-time deformation field generation. Therefore, the deformation field reconstruction framework proposed in this invention combines high accuracy and real-time performance, providing effective support for engineering applications.

[0063] To further verify the effectiveness of the present invention, in the generation process without the CFG guidance mechanism, only the input construction and operation data were adjusted for condition generation. Figure 8 The results of 10 reconstructions under different construction and operation data were statistically analyzed. Overall, as dam construction and operation progressed, the reconstructed deformation fields at each section gradually increased, consistent with the physical law of dam deformation increasing over time. However, the generated deformation values ​​under these conditions were high, with most maximum settlement values ​​exceeding 5 meters, more than 1 meter higher than the finite element analysis results and the deformation field reconstruction results under complete guided conditions. This makes it difficult to directly meet the accuracy requirements of rockfill dam deformation analysis. Therefore, the DDPM model in this invention mainly relies on deformation monitoring data for guided sampling, with operational data introduced as an auxiliary condition to enhance the physical rationality of the generated deformation field, making the magnitude and evolution trend of the reconstructed deformation closer to actual working conditions.

[0064] As the dam's service life extends, monitoring instruments gradually age and even fail, further reducing the amount of usable monitoring data. Therefore, to evaluate the robustness of this framework under conditions of gradually decreasing monitoring data, this invention conducted deformation field reconstruction experiments on the 3-3 section and the core wall section, involving a step-by-step reduction in the number of measuring points. Specifically, in the initial stage, 10 and 5 monitoring points were selected on the 3-3 section and the core wall section, respectively. In each round of the experiment, 2 and 1 measuring points were successively removed from the two sections for reconstruction. To evaluate the consistency of structural information between the generated deformation field and the original deformation field, a structural similarity index (SSIM) was introduced. The SSIM value ranges from [0, 1], with higher values ​​indicating higher similarity between the two images. Figure 9 The SSIM index of the reconstructed results for two cross-sections is shown under different numbers of measuring points. Overall, the SSIM of all reconstructed results is greater than 0.75, indicating that they maintain good consistency with the reference deformation field in terms of structural morphology. This result verifies that the conditional DDPM model proposed in this invention has good robustness under different numbers of measuring points. This is mainly due to the model effectively learning the spatial distribution pattern of the finite element deformation field during the training phase, enabling it to generate physically consistent and high-quality deformation field images during sampling. In summary, the reconstruction framework proposed in this invention demonstrates good adaptability in practical engineering and can provide reliable support for the full-process deformation perception and safety monitoring of rockfill dams during operation.

[0065] The implementation of the various embodiments of the present invention is based on programmed processing through a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a system for constructing a deformable digital twin of a rockfill dam based on generative AI. This system is used to execute the method for constructing a deformable digital twin of a rockfill dam based on generative AI in the above method embodiments.

[0066] The system comprises: a first main module for constructing a finite element numerical model based on deformation monitoring data and operational data of the rockfill dam, and obtaining multiple sets of deformation field simulation samples using a parameter sampling strategy; a second main module for extracting two-dimensional deformation field data of key monitoring sections based on the obtained deformation field simulation samples, extracting depth features of the two-dimensional deformation field images through a deep neural network, and classifying and labeling the simulation samples using an unsupervised clustering method; a third main module for constructing a conditional denoising diffusion probability model, using the classified two-dimensional deformation field simulation samples, corresponding category labels, and operational data used during simulation as conditional inputs for model training, and outputting a controllably generated deformation field reconstruction model; and a fourth main module for constructing a combinatorial optimization framework by combining actual deformation monitoring data and simulation data of corresponding sections, searching for the optimal data category, and inputting the searched optimal data category and operational data into the controllably generated deformation field reconstruction model, outputting the reconstructed deformation field of the rockfill dam.

[0067] The present invention provides a digital twin construction system for rockfill dam deformation based on generative AI. This system addresses the current situation where, under the condition of sparse monitoring data, prior information of finite element physics and engineering operation data are integrated to achieve high-precision and high-efficiency reconstruction of the dam deformation field and improve the ability to thoroughly perceive the deformation of rockfill dams across the entire domain. By employing the aforementioned modules, generative diffusion modeling, category guidance mechanism and continuous condition vector jointly control the generation process, the accuracy, efficiency and physical consistency of deformation reconstruction are significantly improved.

[0068] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The only difference is that corresponding functional modules are set. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art can improve the modules in the above system embodiments by referring to the specific technical solutions in other method embodiments and combining technical features to obtain corresponding technical means and technical solutions composed of these technical means, on the basis of the above system embodiments, and on the premise of ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.

[0069] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a device for constructing a deformable digital twin of a rockfill dam based on generative AI, including a memory and a processor. The memory stores program instructions that are executed by the processor, and the processor calls the program instructions to execute the method for constructing a deformable digital twin of a rockfill dam based on generative AI.

[0070] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the described method for constructing a deformable digital twin of a rockfill dam based on generative AI.

[0071] In summary, the core of this invention lies in proposing a digital twin technology framework for rockfill dam deformation based on generative artificial intelligence and multi-source data fusion. It employs a Conditional Denoising Diffusion Probability Model (CDDPM) and constructs a conditional sampling mechanism through classifier-free guidance (CFG) and continuous conditional vectors to efficiently fuse finite element simulation data, monitoring data, and operational data. Furthermore, it combines ResNet-18 residual neural network and K-means clustering to perform unsupervised classification of finite element data, introducing combinatorial optimization to identify the finite element data category that best matches the monitoring data, thereby achieving conditionally guided deformation field generation. This framework has been applied and validated on the world's highest existing rockfill dam, the Lianghekou Dam (303 meters). The results show that the proposed method for constructing a digital twin of rockfill dam deformation based on generative AI can efficiently reconstruct rockfill dam deformation with high accuracy and real-time performance, significantly improving the comprehensive perception capability of rockfill dam deformation and providing key technical support for its safe operation.

[0072] This invention significantly improves the accuracy, efficiency, and physical consistency of deformation reconstruction by jointly controlling the generation process through generative diffusion modeling, a category-guided mechanism (CFG), and continuous conditional vectors. Compared with traditional finite element inversion methods, a single reconstruction only takes about 3 seconds after training, improving efficiency by about four orders of magnitude and demonstrating real-time response capability. Even when the number of actual measurement points gradually decreases, the structural consistency index (SSIM) remains above 0.75, exhibiting strong robustness and good generalization ability. This method has been validated in the Lianghekou rockfill dam project, achieving a reconstruction accuracy of 92.8%, demonstrating its adaptability and engineering practical value under complex conditions. This invention improves the thorough perception of rockfill dam deformation and can provide important support for the deformation safety assessment of rockfill dams.

[0073] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0074] In this patent, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a digital twin of a rockfill dam deformation based on generative AI, characterized in that, include: A finite element numerical model was constructed based on deformation monitoring data and operational data of the rockfill dam, and multiple sets of deformation field simulation samples were obtained by adopting a parameter sampling strategy. Based on the obtained deformation field simulation samples, two-dimensional deformation field data of key monitoring sections are extracted, and depth features of two-dimensional deformation field images are extracted through deep neural networks. Unsupervised clustering methods are used to classify and label the simulation samples. A conditional denoising diffusion probability model is constructed. The model is trained by using the classified two-dimensional deformation field simulation samples, the corresponding category labels, and the running data used in the simulation as conditional inputs, and outputs a controllable deformation field reconstruction model. By combining actual deformation monitoring data with simulation data of the corresponding cross sections, a combined optimization framework is constructed to search for the optimal data category. The optimal data category and the running data are then input into a controllably generated deformation field reconstruction model, which outputs the reconstructed deformation field of the rockfill dam.

2. The method for constructing a deformable digital twin of a rockfill dam based on generative AI according to claim 1, characterized in that, Based on the obtained deformation field simulation samples, after extracting the two-dimensional deformation field data of the key monitoring sections, the following is also included: The extracted two-dimensional deformation field data is uniformly interpolated and mapped to an image matrix with a fixed resolution. The deformation values ​​are then converted into single-channel grayscale images using standardization processing.

3. The method for constructing a deformable digital twin of a rockfill dam based on generative AI according to claim 2, characterized in that, Depth features of two-dimensional deformable field images are extracted using deep neural networks, including: Using a pre-trained ResNet-18 network as the backbone, the fully connected classification layers at the ends are removed, and only the convolutional layers and global average pooling layers are retained to extract the depth feature vector of each two-dimensional deformation field image.

4. The method for constructing a deformable digital twin of a rockfill dam based on generative AI according to claim 3, characterized in that, The classification and labeling of simulated samples are performed using unsupervised clustering methods, including: The depth feature vectors corresponding to all two-dimensional deformation field images are input into the K-means clustering algorithm to perform unsupervised clustering operations, classify the simulation samples, and assign a unique class label to each simulation sample.

5. The method for constructing a deformable digital twin of a rockfill dam based on generative AI according to claim 1, characterized in that, The training of the conditional denoising diffusion probability model includes: Gaussian noise is gradually added to the simulated deformation field data of each section of the rockfill dam through a forward diffusion process to obtain a noisy image; The noisy image is input into the U-Net network and trained to predict the noise components added to the original two-dimensional deformable field image; The trained U-Net network is used to gradually remove noise during the backdiffusion process, reconstructing the original deformable field image.

6. The method for constructing a deformable digital twin of a rockfill dam based on generative AI according to claim 5, characterized in that, The method further includes: Discrete category labels are encoded into high-dimensional vectors through a fully connected layer, and a category guidance mechanism is used to randomly mask the category labels with a set probability during the training phase. The running data is transformed into feature representations through fully connected layers and fused with the category label embeddings under the category guidance mechanism to jointly guide the decoding path of the U-Net network.

7. The method for constructing a deformable digital twin of a rockfill dam based on generative AI according to claim 1, characterized in that, Combining actual deformation monitoring data with simulation data of the corresponding cross sections, a combined optimization framework is constructed to search for the optimal data category. The searched optimal data category and the running data are then input into a controllably generated deformation field reconstruction model, including: The multi-category data extracted from the finite element simulation samples are used to construct a candidate set. A combined optimization model is introduced, and several categories that are closest to the measured deformation values ​​are selected as conditions for generation guidance. The optimal data category is obtained by using a genetic algorithm to optimize the combined optimization model. The labels of the optimal data categories are combined with the running data to form a conditional vector, which is then input into a trained conditional denoising diffusion probability model for conditional sampling to obtain generative deformation field results that match the measured data.

8. A system for constructing a deformable digital twin of a rockfill dam based on generative AI, used to implement the method described in any one of claims 1 to 7, characterized in that, The system includes: The first main module is used to construct a finite element numerical model based on the deformation monitoring data and operation data of the rockfill dam, and to obtain multiple sets of deformation field simulation samples by adopting a parameter sampling strategy. The second main module is used to extract two-dimensional deformation field data of key monitoring sections based on the obtained deformation field simulation samples, extract the depth features of the two-dimensional deformation field images through deep neural networks, and classify and label the simulation samples using unsupervised clustering methods. The third main module is used to construct a conditional denoising diffusion probability model. It uses the classified two-dimensional deformation field simulation samples, the corresponding category labels, and the running data used in the simulation as conditional inputs to train the model and outputs a controllable deformation field reconstruction model. The fourth main module is used to combine actual deformation monitoring data with simulation data of the corresponding cross sections to construct a combined optimization framework, search for the optimal data category, and input the searched optimal data category and the running data into the controllably generated deformation field reconstruction model to output the reconstructed rockfill dam deformation field.

9. A device for constructing a deformable digital twin of a rockfill dam based on generative AI, characterized in that, The device includes a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to execute the method for constructing a deformable digital twin of a rockfill dam based on generative AI as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the method for constructing a deformable digital twin of a rockfill dam based on generative AI as described in any one of claims 1 to 7.

Images