A soil and rock dam deformation field reconstruction method based on a diffusion model

Abstract

The application discloses a soil and rock dam deformation field reconstruction method based on a diffusion model, belongs to the field of water conservancy engineering safety monitoring and engineering intelligent analysis, and comprises the following steps: constructing a soil and rock dam random finite element deformation field data set, performing grid mapping, geometric mask processing and global normalization on the deformation field sample, and forming a two-dimensional deformation field training sample; training the diffusion model by using the training sample, wherein an active sampling strategy is adopted in the training stage, and high-information samples are preferentially selected according to sample information quantity to participate in training; in the sampling stage, sparse monitoring data is mapped to a section grid coordinate, a monitoring consistency loss function is constructed, and the gradient thereof is embedded into an inverse diffusion sampling process as a guide item to dynamically correct the sampling state, and a reconstructed deformation field is generated. The application can realize high-precision and real-time deformation field reconstruction of a key section of a soil and rock dam under a sparse monitoring condition, and significantly improves the expression ability of the model to a complex deformation mode and the diversity of generated results.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project safety monitoring and intelligent engineering analysis technology, and in particular relates to a method for reconstructing the deformation field of earth-rock dams based on a diffusion model. Background Technology

[0002] Earth-rock dams, due to their strong adaptability to terrain and geological conditions, convenient material availability, and good engineering economy, have become an important dam type widely used in large-scale water conservancy and hydropower projects. With the continuous improvement of the construction level of high earth-rock dams and the continuous increase in dam height, the nonlinear deformation characteristics exhibited during construction, water impoundment, and long-term operation have become more significant. Dam deformation reflects both the structural stress and material response characteristics and is directly related to the safety of the seepage prevention system, dam stability, and service status assessment. Currently, the perception of earth-rock dam deformation mainly relies on two technical means: in-situ monitoring and numerical analysis. In-situ monitoring typically uses monitoring instruments such as settlement meters and displacement gauges to obtain response information of local locations within the dam body, providing direct evidence for judging the operational status. Numerical analysis mainly involves establishing a finite element model of the dam body and combining material parameters, construction processes, and water impoundment history to numerically simulate and analyze dam deformation. These two methods complement each other, forming the main technical foundation for the current deformation analysis of earth-rock dams and playing a crucial role in ensuring the safe operation of dams.

[0003] However, existing technologies still have significant shortcomings in engineering applications. On the one hand, the number of monitoring points inside the dam body is usually limited, and due to factors such as installation conditions, external interference, equipment aging, and local damage, monitoring data often exhibits characteristics such as high dispersion, insufficient spatial coverage, and poor continuity. This makes it difficult to directly reflect the continuous deformation distribution across key sections or even the entire dam area, and existing monitoring methods cannot support a thorough understanding of the deformation state across the entire dam body. On the other hand, while traditional finite element analysis has strong physical interpretation capabilities, its calculation results typically rely on parameter selection and inversion processes. To improve the consistency between calculation results and measured information, repeated parameter inversion and numerical calculations are often required, which is not only time-consuming but also highly sensitive to initial parameters, boundary conditions, and modeling accuracy. This makes it difficult to meet the practical needs of dynamic perception and real-time analysis in scenarios requiring high-frequency updates and rapid responses during operation. In recent years, some studies have attempted to use deep learning methods to predict or complete dam deformation, but most methods still primarily focus on time series analysis of monitoring points or local variable regression. Their output results are usually limited to single points or a small number of measuring points, making it difficult to directly recover a deformation field with spatial continuity. While diffusion models have shown promise in physical field reconstruction tasks, in the scenario of earth-rock dam deformation field reconstruction, the strong similarity between finite element-generated samples can easily lead to overlearning of a few dominant modes, thus weakening the model's ability to express complex deformation features. Furthermore, there is a lack of mature and targeted solutions for effectively integrating sparse monitoring data into the inverse diffusion sampling process to ensure that the generated results both conform to the prior laws of dam deformation and closely approximate measured information. Therefore, how to achieve rapid and accurate reconstruction of the deformation field of key sections of earth-rock dams by fully combining the physical priors formed by finite element simulation with actual monitoring constraints under conditions of sparse or even locally missing monitoring information has become an urgent technical problem to be solved in the field of earth-rock dam safety monitoring and intelligent analysis. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for reconstructing the deformation field of earth-rock dams based on a diffusion model, thereby resolving the issues present in the existing technologies.

[0005] To achieve the above objectives, this invention provides a method for reconstructing the deformation field of earth-rock dams based on a diffusion model, comprising: A random finite element deformation field dataset for earth-rock dams is constructed, which contains deformation field samples at multiple time points and multiple key monitoring sections. The deformation field samples are rasterized, geometrically masked, and globally normalized to form two-dimensional deformation field training samples. The diffusion model is trained using the two-dimensional deformation field training samples to learn the prior distribution of the deformation field of the earth-rock dam, thus obtaining a trained diffusion model. During the training phase, an active sampling strategy is adopted, prioritizing high-information samples based on their information content. This active sampling strategy is used to enhance the model's ability to express complex deformation patterns. The trained diffusion model is used for conditional sampling reconstruction to obtain sparse monitoring data at the target time. The sparse monitoring data is mapped to cross-sectional grid coordinates, and a monitoring consistency loss function is constructed. The gradient of the loss function is used as a guiding term to be embedded in the inverse diffusion sampling process. The sampling state at each step is dynamically corrected to generate a reconstructed deformation field.

[0006] Preferably, the process of constructing a random finite element deformation field dataset for an earth-rock dam includes: establishing a three-dimensional finite element model based on the geometric structure, material zoning, foundation conditions, construction and filling process, and water storage operation process of the earth-rock dam; using the three-dimensional finite element model to randomly sample the dam material parameters and perform finite element numerical analysis to extract two-dimensional deformation field data of key monitoring sections.

[0007] Preferably, the rasterization mapping is as follows: the discrete two-dimensional deformation field data output by the finite element method is used to establish a regular grid covering the target cross section and mapped to a two-dimensional matrix with a fixed resolution; the geometric masking processing is as follows: invalidation is applied to the external region of the dam body according to the geometric boundary of the target cross section; the global normalization is as follows: the deformation value is linearly mapped to the interval [-1,1] based on the global minimum and maximum values ​​of all samples.

[0008] Preferably, the process of training the diffusion model using the two-dimensional deformation field training samples includes: establishing a forward diffusion process, gradually adding Gaussian noise to the original deformation field samples; constructing a denoising network to predict the added noise, and optimizing it using the mean square error between the predicted noise and the real noise as the loss function.

[0009] Preferably, the denoising network adopts the U-Net architecture, which includes a multi-scale convolution module, residual connections, and a temporal embedding module, for simultaneously extracting global distribution features and local detail features of the deformation field.

[0010] Preferably, the active sampling strategy includes: during the training process, calculating the expected prediction error of candidate samples as an information score based on the current diffusion model; prioritizing the selection of samples to participate in training according to the information score from high to low, and periodically updating the candidate sample pool.

[0011] Preferably, the formula for the information scoring is: in, For the sample In the diffusion step The result after adding noise To predict noise for the model.

[0012] Preferably, the monitoring consistency loss function is constructed as follows: during the reverse diffusion process, the estimated value of the intermediate clean deformation field is recovered based on the current noisy state, the predicted value of the estimated value at the corresponding grid position of each monitoring point is extracted, and the sum of squares between the predicted value and the normalized monitoring value is calculated.

[0013] Preferably, embedding the gradient of the loss function as a guiding term into the inverse diffusion sampling process involves: calculating the gradient of the monitoring consistency loss function with respect to the current noisy state, performing gradient descent correction on the current noisy state using a preset guiding weight coefficient, and then using the corrected state to perform standard inverse diffusion update.

[0014] Preferably, the reconstructed deformation field is used for full-domain deformation perception, safety assessment, monitoring data completion, anomaly identification, or risk warning during the operation of earth-rock dams.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs a dataset of stochastic finite element deformation fields of earth-rock dams containing multiple time points and multiple key monitoring sections. This enables the prior distribution learned by the diffusion model to cover deformation modes under various material parameter combinations and loading conditions. Combined with an active sampling strategy to focus on learning complex modes, the trained diffusion model can generate a reconstructed deformation field that matches the current working conditions when faced with actual monitoring data during operation, thus exhibiting good generalization ability.

[0016] This invention performs geometric masking on deformation field samples and invalidates the external region of the dam body, so that the training and sampling process of the diffusion model focuses on the effective region of the dam body, avoiding the interference of the background region on the model's learning of the spatial distribution of the deformation field, and improving the model's attention to the deformation characteristics of the dam body and the learning accuracy.

[0017] This invention adopts a global normalization method, which maps deformation values ​​to a unified interval based on the global minimum and maximum values ​​of all samples. This preserves the relative deformation amplitude relationship between samples under different working conditions and at different times, enabling the diffusion model to better learn the global distribution law of the earth-rock dam deformation field and improve the model's ability to perceive deformation amplitude.

[0018] This invention employs an active sampling strategy during the training phase, prioritizing high-information samples for model training based on their information content, while retaining some random samples to maintain diversity. This technical feature enables the diffusion model to focus on samples with high information density and complex deformation patterns during training, avoiding overlearning of simple or repetitive patterns that dominate in finite element simulation samples. This enhances the model's ability to express complex deformation features and improves the diversity of generated deformation field results.

[0019] The active sampling strategy dynamically selects high-information samples through an information scoring function and prioritizes their use for model parameter updates. This reduces the redundant contribution of low-information samples to the training process, resulting in better model performance under the same number of training rounds, thereby improving training efficiency and sample utilization efficiency.

[0020] This invention maps sparse monitoring data to cross-sectional grid coordinates, constructs a monitoring consistency loss function, and embeds its gradient as a guiding term into the inverse diffusion sampling process. This dynamically corrects the sampling state at each step, ensuring that the generated reconstructed deformation field maintains a high degree of consistency with the measured data at the monitoring points, while simultaneously satisfying the prior distribution of earth-rock dam deformation learned by the diffusion model. This technique directly achieves fine-grained online fusion of monitoring constraints and the generation process, significantly improving the reconstruction accuracy of the deformation field under sparse monitoring conditions.

[0021] Because the gradient of the monitoring consistency loss function is directly embedded in the inverse diffusion sampling process, there is no need for additional parameter inversion, combinatorial optimization search, or multiple finite element calculations before sampling. Constraint sampling can be completed simply by performing gradient correction in each step of the inverse diffusion. Compared with traditional methods that rely on repeated finite element calculations or offline search for optimal conditions, this invention significantly reduces the time overhead of deformation field reconstruction during operation, has good real-time response capabilities, and is suitable for the dynamic sensing needs of earth-rock dams during operation. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the overall process of the deformation field reconstruction method for earth-rock dams based on a diffusion model, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the deformation field generation process of an earth-rock dam based on a diffusion model, according to an embodiment of the present invention. Figure 2 (a) is a schematic diagram of the forward diffusion noise addition process. Figure 2 (b) is a schematic diagram of the reverse denoising and reconstruction process; Figure 3 This is a schematic diagram of the finite element model and key monitoring sections of an earth-rock dam according to an embodiment of the present invention. Figure 3 (a) is a schematic diagram of a three-dimensional finite element model of an earth-rock dam. Figure 3 (b) is a schematic diagram showing the location of key monitoring sections; Figure 4 This is a schematic diagram illustrating the construction and preprocessing process of random finite element deformation field samples according to an embodiment of the present invention, wherein, Figure 4 (a) is a schematic diagram of the generation of random finite element deformation field samples. Figure 4 (b) is a schematic diagram of the variance distribution of the deformation field samples. Figure 4 (c) is a schematic diagram of deformation field rasterization and geometric masking processing. Figure 4 (d) is a schematic diagram of the normalized two-dimensional deformation field training samples; Figure 5 This is a schematic diagram of the monitoring point distribution for deformation field reconstruction according to an embodiment of the present invention, wherein, Figure 5 (a) is a schematic diagram of the distribution of monitoring points in section 3-3. Figure 5 (b) is a schematic diagram showing the distribution of monitoring points along the core wall section; Figure 6 This is a schematic diagram of the deformation field reconstruction results of section 3-3 in an embodiment of the present invention, wherein, Figure 6 (a) is the iterative evolution diagram of the guiding gradient field. Figure 6 (b) is a diagram showing the reconstructed deformation field. Figure 6 (c) is a comparison chart of the reconstruction results and the finite element results. Figure 6 (d) is a graph showing the reconstruction accuracy evaluation results; Figure 7 This is a schematic diagram of the wall cross-section deformation field reconstruction results according to an embodiment of the present invention, wherein, Figure 7 (a) is the iterative evolution diagram of the guiding gradient field. Figure 7 (b) is a diagram showing the reconstructed deformation field. Figure 7 (c) is a comparison chart of the reconstruction results and the finite element results. Figure 7 (d) is a graph showing the results of the reconstruction accuracy evaluation. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0025] Example 1 like Figure 1As shown, this embodiment provides a method for reconstructing the deformation field of an earth-rock dam based on a diffusion model, including: S1: Construct a random finite element deformation field dataset for earth-rock dams. The dataset contains deformation field samples from multiple time points and multiple key monitoring sections. Furthermore, the process of constructing a random finite element deformation field dataset for earth-rock dams includes: establishing a three-dimensional finite element model based on the geometric structure, material zoning, foundation conditions, construction and filling process, and water storage operation process of the earth-rock dam; using the three-dimensional finite element model to randomly sample the dam material parameters and perform finite element numerical analysis to extract two-dimensional deformation field data of key monitoring sections.

[0026] Specifically, step S1 includes: S1.1: Establish a three-dimensional finite element model based on the dam geometry, material zoning, foundation conditions, construction and filling process, and water storage operation process of the target earth-rock dam; S1.2: Select a constitutive model to describe the mechanical behavior of earth-rock dam materials. The constitutive model includes, but is not limited to, the Duncan-Chang EB model, the creep model, and the wetting deformation model. S1.3: The constitutive parameters are perturbed using a random field method or numerical sampling method to generate multiple sets of random parameter fields. Each set of random parameter fields is then input into a three-dimensional finite element model for calculation to obtain the corresponding dam deformation response. The random sampling can reflect the uncertainty and spatial discreteness of the earth-rock dam material parameters, making the generated dataset more representative. S1.4: Extract at least one two-dimensional deformation field data of a key monitoring section from the finite element calculation results. The two-dimensional deformation field data consists of spatial coordinates and corresponding deformation values, and is represented as follows: ; S1.5: Collect two-dimensional deformation field data of different parameter samples, different calculation times and different key sections to form a random finite element deformation field dataset of earth-rock dam for diffusion model training. The sample dataset obtained can be used as the training basis for diffusion model to learn the prior distribution of earth-rock dam deformation.

[0027] The project selected in this embodiment is the Lianghekou Core Rockfill Dam. The Lianghekou Hydropower Station is a key control project in the cascade development of the Yalong River basin. The Lianghekou Core Rockfill Dam is 303 m high and has been filled to the design elevation. Various monitoring devices, including water-tube settlement gauges, flexible inclinometers, and electromagnetic settlement rings, are installed inside the dam body to provide in-situ monitoring data support for deformation analysis during operation. This embodiment focuses on selecting section 3-3 of the dam body and the core wall section as key reconstruction objects. Section 3-3 is used to characterize the overall deformation state of typical areas of the dam body, while the core wall section is used to characterize the deformation response of key seepage prevention structures.

[0028] In this embodiment, a three-dimensional finite element model of the Lianghekou core rockfill dam is first established based on its actual geometry, material zoning, construction and filling process, and water storage operation. The finite element model is calculated using ABAQUS software combined with a user-defined material subroutine. The element type is C3D8, containing 146,691 nodes and 150,776 elements. The total number of finite element loading steps is 102, with the first 59 steps primarily simulating the dam filling process, and subsequent loading steps reflecting the water storage operation process. Considering that this invention mainly focuses on the reconstruction of the deformation field during operation, this embodiment extracts the finite element calculation results from the operation phase as the basis for training samples. Figure 3 The finite element model of the Lianghekou rockfill dam and the positional relationships of key monitoring sections are presented. Figure 3 (a) in the figure is a three-dimensional finite element model of an earth-rock dam. Figure 3 (b) in the figure represents the location of the key monitoring section.

[0029] To improve the coverage of the training samples with the complex deformation modes of actual earth-rock dams, this embodiment uses random field analysis for perturbation sampling of material parameters. In the finite element analysis, the Duncan-Chang EB model, the exponential decay empirical creep model, and the three-parameter wetting deformation model are selected to describe the nonlinear instantaneous deformation, long-term aging deformation, and water-wetting deformation behavior of the rockfill. The random parameter values ​​are set to 0.7–1.3 times the experimental parameter range, and 200 sets of random parameter fields are generated. Figure 4 (a) in the figure is a schematic diagram of the generation of random finite element deformation field samples. Based on the above random parameter field and multi-time loading step calculation results, a total of about 8.8×103 sets of finite element two-dimensional deformation field samples are generated for subsequent diffusion model training.

[0030] S2: Perform rasterization mapping, geometric masking, and global normalization on the deformation field samples to form two-dimensional deformation field training samples; Furthermore, the rasterization mapping is as follows: the discrete two-dimensional deformation field data output by the finite element method is used to establish a regular grid covering the target cross section and mapped to a two-dimensional matrix with a fixed resolution; the geometric masking processing is as follows: invalidation is applied to the external region of the dam body according to the geometric boundary of the target cross section; the global normalization is as follows: the deformation value is linearly mapped to the interval [-1,1] based on the global minimum and maximum values ​​of all samples.

[0031] Specifically, step S2 includes: S2.1: Discrete two-dimensional deformation field data output from the finite element method. Establish a regular mesh covering the target cross-section, and map the discrete node deformation values ​​to a resolution of [resolution missing]. The two-dimensional matrix, through regular grid mapping, can transform the irregular discrete node field into an image representation with a unified pixel structure, thereby adapting to the input requirements of U-Net in the diffusion model; S2.2: Construct a geometric mask based on the geometric boundary of the target cross section of the earth-rock dam to mark the invalid areas outside the dam body, so that the model training and sampling process focuses on the effective area of ​​the dam body. The geometric mask is used to remove invalid areas outside the dam body to avoid background areas from interfering with model training. S2.3: Calculate the global minimum value of the deformed values ​​across the entire sample range. and global maximum value And the deformation value at any position is determined as follows: Normalization is performed. Global normalization preserves the relative deformation amplitude relationships between different samples, improving the model's ability to learn global deformation patterns. (1) in, These are the normalized raster values, with a normalization interval of [missing information]. ; S2.4: Represent the normalized two-dimensional deformable field as a single-channel tensor , as input samples for the diffusion model.

[0032] After obtaining the finite element samples, this embodiment further preprocesses the two-dimensional deformation field data of the key sections. Specifically, discrete deformation data of section 3-3 and the core wall section are extracted from each set of finite element results and mapped onto a regular grid with a fixed resolution to form a two-dimensional single-channel deformation field image of uniform size. Subsequently, a geometric mask is applied according to the actual boundary of the section to retain only the effective area of ​​the dam body and remove the invalid background area to reduce the interference of irrelevant information on model training. Afterward, all samples are processed using a uniform global normalization method to ensure that different samples are within a consistent numerical scale range, thereby forming standardized training samples suitable for the input of the diffusion model. Figure 4 (c) in the diagram is a schematic diagram of the deformation field rasterization and geometric masking process. Figure 4 (d) in the figure is a schematic diagram of the normalized two-dimensional deformation field training samples.

[0033] To verify the applicability of the constructed sample set to model training, this embodiment also statistically analyzed the deformation variance distribution of all samples on the target cross-section. Taking cross-section 3-3 as an example, Figure 4 The variance distribution of the deformation field samples shown in (b) indicates that the deformation response at each location has a certain degree of dispersion and difference, with the maximum variance reaching 0.54. This suggests that the random finite element samples can reflect a variety of potential deformation modes and have good training support capabilities.

[0034] S3: The diffusion model is trained using the two-dimensional deformation field training samples to learn the prior distribution of the deformation field of the earth-rock dam and obtain a trained diffusion model; wherein an active sampling strategy is adopted during the training phase, and high-information samples are selected first to participate in the training based on the amount of sample information; the active sampling strategy is used to enhance the model's ability to express complex deformation modes. Furthermore, the process of training the diffusion model using the two-dimensional deformation field training samples includes: establishing a forward diffusion process, gradually adding Gaussian noise to the original deformation field samples; constructing a denoising network to predict the added noise, and optimizing it using the mean square error between the predicted noise and the real noise as the loss function.

[0035] Furthermore, the denoising network adopts the U-Net architecture, which includes a multi-scale convolution module, residual connections, and a temporal embedding module, for simultaneously extracting the global distribution features and local detail features of the deformation field.

[0036] Furthermore, the active sampling strategy includes: during the training process, calculating the expected prediction error of candidate samples as an information score based on the current diffusion model; prioritizing the selection of samples for training according to the information score from high to low, and periodically updating the candidate sample pool.

[0037] Specifically, step S3 includes: S3.1: Establish the forward diffusion process, in the diffusion step Downward original two-dimensional deformation field sample Gaussian noise is gradually added to obtain noisy samples. The stepwise noise-adding process is used to perturb the original deformed field sample to an approximate Gaussian distribution in order to construct a reversible diffusion generation process; S3.2: The forward diffusion process satisfies: (2) in, For the first The noise scheduling coefficient for each diffusion step. It is the identity matrix; S3.3: The forward diffusion process can be equivalently represented as: (3) in, , , ; S3.4: Construct a denoising network to predict noise terms The model is trained using the following objective function: by minimizing the difference between the actual noise and the predicted noise, the model can learn the denoising rules under different diffusion steps. (4) S3.5: The denoising network adopts a U-Net structure or other convolutional neural network structure with multi-scale feature extraction and reconstruction capabilities to gradually restore the original deformation field distribution during the reverse diffusion process. The multi-scale network structure is beneficial for simultaneously extracting global distribution features and local detail features in the deformation field of earth-rock dam.

[0038] S3.6: During model training, candidate samples are drawn from the candidate sample pool. And the information content of the candidate samples is evaluated based on the current diffusion model; S3.7: The information scoring function of the sample is defined as follows: (5) in, For the sample In the diffusion step The result after adding noise Predict noise for the model; S3.8: According to the information scoring function The model prioritizes high-information samples for training and randomly selects some samples from the remaining samples to form a training subset. This balances sample difficulty and diversity, making the model training more focused on samples with high information content, large error, or complex features, while avoiding overly homogeneous training samples. S3.9: Periodically update the candidate sample pool and training subset so that the model focuses on more representative deformation field samples during training, thereby improving the model's ability to express complex deformation patterns and the diversity of generated results.

[0039] In this embodiment, the preprocessed two-dimensional deformation field samples are input into the denoising diffusion probability model. Gaussian noise is gradually added during the forward diffusion process, and the denoising network is used to learn the noise recovery rules under different diffusion steps during the reverse training process, thereby establishing a mapping relationship from the random noise distribution to the actual deformation field distribution of the earth-rock dam. Figure 2 As shown, Figure 2 (a) in the diagram is a schematic diagram of the forward diffusion noise addition process. Figure 2 (b) in the diagram illustrates the reverse denoising and reconstruction process. Simultaneously, an active sampling strategy is introduced during the training phase, prioritizing samples with higher information content and more complex features for training. This improves sample utilization efficiency, enhances the model's ability to express complex deformation patterns, and increases the diversity of generated results.

[0040] A denoising diffusion probability model is adopted as the core generative model for reconstructing the deformation field of an earth-rock dam. The forward diffusion process progressively adds Gaussian noise to the real deformation field samples, gradually degenerating them into an approximately pure noise distribution. The reverse denoising process learns the denoising map progressively through a neural network, recovering the deformation field conforming to the prior distribution from the random noise. The model training framework is implemented based on Python 3.9.21 and Torch 2.6.0 + cu128. The hardware used for training and sampling is equipped with an Intel Core i7-12700 processor and an NVIDIA GeForce RTX 4080 SUPER graphics card.

[0041] To enhance the model's ability to represent complex deformation patterns, this embodiment introduces an active sampling strategy during the training phase. This involves prioritizing samples with higher information content and more complex deformation patterns from the candidate sample pool for training, while retaining some random samples to maintain overall sample diversity. This approach mitigates the adverse effects of excessive local morphological similarity among finite element samples on model training, allowing the model to focus more on high-information-density samples, thereby improving the richness and stability of the final deformation field generation results. Figure 1 In the overall process shown, the active sampling strategy is embedded as an important part of the model training phase.

[0042] S4: Use the trained diffusion model to perform conditional sampling reconstruction, obtain sparse monitoring data at the target time, map the sparse monitoring data to cross-sectional grid coordinates, construct a monitoring consistency loss function, embed the gradient of the loss function as a guiding term into the inverse diffusion sampling process, dynamically correct the sampling state at each step, and generate a reconstructed deformation field.

[0043] Furthermore, the monitoring consistency loss function is constructed as follows: during the reverse diffusion process, the estimated value of the intermediate clean deformation field is recovered based on the current noisy state, the predicted value of the estimated value at the corresponding grid position of each monitoring point is extracted, and the sum of squares between the predicted value and the normalized monitoring value is calculated.

[0044] Furthermore, embedding the gradient of the loss function as a guiding term into the inverse diffusion sampling process is as follows: calculate the gradient of the monitoring consistency loss function with respect to the current noisy state, perform gradient descent correction on the current noisy state with a preset guiding weight coefficient, and then perform standard inverse diffusion update using the corrected state.

[0045] Furthermore, the reconstructed deformation field is used for full-domain deformation perception, safety assessment, monitoring data completion, anomaly identification, or risk warning during the operation of earth-rock dams.

[0046] Specifically, step S4 includes: S4.1: Obtain sparse monitoring data at the target time, including the physical coordinates of the monitoring points and the corresponding monitoring deformation values, and map the physical coordinates of the monitoring points to the corresponding positions in the two-dimensional grid. The mapping relationship is used to establish a one-to-one correspondence between the physical space of the monitoring points and the two-dimensional grid space. S4.2: In the reverse diffusion... Step, based on the current noisy state and the noise term predicted by the denoising network Restore the corresponding intermediate clean deformation field estimate. Its expression is: (6) S4.3: Constructing the observation operator Used to extract the estimated value of the intermediate cleaning deformation field. The predicted values ​​at the corresponding grid positions of each monitoring point are defined, and the monitoring consistency loss function is defined as follows: The monitoring consistency loss function is used to quantify the degree of deviation between the current generated result and the measured monitoring data. (7) in, For the number of monitoring points, For the first Predicted deformation values ​​at each monitoring point location. This corresponds to the normalized monitoring value; S4.4: Apply the monitoring consistency loss function to the current noisy state. Gradient correction is performed, which dynamically compresses the state after error correction between the generated result and the monitored value in each inverse diffusion step, as shown below: (8) in, To monitor and guide the weighting coefficient; S4.5: Based on the corrected state Perform standard inverse diffusion update to obtain the state for the next diffusion step. The process is repeated until all diffusion steps are completed, and finally the reconstructed deformation field of the target section is output, thereby realizing the constrained reconstruction of the global deformation field of the target section.

[0047] Standard back diffusion update satisfies: (9) The random perturbation term is used to maintain the probabilistic generation characteristics of the sampling process. , , , This is the noise scale parameter associated with the diffusion step.

[0048] This embodiment introduces a monitoring consistency gradient guide during the sampling phase to reconstruct the deformation field under sparse monitoring constraints. Sparse monitoring data at the target time is acquired, and the physical coordinates of the monitoring points are mapped to the two-dimensional grid coordinates of the target section. During the backdiffusion sampling process, the predicted values ​​of the corresponding positions of the monitoring points are extracted based on the current intermediate generation results. A monitoring consistency loss function is constructed, and its gradient is embedded as a guiding term in each step of the backdiffusion update process to dynamically correct the sampling state. This generates a target section reconstructed deformation field that conforms to both the prior distribution of earth-rock dam deformation and the measured monitoring values.

[0049] This embodiment further utilizes actual monitoring data during operation to reconstruct the deformation field at the target time. For the target time, the measured deformation values ​​of the monitoring points on section 3-3 and the core wall section are first extracted, and the physical coordinates of the monitoring points are mapped to the corresponding two-dimensional grid coordinates. Figure 5 The distribution of monitoring points at two key cross-sections is presented, among which... Figure 5 (a) shows the distribution of monitoring points at section 3-3. Figure 5 (b) shows the distribution of monitoring points along the core wall cross-section. By establishing the correspondence between monitoring points and the locations of two-dimensional grids, constraints can be applied to the generated results directly using sparse monitoring values ​​during the sampling stage.

[0050] Subsequently, during the backdiffusion sampling process, a monitoring consistency loss is constructed based on the difference between the predicted and measured values ​​at each monitoring point location in the current generated state. Gradient correction is applied to the current sampling state before each backdiffusion update. In this way, the generated trajectory no longer relies solely on the deformation prior distribution learned during the training phase, but is also dynamically guided by the measured monitoring values. After multiple iterations, the final output is a cross-sectional reconstruction result that satisfies both the statistical prior of the earth-rock dam deformation field and closely matches the monitoring data at the target time.

[0051] Figure 6 The deformation field reconstruction results of section 3-3 at typical times are shown. Figure 6 Figure (a) shows the evolution of the guiding gradient field during the sampling iteration process. It can be seen that as the reverse diffusion process progresses, the monitoring constraint continuously accumulates the corrective effect on the local region, and the gradient field gradually converges from the initial perturbation state to the stable state. Figure 6 (b) in the figure represents the final reconstructed deformation field of section 3-3. Figure 6 (c) shows the corresponding finite element analysis results. As can be seen from the comparison, the deformation field generated by the method in this embodiment maintains a high degree of consistency with the finite element analysis results in terms of overall spatial distribution trend, local settlement characteristics, and regional deformation continuity. Figure 6(d) The reconstruction accuracy evaluation results are given. The error between the reconstructed value and the measured value at the monitoring point is statistically analyzed. The results show that this embodiment can achieve a low reconstruction error and maintain good stability in section 3-3.

[0052] Figure 7 The results of deformation field reconstruction of the core wall section are presented. (Compared to...) Figure 6 similar, Figure 7 (a) reflects the gradual evolution of the monitoring-guided gradient field during the reverse diffusion process, indicating that the monitoring data can continuously participate in and correct the deformation field generation trajectory. Figure 7 (b) in the figure represents the final core wall section reconstruction deformation field result. Figure 7 (c) in the diagram represents the corresponding finite element analysis result. Figure 7 As can be seen from (b) and (c) in this embodiment, the method can not only maintain the continuity of the overall deformation field of the core wall section, but also recover the deformation amplitude and spatial distribution characteristics of the key area well. Figure 7 (d) in the figure gives the reconstruction accuracy evaluation results of the core wall section, indicating that the method of the present invention also has high reconstruction accuracy and stability in the core wall section.

[0053] As can be seen from the above implementation process, the diffusion model-based deformation field reconstruction method for earth-rock dams provided in this embodiment can achieve rapid and high-precision reconstruction of the 3-3 section and the core wall section under sparse monitoring conditions during the operation of the Lianghekou core-wall rockfill dam. Application results show that the method takes approximately 30 seconds for a single section reconstruction and achieves an overall reconstruction accuracy of approximately 90.7%, which can well meet the engineering requirements for rapid perception, continuous section-level reconstruction, and full-domain dynamic analysis of the dam's deformation field during operation. Compared with methods that can only predict single points, this invention can recover a continuous two-dimensional deformation field from discrete monitoring information; compared with traditional finite element analysis that relies on repeated parameter updates, this invention significantly improves the reconstruction efficiency during operation; and because it integrates the prior knowledge of random finite element samples with a monitoring consistency guidance mechanism, its results have both good physical rationality and data consistency.

[0054] This embodiment verifies the feasibility and effectiveness of the method of the present invention in the task of reconstructing the deformation field during the operation of high earth-rock dams, demonstrating that it can provide reliable technical support for the whole-domain deformation perception, safety assessment and risk warning during the operation of earth-rock dams.

[0055] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for reconstructing the deformation field of an earth-rock dam based on a diffusion model, characterized in that, Includes the following steps: A random finite element deformation field dataset for earth-rock dams is constructed, which contains deformation field samples at multiple time points and multiple key monitoring sections; The deformation field samples are rasterized, geometrically masked, and globally normalized to form two-dimensional deformation field training samples. The diffusion model is trained using the two-dimensional deformation field training samples to learn the prior distribution of the deformation field of the earth-rock dam, and a trained diffusion model is obtained. During the training phase, an active sampling strategy is employed, prioritizing high-information samples for training based on their information content; this active sampling strategy is used to enhance the model's ability to express complex deformation patterns. The trained diffusion model is used for conditional sampling reconstruction to obtain sparse monitoring data at the target time. The sparse monitoring data is mapped to cross-sectional grid coordinates, and a monitoring consistency loss function is constructed. The gradient of the loss function is used as a guiding term to be embedded in the inverse diffusion sampling process. The sampling state at each step is dynamically corrected to generate a reconstructed deformation field.

2. The method according to claim 1, characterized in that, The process of constructing a random finite element deformation field dataset for earth-rock dams includes: establishing a three-dimensional finite element model based on the geometric structure, material zoning, foundation conditions, construction and filling process, and water storage operation process of the earth-rock dam; using the three-dimensional finite element model to randomly sample the dam material parameters and perform finite element numerical analysis to extract two-dimensional deformation field data of key monitoring sections.

3. The method according to claim 1, characterized in that, The rasterization mapping is as follows: the discrete two-dimensional deformation field data output by the finite element method is used to establish a regular grid covering the target cross section and mapped to a two-dimensional matrix with a fixed resolution; the geometric masking process is as follows: invalidation is applied to the external region of the dam body according to the geometric boundary of the target cross section; the global normalization is as follows: the deformation value is linearly mapped to the interval [-1,1] based on the global minimum and maximum values ​​of all samples.

4. The method according to claim 1, characterized in that, The process of training the diffusion model using the two-dimensional deformable field training samples includes: establishing a forward diffusion process, gradually adding Gaussian noise to the original deformable field samples; constructing a denoising network to predict the added noise, and optimizing the model using the mean square error between the predicted noise and the actual noise as the loss function.

5. The method according to claim 4, characterized in that, The denoising network adopts the U-Net architecture, which includes a multi-scale convolution module, residual connections, and a temporal embedding module, used to simultaneously extract the global distribution features and local detail features of the deformation field.

6. The method according to claim 4, characterized in that, The active sampling strategy includes: during training, calculating the expected prediction error of candidate samples as an information score based on the current diffusion model; selecting samples to participate in training in order of information score from high to low; and periodically updating the candidate sample pool.

7. The method according to claim 6, characterized in that, The formula for the information scoring is: in, For the sample In the diffusion step The result after adding noise To predict noise for the model.

8. The method according to claim 1, characterized in that, The monitoring consistency loss function is constructed as follows: during the reverse diffusion process, the estimated value of the intermediate clean deformation field is recovered based on the current noisy state, the predicted value of the estimated value at the corresponding grid position of each monitoring point is extracted, and the sum of squares between the predicted value and the normalized monitoring value is calculated.

9. The method according to claim 8, characterized in that, The gradient of the loss function is used as a guiding term to embed the inverse diffusion sampling process as follows: calculate the gradient of the monitoring consistency loss function with respect to the current noisy state, perform gradient descent correction on the current noisy state with a preset guiding weight coefficient, and then perform standard inverse diffusion update using the corrected state.

10. The method according to claim 1, characterized in that, The reconstructed deformation field is used for full-domain deformation perception, safety assessment, monitoring data completion, anomaly identification, or risk warning during the operation of earth-rock dams.

Images