Long-term dam working behavior sensing method based on deformation-stress mapping analysis

Abstract

This invention relates to the field of hydraulic engineering technology and provides a long-term perception method for the operational performance of dams based on deformation-stress mapping analysis. The method includes: acquiring target environmental data of the dam in real-time monitoring; inputting the target environmental data into a deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model; the deep learning model serves as a student model, trained through knowledge transfer from a teacher model based on target environmental data samples and their corresponding three-dimensional stress field labels; the teacher model includes a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model. This invention trains the deep learning model (serving as a student model) using a teacher model comprising four mapping models, enabling the deep learning model to not only accurately characterize the relationship between deformation and stress but also capture the nonlinear correlation between target environmental data, deformation, and the three-dimensional stress field, achieving online and accurate long-term perception of dam stress.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering technology, and in particular to a long-term sensing method for the operational performance of dams based on deformation-stress mapping analysis. Background Technology

[0002] As a core structure in water conservancy projects, gravity dams directly affect the safety of the reservoir area and the lives and property of people downstream. Deformation and stress are key indicators reflecting the dam's performance. Deformation monitoring characterizes the overall or local morphological changes of the dam body, while stress monitoring directly reflects the mechanical state of the dam materials. The combined analysis of these two methods is the core basis for dam safety assessment and risk warning. However, dams are typically designed for a lifespan of 100 to 200 years, while internal stress and / or strain monitoring instruments are mostly buried and have a service life of only 30 to 50 years, far shorter than the dam's design life. Once these instruments fail, internal stress information cannot be directly obtained, severely impacting the continuity and accuracy of dam safety assessments.

[0003] In recent years, high-precision Global Navigation Satellite System (GNSS) and laser ranging technologies for monitoring external deformation have become increasingly mature, and external instruments (instruments for external observation) are easy to replace, making it a feasible way to infer stress from deformation data for long-term perception of the internal stress state of dams.

[0004] In existing technologies, although the finite element inversion method can deduce stress distribution from deformation data, there is a complex nonlinear mapping relationship between deformation and stress. It is affected by multiple factors such as dam structure characteristics, load boundary conditions, material properties, temperature and seepage, making it difficult for traditional deduction methods to accurately characterize the relationship between deformation and stress. Moreover, under actual conditions such as long-term operation, complex loads and material deterioration, it is easily affected by the uncertainty of finite element model parameters and boundary condition errors, resulting in insufficient calculation accuracy and long calculation time, which cannot meet the requirements of real-time performance and long-term effectiveness. Summary of the Invention

[0005] This invention provides a long-term sensing method for dam performance based on deformation-stress mapping analysis, which solves the problems that traditional methods of inferring stress distribution from deformation data are difficult to accurately characterize the relationship between deformation and stress, and cannot meet the requirements of real-time and long-term stress estimation.

[0006] This invention provides a long-term sensing method for the operational performance of dams based on deformation-stress mapping analysis, comprising the following steps: Acquire target environmental data for real-time monitoring of the dam; The target environment data is input into a deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model. The deep learning model is used as a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes: a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model. The first mapping model was trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam; The second mapping model is trained based on the measured deformation samples of each monitoring point and the second deformation labels of each representative node in the dam; The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model. The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

[0007] According to the present invention, a long-term perception method for dam performance based on deformation-stress mapping analysis is provided, wherein the training method of the first mapping model is as follows: The target environment data sample is input into the first mapping model to obtain the predicted deformation data of each monitoring point output by the first mapping model. The predicted deformation data and the first deformation label are substituted into the first loss function. When the first loss function converges, the training of the first mapping model is completed.

[0008] According to the present invention, a long-term sensing method for the operational performance of a dam based on deformation-stress mapping analysis is provided, wherein the target environmental data samples are acquired in the following manner: Acquire multiple sets of source environment data samples and the measured deformation data and measured stress data of each monitoring point in the dam corresponding to each set of source environment data samples, and use the measured deformation data as the first deformation label; Calculate the deformation correlation coefficient and stress correlation coefficient between each type of environmental data in the source environmental data sample and the measured deformation data and measured stress data, respectively, and select the environmental data whose deformation correlation coefficient and stress correlation coefficient are both greater than the correlation coefficient threshold as the target environmental data sample.

[0009] According to the present invention, a long-term sensing method for dam performance based on deformation-stress mapping analysis is provided, wherein the training method of the second mapping model is as follows: The second deformation labels of each representative node in the dam are obtained by inversion based on the pre-constructed dam simulation model. The measured deformation samples at the monitoring points are input into the second mapping model to obtain the second deformation prediction data output by the second mapping model. The second deformation prediction data and the second deformation label are substituted into the second loss function. When the second loss function converges, the training of the second mapping model is completed.

[0010] According to the present invention, a long-term sensing method for dam performance based on deformation-stress mapping analysis is provided, and the representative node selection method is as follows: Based on the inversion of the pre-built dam simulation model, the overall deformation calculation results and stress calculation results of the dam under multiple working conditions are obtained; Based on the physical characteristics of each region of the dam, the mesh of the dam simulation model whose deformation calculation results are greater than or equal to the deformation threshold and whose stress calculation results are greater than or equal to the stress threshold are selected as the first candidate node set. Principal component analysis is used to extract the core spatiotemporal features of the deformation field characterized by deformation calculation results. Similar deformation mode regions are divided by spectral clustering, and a predetermined number of grids are selected as the second candidate node set in each similar deformation mode region. The nodes in the first candidate node set and the second candidate node set are determined as the representative nodes.

[0011] According to the present invention, a long-term perception method for dam performance based on deformation-stress mapping analysis is provided, wherein the training method of the third mapping model is as follows: Deformation samples and stress labels of representative nodes in the dam are obtained by inversion based on the pre-built dam simulation model; The deformation samples of each representative node are input into the third mapping model to obtain the predicted stress data of each representative node output by the third mapping model; the predicted stress data and the stress labels are substituted into the third loss function, and the training of the third mapping model is completed when the third loss function converges.

[0012] According to the present invention, a long-term perception method for dam performance based on deformation-stress mapping analysis is provided, wherein the training method of the fourth mapping model is as follows: Stress samples of representative nodes in the dam and three-dimensional stress field labels of the dam are obtained by inversion based on the pre-built dam simulation model. The stress sample is input into the fourth mapping model to obtain the predicted stress field output by the fourth mapping model. The predicted stress field and the three-dimensional stress field label are substituted into the fourth loss function. When the fourth loss function converges, the training of the fourth mapping model is completed.

[0013] According to the present invention, a long-term perception method for dam operational performance based on deformation-stress mapping analysis is provided, wherein the dam simulation model is constructed as follows: A three-dimensional finite element model including the main body of the dam, the dam heel, and the dam toe was constructed using finite element software. The mesh element size of the key parts of the dam heel, dam toe, and orifice area was set to 1 / 5 to 1 / 3 of the mesh element size of the main body of the dam. Based on the source environment data samples and corresponding deformations and stresses from multiple sets of historical monitoring under different working conditions, the key parameters of the finite element model are inverted. Under the conditions of the key parameters and the source environment data sample, if the deviations between the simulation results of deformation and stress obtained from the three-dimensional finite element model and their respective measured results are less than or equal to a preset deviation threshold, the three-dimensional finite element model is determined to be the dam simulation model.

[0014] The present invention also provides a long-term sensing device for the operational performance of a dam based on deformation-stress mapping analysis, comprising the following modules: The environmental data acquisition module is used to acquire target environmental data for real-time monitoring of the dam. The model execution module is used to input the target environment data into the deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model. The deep learning model is used as a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes: a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model. The first mapping model was trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam; The second mapping model is trained based on the measured deformation samples of each monitoring point and the second deformation labels of each representative node in the dam; The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model. The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the long-term sensing method for dam performance based on deformation-stress mapping analysis as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the long-term sensing method for dam performance based on deformation-stress mapping analysis as described above.

[0017] The present invention provides a long-term perception method for dam performance based on deformation-stress mapping analysis. By using a teacher model including the four mapping models mentioned above to train a deep learning model as a student model through knowledge transfer, a deep learning model is obtained that can map target environmental data to the three-dimensional stress field of the dam. This allows the deep learning model to not only accurately characterize the relationship between deformation and stress, but also capture the nonlinear correlation between target environmental data, deformation, and the three-dimensional stress field. Thus, it can accurately predict the three-dimensional stress field of the dam based on real-time monitored target environmental data. Moreover, the deep learning model only needs real-time monitored target environmental data and no longer depends on various internal sensing devices, thereby realizing online and long-term prediction of dam stress. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the long-term sensing method for dam performance based on deformation-stress mapping analysis provided by the present invention.

[0020] Figure 2 This is a structural diagram of the dam simulation model in the long-term perception method of dam working performance based on deformation-stress mapping analysis provided by the present invention.

[0021] Figure 3 This is a simulation result diagram of the dam body temperature inversion in the dam simulation model of the dam simulation model in the long-term sensing method of dam working performance based on deformation-stress mapping analysis provided by the present invention.

[0022] Figure 4 This is a simulation result diagram of stress inversion of the dam simulation model in the long-term perception method of dam working performance based on deformation-stress mapping analysis provided by this invention.

[0023] Figure 5 This is a schematic diagram of representative node selection in the long-term sensing method for dam performance based on deformation-stress mapping analysis provided by the present invention.

[0024] Figure 6This is a comparison diagram of the stress process line derived from the known deformation of representative nodes and the stress process line calculated by the dam simulation model in the long-term perception method of dam working performance based on deformation-stress mapping analysis provided by this invention.

[0025] Figure 7 The figure shows a comparison between the simulated three-dimensional stress field and the three-dimensional stress field derived by the deep learning model in the long-term perception method of dam working performance based on deformation-stress mapping analysis provided by the present invention. (a) is a schematic diagram of the simulated three-dimensional stress field, and (b) is a schematic diagram of the three-dimensional stress field derived by the deep learning model.

[0026] Figure 8 This is a schematic diagram of the structure of the long-term sensing device for dam performance based on deformation-stress mapping analysis provided by the present invention.

[0027] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] The long-term sensing method for dam performance based on deformation-stress mapping analysis, as described in this invention, is as follows: Figure 1 As shown, it includes the following steps S110 and S120.

[0030] Step S110: Obtain target environmental data for real-time monitoring of the dam. For example, the target environmental data includes: air temperature and reservoir water level, which can be obtained in real-time through corresponding temperature sensors and water level sensors.

[0031] Step S120: Input the target environment data into the deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model. The deep learning model serves as a student model, trained through knowledge transfer from the teacher model based on the target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes: a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model.

[0032] The first mapping model is trained based on target environmental data samples and the first deformation labels of each monitoring point in the dam. The monitoring points are the locations of pre-embedded deformation and stress sensors within the dam body. Using the first mapping model, the first deformation data of each monitoring point in the dam body can be predicted based on real-time monitored target environmental data.

[0033] The second mapping model is trained based on the measured deformation samples from each monitoring point and the second deformation labels of each representative node in the dam. The representative nodes are pre-selected nodes within the dam; that is, the representative nodes correspond to the key stress-bearing areas of the dam. The second mapping model can predict the second deformation data of each representative node based on the real-time deformation data monitored by each monitoring point, or predict the second deformation data based on the first deformation data predicted by the first mapping model. Through the representative nodes and the second mapping model, the sparse monitoring point data is extended to the global stress field, overcoming the limitation of uneven monitoring point distribution.

[0034] The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-constructed dam simulation model. The stress data of each representative node can be predicted using the third mapping model based on the second deformation data predicted by the second mapping model.

[0035] The fourth mapping model is trained based on stress samples and three-dimensional stress field labels of each representative node. These stress samples and labels are derived from a pre-constructed dam simulation model. The fourth mapping model can predict the overall three-dimensional stress field of the dam based on the stress data of each representative node predicted by the third mapping model.

[0036] The core role of the teacher model is to output highly reliable three-dimensional stress field and intermediate process knowledge (such as characteristic distribution of representative node deformation and representative node stress), providing dual learning basis for the student model.

[0037] Four mapping models constitute a complete teacher link from target environment data to the 3D stress field: target environment data → (first mapping model) first deformation data of monitoring points → (second mapping model) second deformation data of representative nodes → (third mapping model) stress data of representative nodes → (fourth mapping model) 3D stress field. That is, the teacher model is a model formed by the overall link of the four mapping models and can be pre-trained. The final output of this link, the 3D stress field, is completely consistent with the target 3D stress field directly predicted by the student model, and can be used as a hard label for distillation; at the same time, the intermediate links in the link (such as the stress characteristics of representative nodes predicted by the third mapping model) can be used as soft labels, allowing the student model to learn the teacher's reasoning logic.

[0038] When training the student model, the teacher model can be frozen, with soft labels q. teacher As parameters of the distillation loss function, the hard labels are used as parameters of the student loss function. In the distillation loss function, the output of the student model (which can be softened by temperature T) should be as close as possible to the soft target distribution of the teacher model. This is commonly measured by KL divergence or cross-entropy. Therefore, the distillation loss function is: L soft =T 2 ·KL(q teacher ||q student ).

[0039] q student This represents the predicted three-dimensional stress field of the student model, and the student loss function is the output of the student model (p). student ) and real labels (y true Cross-entropy loss: L hard =CE(y true ,p student ).

[0040] The student model training is complete when the total loss function converges, where the total loss function is: L=αL soft +(1-α)L hard .

[0041] Where α is a hyperparameter, which can take the value 0.7.

[0042] The learning objective of the student model is to skip the intermediate steps of the teacher's multi-step reasoning and directly learn the mapping relationship between environmental data and three-dimensional stress field by distillation. However, the knowledge support behind it comes from the high-precision output (three-dimensional stress field) and intermediate features of the teacher model's complete chain.

[0043] This embodiment of the long-term perception method for dam performance based on deformation-stress mapping analysis trains a deep learning model (which serves as a student model) with a teacher model that includes the four mapping models mentioned above. This results in a deep learning model that can map target environmental data to the dam's three-dimensional stress field. This deep learning model can not only accurately characterize the relationship between deformation and stress, but also capture the nonlinear correlation between target environmental data, deformation, and the three-dimensional stress field. Thus, it can accurately predict the dam's three-dimensional stress field based on real-time monitored target environmental data. Moreover, the deep learning model only requires real-time monitored target environmental data and no longer depends on various internal sensing devices. Therefore, it achieves online and long-term prediction of dam stress.

[0044] In some embodiments, the training method of the first mapping model is as follows: input the target environment data sample into the first mapping model to obtain the predicted deformation data of each monitoring point output by the first mapping model, substitute the predicted deformation data and the first deformation label into the first loss function, and the first mapping model is trained when the first loss function converges.

[0045] The first mapping model is an environment-monitoring point deformation model. For example, the first mapping model can use a gated recurrent unit (GRU) neural network. The input historical target environmental data samples include reservoir water level and temperature (e.g., time series data of the past 5 years). The output is the predicted deformation data of each monitoring point. The GRU neural network strengthens the weight of the target environmental data of the past 5 years by introducing an attention mechanism. The ratio of the training set, validation set and test set of the first mapping model is 7:2:1, and the mean square error (MSE) of the validation set is ≤0.005.

[0046] In some embodiments, the target environment data sample is obtained in the following ways: Multiple sets of source environmental data samples and the measured deformation and stress data of each monitoring point in the dam corresponding to each set of source environmental data samples are acquired, with the measured deformation data serving as the first deformation label. Specifically, the source environmental data samples include historical data such as reservoir water level, air temperature, and reservoir water temperature, which are collected through corresponding sensors. The measured deformation and stress data of each monitoring point are historical data collected by deformation sensors and stress sensors pre-embedded in the dam.

[0047] Calculate the deformation correlation coefficient and stress correlation coefficient between each type of environmental data in the source environmental data sample and the measured deformation data and measured stress data, respectively, and select the environmental data whose deformation correlation coefficient and stress correlation coefficient are both greater than the correlation coefficient threshold as the target environmental data sample.

[0048] Specifically, to extract the temporal and spatial correlation characteristics of each source environmental data sample and the deformation and stress at each monitoring point, a sliding window statistical feature extraction method combined with the autocorrelation function (ACF) can be used to extract temporal features from the multi-source monitoring data of the dam, namely the temporal sequences of each source environmental data sample (reservoir water level, air temperature, and reservoir water temperature, etc.), deformation, and stress. A spatial neighborhood statistical method combined with the distance weighting method can be used to extract spatial correlation characteristics from each source environmental data sample (reservoir water level, air temperature, and reservoir water temperature, etc.), deformation, and stress, based on the spatial distribution of each monitoring point.

[0049] Based on the temporal and spatial correlation characteristics of each source environmental data sample, deformation, and stress, correlation analysis (e.g., Pearson correlation coefficient) is used to calculate the correlation coefficient between each source environmental data sample and the dam's performance (including deformation and stress). A higher correlation coefficient indicates that the source environmental data is a key driving factor for the dam's performance. For example, the correlation coefficient threshold is 0.5.

[0050] It should be noted that the preprocessing of the source environmental data samples is included before calculating the correlation coefficient. Missing data repair and time series alignment: Reconstruct missing data in source environmental data samples to ensure that the missing rate of a single monitoring sequence is repaired to an acceptable range for engineering applications; Synchronize data with different sampling frequencies to the standard time axis through a unified time grid mapping and sliding window alignment mechanism.

[0051] Scale unification and standardization: Dynamic standardization with sliding window (window size of 5~10 days) combined with minimum-maximum normalization is adopted to map all source environmental data samples to the [0,1] interval, so as to achieve unification of multi-parameter dimensions and distribution adaptation.

[0052] Signal decomposition and noise suppression: The monitoring sequence corresponding to each source environmental data sample is decomposed into multiple frequency components by adaptive mode decomposition method, separating long-term trends, periodic fluctuations and high-frequency noise, removing non-structural effects such as instrument drift and environmental response, and strengthening information related to the structural behavior of the dam.

[0053] Data drift correction: Construct an integrated "detection-estimation-correction" process, calculate 3 times the standard deviation based on the denoised data as the threshold for abrupt change detection, identify the drift interval, and achieve accurate correction of the drift amount based on the correlation between adjacent measurement points.

[0054] By performing preprocessing such as missing value repair, noise suppression, and drift correction on the source environmental data samples, the continuity and consistency of the source environmental data samples are ensured, avoiding the impact of missing samples and noise on model training. This lays a solid foundation for intelligent analysis, enabling the finally trained deep learning model to accurately predict stress.

[0055] In some embodiments, the training method of the second mapping model is as follows: The second deformation label of each representative node in the dam is obtained by simulation based on the pre-constructed dam simulation model.

[0056] After establishing the dam simulation model, typical working conditions such as normal water storage, sudden water level drop (5~15m), extreme high temperature (35~40℃), extreme low temperature (-15~-5℃), and earthquakes of magnitude VII~VIII are designed and full-process simulation calculations are carried out. For each working condition, 30~50 sets of deformation-stress data are output, generating no less than 300 sets of high-confidence samples. That is, by simulating the deformation calculation results and stress calculation results of the dam under different working conditions through the dam simulation model, the deformation calculation results of each representative node in the dam can be extracted from the overall deformation calculation results and stress calculation results of the dam as a second deformation label.

[0057] The measured deformation samples at the monitoring points are input into the second mapping model to obtain the second deformation prediction data output by the second mapping model. The second deformation prediction data and the second deformation label are substituted into the second loss function. When the second loss function converges, the training of the second mapping model is completed.

[0058] The second mapping model is a monitoring point-representative node deformation model. For example, the second mapping model can be a graph convolutional neural network (GCN / GAT) that fuses spatiotemporal features to realize the spatial transfer of sparse monitoring point deformation to dense representative node deformation, with a test set mean absolute error (MAE) ≤ 0.2 mm.

[0059] In some embodiments, the dam simulation model is constructed as follows: A three-dimensional finite element model of the dam body, including the main body, heel, and toe, was constructed using finite element software. The mesh element size for key components such as the heel, toe, and orifice area was set to 1 / 5 to 1 / 3 of the mesh element size for the main body. The main body of the dam includes its seepage prevention curtain, galleries, and detailed structures such as transverse and longitudinal joints. Specifically, for example... Figure 2 As shown, based on the dam structural design drawings, a three-dimensional finite element model of a typical dam section was established using finite element software. The main body of the dam body uses hexahedral mesh elements. For key parts such as the dam heel, dam toe, and orifice area, the mesh element size is 1 / 5 to 1 / 3 of the main body mesh element size. For example, the mesh element size of the main body of the dam body is 5 to 10 m, while the mesh of the dam heel, dam toe, and orifice area is refined to 1 to 3 m.

[0060] Based on historical monitoring of multiple sets of source environment data samples under different operating conditions, along with corresponding deformations and stresses, key parameters of the finite element model are inverted. Specifically, combining historical monitoring of multiple sets of source environment data samples under different operating conditions (≥100 sets, covering various operating conditions, such as normal water storage, and extreme operating conditions, such as sudden drop in water level, extreme temperature, etc.), along with corresponding deformations and stresses, genetic algorithms or particle swarm optimization algorithms can be used to optimize the finite element model parameters, and key parameters such as surface heat dissipation coefficient, elastic modulus, and linear expansion coefficient are obtained through inversion.

[0061] Under the conditions of the key parameters and the source environmental data samples, if the deviations between the simulation results of deformation and stress obtained from the three-dimensional finite element model and their respective measured results are less than or equal to a preset deviation threshold, then the three-dimensional finite element model is determined to be the dam simulation model. Specifically, it is ensured that, under the conditions of the key parameters and the source environmental data samples, the deviations between the simulation results of deformation and stress and their respective measured data are ≤15%, such as... Figure 3 and 4 As shown.

[0062] In some embodiments, the above-mentioned representative node selection method is as follows: Based on a pre-constructed dam simulation model, the overall deformation and stress calculation results of the dam under multiple working conditions were obtained. Since the main body of the dam in the simulation model is constructed using hexahedral elements, each element can be understood as a three-dimensional mesh. The overall deformation calculation results of the dam include the coordinates of each three-dimensional mesh and the corresponding deformation calculation results, while the overall stress calculation results of the dam include the coordinates of each three-dimensional mesh and the corresponding stress calculation results.

[0063] Based on the physical characteristics of different regions of the dam, meshes from the dam simulation model whose deformation calculation results are greater than or equal to the deformation threshold and whose stress calculation results are greater than or equal to the stress threshold are selected as the first candidate node set. For example, the selection criteria are that the stress threshold is ≥ 70% of the stress design value and the deformation threshold is ≥ 1.5 times the average value of the deformation calculation results. This allows the meshes of high-stress regions such as the dam heel, dam toe, and orifice edge to be selected as the physical key nodes, i.e., the first candidate node set.

[0064] Principal component analysis (PCA) is used to extract the core spatiotemporal features of the deformation field characterized by deformation calculation results. Based on the mean, standard deviation, trend slope, coefficient of variation, and interquartile range of the deformation calculation results, spectral clustering is used to divide similar deformation mode regions. A predetermined number of grids are selected as the second candidate node set in each similar deformation mode region. For example, PCA is performed on the simulated deformation calculation results to extract the first 3-5 principal components (cumulative contribution rate ≥ 85%). Spectral clustering is then used to divide the dam body into 5-8 similar deformation mode regions, and 3-6 representative nodes are selected in each similar deformation mode region. The nodes in the first candidate node set and the second candidate node set are determined as the representative nodes. Specifically, duplicate nodes in the first candidate node set and the second candidate node set are removed to form representative nodes that account for 5% to 10% of the total number of nodes in the dam body and cover the key stress areas. For example: Figure 5 As shown, a set of 30 to 50 representative nodes will eventually be formed, covering the key stress areas and deformation characteristic areas of the dam body.

[0065] In some embodiments, the training method of the third mapping model is as follows: Deformation samples and stress labels for representative nodes in the dam are obtained based on a pre-constructed dam simulation model. Specifically, the deformation calculation results of representative nodes in the deformation calculation results obtained from the dam simulation model inversion are used as deformation samples, and the stress calculation results of representative nodes in the stress calculation results obtained from the dam simulation model inversion are used as stress labels.

[0066] The deformation samples of each representative node are input into the third mapping model to obtain the predicted stress data of each representative node output by the third mapping model; the predicted stress data and the stress labels are substituted into the third loss function, and the training of the third mapping model is completed when the third loss function converges.

[0067] The third mapping model is a deformation-stress model representing the nodes. For example, the third mapping model can be a GNN-EPINN (Graph Neural Network based Efficient Physics-Informed Neural Network) hybrid model. The input is the deformation data of each representative node, and the output is the stress data of the corresponding representative node. A multi-objective loss function (data fitting loss + potential energy loss + boundary condition loss) is designed. The model is trained for 300-800 iterations, and the relative error of stress prediction on the test set is ≤8%. The stress process lines derived from the known deformation of the representative nodes are compared with the stress process lines calculated by the simulation model. Figure 6 As shown, by Figure 6 It can be seen that the stress of the representative node predicted by the third mapping model is basically consistent with the stress calculated by the simulation model, indicating that the third mapping model has accurate prediction ability.

[0068] In some embodiments, the fourth mapping model is trained as follows: The stress samples of representative nodes in the dam and the three-dimensional stress field labels of the dam are obtained by simulation based on a pre-constructed dam simulation model. That is, the stress calculation results of representative nodes in the stress calculation results obtained by inverting the dam simulation model are used as stress samples.

[0069] The stress sample is input into the fourth mapping model to obtain the predicted stress field output by the fourth mapping model. The predicted stress field and the three-dimensional stress field label are substituted into the fourth loss function. When the fourth loss function converges, the training of the fourth mapping model is completed.

[0070] The fourth mapping model is a stress-3D stress field model representing the nodes. For example, the fourth mapping model can be a model using Graph Attention Network (GAT) combined with Radial Basis Function (RBF) interpolation algorithm, or a RBF interpolation algorithm model. That is, the fourth mapping model can be understood as an interpolation module that generates deformation and stress cloud maps through RBF interpolation. Based on the stress samples of the representative nodes and the topological relationship of the dam's finite element mesh, it reconstructs the 3D stress field. Its loss function includes data fitting loss (the deviation between the reconstructed and simulated 3D stress field values, which can be understood as the deviation between the interpolated and reconstructed 3D stress field and the simulated 3D stress field) and smoothing loss (stress gradient constraints of adjacent mesh elements). After training, the average relative error between the reconstructed and simulated 3D stress fields is ≤10%. Figure 7 As shown in (a) and (b), the simulated three-dimensional stress field is compared with four mapping models ( Figure 7 The comparison results of the three-dimensional stress field derived by the proxy model in the model show that the two are basically consistent, indicating that the three-dimensional stress field of the dam can be accurately predicted through the four mapping models, thus effectively guiding the training of the deep learning model as a student model. The trained deep learning model can accurately perceive the three-dimensional stress field of the dam.

[0071] The following describes the long-term sensing device for dam performance based on deformation-stress mapping analysis provided by the present invention. The long-term sensing device for dam performance based on deformation-stress mapping analysis described below can be referred to in correspondence with the long-term sensing method for dam performance based on deformation-stress mapping analysis described above.

[0072] The long-term sensing device for dam performance based on deformation-stress mapping analysis, as described in this invention, is an example of such a device. Figure 8 As shown, it includes: The environmental data acquisition module 810 is used to acquire target environmental data for real-time monitoring of the dam.

[0073] The model execution module 820 is used to input the target environment data into the deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model.

[0074] The deep learning model is a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model.

[0075] The first mapping model is trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam.

[0076] The second mapping model is trained based on the deformation samples measured at each monitoring point and the second deformation labels of each representative node in the dam.

[0077] The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model.

[0078] The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

[0079] In some embodiments, the first mapping model is trained as follows: The target environment data sample is input into the first mapping model to obtain the predicted deformation data of each monitoring point output by the first mapping model. The predicted deformation data and the first deformation label are substituted into the first loss function. When the first loss function converges, the training of the first mapping model is completed.

[0080] In some embodiments, the target environment data sample is obtained in the following manner: Acquire multiple sets of source environment data samples and the measured deformation data and measured stress data of each monitoring point in the dam corresponding to each set of source environment data samples, and use the measured deformation data as the first deformation label; Calculate the deformation correlation coefficient and stress correlation coefficient between each type of environmental data in the source environmental data sample and the measured deformation data and measured stress data, respectively, and select the environmental data whose deformation correlation coefficient and stress correlation coefficient are both greater than the correlation coefficient threshold as the target environmental data sample.

[0081] In some embodiments, the training method of the second mapping model is as follows: The second deformation label of each representative node in the dam is obtained by simulation based on the pre-constructed dam simulation model. The measured deformation samples at the monitoring points are input into the second mapping model to obtain the second deformation prediction data output by the second mapping model. The second deformation prediction data and the second deformation label are substituted into the second loss function. When the second loss function converges, the training of the second mapping model is completed.

[0082] In some embodiments, the representative node is selected as follows: Based on the pre-built dam simulation model, the overall deformation and stress calculation results of the dam under multiple working conditions were obtained. Based on the physical characteristics of each region of the dam, the mesh of the dam simulation model whose deformation calculation results are greater than or equal to the deformation threshold and whose stress calculation results are greater than or equal to the stress threshold are selected as the first candidate node set. Principal component analysis is used to extract the core spatiotemporal features of the deformation field characterized by deformation calculation results. Similar deformation mode regions are divided by spectral clustering, and a predetermined number of grids are selected as the second candidate node set in each similar deformation mode region. The nodes in the first candidate node set and the second candidate node set are determined as the representative nodes.

[0083] In some embodiments, the training method of the third mapping model is as follows: Deformation samples and stress labels of representative nodes in the dam were obtained by simulating a pre-built dam simulation model. The deformation samples of each representative node are input into the third mapping model to obtain the predicted stress data of each representative node output by the third mapping model; the predicted stress data and the stress labels are substituted into the third loss function, and the training of the third mapping model is completed when the third loss function converges.

[0084] In some embodiments, the fourth mapping model is trained as follows: Based on the pre-built dam simulation model, stress samples of representative nodes in the dam and three-dimensional stress field labels of the dam were obtained through simulation. The stress sample is input into the fourth mapping model to obtain the predicted stress field output by the fourth mapping model. The predicted stress field and the three-dimensional stress field label are substituted into the fourth loss function. When the fourth loss function converges, the training of the fourth mapping model is completed.

[0085] In some embodiments, the dam simulation model is constructed as follows: A three-dimensional finite element model including the main body of the dam, the dam heel, and the dam toe was constructed using finite element software. The mesh element size of the key parts of the dam heel, dam toe, and orifice area was set to 1 / 5 to 1 / 3 of the mesh element size of the main body of the dam. Based on the source environment data samples and corresponding deformations and stresses from multiple sets of historical monitoring under different working conditions, the key parameters of the finite element model are inverted. Under the conditions of the key parameters and the source environment data sample, if the deviations between the simulation results of deformation and stress obtained from the three-dimensional finite element model and their respective measured results are less than or equal to a preset deviation threshold, the three-dimensional finite element model is determined to be the dam simulation model.

[0086] Figure 9 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 9 As shown, the electronic device may include: a processor 910, a communication interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communication interface 920, and the memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions in the memory 930 to execute a long-term sensing method for the operational performance of a dam based on deformation-stress mapping analysis. This method includes: Obtain target environmental data for real-time monitoring of the dam.

[0087] The target environment data is input into a deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model.

[0088] The deep learning model is a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model.

[0089] The first mapping model is trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam.

[0090] The second mapping model is trained based on the deformation samples measured at each monitoring point and the second deformation labels of each representative node in the dam.

[0091] The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model.

[0092] The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

[0093] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the long-term sensing method for dam performance based on deformation-stress mapping analysis provided by the above methods. The method includes: Obtain target environmental data for real-time monitoring of the dam.

[0095] The target environment data is input into a deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model.

[0096] The deep learning model is a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model.

[0097] The first mapping model is trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam.

[0098] The second mapping model is trained based on the deformation samples measured at each monitoring point and the second deformation labels of each representative node in the dam.

[0099] The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model.

[0100] The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

[0101] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the long-term sensing method for dam performance based on deformation-stress mapping analysis provided by the methods described above, the method comprising: Obtain target environmental data for real-time monitoring of the dam.

[0102] The target environment data is input into a deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model.

[0103] The deep learning model is a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model.

[0104] The first mapping model is trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam.

[0105] The second mapping model is trained based on the deformation samples measured at each monitoring point and the second deformation labels of each representative node in the dam.

[0106] The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model.

[0107] The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence 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 ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] All actions involving the acquisition of signal information or data in this invention are carried out in compliance with the relevant data protection laws and policies of the country where the device is located, and with the authorization granted by the owner of the device.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A long-term sensing method for the operational performance of dams based on deformation-stress mapping analysis, characterized in that, include: Acquire real-time environmental data for monitoring the dam; The target environment data is input into a deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model. The deep learning model is used as a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes: a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model. The first mapping model was trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam; The second mapping model is trained based on the measured deformation samples of each monitoring point and the second deformation labels of each representative node in the dam; The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model. The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

2. The long-term sensing method for dam performance based on deformation-stress mapping analysis according to claim 1, characterized in that, The training method for the first mapping model is as follows: The target environment data sample is input into the first mapping model to obtain the predicted deformation data of each monitoring point output by the first mapping model. The predicted deformation data and the first deformation label are substituted into the first loss function. When the first loss function converges, the training of the first mapping model is completed.

3. The long-term sensing method for dam performance based on deformation-stress mapping analysis according to claim 2, characterized in that, The target environment data sample is obtained in the following way: Acquire multiple sets of source environment data samples and the measured deformation data and measured stress data of each monitoring point in the dam corresponding to each set of source environment data samples, and use the measured deformation data as the first deformation label; Calculate the deformation correlation coefficient and stress correlation coefficient between each type of environmental data in the source environmental data sample and the measured deformation data and measured stress data, respectively, and select the environmental data whose deformation correlation coefficient and stress correlation coefficient are both greater than the correlation coefficient threshold as the target environmental data sample.

4. The long-term sensing method for dam operational performance based on deformation-stress mapping analysis according to claim 1, characterized in that, The training method for the second mapping model is as follows: The second deformation label of each representative node in the dam is obtained by simulation based on the pre-constructed dam simulation model. The measured deformation samples at the monitoring points are input into the second mapping model to obtain the second deformation prediction data output by the second mapping model. The second deformation prediction data and the second deformation label are substituted into the second loss function. When the second loss function converges, the training of the second mapping model is completed.

5. The long-term sensing method for dam performance based on deformation-stress mapping analysis according to claim 4, characterized in that, The method for selecting representative nodes is as follows: Based on the pre-built dam simulation model, the overall deformation and stress calculation results of the dam under multiple working conditions were obtained. Based on the physical characteristics of each region of the dam, the mesh of the dam simulation model whose deformation calculation results are greater than or equal to the deformation threshold and whose stress calculation results are greater than or equal to the stress threshold are selected as the first candidate node set. Principal component analysis is used to extract the core spatiotemporal features of the deformation field characterized by deformation calculation results. Similar deformation mode regions are divided by spectral clustering, and a predetermined number of grids are selected as the second candidate node set in each similar deformation mode region. The nodes in the first candidate node set and the second candidate node set are determined as the representative nodes.

6. The long-term sensing method for dam performance based on deformation-stress mapping analysis according to claim 1, characterized in that, The training method for the third mapping model is as follows: Deformation samples and stress labels of representative nodes in the dam were obtained by simulating a pre-built dam simulation model. The deformation samples of each representative node are input into the third mapping model to obtain the predicted stress data of each representative node output by the third mapping model; the predicted stress data and the stress labels are substituted into the third loss function, and the training of the third mapping model is completed when the third loss function converges.

7. The long-term sensing method for dam performance based on deformation-stress mapping analysis according to claim 1, characterized in that, The training method for the fourth mapping model is as follows: Based on the pre-built dam simulation model, stress samples of representative nodes in the dam and three-dimensional stress field labels of the dam were obtained through simulation. The stress sample is input into the fourth mapping model to obtain the predicted stress field output by the fourth mapping model. The predicted stress field and the three-dimensional stress field label are substituted into the fourth loss function. When the fourth loss function converges, the training of the fourth mapping model is completed.

8. The long-term sensing method for dam operational performance based on deformation-stress mapping analysis according to any one of claims 3 to 7, characterized in that, The dam simulation model is constructed as follows: A three-dimensional finite element model including the main body of the dam, the dam heel, and the dam toe was constructed using finite element software. The mesh element size of the key parts of the dam heel, dam toe, and orifice area was set to 1 / 5 to 1 / 3 of the mesh element size of the main body of the dam. Based on the source environment data samples and corresponding deformations and stresses from multiple sets of historical monitoring under different working conditions, the key parameters of the finite element model are inverted. Under the conditions of the key parameters and the source environment data sample, if the deviations between the simulation results of deformation and stress obtained from the three-dimensional finite element model and their respective measured results are less than or equal to a preset deviation threshold, the three-dimensional finite element model is determined to be the dam simulation model.

9. A long-term sensing device for the operational performance of a dam based on deformation-stress mapping analysis, characterized in that, include: The environmental data acquisition module is used to acquire target environmental data for real-time monitoring of the dam. The model execution module is used to input the target environment data into the deep learning model to obtain the three-dimensional stress field of the dam output by the deep learning model. The deep learning model is used as a student model, which is trained by knowledge transfer from a teacher model based on target environment data samples and their corresponding three-dimensional stress field labels. The teacher model includes: a first mapping model, a second mapping model, a third mapping model, and a fourth mapping model. The first mapping model was trained based on the target environment data samples and the first deformation labels of each monitoring point in the dam; The second mapping model is trained based on the measured deformation samples of each monitoring point and the second deformation labels of each representative node in the dam; The third mapping model is trained based on the deformation samples and stress labels of each representative node, wherein the deformation samples and stress labels of each representative node are obtained based on a pre-built dam simulation model. The fourth mapping model is trained based on the stress samples and three-dimensional stress field labels of each representative node, wherein the stress samples and three-dimensional stress field labels of each representative node are obtained based on a pre-constructed dam simulation model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the long-term sensing method for dam performance based on deformation-stress mapping analysis as described in any one of claims 1 to 7.

Images