Concrete dam deformation early warning method based on PDM-GAR model

Abstract

The invention provides a concrete dam deformation early warning method based on a PDM-GAR model, and the method comprises the steps: extracting and obtaining the multi-scale feature information of historical monitoring data through a PDM module of the PDM-GAR model, carrying out the weighted regression of environment variables of all scales through a GAR module, and precisely simulating the deformation value of each measurement point; according to the method, a dam safety evaluation system is constructed based on a simulation result of a PDM-GAR model and a historical residual error in a training process of the PDM-GAR model; starting from artificial intelligence simulation residual errors of the single measuring points, the early warning level of the single measuring points is evaluated step by step, and finally comprehensive evaluation of the safety early warning level of the whole dam is achieved; according to the dam deformation simulation method based on the PDM-GAR, the problem of single-point to overall safety evaluation is effectively solved, the simulation result of the PDM-GAR is combined with the health state of the dam through the dam safety early warning algorithm based on measuring point-overall grading, and the problems that an existing dam deformation simulation method is low in simulation precision and poor in generalization performance are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring and early warning, and in particular to a deformation early warning method for concrete dams based on the PDM-GAR model. Background Art

[0002] As an important part of water conservancy projects, the stability and safety of reservoir concrete dams are directly related to the safety of people's lives and property in downstream areas. Therefore, accurately simulating the deformation of concrete dams and evaluating the safety status of dams based on the simulation results are of great significance for disaster prevention and reduction and ensuring the safe operation of dams.

[0003] As a dynamically changing complex system, the operating environment of concrete dams is complex and changeable, and is affected by many factors. The operating state of the dam usually manifests in the forms of deformation, stress, displacement, seepage, seepage pressure, and cracking. Among them, deformation, as the most intuitive reflection of the operating state of concrete gravity dams, is closely related to variables such as temperature, water pressure, and time. The changes in the operating state and safety of the dam will be revealed through deformation data. Therefore, in-depth analysis and accurate simulation of deformation laws are important means to ensure the safety of the dam.

[0004] With the rapid development of deep learning technology, more and more artificial intelligence methods have been used for the analysis, simulation, and monitoring of dam deformation. For example, Wei Bowen et al. proposed a combined simulation method based on BP-ARIMA, which corrected the results by superimposing the residual prediction term and the regression model simulation value; Xiang Xinghua et al. proposed a simulation method based on IPSO-Transformer, which used an improved particle swarm algorithm to optimize the model training process. However, these methods have certain limitations in processing historical monitoring data. Most of them only input the average values of variables such as water temperature and air temperature into the model simply, and cannot deeply explore the deformation laws of the dam under different monitoring environments in historical data, resulting in insufficient simulation accuracy. In addition, due to the structural characteristics of traditional time series recurrent neural network models, it is difficult to capture long-term dependence relationships, and there are also problems of gradient disappearance or gradient explosion, which further limits their application effects. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a deformation early warning method for concrete dams based on the PDM-GAR model. The present invention significantly improves the accuracy of deformation simulation, realizes hierarchical diagnosis from single measuring points to the overall dam, and provides a more comprehensive dam safety assessment result through layer-by-layer information fusion and reasoning.

[0006] The technical solution of the present invention is: a deformation early warning method for concrete dams based on the PDM-GAR model, including the following steps:

[0007] S1), Obtain the historical data of the dam over a period of time, preprocess it, and divide it into a training set, a validation set, and a test set;

[0008] S2), Construct a concrete dam deformation simulation hybrid network model PDM-GAR; use the training set for training, and then use the validation set and the test set to verify and test the trained concrete dam deformation simulation hybrid network model PDM-GAR;

[0009] S3), Calculate the deformation residuals of all horizontal monitoring points of the dam in the training set, validation set, and test set of the concrete dam deformation simulation hybrid network model PDM-GAR; store the historical residuals of each point one by one, and construct a residual library for each monitoring point;

[0010] S4), Sort the residual library of each monitoring point from small to large to generate a cumulative distribution knowledge library CDF of the residuals of the monitoring points;

[0011] S5), Real-time obtain the dam monitoring data and input it into the trained concrete dam deformation simulation hybrid network model PDM-GAR. The concrete dam deformation simulation hybrid network model PDM-GAR uses environmental variable data to simulate the deformation values of each point of the dam, and combines the real deformation values to calculate the current simulated deformation residuals;

[0012] S6), Perform safety scoring according to the simulated deformation residuals of each monitoring point and its corresponding cumulative distribution knowledge library;

[0013] S7), Evaluate the warning level of each monitoring point according to the safety score and the warning rules;

[0014] S8), Use the weighted scoring algorithm for monitoring point zoning to evaluate the overall safety status of the dam;

[0015] S9), Further determine the overall safety warning level of the dam according to the warning rules and the overall safety score of the dam.

[0016] Preferably, in step S1), the historical data X of the dam includes the water level H, the air temperature T air , the water temperature T water , and the time component θ; which is expressed as:

[0017]

[0018] Among them, is the temperature of the j-th water temperature monitoring point.

[0019] Preferably, in step S2), the concrete dam deformation simulation hybrid network model PDM-GAR includes a past decomposition mixing module PDM and a gated attention residual network module GAR; the past decomposition mixing module PDM decomposes into 4 time series with different scales, and the gated attention residual network module GAR obtains the contribution components of each scale.

[0020] Preferably, in step S2), the past decomposition mixing module PDM includes a downsampling module, multiple auto-vector pooling layers, and two feedforward layers; the downsampling module decomposes the environmental variables of historical temperature, water temperature, and upstream water level height into 4 different time scales x={x0, x1, x2, x3};

[0021] Then, the auto-vector pooling layer is used to decompose the environmental variables of 4 different scales into seasonal components x s and trend components x t , that is:

[0022] x t = AvgPool(Padding(x));

[0023] x s = x - x t ; (2)

[0024] In the formula, AvgPool represents pooling processing using the auto-vector pooling layer, Padding represents the padding operation to keep the variable length unchanged; x represents time series with different scales;

[0025] Then, the trend components x t decomposed from 4 different scales are mixed in a top-down manner, while the seasonal components x s are mixed in a bottom-up manner; after mixing, it is processed through the feedforward layer, and finally the fused multi-scale mixed data is obtained.

[0026] Preferably, in step S2), the feedforward layer consists of two linear layers, and the GELU activation function is used for non-linear transformation between the two linear layers to enhance the feature expression ability.

[0027] Preferably, in step S2), the gated attention residual network module GAR includes multiple 4-branch circuits, and each branch circuit corresponds to processing the scale mixed data output by the past decomposition mixing module PDM.

[0028] Preferably, in step S2), each branch of the gated attention residual network module GAR includes a multi-layer GRU module, a multi-head attention module, and a residual connection module; the residual connection module consists of one or two mean pooling layers and two linear layers.

[0029] Preferably, in step S2), in each branch of the gated attention residual network module GAR, a multi-layer GRU module is first used to capture the temporal dependency of the dam deformation on the environmental variables, and the weighting process is:

[0030] r t =σ(W xr x t +W hr h t-1 +b r ) (3)

[0031] z t =σ(W xz x t +W hz h t-1 +b z ) (4)

[0032]

[0033] Where r t 、z t 、h t 、 Represent the reset gate, update gate, hidden state, and candidate hidden state respectively; σ represents the sigmoid activation function; ⊙ represents the Hadamard product; tanh(·) is the tanh activation function, W xr 、W hr is the reset gate weight matrix, W xz 、W hz To update the gate weight matrix, W xh 、W hh is the hidden layer weight matrix, b r 、b z and b h are reset gate, update gate and hidden layer bias vector respectively; x t is the trend component.

[0034] Preferably, in step S2), a multi-head attention module is used to perform a linear transformation on each tensor X output after processing by the multi-layer GRU module, and project it into a query Q, a key K, and a value V, that is:

[0035] Q=XW Q ,K=XW K ,V=XW V (7)

[0036] Where X is the tensor output by the multi-layer GRU module; Transformation matrices representing query, key, and value respectively; represents the dimension of the key vector; H is the hidden layer size, and h is the number of attention heads;

[0037] And calculate the dot product attention for each sample:

[0038]

[0039] In the formula, softmax is the activation function, T represents the transposition operation;

[0040] For the multi-head attention mechanism, h attention heads are used, and the output of each head is:

[0041] head i =Attenttion(Q i ,K i ,V i ) (9)

[0042] Where, represents the i-th attention head; Q i ,K i ,V i denote the query, key, and value matrices of the i-th attention head respectively;

[0043] Finally, the outputs of all attention heads are concatenated and linearly transformed to generate the final output:

[0044] MutiHead=Concat(head1,…,head h )W O (10)

[0045] In the formula, MutiHead represents multi-head attention; Concat represents the concatenation operation; is the linear transformation weight matrix after splicing; is the linear transformation weight matrix after splicing.

[0046] Preferably, in step S3), the deformation residual simulated by the concrete dam deformation simulation hybrid network model PDM-GAR is expressed as:

[0047]

[0048] Where δ is the deformation residual; y i is the deformation observation value of the i-th horizontal monitoring point; is the deformation prediction value of the concrete dam deformation simulation hybrid network model PDM-GAR at the i-th horizontal monitoring point.

[0049] Preferably, in step S6), according to the simulated deformation residuals of each measuring point and its corresponding cumulative distribution knowledge base, a safety score score is carried out; specifically as follows:

[0050] S61) Compare the simulated deformation residuals of each measuring point on the current day with the cumulative distribution in its residual library to obtain the quantile corresponding to the deformation residuals; the quantile δ cdf The calculation formula of is:

[0051]

[0052] In the formula, δ is the deformation residual simulated by the concrete dam deformation simulation hybrid network model PDM-GAR; index(·) is to obtain the sequence number of the measuring point residual in the residual library; n is the number of data in the residual library; δ cdf is the quantile of the residual in the residual library;

[0053] S62) Based on the quantile and the score calculation formula, a score is carried out. The score calculation formula is:

[0054]

[0055] Among them, the scoring rules are set in combination with the normal distribution and the distribution of the residual library, and the cumulative distribution values corresponding to each standard deviation of the normal distribution are used to determine the segmentation criteria for scoring.

[0056] Preferably, in step S6), if the simulated deformation residual exceeds the range of the residual library, the measuring point state is extremely unsafe and the score is a fixed value of 55 points.

[0057] Preferably, in step S7), the early warning rules are as follows:

[0058] If the cumulative distribution probability warning index is [0.159, 0.977], and the score range is [100, 90), the warning level is "normal";

[0059] If the cumulative distribution probability warning index is (0.023, 0.159) ∪ (0.841, 0.977), and the score range is [100, 90), the warning level is "fourth-level warning";

[0060] If the cumulative distribution probability warning index is (0.001, 0.023) ∪ (0.977, 0.999), and the score range is [80, 70), the warning level is "third-level warning";

[0061] If it is outside the residual library sequence of the cumulative distribution probability warning index and the score range is [0, 60], the warning level is "Level 1 Warning".

[0062] Preferably, in step S8), the weighted scoring algorithm for measuring point zoning is used to evaluate the overall safety status of the dam, which specifically includes the following steps:

[0063] S81), Using the Ward hierarchical clustering method, based on the historical deformation data of all deformation monitoring points of the dam, the measuring points are clustered into N categories;

[0064] S82), Calculate the within-class variance of each category and use it as the weight w of the corresponding category;

[0065] S83), Calculate the overall safety score score of the dam according to the following formula dam , that is:

[0066]

[0067] In the formula, score i is the safety score of measuring point i, and w i is the weight value of the category corresponding to measuring point i.

[0068] The beneficial effects of the present invention are:

[0069] 1. Based on the historical environmental variable information of dam monitoring, the present invention extracts and obtains the multi-scale feature information of historical monitoring data through the PDM module, and uses the gated attention residual network module GAR to perform weighted regression on environmental variables at each scale, accurately simulating the deformation values of each measuring point;

[0070] 2. Based on the simulation results of the PDM-GAR model and the historical residuals during its training process, the present invention constructs a complete dam safety assessment system; starting from the artificial intelligence simulation residuals of a single measuring point, the warning level of a single measuring point is gradually evaluated, and finally the comprehensive evaluation of the overall dam safety warning level is realized; this system effectively solves the problem of safety assessment from a single point to the whole, providing a scientific basis for the comprehensive monitoring and safety guarantee of the dam operation status;

[0071] 3. The present invention solves the problems of low simulation accuracy and poor generalization performance in existing dam deformation simulation methods. Through the dam safety warning algorithm based on measuring point - overall grading, the simulation results of PDM-GAR are combined with the dam health status, thereby providing an efficient solution for the real-time monitoring and safety assessment of the dam, and significantly reducing the dam maintenance cost. Description of the Drawings

[0072] Figure 1 is a schematic flow chart of the warning method of the present invention;

[0073] Figure 2 It is the structural framework diagram of the PDM-GAR model of the present invention;

[0074] Figure 3 It is the structural framework diagram of the past decomposition and mixing PDM module of the present invention;

[0075] Figure 4 It is the distribution map of dam deformation monitoring points in the embodiment of the present invention;

[0076] Figure 5 It is the time series diagram of the results of the PDM-GAR model of the present invention and other statistical models;

[0077] Figure 6 It is the time series diagram of the ablation experiment results of the PDM-GAR model of the present invention;

[0078] Figure 7 It is the residual distribution diagram of the PDM-GAR model of the present invention and the comparison model;

[0079] Figure 8 It is the schematic diagram of the partition of dam measuring points in the embodiment of the present invention. Detailed implementation manners

[0080] The following further describes the detailed implementation manners of the present invention with reference to the accompanying drawings:

[0081] As Figure 1 shown, this embodiment provides a concrete dam deformation early warning method based on the PDM-GAR model, including the following steps:

[0082] S1), Obtain the historical data of the dam in the past 96 days, and after preprocessing, divide it into a training set, a validation set, and a test set;

[0083] The historical data X of the dam includes water level H, air temperature T air , water temperature T water , and time component θ; expressed as:

[0084]

[0085] Among them, is the temperature of the jth water temperature measuring point.

[0086] S2), Construct a concrete dam deformation simulation hybrid network model PDM-GAR; and use the training set for training, and then use the validation set and the test set to verify and test the trained concrete dam deformation simulation hybrid network model PDM-GAR;

[0087] As Figure 2As shown, the concrete dam deformation simulation hybrid network model PDM-GAR includes a past decomposition mixing module PDM and a gated attention residual network module GAR; it is decomposed into four time series of different scales by the past decomposition mixing module PDM, and the contribution components of each scale are obtained through the gated attention residual network module GAR.

[0088] As Figure 3 shown, the past decomposition mixing module PDM includes a downsampling module, multiple auto-vector pooling layers, and two feedforward layers; the downsampling module decomposes the environmental variables of historical air temperature, water temperature, and upstream water level height into four different time scales x = {x0, x1, x2, x3};

[0089] Then, the auto-vector pooling layer is used to decompose the environmental variables of the four different scales into seasonal components x s and trend components x t , that is:

[0090] x t = AvgPool(Padding(x));

[0091] x s = x - x t ; (2)

[0092] In the formula, AvgPool represents pooling processing using the auto-vector pooling layer, Padding represents the padding operation to keep the variable length unchanged; x represents the time series of different scales;

[0093] Then, the trend components x t decomposed from the four different scales are mixed in a top-down manner, while the seasonal components x s are mixed in a bottom-up manner; after mixing, it is processed through the feedforward layer, and finally the fused multi-scale mixed data is obtained. The feedforward layer consists of two linear layers, and the GELU activation function is used for non-linear transformation between the two linear layers to enhance the feature expression ability.

[0094] As Figure 2 shown, the gated attention residual network module GAR includes multiple four branches, and each branch corresponds to processing the scale mixed data output by the past decomposition mixing module PDM. Each branch of the gated attention residual network module GAR includes multiple GRU modules, a multi-head attention module, and a residual connection module; the residual connection module consists of one or two mean pooling layers and two linear layers.

[0095] In each branch of the gated attention residual network module GAR, a multi-layer GRU module is first used to capture the temporal dependence of dam deformation on environmental variables. The weighting process is as follows:

[0096] r t =σ(W xr x t +W hr h t-1 +b r ) (3)

[0097] z t =σ(W x2 x t +W hz h t-1 +b z ) (4)

[0098]

[0099] In the formula, r t , z t , h t , respectively represent the reset gate, update gate, hidden state, and candidate hidden state; σ represents the sigmoid activation function; ⊙ represents the Hadamard product; tanh(·) is the tanh activation function, W xr , W hr are the reset gate weight matrices, W xz , W hz are the update gate weight matrices, W xh , W hh are the hidden layer weight matrices, b r , b z and b h are the reset gate, update gate, and hidden layer bias vectors respectively; x t is the trend component.

[0100] Use the multi-head attention module to perform a linear transformation on each tensor X output after being processed by the multi-layer GRU module, and project it into the query Q, key K, and value V, that is:

[0101] Q=XW Q ,K=XW K ,V=XW V (7)

[0102] In the formula, X is the tensor output by the multi-layer GRU module; respectively represent the transformation matrices of the query, key, and value;[[ID=8%]] represents the dimension of the key vector; H is the hidden layer size, and h is the number of attention heads;

[0103] And calculate the dot-product attention for each sample:

[0104]

[0105] Where softmax is the activation function and T represents the transpose operation;

[0106] For the multi-head attention mechanism, h attention heads are used, and the output of each head is:

[0107] head i = Attenttion(Q i , K i , V i ) (9)

[0108] Where represents the i-th attention head; Q i , K i , V i represent the query, key, and value matrices of the i-th attention head respectively;

[0109] Finally, the outputs of all attention heads are concatenated and the final output result is generated through a linear transformation:

[0110] MutiHead = Concat(head1,…,head h )W O (10)

[0111] Where MutiHead represents multi-head attention; Concat represents the concatenation operation; is the weight matrix of the linear transformation after concatenation.

[0112] During the model training process, the Adam optimization algorithm is used to optimize the internal parameters of the model; and the optimal hyperparameter combination, including the input time length, network depth, and learning rate, is determined through the grid search method. Then the model is evaluated using the test set.

[0113] During the model evaluation stage, three key metrics are calculated to compare the model performance, namely the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R 2 ), to comprehensively evaluate the simulation accuracy and performance of the PDM-GAR model, which are respectively:

[0114]

[0115] Where n is the number of data sets; y i is the deformation observation value of the i-th horizontal monitoring point; is the deformation prediction value of the PDM-GAR hybrid network model for the deformation simulation of the concrete dam at the i-th horizontal monitoring point; is the average value of the observed deformation values of all horizontal monitoring points.

[0116] S3), calculate the deformation residuals of all horizontal monitoring points of the dam in the training set, validation set, and test set by the PDM-GAR hybrid network model for the deformation simulation of the concrete dam; and store the historical residuals of each point one by one to construct the residual library of each monitoring point;

[0117] The deformation residuals simulated by the PDM-GAR hybrid network model for the deformation simulation of the concrete dam are expressed as:

[0118]

[0119] In the formula, δ is the deformation residual; y i is the observed deformation value of the i-th horizontal monitoring point; is the deformation prediction value of the PDM-GAR hybrid network model for the deformation simulation of the concrete dam at the i-th horizontal monitoring point.

[0120] S4), sort the residual library of each monitoring point from small to large to generate the cumulative distribution knowledge library CDF of the monitoring points;

[0121] S5), obtain the dam monitoring data in real time and input it into the trained PDM-GAR hybrid network model for the deformation simulation of the concrete dam. The PDM-GAR hybrid network model for the deformation simulation of the concrete dam uses the environmental variable data to simulate the deformation values of each point of the dam, and calculates the current simulated deformation residuals in combination with the real deformation values;

[0122] S6), perform a safety score score according to the simulated deformation residuals of each monitoring point and its corresponding cumulative distribution knowledge library; specifically as follows:

[0123] S61), compare the simulated deformation residuals of each monitoring point on the current day with the cumulative distribution in its residual library to obtain the quantile corresponding to the deformation residuals; the quantile δ cdf The calculation formula of is:

[0124]

[0125] In the formula, δ is the deformation residual simulated by the PDM-GAR hybrid network model for the deformation simulation of the concrete dam; index(·) is to obtain the sequence number of the monitoring point residual in the residual library; n is the number of data in the residual library; δ cdf is the quantile of the residual in the residual library;

[0126] S62), perform scoring based on the quantile and the score calculation formula. The score calculation formula is:

[0127]

[0128] Among them, the scoring rules are set by combining the normal distribution and the distribution of the residual library, and the cumulative distribution values corresponding to the standard deviations of the normal distribution are used to determine the sectional criteria for scoring.

[0129] If the simulated deformation residual exceeds the range of the residual library, the safety score of the measuring point is a fixed value of 55 points.

[0130] S7) Evaluate the warning level of each measuring point according to the safety score and the warning rules; the warning rules are as follows:

[0131]

[0132] S8) Evaluate the safety status of the dam as a whole by using the weighted scoring algorithm for measuring point zoning; specifically, it includes the following steps:

[0133] S81) Adopt the Ward hierarchical clustering method, and cluster the measuring points into N categories according to the historical deformation data of all deformation monitoring points of the dam;

[0134] S82) Calculate the within-class variance of each category and take it as the weight w of the corresponding category;

[0135] S83) Calculate the overall safety score score of the dam according to the following formula dam , that is:

[0136]

[0137] In the formula, score i is the safety score of measuring point i, and w i is the weight value of the category corresponding to measuring point i.

[0138] S9) Further determine the overall safety warning level of the dam according to the warning rules and the overall safety score of the dam.

[0139] Embodiment 2

[0140] In this embodiment, a certain concrete gravity dam is taken as the research object, and the specific parameters are as follows: the elevation of the dam crest is 156.30 m, the length of the dam crest is 205.5 m, the maximum dam height is 71.3 m, and the catchment area above the dam site is 1990 km 2 , and the total reservoir capacity is 172 million m 3 . The dam is equipped with monitoring items including deformation, seepage pressure, seepage flow, air temperature and water temperature, etc. The monitoring equipment for monitoring the dam displacement mainly includes vertical plumb lines, inverted plumb lines and tension wires, etc. For the specific layout scheme, see Figure 4 .

[0141] In this embodiment, the automated monitoring data of measurement points EX1-2, EX1-6, EX2-4, and EX2-8 from October 14, 2007 to November 28, 2023 are selected, and the data are divided into a training set, a validation set, and a test set according to a ratio of 13:1:1. The training set is used for model training, the validation set is used for hyperparameter optimization and model tuning during the training process, and the test set is used as an independent unknown data set to detect the final performance of the model. The training set and the validation set are input into the PDM-GAR model. The model reads the data of consecutive S days to simulate the deformation data of the next 1 day, and is decomposed into 4 time scales by the PDM module. The contribution components of each scale are obtained by the gated attention residual network module GAR. In the model training stage, in a batch processing manner, the model reads a batch of data of size B to form a tensor of shape B×S×F (where F represents the dimension of the input environmental variables); during the model optimization process, the Adam optimization algorithm is used to optimize the internal parameters of the model; the optimal hyperparameter combination, including the input time length, network depth, and learning rate, etc., is determined by the grid search method. The final result shows that the best simulation effect can be achieved by using the input containing 96-day environmental variable information. After determining the optimal structure of the PDM-GAR model, the performance of the model is evaluated using the test set. At the same time, other comparison models are trained in the same way, including: PDM-CNN-LSTM, PDM-CNN-GRU, CNN-LSTM, CNN-GRU, and GRU, and their performance is compared on the test set. The performance evaluation indicators include MAE, RMSE, and coefficient of determination (R 2 ), and the test results are shown in Table 1:

[0142] Table 1 Test results of Example 1 and comparison models

[0143]

[0144] As can be seen from Table 1, the PDM-GAR model proposed in Example 1 is superior to the comparison models in all indicators, indicating its superiority in the dam deformation simulation task.

[0145] In the model accuracy evaluation stage, in order to better verify the effectiveness of PDM-GAR, Example 1 compared the results of traditional statistical models, deep learning models, and their fusion models, and conducted ablation experiments on PDM-GAR to verify the role of each module. Finally, comparative analysis was carried out through the residual scatter distributions of different models in the test set, as shown in Figure 5 (a), (b), (c), (d); Figure 6 (a), (b), (c), (d) and Figure 7Among (a), (b), (c), and (d), where (a), (b), (c), and (d) respectively correspond to measurement points EX1-2, EX1-6, EX2-4, and EX2-8. The results show that the PDM-GAR model is superior to other models in both simulation accuracy and generalization ability, and can effectively and accurately simulate future dam deformations.

[0146] On a certain day, the simulation residual of measurement point EX1-02 is -0.51 mm, its ranking in the cumulative distribution library is the 39th, and the corresponding quantile is 0.0067. The safety score of this measurement point on that day is 72.3 points. Combining with the early warning rules, a level-IV early warning is required for this measurement point on that day.

[0147] According to the Ward hierarchical clustering method, each measurement point is clustered into several partitions, namely partitions 1-6, based on its historical deformation characteristics. The details of the partitions of the exemplary dam are as Figure 8 shown. The variances of the historical deformation amounts corresponding to the partitions are: 0.27 mm, 2.07 mm, 5.21 mm, 1.71 mm, 0.69 mm, and 1.49 mm respectively. In order to reflect the importance of unstable measurement points, higher weights are assigned to partitions with larger variances. Combining the safety scores of each measurement point and the partition weights, the overall safety score of the dam is calculated. The overall score of the dam on that day is 81.2 points, and the corresponding overall safety early warning level of the dam is level-IV early warning.

[0148] The above embodiments and descriptions in the specification only illustrate the principles and the best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. A method for early warning of concrete dam deformation based on the PDM-GAR model, characterized in that, It includes the following steps: S1), Obtain the historical data of the dam over a period of time, preprocess it, and divide it into a training set, a validation set, and a test set; S2), Construct a concrete dam deformation simulation hybrid network model PDM-GAR; use the training set for training, and then use the validation set and the test set to verify and test the trained concrete dam deformation simulation hybrid network model PDM-GAR; S3), Calculate the deformation residuals of all horizontal monitoring points of the dam in the training set, validation set, and test set of the concrete dam deformation simulation hybrid network model PDM-GAR; store the historical residuals of each point one by one, and construct a residual library for each monitoring point; S4), Sort the residual library of each monitoring point from small to large to generate a cumulative distribution knowledge library CDF of the residuals of the monitoring points; S5), Obtain the dam monitoring data in real time and input it into the trained concrete dam deformation simulation hybrid network model PDM-GAR. The concrete dam deformation simulation hybrid network model PDM-GAR uses environmental variable data to simulate the deformation values of each point of the dam, and combines the real deformation values to calculate the current simulated deformation residuals; S6), Perform safety scoring according to the simulated deformation residuals of each monitoring point and its corresponding cumulative distribution knowledge library; S7), Evaluate the warning level of each monitoring point according to the safety score and the warning rules; S8), Use the weighted scoring algorithm for monitoring point zoning to evaluate the overall safety status of the dam; S9), Further determine the overall safety warning level of the dam according to the warning rules and the overall safety score of the dam.

2. The concrete dam deformation early warning method based on the PDM-GAR model according to claim 1, characterized in that: In step S1), the historical data X of the dam includes the water level H, the air temperature T air , the water temperature T water , and the time component θ; which is expressed as: Among them, is the temperature of the j-th water temperature measurement point.

3. A method for deformation early warning of concrete dams based on the PDM-GAR model according to claim 1, characterized in that: In step S2), the concrete dam deformation simulation hybrid network model PDM-GAR includes a past decomposition hybrid module PDM and a gated attention residual network module GAR; it is decomposed into 4 time series of different scales through the past decomposition hybrid module PDM, and the contribution components of each scale are obtained through the gated attention residual network module GAR.

4. A deformation early warning method for concrete dams based on the PDM-GAR model according to claim 3, characterized in that: In step S2), the past decomposition hybrid module PDM includes a downsampling module, multiple auto-vector pooling layers, and two feedforward layers; the downsampling module decomposes the environmental variables of historical temperature, water temperature, and upstream water level height into 4 different time scales: x = {x0, x1, x2, x3}; Then, the automatic vector pooling layer is used to decompose the environmental variables at four different scales into the seasonal component x s and the trend component x t , that is: x t = AvgPool(Padding(x)); x s = x - x t ; (2) In the formula, AvgPool represents pooling processing using the auto-vector pooling layer, Padding represents a padding operation to keep the variable length unchanged; x represents time series of different scales; Then, the trend components x decomposed at 4 different scales t are mixed in a top-down manner, while the seasonal components x s are mixed in a bottom-up manner; after the mixing is completed, it is processed through the feedforward layer, and finally the fused multi-scale mixed data is obtained.

5. The concrete dam deformation early warning method based on the PDM-GAR model according to claim 3, characterized in that: In step S2), the gated attention residual network module GAR includes multiple 4 branches, and each branch corresponds to processing the scale-mixed data output by the past decomposition hybrid module PDM; Each branch of the gated attention residual network module GAR includes multiple GRU modules, a multi-head attention module, and a residual connection module; the residual connection module consists of two mean pooling layers and two linear layers.

6. The concrete dam deformation early warning method based on the PDM-GAR model according to claim 5, characterized in that: In step S2), in each branch of the gated attention residual network module GAR, first use multiple GRU modules to capture the time-dependent relationship between the dam deformation and the environmental variables, and its weighting process is: r t = σ(W xr x t + W hr h t-1 + b r ) (3) z t = σ(W xz x t + W hz h t-1 + b z ) (4) where r t , z t , h t , and represent the reset gate, update gate, hidden state, and candidate hidden state respectively; σ represents the sigmoid activation function; ⊙ represents the Hadamard product; tanh(·) is the tanh activation function, W xr , W hr are the reset gate weight matrices, W xz , W hz are the update gate weight matrices, W xh , W hh are the hidden layer weight matrices, b r , b z and b h are the reset gate, update gate, and hidden layer bias vectors respectively; x t is the trend component; Use a multi-head attention module to perform a linear transformation on each tensor X output after being processed by a multi-layer GRU module, and project it into a query Q, a key K, and a value V, that is: Q = XW Q , K = XW K , V = XW V (7) where X is the tensor output by the multi-layer GRU module; respectively represent the transformation matrices of the query, key, and value; represents the dimension of the key vector; H is the hidden layer size, and h is the number of attention heads; And calculate the dot-product attention for each sample: In the formula, softmax is the activation function, and T represents the transpose operation; For the multi-head attention mechanism, use h attention heads, and the output of each head is: head i = Attention(Q i , K i , V i ) (9) In the formula, represents the i-th attention head; Q i , K i , V i represent the query, key, and value matrices of the i-th attention head respectively; Finally, concatenate the outputs of all attention heads and generate the final output result through a linear transformation: MutiHead=Concat(head1,…,head h )W O (10) Wherein, MutiHead represents multi-head attention; Concat represents the concatenation operation; is the weight matrix of the linear transformation after concatenation.

7. A deformation early warning method for concrete dams based on the PDM-GAR model according to claim 1, characterized in that: In step S3), the deformation residual simulated by the concrete dam deformation simulation hybrid network model PDM-GAR is expressed as: where δ is the deformation residual; y i is the deformation observation value of the i-th horizontal monitoring point; is the deformation prediction value of the PDM-GAR of the concrete dam deformation simulation hybrid network model for the i-th horizontal monitoring point.

8. A method for early warning of concrete dam deformation based on the PDM-GAR model according to claim 1, characterized in that: In step S6), perform a safety score score according to the simulated deformation residual of each measuring point and its corresponding cumulative distribution knowledge base; specifically as follows: S61), comparing the simulated deformation residuals of each measurement point on the current day with the cumulative distribution in its residual library to obtain the quantile corresponding to the deformation residual; the quantile δ cdf is calculated by the following formula: where δ is the deformation residual simulated by the PDM-GAR of the concrete dam deformation simulation hybrid network model; index(·) is to obtain the sequence number of the measured point residual in the residual library; n is the number of data in the residual library; δ cdf is the quantile of the residual in the residual library; S62)、Perform scoring based on the quantile and the score calculation formula. The score calculation formula is: Among them, the scoring rule is set by combining the normal distribution and the distribution of the residual library, and the cumulative distribution values corresponding to the standard deviations of the normal distribution are used to determine the segmentation criteria for scoring.

9. A method for early warning of concrete dam deformation based on the PDM-GAR model according to claim 1, characterized in that: In step S7), the warning rule is as follows: If the cumulative distribution probability warning index is [0.159, 0.977], and the score range is [100, 90), then the warning level is "normal"; If the cumulative distribution probability warning index is (0.023, 0.159) ∪ (0.841, 0.977), and the score range is [100, 90), then the warning level is "Level 4 warning"; If the cumulative distribution probability warning index is (0.001, 0.023) ∪ (0.977, 0.999), and the score range is [80, 70), then the warning level is "Level 3 warning"; If the cumulative distribution probability warning index is outside the residual library sequence, and the score range is [0, 60], then the warning level is "Level 1 warning".

10. A method for early warning of concrete dam deformation based on the PDM-GAR model according to claim 1, characterized in that: In step S8), use the measuring point partition weighted scoring algorithm to evaluate the overall safety status of the dam, specifically including the following steps: S81)、Adopt the Ward hierarchical clustering method, and cluster the measuring points into N categories according to the historical deformation data of all deformation monitoring points of the dam; S82)、Calculate the within-class variance of each category and use it as the weight w of the corresponding category; S83) Calculate the safety score score of the overall dam according to the following formula dam , that is: where score i is the safety score of measuring point i, and w i is the weight value of the corresponding category of measuring point i.

Images