Multi-field coupling concrete dam parameter intelligent inversion method, system and device based on deep learning and medium

Abstract

The invention discloses a multi-field coupling concrete dam parameter intelligent inversion method, system and device based on deep learning and a medium, and belongs to the technical field of dam engineering safety monitoring and computational mechanics crossing. The geometry and boundary modeling unit is used for uniformly expressing a dam body-dam foundation geometric region and a boundary type; simultaneously representing a state field and a parameter field by utilizing PINN inversion, calculating a physical residual error, and implementing parameter physical feasibility constraint and prior regularization; the organization optimizer, residual drive sampling and weight self-adaption are used for training scheduling, and result review, visualization and data export are carried out through the verification and output unit. According to the method, end-to-end parameter field inversion is realized on the premise of not depending on frequent mesh generation, the calculation cost of multiple forward modeling is reduced, the non-uniqueness of inversion is inhibited, and the method has higher recognition capability for engineering concerned areas such as permeation channels and degraded zones.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of dam engineering safety monitoring and computational mechanics, specifically to a method, system, equipment, and medium for intelligent inversion of multi-field coupled concrete dam parameters based on deep learning. Background Technology

[0002] Several common methods exist for intelligent inversion of parameters in multi-field coupled concrete dams, each with its own technical limitations:

[0003] The typical process of traditional inversion based on finite element method (FEM) is as follows: a force-permeability-thermal coupled finite element forward model is used, and gradient or intelligent optimization algorithms (GA / PSO / DE, etc.) are used to iteratively search in the parameter space, with the observation and simulation deviation as the objective function for optimization.

[0004] However, the shortcomings of traditional inversion are: (1) high computational cost of forward modeling (multiple meshing, assembly and solution), and slow convergence in multi-parameter, high-dimensional space; (2) obvious ill-conditioning / non-uniqueness, requiring a large amount of regularization, prior knowledge and human experience; (3) strong coupling and nonlinear boundaries (such as temperature-seepage coupling convection terms, contact / interface permeability changes) introduce more computational burden; (4) limited ability to reconstruct spatially heterogeneous parameter fields (such as zonal deterioration, dam heel seepage channels).

[0005] Bayesian / collective homogenization methods (MCMC, EnKF / ES-MDA, etc.): can provide uncertainty assessment, but the sampling cost is huge and highly dependent on forward modeling efficiency; when the coupling mechanism is complex, the monitoring is sparse, and the prior is incomplete, the convergence and reliability are limited.

[0006] Pure data-driven machine learning: It uses monitoring data to regress parameters / states, which is difficult to generalize and difficult to ensure physical consistency. It is sensitive to abnormal data and difficult to handle situations outside of normal operating conditions. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, this invention aims to construct an inversion framework that deeply integrates multi-field coupled physical laws with measured monitoring data without the need for frequent meshing and assembly. This framework reduces forward modeling costs, suppresses inversion ill-conditions, fully incorporates thermal-permeability-mechanical coupling, and accurately identifies the spatially heterogeneous parameter fields of the dam body and foundation under sparse monitoring conditions, thereby providing a reliable parameter basis for the construction and condition assessment of digital twins.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based intelligent inversion method for multi-field coupled concrete dam parameters, comprising,

[0010] The system accesses multi-field coupled concrete dam data through the data access and cleaning unit, and the geometry and boundary modeling unit uniformly represents the geometric region and boundary type of the dam body-dam foundation. The system uses PINN inversion to simultaneously represent the state field and parameter field and calculate the physical residuals, and implements parameter physical feasibility constraints and prior regularization. The system organizes optimizers, residual-driven sampling and weight adaptive training scheduling, and the verification and output units perform result verification, visualization and data export.

[0011] As a preferred embodiment of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described in this invention, the multi-field coupled concrete dam parameters include monitoring data of the concrete dam that are accessed and managed, including first monitoring data, second monitoring data, third monitoring data, fourth monitoring data and fifth monitoring data;

[0012] The monitoring data were standardized in units and aligned in time, and the sensor positions were calibrated in the same coordinate system.

[0013] As a preferred embodiment of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described in this invention, the unified expression includes: modeling the geometric region and boundary type of the dam body-dam foundation based on monitoring data, and uniformly encoding the computational domain and boundary using a signed distance function;

[0014] A general form is used to describe the heat transfer relationship between the reservoir water and the air side to ensure the physical consistency of the engineering boundary conditions.

[0015] As a preferred embodiment of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described in this invention, the PINN inversion includes: constructing a physical information neural network, including a state field network and a parameter field network, and using a multilayer perceptron to simultaneously represent the state field and the parameter field and share the backbone layer.

[0016] The state field network outputs displacement, pore water pressure, and temperature;

[0017] The parameter field network outputs spatially heterogeneous material parameters and percolation thermal parameters, and ensures the positive definiteness and range constraints of the parameters through function transformation.

[0018] As a preferred embodiment of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described in this invention, the implementation parameter physical feasibility constraints and prior regularization include constructing a loss function, including multi-field strong form residuals, residuals of boundary conditions and initial conditions, residuals of monitoring data, and prior regularization terms, and approximating volume integrals and boundary integrals through sampling points.

[0019] Based on the state field network and parameter field network inverted by PINN, multi-field strength form residuals are defined, and monitoring data residuals and prior regularization terms are further introduced;

[0020] The multi-field residuals are calculated as follows:

[0021]

[0022] Calculate the deviation between the boundary conditions and the initial conditions:

[0023]

[0024] The residual of the monitoring data is measured as follows:

[0025]

[0026] Parameter priors and regularization using H 1 The combined form of smoothing, total variation, and box constraints is represented as follows:

[0027]

[0028] All values ​​are approximated by volume integrals at the sampling points and boundary integrals, which together constitute the joint loss:

[0029]

[0030] Where, r mech For momentum balance residuals, Let σ be the divergence of the stress tensor. Let ρ be the density, b be the force vector, and r be the force term. hyd For the seepage mass conservation residual, S s For the specific water storage coefficient field, p t Let α be the time derivative of the pore water pressure at time t. B For the Biot coefficient field, ε v,t For the volumetric strain at time t, To take the divergence of the vector field within the brackets, where k is the permeability field and μ is the dynamic viscosity, Let q be the density of water, g be the acceleration vector due to gravity, and q be the velocity vector. p For seepage source and sink terms, r therm For temperature residual, This is the equivalent volumetric heat capacity term for the medium. The term represents the heat-carrying convection term of seepage, where v is the Darcy velocity. Let κ be the divergence of the thermal conductivity term, κ be the equivalent thermal conductivity field, and q be the thermal conductivity term. T For body heat source / sink; Let u be the boundary / initial residual of displacement, pore water pressure, and temperature, and u be the displacement field. D Γ is the displacement boundary value given at the displacement Dirichlet boundary.u The boundary subset to which the displacement Dirichlet condition is applied; p is the pore water pressure, p D For the pressure boundary value given at the pressure Dirichlet boundary, Γ p For the boundary subset of the applied pressure Dirichlet condition, T is the temperature. D Γ is the temperature boundary value given at the temperature Dirichlet boundary. T For the boundary subset to which the temperature Dirichlet condition is applied; r data To monitor the residual vector of the data, For the observation operator, y obs Let R be the observed data vector; R is the joint prior and regularization term. λ TV , λ box These are the non-negative weight coefficients corresponding to the three types of regularization; Ω is the computational domain. Let θ be the gradient of the parameter field; ||·|| is the Euclidean / Frobenius norm, Φ box (θ) is the box constraint penalty function; For the total loss, w m w h w t w b w d r represents the five non-negative weights corresponding to different residuals. Γ Let ||r| be the boundary / initial value residual vector. Γ || 2 Denotes its square norm; <·> r 、<·> S 、<·> Γ Let S represent the sampling mean of the volume integral, boundary integral, and data error using the in-domain sampling points, boundary sampling points, and data sample set S, respectively.

[0031] The beneficial effects of the preferred technical solution in the embodiments of the present invention are as follows: by constructing a loss function that includes multi-field physical residuals, boundary conditions, monitoring data and prior regularization, the physical feasibility and numerical stability of the inversion parameters are ensured, and the occurrence of non-physical solutions is avoided.

[0032] As a preferred embodiment of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described in this invention, the training schedule includes: automatically differentiating and calculating the strong form residuals, boundary / initial value residuals and monitoring data residuals of the multi-field equations, and performing joint optimization under the constraint of the regularization term; during optimization, using an adaptive gradient method based on mini-batch to find the best value at a larger learning rate, and using a quasi-Newton method to refine the parameters at a smaller learning rate.

[0033] Adaptive adjustment is performed based on the gradient norm, which involves calculating the ratio of the gradient norm of all tasks to the average gradient norm and updating the weight coefficients w based on the current ratio. i :

[0034]

[0035] in, Let G be the weight of the i-th task in the k+1-th iteration. i Let be the gradient norm of the i-th task under the current weights; Let be the average of the gradient norms for each task, and α be the adjustment exponent. Let L be the gradient of the network parameter set Θ; ||·|| is the gradient vector L. 2 Norm, Let θ be the i-th sub-loss, and Θ be all trainable parameters;

[0036] After every few rounds of training, the sample points are enriched and migrated according to the residual distribution, and adaptive sampling is performed on high residual regions, with the ratio of in-domain sampling to boundary sampling maintained in the range of 20% to 30%.

[0037] The training process terminates when both the verification point error and the physical residual reach the target. The resulting parameter field and state field can then be exported from the verification and output unit to a standard engineering format for verification and display.

[0038] The preferred technical solution in the embodiments of the present invention has the following advantages: by outputting results in multiple formats and performing cross-validation, it provides reliable inversion results for engineering applications and supports iterative improvement and visualization analysis.

[0039] As a preferred embodiment of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described in this invention, the verification and output unit includes outputting the inverted parameter field and state field, and exporting them in a specified format for engineering verification and visualization.

[0040] After training convergence, the output includes contour lines and slices of the spatially heterogeneous parameter field, fitted curves of key monitoring points, and residual statistics.

[0041] The parameter field and state field are exported as VTK, GeoTIFF or comma-separated format, displayed in visualization software, and cross-validated with the external finite element model. If there is a systematic deviation between the external finite element model and the system in a local area, the deviation area is taken as the focus of a new round of sampling for iterative improvement.

[0042] The preferred technical solution in the embodiments of the present invention has the following advantages: by outputting results in multiple formats and performing cross-validation, it provides reliable inversion results for engineering applications and supports iterative improvement and visualization analysis.

[0043] Another objective of this invention is to provide an intelligent inversion system for multi-field coupled concrete dam parameters based on deep learning.

[0044] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a deep learning-based intelligent inversion system for multi-field coupled concrete dam parameters, comprising: a data access and cleaning unit, a geometry and boundary modeling unit, a PINN inversion solution unit, a constraint and regularization unit, a training and scheduling unit, and a verification and output unit;

[0045] The data access and cleaning unit is used to access and manage temperature, seepage pressure, displacement, reservoir water level, and meteorological records.

[0046] The geometry and boundary modeling unit is used to uniformly represent the geometric region of the dam body-dam foundation and its boundary types;

[0047] The PINN inversion solution unit is used to simultaneously represent the state field and parameter field and calculate the physical residual;

[0048] The constraint and regularization unit is used to implement parametric physical feasibility constraints and prior regularization;

[0049] The training scheduling unit is used to organize the optimizer, residual-driven sampling, and weight adaptation.

[0050] The verification and output unit is used for result verification, visualization, and data export.

[0051] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning.

[0052] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning.

[0053] The beneficial effects of this invention are as follows: By directly embedding the control equations of the heat-seepage-mechanical coupling into a neural representation with strong residuals, and jointly minimizing them with the consistency requirements of monitoring data in the same loss function, this invention achieves end-to-end parameter field inversion without relying on frequent mesh partitioning. The above design significantly reduces the computational cost of multiple forward models and suppresses the non-uniqueness of the inversion through hard-constraint boundary encoding and prior regularization. Furthermore, the neural representation of the parameter field allows for the continuous or partitioned characterization of the spatial heterogeneous distribution of the dam body and foundation, thereby providing a higher identification capability for engineering areas of interest such as seepage channels and deterioration zones.

[0054] The key technology of this invention lies in using the strong-form residuals of multi-field coupling equations as the core, and unifying momentum balance and coupled constitutive model, seepage mass conservation and Darcy's law, heat conduction including convection terms, and the consistency of boundary / initial conditions and monitoring data into a single joint loss. The parameter field adopts a neural representation with positive definiteness and range constraints, and the multi-task weights are adaptively adjusted through gradient norm, while the hard constraints of Dirichlet conditions are achieved using distance functions. The above key points form the core defining features of the system-independent and method-independent claims, and the partitioned-continuous hybrid parameter expression and residual-driven adaptive sampling are preferred additional features. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 The above is a flowchart of an intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning, provided in one embodiment of the present invention. Detailed Implementation

[0057] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0058] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a deep learning-based intelligent inversion method for multi-field coupled concrete dam parameters, including:

[0059] S100: The multi-field coupled concrete dam data is accessed through the data access and cleaning unit, and the geometry and boundary modeling unit uniformly represents the geometric region and boundary type of the dam body-dam foundation.

[0060] S200: Utilize PINN inversion to simultaneously represent the state field and parameter field and calculate the physical residuals, and implement parametric physical feasibility constraints and prior regularization;

[0061] S300, organization optimizer, residual-driven sampling and weight adaptation are used for training and scheduling, and the results are verified, visualized and exported through the validation and output units.

[0062] It should be noted that existing techniques typically rely on a separate, step-by-step inversion strategy, such as first inverting thermal parameters based on the temperature field and then inverting mechanical parameters based on the displacement field. This method artificially severs the inherent coupling effects between multiple physics fields, leading to physical inconsistencies in the inversion results. Furthermore, traditional methods heavily depend on grid discretization and iterative forward calculations, resulting in high computational costs and difficulty in effectively incorporating complex prior knowledge. They often get trapped in local optima, limiting the reliability and engineering applicability of the inversion results.

[0063] Therefore, to address the aforementioned problems, through steps S100-S300, this invention, with Physical Information Neural Network (PINN) as its core, provides a system and method for intelligent inversion of parameter fields (elastic modulus, two-way Poisson's ratio or equivalent parameters, two-way / anisotropic permeability coefficient, thermal conductivity, specific heat capacity, Biot coefficient, specific water storage coefficient, etc.) of concrete gravity dams / arch dams and their foundations. This invention achieves end-to-end identification of spatially heterogeneous parameter fields by jointly optimizing the physical residuals embedded in the control equations and the fitting residuals of monitoring data, without requiring explicit mesh generation.

[0064] Example 2, refer to Figure 1 This is one embodiment of the present invention, which provides a deep learning-based intelligent inversion method for multi-field coupled concrete dam parameters, including:

[0065] In this embodiment of the invention, in S100, multi-field coupled concrete dam data is accessed through the data access and cleaning unit, and the geometry and boundary modeling unit uniformly represents the geometric region and boundary type of the dam body-dam foundation, including the following steps S101-S102:

[0066] S101. Access and manage the monitoring data of the concrete dam, including the first monitoring data, the second monitoring data, the third monitoring data, the fourth monitoring data, and the fifth monitoring data;

[0067] The first monitoring data is temperature, the second is seepage pressure, the third is displacement, the fourth is reservoir water level, and the fifth is meteorological records.

[0068] In the embodiments of the present invention, the monitoring data are standardized in units and aligned in time, and the sensor positions are calibrated in the same coordinate system.

[0069] In an optional implementation, the monitoring data processing in S101 can be based on the unit unification and time alignment of the pre-calibrated location, access the monitoring data of the multi-field coupled concrete dam, automatically identify the data unit and convert it into a unified standard unit, interpolate or resample the timestamp so that all data sequences have the same time interval and reference, and use the coordinate information pre-stored in the system to directly apply the sensor position to the geometric modeling.

[0070] In another optional implementation, the monitoring data processing in S101 can also be based on the unification of units and time alignment of the relative coordinate system. After the monitoring data is accessed, the data units are manually or semi-automatically unified to commonly used units. The data is synchronized to a common time axis through a time series alignment tool, missing values ​​are filled, and the relative position coordinates of the sensors are calibrated with the key points of the dam body (such as the dam crest or the center of the dam foundation) as the reference origin, instead of the absolute coordinate system.

[0071] In an embodiment of the present invention, S102, based on monitoring data, the geometric region and boundary type of the dam body-dam foundation are modeled. During the geometric and boundary modeling process, the computational domain and boundary are uniformly encoded using a signed distance function (SDF) so that hard or weak constraints can be applied to the Dirichlet, Neumann and Robin boundary conditions in subsequent training.

[0072] For the boundary heat transfer problem, the system uses a general form to describe the heat transfer relationship between the reservoir water and the air side, ensuring the physical consistency of the engineering boundary conditions:

[0073]

[0074] This is the Robin boundary condition, also known as the convection / heat transfer boundary condition, which is generally written as the normal diffusion flux on the boundary. Heat exchange term with the environment γ(φ-φ) ∞ )equal;

[0075] The unified expression for the heat transfer relationship between the reservoir water and the air side in this invention can be achieved by selecting different γ and φ values ​​at different boundary sections. ∞ The terms represent heat exchange with reservoir water or air, respectively, thereby ensuring the physical consistency of engineering boundary conditions; this invention corresponds to the Robin case.

[0076] - is a leading negative sign, which, together with n, forms -n·(), the notation for the outward normal orientation according to this invention. The negative sign ensures that the outward normal diffusion flux is positive. n is the unit outward normal vector at the boundary, its direction pointing from inside the computational domain to outside the boundary. · is the dot product, representing the inner product of the normal vector and the flux vector, i.e., the normal flux. A is the transport / diffusion coefficient tensor (symmetric positive definite). In the boundary heat transfer problem that this invention focuses on, A represents thermal conductivity, which can degenerate into scalar thermal conductivity in isotropic cases, with units of W / (m·K). Let be the gradient operator. Calculating the spatial gradient of a scalar field φ yields a vector composed of partial derivatives along each coordinate direction. φ is a scalar field. The diffusion flux vector is determined by the transport property A of the medium and the field gradient. Jointly determined. γ is the heat transfer coefficient or Robin coefficient, characterizing the intensity of heat exchange between the boundary and the external environment, with units of W / (m²). 2 ·K). φ ∞ As an environmental reference value, in the scenario of this invention, the ambient temperature is considered to be φ, which represents the heat exchange between the upstream and reservoir water. ∞ The intake water temperature is γ, which is the heat transfer coefficient on the intake side; the downstream heat exchange with the air is φ. ∞ Take the air temperature, and γ as the air-side heat transfer coefficient. R For the boundary subset to which Robin conditions are applied, onΓ R This equation applies only at the boundary Γ R established on.

[0077] Based on engineering requirements, select the set of parameters to be identified, including at least the elastic modulus, isotropic or anisotropic permeability coefficient, and equivalent thermal conductivity; at the same time, the Biot coefficient and specific water storage coefficient can be introduced to fully characterize the coupling behavior.

[0078] This selection action directly determines the output dimension and transformation type of the parametric field network, and ensures its positive value and range constraints based on the parametric field network.

[0079] In an optional implementation, the unified representation in S102 can be based on the geometric and boundary encoding of the boundary mesh and attribute mapping. Based on monitoring data, a boundary mesh representation of the dam body and dam foundation is generated, directly defining the geometric region and boundary location. At the mesh nodes or surfaces, the boundary type (such as Dirichlet, Neumann, Robin) is specified through attribute mapping, and related parameters (such as heat transfer coefficient and environmental values) are stored to achieve unified boundary management. In subsequent training, boundary conditions are directly applied through mesh attributes, such as fixing the field value at the Dirichlet boundary, defining the flux at the Neumann boundary, and defining the heat transfer relationship at the Robin boundary, and ensuring the continuity of boundary conditions through interpolation methods. However, the boundary mesh representation is difficult to integrate seamlessly with the deep learning framework in parameter inversion, requiring additional mesh processing steps, and is not suitable for irregular boundaries.

[0080] In this embodiment of the invention, S200 utilizes PINN inversion to simultaneously represent the state field and parameter field and calculates the physical residual, implementing parametric physical feasibility constraints and prior regularization, including the following steps S201-S202:

[0081] S201. Construct a physical information neural network, including a state field network and a parameter field network, and use a multilayer perceptron to simultaneously represent the state field and the parameter field and share the backbone layer.

[0082] In an embodiment of the present invention, a multilayer perceptron is used to simultaneously represent the state field and the parameter field and share a backbone layer, including the following steps A1-A2:

[0083] A1. During the network construction and initialization phase, the system uses a multilayer perceptron to simultaneously represent the state field and parameter field, and shares the backbone layer to improve sample efficiency.

[0084] The SDF provided by the geometry and boundary modeling unit allows Dirichlet conditions to be embedded as hard constraints:

[0085]

[0086] It should be noted that the geometry and boundary modeling unit provides a signed distance function s(x)(SDF), embedding the Dirichlet condition as a hard constraint on the Dirichlet boundary, s(x) = 0, thereby making φ(x) = φ b (x) is satisfied exactly at the boundary; within the domain, it is satisfied by... The field quantity can be freely adjusted, so there is no need to apply penalty weights to such boundaries during training.

[0087] Where φ(x) is the value of the target field quantity at position x, representing a Dirichlet-type boundary assignment object for any component in the state field (such as displacement component, pore pressure p, temperature T, etc.).b s(x) is the extension / interpolation function of the boundary given values ​​(i.e., Dirichlet data); at the Dirichlet boundary, it is consistent with the specified value of φ. s(x) is the SDF, where s(x) = 0 at the target Dirichlet boundary, and is the signed distance inside / outside the domain, used to ensure hard constraints. This is a free field function, output by a neural network, used to adjust the shape and size of φ within the domain, without disrupting the boundary condition φ = φ. b The hard constraints. x is a spatial position vector.

[0088] A2. Thus, there is no need to introduce penalty weights for this type of boundary during training, and the remaining Neumann and Robin conditions are implemented through the corresponding boundary residuals.

[0089] In an optional implementation, in S201, a multilayer perceptron can be used to simultaneously represent the state field and parameter field while sharing a backbone layer. Two completely independent multilayer perceptron networks can be constructed, one dedicated to representing the state field and the other to representing the parameter field. These two networks do not share any layers or parameters, receive the same input coordinate data, and independently output their respective field quantities. During training, the state field network and the parameter field network are coupled through a physical residual loss function to enforce the satisfaction of the governing equations and boundary conditions. The Dirichlet condition is embedded in the form of a signed distance function provided by the geometry and boundary modeling unit as a hard constraint to ensure that the boundary values ​​are accurately satisfied. The Neumann and Robin conditions are implemented by calculating the boundary residuals and adding them to the loss function. However, this implementation, due to network separation and lack of parameter sharing, results in low training sample efficiency.

[0090] In another alternative implementation, S201 can use a multilayer perceptron to simultaneously represent the state field and parameter field and share the backbone layer. This can be achieved by constructing a multilayer perceptron network with a shared input layer but independent hidden layers. The input layer is shared, but the subsequent hidden layers are divided into two independent branches: one for the state field output and the other for the parameter field output. The shared input layer extracts basic features, and then each branch performs specialized processing. The state field branch outputs state variables, and the parameter field branch outputs material parameters. During training, the number of parameters is reduced by sharing the input layer, but the branch structure allows the field quantities to learn independently. Dirichlet hard constraints are implemented using SDF, and other boundary conditions are handled through residuals; however, the shared input layer in this implementation limits the network's ability to capture the complex nonlinear relationships between the state field and parameter field.

[0091] The state field network outputs displacement, pore water pressure, and temperature; specifically:

[0092] In the PINN inversion solution unit, the system constructs a state field network to output displacement u, pore water pressure p, and temperature T, and a parameter field network to output the spatial distribution of material and seepage thermal parameters. The state field must simultaneously satisfy momentum equilibrium and coupled constitutive relations:

[0093]

[0094] The first formula is the momentum balance equation under static conditions, indicating that when there is no inertial term in the domain, the stress divergence and body force are in balance, and the resultant force is zero. The second formula is the coupled constitutive relation that unifies the pore pressure effect and thermal expansion strain into the stress calculation. First, the linear relationship between the effective stress σ' and the mechanical strain-thermal strain difference is obtained from the second formula. Then, the influence of pore water pressure (Biot's effective stress principle) is superimposed back to the total stress σ from the first formula.

[0095] σ is the total stress tensor, ρ is the density, which varies with position; b is the body force vector, such as the body force caused by gravitational acceleration; 0 is the zero vector, representing local equilibrium; σ' is the effective stress tensor, the stress excluding pore pressure; α B (x) represents the Biot coefficient field, whose value varies with spatial position x, measuring the intensity of the influence of pore pressure on stress; p is the pore water pressure; I is the identity tensor, a unit second-order tensor; C(x) is the elastic stiffness tensor field, which varies spatially heterogeneously with position x. When isotropic, it is related to the elastic modulus / Poisson's ratio; when anisotropic, it is a fourth-order tensor; α T (x) represents the thermal expansion coefficient field, which varies with x; T is the temperature, T0 is the reference temperature, and x is the spatial position vector, reflecting the above parameter fields C(x) and α. B (x), α T Spatial heterogeneity of (x); ε is the small strain tensor, defined by the symmetric gradient of displacement u:

[0096]

[0097] This allows us to incorporate the effects of pore pressure and thermal strain on the stress field into the same framework.

[0098] The mass conservation equation and Darcy's law for the seepage field are expressed as follows:

[0099]

[0100] It should be noted that this is a first-order mass conservation equation, with the change in the water storage term. With the volumetric strain rate term caused by consolidation The sum, minus the divergence of Darcy flux, equals the volumetric source term q. p (x,t); The formula containing v(x,t) is Darcy's law in vector form, representing the seepage velocity (or Darcy velocity) v(x,t) and the driving force. Proportional, with a proportionality constant of (Ratio of permeability to dynamic viscosity), the negative sign indicates that the flow direction is opposite to the pressure drop direction.

[0101] Among them, S s (x) is the specific water storage coefficient field, which varies with the spatial location x and measures the change in water content per unit volume of porous medium under a unit pressure change. In this invention, it is used as one of the parameters to be identified. Let p be the time derivative of the pore water pressure, representing the rate of change with time. For volumetric strain ε v The time derivative, ε v For the strain tensor trace, ε v =tr(ε); To take the divergence of the vector field within the brackets, k(x) is the permeability tensor field, which varies with x; μ is the dynamic viscosity. The force within the square brackets is the combined driving force of the pressure gradient and gravity. The gradient of pore pressure; ρ is the density of water; g is the acceleration vector due to gravity; q p (x,t) represents the source and sink terms, the injection and drainage intensity per unit volume, which varies with space x and time t. x,t represent spatial position and time; v(x,t) is the Darcy velocity vector field, which varies with x,t; - indicates the flow direction from high pressure to low pressure. The potential energy gradient correction due to gravity is given, and the square brackets [] are used to aggregate the driving force terms.

[0102] The temperature field satisfies the following when considering seepage and convection terms:

[0103]

[0104] It should be noted that the energy balance (convection-diffusion) equation for the temperature field considering the seepage and convection terms includes the equivalent volumetric heat capacity term of the medium. With permeation heat-carrying items The sum, minus the thermal flux divergence equals the body heat source q T (x,t);

[0105] Where ρ is the density of the solid / equivalent medium, and c is the specific heat capacity of the solid / equivalent medium. c is the time derivative of temperature T. w The specific heat capacity of water, The product of Darcy velocity v and temperature gradient represents convective heat transfer along the flow direction. q represents the divergence of the thermal conductivity term; κ(x) is the equivalent thermal conductivity field, which varies with position x; T (x,t) represents the volumetric heat source / sink, which varies with space and time.

[0106] The parameter field network outputs spatially heterogeneous material parameters and percolation thermal parameters, and ensures the positive definiteness and range constraints of the parameters through function transformation; specifically:

[0107] The positive definiteness and range constraints of the parameter field are achieved through a combination of neural representation and function transformation, thereby obtaining a physically feasible and differentiable spatial heterogeneous characterization, namely, the elastic modulus field, the anisotropic representation of thermal conductivity, the isotropic representation of permeability, and the thermal expansion coefficient field:

[0108] E(x)=E min +softplus(φ E (x;ψ E ))

[0109]

[0110] κ(x)=κ min +softplus(φ κ (x;ψ κ ))

[0111] α T (x)=α T,min +σ(φ T (x))

[0112] In the elastic modulus field, E(x) represents the elastic modulus field, and E... min The lower bound of E is physically non-negative; softplus(·) is ln(1+e) / 2. z Positive value mapping in the form of φ ensures that the output is >0; E (x;ψ E ) represents a latent scalar generated by a parametric field network; ψ E Its network parameters;

[0113] In the anisotropic representation of thermal conductivity, k(x) is the equivalent thermal conductivity tensor field, exp(·) is the exponential mapping, ensuring that the coefficients are positive or positive definite; φ k (x;ψ k ) represents the scalar / vectorized latent parameters of the network output; ψ k Its parameters are given by R, which is an orthogonal rotation matrix. Used to specify the principal axis direction; diag(k1,k2,k3) is the diagonal principal value matrix; k1, k2, and k3 are the principal thermal rates. This is the transpose of R;

[0114] In the isotropic representation of permeability, κ(x) represents the permeability field. min This is the lower limit for κ;

[0115] Among them, in the coefficient of thermal expansion field, α T (x) represents the thermal expansion coefficient field, α T,min For α T The lower bound of σ(·) is the Sigmoid function mapped to the (0,1) interval, and then the lower bound is achieved by translation; φ T (x) is the latent scalar of the network output, α T The range of values ​​for the Biot coefficient field in (x) is:

[0116] α B (x)∈(0,1]

[0117] α B (x) is the Biot coefficient field, which is physically confined to the open interval (0,1].

[0118] S202. Construct a loss function, including residuals in the form of multiple field strengths, residuals of boundary conditions and initial conditions, residuals of monitoring data, and prior regularization terms, and approximate volume integrals and boundary integrals through sampling points.

[0119] Based on the state field network and parameter field network inverted by PINN, multi-field strength form residuals are defined, and monitoring data residuals and prior regularization terms are further introduced;

[0120] The multi-field residuals are calculated as follows:

[0121]

[0122] Furthermore, when all physical equations are satisfied, the residual (r) mech ,r hyd ,r therm The value is zero; during training, the parameters and states are inverted by minimizing the joint loss of these residuals and the residuals of the monitoring data.

[0123] Where, r mech The momentum balance residual comes from The strong form; Let σ be the divergence of the stress tensor. Let ρ be the density, b be the force vector, and r be the force term. hyd For the seepage mass conservation residual, S s For the specific water storage coefficient field, p t Let α be the time derivative of the pore water pressure at time t. B For the Biot coefficient field, ε v,t Let ε be the volumetric strain at time t. v =tr(ε); To take the divergence of the vector field within the brackets, applied to the Darcy flux; k is the permeability field, μ is the dynamic viscosity. This is the driving force term, which is the pore pressure gradient minus the gravity term; Let q be the density of water, g be the acceleration vector due to gravity, and q be the velocity vector. p For seepage source and sink terms, r therm For temperature residual, This is the equivalent volumetric heat capacity term for the medium. The term represents the heat-carrying convection term of seepage, where v is the Darcy velocity. Let the divergence of the thermal conductivity term be denoted as . For the equivalent thermal conductivity field, q T For body heat source / sink;

[0124] Calculate the deviation between the boundary conditions and the initial conditions:

[0125]

[0126] in, The boundary / initial residuals for displacement, pore water pressure, and temperature are given; each residual is zero when the corresponding conditions are strictly satisfied. u represents the displacement field, output from the state field network. D Γ is the displacement boundary value given at the displacement Dirichlet boundary. u For the boundary subset to which the displacement Dirichlet condition is applied; onΓ u This indicates that the equation holds only on this boundary; p is the pore water pressure, output by the state field network; p D For the pressure boundary value given at the pressure Dirichlet boundary, Γ p For the boundary subset of the applied Dirichlet conditions, T is the temperature, output by the state field network; T D Γ is the temperature boundary value given at the temperature Dirichlet boundary. T For the boundary subset where the temperature Dirichlet condition is applied;

[0127] The residual of the monitoring data is measured as follows:

[0128]

[0129] Where, r data The residual vector of the monitoring data represents the error in the observation space; As an observation operator, it maps the internal field quantities u, p, and T of the model to observed quantities (such as sampling / interpolation at sensor locations, extraction of a component, spatial-temporal averaging, unit and scale transformations, etc.), thus placing them in the same space and dimensions as the measured data. obs For the observation data vector (i.e., monitoring records), and They have consistent dimensions and physical units.

[0130] Parameter priors and regularization using H 1The combined form of smoothing, total variation, and box constraints is represented as follows:

[0131]

[0132] Furthermore, for all regularized parameter field sets Θ (such as E, k, ... α B ,α T ,S s The combined prior / regularization applied (selected as needed) is: The item is H 1 Smoothing / Tikhonov regularization suppresses drastic gradients; λ TV The term in question is a total variation (TV) regularizer, which encourages piecewise smoothing / boundary preservation; λ box Φ box The term containing (θ) is the box constraint penalty, which restricts the parameters to the physically feasible range.

[0133] Where R is the joint prior / regularization term, and θ∈Θ represents the term for each field θ in the parameter field set Θ. , λ TV , λ box These are the non-negative weight coefficients corresponding to the three types of regularization; Ω is the computational domain. Let θ be the gradient of the parameter field; ||·|| is the Euclidean / Frobenius norm, Φ box (θ) is the box constraint penalty function; it takes a positive value when θ falls outside the given interval (that is, it can realize soft constraint or barrier constraint), and takes zero or minimal value inside the interval; dΩ is the volume element.

[0134] All values ​​are approximated by volume integrals at the sampling points and boundary integrals, which together constitute the joint loss:

[0135]

[0136] It should be noted that the three types of strong form residuals (mechanical, seepage, and thermal), boundary / initial value residuals, and monitoring data residuals are weighted and summed, and combined with the prior regularization R to form a joint loss. The volume and boundary integrals are approximated by sampling point volume integrals and boundary integrals.

[0137] in, For the total loss, w m w h w t w b w d r represents the five non-negative weights corresponding to different residuals. Γ The boundary / initial value residual vector contains Equal components; ||r Γ || 2 Denotes its square norm; <·>r 、<·> S 、<·> Γ Let S represent the sampling approximations / means of the volume integral, boundary integral, and data error using the in-domain sampling points, boundary sampling points, and data sample set S, respectively.

[0138] In an embodiment of the present invention, in S300, the organization optimizer, residual-driven sampling, and weight adaptive training scheduling are performed, and the results are verified, visualized, and exported through the verification and output unit, including the following steps S301-S302:

[0139] S301. The training schedule automatically differentiates and calculates the strong form residuals, boundary / initial value residuals and monitoring data residuals of the multi-field equations, and performs joint optimization under the constraint of the regularization term. During optimization, the adaptive gradient method based on mini-batch is used to find the best under a large learning rate, and the quasi-Newton method is used to refine the parameters under a small learning rate.

[0140] Adaptive adjustment is performed based on the gradient norm, which involves calculating the ratio of the gradient norm of all tasks to the average gradient norm and updating the weight coefficients w based on the current ratio. i :

[0141]

[0142] It should be noted that if the gradient norm G of a certain task i Above average The weights will be increased or decreased accordingly to balance the training, with the intensity controlled by α.

[0143] in, Let G be the weight of the i-th task in the k+1-th iteration (i∈{mech,hyd,therm,b,d}). i Let the gradient norm of the i-th task under the current weights be defined as follows: This represents the average of the gradient norms for each task. α is the adjustment index, 0≤α≤1, which controls the adaptive speed / force; Let L be the gradient of the network parameter set Θ; ||·|| is the gradient vector L. 2 Norm; For the i-th sub-loss, Θ represents all trainable parameters, which is the set of parameters for both the state field network and the parameter field network.

[0144] After every few rounds of training, the sample points are enriched and migrated according to the residual distribution, and adaptive sampling is performed on high residual regions, with the ratio of in-domain sampling to boundary sampling maintained in the range of 20% to 30%.

[0145] The training process terminates when both the verification point error and the physical residual reach the target. The resulting parameter field and state field can then be exported from the verification and output unit to a standard engineering format for verification and display.

[0146] It should be further explained that, in terms of network structure, state field networks and parameter field networks can adopt multilayer perceptrons with 6-10 layers and 64-256 neurons per layer.

[0147] For fields containing high-frequency gradients, structures with sinusoidal or SiLU activation are more conducive to convergence.

[0148] Regarding the optimizer, an adaptive gradient method is initially used for coarse search, with a learning rate of 10. -3 Up to 10 -4 Within the interval, the cosine annealing method is used to decrease the value, and then a quasi-Newton method is switched to achieve secondary convergence.

[0149] Gradients can be clipped in the range of 1 to 5 at any stage to prevent gradient explosion.

[0150] Regarding sample selection, initially, 10 points were projected both within the domain and at the boundary. 4 The magnitude ranges from several thousand to several thousand, and the high residual subdomains are gradually encrypted based on the residual hot zones;

[0151] Maintaining the ratio between boundary sampling and intra-domain sampling within the range of approximately 20% to 30% can balance boundary and intra-domain constraints.

[0152] To eliminate numerical scaling bias caused by differences in multiple physical dimensions, the data is normalized according to feature length, velocity, pressure and temperature before entering the network.

[0153] Small-scale deep ensembles or evidence learning are introduced to provide pixel-level uncertainty estimates, which are not involved in the main loss calculation but can be used for subsequent engineering interpretation and region screening.

[0154] In an embodiment of the present invention, step S302, outputting the inverted parameter field and state field and exporting them in a specified format for engineering verification and visualization, includes the following steps B1-B2:

[0155] B1. After training convergence, the output includes contour lines and slices of the spatially heterogeneous parameter field, fitted curves of key monitoring points, and residual statistics.

[0156] B2. Export the parameter field and state field as VTK, GeoTIFF or comma-separated format, display them in visualization software, and cross-validate them with the external finite element model. If there is a systematic deviation between the external finite element model and this system in a local area, the deviation area will be the focus of a new round of sampling for iterative improvement.

[0157] In an optional implementation, data export in S302 can be based on standard engineering data exchange formats and lightweight visualization tools. After training convergence, the parameter field and state field data are exported to standard engineering data exchange formats (such as CGNS or XDMF). Lightweight open-source visualization tools (such as ParaView or VisIt) are used to directly load the exported files to generate two-dimensional slices and three-dimensional contour maps of the dam body-dam foundation area, and the locations of monitoring points are overlaid to display the fitting situation. The time series data of key monitoring points are saved as simple text formats (such as TXT or CSV) for plotting fitting curves and residual statistics in general charting software (such as Python Matplotlib or Excel), avoiding reliance on professional software licenses.

[0158] In another optional implementation, the data export in S302 can also be integrated into a web-based interactive reporting system. After training convergence, the parameter field and state field data are converted into a lightweight mesh format (such as GLTF or OBJ) and embedded into the web-based interactive report. JavaScript libraries (such as Three.js) are used to implement 3D model rotation, scaling, and parameter switching within the browser. Monitoring data residuals and inversion results are stored in JSON format. The fitted curves and residual distributions of key points are displayed through dynamic charts on the web page (such as D3.js). Users can interactively query parameter values ​​at specific locations. The export function provides one-click generation of PDF reports, including static screenshots and statistical summaries, which facilitates rapid distribution and documentation. At the same time, the download link for the original data is retained for in-depth analysis.

[0159] Example 3 is an embodiment of the present invention, which provides an intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0160] Experiment 1: Two-dimensional gravity dam;

[0161] In a composite example, the computational domain is a typical gravity dam profile, the reservoir water level varies with a seasonal curve, the dam base is fixed, and the rest of the outer surface is free.

[0162] The actual parameter field is set such that the elastic modulus near the dam heel is higher than that of the upper part of the dam body, and a weakly permeable zone is arranged below the dam heel; the thermal conductivity varies slowly between 2.0 and 2.5 W / (m·K).

[0163] The monitoring data consists of thermometers, piezometers, and displacement gauges, and is collected using a combination of synchronous and asynchronous sampling.

[0164] The inverted elastic modulus distribution deviates from the true value of the structural boundary position by less than 2m. The center position and width of the weakly permeable zone are accurately reconstructed, and the seasonal amplitude and phase of the temperature field match the true value.

[0165] This experiment demonstrates the applicability of the method under conditions of thermo-permeability-mechanical coupling and spatial heterogeneity.

[0166] Experiment 2: Three-dimensional arch dam, engineering data;

[0167] In actual engineering, monitoring data and operating events over many years are imported, and the parameters are mainly selected based on elastic modulus, permeability tensor and thermal conductivity, and the "zoning constant + smooth transition" is adopted according to the geological zoning of the dam body and dam foundation.

[0168] The upstream surface adopts a heat exchange boundary, and the downstream convective heat transfer coefficient is corrected according to the wind speed; events such as gate opening and closing are loaded in a time-period manner.

[0169] During the training phase, seepage and thermal parameters are first identified, and then joint training is conducted to avoid task competition.

[0170] The output includes a parameter zoning map, displacement fitting curves for key points, and a histogram of seepage pressure residuals. The parameter field is provided in VTK format for third-party platforms to verify and display.

[0171] Example 4 is an embodiment of the present invention. The above is an illustrative scheme of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning. It should be noted that the technical solution of the intelligent inversion system for multi-field coupled concrete dam parameters based on deep learning belongs to the same concept as the technical solution of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described above. For details not described in detail in the technical solution of the intelligent inversion system for multi-field coupled concrete dam parameters based on deep learning in this embodiment, please refer to the description of the technical solution of the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning described above.

[0172] This embodiment provides a deep learning-based intelligent inversion system for multi-field coupled concrete dam parameters, including: a data access and cleaning unit, a geometry and boundary modeling unit, a PINN (deep learning) inversion solution unit, a constraint and regularization unit, a training and scheduling unit, and a verification and output unit.

[0173] The system provided by this invention is functionally composed of several mutually cooperating units;

[0174] The data access and cleaning unit is used to access and manage temperature, seepage pressure, displacement, reservoir water level, and meteorological records.

[0175] The geometry and boundary modeling unit is used to uniformly represent the geometric region of the dam body and dam foundation and its boundary types;

[0176] The PINN inversion solver is used to simultaneously represent the state field and parameter field and calculate the physical residual.

[0177] The constraint and regularization unit is used to implement parametric physical feasibility constraints and prior regularization;

[0178] The training scheduling unit is used to organize the optimizer, residual-driven sampling, and weight adaptation;

[0179] The verification and output unit is used for result verification, visualization, and data export.

[0180] All units are coupled to each other through data flow and control flow, with the PINN inversion solution unit undertaking the core computational functions in the system.

[0181] This embodiment also provides an electronic device applicable to the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning as proposed in the above embodiment.

[0182] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning as proposed in the above embodiment.

[0183] The storage medium proposed in this embodiment and the intelligent inversion method for multi-field coupled concrete dam parameters based on deep learning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0184] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0185] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based intelligent inversion method for multi-field coupled concrete dam parameters, characterized in that: The application relates to a multi-field coupling concrete dam parameter inversion method based on physics-informed neural network (PINN). The method comprises the following steps: accessing multi-field coupling concrete dam data through a data access and cleaning unit, and uniformly expressing dam body-dam foundation geometric regions and boundary types through a geometric and boundary modeling unit; The method simultaneously expresses state fields and parameter fields through PINN inversion and calculates physical residuals, implements parameter physical feasibility constraints and prior regularization, and organizes a training optimizer, residual-driven sampling and weight self-adaptation to perform training scheduling, and performs result review, visualization and data export through a verification and output unit. The multi-field coupling concrete dam parameters comprise monitoring data of a concrete dam, including first monitoring data, second monitoring data, third monitoring data, fourth monitoring data and fifth monitoring data; 2. The deep learning-based multi-field coupled concrete dam parameter intelligent inversion method according to claim 1, characterized in that: The monitoring data is uniformly processed in units and time, and the sensor positions are calibrated in the same coordinate system. The uniform expression comprises modeling dam body-dam foundation geometric regions and boundary types based on the monitoring data, and uniformly encoding the calculation domain and the boundary through a signed distance function; 3. The deep learning-based multi-field coupled concrete dam parameter intelligent inversion method according to claim 2, characterized in that: The heat exchange relationship between the water and the air on the engineering boundary is described in a general form, and the physical consistency of the engineering boundary conditions is ensured. The PINN inversion comprises constructing a physics-informed neural network, including a state field network and a parameter field network, to simultaneously express the state fields and the parameter fields through a multilayer perceptron and share a backbone layer; 4. The deep learning-based multi-field coupled concrete dam parameter intelligent inversion method according to claim 3, characterized in that: The state field network outputs displacement, pore water pressure and temperature; The parameter field network outputs spatially heterogeneous material parameters and seepage thermal parameters, and ensures the positive definiteness and range constraints of the parameters through function transformation. The implementation of the parameter physical feasibility constraints and the prior regularization comprises constructing a loss function, including multi-field strong form residuals, boundary condition and initial value condition residuals, monitoring data residuals and prior regularization terms, and approximating volume integrals and boundary integrals through sampling points; 5. The deep learning-based multi-field coupled concrete dam parameter intelligent inversion method according to claim 4, characterized in that: The multi-field strong form residuals are defined according to the state field network and the parameter field network of the PINN inversion, and the monitoring data residuals and the prior regularization terms are further introduced; The multi-field residuals are calculated as follows: The boundary and initial value condition deviations are calculated as follows: The monitoring data residuals are measured as follows: All values are approximated through volume integrals and boundary integrals of sampling points to form a joint loss: Parameter priors and regularization The combination of smoothness, total variation and box constraints is represented as: The training scheduling comprises training scheduling to automatically differentiate the strong form residuals of the multi-field equation, the boundary / initial value residuals and the monitoring data residuals, and completing joint optimization under the constraint of the regularization term, wherein, during optimization, an adaptive gradient method based on a small batch is used to search for optimization at a relatively large learning rate, and a quasi-Newton method is used to fine-tune the parameters at a small learning rate; where, is the momentum balance residual, is the divergence of the stress tensor ; is the body force term, is the density, is the body force vector, is the seepage mass conservation residual, is the specific storage field, is the time derivative of the pore water pressure at t, is the Biot coefficient field, is the volumetric strain at t, is the divergence of the vector field in the brackets, is the permeability field, is the dynamic viscosity, is the density of water, is the gravitational acceleration vector; is the seepage source / sink term, is the temperature residual, is the medium equivalent volumetric heat capacity term, is the seepage heat advection term, is the Darcy velocity, is the divergence of the heat conduction term, is the equivalent thermal conductivity field, is the volumetric heat source / sink; , , are the boundary / initial residual of displacement, pore water pressure and temperature, is the displacement field, is the displacement boundary value given at the displacement Dirichlet boundary, is the boundary subset where the displacement Dirichlet condition is imposed; is the pore water pressure, is the pressure boundary value given at the pressure Dirichlet boundary, is the boundary subset where the pressure Dirichlet condition is imposed, is the temperature, is the temperature boundary value given at the temperature Dirichlet boundary, is the boundary subset where the temperature Dirichlet condition is imposed; is the monitoring data residual vector, is the observation operator, is the observation data vector; is the joint prior and regularization term, is the non-negative weight coefficient corresponding to the three types of regularization; is the computational domain, is the parameter field the gradient of is the Euclidean / Frobenius norm, is the box constraint penalty function; is the total loss, are five non-negative weights corresponding to different residuals, is the boundary / initial residual vector, denotes its squared norm; , , is the set of domain interior sampling points, boundary sampling points and data samples respectively sampled mean of volume integral, boundary integral and data error.

6. The deep learning-based multi-field coupled concrete dam parameter intelligent inversion method according to claim 5, characterized in that: After every several rounds of training, the sample points are densely filled and migrated according to the residual distribution, and adaptive sampling is performed on the high residual area, wherein the proportion of the domain sampling and the boundary sampling is maintained in the range of 20% to 30%. Adaptive adjustment according to the gradient norm, that is, by calculating the ratio of the gradient norm of all tasks to the average gradient norm, and updating the weight coefficient according to the current ratio : in, , For the first During the nth iteration Each task weight For the first The gradient norm of each task under the current weights; This represents the average of the gradient norms for each task. To adjust the index, For the network parameter set The gradient; For the gradient vector Norm, For the first A loss of size, For all trainable parameters; The training process is terminated when the validation point error and the physical residual meet the requirements at the same time, and the obtained parameter field and state field can be exported to a standard engineering format by the verification and output unit for review and display. The verification and output unit comprises outputting the parameter field and the state field obtained through inversion, and exporting the parameter field and the state field to a specified format for engineering verification and visualization; 7. The deep learning-based multi-field coupled concrete dam parameter intelligent inversion method according to claim 6, characterized in that: After training convergence, the output includes contour lines and slices of the spatially heterogeneous parameter field, fitting curves of key monitoring points and residual statistical charts. ​ The parameter field and the state field are derived as VTK, GeoTIFF or comma-separated format, displayed in visualization software, and cross-verified with external finite element models; if there is systematic deviation between the external finite element model and the system in the local area, the deviation area is taken as the focus of the next round of sampling for iterative improvement.

8. The system for intelligent inversion of parameters of a multi-field coupled concrete dam based on deep learning, applying the method for intelligent inversion of parameters of a multi-field coupled concrete dam based on deep learning according to any one of claims 1-7, characterized in that, The method comprises the following steps: a data access and cleaning unit, a geometry and boundary modeling unit, a PINN inversion solving unit, a constraint and regularization unit, a training scheduling unit, and a verification and output unit; The data access and cleaning unit is used for accessing and managing temperature, osmotic pressure, displacement, and reservoir water level and weather records. The geometry and boundary modeling unit is used for unified expression of dam body-dam foundation geometric region and its boundary type. The PINN inversion solving unit is used for simultaneous expression of state field and parameter field and calculation of physical residual. The constraint and regularization unit is used for implementation of parameter physical feasibility constraint and prior regularization. The training scheduling unit is used for organization of optimizer, residual driven sampling and weight self-adaptation. The verification and output unit is used for result review, visualization and data export. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the deep learning-based multi-field coupled concrete dam parameter intelligent inversion method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the deep learning-based multi-field coupled concrete dam parameter intelligent inversion method in any one of claims 1 to 7.

Images