An intelligent monitoring and diagnosis system for water conservancy projects based on digital twinning
By combining multidimensional mapping perception and twin model dynamic calibration with structural health diagnosis, the problem of high false alarm rate of water conservancy engineering detection system under complex hydrological conditions has been solved, and early identification and accurate diagnosis of structural damage have been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUNNAN YEXIN PLANNING & DEV GRP CO LTD
- Filing Date
- 2026-06-22
- Publication Date
- 2026-07-21
AI Technical Summary
Existing water conservancy engineering monitoring systems lack in-depth analysis of the coupling mechanism of multidimensional physical fields, making it difficult to distinguish between sensor zero-point drift and structural creep damage under complex hydrological conditions. This results in a high false alarm rate and makes it impossible to achieve accurate condition monitoring around the clock.
A multidimensional mapping sensing module is used to synchronously collect data and perform spline interpolation alignment. A twin model dynamic calibration module is used to adjust model parameters through gradient descent optimization algorithm. Combined with a structural health diagnosis module, the structural status is determined in real time, and a damage location report is generated.
It enables early identification and trend prediction of hidden damage to water conservancy projects, and significantly improves the anti-interference ability and diagnostic confidence of the detection system.
Smart Images
Figure CN122432948A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital twin simulation technology, and in particular to an intelligent monitoring and diagnostic system for water conservancy projects based on digital twins. Background Technology
[0002] The field of digital twin simulation technology involves using data such as physical models, sensor updates, and operational history to map data in a virtual space, thereby reflecting the entire lifecycle of the corresponding physical equipment. Traditional hydraulic engineering monitoring systems rely on physical sensors deployed on dams or gates to collect stress, displacement, and seepage pressure data, setting fixed safety thresholds for comparison and alarms, and using regular manual inspections to verify the authenticity of the sensor data and the structural integrity.
[0003] In practical operation, existing technologies lack in-depth analysis of the coupling mechanism of multidimensional physical fields, which makes it impossible to accurately isolate the nonlinear effects of environmental load fluctuations on structural response by relying solely on static thresholds. Furthermore, it is difficult to distinguish between the zero-point drift of the sensor itself and the minute creep damage of the structure. This results in a high false alarm rate caused by missing minor hidden dangers under complex hydrological conditions or by environmental interference, making it impossible to achieve accurate status visibility around the clock. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a digital twin-based intelligent monitoring and diagnosis system for water conservancy projects.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a digital twin-based intelligent monitoring and diagnosis system for water conservancy projects, comprising:
[0006] The multidimensional mapping perception module is used to synchronously collect environmental load driving data and key node response data at the site of the water conservancy project. Based on the timestamp, spline interpolation alignment is performed on the heterogeneous environmental load driving data and key node response data, and the aligned data is mapped to the three-dimensional spatial grid nodes of the virtual model to construct the current state input matrix.
[0007] The twin model dynamic calibration module is used to import the current state input matrix into a pre-constructed high-fidelity hydraulic finite element model, solve the theoretical physical field distribution data using the physical control equation, construct a loss function with the convergence objective of minimizing the weighted Euclidean distance between the key node response data and the theoretical physical field distribution data through the gradient descent optimization algorithm, solve the permeability coefficient and elastic modulus parameters of the model based on the back-inverse iteration of the loss function, and output the constitutive parameter correction vector.
[0008] The structural health diagnosis module is used to extract the time series features of the constitutive parameter correction vector, construct the parameter evolution phase space trajectory of the constitutive parameter correction vector in the multidimensional phase space, calculate the evolution rate magnitude in the tangent direction of the parameter evolution phase space trajectory, compare the evolution rate magnitude with the preset nonlinear safety threshold surface in dynamic spatial position, determine the structural state based on the comparison result, and generate a structural damage location report when the determination result exceeds the limit.
[0009] As a further aspect of the present invention, the multidimensional mapping perception module specifically includes:
[0010] The source data aggregation submodule is used to build a heterogeneous sensor network interface to receive environmental variables, including upstream water level and rainfall intensity, in real time, and simultaneously acquire structural response indicators, including dam seepage pressure and displacement. It also uses a network time protocol server to stamp all acquired raw data packets with a unified high-precision timestamp.
[0011] The timing synchronization alignment submodule is used to analyze the sampling frequency difference between the environmental load driving data and the key node response data, reconstruct the continuous change curve of sparse low-frequency data using cubic B-spline interpolation algorithm, and resample it onto the dense time axis of high-frequency data to generate a synchronized time series dataset.
[0012] The spatial grid mapping submodule is used to establish the coordinate transformation matrix between the physical sensor positions and the virtual finite element grid nodes. It uses shape function interpolation to project the synchronized time series data onto the corresponding nodes of the three-dimensional grid and assembles the spatial distribution vector to generate the current state input matrix.
[0013] As a further aspect of the present invention, the twin model dynamic calibration module specifically includes:
[0014] The forward solution submodule of the physical field is used to call the finite element kernel solver to load the current state input matrix as boundary conditions and body loads, solve the partial differential equations controlling seepage and stress field based on the current material parameters, and calculate the theoretical response value at the corresponding node of the sensor to generate the theoretical physical field distribution data.
[0015] The residual objective construction submodule is used to calculate the deviation vector between the key node response data and the theoretical physical field distribution data, assign confidence weights to different monitoring points according to sensor accuracy and spatial importance, and construct a scalar loss function that aims to minimize the weighted Euclidean distance and includes a regularization term.
[0016] The parameter inversion optimization submodule is used to calculate the sensitivity Jacobian matrix of the loss function with respect to material properties using the gradient descent optimization algorithm, iteratively update the permeability coefficient and elastic modulus along the negative gradient direction until the loss value is lower than the preset convergence threshold, and output the constitutive parameter correction vector.
[0017] As a further aspect of the present invention, the structural health diagnosis module specifically includes:
[0018] The phase space reconstruction analysis submodule is used to apply the time delay embedding theorem to the time series of the constitutive parameter correction vector, calculate the optimal delay time using the mutual information method and determine the embedding dimension using the spurious neighbor method, and construct a high-dimensional state vector to form the phase space trajectory of the parameter evolution.
[0019] The evolution feature calculation submodule is used to perform differential geometric analysis on the reconstructed parameter evolution phase space trajectory, calculate the tangential vector of the trajectory point at the current time, and solve the Euclidean norm of the tangential vector to quantify the dynamic change rate of the system state and generate the evolution rate modulus.
[0020] The state deviation determination submodule is used to map the calculated evolution rate magnitude to the domain of the pre-trained nonlinear safety threshold surface, determine whether the current state point is located within the safe area enclosed by the nonlinear safety threshold surface, and trigger an alarm mechanism when the state point crosses the boundary to generate the structural damage location report.
[0021] As a further aspect of the present invention, the data spatiotemporal alignment operation performed by the timing synchronization alignment submodule specifically includes:
[0022] Acquire the environmental load driving data and the key node response data with original timestamps, identify the signal stream with the highest sampling frequency as the main reference clock reference, and calculate the time lag and frequency ratio of other heterogeneous data streams relative to the main reference clock reference.
[0023] A sliding window mechanism is used to extract local data segments. Within each window, a cubic spline interpolation function that can maintain the continuity of the first and second derivatives of the data is constructed. The interpolation function is resampled and calculated based on the discrete time points of the main reference clock to fill the information gaps caused by low-frequency sampling.
[0024] Outlier detection and smoothing are performed on all resampled data streams to remove outlier noise caused by sensor malfunctions. The processed multidimensional data are then merged in time-aligned order to generate a synchronous monitoring sequence.
[0025] As a further aspect of the present invention, the process by which the spatial grid mapping submodule establishes the current state input matrix specifically includes:
[0026] Read the three-dimensional mesh topology data of the virtual model and the actual installation coordinates of the physical sensors, and use the nearest neighbor search algorithm to identify the cell to which each sensor belongs and the number of its adjacent nodes in the mesh space;
[0027] Based on the finite element isoparametric element theory, the local coordinates of the sensor position relative to the node of the element are calculated, and an interpolation mapping relationship including shape function weights is constructed to allocate the physical quantity values measured by the sensor to the corresponding mesh node degrees of freedom.
[0028] Data mapping is completed by traversing all monitoring points. For areas where no sensors are deployed, Kriging interpolation is used to deduce node values. Finally, the current state input matrix is generated according to the input format specifications of the finite element solver.
[0029] As a further aspect of the present invention, the process of constructing the loss function by the residual target construction submodule specifically includes:
[0030] The theoretical physical field distribution data and the real-time monitoring response data of the key nodes are obtained in the current iteration step. The basic weight coefficient of each monitoring point is determined according to the reciprocal of the measurement standard deviation of the sensors at each monitoring point, and a weighted matrix is generated by combining the structural vulnerability factor of the area where the monitoring point is located.
[0031] By introducing a Tikhonov regularization term to constrain the smoothness of the parameter space, and combining it with the weighting matrix, a loss function expression of the following form is constructed:
[0032] ;
[0033] in, Represents the value of the loss function. This represents the constitutive parameter correction vector to be optimized. Represents the total number of monitoring points. Representing the The overall weight of each monitoring point Representing the The key node response data of each point. This represents the corresponding theoretical physical field distribution data. Represents the regularization coefficient. This represents the prior parameter vector from the previous time step.
[0034] As a further aspect of the present invention, the reverse iterative solution process performed by the parameter inversion optimization submodule specifically includes:
[0035] Based on the adjoint variable method or the finite difference method, the gradient vector of the loss function with respect to constitutive parameters such as permeability coefficient and elastic modulus is calculated, and the inverse matrix of the Hessian matrix is approximated by the BFGS formula of the quasi-Newton method to correct the search direction.
[0036] Perform a line search operation along the corrected search direction to determine the optimal step size that allows the loss function to decrease sufficiently. Update the current constitutive parameter estimate accordingly and check whether the updated parameter is within the preset constraint range that conforms to physical meaning.
[0037] Determine whether the decrease in loss function or the update of parameters in the current iteration step is less than the preset convergence tolerance. If not, feed the updated parameters back to the forward solution submodule of the physical field for the next round of calculation. If the condition is met, terminate the iteration and output the constitutive parameter correction vector.
[0038] As a further aspect of the present invention, the process by which the phase space reconstruction analysis submodule constructs the parameter evolution phase space trajectory specifically includes:
[0039] Extract the key components characterizing structural stiffness and permeability from the constitutive parameter correction vector, calculate the autocorrelation function of the normalized time series, and select the component whose autocorrelation function first decreases to its initial value. The time length at which the delay point is taken as the optimal delay time;
[0040] Calculate the proportion of false neighbor points that changes with the increase of the embedding dimension, and determine the dimension when the proportion converges to zero as the minimum embedding dimension. Construct a high-dimensional delay coordinate vector containing historical state information based on the optimal delay time and the minimum embedding dimension.
[0041] By connecting all the constructed high-dimensional delayed coordinate vectors in chronological order, a continuous geometric curve describing the evolution path of the system's dynamic behavior is generated in the multidimensional phase space, thus establishing the phase space trajectory of the parameter evolution.
[0042] As a further aspect of the present invention, the construction process of the nonlinear safety threshold surface in the state deviation determination submodule specifically includes:
[0043] Collect the historical constitutive parameter correction vectors generated during the healthy operation cycle of the water conservancy project, and use the phase space reconstruction analysis submodule to transform them into a set of historical healthy phase space trajectory points;
[0044] The distribution pattern of the historical health phase space trajectory points set in the multidimensional feature space is learned by using a single-class support vector machine algorithm. The data is then mapped to a higher-dimensional space using a Gaussian kernel function to find the optimal hyperplane and establish the nonlinear decision boundary that encloses all health state points.
[0045] The nonlinear decision boundary is inversely mapped back to the original phase space to generate a closed surface, and the closed surface is solidified into the nonlinear safety threshold surface for real-time online monitoring and comparison.
[0046] Compared with the prior art, the advantages and positive effects of the present invention are as follows:
[0047] In this invention, by constructing a constitutive parameter correction vector and tracking its evolution trajectory in real time, the structural response deviation is transformed into a dynamic adjustment of the model's physical parameters using inversion correction logic. This accurately quantifies the substantial changes in the structural stiffness or permeability, effectively eliminating the interference of elastic deformation caused solely by water pressure or temperature fluctuations. It enables early identification and trend prediction of hidden damage to water conservancy projects, significantly improving the anti-interference capability and diagnostic confidence of the detection system. Attached Figure Description
[0048] Figure 1 This is a flowchart illustrating the principle of the digital twin-based water conservancy engineering detection system of the present invention.
[0049] Figure 2 This is a detailed flowchart of the operation of the multidimensional mapping perception module of the present invention;
[0050] Figure 3 This is a detailed flowchart of the operation of the twin model dynamic calibration module of the present invention;
[0051] Figure 4 This is a detailed flowchart of the structural health diagnosis module of the present invention. Detailed Implementation
[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.
[0053] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.
[0054] Please see Figure 1 and Figure 2 This invention provides a technical solution: a digital twin-based intelligent monitoring and diagnosis system for water conservancy projects, comprising:
[0055] The multidimensional mapping perception module is used to synchronously collect environmental load driving data and key node response data at the site of water conservancy projects. Based on timestamps, it performs spline interpolation alignment operations on heterogeneous environmental load driving data and key node response data, and maps the aligned data to the three-dimensional spatial grid nodes of the virtual model to construct the current state input matrix.
[0056] The multidimensional mapping perception module specifically includes:
[0057] The source data aggregation submodule is used to build a heterogeneous sensor network interface to receive environmental variables, including upstream water level and rainfall intensity, in real time, and simultaneously acquire structural response indicators, including dam seepage pressure and displacement. It also uses a network time protocol server to stamp all acquired raw data packets with a unified high-precision timestamp.
[0058] The time-series synchronization and alignment submodule is used to analyze the sampling frequency difference between environmental load-driven data and key node response data, reconstruct the continuous change curve of sparse low-frequency data using cubic B-spline interpolation algorithm, and resample it onto the dense time axis of high-frequency data to generate a synchronized time series dataset.
[0059] The spatial mesh mapping submodule is used to establish the coordinate transformation matrix between the physical sensor positions and the virtual finite element mesh nodes. It uses shape function interpolation to project the synchronized time series data onto the corresponding nodes of the three-dimensional mesh and assembles the spatial distribution vector to generate the current state input matrix.
[0060] The data spatiotemporal alignment operations performed by the timing synchronization and alignment submodule specifically include:
[0061] Acquire environmental load driving data and key node response data with original timestamps, identify the signal stream with the highest sampling frequency as the main reference clock reference, and calculate the time lag and frequency ratio of other heterogeneous data streams relative to the main reference clock reference.
[0062] A sliding window mechanism is used to extract local data segments. Within each window, a cubic spline interpolation function that can maintain the continuity of the first and second derivatives of the data is constructed. The interpolation function is resampled and calculated based on the discrete time points of the main reference clock to fill the information gaps caused by low-frequency sampling.
[0063] Outlier detection and smoothing are performed on all resampled data streams to remove outlier noise caused by sensor malfunctions. The processed multidimensional data are then merged in time-aligned order to generate a synchronous monitoring sequence.
[0064] The process of establishing the current state input matrix in the spatial grid mapping submodule specifically includes:
[0065] Read the three-dimensional mesh topology data of the virtual model and the actual installation coordinates of the physical sensors, and use the nearest neighbor search algorithm to identify the cell to which each sensor belongs and the number of its adjacent nodes in the mesh space;
[0066] Based on the finite element isoparametric element theory, the local coordinates of the sensor position relative to the node of the element are calculated, and an interpolation mapping relationship including shape function weights is constructed to allocate the physical quantity values measured by the sensor to the corresponding mesh node degrees of freedom.
[0067] Data mapping is completed by traversing all monitoring points. For areas where no sensors are deployed, the Kriging interpolation method is used to deduce the node values. Finally, the current state input matrix is generated according to the input format specification of the finite element solver.
[0068] High-precision sensor networks deployed at water conservancy project sites (such as concrete gravity dams) first initiate a full data acquisition process. Fiber optic grating sensor arrays read strain data from key sections of the dam body in real time at a sampling frequency of 100Hz. Simultaneously, vibrating wire piezometers collect seepage pressure data from the dam foundation and body at a frequency of 10Hz. These two types of data constitute the critical node response data. In parallel, meteorological and hydrological stations located in the reservoir area collect upstream water level and rainfall intensity data at 1-minute and 10-minute intervals, respectively, using radar level gauges and tipping bucket rain gauges, forming environmental load-driven data. All front-end acquisition devices are connected to the field data acquisition unit (DAQ) via industrial Ethernet. The high-precision crystal oscillator clock source built into the DAQ synchronizes with the NTP server of the BeiDou satellite timing system every 60 seconds, ensuring that all data packets are tagged with a unified timestamp with an error of no more than 1 millisecond in the header.
[0069] To address the "data fragmentation" problem caused by the significant differences in sampling frequencies of the aforementioned data sources (from 100Hz to 0.0016Hz), a time-series synchronization and alignment operation is performed. The system selects the time axis of the strain data with the highest sampling frequency (100Hz) as the primary reference clock. For low-frequency water level (once every minute) and rainfall (once every 10 minutes) data, instead of using simple linear interpolation to avoid introducing angular errors, a cubic B-spline interpolation function is constructed. A sliding time window of 600 seconds is selected, and all discrete water level observation points within the window are extracted as control points. The coefficients of the B-spline basis function are calculated using the DeBoor algorithm. During the calculation process, node vectors are set to ensure the continuity of the first and second derivatives of the interpolation curve at the connection points, thereby realistically restoring the smooth physical characteristics of water level changes. Each 10-millisecond time point on the primary reference clock is substituted into the constructed spline function to calculate the corresponding interpolated water level data, thus reconstructing the sparse low-frequency signal into a dense high-frequency signal.
[0070] The cubic B-spline interpolation function mentioned above is a smooth curve fitting tool commonly used in mathematical numerical analysis. It approximates discrete data points through a piecewise defined polynomial function and maintains the continuity of the second derivative at the connection points, which can effectively avoid Runge's phenomenon and truly reflect the continuous change trend of physical quantities.
[0071] After time alignment, the data stream enters the spatial mapping stage. The 3D finite element mesh topology file of the hydraulic engineering digital twin is retrieved, containing 150,000 tetrahedral elements and 28,000 node coordinates. Using the KD-Tree nearest neighbor search algorithm, the nearest element in the virtual mesh is found based on the physical sensor's installation coordinates (x, y, z). For example, for a piezometer P1 (coordinates: 150.5, 20.0, 5.0) installed at the dam heel, the algorithm locates it inside tetrahedral element E4502. Then, isoparametric element shape functions are called to calculate the local natural coordinates of point P1 relative to the four vertices of element E4502. Through shape function matrix operations, the measured seepage pressure value of P1 is weighted and distributed to the four nodes of this element. For areas without sensors, the Kriging spatial interpolation algorithm is used to extrapolate the initial state values of unknown nodes based on the variogram model of known points. Finally, the initial values of water level load, temperature boundary conditions and measured physical quantities on all grid nodes are assembled strictly according to the degree of freedom order required by the finite element solver to generate a current state input matrix with a dimension of 84000×1 (assuming 3 degrees of freedom for each node).
[0072] To verify the effectiveness of data preprocessing, measured data from a heavy rainfall event were selected for testing. Table 1 shows a comparison of the errors of cubic B-spline interpolation using this scheme and traditional linear interpolation in recovering missing data. Experiments show that cubic B-spline interpolation reduces the root mean square error (RMSE) by more than 85%, significantly preserving the fluctuation characteristics of the signal, laying the foundation for subsequent high-precision simulations.
[0073] Table 1 Comparison of Data Interpolation Algorithm Accuracy
[0074] Linear interpolation 12.5 15.8 28 Lagrange interpolation 8.4 10.2 34 Cubic B-spline interpolation (this scheme) 1.8 2.3 48
[0075] As shown in Table 1, the cubic B-spline interpolation method used in this scheme has a significantly better error index than traditional methods when processing non-uniform sampling data, and can accurately reproduce subtle changes in environmental load.
[0076] Please see Figure 1 and Figure 3The twin model dynamic calibration module is used to import the current state input matrix into a pre-built high-fidelity hydraulic finite element model, solve the theoretical physical field distribution data using the physical control equation, construct a loss function with the convergence objective of minimizing the weighted Euclidean distance between the key node response data and the theoretical physical field distribution data through the gradient descent optimization algorithm, and solve the permeability coefficient and elastic modulus parameters of the model based on the back-inverse iteration of the loss function, and output the constitutive parameter correction vector.
[0077] The twin model dynamic calibration module specifically includes:
[0078] The forward solution submodule of the physics field is used to call the finite element kernel solver to load the current state input matrix as boundary conditions and body loads, solve the partial differential equations controlling seepage and stress field based on the current material parameters, and calculate the theoretical response values at the corresponding nodes of the sensor to generate theoretical physics field distribution data.
[0079] The residual objective construction submodule is used to calculate the deviation vector between the response data of key nodes and the distribution data of theoretical physical fields. Based on the sensor accuracy and spatial importance, it assigns confidence weights to different monitoring points and constructs a scalar loss function that aims to minimize the weighted Euclidean distance and includes a regularization term.
[0080] The parameter inversion optimization submodule is used to calculate the sensitivity Jacobian matrix of the loss function with respect to material properties using the gradient descent optimization algorithm, iteratively update the permeability coefficient and elastic modulus along the negative gradient direction until the loss value is lower than the preset convergence threshold, and output the constitutive parameter correction vector.
[0081] The process of constructing the loss function in the residual objective construction submodule specifically includes:
[0082] Obtain the theoretical physical field distribution data and the real-time monitoring key node response data under the current iteration step, determine the basic weight coefficient of each monitoring point based on the reciprocal of the measurement standard deviation of the sensors at each monitoring point, and generate a weighted matrix by combining the structural vulnerability factor of the area where the monitoring point is located.
[0083] By introducing a Tikhonov regularization term to constrain the smoothness of the parameter space, and combining it with the weighting matrix, a loss function expression of the following form is constructed:
[0084] ;
[0085] in, Represents the value of the loss function. This represents the constitutive parameter correction vector to be optimized. Represents the total number of monitoring points. Representing the The overall weight of each monitoring point Representing the Key node response data at each point This represents the corresponding theoretical physical field distribution data. Represents the regularization coefficient. This represents the prior parameter vector from the previous time step.
[0086] The reverse iterative solution process executed by the parameter inversion optimization submodule specifically includes:
[0087] Based on the adjoint variable method or the finite difference method, the gradient vector of the loss function with respect to constitutive parameters such as permeability coefficient and elastic modulus is calculated, and the inverse matrix of the Hessian matrix is approximated by the BFGS formula of the quasi-Newton method to correct the search direction.
[0088] Perform a line search operation along the corrected search direction to determine the optimal step size that allows the loss function to decrease sufficiently. Update the current constitutive parameter estimate accordingly and check whether the updated parameter is within the preset constraint range that conforms to physical meaning.
[0089] Determine whether the decrease in loss function or the update of parameters in the current iteration step is less than the preset convergence tolerance. If not, feed the updated parameters back to the forward physics submodule for the next round of calculation. If the condition is met, terminate the iteration and output the constitutive parameter correction vector.
[0090] The current state input matrix is used as boundary conditions and initial loads, and loaded into the finite element calculation kernel based on porous media fluid-structure interaction theory. The kernel solver constructs a system of partial differential equations based on Darcy's Law and Hooke's Law, and utilizes the parallel computing capabilities of the core processor (CPU) to assemble and solve the stiffness matrix of 150,000 elements. This forward solution process outputs theoretical distribution data for the entire field, from which theoretical seepage pressure and displacement values for 50 key nodes precisely corresponding to the locations of on-site sensors are extracted and denoted as theoretical physical field distribution data.
[0091] To eliminate the deviation between the model parameters and the actual physical state, a loss function is constructed that includes data residuals and a regularization term. The construction logic of this loss function is as follows: ;
[0092] in, This represents the calculated value of the loss function; Represents the total number of monitoring points; Representing the The weighting coefficients for each monitoring point are used to balance the measurement uncertainties of different types of sensors; Representing the Measured key node response data at each point; Represents based on current parameters The calculated distribution data of theoretical physical fields; This represents the regularization coefficient, used to constrain the smoothness of parameters; This represents the normalized constitutive parameter correction vector to be inverted; This represents the normalized prior parameter vector from the previous time step.
[0093] Before performing specific loss function calculations, the physical parameters must be dimensionless to eliminate the permeability coefficient (order of magnitude). ) and elastic modulus (order of magnitude) The permeability coefficient has a significant difference in dimensions between the two. In this embodiment, the permeability coefficient is divided by... Divide the elastic modulus by So that all their values fall into Within the interval. Assuming at the current moment, for monitoring point 1 (high-precision piezometer), its weight... Set to 100; measured seepage pressure value at this point. The value is 50.2 kPa, while the model's forward calculation value is... The value is 48.5 kPa. Substituting this value into the first term (residual term) of the loss function, the calculation is as follows: If the regularization term is calculated to be 0.001 at this point (assuming the parameter variation is minimal), then the total loss function value for this step is... It is approximately 144.501. This value precisely quantifies the state deviation between the current digital twin model and the physical engineering.
[0094] Efficiently calculate the loss function using the adjoint method. Regarding parameter vectors The gradient is calculated. The gradient direction indicates the steepest descent direction for parameter adjustment. A BFGS quasi-Newton optimization algorithm is employed, utilizing historical gradient information to approximate the inverse of the Hessian matrix, thus obtaining an accurate search direction without directly calculating the second derivative. After determining the direction, a Wolfe line search strategy is executed to find the step size that results in the fastest decrease in the loss function value.
[0095] The aforementioned BFGS quasi-Newton optimization algorithm is an iterative algorithm widely used in unconstrained nonlinear optimization problems. Its core lies in iteratively updating the inverse of the approximate Hessian matrix, thereby avoiding the huge amount of computation required to directly calculate the second derivative. It has the characteristics of fast convergence speed and high numerical stability.
[0096] The convergence criterion is set as the rate of change of the loss function value being less than... Or the parameter update amount is less than 1%. After about 15 to 20 iterations, the algorithm converges. For example, the normalized permeability coefficient changes from the initial 1.0 (corresponding to...) The value was revised to 1.25 (corresponding to...). The normalized elastic modulus was corrected from 28.0 (corresponding to 28.0 GPa) to 27.4 (corresponding to 27.4 GPa). These two corrected parameters reduce the average error between the finite element model output and the measured data to within 0.5%. This set of calibrated parameters is then denormalized and encapsulated into a constitutive parameter correction vector, serving as feature information reflecting the true physical properties of the current structure.
[0097] Experimental data show that, even with 5% Gaussian white noise, the parameter error obtained by this method is still controlled within 3%, while the error of the traditional least squares method without regularization is as high as 15%. This result proves the effectiveness of introducing weighted residuals and Tikhonov regularization terms.
[0098] Please see Figure 1 and Figure 4 The structural health diagnosis module is used to extract the time series features of the constitutive parameter correction vector, construct the parameter evolution phase space trajectory of the constitutive parameter correction vector in the multidimensional phase space, calculate the evolution rate magnitude in the tangent direction of the parameter evolution phase space trajectory, compare the evolution rate magnitude with the preset nonlinear safety threshold surface in dynamic spatial position, determine the structural state based on the comparison results, and generate a structural damage location report when the determination result exceeds the limit.
[0099] The structural health diagnosis module specifically includes:
[0100] The phase space reconstruction analysis submodule is used to apply the time delay embedding theorem to the time series of constitutive parameter correction vectors, calculate the optimal delay time using the mutual information method and determine the embedding dimension using the spurious neighbor method, and construct a high-dimensional state vector to form the parameter evolution phase space trajectory.
[0101] The evolution feature calculation submodule is used to perform differential geometric analysis on the reconstructed parameter evolution phase space trajectory, calculate the tangential vector of the trajectory point at the current time, and solve the Euclidean norm of the tangential vector to quantify the dynamic change rate of the system state and generate the evolution rate modulus.
[0102] The state deviation determination submodule is used to map the calculated evolution rate magnitude to the domain of a pre-trained nonlinear safety threshold surface, determine whether the current state point is located within the safe area enclosed by the nonlinear safety threshold surface, and trigger an alarm mechanism to generate a structural damage location report when the state point crosses the boundary.
[0103] The process of constructing the parameter evolution phase space trajectory by the phase space reconstruction analysis submodule specifically includes:
[0104] Extract the key components characterizing structural stiffness and permeability from the constitutive parameter correction vector, calculate the autocorrelation function of the normalized time series, and select the component whose autocorrelation function first decreases to its initial value. The time length at which the delay point is taken as the optimal delay time;
[0105] Calculate the proportion of false neighbors that changes with the increase of the embedding dimension, and determine the dimension when the proportion converges to zero as the minimum embedding dimension. Construct a high-dimensional delay coordinate vector containing historical state information based on the optimal delay time and the minimum embedding dimension.
[0106] By connecting all the constructed high-dimensional delayed coordinate vectors in chronological order, a continuous geometric curve describing the evolution path of the system's dynamic behavior is generated in the multidimensional phase space, thus establishing the parameter evolution phase space trajectory.
[0107] The construction process of the nonlinear safety threshold surface in the state deviation determination submodule specifically includes:
[0108] Collect the historical constitutive parameter correction vectors generated during the healthy operation cycle of water conservancy projects, and use the phase space reconstruction analysis submodule to transform them into a set of historical healthy phase space trajectory points;
[0109] A single-class support vector machine algorithm is used to learn the distribution pattern of historical health phase space trajectory points in a multi-dimensional feature space. A Gaussian kernel function is used to map the data to a higher-dimensional space to find the optimal hyperplane and establish the nonlinear decision boundary that encloses all health state points.
[0110] The nonlinear decision boundary is inversely mapped back to the original phase space to generate a closed surface, which is then solidified into a nonlinear safety threshold surface for real-time online monitoring and comparison.
[0111] The system continuously receives the constitutive parameter correction vector (such as elastic modulus E and permeability coefficient K) output from step S2 and stores it as time series data. To capture the nonlinear dynamic characteristics of structural performance degradation, phase space reconstruction is performed on this time series. First, the optimal time delay is determined using the mutual information method. Calculate the sequence and The mutual information is selected by choosing the delay time when the mutual information first reaches its minimum value, for example, calculated as follows: (Corresponding to 5 sampling periods). Next, the minimum embedding dimension is determined using the False Nearest Neighbors (FNN) method. As the dimension increases, when the proportion of false neighbors drops below 1%, the embedding dimension is determined to be 4. Based on this, a high-dimensional state vector is constructed. Connecting the state vectors at consecutive moments in four-dimensional space forms the parameter evolution phase space trajectory that describes the system's evolution.
[0112] For this trajectory, its tangential vector at the current moment is calculated, and its Euclidean norm, i.e., the evolution rate modulus, is obtained. This modulus physically represents the "rate" of structural parameter degradation. For example, during normal aging, the elastic modulus decays extremely slowly, and the calculated evolution rate modulus may remain at a low level of 0.001 / day; however, when sudden damage (such as crack propagation) occurs, the parameters will change drastically, causing the evolution rate modulus to soar to over 0.5 / day.
[0113] The judgment process no longer relies on a single fixed threshold, but instead compares against a pre-trained nonlinear safety threshold surface. This surface is trained using a One-Class Support Vector Machine (SVM) algorithm based on massive parameter data from the past 5 years of healthy operating conditions. The SVM uses a Gaussian kernel function to map the data to a high-dimensional feature space, searching for a hyperspherical boundary that can compactly enclose all healthy sample points.
[0114] The aforementioned One-Class SVM algorithm is an unsupervised learning algorithm mainly used for anomaly detection. Its principle is to map the target class data to a high-dimensional feature space and find a hyperplane or hypersphere in this space to separate the target data from the origin, thereby defining the boundary range of normal data.
[0115] The currently calculated state points (including parameter values and their evolution rates) are mapped to this feature space. If a point falls inside the hypersphere, the structure is considered to be in a "healthy" state; if a point falls outside the hypersphere, its Mahalanobis distance to the boundary is calculated as a quantitative indicator of the degree of damage. For example, if, during a monitoring session, the distance of a state point from the boundary exceeds 3.0 (3 standard deviations), the system immediately classifies it as "abnormal".
[0116] Once a limit is exceeded, the system automatically triggers damage location logic. By tracing back to the sensitive areas that caused the abnormal parameter changes, the specific grid cell number is identified. For example, if an abnormal increase in the permeability coefficient of a cell at an elevation of 200 meters on the left bank of the dam is found, the system automatically generates a structural damage location report containing the damage location (coordinates, cell number), the type of abnormal parameter (sudden increase in permeability coefficient), the severity (severe), and recommended measures. Within 500 milliseconds, an alarm is pushed to the monitoring screen at the dam safety management center and the handheld terminals of relevant personnel via a dedicated 4G / 5G industrial router.
[0117] Table 2 lists the comparative test results of this scheme and the traditional threshold method in damage diagnosis. In 100 simulated micro-crack damage scenarios, the detection rate of this scheme reached 98%, and the average alarm lead time was 120 hours earlier than the traditional method, which fully verifies the sensitivity of the dynamic diagnostic technology based on phase space trajectory.
[0118] Table 2 Comparison of Performance of Damage Diagnosis Methods
[0119] Fixed threshold method 15.2 22.5 12 >20 Statistical process control (SPC) 8.5 12.0 48 10-20 This scheme (phase space SVM) 1.2 2.0 120 <5
[0120] As shown in Table 2, this solution is significantly better than existing technologies in terms of both false alarm rate and false negative rate, and can detect potential hazards much earlier, thus having extremely high engineering application value.
[0121] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.
Claims
1. A smart monitoring and diagnostic system for water conservancy projects based on digital twins, characterized in that, include: The multidimensional mapping perception module is used to synchronously collect environmental load driving data and key node response data at the site of the water conservancy project. Based on the timestamp, spline interpolation alignment is performed on the heterogeneous environmental load driving data and key node response data, and the aligned data is mapped to the three-dimensional spatial grid nodes of the virtual model to construct the current state input matrix. The twin model dynamic calibration module is used to import the current state input matrix into a pre-constructed high-fidelity hydraulic finite element model, solve the theoretical physical field distribution data using the physical control equation, construct a loss function with the convergence objective of minimizing the weighted Euclidean distance between the key node response data and the theoretical physical field distribution data through the gradient descent optimization algorithm, solve the permeability coefficient and elastic modulus parameters of the model based on the back-inverse iteration of the loss function, and output the constitutive parameter correction vector. The structural health diagnosis module is used to extract the time series features of the constitutive parameter correction vector, construct the parameter evolution phase space trajectory of the constitutive parameter correction vector in the multidimensional phase space, calculate the evolution rate magnitude in the tangent direction of the parameter evolution phase space trajectory, compare the evolution rate magnitude with the preset nonlinear safety threshold surface in dynamic spatial position, determine the structural state based on the comparison result, and generate a structural damage location report when the determination result exceeds the limit.
2. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 1, characterized in that, The multidimensional mapping sensing module specifically includes: The source data aggregation submodule is used to build a heterogeneous sensor network interface to receive environmental variables, including upstream water level and rainfall intensity, in real time, and simultaneously acquire structural response indicators, including dam seepage pressure and displacement. It also uses a network time protocol server to stamp all acquired raw data packets with a unified high-precision timestamp. The timing synchronization alignment submodule is used to analyze the sampling frequency difference between the environmental load driving data and the key node response data, reconstruct the continuous change curve of sparse low-frequency data using cubic B-spline interpolation algorithm, and resample it onto the dense time axis of high-frequency data to generate a synchronized time series dataset. The spatial grid mapping submodule is used to establish the coordinate transformation matrix between the physical sensor positions and the virtual finite element grid nodes. It uses shape function interpolation to project the synchronized time series data onto the corresponding nodes of the three-dimensional grid and assembles the spatial distribution vector to generate the current state input matrix.
3. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 1, characterized in that, The twin model dynamic calibration module specifically includes: The forward solution submodule of the physical field is used to call the finite element kernel solver to load the current state input matrix as boundary conditions and body loads, solve the partial differential equations controlling seepage and stress field based on the current material parameters, and calculate the theoretical response value at the corresponding node of the sensor to generate the theoretical physical field distribution data. The residual objective construction submodule is used to calculate the deviation vector between the key node response data and the theoretical physical field distribution data, assign confidence weights to different monitoring points according to sensor accuracy and spatial importance, and construct a scalar loss function that aims to minimize the weighted Euclidean distance and includes a regularization term. The parameter inversion optimization submodule is used to calculate the sensitivity Jacobian matrix of the loss function with respect to material properties using the gradient descent optimization algorithm, iteratively update the permeability coefficient and elastic modulus along the negative gradient direction until the loss value is lower than the preset convergence threshold, and output the constitutive parameter correction vector.
4. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 1, characterized in that, The structural health diagnosis module specifically includes: The phase space reconstruction analysis submodule is used to apply the time delay embedding theorem to the time series of the constitutive parameter correction vector, calculate the optimal delay time using the mutual information method and determine the embedding dimension using the spurious neighbor method, and construct a high-dimensional state vector to form the phase space trajectory of the parameter evolution. The evolution feature calculation submodule is used to perform differential geometric analysis on the reconstructed parameter evolution phase space trajectory, calculate the tangential vector of the trajectory point at the current time, and solve the Euclidean norm of the tangential vector to quantify the dynamic change rate of the system state and generate the evolution rate modulus. The state deviation determination submodule is used to map the calculated evolution rate magnitude to the domain of the pre-trained nonlinear safety threshold surface, determine whether the current state point is located within the safe area enclosed by the nonlinear safety threshold surface, and trigger an alarm mechanism when the state point crosses the boundary to generate the structural damage location report.
5. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 2, characterized in that, The data spatiotemporal alignment operation performed by the timing synchronization and alignment submodule specifically includes: Acquire the environmental load driving data and the key node response data with original timestamps, identify the signal stream with the highest sampling frequency as the main reference clock reference, and calculate the time lag and frequency ratio of other heterogeneous data streams relative to the main reference clock reference. A sliding window mechanism is used to extract local data segments. Within each window, a cubic spline interpolation function that can maintain the continuity of the first and second derivatives of the data is constructed. The interpolation function is resampled and calculated based on the discrete time points of the main reference clock to fill the information gaps caused by low-frequency sampling. Outlier detection and smoothing are performed on all resampled data streams to remove outlier noise caused by sensor malfunctions. The processed multidimensional data are then merged in time-aligned order to generate a synchronous monitoring sequence.
6. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 2, characterized in that, The process by which the spatial grid mapping submodule establishes the current state input matrix specifically includes: Read the three-dimensional mesh topology data of the virtual model and the actual installation coordinates of the physical sensors, and use the nearest neighbor search algorithm to identify the cell to which each sensor belongs and the number of its adjacent nodes in the mesh space; Based on the finite element isoparametric element theory, the local coordinates of the sensor position relative to the node of the element are calculated, and an interpolation mapping relationship including shape function weights is constructed to allocate the physical quantity values measured by the sensor to the corresponding mesh node degrees of freedom. Data mapping is completed by traversing all monitoring points. For areas where no sensors are deployed, Kriging interpolation is used to deduce node values. Finally, the current state input matrix is generated according to the input format specifications of the finite element solver.
7. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 3, characterized in that, The process of constructing the loss function in the residual objective construction submodule specifically includes: The theoretical physical field distribution data and the real-time monitoring response data of the key nodes are obtained in the current iteration step. The basic weight coefficient of each monitoring point is determined according to the reciprocal of the measurement standard deviation of the sensors at each monitoring point, and a weighted matrix is generated by combining the structural vulnerability factor of the area where the monitoring point is located. By introducing a Tikhonov regularization term to constrain the smoothness of the parameter space, and combining it with the weighting matrix, a loss function expression of the following form is constructed: ； in, Represents the value of the loss function. This represents the constitutive parameter correction vector to be optimized. Represents the total number of monitoring points. Representing the The overall weight of each monitoring point Representing the The key node response data of each point. This represents the corresponding theoretical physical field distribution data. Represents the regularization coefficient. This represents the prior parameter vector from the previous time step.
8. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 3, characterized in that, The reverse iterative solution process executed by the parameter inversion optimization submodule specifically includes: Based on the adjoint variable method or the finite difference method, the gradient vector of the loss function with respect to constitutive parameters such as permeability coefficient and elastic modulus is calculated, and the inverse matrix of the Hessian matrix is approximated by the BFGS formula of the quasi-Newton method to correct the search direction. Perform a line search operation along the corrected search direction to determine the optimal step size that allows the loss function to decrease sufficiently. Update the current constitutive parameter estimate accordingly and check whether the updated parameter is within the preset constraint range that conforms to physical meaning. Determine whether the decrease in loss function or the update of parameters in the current iteration step is less than the preset convergence tolerance. If not, feed the updated parameters back to the forward solution submodule of the physical field for the next round of calculation. If the condition is met, terminate the iteration and output the constitutive parameter correction vector.
9. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 4, characterized in that, The process by which the phase space reconstruction analysis submodule constructs the parameter evolution phase space trajectory specifically includes: Extract the key components characterizing structural stiffness and permeability from the constitutive parameter correction vector, calculate the autocorrelation function of the normalized time series, and select the component whose autocorrelation function first decreases to its initial value. The time length at which the delay point is taken as the optimal delay time; Calculate the proportion of false neighbor points that changes with the increase of the embedding dimension, and determine the dimension when the proportion converges to zero as the minimum embedding dimension. Construct a high-dimensional delay coordinate vector containing historical state information based on the optimal delay time and the minimum embedding dimension. By connecting all the constructed high-dimensional delayed coordinate vectors in chronological order, a continuous geometric curve describing the evolution path of the system's dynamic behavior is generated in the multidimensional phase space, thus establishing the phase space trajectory of the parameter evolution.
10. The intelligent monitoring and diagnosis system for water conservancy projects based on digital twins according to claim 4, characterized in that, The construction process of the nonlinear safety threshold surface in the state deviation determination submodule specifically includes: Collect the historical constitutive parameter correction vectors generated during the healthy operation cycle of the water conservancy project, and use the phase space reconstruction analysis submodule to transform them into a set of historical healthy phase space trajectory points; The distribution pattern of the historical health phase space trajectory points set in the multidimensional feature space is learned by using a single-class support vector machine algorithm. The data is then mapped to a higher-dimensional space using a Gaussian kernel function to find the optimal hyperplane and establish the nonlinear decision boundary that encloses all health state points. The nonlinear decision boundary is inversely mapped back to the original phase space to generate a closed surface, and the closed surface is solidified into the nonlinear safety threshold surface for real-time online monitoring and comparison.