A method and system for intelligent learning and reasoning optimization of a heart wall dam two-dimensional section structure

Abstract

The application discloses a kind of core wall dam two-dimensional section structure intelligent learning and reasoning optimization method and system, it is related to hydraulic engineering structure optimization technical field, method includes: based on the multi-source engineering data acquisition system of core wall rockfill dam, dam section structure design scheme set is constructed;Interaction framework of intelligent agent and finite element environment is constructed, and intelligent agent is iterated through the interaction with the state of finite element environment, action strategy is optimized, and the intelligent agent of training completion is obtained.The field deformation monitoring data collected by the settlement type monitoring sensor arranged in the core wall dam field monitoring space is used to correct the dam body mechanical deformation response field, improve the accuracy of the dam body mechanical deformation response field, improve the training accuracy of the intelligent agent, improve the optimization efficiency of the core wall dam two-dimensional section structure, realize the automation, intelligentization and high efficiency of the core wall rockfill dam structure design, provide important technical support for the intelligent structure design of hydraulic engineering, with remarkable practicality and wide application prospect.

Description

Technical Field

[0001] This invention relates to the fields of hydraulic structure design and artificial intelligence technology, specifically to an intelligent learning and reasoning optimization method and system for two-dimensional cross-sectional structures of core wall dams. Background Technology

[0002] Core-wall rockfill dams, or core-wall dams for short, are a major type of dam in water conservancy engineering. The structural design of their core wall and other sections directly affects the dam's seepage prevention safety, deformation coordination, and overall stability. Traditional core-wall dam structural design mainly relies on engineers' engineering experience and specifications. This approach requires a significant amount of time for tedious model building, mesh generation, parameter assignment, and result extraction during finite element method (FEM) verification, resulting in long structural design cycles, high labor costs, and difficulty in obtaining globally optimal solutions.

[0003] In recent years, intelligent optimization algorithms have been gradually applied in fields such as structural design and engineering parameter identification. However, existing methods still face many challenges in the structural design and optimization of core wall dams, especially under the coupled effects of multiple objectives and constraints such as dam slope stability, settlement control, core wall hydraulic gradient, and stress safety. Traditional heuristic optimization methods are prone to problems such as low search efficiency and limited local optimization capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent learning and reasoning optimization method and system for two-dimensional cross-sectional structures of core wall dams, which can effectively solve the aforementioned problems.

[0005] The technical solution adopted in this invention is as follows:

[0006] In a first aspect, the present invention provides an intelligent learning and reasoning optimization method for two-dimensional cross-sectional structures of core wall dams, comprising the following steps:

[0007] Step S1: Based on the multi-source engineering data acquisition system for core-wall rockfill dams, construct a set of dam cross-sectional structural design schemes; each dam cross-sectional structural design scheme includes dam cross-sectional geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data.

[0008] Step S2: Construct the interaction framework between the DQN agent and the finite element environment, including the finite element modeling and evaluation model and the DQN agent;

[0009] The finite element modeling and evaluation model takes the dam cross-sectional structure design scheme as input, and obtains a finite element mechanical response state that fully reflects the characteristics of the finite element mechanical response through finite element modeling, dam mechanical response simulation, dam mechanical response monitoring assimilation correction based on sparse on-site deformation monitoring data, and dam mechanical response time series prediction. Combining the constraint judgment results and the finite element mechanical response state, the finite element environment state is obtained as the output.

[0010] The DQN agent discretizes the adjustment range of the core wall geometric parameters into the agent's action space, constructs an agent environment state vector based on the finite element environment state, and constructs an agent reward function with the goal of minimizing the dam deformation response characteristic value and combining the constraint judgment results. The DQN agent optimizes its action strategy through interactive iteration with the finite element environment state to obtain a trained DQN agent.

[0011] Step S3: Using the trained DQN agent, the two-dimensional cross-sectional structure optimization design of the dam body is carried out based on the dam body cross-sectional structure design scheme given in the new project of the heart wall dam.

[0012] Secondly, the present invention provides an intelligent learning and reasoning optimization system for two-dimensional cross-sectional structures of core wall dams, comprising:

[0013] The dam cross-section structure design scheme set construction module is used to construct a dam cross-section structure design scheme set based on the multi-source engineering data acquisition system for core-wall rockfill dams. Each dam cross-section structure design scheme includes dam cross-section geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data.

[0014] The interaction framework between DQN agents and the finite element environment includes the finite element modeling and evaluation model and the DQN agent.

[0015] The finite element modeling and evaluation model takes the dam cross-sectional structure design scheme as input, and obtains a finite element mechanical response state that fully reflects the characteristics of the finite element mechanical response through finite element modeling, dam mechanical response simulation, dam mechanical response monitoring assimilation correction based on sparse on-site deformation monitoring data, and dam mechanical response time series prediction. Combining the constraint judgment results and the finite element mechanical response state, the finite element environment state is obtained as the output.

[0016] The DQN agent discretizes the adjustment range of the core wall geometric parameters into the agent's action space, constructs an agent environment state vector based on the finite element environment state, and constructs an agent reward function with the goal of minimizing the dam deformation response characteristic value and combining the constraint judgment results. The DQN agent optimizes its action strategy through interactive iteration with the finite element environment state to obtain a trained DQN agent.

[0017] The dam body two-dimensional cross-section optimization module is used to carry out the two-dimensional cross-section structural optimization design of the dam body based on the dam body cross-section structural design scheme given in the new core wall dam project, using a trained DQN agent.

[0018] The intelligent learning and reasoning optimization method and system for two-dimensional cross-sectional structures of core wall dams provided by this invention have the following advantages:

[0019] By collecting on-site deformation monitoring data from settlement monitoring sensors deployed in the on-site monitoring space of the core wall dam, the mechanical deformation response field of the dam body is corrected, improving the accuracy of the mechanical deformation response field of the dam body, thereby improving the accuracy of training the DQN agent and increasing the optimization efficiency of the two-dimensional cross-sectional structural parameters of the core wall dam. This realizes the automation, intelligence and efficiency of the structural design of the core wall rockfill dam, and provides important technical support for the intelligent structural design of water conservancy projects. Attached Figure Description

[0020] Figure 1 A flowchart of the intelligent learning and reasoning optimization method for the two-dimensional cross-sectional structure of the core wall dam provided by the present invention;

[0021] Figure 2 This is a schematic diagram of the two-dimensional cross-sectional parameterized structure of the core wall dam provided by the present invention;

[0022] Figure 3 The diagram illustrates the interaction framework between the DQN agent and the finite element environment provided by this invention.

[0023] Figure 4 The convergence curves of the reward function and loss function of the DQN deep reinforcement learning algorithm provided by this invention during the training process;

[0024] Figure 5 The image shows a two-dimensional cross-sectional view of the core wall in a typical case of optimized structural design provided by this invention. Detailed Implementation

[0025] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0026] This invention provides an intelligent learning and reasoning optimization method and system for the two-dimensional cross-sectional structure of a core-wall dam. By collecting on-site deformation monitoring data from settlement monitoring sensors deployed in the core-wall dam's on-site monitoring space, the method corrects the dam's mechanical deformation response field, improving its accuracy. This, in turn, enhances the accuracy of training the DQN agent and increases the optimization efficiency of the core-wall dam's two-dimensional cross-sectional structural parameters. This achieves automation, intelligence, and high efficiency in the design of core-wall rockfill dam structures, providing crucial technical support for intelligent structural design in water conservancy projects. It possesses significant practicality and broad application prospects.

[0027] See Figure 1 This invention provides an intelligent learning and reasoning optimization method for two-dimensional cross-sectional structures of core wall dams, comprising the following steps:

[0028] Step S1: Based on the multi-source engineering data acquisition system for core-wall rockfill dams, construct a set of dam cross-sectional structural design schemes; each dam cross-sectional structural design scheme includes dam cross-sectional geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data.

[0029] Specifically, the structural design scheme for each dam section is characterized as follows: ; The data includes dam cross-sectional geometric parameters, material zoning characteristics, engineering load conditions, and sparse field deformation monitoring data. The dam cross-sectional geometric parameters include core wall geometric parameters and other material zoning geometric parameters. The engineering load conditions include construction loading paths and water level change paths. The water level change paths include normal storage level, design flood level, check flood level, construction period water level, stable seepage condition water level, and rapid precipitation condition water level. The sparse field deformation monitoring data includes a field monitoring data matrix in the sparse field monitoring space and characteristics of each monitoring point, such as monitoring point location, type, and missing measurement markers.

[0030] The specific implementation method of this step includes steps S11 to S12:

[0031] Step S11: Establish a multi-source engineering data acquisition system for intelligent optimization of core-wall rockfill dams: collect two-dimensional cross-sectional structural design drawings, structural design data, geological survey data, material test results, water level and working condition data, and sparse on-site deformation monitoring data during the construction period, water storage period, and operation period of typical core-wall rockfill dam projects.

[0032] The structural design data includes dam height, dam crest width, upstream and downstream slope ratios, coordinates of control points on the dam's outer contour, core wall top width, core wall upstream and downstream slope ratios, core wall bottom width, filter layer thickness, filter layer slope ratio, transition layer thickness, transition layer slope ratio, upstream and downstream rockfill area range, and foundation boundary dimensions.

[0033] The material parameters include the density, elastic modulus, Poisson's ratio, cohesion, internal friction angle, permeability coefficient, and constitutive model parameters of the core wall material, filter material, transition material, riprap material, and foundation material.

[0034] The water level data includes normal storage level, design flood level, check flood level, construction period water level, stable seepage condition water level, and rapid precipitation condition water level.

[0035] The sparse field deformation monitoring data includes field monitoring data and its spatial registration information. The field monitoring data includes at least one of the following: settlement monitoring points, horizontal displacement monitoring points, inclinometer tubes, plumb lines, stratified settlement, piezometers, and fiber optic monitoring lines; the spatial registration information includes the coordinates of the monitoring points or monitoring lines, the corresponding finite element unit number, the monitoring type, the monitoring time, the missing measurement marker, and the monitoring residual covariance.

[0036] Step S12: Establish a parameterized data representation mechanism for the two-dimensional cross-sectional structure design. This involves uniformly encoding the dam height, dam crest width, upstream and downstream slope ratios, core wall crest width, upstream and downstream slope ratios of the core wall, filter layer thickness and slope ratio, transition layer thickness and slope ratio, rockfill zone parameters, foundation layering parameters, reservoir water level parameters, and sparse on-site deformation monitoring data into structured feature vectors. This forms the dam cross-sectional structure design scheme, which is used for subsequent finite element automatic modeling and deep reinforcement learning model training. Furthermore, the dam cross-sectional structure design scheme can include dam safety evaluation index thresholds, including the maximum dam settlement threshold, dam slope stability safety factor threshold, maximum hydraulic gradient of the core wall, and hydraulic fracturing judgment results determined based on the maximum effective minor principal stress of the core wall.

[0037] Step S2: Construct the interaction framework between the DQN agent and the finite element environment, including the finite element modeling and evaluation model and the DQN agent;

[0038] The finite element modeling and evaluation model takes the dam cross-sectional structure design scheme as input, and obtains a finite element mechanical response state that fully reflects the characteristics of the finite element mechanical response through finite element modeling, dam mechanical response simulation, dam mechanical response monitoring assimilation correction based on sparse on-site deformation monitoring data, and dam mechanical response time series prediction. Combining the constraint judgment results and the finite element mechanical response state, the finite element environment state is obtained as the output.

[0039] The DQN agent discretizes the adjustment range of the core wall geometric parameters into the agent's action space, constructs an agent environment state vector based on the finite element environment state, and constructs an agent reward function with the goal of minimizing the dam deformation response characteristic value and combining the constraint judgment results. The DQN agent optimizes its action strategy through interactive iteration with the finite element environment state to obtain a trained DQN agent.

[0040] This step mainly includes two parts: the finite element modeling and evaluation model and the DQN agent. The principles of the finite element modeling and evaluation model and the DQN agent are described in detail below:

[0041] (I) Finite Element Modeling Evaluation Model

[0042] The finite element modeling and evaluation model is used for each dam cross-section structural design scheme to output the finite element environment state for DQN agent perception. It consists of two parts. First, through finite element modeling, dam mechanical response simulation, dam mechanical response monitoring assimilation correction based on sparse on-site deformation monitoring data, and dam mechanical response time series prediction, a finite element mechanical response state that comprehensively reflects the characteristics of the finite element mechanical response is obtained. Second, by combining the constraint judgment results and the finite element mechanical response state, the finite element environment state is obtained as the output.

[0043] (1) Finite element modeling and evaluation model to obtain finite element mechanical response state

[0044] Because core-wall dams typically deploy monitoring equipment such as settlement, horizontal displacement, inclinometers, plumb bobs, piezometers, or fiber optics during construction, impoundment, and operation, on-site monitoring data can reflect the actual deformation response of the dam body. However, the number of monitoring points is limited and their spatial distribution is sparse. While the finite element model can provide a complete spatial field, it is subject to deviations from the measured response due to material parameters, constitutive models, construction process simulations, and boundary conditions. If DQN optimization only uses the maximum settlement obtained from finite element simulation as the evaluation criterion, it may not fully reflect the corrective effect of the actual monitoring response on the safety of the structural design. Therefore, this invention introduces an intelligent dam deformation correction mechanism based on monitoring data assimilation into the DQN environmental assessment algorithm. This allows the finite element prior field to be dynamically corrected based on on-site monitoring data, and the posterior deformation field is used for reward functions and constraint determination.

[0045] Step A1, the finite element modeling evaluation model includes a core wall dam finite element modeling model, a dam mechanical response simulation model, a pre-trained core wall dam deformation field time series prediction model, and a feature extraction model;

[0046] Step A2, for the input dam cross-sectional structure design scheme A finite element model of a core-wall dam was adopted, based on the geometric parameters of the dam cross-section. Perform parametric modeling and partition features according to material. Automatically assign constitutive parameters, permeability parameters, and strength parameters to each material zone of the cross-section, and model the corresponding core-wall dam finite element model. ;

[0047] As one implementation method, see [link / reference] Figure 2 Based on the geometric parameters of the dam cross section When performing parametric modeling, a two-stage geometric analytical algorithm of core-driven and layer-by-layer expansion is adopted to deduce the coordinates of geometric control points of each zone, and to map the deduced geometric boundaries and foundation boundary conditions of each zone of the full section into a recognizable two-dimensional finite element model.

[0048] Two-stage geometric analytical algorithm:

[0049] In the first stage, the current top width of the heart wall is input. The slope of the heart wall Downstream slope ratio of the core wall

[0050] As the core driving variable, combined with the known elevation of the top of the core wall and bottom elevation The coordinates of the four vertices of the core wall region are calculated using analytical geometry to generate the core wall geometric region.

[0051] In the second stage, the boundary of the core wall geometric region generated in the first stage is used as the starting benchmark. Combined with the top elevation of each zone, the thickness of the filter layer at the top is determined according to the preset thickness. and the thickness of the top of the transition layer A horizontal equidistant offset calculation is performed outward to determine the coordinates of the boundary nodes at the top of each internal functional zone. Then, starting from the aforementioned top boundary nodes, geometric rays are drawn downward according to the preset slope ratio of each layer, and the intersection points of the rays with the dam body base contour line are solved to obtain the coordinates of the control nodes at the bottom of each zone, thereby generating the complete geometric boundaries of the internal filter layer and transition layer. Further, the outermost determined zone boundary is spatially closed with the fixed dam body outer contour line to automatically define and generate the actual range of the upstream and downstream rockfill areas. Finally, the complete set of geometric boundaries of each zone across the entire cross section derived above is converted into a mesh model of a finite element model that can be recognized by finite element software, and automated mesh generation is performed.

[0052] The coordinates of the intersection of the vertical centerline of the dam axis and the elevation line of the top of the core wall As the coordinate reference point Assume the elevation difference of the core wall The coordinates of the four control vertices of the heart wall can be represented as:

[0053]

[0054]

[0055]

[0056]

[0057] In the formula, For the width of the top of the heart wall, For the slope ratio of the heart wall, For the downstream slope ratio of the core wall, and These are the top and bottom elevations of the core wall, respectively. , , and These are the identifiers for the four core wall control vertices.

[0058] After the geometric model is generated, the system automatically completes the regional partitioning, material assignment, contact or interface relationship setting, boundary condition setting, and mesh generation. For critical areas such as the core wall, filter layer, and transition layer, locally refined meshing is preferred; for the foundation or rockfill area far from the core area of ​​the dam, a relatively coarse mesh can be used to balance computational accuracy and efficiency.

[0059] The finite element model can automatically set analysis steps according to various working conditions, such as construction filling (corresponding to waterless or low water levels during construction), water storage, stable seepage (corresponding to normal water storage levels), dam slope stability, and special loads (corresponding to design / check flood levels, sudden water level drops, and earthquakes), and call the finite element solver to perform calculations. After the calculation is completed, information such as deformation field, stress field, hydraulic gradient field, and dam slope stability safety factor is automatically extracted from the result file and converted into standardized evaluation indicators that can be called by the deep reinforcement learning environment.

[0060] Step A3: Using the dam body mechanical response simulation model, the finite element model of the core wall dam is... Loading engineering load conditions Finite element response simulation was performed to obtain the complete finite element response field. Including the finite element complete deformation response field Finite element complete stable response field Finite element complete seepage response field and finite element complete stress response field ;

[0061] It should be noted that, for deformation response, stability response, seepage response, and stress response, deformation monitoring data reflecting deformation response is easier to obtain in actual engineering monitoring. Therefore, this invention uses on-site deformation monitoring data from actual engineering monitoring, such as settlement data, to correct the complete deformation response field of the finite element method. It uses minimizing dam settlement, which represents deformation characteristics, as the objective function, while stability response, seepage response, and stress response serve as constraints. Thus, by using readily available on-site deformation monitoring data, it achieves the advantages of improving the accuracy and efficiency of agent training.

[0062] Step A4: A pre-trained time-series prediction model for the deformation field of the core wall dam is used, based on sparse on-site deformation monitoring data. Response field of complete deformation of finite element The model of the core wall dam is realized by making corrections in the latent space and loading various engineering load conditions. Comprehensive simulation analysis under multiple working conditions was conducted to extract the maximum settlement characteristic value of the dam body, which reflects the deformation response characteristics of the dam body. ;

[0063] Step A42, Initial Correction Stage: Response Field of Finite Element Complete Deformation By performing monitoring and assimilation correction, the complete deformation response field of the finite element after monitoring and assimilation is obtained. ,include:

[0064] Step A421, Sparse Field Deformation Monitoring Data Including field monitoring data matrix in sparse field monitoring space Based on the observation operator, the complete deformation response field of the finite element is obtained. Mapping to a sparse field monitoring space to obtain a matrix of field monitoring data. Comparable prior monitoring prediction matrix And based on the field monitoring data matrix With the prior monitoring prediction matrix Construct monitoring residual matrix based on the difference between them The monitoring residual matrix is ​​processed by a pre-trained monitoring encoder. Feature extraction is performed to obtain sparse monitoring and correction features. ;

[0065] Step A422: Use a pre-trained shared VAE encoder to input the complete deformation response field of the finite element method. Encoded as prior latent variables in a low-dimensional latent space The expression is: , This represents a shared VAE encoder; based on the sparse monitoring correction features in a low-dimensional latent space, a pre-trained latent variable updater is used. Correcting the prior latent variables , obtain the posterior latent variables The expression is: , This represents a latent variable updater. As an optional implementation, the latent variable updater first obtains the latent variable corrections. Then use the formula Obtain posterior latent variables .

[0066] This step involves updating latent variables within the latent space. The fusion update within the latent space is achieved through a latent variable updater, which takes prior latent variables and sparse monitoring correction features as inputs, updating the prior latent variables to posterior latent variables. Therefore, this invention ensures that the correction of the complete stress-deformation field by monitoring data occurs within the latent space. The core of this step lies in the sparse monitoring correction features extracted from the monitoring residuals. Correcting prior latent variables Instead of generating a complete stress-deformation field directly from sparse monitoring points.

[0067] Step A423, the posterior latent variable The input is fed into a pre-trained shared VAE decoder, which converts the posterior latent variable... Reconstructed into a complete finite element deformation response field after monitoring assimilation The monitored assimilated finite element complete deformation response field That is, the response field to the complete deformation of the finite element. The correction results;

[0068] In this step, the shared VAE decoder can employ at least one of the following: a direct decoding structure, a residual decoding structure, or a modal residual decoding structure. In the direct decoding structure, the decoding and reconstruction process can be represented as follows: In this approach, the posterior latent variables are input into the decoder of the shared variational autoencoder to obtain the complete finite element deformation response field.

[0069] Step A43, Time Series Prediction Stage: Apply different engineering load conditions to perform time series prediction, and obtain the finite element complete deformation response field prediction sequence, including:

[0070] Based on the dam body cross-sectional structure design scheme Engineering load conditions Based on this, the engineering load conditions and sparse on-site deformation monitoring data of the same dam body in the next engineering stage are obtained by searching the dam body cross-sectional structural design scheme set; the assimilated finite element complete deformation response field is then used. The engineering load conditions for the next engineering stage are input into the autoregressive propagation model to predict the finite element complete deformation response field for the next engineering stage. Then, the monitoring assimilation correction method in step A42 is used to obtain the monitored and assimilated finite element complete deformation response field for the next engineering stage. This process is repeated continuously to achieve rolling prediction and obtain the finite element complete deformation response field prediction sequence.

[0071] In this invention, the autoregressive propagation model can be any one or a combination of GRU, Transformer, Temporal Convolution, Neural ODE, or PINN state-space model.

[0072] Step A44, Feature Extraction: The finite element complete deformation response field after monitoring and assimilation from the initial correction stage is extracted. Feature extraction was performed to obtain the maximum settlement characteristic value of the dam body in the initial correction stage. Extract the monitoring residual matrix from the initial correction stage. Feature extraction was performed on the finite element complete deformation response field prediction sequence to obtain the characteristic value of the maximum settlement of the dam body in the time-series prediction stage. ;

[0073] Step A45, Feature Fusion: For the maximum settlement characteristic values ​​of the dam body in the initial correction stage... Monitoring residual matrix during the initial correction phase and the characteristic value of maximum settlement of the dam body during the time-series prediction stage. Feature fusion is performed, and prediction uncertainties are considered to obtain the maximum settlement characteristic value of the dam body, which reflects the deformation response characteristics of the dam body. .

[0074] As one implementation method, the characteristic value of the maximum settlement of the dam body Calculate using the following formula:

[0075]

[0076] in: This represents the maximum allowable settlement of the dam body. express Transitions; Indicates the monitoring residual covariance; Indicates the number of valid monitoring points; This indicates uncertainty in forecasting; The weights are for the corresponding items.

[0077] Step A5: Using a feature extraction model, the finite element complete stable response field is analyzed. Finite element complete seepage response field and finite element complete stress response field Feature extraction is performed to obtain the characteristic value of the dam slope stability safety factor, which reflects the stability response characteristics of the dam body. The maximum hydraulic gradient characteristic value of the core wall, reflecting the seepage response characteristics of the dam body. And the maximum effective minor principal stress of the core wall, which reflects the stress response characteristics of the dam body. ;

[0078] Step A6: Generate the finite element mechanical response state .

[0079] In the finite element modeling and evaluation model used in this embodiment, it is necessary to pre-train the shared variational autoencoder composed of a shared VAE encoder and a shared VAE decoder. The training method is as follows:

[0080] During training, for finite element complete deformation response field samples The shared VAE encoder will sample the complete deformation response field of the finite element method. Mapped to latent variables in a low-dimensional latent space The latent variable distribution parameters, including the mean and standard deviation The expression is , Indicates a shared VAE encoder; based on the mean. and standard deviation The latent variables are obtained by reparameterization. The expression is: ;in, This represents a random perturbation vector sampled from a standard multivariate normal distribution. This represents a standard multivariate normal distribution with a mean of 0 and a covariance matrix of identity matrix I, where I represents the identity matrix. This represents element-wise multiplication;

[0081] During training, the shared VAE decoder will store latent variables in the low-dimensional latent space. Reconstruction yields the reconstructed finite element complete deformation response field. The expression is , Indicates a shared VAE decoder;

[0082] The training loss of the shared variational autoencoder Represented as:

[0083]

[0084]

[0085] in: The reconstruction loss of the complete deformation response field in the finite element method; For latent space regularization constraints; Indicates the regularization weight; Indicates KL divergence; This represents the finite element complete deformation response field generated by the shared VAE encoder based on the input. latent variables obtained The approximate posterior distribution, This represents the pre-defined prior distribution of latent variables.

[0086] Furthermore, in this invention, after training with a shared VAE encoder and a shared VAE decoder, to enable the prediction model of this invention to learn the systematic errors that may occur in the finite element model of a real engineering center wall dam, a biased prior field and a target field pairing sample are constructed. The biased prior field can be generated through high- and low-precision finite element model pairing, dam height scale deviation, material parameter perturbation, construction path deviation, water level response sensitivity deviation, creep or wetting parameter deviation, and latent space perturbation. Among these, the dam height scale deviation can be used to simulate systematic errors such as underestimating the calculated displacement of high dams and overestimating the calculated displacement of low dams.

[0087] It should be noted that the paired samples of the target field and the biased prior field are only used in the offline training phase to enable the monitoring encoder and the latent variable updater to learn the mapping relationship of the prior latent variables corrected by the monitoring residuals. In the online application phase, the pseudo-field monitoring data matrix obtained by mapping the target field to the sparse field monitoring space is replaced by the field monitoring data. The monitoring residual is directly formed by the field monitoring data and the current prior monitoring prediction value to update the current prior latent variables posteriorly, so as to realize the assimilation of monitoring data to update the prior latent variables mapped to the prior deformable field and realize the correction of the finite element model error.

[0088] The specific training method is as follows:

[0089] Step D1: Construct paired samples with the biased prior field and the target field;

[0090] For each dam cross-section structural design scheme in the dam cross-section structural design scheme set , The data includes the dam's cross-sectional geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data. A complete finite element deformation response field is obtained through a finite element modeling model of the core-wall dam and a dam mechanical response simulation model. The design scheme of the dam cross-section was incorporated, thereby constructing a multi-condition finite element complete deformation response field sample library; the finite element complete deformation response field in each sample As the target field, it is re-represented as The target field is constructed using a perturbation algorithm. Paired biased prior fields ;

[0091] Biased a priori field and target field The relationship between them can be represented as , This indicates a systematic deviation introduced by material parameters, construction path, water level conditions, scale effects, or latent space disturbances.

[0092] As an example, an engineering condition perturbation algorithm or a latent space perturbation algorithm can be used to obtain the target field. Paired biased prior fields :

[0093] Engineering condition disturbance algorithm: for engineering load conditions According to the set deviation strength parameters To perform the perturbation, use the deviation generation function. Generate system deviation The expression is ; to system bias Acting on the target field , obtain the target field Paired biased prior fields ;

[0094] The latent space perturbation algorithm is expressed as:

[0095]

[0096]

[0097] target field The input is sent to the shared VAE encoder, which will then input the target field. Encoding target latent variables in a low-dimensional latent space ;

[0098] For target latent variables By perturbing the target latent variable, we can obtain the perturbed target latent variable. The expression is: ;in, Indicates the intensity of the disturbance; Let represent the latent space perturbation term, and let represent the random perturbation vector sampled from the multivariate normal distribution. This indicates that the mean is 0 and the covariance matrix is... The multivariate normal distribution, It is an identity matrix.

[0099] Using the shared VAE decoder, the perturbed target latent variables are... Reconstructed into a complete finite element deformation response field, which serves as the target field. Paired biased prior fields The expression is ;, This indicates a shared VAE decoder.

[0100] Step D2, using the same observation operator , respectively target field and biased a priori fields Mapping to the sparse field monitoring space yields the target monitoring prediction matrix for the sparse field monitoring space. And biased monitoring prediction matrix The expression is and ;

[0101] Step D3: Calculate the target monitoring prediction matrix. As a pseudo-site monitoring data matrix, the biased monitoring prediction matrix will be used. As the prior monitoring and prediction matrix, the formula is used. Obtain the pseudo-monitoring residual matrix The data is then input into the monitoring encoder to obtain pseudo-sparse monitoring correction features. ;

[0102] Step D4, respectively target fields and biased a priori fields The input is sent to the shared VAE encoder, which will then input the target field. Encoding target latent variables in a low-dimensional latent space There will be a biased prior field Encoded as biased prior latent variables in a low-dimensional latent space ;

[0103] Step D5, adjust the pseudo-sparse monitoring correction features. and the biased prior latent variables The latent variable updater is input, and the latent variable updater learns the pseudo-sparse monitoring and correction features. For the biased prior latent variables The latent variable correction amount is used to update the biased prior latent variable. Obtain biased posterior latent variables Through training, the biased posterior latent variables are made... Continuously approaching the target latent variable The monitoring encoder and the latent variable updater are trained to obtain the trained latent variable correction model.

[0104] (2) Combining the constraint determination results and the finite element mechanical response state, the finite element environment state is obtained as the output:

[0105] Step B1: Establish a multi-level constraint system for the core wall dam, including dam cross-sectional geometric constraints, dam deformation constraints, dam slope stability constraints, dam seepage constraints, and dam stress constraints.

[0106] Establish geometric constraints for the dam cross-section: the core wall, filter layer, and transition layer must not cross the dam boundary; the thickness of each functional zone must not be less than the specification limit; the dam slope ratio must meet stability requirements; the core wall, filter layer, and transition layer should be continuously formed and not intersect; the core wall area and bottom width must not be too small and must meet seepage control requirements. Therefore, when the finite element model shows geometric invalidity, mesh generation failure, calculation non-convergence, or abnormal results, the current structural design state is marked as infeasible, and the corresponding infeasibility flag and penalty information are returned to the finite element environment.

[0107] In practical applications, the geometric constraints of the dam cross-section can be expressed as:

[0108]

[0109]

[0110]

[0111] in, This refers to the actual thickness of each functional zone, such as the filter layer and the transition layer. The minimum thickness required by the specification; and These are the two-dimensional cross-sectional area of ​​the core wall and the minimum requirements for meeting seepage prevention requirements; and These refer to the bottom width of the core wall and the minimum limit required to meet seepage prevention requirements.

[0112] Establish a stability constraint judgment system: characteristic value of dam slope stability safety factor calculated based on strength reduction method Not less than the minimum allowable value of the dam slope stability safety factor ,Right now ;

[0113] Establish a deformation constraint judgment system: characteristic value of maximum settlement of dam body Not exceeding the maximum allowable settlement of the dam body ,Right now ;

[0114] Establish a seepage constraint determination system: characteristic value of the maximum hydraulic gradient of the core wall. Not exceeding the maximum allowable hydraulic gradient of the core wall ,Right now ;

[0115] Establish a stress constraint determination system: through Indicates the maximum effective minor principal stress of the core wall. Whether the hydraulic fracturing risk assessment item derived from tensile stress appears; in this invention, according to the sign convention that tensile stress is positive and compressive stress is negative, when When the maximum effective minor principal stress of the core wall does not exhibit tensile stress, this can be used as a criterion for determining whether hydraulic splitting has occurred. =1; when When the maximum effective minor principal stress of the core wall reaches tensile stress, it serves as the criterion for determining whether hydraulic fracturing has occurred. ;

[0116] Step B2: Determine whether the dam cross-sectional structural design scheme conforms to the geometric constraints of the dam cross-section and the degree of deviation from the relative constraint boundaries; determine the maximum settlement characteristic value of the dam body respectively. Characteristic value of dam slope stability safety factor eigenvalues ​​of the maximum hydraulic gradient of the core wall Maximum effective minor principal stress of the core wall The constraint determination result is obtained by determining whether the dam body deformation constraint, dam slope stability constraint, dam body seepage constraint, and dam body stress constraint are met, as well as the degree of deviation from the relative constraint boundary.

[0117] Step B3, the constraint determination result and the finite element mechanical response state By splicing the data, we obtain the finite element environment state as the output.

[0118] (ii) DQN agent

[0119] In this invention, a parametric optimization model is established with the core wall region as the core: the top width of the core wall and the upstream and downstream slope ratio of the core wall are used as optimization variables. An optimization objective function is established: minimizing dam settlement is the core objective, and a weighted summation mechanism is used for comprehensive evaluation, combining the geometric characteristics of the core wall, the dam slope stability safety factor, the hydraulic gradient of the core wall, and the hydraulic fracturing constraint conditions. A multi-level constraint system for the core wall dam is established.

[0120] Furthermore, this invention constructs the optimization process of the two-dimensional cross-section of the core wall dam as a deep reinforcement learning problem, with the finite element modeling evaluation model as the environment, the DQN agent as the decision-making subject, the adjustment of the core wall structure design parameters as the action, and the evaluation results and constraint satisfaction of the finite element modeling evaluation model as environmental feedback.

[0121] A DQN agent is constructed based on a parameterized optimization model centered on the heart-wall region. The DQN agent optimizes its action strategy through iterative interaction with the finite element environment state, resulting in a trained DQN agent, including:

[0122] Step C1, Define the agent's action space :

[0123] in, For the variation in the width of the top of the core wall, For the change of the slope ratio of the heart wall, This refers to the change in the downstream slope ratio of the core wall.

[0124] Therefore, the top width, upstream slope ratio, and downstream slope ratio of the core wall geometric parameters are taken as the objects of action adjustment; the adjustment range of the top width, upstream slope ratio, and downstream slope ratio of the core wall is discretized into a finite set of actions to form the action space of the intelligent agent; the actions of the intelligent agent include: performing an operation to increase, decrease, or remain unchanged on the top width of the core wall by a set range, and performing an operation to increase, decrease, or remain unchanged on the upstream slope ratio of the core wall by a set range, and performing an operation to increase, decrease, or remain unchanged on the downstream slope ratio of the core wall by a set range;

[0125] Therefore, in each decision step, the DQN agent selects a specific action from the action space according to the policy and applies it to the current dam structure design parameters.

[0126] Furthermore, to improve the agent's search efficiency in the discrete action space, an action masking strategy is introduced during the action selection phase: an anti-oscillation mask is set to prevent the agent from outputting actions that are completely opposite to the previous decision step, thus eliminating invalid reciprocal oscillations in the A→B→A pattern; simultaneously, a stagnation suppression mask is set to forcibly block the option when the agent continuously executes "remain unchanged" actions to a preset threshold, thereby encouraging space exploration. The agent only selects the optimal action probabilistically based on the current state from the subset of legal actions filtered by the mask.

[0127] Step C2, constructing the agent's environment state vector: Normalize the initial dam cross-sectional structure design scheme, dam characteristic parameters, the current decision-step dam cross-sectional structure design scheme, and the current decision-step finite element environment state, then combine and splice them into the agent's environment state vector for the current decision step. Dam characteristic parameters can include dam cross-sectional geometric features and material zoning features, etc.

[0128] Step C3, construct the agent reward function that considers the penalty mechanism:

[0129] When constructing the reward function, the maximum settlement reduction of the dam body is used as the primary positive reward, while geometric constraints, stability constraints, deformation constraints, seepage constraints, and stress constraints are also considered. Based on the feedback state (feasible or infeasible) of the finite element environment, a piecewise function is used to comprehensively evaluate the immediate reward of the agent's current action. The specific formula can be expressed as:

[0130]

[0131] Its meaning is: at any decision step For the environment state vector If all constraints are satisfied, then the environment state vector If a given constraint is not satisfied, the property is located in the feasible region; otherwise, if any constraint is not satisfied, the property is located in the infeasible region.

[0132] In the decision-making step The DQN agent selects from the action space according to the action policy. Select Action It drives the finite element modeling and evaluation model to output the finite element environmental state, and obtains the decision steps. Environment state vector Identify environmental state vectors Located in the feasible or infeasible region, and based on the environment state vector Compared to the environment state vector Calculate immediate rewards based on state changes. ;

[0133] if If all environmental state vectors lie within the feasible region, then the multi-objective comprehensive score under the feasible state will be calculated. As an instant reward Multi-objective comprehensive score under feasible conditions The calculation formula is:

[0134]

[0135] in: For the current decision-making steps The change in the maximum settlement characteristic value of the dam body compared to the maximum settlement characteristic value of the dam body under the initial dam body cross-section structural design scheme; Indicates the current decision step The maximum settlement characteristic value of the dam body compared to the previous decision step -1 represents the change in the maximum settlement characteristic value of the dam body; This indicates the maximum allowable hydraulic gradient of the core wall. and current decision-making steps Maximum hydraulic gradient eigenvalue of the core wall The difference; Indicates the current decision step Characteristic value of dam slope stability safety factor Minimum allowable safety factor for dam slope stability The difference; Indicates the current decision step Evaluation items for the geometry or quantity of the core wall; Indicates the current decision step The hydraulic fracturing risk assessment item is based on the maximum effective minor principal stress of the core wall. Sure; These are the weight coefficients for the corresponding items.

[0136] if If the environment state vector moves from the infeasible region to the feasible region, then the formula is used. Calculate instant rewards ;in, To guide DQN agents to successfully cross constraint boundaries and enter feasible regions from infeasible regions, rewards are awarded. This is the corresponding score scaling factor;

[0137] if If the environment state vector moves from the feasible region to the infeasible region, then the formula is used. Calculate instant rewards ; A fixed boundary violation penalty constant is used when the dam cross-sectional structure design scheme falls from the feasible region into the infeasible region.

[0138] if If the environment state vector enters the infeasible region from the infeasible region, then the formula is used. Calculate instant rewards ; The basic penalty term set for DQN agents when exploring infeasible domains; This is a dynamic penalty function; if the current action causes the sum of the penalty terms for breach of contract to increase... If the amount is reduced, the penalty will be reduced; if the total amount of penalties for breach of contract is reduced... If the penalty is increased, the DQN agent will be dynamically guided to move closer to the feasible domain boundary.

[0139] Step C4, the DQN agent is based on the action space The environment state vector and reward function interact and iterate with the finite element environment state to optimize the action strategy, resulting in a trained DQN agent.

[0140] DQN network training and parameter update mechanism: The agent outputs actions based on the current state to obtain new geometric parameters of the dam cross section, and inputs them into the finite element modeling evaluation model. The finite element modeling evaluation model recalculates to obtain the new state and reward. By continuously interacting with the finite element environment, the policy network parameters are updated until the model converges, and the optimal policy for the core wall dam structure design is obtained.

[0141] As one specific implementation, the DQN agent uses an improved deep Q-network for policy optimization; the improved deep Q-network includes a state input layer, a shared feature layer, a state value branch, an action advantage branch, and a fusion layer;

[0142] The state input layer is used to input the environment state vector. ;

[0143] The shared feature layer, for the environmental state vector A unified nonlinear mapping is performed to obtain shared hidden features. Shared hidden features Input the state value branch and the action advantage branch respectively;

[0144] The state value branch is based on the shared hidden features. Obtain the state value function ;

[0145] The action advantage branch, based on shared hidden features Obtain the advantage function ;

[0146] The fusion layer, for the state value function and dominance function By fusing the components, we obtain the Q-value:

[0147]

[0148] in: For decision-making steps At that time, in the environment state vector Next action Action value function; The state value function is only related to the current environment state vector. Related; The dominance function represents the current action. The relative merits of average value relative to state; Environment state vector The set of valid actions under the mask; To meet Each mask is a valid action; For action The relative merits of average value relative to state; These are the network parameters for the action value function, state value branch, and action advantage branch, respectively.

[0149] As one implementation method, the DQN agent uses the environment state vector As input, the Q-values ​​corresponding to each discrete action are used as output. During training, the DQN agent adjusts its state based on the current environment state vector. Select Action It automatically modifies the geometric parameters of the core wall, calls the finite element modeling and evaluation model to remodel and calculate, and obtains a new environmental state vector. and rewards Furthermore, the parameters of the deep Q-network are updated and improved by utilizing experience replay and target network mechanisms.

[0150] During training, the experience playback pool simultaneously stores the complete deformation response field of the finite element method. And to monitor the finite element complete deformation response field after assimilation. Among them, the finite element complete deformation response field... The a priori deformation response field of the finite element method; monitoring the assimilated complete deformation response field of the finite element method. This represents the finite element a posteriori deformation response field. When there is a significant difference between the finite element prior evaluation and the finite element a posteriori evaluation for the same design action, the empirical sample is preferentially extracted to train the DQN policy network to improve the agent's ability to identify the differences between finite element system biases and actual field responses.

[0151] During training, different core wall dam structural design cases can be selected from the sample database as the initial state. The initial structural design parameters can also be disturbed within a reasonable range to enhance the model's adaptability to different dam heights, material parameters, foundation conditions, and water level conditions.

[0152] Training stops when the number of training rounds reaches a preset value, the reward function stabilizes, the improvement of the objective function is less than a set threshold, or no better feasible structural design appears after several consecutive steps, thus obtaining the DQN policy network for optimizing the two-dimensional cross-sectional structure design of the core wall dam.

[0153] Step S3: Using the trained DQN agent, the two-dimensional cross-sectional structure optimization design of the dam body is carried out based on the dam body cross-sectional structure design scheme given in the new project of the heart wall dam.

[0154] Initial state mapping for new projects: Extract the initial geometric and material parameters of the two-dimensional cross-sectional structure design of the new core wall dam project to be optimized, obtain the finite element environment state by calling the finite element modeling evaluation model, and map it into the initial intelligent agent environment state vector;

[0155] Fast action output based on optimal policy: The current agent environment state vector is input into the trained DQN policy network. The DQN policy network directly calculates the Q value of each optional action through forward propagation based on the learned global optimal policy, and outputs the optimal adjustment action with the maximum Q value using a greedy policy.

[0156] Automated iteration and convergence determination: The optimal action output by DQN is automatically updated to update the geometric parameters of the core wall, driving the finite element modeling evaluation model to recalculate and update the agent's environmental state vector, and the state update and action reasoning process is executed cyclically.

[0157] Optimization output: When the structural design parameters remain unchanged for multiple consecutive steps or the preset convergence condition is reached, the iteration is terminated. The final two-dimensional cross-sectional structural design scheme of the core wall dam that satisfies the engineering constraints and minimizes the dam body settlement is output, along with the corresponding finite element calculation results. For example, the optimized core wall top width, the upstream and downstream slope ratio of the core wall, the core wall boundary coordinates, the two-dimensional cross-sectional partition diagram, the finite element calculation results, and the optimization iteration process record.

[0158] This invention provides an intelligent learning and reasoning optimization method for two-dimensional cross-sectional structures of core-wall dams. Using the GYY core-wall rockfill dam structural design project as an example, the implementation of this invention is illustrated. The complete intelligent optimization process is as follows: Figure 1 As shown.

[0159] S101: Constructing the Heart-Wall Rockfill Dam Dataset:

[0160] Obtain the geometric parameters, material physical and mechanical parameters, and water level parameters of existing core-wall rockfill dam projects. Establish a sample database of core-wall rockfill dams, storing two-dimensional typical cross-sections of core-wall rockfill dam projects in DXF file format, and storing the geometric parameters, material physical and mechanical parameters, and water level parameters in JSON file format.

[0161] S102: Establish a fully automated modeling and mechanical index simulation module based on finite element software.

[0162] Based on the Python secondary development API of finite element method (FEM) software, a fully automated modeling and result extraction script is constructed. After reading the corresponding DXF and JSON files, the script automatically generates the geometric boundaries of the core wall region, filter layer, transition layer, upstream and downstream rockfill zones, and foundation region within the FEM software. Figure 2 The diagram shows a two-dimensional parametric structural diagram of the core-wall dam. Subsequently, based on the material database, constitutive models (such as hardened soil models or Mohr-Coulomb models) and permeability and strength parameters are automatically assigned to each zone. Finite element meshes are automatically generated, and construction filling, water storage, and steady-state seepage conditions are set and submitted for calculation. After the simulation is completed, the script automatically extracts the maximum settlement of the dam body, the slope stability safety factor, the effective minor principal stress of the core wall, and the maximum hydraulic gradient of the core wall under each condition.

[0163] S103: Establishing Constraints and Objectives for Optimizing the Core Wall Region

[0164] (1) Optimization variables: top width of the core wall, upstream and downstream slope ratio of the core wall;

[0165] (2) Optimization Constraints: Establish a multi-level engineering constraint system. Geometric constraints require that the core wall, filter layer, and transition layer maintain continuous formation and must not cross the external boundary of the dam body. The thickness of each zone must not be less than the limit specified in the code, and the area of ​​the core wall must not be too small and must meet the seepage control requirements. Stability constraints require that the dam slope stability safety factor calculated based on the strength reduction method is greater than the allowable value specified in the code. Deformation constraints require that the maximum settlement of the dam body is less than the allowable value specified in the code. Seepage constraints require that the maximum hydraulic gradient of the core wall is less than the allowable hydraulic gradient. Stress constraints require that no hydraulic splitting occurs inside the core wall. If any constraint is not met, the structural design scheme is deemed infeasible.

[0166] (3) Optimization objective: The core objective is to reduce the maximum settlement of the dam body. The objective function is constructed by combining geometric constraints, stability constraints, deformation constraints, seepage constraints and stress constraints, and using a weighted summation mechanism.

[0167] S104: Establishing an iterative solution and training framework for optimization strategies based on the DQN deep reinforcement learning algorithm.

[0168] Constructing an interaction framework between DQN agents and the finite element environment, such as... Figure 3 As shown:

[0169] (1) State space: Extract the feature vector of the geometric parameters of the core wall in the current iteration step, and combine it with the maximum settlement value, safety factor value, hydraulic gradient value of the core wall and constraint satisfaction state fed back by the finite element software. After normalization, the vector is spliced ​​into a state vector and input into the DQN neural network.

[0170] (2) Action space: The adjustment step size of the top width of the core wall and the adjustment range of the slope ratio of the core wall are discretized into a finite set of action instructions.

[0171] (3) Reward function: If the finite element software shows that the maximum settlement of the dam body decreases and all engineering constraints are met after the agent performs a certain action, a positive reward is given; if the settlement increases, a smaller penalty is given; if the geometric or mechanical constraints are violated, a strong negative penalty is given.

[0172] (4) Network Training: The agent outputs an action based on the current state, automatically reconstructs the two-dimensional cross-sectional model of the heart wall dam, and re-acquires the new state and reward, storing the experience in the experience replay pool. Gradient descent is used to update the Q-network parameters by continuously extracting experience. As the number of iterations increases, the reward function and loss function gradually converge, such as... Figure 4 The figure shows the convergence curves of the reward function and loss function of the DQN deep reinforcement learning algorithm during the training process; the DQN model learns the globally optimal strategy for adjusting the heart wall dam parameters.

[0173] S105: A Heart-Bridge Dam Structure Design Optimization Method Based on Post-Training Deep Reinforcement Learning

[0174] For the GYY core wall dam project to be optimized, the initial structural design parameters are input, with a core wall top width of 6.0m and an upstream-downstream slope ratio of 0.20. The DQN algorithm achieves efficient global optimization of structural design parameters.

[0175] Structural optimization design results as follows Figure 5As shown, this is a two-dimensional cross-section of the core wall resulting from the structural optimization design of a typical case. The red area represents the core wall region after optimization using the DQN algorithm, while the blue area represents the core wall region of the original actual structural design. Under the condition of satisfying the optimization constraints, the optimized core wall region has a top width of 3.6m, an upstream-downstream slope ratio of 0.20, and the core wall area is reduced by 11.65% compared to before optimization. The maximum settlement of the dam body is improved by 3.40% compared to before optimization.

[0176] The beneficial effects of this invention are as follows: By constructing a parameterized sample database and an automated finite element simulation process for core-wall rockfill dams, effective integration between structural design data and dam mechanical response simulation is achieved; by establishing an optimization model with a multi-level constraint system, the standardization and systematic nature of the core-wall dam structural design optimization process are improved; by introducing the DQN deep reinforcement learning algorithm, automatic iterative updates and strategy learning of core-wall dam structural design parameters in the finite element environment are realized, effectively reducing the workload of manual trial calculations, shortening the structural design cycle, and improving optimization efficiency; by applying the trained DQN agent to the design of new engineering structures, rapid optimization of new core-wall dam schemes is achieved. Through the on-site deformation monitoring data collected by settlement monitoring sensors deployed in the core-wall dam on-site monitoring space, the mechanical deformation response field of the dam body is corrected, improving the accuracy of the dam body's mechanical deformation response field, thereby improving the accuracy of DQN agent training and increasing the optimization efficiency of the two-dimensional cross-sectional structural parameters of the core-wall dam. This achieves automation, intelligence, and efficiency in the structural design of core-wall rockfill dams, providing important technical support for intelligent structural design in water conservancy projects, and demonstrating significant practicality and broad application prospects.

[0177] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent learning and reasoning optimization of two-dimensional cross-sectional structures of a core-wall dam, characterized in that, Includes the following steps: Step S1: Based on the multi-source engineering data acquisition system for core-wall rockfill dams, construct a set of dam cross-sectional structural design schemes; each dam cross-sectional structural design scheme includes dam cross-sectional geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data. Step S2: Construct the interaction framework between the DQN agent and the finite element environment, including the finite element modeling and evaluation model and the DQN agent; The finite element modeling and evaluation model takes the dam cross-sectional structure design scheme as input, and obtains a finite element mechanical response state that fully reflects the characteristics of the finite element mechanical response through finite element modeling, dam mechanical response simulation, dam mechanical response monitoring assimilation correction based on sparse on-site deformation monitoring data, and dam mechanical response time series prediction. Combining the constraint judgment results and the finite element mechanical response state, the finite element environment state is obtained as the output. The DQN agent discretizes the adjustment range of the core wall geometric parameters into the agent's action space, constructs an agent environment state vector based on the finite element environment state, and constructs an agent reward function with the goal of minimizing the dam deformation response characteristic value and combining the constraint judgment results. The DQN agent optimizes its action strategy through interactive iteration with the finite element environment state to obtain a trained DQN agent. Step S3: Using the trained DQN agent, the two-dimensional cross-sectional structure optimization design of the dam body is carried out based on the dam body cross-sectional structure design scheme given in the new project of the heart wall dam.

2. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 1, characterized in that, The structural design scheme of each dam section is characterized as follows: ; These include the dam body cross-sectional geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data. The dam body cross-sectional geometric parameters include the core wall geometric parameters and other material zoning geometric parameters. The engineering load conditions include the construction loading path and the water level change path. The water level change path includes the normal storage water level, design flood level, check flood level, construction period water level, stable seepage condition water level, and rapid precipitation condition water level.

3. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 1, characterized in that, The finite element modeling and evaluation model yields the following finite element mechanical response states: Step A1, the finite element modeling evaluation model includes a core wall dam finite element modeling model, a dam mechanical response simulation model, a pre-trained core wall dam deformation field time series prediction model, and a feature extraction model; Step A2, for the input dam cross-sectional structure design scheme A finite element model of a core-wall dam was adopted, based on the geometric parameters of the dam cross-section. Perform parametric modeling and partition features according to material. Automatically assign constitutive parameters, permeability parameters, and strength parameters to each material zone of the cross-section, and obtain the corresponding finite element model of the core wall dam. ; Step A3: Using the dam body mechanical response simulation model, the finite element model of the core wall dam is... Loading engineering load conditions Finite element response simulation was performed to obtain the complete finite element response field. Including the finite element complete deformation response field Finite element complete stable response field Finite element complete seepage response field and finite element complete stress response field ; Step A4: A pre-trained time-series prediction model for the deformation field of the core wall dam is used, based on sparse on-site deformation monitoring data. Response field of complete deformation of finite element The model of the core wall dam is realized by making corrections in the latent space and loading various engineering load conditions. Comprehensive simulation analysis under multiple working conditions was conducted to extract the maximum settlement characteristic value of the dam body, which reflects the deformation response characteristics of the dam body. ; Step A5: Using a feature extraction model, the finite element complete stable response field is analyzed. Finite element complete seepage response field and finite element complete stress response field Feature extraction is performed to obtain the characteristic value of the dam slope stability safety factor, which reflects the stability response characteristics of the dam body. The maximum hydraulic gradient characteristic value of the core wall, reflecting the seepage response characteristics of the dam body. And the maximum effective minor principal stress of the core wall, which reflects the stress response characteristics of the dam body. ; Step A6: Generate the finite element mechanical response state .

4. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 3, characterized in that, Step A4 includes: Step A41, the pre-trained heart wall dam deformation field time series prediction model includes an observation operator, a pre-trained shared VAE encoder, a pre-trained shared VAE decoder, a monitoring encoder, a latent variable updater, and an autoregressive propagation model; Step A42, Initial Correction Stage: Response Field to Complete Deformation of Finite Element By performing monitoring and assimilation correction, the complete deformation response field of the finite element method after monitoring and assimilation is obtained. ,include: Step A421, Sparse Field Deformation Monitoring Data Including field monitoring data matrix in sparse field monitoring space Based on the observation operator, the complete deformation response field of the finite element is obtained. Mapping to a sparse field monitoring space to obtain a matrix of field monitoring data. Comparable prior monitoring prediction matrix And based on the field monitoring data matrix With the prior monitoring prediction matrix Construct monitoring residual matrix based on the difference between them The monitoring residual matrix is ​​processed by a pre-trained monitoring encoder. Feature extraction is performed to obtain sparse monitoring and correction features. ; Step A422: Use a pre-trained shared VAE encoder to input the complete deformation response field of the finite element method. Encoded as prior latent variables in a low-dimensional latent space ; using a pre-trained latent variable updater to perform correction based on the sparse monitoring features in a low-dimensional latent space. Correcting the prior latent variables , obtain the posterior latent variables ; Step A423, the posterior latent variable The input is fed into a pre-trained shared VAE decoder, which converts the posterior latent variable... Reconstructed into a complete finite element deformation response field after monitoring assimilation ; The monitored assimilated finite element complete deformation response field That is, the response field to the complete deformation of the finite element. The correction results; Step A43, Time Series Prediction Stage: Apply different engineering load conditions to perform time series prediction, and obtain the finite element complete deformation response field prediction sequence, including: Based on the dam body cross-sectional structure design scheme Engineering load conditions Based on this, the engineering load conditions and sparse on-site deformation monitoring data of the same dam body in the next engineering stage are obtained by searching the dam body cross-sectional structural design scheme set; the assimilated finite element complete deformation response field is then used. The engineering load conditions for the next engineering stage are input into the autoregressive propagation model to predict the finite element complete deformation response field for the next engineering stage. Then, the monitoring assimilation correction method in step A42 is used to obtain the monitored and assimilated finite element complete deformation response field for the next engineering stage. This process is repeated continuously to achieve rolling prediction and obtain the finite element complete deformation response field prediction sequence. Step A44, Feature Extraction: The finite element complete deformation response field after monitoring and assimilation from the initial correction stage is extracted. Feature extraction was performed to obtain the maximum settlement characteristic value of the dam body in the initial correction stage. Extract the monitoring residual matrix from the initial correction stage. Feature extraction was performed on the finite element complete deformation response field prediction sequence to obtain the maximum settlement characteristic value of the dam body in the time-series prediction stage. ; Step A45, Feature Fusion: For the maximum settlement characteristic values ​​of the dam body in the initial correction stage... Monitoring residual matrix during the initial correction phase and the characteristic value of maximum settlement of the dam body during the time-series prediction stage. Feature fusion is performed, and prediction uncertainties are considered to obtain the maximum settlement characteristic value of the dam body, which reflects the deformation response characteristics of the dam body. .

5. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 4, characterized in that, When pre-training a shared variational autoencoder composed of a shared VAE encoder and a shared VAE decoder, the training method is as follows: During training, for finite element complete deformation response field samples The shared VAE encoder will sample the complete deformation response field of the finite element method. Mapped to latent variables in a low-dimensional latent space The latent variable distribution parameters, including the mean and standard deviation The expression is , Indicates a shared VAE encoder; based on the mean. and standard deviation The latent variables are obtained by reparameterization. The expression is: ;in, This represents a random perturbation vector sampled from a standard multivariate normal distribution. This represents a standard multivariate normal distribution with a mean of 0 and a covariance matrix of identity matrix I, where I represents the identity matrix. This represents element-wise multiplication; During training, the shared VAE decoder will store latent variables in the low-dimensional latent space. Reconstruction yields the reconstructed finite element complete deformation response field. The expression is , Indicates a shared VAE decoder; The training loss of the shared variational autoencoder Represented as: ； ； in: The reconstruction loss of the complete deformation response field in the finite element method; Latent space regularization constraint; Indicates the regularization weight; Indicates KL divergence; This represents the finite element complete deformation response field generated by the shared VAE encoder based on the input. latent variables obtained The approximate posterior distribution, This represents the pre-defined prior distribution of the latent variables.

6. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 1, characterized in that, The finite element environment state, obtained by combining the constraint determination results and the finite element mechanical response state, is specifically as follows: Step B1: Establish a multi-level constraint system for the core wall dam, including dam cross-sectional geometric constraints, dam deformation constraints, dam slope stability constraints, dam seepage constraints, and dam stress constraints. Step B2: Determine whether the dam cross-sectional structure design scheme conforms to the geometric constraints of the dam cross-section and the degree of deviation from the relative constraint boundary; Determine the maximum settlement characteristic value of the dam body respectively Characteristic value of dam slope stability safety factor eigenvalues ​​of the maximum hydraulic gradient of the core wall Maximum effective minor principal stress of the core wall The constraint determination result is obtained by determining whether the dam body deformation constraint, dam slope stability constraint, dam body seepage constraint, and dam body stress constraint are met, as well as the degree of deviation from the relative constraint boundary. Step B3, the constraint determination result and the finite element mechanical response state By splicing the data, we obtain the finite element environment state as the output.

7. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 1, characterized in that, The DQN agent optimizes its action strategy through interactive iteration with the finite element environment state, resulting in a trained DQN agent, including: Step C1, Define the agent's action space The top width, upstream slope ratio, and downstream slope ratio of the core wall geometry parameters are used as the objects of action adjustment; the adjustment range of the top width, upstream slope ratio, and downstream slope ratio of the core wall is discretized into a finite set of actions to form the action space of the intelligent agent; the intelligent agent actions include: performing operations to increase, decrease, or remain unchanged the top width of the core wall by a set range, and performing operations to increase, decrease, or remain unchanged the upstream slope ratio of the core wall by a set range, and performing operations to increase, decrease, or remain unchanged the downstream slope ratio of the core wall by a set range. Step C2, construct the agent environment state vector: normalize the initial dam cross-section structure design scheme, dam characteristic parameters, the dam cross-section structure design scheme of the current decision step, and the finite element environment state of the current decision step, and combine them into the agent environment state vector of the current decision step. Step C3, construct the agent reward function: ； Its meaning is: at any decision step For the environment state vector If all constraints are satisfied, then the environment state vector If a given constraint is not satisfied, the property is located in the feasible region; otherwise, if any constraint is not satisfied, the property is located in the infeasible region. In the decision-making step The DQN agent selects from the action space according to the action policy. Select Action It drives the finite element modeling and evaluation model to output the finite element environmental state, and obtains the decision steps. Environment state vector Identify environmental state vectors Located in the feasible or infeasible region, and based on the environment state vector Compared to the environment state vector Calculate immediate rewards based on state changes. ; if If all environmental state vectors lie within the feasible region, then the multi-objective comprehensive score under the feasible state will be calculated. As an instant reward ; if If the environment state vector moves from the infeasible region to the feasible region, then the formula is used. Calculate instant rewards ;in, To guide DQN agents to successfully cross constraint boundaries and enter feasible regions from infeasible regions, rewards are awarded. This is the corresponding score scaling factor; if If the environment state vector moves from the feasible region to the infeasible region, then the formula is used. Calculate instant rewards ; A fixed boundary violation penalty constant is used when the dam cross-sectional structure design scheme falls from the feasible region into the infeasible region. if If the environment state vector enters the infeasible region from the infeasible region, then the formula is used. Calculate instant rewards ; The basic penalty term set for DQN agents when exploring infeasible domains; This is a dynamic penalty function; if the current action causes the sum of the penalty terms for breach of contract to increase... If the amount is reduced, the penalty will be reduced; if the total amount of penalties for breach of contract is reduced... If the penalty is increased, the DQN agent will be dynamically guided to move closer to the feasible domain boundary. Step C4, the DQN agent is based on the action space The environment state vector and reward function interact and iterate with the finite element environment state to optimize the action strategy, resulting in a trained DQN agent.

8. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 7, characterized in that, Multi-objective comprehensive score under feasible conditions The calculation formula is: ； in: For the current decision-making steps The change in the maximum settlement characteristic value of the dam body compared to the maximum settlement characteristic value of the dam body under the initial dam body cross-sectional structural design scheme; Indicates the current decision step The maximum settlement characteristic value of the dam body compared to the previous decision step -1 represents the change in the maximum settlement characteristic value of the dam body; This indicates the maximum allowable hydraulic gradient of the core wall. and current decision-making steps Maximum hydraulic gradient eigenvalue of the core wall The difference; Indicates the current decision step Characteristic value of dam slope stability safety factor Minimum allowable safety factor for dam slope stability The difference; Indicates the current decision step Evaluation items for the geometry or quantity of the core wall; Indicates the current decision step The hydraulic fracturing risk assessment item is based on the maximum effective minor principal stress of the core wall. Sure; These are the weight coefficients for the corresponding items.

9. The intelligent learning and reasoning optimization method for a two-dimensional cross-sectional structure of a core wall dam according to claim 1, characterized in that, The DQN agent uses an improved deep Q network for policy optimization; the improved deep Q network includes a state input layer, a shared feature layer, a state value branch, an action advantage branch, and a fusion layer. The state input layer is used to input the environment state vector. ; The shared feature layer, for the environmental state vector A unified nonlinear mapping is performed to obtain shared hidden features. Shared hidden features Input the state value branch and the action advantage branch respectively; The state value branch is based on the shared hidden features. Obtain the state value function ; The action advantage branch, based on shared hidden features Obtain the advantage function ; The fusion layer, for the state value function and dominance function By fusing the components, we obtain the Q-value: ； in: For decision-making steps At that time, in the environment state vector Next action Action value function; The state value function is only related to the current environment state vector. Related; The dominance function represents the current action. The relative merits of average value relative to state; Environment state vector The set of valid actions under the mask; To meet Each mask is a valid action; For action The relative merits of average value relative to state; These are the network parameters for the action value function, state value branch, and action advantage branch, respectively.

10. A smart learning and reasoning optimization system for two-dimensional cross-sectional structures of a core-wall dam, characterized in that, include: The dam body cross-section structure design scheme set construction module is used to construct a dam body cross-section structure design scheme set based on the multi-source engineering data acquisition system for core-wall rockfill dams. Each dam section structural design scheme includes dam section geometric parameters, material zoning characteristics, engineering load conditions, and sparse on-site deformation monitoring data; The interaction framework between DQN agents and the finite element environment includes the finite element modeling and evaluation model and the DQN agent. The finite element modeling and evaluation model takes the dam cross-sectional structure design scheme as input, and obtains a finite element mechanical response state that fully reflects the characteristics of the finite element mechanical response through finite element modeling, dam mechanical response simulation, dam mechanical response monitoring assimilation correction based on sparse on-site deformation monitoring data, and dam mechanical response time series prediction. Combining the constraint judgment results and the finite element mechanical response state, the finite element environment state is obtained as the output. The DQN agent discretizes the adjustment range of the core wall geometric parameters into the agent's action space, constructs an agent environment state vector based on the finite element environment state, and constructs an agent reward function with the goal of minimizing the dam deformation response characteristic value and combining the constraint judgment results. The DQN agent optimizes its action strategy through interactive iteration with the finite element environment state to obtain a trained DQN agent. The dam body two-dimensional cross-section optimization module is used to carry out the two-dimensional cross-section structural optimization design of the dam body based on the dam body cross-section structural design scheme given in the new core wall dam project, using a trained DQN agent.

Images