Concrete dam multi-measuring-point deformation depth prediction method based on data deconstruction

Abstract

The invention discloses a concrete dam multi-measuring-point deformation depth prediction method based on data deconstruction, and belongs to the technical field of dam operation safety monitoring and management. A concrete dam multi-measuring-point deformation monitoring sequence is decomposed into a plurality of relatively stable subsequences with different frequencies through MVMD, and the complexity of the multi-measuring-point deformation sequence is reduced; in combination with a CNN local feature extraction capability and a Transform global dependence attention capability, prediction is carried out on decomposed subsequences, and measurement point deformation data are reconstructed; project examples show that when the concrete dam multi-measuring-point deformation prediction model is constructed, the prediction performance is higher by considering the spatial relevance of the deformation sequence and the similarity of deformation of different measuring points, the calculation precision and efficiency are obviously improved compared with those of a traditional method, the overall error is low, the arch dam deformation behavior can be efficiently tracked, and the method is suitable for large-scale popularization and application. And effective technical support and method reference are provided for accurately predicting the dam operation safety state under the dynamic change of the environmental load.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dam operation safety monitoring and management, and particularly relates to a concrete dam multi-measuring point deformation depth prediction method based on data decomposition. BACKGROUND

[0002] Dam deformation monitoring is an important means to control the deformation state and safe operation of concrete dams. There are numerous safety monitoring points for large concrete dams, which are widely distributed. The massive monitoring data brings difficulties to timely and accurately analyze and evaluate the safety of the dam. At present, the analysis of dam safety monitoring data commonly used in engineering is mainly based on the analysis of single measuring point sequence one by one. Not only is the workload large, but the single measuring point data may also be affected by observation errors and other factors, showing instability and uncertainty. It is difficult to determine whether the abnormal change of one or several measuring point data reflects the main trend of the dam.

[0003] So far, the research on dam deformation prediction methods can be mainly divided into physical methods and statistical methods. Since the mechanism of dam deformation involves complex mechanics, seepage, cracks and chemical action, it is very difficult to establish a physical model that can reflect the above processes. Statistical methods use mathematical methods to establish internal relationships between historical sample data and output deformation before the prediction time according to historical data. In recent years, deep learning has developed rapidly in the field of dam deformation prediction, which can effectively mine the hidden coupling relationship between input parameters and extract the time correlation of features, accurately reflect the power fluctuation characteristics, and thus have better prediction effect. There are many dam safety monitoring projects, and the monitoring information of the measuring points is rich. Using single project or single measuring point information to evaluate the safety and stability of the dam has one-sidedness and cannot truly reflect the overall state of the dam, so the development of multi-measuring point models has a certain trend.

[0004] Due to the strong nonlinearity and strong volatility of dam deformation sequences, in order to improve the prediction accuracy of deformation sequences, using modal decomposition method to reduce the complexity of input data before prediction can significantly improve the prediction effect. However, the above common modal decomposition methods need to decompose each measuring point deformation data one by one when decomposing multi-measuring point deformation data, which is not suitable for simultaneously processing multi-element time series observation information composed of water level, temperature and other environmental quantities and multi-measuring point deformation data. This easily ignores the coupling relationship and correlation of each measuring point deformation value in the time domain and frequency domain. SUMMARY

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies by using MVMD (Multivariate Mode Decomposition) to decompose multi-measurement point sequences into multiple relatively stable subsequences with different frequency scales, thereby reducing the complexity of multi-measurement point deformation sequences. Subsequently, to enhance the model's ability to fit the temporal and spatial features of the sequence, the local feature extraction capability of CNN (Convolutional Neural Network) and the global dependency attention capability of Transformer are combined to predict the decomposed components and reconstruct the measurement point deformation data. This method provides a novel approach for monitoring and predicting dam deformation during operation.

[0006] To achieve the above objectives, this invention provides a method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction, comprising the following steps: S1. Data collection and preprocessing: Collect environmental variable factors and dam deformation data from the dam safety monitoring system, eliminate outliers, and obtain time series of environmental quantities and deformation measurements at multiple monitoring points. S2. Multi-point deformation sequence mode decomposition: Using multivariate variational mode decomposition (MVMD), the correlation between deformation at different measuring points at the same elevation in the time and frequency domains is analyzed. The time-frequency decomposition of the multi-point deformation monitoring sequence is performed synchronously to obtain several sets of stable IMF components belonging to different measuring points, and the training set and test set are divided according to a set ratio. S3. CNN-Transformer Model Training and Evaluation: A CNN-Transformer prediction model is constructed by inputting environmental variables affecting the deformation of the measurement points and the IMF (Intrinsic Mode Components) based on MVMD decomposition, and the coefficient of determination R is used. 2 The root mean square error (RMSE) is used to evaluate the predictive performance of the CNN-Transformer model.

[0007] Preferably, S2 specifically includes: S201. Extract K multi-modulated oscillation signals containing C different measurement points from a multi-dimensional signal based on MVMD; S202. Determine the MVMD optimization objective and bandwidth estimation method; S203. Construct an MVMD-constrained optimization problem model; S204. Derive the augmented Lagrangian function and decompose the optimization problem; S205. Simplify the center frequency update process; S206. Analyze the impact of key MVMD parameters on the decomposition results and introduce the permutation entropy method.

[0008] Preferably, the specific details of extracting K multi-element modulated oscillation signals containing C different measurement points from the multi-element signal based on MVMD in S201 are as follows: Based on MVMD, K predefined multivariate modulated oscillation signals containing C different measuring points of dam deformation data are extracted from the input multivariate signal x(t). The expression is as follows: (1) In the formula, ; These are deformation signals at different measuring points.

[0009] Preferably, the specific details of the MVMD optimization objective and bandwidth estimation method specified in S202 are as follows: The set of multi-modulated oscillation signals in the input data is expressed as follows: (2) The function representing the multivariables that needs optimization in MVMD is defined as follows: (3) In the formula: For partial derivative operations over time; It is a complex exponential function; j is the imaginary unit; The center angular frequency of the k-th component; This is the analytic signal of the kth positive frequency. Estimating the bandwidth of a multi-modulated oscillator signal: by... One-sided spectral shift of all channels And by taking the Frobenius norm of the matrix to estimate the bandwidth of the multivariate modulated oscillation signal, then... The expression is as follows: (4) In the formula, This is the analytical modulation signal corresponding to mode k in channel c.

[0010] Preferably, the specific content of constructing the MVMD constrained optimization problem model in S203 is as follows: By clarifying the MVMD optimization objective and bandwidth estimation method in S202, the constrained optimization problem of MVMD is obtained, expressed as follows: (5) In the formula, It is the set of multi-modulated oscillation signals in channel c; for The set of center frequencies; This represents the original deformation monitoring signal at measurement point c.

[0011] Preferably, the specific details of deriving the augmented Lagrangian function and decomposing the optimization problem in S204 are as follows: Construct the augmented Lagrange function, as shown in the following expression: (6) In the formula, As a penalty factor; For Lagrange multipliers; The complex optimization problem shown in equation (6) is transformed into several simpler suboptimal problems by using the alternating direction multiplier method, as shown in the following expressions: (7) (8) (9) In the formula, n is the time step; n is the number of iterations; This refers to the signal components of the c-th measurement point and the k-th mode after the (n+1)-th iteration update; These are the modal components whose numbers are less than k after the (n+1)th iteration update; For modal components whose numbers are greater than or equal to k after the nth iteration update; These are all the center frequencies after the nth iteration update; The Lagrange multipliers are those updated in the nth iteration. The center angular frequency of the k-th component after the (n+1)-th iteration update; These are the signal components of all channels after the (n+1)th iteration update; The center frequency whose number is less than k after the (n+1)th iteration update; This represents the Lagrange multiplier after the (n+1)th iteration update; The Lagrange multiplier for the c-th channel after the nth iteration update; Equation (7) is equivalent to equation (10), and the expression is as follows: (10) In the formula, and These represent the corresponding time-domain signal and its frequency-domain signal after Fourier transform, respectively. Indicates the center frequency. Indicates the corresponding central angle is The signal components at the c-th measurement point and the i-th mode; This represents the variable that minimizes the augmented Lagrange function.

[0012] Preferably, the simplified center frequency update process in S205 is as follows: The center frequency update process in equation (8) is simplified to equation (11), as follows: (11) According to Plancherel's theorem of the inner product of time-domain and frequency-domain functions, equation (11) is equivalent to equation (12), as shown below: (12) When the first derivative of equation (12) is 0, to minimize the sum of the quadratic functions, we have: (13) In the formula, This is expressed as the frequency of the corresponding center point after the nth iteration update. The signal components at the c-th measurement point and the k-th mode.

[0013] Preferably, the specific details of introducing the permutation entropy method in S206 to analyze the impact of key MVMD parameters on the decomposition results are as follows: The MVMD decomposition result depends on K and α. The value of K determines the accuracy of the final decomposition result. If the value of K is too small, the mode decomposition will be incomplete. If the value of K is too large, over-decomposition will occur. The value of α has an impact on MVMD performance that is related to the signal and noise. Introducing the permutation entropy method, the expression is as follows: (14) In the formula: For the first The probability of each symbol appearing; The complexity and randomness of the time series; The larger the value, the greater the time series detail. The smaller the value, the more regular the time series.

[0014] Preferably, the specific content of S3 is as follows: S301. CNN is used to extract features from dam deformation monitoring data. The CNN architecture includes convolutional layers, pooling layers and fully connected layers. S302. The convolutional layer captures spatial patterns of data through mathematical convolution operations, which is used to extract the local spatiotemporal correlation features between dam deformation data and environmental quantities, and generate feature maps. S303 and the pooling layer downsample the feature map, reducing network complexity by compressing the data dimension and preserving the invariance of key features; the fully connected layer flattens the abstract features and maps them to predicted values ​​through nonlinear transformation, outputting continuous deformation in deformation prediction tasks. S304. Extract feature values ​​and refine and optimize them; S305. Introducing the Transformer encoder module, which coordinates local feature extraction and global dependency modeling to improve the accuracy of deformation inflection point prediction; S306. Based on the Transformer architecture, improve task adaptability and optimize the input and output layer design to enhance the accuracy of arch dam deformation prediction, including: removing the position encoding layer and simplifying the output layer normalization; S307. Generate an improved Transformer that maintains the encoder-decoder architecture. The encoder consists of an input layer and multiple stacked processing units. The decoder adopts the same stacked structure and includes three types of attention mechanisms: a self-attention mechanism to resolve the temporal correlation within the deformation sequence; a masked attention mechanism to ensure causal constraints in the prediction process; and a cross-module attention mechanism to bridge the dynamic coupling relationship between environmental quantities and deformation quantities. The encoder inputs historical environmental quantities and dam deformation monitoring data, and the decoder gradually receives future environmental quantities and intermediate deformation prediction results. Through iterative iteration, a multi-step deformation prediction sequence is generated, and the prediction value of each step is fed back to the decoder in real time for subsequent calculations.

[0015] Preferably, the specific details of extracting feature values ​​and refining and optimizing the feature values ​​in S304 are as follows: S3041, Feature extraction stage: The first convolutional layer captures the short-range spatial dependency of deformed data, which is then passed to the second convolutional layer after ReLU activation; S3042, Feature Deepening Stage: Secondary convolution extracts high-order feature patterns, and ReLU is used to enhance non-linear expressive power; S3043, Feature Optimization Stage: The max pooling layer compresses redundant information and retains significant features that are sensitive to dam safety diagnosis; the batch normalization layer stabilizes the feature distribution, accelerates model convergence, and suppresses gradient anomalies.

[0016] Therefore, the present invention employs the above-mentioned data-based multi-point deformation depth prediction method for concrete dams, which has the following beneficial effects: (1) Sequence decomposition and dimensionality reduction: MVMD is used to decompose multi-point sequences into multiple relatively stable subsequences with different frequency scales, thereby reducing the complexity of multi-point deformed sequences.

[0017] (2) Model fusion and fitting: Combining the local feature extraction capability of CNN with the global dependency attention capability of Transformer, the decomposed components are predicted and the measurement point deformation data are reconstructed.

[0018] (3) Predictive value: It enables the synchronous prediction of deformation sequences at multiple measuring points, providing strong support for subsequent deformation prediction of dam and foundation during operation.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1This is a flowchart of the method for predicting the deformation depth of concrete dams based on data deconstruction, as described in this invention. Figure 2 Layout diagram of arch dam monitoring facilities; Figure 3 This is a graph showing the changes in upstream water level and local temperature. Figure 4 Deformation time sequence diagrams for different measuring points; Figure 5 for A graph showing how the value changes with the value of K; Figure 6 The diagram shows the decomposition of MVMD, where (a) is the TCN9 horizontal displacement monitoring sequence, (b) is the TCN4 horizontal displacement monitoring sequence, and (c) is the TCN16 horizontal displacement monitoring sequence. Figure 7 The following are the model prediction results: (a) is the deformation prediction result of TCN4 measuring point, (b) is the scatter plot of the prediction residual of TCN4 measuring point, (c) is the deformation prediction result of TCN9 measuring point, (d) is the scatter plot of the prediction residual of TCN9 measuring point, (e) is the deformation prediction result of TCN16 measuring point, and (f) is the scatter plot of the prediction residual of TCN16 measuring point. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0022] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0023] The following is combined with Figures 1-7 The embodiments of the present invention will be described in detail below.

[0024] Example 1 This invention provides a method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction, comprising the following steps: S1. Data collection and preprocessing: Collect environmental variable factors and dam deformation data from the dam safety monitoring system, eliminate outliers, and obtain time series of environmental quantities and deformation measurements at multiple monitoring points. S2. Multi-point deformation sequence mode decomposition: Using multivariate variational mode decomposition (MVMD), the correlation between deformation at different measuring points at the same elevation in the time and frequency domains is analyzed. The time-frequency decomposition of the multi-point deformation monitoring sequence is performed synchronously to obtain several sets of stable IMF components belonging to different measuring points, and the training set and test set are divided according to a set ratio. S2 specifically includes: S201. Extracting K multi-element modulated oscillation signals containing C different measurement points from a multi-element signal based on MVMD, the specific content is as follows: Based on MVMD, a predefined multivariate modulated oscillation signal Uk(t) containing C different measuring points of dam deformation data is extracted from the input multivariate signal x(t). The expression is as follows: (1) In the formula, ; These are deformation signals at different measuring points.

[0025] S202. Determine the MVMD optimization objective and bandwidth estimation method, the details of which are as follows: The set of multi-modulated oscillation signals in the input data is expressed as follows: (2) The requirement is to minimize the sum of the bandwidths of the extracted patterns while accurately reconstructing the original signal. To this end, based on the analytic vectors represented by equation (3), The bandwidth can be obtained through its analytic vector. The gradient function is estimated using the L2 norm. The multivariate function to be optimized in MVMD is then expressed as follows: (3) In the formula: For partial derivative operations over time; It is a complex exponential function; j is the imaginary unit; The center angular frequency of the k-th component; This is the analytic signal of the kth positive frequency. By One-sided spectral shift of all channels And by taking the Frobenius norm of the matrix to estimate the bandwidth of the multivariate modulated oscillation signal, then... The expression is as follows: (4) In the formula, This is the analytical modulation signal corresponding to mode k in channel c.

[0026] S203. Construct an MVMD constrained optimization problem model, the details of which are as follows: By clarifying the MVMD optimization objective and bandwidth estimation method in S202, the constrained optimization problem of MVMD is obtained, expressed as follows: (5) In the formula, It is the set of multi-modulated oscillation signals in channel c; for The set of center frequencies.

[0027] S204. Derive the augmented Lagrangian function and decompose the optimization problem. The details are as follows: The corresponding augmented Lagrangian function is expressed as follows: (6) In the formula, As a penalty factor; For Lagrange multipliers; The complex optimization problem shown in equation (6) is transformed into several simpler suboptimal problems by using the alternating direction multiplier method, as shown in the following expressions: (7) (8) (9) In the formula, n is the time step; n is the number of iterations; This refers to the signal components of the c-th measurement point and the k-th mode after the (n+1)-th iteration update; These are the modal components whose numbers are less than k after the (n+1)th iteration update; For modal components whose numbers are greater than or equal to k after the nth iteration update; These are all the center frequencies after the nth iteration update; The Lagrange multipliers are those updated in the nth iteration. The center angular frequency of the k-th component after the (n+1)-th iteration update; These are the signal components of all channels after the (n+1)th iteration update; The center frequency whose number is less than k after the (n+1)th iteration update; This represents the Lagrange multiplier after the (n+1)th iteration update; The Lagrange multiplier for the c-th channel after the nth iteration update; Formula (7) can be equivalent to: (10) In the formula, and These represent the corresponding time-domain signal and its frequency-domain signal after Fourier transform, respectively. Indicates the center frequency. Indicates the corresponding central angle is The signal components at the c-th measurement point and the i-th mode; This represents the variable that minimizes the augmented Lagrange function.

[0028] S205. Simplify the center frequency update process, the details of which are as follows: The center frequency update process shown in formula (8) can be simplified as follows: (11) According to Plancherel's theorem of the inner product of time-domain and frequency-domain functions, equation (11) is equivalent to: (12) Let the first derivative of formula (12) be 0, so that the sum of the quadratic functions is minimized, then we have: (13) In the formula, This is expressed as the frequency of the corresponding center point after the nth iteration update. The signal components at the c-th measurement point and the k-th mode.

[0029] S206. Analyze the impact of key MVMD parameters (K, α) on the decomposition results and introduce the permutation entropy method. The specific content is as follows: The quality of MVMD decomposition results mainly depends on K and α. The value of K determines the accuracy of the final decomposition result. If the value of K is too small, incomplete mode decomposition will occur; if the value of K is too large, over-decomposition will occur, failing to achieve the desired effect. The influence of the value of α on MVMD performance is more complex and closely related to signal and noise. Therefore, the settings of K and α are crucial for MVMD. Permutation entropy, proposed by Bandt et al., is a method for detecting the complexity and dynamical abrupt changes in time series. This method boasts advantages such as computational simplicity, robustness, and high computational efficiency, and its expression is as follows: (14) In the formula: For the first The probability of each symbol appearing; The complexity and randomness of the time series; The larger the value, the greater the time series detail. The smaller the value, the more regular the time series.

[0030] S3. Training and Evaluation of the CNN-Transformer Model: A CNN-Transformer prediction model is constructed by inputting environmental variables affecting the deformation of the measurement points and the IMF components based on MVMD decomposition, and the coefficient of determination R is used. 2 The root mean square error (RMSE) is used to evaluate the predictive performance of the CNN-Transformer model. The specific details of S3 are as follows: S301. Feature extraction of deformation monitoring data is performed using a convolutional neural network (CNN). A typical CNN architecture includes core components such as convolutional layers, pooling layers, and fully connected layers.

[0031] S302. The convolutional layer captures spatial patterns in the data through mathematical convolution operations: the convolutional kernel slides across the input sequence, performs a weighted dot product on the data within the local receptive field, and superimposes a bias term to generate a feature map. This architecture is used to extract the local spatiotemporal correlation features between dam deformation and environmental quantities (water level, temperature, etc.).

[0032] S303 and the pooling layer downsample the feature map, reducing network complexity by compressing the data dimension while preserving the invariance of key features; the fully connected layer flattens the abstract features and maps them to predicted values ​​through nonlinear transformation, outputting continuous deformation in deformation prediction tasks.

[0033] S304. Extract feature values ​​and refine and optimize them, as detailed below: S3041, Feature extraction stage: The first convolutional layer captures the short-range spatial dependency of deformed data, which is then passed to the second convolutional layer after ReLU activation; S3042, Feature Deepening Stage: Secondary convolution extracts high-order feature patterns, and ReLU is used to enhance non-linear expressive power; S3043, Feature Optimization Stage: The max pooling layer compresses redundant information and retains significant features that are sensitive to dam safety diagnosis; the batch normalization layer stabilizes the feature distribution, accelerates model convergence, and suppresses gradient anomalies.

[0034] This design significantly expands the model's receptive field, effectively extracting local abrupt changes in the dam deformation process (such as deformation inflection points caused by a sudden drop in water level), and providing highly discriminative feature representations for time series modeling.

[0035] S305. Although CNNs excel at extracting local features, their limited receptive field makes it difficult to capture long-term temporal dependencies in dam deformation (such as the hysteresis deformation effect caused by periodic fluctuations in reservoir water levels). In contrast, Transformers, with their global modeling capabilities, have become the mainstream architecture for time series analysis. They explicitly establish the correlation between any time step through a self-attention mechanism, enabling in-depth analysis of the cross-temporal coupling relationship between deformation and environmental factors. Therefore, this invention introduces a Transformer encoder module to coordinate local feature extraction and global dependency modeling, significantly improving the accuracy of deformation inflection point prediction.

[0036] S306. Based on the Transformer architecture, task adaptability improvements were made, and the input / output layer design was optimized to enhance the accuracy of arch dam deformation prediction. These improvements included: removal of the positional encoding layer: In natural language processing, positional encoding is used to parse unordered word order relationships. However, arch dam deformation sequences exhibit significant periodicity (e.g., annual temperature-deformation response), and their spatiotemporal order follows deterministic physical mechanisms. Retaining positional encoding may introduce noise interference, diluting the effective expression of key deformation features; output layer normalization and simplification: The original Softmax layer was removed because this design was specifically designed to serve the probabilistic output of classification tasks. Arch dam deformation prediction after material parameter updates is a regression problem; using a linear output layer is more suitable for the continuous prediction requirements of deformation values.

[0037] S307, the improved Transformer maintains the encoder-decoder architecture, where the encoder consists of an input layer and multiple stacked processing units. The decoder uses the same stacked structure and includes three types of attention mechanisms: self-attention to resolve temporal correlations within the deformation sequence; masked attention to ensure causal constraints in the prediction process; and cross-module attention to bridge the dynamic coupling relationship between environmental variables and deformation variables. The encoder inputs historical environmental variables and deformation monitoring data, while the decoder progressively receives future environmental variables and intermediate deformation prediction results. Multi-step deformation prediction sequences are generated through iterative iteration, with each prediction value fed back to the decoder in real time for subsequent calculations.

[0038] Example 2 A method for predicting the deformation depth of concrete dams based on data deconstruction, the flowchart of which is as follows: Figure 1 As shown in the example, this embodiment uses a concrete double-curvature arch dam located in Panzhihua City, Sichuan Province. The maximum dam height is 240m, ranking first in Asia and third in the world. The crest elevation is 1188.5m, the normal reservoir water level is 1200m, and the reservoir capacity is 5.8 billion cubic meters. 3 Adjusting reservoir capacity by 3.37 billion cubic meters 3 The design flow rate is 7400 m³ / s. 3 / s, the verification flow rate is 7600m³ / s. 3 The dam consists of 39 sections.

[0039] To ensure the operational safety and design verification of the concrete arch dam and underground powerhouse, a comprehensive monitoring system, including environmental monitoring, deformation monitoring, seepage monitoring, stress-strain monitoring, and temperature monitoring, has been installed on the surfaces and interiors of the main structures. To accurately monitor the deformation effects of the concrete dam, the dam deformation monitoring system primarily uses vertical lines, inverted vertical lines, dam crest tension lines, and line-of-sight. Eight inverted vertical lines and ten vertical lines are arranged within the dam body, forming five groups of vertical and inverted vertical lines, totaling 20 measuring points. Figure 2 As shown in the figure, TCN1-TCN20 represent the vertical measuring point numbers, EX1-EX7 represent the tension wire measuring point numbers, ①- This indicates that the dam is divided into 39 sections. Monitoring is conducted using both manual and automated methods, with deformation parallel and perpendicular to the dam axis monitored using a plumb line method. The process is shown below: This invention uses monitoring data at the same elevation as an example to establish a model, selecting the deformation of the dam body along the river at an elevation of 1171.05m as the research object. Three known vertical measuring points, TCN4, TCN9, and TCN16, are selected. Measuring points TCN4 and TCN16 are located in the non-overflow dam section, while measuring point TCN9 is located in the overflow dam section. A total of 1096 data sets were selected from January 1, 2019 to December 31, 2022. Of these, 1000 data sets from January 1, 2019 to September 26, 2021 were used as training samples, and 100 data sets from September 26, 2021 to December 31, 2022 were selected as prediction samples. The equipment used to measure the deformation at the three measuring points is a vertical coordinate instrument, and monitoring is conducted automatically twice a day. Figure 3 This data represents the measured upstream water level and local temperature during this time period. Figure 4 Deformation time history diagrams for three measuring points: TCN4, TCN9, and TCN16. Figure 4 It can be seen that the curves formed by the data measured at the three measuring points rise and fall together, showing a basically consistent trend. Even in some special periods (such as from July 2, 2020 to July 22, 2020), when there is a sudden increase in measured values, the three curves still maintain the same trend. This shows that there is a certain correlation between the deformation values ​​of each point when the dam body is in the same direction and at the same elevation. The observed signal x(t) is decomposed using MVMD. First, the variational mode number K is determined. With the default value of α=2000, the channels are calculated for different K values. For example... Figure 5 As shown. By Figure 5 It can be seen that among the deformation signals of the three different measuring points, The value of K shows a clear turning point when K is 3, therefore K=3 is determined; Choose the penalty factor α value. When K=3, select different α values ​​within [100, 10000]. After trial and error, we can obtain α=500. The IMF decomposition results of the three measurement points obtained from this are as follows: Figure 6 As shown; The CNN-Transformer prediction model has multiple parameters, and the selection of these parameters directly affects the final prediction performance. This invention utilizes orthogonal experimental design to select and optimize the parameters of the CNN-Transformer, and the final complete network parameters are shown in Table 1. Table 1 CNN-Transformer Network Parameters

[0040] To explore the effectiveness of MVMD decomposition in improving model prediction accuracy, this paper uses the IMF components after MVMD decomposition as input to the CNN-Transformer model. The results are compared with the deformation sequence of measurement points without decomposition. Figure 7 As shown, from Figure 7 As can be seen from (a), (c) and (e) in the figure, after training, the predicted value after decomposing and reconstructing the monitoring sequence is close to the actual deformation of the measuring point. The curves of the two changing over time have a high degree of consistency, and the predicted result curve has a high degree of overlap with the measured sinking velocity curve. Figure 7 Figures (b), (d), and (f) more clearly demonstrate the closeness between the predicted results after MVMD decomposition and the actual values; the scatter points basically fall on the dotted lines in the graph. Table 2 summarizes the evaluation index values ​​of the predicted results for the training set, test set, and all samples. As can be seen from Table 1, the RMSE of all three models is less than 0.42 and R0 is less than 0.42. 2 All values ​​are greater than 0.90, indicating good prediction performance and a small overall error in the predicted values. Furthermore, the RMSE of the combined MVMD and CNN-Transformer model is closer to 0 than that of the CNN-Transformer model alone. 2 The value is closer to 1. This indicates that the constructed network model fully recognizes the data features of the training set and is well applied to the test set for deformable value prediction. Although no overfitting prevention measures were taken, the model still shows strong generalization ability. To further assess the effectiveness of the proposed prediction model, the performance of three different prediction models was calculated, primarily including the predicted RMSE and R0. 2 The results are shown in Table 2. The table shows that the proposed model has the lowest RMSE error compared to other models, and the prediction accuracy R0 is also the highest. 2It also reached its highest value. This indicates that MVMD effectively reduced the interference of dam deformation data complexity on prediction accuracy, enabling deep learning models to more effectively extract deeper features from the data.

[0041] Table 2 Statistical indicators of the models built at different measurement points

[0042] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction, characterized in that, Includes the following steps: S1. Data collection and preprocessing: Collect environmental variable factors and dam deformation data from the dam safety monitoring system, eliminate outliers, and obtain time series of environmental quantities and deformation measurements at multiple monitoring points. S2. Multi-point deformation sequence mode decomposition: Using multivariate variational mode decomposition (MVMD), the correlation between deformation at different measuring points at the same elevation in the time and frequency domains is analyzed. The time-frequency decomposition of the multi-point deformation monitoring sequence is performed synchronously to obtain several sets of stable IMF components belonging to different measuring points, and the training set and test set are divided according to a set ratio. S3. Training and Evaluation of the CNN-Transformer Model: A CNN-Transformer prediction model is constructed by inputting environmental variables affecting the deformation of the measurement points and the IMF components based on MVMD decomposition, and the coefficient of determination R is used. 2 The root mean square error (RMSE) is used to evaluate the predictive performance of the CNN-Transformer model.

2. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 1, characterized in that, S2 specifically includes: S201. Extract K multi-modulated oscillation signals containing C different measurement points from a multi-dimensional signal based on MVMD; S202. Determine the MVMD optimization objective and bandwidth estimation method; S203. Construct an MVMD-constrained optimization problem model; S204. Derive the augmented Lagrangian function and decompose the optimization problem; S205. Simplify the center frequency update process; S206. Analyze the impact of key MVMD parameters on the decomposition results and introduce the permutation entropy method.

3. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 2, characterized in that, The specific details of extracting K multi-element modulated oscillation signals containing C different measurement points from a multi-element signal based on MVMD in S201 are as follows: Based on MVMD, K predefined multivariate modulated oscillation signals containing C different measuring points of dam deformation data are extracted from the input multivariate signal x(t). The expression is as follows: （1） In the formula, ; These are deformation signals at different measuring points.

4. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 3, characterized in that, S202 specifies the MVMD optimization objective and bandwidth estimation method as follows: The set of multi-modulated oscillation signals in the input data is expressed as follows: （2） The function representing the multivariables that needs optimization in MVMD is defined as follows: （3） In the formula: For partial derivative operations over time; It is a complex exponential function; j is the imaginary unit; The center angular frequency of the k-th component; This is the analytic signal of the kth positive frequency. Estimating the bandwidth of a multi-modulated oscillator signal: by... One-sided spectral shift of all channels The bandwidth of the multivariate modulated oscillation signal is estimated by taking the Frobenius norm of the matrix. The expression is as follows: （4） In the formula, This is the analytical modulation signal corresponding to mode k in channel c.

5. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 4, characterized in that, The specific details of constructing the MVMD-constrained optimization problem model in S203 are as follows: By clarifying the MVMD optimization objective and bandwidth estimation method in S202, the constrained optimization problem of MVMD is obtained, expressed as follows: （5） In the formula, It is the set of multi-modulated oscillation signals in channel c; for The set of center frequencies; This represents the original deformation monitoring signal at measurement point c.

6. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 5, characterized in that, The specific details of deriving the augmented Lagrange function and decomposing the optimization problem in S204 are as follows: Constructing the augmented Lagrangian function The expression is as follows: （6） In the formula, As a penalty factor; For Lagrange multipliers; The complex optimization problem shown in equation (6) is transformed into several simpler suboptimal problems by using the alternating direction multiplier method, as shown in the following expressions: （7） （8） （9） In the formula, n is the time step; n is the number of iterations; This refers to the signal components of the c-th measurement point and the k-th mode after the (n+1)-th iteration update; These are the modal components whose numbers are less than k after the (n+1)th iteration update; For modal components whose numbers are greater than or equal to k after the nth iteration update; These are all the center frequencies after the nth iteration update; The Lagrange multipliers are those updated in the nth iteration. The center angular frequency of the k-th component after the (n+1)-th iteration update; These are the signal components of all channels after the (n+1)th iteration update; The center frequency whose number is less than k after the (n+1)th iteration update; This represents the Lagrange multiplier after the (n+1)th iteration update; The Lagrange multiplier for the c-th channel after the nth iteration update; Equation (7) is equivalent to equation (10), and the expression is as follows: （10） In the formula, and These represent the corresponding time-domain signal and its frequency-domain signal after Fourier transform, respectively. Indicates the center frequency. Indicates the corresponding central angle is The signal components at the c-th measurement point and the i-th mode; This represents the variable that minimizes the augmented Lagrange function.

7. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 6, characterized in that, The specific details of the simplified center frequency update process in S205 are as follows: The center frequency update process in equation (8) is simplified to equation (11), as follows: （11） According to Plancherel's theorem of the inner product of time-domain and frequency-domain functions, equation (11) is equivalent to equation (12), as shown below: （12） When the first derivative of equation (12) is 0, to minimize the sum of the quadratic functions, we have: （13） In the formula, The frequency of the corresponding center point after the nth iteration update is expressed as... The signal components at the c-th measurement point and the k-th mode.

8. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 7, characterized in that, The specific details of introducing the permutation entropy method in S206 to analyze the impact of key MVMD parameters on decomposition results are as follows: The MVMD decomposition result depends on K and α. The value of K determines the accuracy of the final decomposition result. If the value of K is too small, the mode decomposition will be incomplete. If the value of K is too large, over-decomposition will occur. The value of α has an impact on MVMD performance that is related to the signal and noise. Introducing the permutation entropy method, the expression is as follows: （14） In the formula: For the first The probability of each symbol appearing; The complexity and randomness of the time series; The larger the value, the greater the time series detail. The smaller the value, the more regular the time series.

9. The method for predicting the deformation depth of concrete dams at multiple measuring points based on data deconstruction according to claim 1, characterized in that, The specific details of S3 are as follows: S301. CNN is used to extract features from dam deformation monitoring data. The CNN architecture includes convolutional layers, pooling layers and fully connected layers. S302. The convolutional layer captures spatial patterns of data through mathematical convolution operations, which is used to extract the local spatiotemporal correlation features between dam deformation data and environmental quantities, and generate feature maps. S303 and pooling layers downsample the feature maps, reducing network complexity by compressing data dimensions; fully connected layers flatten abstract features and map them to predicted values ​​through nonlinear transformations, outputting continuous deformation amounts in deformation prediction tasks. S304. Extract feature values ​​and refine and optimize them; S305. Introducing the Transformer encoder module, which coordinates local feature extraction and global dependency modeling to improve the accuracy of deformation inflection point prediction; S306. Improve task adaptability based on Transformer architecture, including: removing the positional encoding layer and simplifying output layer normalization; S307. Generate an improved Transformer that maintains the encoder-decoder architecture. The encoder consists of an input layer and multiple stacked processing units. The decoder adopts the same stacked structure and includes three types of attention mechanisms: a self-attention mechanism to resolve the temporal correlation within the deformation sequence; a masked attention mechanism to ensure causal constraints in the prediction process; and a cross-module attention mechanism to bridge the dynamic coupling relationship between environmental quantities and deformation quantities. The encoder inputs historical environmental quantities and dam deformation monitoring data, and the decoder gradually receives future environmental quantities and intermediate deformation prediction results. Through iterative iteration, a multi-step deformation prediction sequence is generated, and the prediction value of each step is fed back to the decoder in real time for subsequent calculations.

10. The method for predicting the deformation depth of a concrete dam based on data deconstruction according to claim 9, characterized in that, The specific steps for extracting, refining, and optimizing feature values ​​in S304 are as follows: S3041, Feature extraction stage: The first convolutional layer captures the short-range spatial dependency of deformed data, which is then passed to the second convolutional layer after ReLU activation; S3042, Feature Deepening Stage: Secondary convolution extracts high-order feature patterns, and ReLU is used to enhance non-linear expressive power; S3043, Feature Optimization Stage: The maximum pooling layer compresses redundant information and retains significant features that are sensitive to dam safety diagnosis; Batch normalization layers stabilize feature distribution, accelerate model convergence, and suppress gradient anomalies.

Images