Rockfill dam deformation analysis model updating method considering damming material parameter correlation

Abstract

The invention discloses a rock-fill dam deformation analysis model updating method considering damming material parameter correlation, which comprises the following steps: firstly, collecting triaxial test parameters of a plurality of rock-fill dams with different dam types and lithology in the world, quantifying parameter correlation strength, constructing a priori knowledge model in combination with a variational auto-encoder, and mining a parameter internal correlation mode to provide physical constraints; secondly, generating a parameter sample by adopting a Latin hypercube sampling mode, constructing a data set in combination with finite element simulation, training a deep neural network agent model fused with spatial-temporal characteristics, and realizing high-precision mapping of damming material parameters and global deformation response; then introducing an elite archiving strategy to improve a multi-objective optimization algorithm NSGA-III; then DPK-VAE, a DNN proxy model and an improved NSGA-III are integrated, a multi-objective optimization framework is formed, and parameter optimization of the constitutive model of the damming material is completed so as to achieve updating of a rock-fill dam deformation analysis model; and finally, calculating a global deformation field and key safety indexes based on the updating model, and realizing thorough perception of the deformation character and the operation state of the dam body.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy engineering and geotechnical engineering technology, and in particular relates to deformation analysis and safety assessment technology for high rockfill dam projects. Background Technology

[0002] Earth-rock dams, as core infrastructure for flood control, power generation, and water resource allocation, require life-cycle deformation safety monitoring and analysis for project stability. Due to their economic advantages, terrain adaptability, and seismic performance, earth-rock dams have become a widely used dam type globally. However, the complexity of the mechanical properties of dam materials and the dynamic changes in construction and operation conditions make accurate characterization of dam deformation behavior a significant challenge. Traditional deformation analysis of rockfill dams mainly relies on numerical inversion methods, using finite element simulation combined with monitoring data to correct constitutive parameters and improve simulation accuracy. However, traditional finite element methods have low computational efficiency during parameter updates, making it difficult to meet the needs of large-scale iterative calculations. Furthermore, parameter inversion methods often simplify the problem to a purely numerical optimization problem, ignoring the physical constraints between parameters. This results in mathematically feasible optimization results but lacking physical rationality, affecting the engineering applicability and predictive reliability of the model.

[0003] In recent years, the rapid development of deep learning technology has driven research on parameter update methods based on surrogate models, significantly improving the efficiency of safety analysis for earth-rock dams. However, existing research has largely focused on improving computational accuracy and efficiency, with insufficient attention paid to the constraint mechanisms of the physical meaning of parameters. Purely data-driven neural networks and black-box optimization strategies may lead to optimization paths deviating from true physical laws during iteration, resulting in solution spaces that are "mathematically feasible but physically distorted." Furthermore, while probabilistic inversion methods based on Bayesian theory can assess parameter uncertainty, their reliance on prior parameter distribution assumptions based on engineering experience fails to fully exploit the potential dependencies between parameters, leading to a lack of uniqueness and engineering interpretability in the optimization results. Therefore, how to effectively integrate physical mechanisms with data-driven methods to construct deformation analysis models that combine predictive accuracy and physical interpretability has become a key scientific problem urgently needing to be solved in the field of safety assessment of high earth-rock dams. Summary of the Invention

[0004] To address key issues in parameter updating during deformation analysis of rockfill dams, including the low computational efficiency of traditional finite element methods (FEM) which fails to meet iterative requirements, and the physical distortion of optimization results caused by existing parameter inversion methods neglecting the physical correlations between parameters, this invention proposes a model updating method for rockfill dam deformation analysis that considers the correlation of dam material parameters. This method combines deep learning technology with the physical mechanisms of geotechnical materials to construct a physically constrained multi-objective optimization algorithm framework for parameter updating and prediction in deformation analysis and safety monitoring of high earth-rock dams. By exploring the intrinsic correlations between dam material parameters, this method achieves a deep integration of engineering physics laws and data-driven methods.

[0005] According to one aspect of the present invention, a method for updating a deformation analysis model of a rockfill dam that considers the correlation of dam material parameters is provided, comprising:

[0006] S1: By collecting triaxial test parameters from multiple rockfill dam projects around the world, the correlation between parameters is quantified, and a variational autoencoder (VAE) is used to learn the parameter association patterns to construct a prior knowledge model. This model provides physical supervision for subsequent optimization and ensures the rationality of parameter updates.

[0007] S2: Samples are generated in the parameter space using Latin hypercube sampling and random sampling methods, and a dataset is constructed by combining finite element simulation. Based on this dataset, a deep neural network (DNN) surrogate model that integrates multiple physical features is trained to achieve a high-precision mapping from dam material parameters to the spatiotemporal deformation response of the structure across the entire domain.

[0008] S3: An elite archiving strategy is introduced to improve the multi-objective optimization algorithm NSGA-III, enhancing the robustness of the method in high-dimensional objective optimization problems, while ensuring the continued stability of physical constraints.

[0009] S4: Integrate the prior knowledge model (DPK-VAE), the DNN surrogate model, and the improved NSGA-III algorithm to form a multi-objective optimization framework for the physical constraints of dam materials. Through this framework, optimize the constitutive model parameters of the dam materials and complete the update of the deformation analysis model of the rockfill dam.

[0010] As a further technical solution, the method also includes:

[0011] S5: Based on the updated deformation analysis model, calculate the global deformation field and key safety indicators to achieve a thorough understanding of the dam's deformation characteristics and operational status, providing a scientific basis for the engineering's deformation safety assessment and safety decision-making.

[0012] Furthermore, step S1 introduces the use of VAE to learn the correlation structure of dam construction material parameters, the main steps of which include:

[0013] S1.1: Collect triaxial test parameters of rockfill dams of different types and scales from both domestic and international sources, covering the two main dam types: concrete-faced rockfill dams and core-walled rockfill dams. From a geological perspective, the dataset should include different lithological types to reflect the typicality and diversity of the test parameters under geological conditions. Data sources include not only engineering technical reports and relevant literature, but also other geotechnical engineering data accumulated by the experimental unit.

[0014] S1.2: First, construct the correlation coefficient matrix and use the Pearson correlation coefficient to evaluate the correlation strength of each parameter pair. The calculation formula is as follows: , In the formula: p is the sample mean of parameter i. i (k) Let i be the value of parameter i on the k-th sample.

[0015] S1.3: Based on the combination of highly correlated parameters, a proportional feature is further constructed to characterize the correlation structure between parameters. Specifically, for a set of parameter pairs with significant correlation (p... i , p j The proportional characteristic function is defined as follows: , In the formula: f ij (P), g n For parameter scaling characteristics, p i and p j For significantly correlated parameter components, n is the index of strongly correlated parameter combinations. This feature form possesses good mathematical properties and practical interpretability. On the one hand, the logarithmic ratio eliminates the influence of the original parameter dimensions and scale, improving the feature's generality and robustness; on the other hand, if there is an exponential or power-law relationship between parameter pairs, this form can effectively reveal the underlying structural dependencies.

[0016] S1.4: To effectively utilize parameter correlation features to guide the parameter update process, a variational autoencoder (VAE) is used to construct a prior knowledge model. A VAE is a generative model that integrates deep learning and Bayesian inference, capable of learning the latent representation of input data and generating new samples. Compared to traditional autoencoders (AEs), VAEs introduce variational inference and probability distributions, giving the latent space good continuity and structure.

[0017] In VAE, assuming the original input data is x, its distribution is p(x), and z is the latent variable extracted by the encoder, This is the reconstructed data for the decoder. According to Bayes' theorem, the encoder p(z | x) can be expressed as: , In the formula: p(x | z) is the likelihood function, which represents the probability of reconstructing the original input data x given the latent variable z. p(z) is the prior distribution of the latent variable z, usually set as a standard normal distribution N~(0, 1). Since p(x) and p(x | z) are difficult to determine directly, VAE uses variational inference to make the variational posterior distribution q of the encoder output... ϕ (z |x) approximates the true posterior distribution p(z |x), where A Gaussian distribution is typically used. For encoder parameters, the objective of variational inference is to minimize the KL divergence, which is calculated as follows: .

[0018] By minimizing the KL divergence, the distribution of the encoder output approximates the true posterior distribution. This process is equivalent to maximizing the lower bound of evidence, ELBO, which is calculated as follows: , In the formula: Let θ be the posterior distribution of the latent variable z after decoder parameterization, and θ be the decoder parameters. The first term of ELBO is the reconstruction loss, which measures the difference between the original data x and the reconstructed data. The difference between them is usually calculated using the mean squared error. The second term is the regularization term. This is used to constrain the distribution characteristics of latent variables.

[0019] An improved β-VAE model is adopted, which adjusts the weight of the KL divergence term by introducing the hyperparameter β, thereby achieving a balance between reconstruction accuracy and potential spatial structure. Its objective function is: .

[0020] Based on the above theory, a generative model VAE based on deep learning is constructed as a prior knowledge model (DeepPrior Knowledge-VAE, DPK-VAE). This model adopts a modular fully connected network architecture, consisting of a probabilistic encoder and a generative decoder forming a symmetric deep neural network. The encoder extracts high-order features of the input data step by step through multi-layer data feature extraction modules, using nonlinear transformations, and outputs the mean and log-variance of the latent variables. The decoder adopts a mirror-symmetric structure with the encoder, inversely mapping the latent variables to the data space to achieve accurate reconstruction of the input data.

[0021] Furthermore, in step S2, parameter samples are generated using Latin hypercube sampling and random sampling methods, and a training dataset for the surrogate model is constructed using finite element simulation. The trained DNN surrogate model replaces the finite element forward calculation in subsequent model updates to improve optimization efficiency. The specific steps are as follows:

[0022] S2.1: Random sampling (RS) is a method for generating random numbers based on probability distributions, widely used in statistical computing, machine learning, Bayesian inference, optimization problems, and other fields. The direct sampling method using a uniform distribution has the following mathematical expression: , In the formula: u is a random number between [0, 1] generated by a standard random number generator, x is a sampled value within the target interval [a, b], a is the lower bound of the distribution of variable x, and b is the upper bound of the distribution of variable x. This method has the advantages of simple implementation and high computational efficiency, and is suitable for large-scale datasets. However, its applicability in high-dimensional parameter spaces is limited, and problems such as uneven sample distribution and insufficient sampling in local regions may occur, affecting convergence.

[0023] Latin Hypercube Sampling (LHS) is an advanced multidimensional parameter space sampling method that demonstrates significant advantages in fields such as uncertainty quantification and engineering simulation. Its core idea is to uniformly divide the parameter interval of each dimension into N non-overlapping sub-intervals, randomly select a sample point within each sub-interval, and then form the final sampling result through random permutations and combinations. The mathematical expression is: , In the formula: x ij This represents the i-th sampling point in the j-th dimension. It is the inverse function of the probability distribution function in the j-th dimension, U ij It is a uniform random number within the interval [0, 1].

[0024] The LHS method ensures that each sub-interval of each dimension has exactly one sample point, effectively avoiding sample clustering or omission that may occur with simple random sampling. It can significantly improve the parameter space coverage and enhance estimation accuracy and stability.

[0025] S2.2: The sampled parameter combinations are substituted into the finite element analysis model for calculation. The constitutive models used include an instantaneous deformation constitutive model based on the Duncan-Zhang EB model, a rheological constitutive model based on the Nanshui seven-parameter model, and a humidification constitutive model based on the Shenzhujiang three-parameter humidification model. The calculation tools include commercial software such as ABAQUS and ANSYS, combined with a self-developed finite element constitutive calculation program.

[0026] S2.3: The calculated training samples are proportionally divided into training, test, and validation sets to construct a deep neural network (DNN) surrogate model. This model captures the highly nonlinear relationship between the material parameters and deformation of the rockfill dam through multiple layers of fully connected neurons and updates the weights using the backpropagation algorithm. To further improve model performance, a Bayesian hyperparameter optimization method is adopted to enhance its ability to learn complex nonlinear features.

[0027] The design of the proxy model fully considers the spatiotemporal global characteristics. By replacing the measurement point number with the coordinate information of the measurement point number, spatial information is given to the measurement point; the material parameter partitions are mapped to the measurement point information of the corresponding partitions to reduce input redundancy; at the same time, the filling and water storage information is introduced to give the model temporal characteristics, thereby outputting the deformation response under specific time and parameters, and comprehensively simulating the entire domain and operation process of the rockfill dam.

[0028] The training performance of the surrogate model is evaluated using root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 The evaluation is conducted using the following formulas (10), (11), and (12): , , , In the formula: This represents the i-th prediction value of the surrogate model in the validation set. This represents the test value, i.e., the result of the finite element simulation.

[0029] Furthermore, in step S3, an elite archiving strategy is introduced to improve the NSGA-III multi-objective optimization algorithm, ensuring the continued stability of physical constraints. The specific steps are as follows:

[0030] S3.1: The NSGA-III algorithm is a multi-objective optimization algorithm based on an elitist strategy. This algorithm effectively maintains population diversity through non-dominated sorting and reference point mechanisms, and can efficiently approximate the Pareto front in high-dimensional objective spaces. When applied to the optimization and updating of parameters in a rockfill dam constitutive model, the key steps of this algorithm include:

[0031] (1) Population initialization: A parent population of size N is generated by random generation or Latin hypercube sampling, and the offspring population is generated by simulating binary crossover and polynomial mutation operators.

[0032] (2) Non-dominated sorting: The parent and offspring populations are merged to form a 2N ensemble population. A fast non-dominated sorting strategy is used to divide individuals into multiple non-dominated levels (F1, F2, … , F i ).

[0033] (3) Environmental selection: NSGA-III innovatively introduces a reference point-based selection mechanism. A set of reference points are evenly distributed in the high-dimensional target space. The vertical distance between individuals and reference points is calculated, and individuals that are close to reference points and evenly distributed are preferentially retained. For uncovered reference points, a niche preservation strategy is adopted to ensure population diversity.

[0034] (4) Iterative optimization: Repeat the above process until the termination condition is met, and finally output a set of Pareto optimal solutions that are widely distributed and have good convergence.

[0035] S3.2: In genetic optimization, population diversity and the preservation of high-quality solutions are crucial for improving the algorithm's convergence speed and solution set quality. However, in parameter optimization guided by regularization terms, phenomena such as fitness rebound of penalty terms and loss of high-quality solutions often occur in the mid-to-late stages. This phenomenon is particularly pronounced in complex engineering optimization problems with high-dimensional parameter spaces. To address this, an elite archiving strategy is designed and introduced, using a dynamic elite library mechanism to stably maintain regularization constraints.

[0036] The core idea of ​​the elite archiving strategy is to dynamically retain and utilize individuals that excel in both the objective function and penalty terms throughout the generations, thereby improving the overall quality and physical feasibility of the solution set. Specifically, the mechanism of this strategy includes the following stages:

[0037] (1) Selection of Elite Candidate Solutions: During each generation of evolution, all individuals in the current population are evaluated based on the comprehensive loss function. The comprehensive loss function integrates the sedimentation prediction error term and the physical constraint penalty term to reflect the individual's performance in objective optimization and physical feasibility. Subsequently, individuals are sorted according to their comprehensive fitness values ​​from smallest to largest, and several optimal individuals are selected as elite candidates for this generation. The comprehensive loss function Φ(x) is calculated as follows: , In the formula, w i Let f be the objective function weight for the i-th partition. i (x) represents the settlement prediction error term for the i-th partition, λ j The penalty term weights for the three stone pile partitions, p j(x) represents the partition penalty value. The weight of the penalty term is set slightly higher than the prediction error weight to strengthen the engineering constraints of the solution set on physical feasibility, ensuring that the parameter solution in the optimization process not only performs well in the objective function, but also meets the physical rationality requirements.

[0038] (2) Elite Archive Update: Merge the current generation of elite candidate solutions with historical elite archive individuals to form a larger set of candidate solutions. If the number of elite individuals in the candidate set exceeds the preset elite archive capacity, then filter again based on the comprehensive loss value, retain the best individuals, and finally form a new generation of elite archive.

[0039] (3) Elite perturbation sampling: In each generation of iteration, an elite perturbation sampling mechanism is adopted. Taking the individuals in the elite archive as the center, a new solution is generated by Gaussian perturbation to replace the bottom 20% of individuals in the current population that are sorted by non-dominance. This effectively prevents excellent individuals from being eliminated due to genetic operations or fluctuations in population diversity.

[0040] Furthermore, in step S4, considering the multi-material zoning characteristics of rockfill dams, a multi-objective optimization framework is constructed for each zoning. The prior knowledge model of dam construction materials (DPK-VAE) and the DNN proxy model are integrated into this framework to achieve the physical rationality update of the model parameters, and finally complete the optimization and update of the deformation analysis model of the rockfill dam.

[0041] S4.1: The objective function of this multi-objective optimization framework consists of two parts: a basic fitness term and a regularization penalty function term. The basic fitness term calculates the predicted values ​​of the model parameters using a surrogate model, and further uses the root mean square relative error (RMSRE) to evaluate the goodness of fit between the predicted values ​​and the measured data. The regularization term uses a penalty function mechanism to quantify the degree of deviation between the model parameters and prior knowledge, thus playing a guiding and constraining role in the optimization process. The final combined objective function is as follows: , In the formula, F(X) is a multi-objective function; F i (X) is the comprehensive objective function corresponding to the i-th material partition; f i (x) represents the settlement prediction error term for the i-th partition, calculated using RMSRE; f θi (x (j) y is the predicted settlement value of the surrogate model at the j-th measuring point in the i-th partition; i (j) N represents the measured settlement value at the j-th measuring point in the i-th partition; i P represents the total number of measurement points in the i-th partition; i (X) is a regularization term that measures the degree of deviation between the parameters and prior knowledge. The loss is calculated using the prior knowledge model DPK-VAE. priorWe obtain that this term consists of the reconstruction loss and the KL divergence; λ j x is the weight of the penalty term for the i-th partition, used to adjust the strength of the penalty term. L With x U The lower and upper bounds of the parameter variables limit the search range for parameter optimization.

[0042] S4.2: When performing iterative optimization using a DNN surrogate model and an improved multi-objective NSGA-III algorithm based on the correlation regularization of dam material parameters, the calculation method can be seen in the pseudocode in Table 1 below:

[0043] Table 1 Improved VAE-DNN-NSGA-III Algorithm

[0044] .

[0045] S4.3: Unlike the conventional NSGA-III algorithm, the improved NSGA-III algorithm introduces an elite archiving mechanism to continuously pass on and spread high-quality individual genes. This enhances the algorithm's local search capability in the neighborhood of high-quality solutions and alleviates problems such as fitness rebound and solution set quality degradation in the later stages of optimization. This strategy provides a robust and efficient intelligent means for optimizing complex engineering parameters.

[0046] Furthermore, step S5, based on the updated and optimized dam material parameters, combined with a high-precision finite element simulation model, conducts deformation simulation calculations throughout the dam's entire life cycle, forming a closed-loop system from parameter updates to safety assessment, thereby achieving a thorough understanding of the dam's deformation state and risk warning.

[0047] S5.1: Through a high-precision finite element simulation model, accurate calculation and dynamic updating of the deformation field across the entire dam body were achieved. Based on the analysis of the calculated deformation field cloud map, the spatial distribution characteristics and temporal evolution of dam deformation can be systematically revealed, providing engineering managers with an intuitive means of monitoring the dam's condition.

[0048] Furthermore, dam crest cracking is closely related to uneven deformation of the dam body. Controlling uneven settlement is crucial for ensuring the structural integrity of high earth-rock dams and reducing the risk of dam failure. Deformation inclination has been widely used to assess the risk of uneven settlement and dam crest cracking; its calculation formula is as follows: , In the formula: γ represents the deformation inclination at points a and b on the dam crest, δ a and δ b The values ​​(mm) represent the settlement deformation at two points at specific times. ab θ is the horizontal distance (mm) between the two points, and θ is the angle between the line connecting the two points and the horizontal plane.

[0049] In engineering practice, γ is usually used c = 1.0% is used as the critical threshold for the risk of dam crest cracking. By calculating the deformation inclination of key parts of the dam and comparing it with the critical threshold, the risk of dam crest cracking can be accurately assessed, with particular attention paid to areas of high deformation inclination under critical conditions such as full storage period, providing a theoretical basis for early warning of dam crest cracking risk. By integrating the results of deformation analysis and risk assessment, a safety assessment conclusion for the entire life cycle of high rockfill dams is formed, clarifying the safety status and potential risk points at different stages, and outputting quantitative deformation parameters and risk indicators. These results provide a scientific basis for dam operation and maintenance decisions and support the optimization of safety management and maintenance schemes throughout the entire life cycle of the project.

[0050] According to one aspect of the present invention, a device for updating a deformation analysis model of a rockfill dam that considers the correlation of dam material parameters is provided, comprising:

[0051] The first main module is used to collect triaxial test parameters of multiple rockfill dam projects around the world, quantify the correlation between parameters, and use a variational autoencoder to learn the parameter association pattern and build a prior knowledge model.

[0052] The second main module is used to generate samples in the parameter space using Latin hypercube sampling and random sampling methods, construct a dataset by combining finite element simulation, and train a deep neural network surrogate model that integrates multiple physical features based on this dataset. The deep neural network surrogate model is used to map from dam material parameters to the spatiotemporal deformation response of the structure in the whole domain.

[0053] The third main module is used to introduce an elite archiving strategy and improve the multi-objective optimization algorithm NSGA-III;

[0054] The fourth main module is used to integrate the prior knowledge model, the deep neural network surrogate model, and the improved multi-objective optimization algorithm NSGA-III to form a multi-objective optimization framework for the physical constraints of the dam material. This framework is used to optimize the constitutive model parameters of the dam material and complete the update of the deformation analysis model of the rockfill dam.

[0055] As a further technical solution, the device also includes:

[0056] The fifth main module is used to calculate the global deformation field and key safety indicators based on the updated deformation analysis model.

[0057] 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 perform the steps of the method for updating the deformation analysis model of a rockfill dam that considers the correlation of dam material parameters.

[0058] Compared with existing technologies, the method for updating the deformation analysis model of rockfill dams that considers the correlation of dam material parameters proposed in this invention has significant theoretical innovation and engineering application value. Its role and significance are mainly reflected in two aspects:

[0059] (1) In terms of the rationality of parameter optimization results, the generative model VAE is used to explore the inherent correlation pattern of parameters and construct physical constraints, which significantly improves the physical interpretability of optimization results and effectively solves the key bottleneck problem that is mathematically feasible but physically distorted.

[0060] (2) In terms of computational efficiency and generalization ability, the spatiotemporal global DNN proxy model not only achieves efficient substitution for finite element calculation, but also enhances the spatiotemporal adaptability of the model through multi-physical feature input, providing an efficient and robust computational tool for the dynamic updating of the deformation analysis model of high rockfill dam.

[0061] In summary, this method provides reliable technical support for the full life-cycle safety management of high rockfill dams, combining physical rationality and predictive accuracy, and has significant engineering practical value for ensuring the safe operation of ultra-high rockfill dams. Attached Figure Description

[0062] Figure 1 A flowchart of a method for updating a deformation analysis model of a rockfill dam that considers the correlation of dam material parameters, provided in an embodiment of the present invention;

[0063] Figure 2 Information on earth-rock dams that provides dam material test parameter data for embodiments of the present invention;

[0064] Figure 3 The results of correlation analysis based on experimental parameters are from an embodiment of the present invention.

[0065] Figure 4 This is a diagram showing the reconstruction effect of the parameter correlation features of the DPK-VAE model in an embodiment of the present invention.

[0066] Figure 5 The prediction performance of the proxy model in this embodiment of the invention on the test set;

[0067] Figure 6 This illustrates the iterative evolution of the fitness and penalty terms of the multi-objective function in this embodiment of the invention.

[0068] Figure 7 This is a comparison of parameter correlation characteristics in the optimization results of the embodiments of the present invention, with and without prior knowledge.

[0069] Figure 8 This is a comparison of parameter correlation coefficients in the optimization results of the embodiments of the present invention, with and without prior knowledge.

[0070] Figure 9This is a comparison of the parameter distribution ranges in the optimization results of the embodiments of the present invention with and without prior knowledge.

[0071] Figure 10 As an embodiment of the present invention, a comparison of the relative errors between the predicted and measured deformations of different deformation analysis models is presented.

[0072] Figure 11 As an embodiment of the present invention, a comparison of finite element analysis results at different cross sections and monitoring points based on updated parameters and experimental parameters is presented.

[0073] Figure 12 Comparison of contour map results of dam crest deformation inclination in this embodiment of the invention. Detailed Implementation

[0074] Deep generative models, such as variational autoencoders (VAEs), have significant advantages in data feature learning and correlation modeling, enabling them to uncover the implicit parameter correlation structures within massive experimental data and provide physical constraints for parameter updates. Integrating parameter correlations as prior knowledge into the optimization process not only improves the physical rationality of model parameters but also avoids the solution space drift problem caused by neglecting the intrinsic relationships between parameters in traditional methods. Based on this, this invention proposes a method for updating a rockfill dam deformation analysis model that considers the correlation of dam material parameters. This method is based on a rockfill dam deformation constitutive model, integrating multi-source monitoring data, finite element simulation, and a deep learning surrogate model. Its core lies in introducing a prior knowledge constraint mechanism based on the correlation of dam material parameters, ensuring that model updates possess both numerical accuracy and physical rationality. Specifically, this mechanism utilizes VAEs to learn and extract the intrinsic correlation structures between constitutive parameters of dam materials from large-scale experimental data and embeds them as physical constraints into the parameter optimization and update process. Furthermore, to ensure stable model updates, a model update strategy that combines stability and efficiency was designed. By improving the multi-objective optimization algorithm (NSGA-III) and introducing an elite archiving strategy, the problem of easy relaxation of physical constraints during high-dimensional parameter optimization was solved. Practical applications show that the deformation analysis model updated using this method exhibits a significantly improved degree of agreement between the finite element results and the monitored values, demonstrating outstanding effectiveness and advancement in deformation prediction accuracy and safety assessment reliability.

[0075] The specific steps of the method proposed in this invention include:

[0076] (1) In the prior knowledge modeling stage of dam material parameter correlation, triaxial test data of multiple rockfill dam projects with different dam types, dam heights and lithologies were collected. Pearson correlation coefficient and p-value were used to identify strongly correlated combinations of parameters and construct a dimensionless correlation feature matrix. The DPK-VAE model was trained using this matrix as input. The deviation between the parameters and prior knowledge was quantified by the loss function of combined reconstruction loss and KL divergence, which verified the model's ability to reconstruct parameter correlation patterns and formed a reliable physical constraint basis.

[0077] (2) The spatiotemporal global DNN proxy model is constructed by generating samples in the parameter space using the LHS method. Combined with the Duncan-Zhang EB model and creep and wetting models, finite element simulation calculations covering the entire process of construction and water storage are carried out to construct a dataset of the mapping relationship between parameters and spatiotemporal deformation. By training a deep neural network model that integrates physical features such as the spatiotemporal coordinates of measuring points and filling elevation, and by using Bayesian optimization to fine-tune the hyperparameters, the model is ensured to efficiently replace finite element calculations and realize end-to-end mapping between parameters and deformation response.

[0078] (3) In the multi-objective parameter update stage guided by prior knowledge, a multi-objective function is constructed for each material partition, which includes the deformation prediction error and the prior knowledge penalty function. The weight of the penalty function is determined by combining Bayesian optimization. An improved NSGA-III algorithm with an elite archiving strategy is adopted. By dynamically retaining elite solutions and replacing inferior individuals with Gaussian perturbation to suppress constraint relaxation, non-dominated sorting and reference point selection are performed on the merged parent and offspring populations. Finally, the optimal parameters that balance accuracy and rationality are selected from the Pareto front to complete the model update.

[0079] (4) The effectiveness of the method was verified from three aspects: parameter correlation consistency, deformation prediction accuracy, and safety risk assessment, which fully verified the engineering applicability of the method.

[0080] 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.

[0081] like Figure 1 As shown, this embodiment uses a deformation analysis model update method for rockfill dams that considers the correlation of dam material parameters provided by the present invention to update the deformation analysis model of an LHK core rockfill dam.

[0082] 1. Constitutive model parameters of dam construction materials

[0083] Triaxial test parameters were collected from 48 rockfill dams of different types and scales both domestically and internationally (detailed information on the constitutive model parameters of the dam materials is shown in Table 2). These parameters cover two main dam types: concrete-faced rockfill dams (CFRD) and core-walled rockfill dams (CWRD), as well as the gradient distribution of dam heights for two types of earth-rock dams, with 200m as the dividing line. From a geological perspective, the dataset includes 18 lithological types, fully demonstrating the typicality and diversity of the engineering geological conditions of the test parameter data. Figure 2 Figures (a) and (b) respectively present statistical results on the maximum dam height and rock properties of these dams. Figure 2 Figure (c) shows the distribution of the seven collected experimental parameters, whose discrete ranges conform to engineering experience. This multi-dimensional, balanced, and statistically robust sample characteristic provides a reliable data foundation for subsequent deep learning models to explore the intrinsic correlations among parameters.

[0084] Table 2 Constitutive model parameters of rockfill dam lining materials .

[0085] 2. Parameter Correlation Feature Extraction

[0086] To investigate the correlation characteristics and quantify the dependencies among the parameters in the riprap test, a correlation coefficient matrix was first constructed based on the collected experimental parameter data. The Pearson correlation coefficient was then used to evaluate the correlation strength of each parameter pair. In the parameter correlation analysis, the Pearson coefficient quantifies the strength and direction of the correlation. To ensure the statistical reliability of the correlation analysis results, a p-value significance test was introduced to determine whether the correlation between each parameter pair was significant. The criteria for determining the statistical significance level were: p ≤ 0.05 indicates significant correlation, p ≤ 0.01 indicates highly significant correlation, and p ≤ 0.001 indicates extremely significant correlation. Figure 3 The correlation analysis results for each parameter variable are presented. Based on a comprehensive consideration of the absolute value of the correlation coefficient (|r|>0.5) and the significance level, several significantly correlated parameter combinations were selected, including: And n-m. These parameters show a clear correlation, indicating that they exhibit a cooperative variation law in the stress-deformation system response.

[0087] Five sets of parameter proportional characteristic factors g were calculated. n(See Equation 2 for specific calculations), forming the correlation feature matrix G = [g1, g2, …, g5]T∈R5. The processed features serve as a quantitative representation of parameter correlation, providing structured input data for subsequent prior knowledge models.

[0088] 3. Prior knowledge model based on variational autoencoder (VAE)

[0089] Based on the above analysis, a deep learning-based generative model VAE (Deep Prior Knowledge-VAE, DPK-VAE) was constructed as the prior knowledge model. The correlation feature matrix G, representing the parameter correlation, was used as the input and training data for the prior knowledge model. During training, a KL divergence annealing strategy was employed to implement a dynamic weight β adjustment mechanism: initially focusing on data reconstruction capabilities, β was gradually increased with each training round to enhance the regularization constraint of the latent space, thereby improving training stability. After training, negative ELBO (denoted as Loss) was used. prior This metric serves as an evaluation indicator to quantify the degree of deviation between model parameters and prior knowledge during the optimization process. A smaller metric indicates a higher consistency between the parameter results and the prior parameter correlation structure. By incorporating this metric into the parameter update process, the optimization is guided towards convergence in a solution space that better conforms to physical laws.

[0090] To verify the effectiveness of the DPK-VAE model, Figure 4 The correlation characteristics of the reconstructed parameters were compared with those of the original experimental data, with a focus on showcasing the reconstruction effect of the proportional characteristic factors of five key parameters. The results show that the distribution of the reconstructed parameter proportional factors is highly consistent with the original distribution, accurately reproducing the correlation characteristics between parameters. This indicates that the DPK-VAE model successfully extracts prior knowledge of parameter correlation, providing a reliable physical constraint mechanism for optimizing the constitutive parameters of dam materials. This ensures that the optimization results conform to actual engineering understanding, effectively compensating for the shortcomings of traditional parameter updates that neglect parameter correlation, and significantly improving the model's interpretability and engineering applicability.

[0091] 4. Finite element sampling and surrogate model training

[0092] Training data was generated based on finite element sampling, and a high-precision surrogate model was established to replace the time-consuming forward simulation calculations. For the four main material zones of the Lianghekou Dam (rockfill zone I, II, III, and core zone), key parameters of the Duncan-Chang EB model, the exponential decay creep model, and the three-parameter wetting model were selected as variables to be inverted (specific parameter information is shown in Table 2). The parameters mainly include... and Among them, parameter c EBThis is typically used for core wall analysis and therefore was not included in the parameter update combination. To simplify the analysis, a scaling factor λ is introduced. creep and λ wetting The creep model parameters (b, c) are adjusted uniformly respectively. creep , d) and humidification model parameters (c w ,n w , b w The final set of parameters to be updated is as follows: and The range of values ​​for each parameter was determined with reference to indoor triaxial test data and engineering experience to ensure coverage of parameter fluctuation ranges in actual engineering projects. The specific ranges are shown in Table 3.

[0093] Table 3 Sampling range of dam construction material parameters .

[0094] 400 parameter combinations were generated in the parameter space using Latin hypercube sampling (LHS). These parameter combinations served as the basis for the input data of the deep neural network (DNN) surrogate model, with each parameter combination corresponding to a complete finite element simulation calculation. The corresponding output dataset was constructed by extracting the deformation values ​​of each monitoring point at different times. The dataset was divided into training, validation, and test sets at a ratio of 70%, 20%, and 10%, respectively. The DNN surrogate model adopted a three-layer hidden layer structure, with 20-100 neurons in each layer, and was optimized using a Bayesian hyperparameter optimization method. The model training process used mean squared error (MSE) as the loss function, combined with Dropout and early stopping strategies to prevent overfitting.

[0095] Figure 5 The training process and test set performance of the proxy models for each partition are demonstrated. The results show that the models can accurately capture the material deformation characteristics, the predicted deformation on the test set is in high agreement with the finite element simulation results, and the models in all four partitions maintain high accuracy.

[0096] 5. Prior knowledge-guided model parameter updates

[0097] By integrating the previously trained prior knowledge model DPK-VAE, the deep neural agent model, and the improved NSGA-III algorithm, the parameters of the constitutive model of the rockfill dam with regularization term were optimized and updated. During the optimization process, the elite archiving mechanism was activated from the generation of the second-generation population, dynamically retaining and utilizing elite solutions through iterative updates, thus achieving the continuous transmission and diffusion of high-quality individual genes. Furthermore, for the screening and processing of multi-source monitoring data of the rockfill dam, measurement data from various monitoring instruments, including electromagnetic settlement rings, flexible inclinometers, and water-tube settlement meters, were ultimately selected for parameter inversion and updating. Detailed information is shown in Table 4.

[0098] Table 4. Number of measurement points used for model updates in each material partition. .

[0099] Note: ESG represents electromagnetic settlement gauge; SAA represents array displacement sensing system; HOSG represents water pipe settlement gauge; SDT represents soil displacement gauge.

[0100] After several trials, the weight of the prior knowledge regularization term was finally set to [2.25, 2, 2.25]. Figure 6 The objective function values ​​F for each material partition are shown. i (x) and penalty value P i The iterative optimization results of (x) are presented. The results show that the fitness value of the objective function for each partition decreases rapidly in the early stages of optimization and gradually stabilizes and converges after 120 iterations. Ultimately, the overall fitness value for each partition stabilizes at around 0.15. Furthermore, the hypervolume index, a key indicator for evaluating the quality of the Pareto solution set, monotonically increases until saturation during the optimization process, indicating that the coverage of the Pareto solution set over the objective space continuously expands during optimization, and the quality and diversity of the solution set improve simultaneously. These results fully validate the reliability and robustness of the improved multi-objective optimization algorithm in solving high-dimensional problems.

[0101] Finally, the equilibrium point decision method was used to select the optimal solution from the high-quality, high-coverage Pareto front as the final result of parameter update. The updated parameter results and the comparison with the experimental parameters are shown in Table 5.

[0102] Table 5 Comparison of Constitutive Model Update Parameters and Experimental Parameters .

[0103] Note: E represents the experimental parameter; U represents the updated parameter.

[0104] 6. Verification based on prior knowledge-guided mechanisms

[0105] The prior knowledge-guided mechanism, as the core of the deformation analysis model update method in this invention, aims to solve the common problem in traditional parameter optimization methods of being "mathematically feasible but physically unreasonable." By introducing a prior knowledge model (DPK-VAE), the potential correlation structure between parameters is learned from a large amount of experimental parameter data of earth-rock dam engineering, and this correlation is used as a physical constraint in the optimization process. To verify the effectiveness of this mechanism, a total of 60 sets of comparative experiments on parameter optimization with and without prior knowledge guidance were conducted.

[0106] Experimental results show that the prior knowledge-guided mechanism significantly improves the parameter optimization results. In K-K... b In the comparison of optimization results for three sets of significantly correlated parameter pairs, n-m and φ-∆φ, Figure 7 ), introducing Pareto front non-dominated solutions guided by prior knowledge ( Figure 7 The red scatter plots and the contour lines of the joint distribution kernel density of the experimental parameter data show a high degree of consistency. Specifically, these non-dominated solutions exhibit a clear linear correlation trend (red regression line), which is close to the overall trend of the joint distribution of the experimental parameters (blue regression line), and has a low degree of dispersion. In contrast, the optimization results without prior knowledge guidance ( Figure 7 (b) The non-dominated solutions of the parameters deviate significantly from the joint distribution pattern of the experimental data, with a significantly weakened correlation and a highly discrete distribution.

[0107] Figure 8 The correlation coefficient analysis further confirmed this. Under the guidance of prior knowledge ( Figure 8 a) Parameter pairs with clear correlation (K-K) b The correlation coefficients of (n-m, φ-∆φ) remained at a high level; while in the optimization results without prior knowledge guidance ( Figure 8 b) The correlation coefficients of these parameter pairs are generally low, and the correlation that should exist between the parameters is obviously lost.

[0108] Figure 9 The distribution characteristics of partial parameter optimization results under two schemes are shown. The optimization results guided by prior knowledge exhibit a narrower distribution range and lower randomness, while the results without prior knowledge have a wider distribution range and greater solution set dispersion. The essence of this difference lies in the fact that prior knowledge, as a physical constraint, effectively suppresses random disturbances in the parameter update process, enabling the optimization algorithm to focus on a physically reasonable and engineering-feasible solution domain.

[0109] 7. Deformation analysis and safety assessment based on the updated model

[0110] Based on the model update results guided by the aforementioned prior knowledge, a systematic deformation characteristic analysis and safety status assessment of the LHK Dam were conducted. By combining a high-precision finite element simulation model, a complete closed loop from deformation analysis model update to safety assessment was constructed, achieving a thorough understanding of the dam's deformation state and risk early warning.

[0111] To systematically evaluate the performance of the model update method, this study calculated the relative errors between predicted deformation and measured values ​​for three schemes: experimental parameters, update parameters without prior knowledge, and update parameters incorporating prior knowledge, using high-quality monitoring data from the four key partitions selected in Table 3. The results are shown in [Table 3]. Figure 10As shown, the overall prediction error based on the experimental parameters was 38.5%, while the updated parameter scheme significantly reduced the overall relative error of the dam to about 10.5%, a reduction of 28%. This result demonstrates that the method proposed in this invention significantly improves the model prediction accuracy by integrating monitoring data with prior knowledge for model updates, while ensuring the physical rationality of the parameter results, effectively solving the problem of neglecting the intrinsic correlation of parameters in traditional parameter updates.

[0112] Figure 11 This paper details the comparison between predicted and measured deformation data at different monitoring sections and points of the dam under two working conditions: one based on experimental parameters and the other on parameters updated based on prior knowledge. Further analysis shows that the model predictions after the digital twin framework update highly match the measured data, while the dam settlement calculated using the original experimental parameters is significantly greater than the measured value. This confirms that the method, through monitoring data-driven model updates, effectively corrects the overestimation of dam deformation by the original experimental parameters, enabling the virtual model to more accurately reflect the deformation behavior of the actual project. Furthermore, the settlement deformation cloud map on the right clearly shows the spatial distribution characteristics of dam settlement deformation during the full storage period (October 2024). Deformation is mainly concentrated in the middle of the dam and near the core wall area, with a reasonable spatial distribution and a certain degree of central symmetry. This distribution pattern is consistent with the typical deformation characteristics of high earth-rock dams, indicating that the digital twin framework method integrating prior knowledge not only improves the accuracy of deformation prediction but also ensures the deformation coordination of different material regions within the dam.

[0113] In the risk analysis of dam crest cracking Figure 12 This paper presents a cloud map showing the deformation gradient distribution along the river crest of the dam. Two strategies—one based on experimental parameters and the other on prior knowledge—were used to guide parameter updates. The predicted results for the dam completion period (December 2021) and the full storage period (October 2024) were compared. Analysis shows that during the dam completion period, the extreme values ​​of the deformation gradient under both conditions are around 0.5%, indicating a small and uniform overall deformation gradient. However, the deformation gradient at the dam crest increases significantly during the full storage period, especially under the experimental parameter conditions (…). Figure 12 The deformation gradient value of (a) reached the critical threshold of 1.1% in the downstream local region, while under the parameter update condition ( Figure 12 (b) The maximum deformation gradient is around 0.9%. Comparison with monitoring data shows that the deformation prediction results of the model based on updated parameters are closer to the measured data, and the deformation gradient is significantly lower than the calculation results based on experimental parameters. This indicates that the model calculation results based on experimental parameters may overestimate the deformation gradient at the dam crest and the risk of dam crest cracking. This model update method, by integrating multi-source monitoring data for prior knowledge-guided model updates, can more realistically reflect the actual deformation state of the physical entity, effectively avoiding conservative bias in risk assessment, and providing a more reliable decision-making basis for engineering safety assessment.

[0114] The implementation of the various embodiments of the present invention is based on programmed processing by 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 device for updating a deformation analysis model of a rockfill dam that considers the correlation of dam construction material parameters. This device is used to execute the method for updating a deformation analysis model of a rockfill dam that considers the correlation of dam construction material parameters in the above method embodiments. The device comprises: a first main module for collecting triaxial test parameters from multiple rockfill dam projects worldwide, quantifying the correlation between parameters, and using a variational autoencoder to learn parameter association patterns and construct a prior knowledge model; a second main module for generating samples in the parameter space using Latin hypercube sampling and random sampling methods, constructing a dataset by combining finite element simulation, and training a deep neural network surrogate model that integrates multiple physical features based on this dataset. The deep neural network surrogate model is used to map from dam material parameters to the spatiotemporal deformation response of the structure across the entire domain; a third main module for introducing an elite archiving strategy and improving the multi-objective optimization algorithm NSGA-III; and a fourth main module for integrating the prior knowledge model, the deep neural network surrogate model, and the improved multi-objective optimization algorithm NSGA-III to form a multi-objective optimization framework for the physical constraints of the dam material. This framework is used to optimize the constitutive model parameters of the dam material and update the deformation analysis model of the rockfill dam.

[0115] The rockfill dam deformation analysis model update device provided in this invention addresses key issues in the parameter update process of rockfill dam deformation analysis, including the low computational efficiency of traditional finite element methods, which makes it difficult to meet iteration requirements, and the physical distortion of optimization results caused by existing parameter inversion methods neglecting the physical correlation between parameters. By employing the aforementioned modules and combining deep learning technology with the physical mechanism of geotechnical materials, a multi-objective optimization algorithm framework with physical constraints is constructed for parameter updating and prediction in deformation analysis and safety monitoring of high earth-rock dams.

[0116] It should be noted that the device embodiments provided by the present invention, in addition to implementing the methods in the above method embodiments, are also used to implement the methods in other method embodiments provided by the present invention. The difference lies only in setting corresponding functional modules, and their principles are basically the same as those of the above device embodiments provided by the present invention. As long as those skilled in the art, based on the above device embodiments, refer to the specific technical solutions in other method embodiments, and obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and improve the device in the above device embodiments under the premise of ensuring the practicality of the technical solutions, they can obtain corresponding device-type embodiments for implementing the methods in other method-type embodiments. For example, based on the content of the above device embodiments, the rockfill dam deformation analysis model update device considering the correlation of dam material parameters provided in the embodiments of the present invention further includes: a fifth main module, used to calculate the global deformation field and key safety indicators based on the updated deformation analysis model.

[0117] 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 steps of the method for updating the deformation analysis model of a rockfill dam that considers the correlation of dam material parameters.

[0118] In summary, this invention proposes a method for updating the deformation analysis model of rockfill dams that considers the correlation of dam material parameters, aiming to solve the problems of low computational efficiency in model updates and physical distortion of parameter optimization results in traditional rockfill dam deformation analysis. This method first collects triaxial test parameters from multiple rockfill dams of different types and lithologies worldwide, quantifies the correlation strength of parameters using the Pearson correlation coefficient, and constructs a prior knowledge model (DPK-VAE) using a variational autoencoder (β-VAE) to mine the intrinsic correlation patterns of parameters and provide physical constraints. Second, it generates parameter samples using Latin hypercube sampling (LHS), constructs a dataset using finite element simulation, and trains a deep neural network (DNN) surrogate model that integrates spatiotemporal features to achieve high-precision mapping between dam material parameters and global deformation response. Subsequently, it introduces an elite archiving strategy to improve the multi-objective optimization algorithm NSGA-III, enhancing the robustness of high-dimensional optimization. Then, it integrates DPK-VAE, the DNN surrogate model, and the improved NSGA-III to form a multi-objective optimization framework, completing the parameter optimization of the dam material constitutive model to update the rockfill dam deformation analysis model. Finally, based on the updated model, it calculates the global deformation field and key safety indicators, achieving a thorough understanding of the dam's deformation characteristics and operational status. Research and verification show that this method can significantly improve the accuracy of deformation prediction, ensure the physical rationality of deformation analysis model updates, and provide scientific support for the safety assessment and decision-making of rockfill dams throughout their entire life cycle.

[0119] 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 updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters, characterized in that, include: Triaxial test parameters from multiple rockfill dam projects worldwide were collected, the correlation between parameters was quantified, and a variational autoencoder was used to learn the parameter association patterns to construct a prior knowledge model. Samples are generated in the parameter space using Latin hypercube sampling and random sampling methods. A dataset is constructed by combining finite element simulation. Based on this dataset, a deep neural network surrogate model that integrates multiple physical features is trained. The deep neural network surrogate model is used to map from dam material parameters to the spatiotemporal deformation response of the structure in the entire domain. An elite archiving strategy was introduced to improve the multi-objective optimization algorithm NSGA-III; By integrating prior knowledge models, deep neural network surrogate models, and the improved multi-objective optimization algorithm NSGA-III, a multi-objective optimization framework for the physical constraints of dam materials is formed. This framework is used to optimize the constitutive model parameters of dam materials and update the deformation analysis model of rockfill dams.

2. The method for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 1, characterized in that, The method further includes: Based on the updated deformation analysis model, the global deformation field and key safety indicators are calculated.

3. The method for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 1, characterized in that, The correlation between parameters is quantified, and a variational autoencoder is used to learn the parameter association patterns to construct a prior knowledge model, including: Construct a correlation coefficient matrix and use the Pearson correlation coefficient to evaluate the correlation strength of each parameter pair; Based on the highly correlated parameter combinations obtained after evaluation, proportional features are constructed to characterize the correlation structure between parameters; A variational autoencoder is used to construct a prior knowledge model.

4. The method for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 1, characterized in that, Samples are generated in the parameter space using Latin hypercube sampling and random sampling methods. A dataset is constructed using finite element simulation. Based on this dataset, a deep neural network surrogate model incorporating multiple physical features is trained, including: The sampled parameter combinations were substituted into the finite element analysis model for calculation. The calculation constitutive models used included the instantaneous deformation constitutive model based on the Duncan-Zhang EB model, the rheological constitutive model based on the Nanshui seven-parameter model, and the humidification constitutive model based on the Shenzhujiang three-parameter humidification model. A deep neural network proxy model is constructed. This model captures the highly nonlinear relationship between the material parameters and deformation of the rockfill dam through a multi-layer fully connected neuron and updates the weights using the backpropagation algorithm. At the same time, the Bayesian hyperparameter optimization method is used to enhance the learning ability of complex nonlinear features. The calculated training samples are divided into training set, test set and validation set according to the proportion, and the constructed deep neural network proxy model is trained, tested and validated.

5. The method for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 1, characterized in that, An elite archiving strategy is introduced to improve the multi-objective optimization algorithm NSGA-III, including: In each generation of evolution, all individuals in the current population are evaluated according to the comprehensive loss function, and the individuals are sorted according to the comprehensive fitness value from smallest to largest. Several optimal individuals are selected as elite candidates for this generation. The comprehensive loss function integrates the settlement prediction error term and the physical constraint penalty term. Merge current generation elite candidate solutions with archived historical elite individuals to form a larger set of candidate solutions; In each iteration, an elite perturbation sampling mechanism is adopted. Taking individuals in the elite archive as the center, a new solution is generated through Gaussian perturbation to replace the individuals in the current population that are sorted by a predetermined proportion according to non-dominance.

6. The method for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 1, characterized in that, By integrating prior knowledge models, deep neural network surrogate models, and the improved multi-objective optimization algorithm NSGA-III, a multi-objective optimization framework for the physical constraints of dam materials is formed. This framework is used to optimize the constitutive model parameters of the dam materials, thereby updating the deformation analysis model of the rockfill dam, including: A multi-objective optimization framework is constructed for each section of the rockfill dam. The objective function of the multi-objective optimization framework includes a basic fitness term and a regularization penalty function term. The basic fitness term is calculated by a surrogate model to obtain the predicted values ​​of the model parameters, and the root mean square relative error is used to evaluate the goodness of fit between the predicted values ​​and the measured data. The regularization term adopts a penalty function mechanism to quantify the degree of deviation between the model parameters and prior knowledge. By integrating prior knowledge models, deep neural network proxy models, and the improved multi-objective optimization algorithm NSGA-III, the parameters of the rockfill dam constitutive model are optimized and updated for the objective function. During the optimization process, the elite archiving mechanism is activated from the generation of the second generation population. Through iterative updates, elite solutions are dynamically retained and utilized to achieve the continuous transmission and diffusion of high-quality individual genes.

7. The method for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 6, characterized in that, Integrating prior knowledge models, deep neural network surrogate models, and the improved multi-objective optimization algorithm NSGA-III, the objective function is optimized and updated using the constitutive model parameters of the rockfill dam. Further steps include: S1, Initialization: Load the DNN proxy model and the prior knowledge model DPK-VAE, load the input data and perform normalization processing, read the measured deformation data and spatiotemporal feature information, and initialize the elite library; S2, Population Generation: An initial population P0 with a size of N is generated using a mixed strategy of Latin hypercube sampling and random sampling. S3, for iterations 1 to G, perform the following iterative optimization: S3.1, for each individual x∈P g : a) Calculate the four objective functions f i (x), the predicted value is calculated by the deep neural network surrogate model of each partition; b) Calculate the correlation regularization term P for dam construction material parameters. j (x) is calculated by reasoning from a prior knowledge model; c) Combine the objective function value and the penalty value to form a multi-objective fitness; S3.2, Update the elite library ε based on the comprehensive score to protect the preset elite solutions; S3.3, select, crossover, and mutate to generate new individuals, and replace αN worst individuals with elite perturbation sampling; S3.4, using the NSGA-III algorithm selection mechanism, updates to the next generation P. g +1; S3.5, Information Statistics and Callback; S4 outputs the Pareto front solution set P* and the optimal dam construction material parameter solution X*.

8. A device for updating a deformation analysis model for a rockfill dam that considers the correlation of dam material parameters, characterized in that: include: The first main module is used to collect triaxial test parameters of multiple rockfill dam projects around the world, quantify the correlation between parameters, and use a variational autoencoder to learn the parameter association pattern and build a prior knowledge model. The second main module is used to generate samples in the parameter space using Latin hypercube sampling and random sampling methods, construct a dataset by combining finite element simulation, and train a deep neural network surrogate model that integrates multiple physical features based on this dataset. The deep neural network surrogate model is used to map from dam material parameters to the spatiotemporal deformation response of the structure in the whole domain. The third main module is used to introduce an elite archiving strategy and improve the multi-objective optimization algorithm NSGA-III; The fourth main module is used to integrate the prior knowledge model, the deep neural network surrogate model, and the improved multi-objective optimization algorithm NSGA-III to form a multi-objective optimization framework for the physical constraints of the dam material. This framework is used to optimize the constitutive model parameters of the dam material and complete the update of the deformation analysis model of the rockfill dam.

9. The equipment for updating the deformation analysis model of a rockfill dam considering the correlation of dam material parameters according to claim 8, characterized in that, The device also includes: The fifth main module is used to calculate the global deformation field and key safety indicators based on the updated deformation analysis model.

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 perform the steps of the method for updating the deformation analysis model of a rockfill dam that considers the correlation of dam material parameters as described in any one of claims 1 to 7.

Images