Gravity dam deformation monitoring method based on intelligent analysis
By combining the CNN-BiLSTM-Attention model and the Northern Eagle optimization algorithm to optimize hyperparameters, the nonlinearity problem of concrete dam deformation monitoring data was solved, and high-precision deformation prediction was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RAILWAY FIRST SURVEY & DESIGN INST GRP
- Filing Date
- 2024-10-31
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to establish simple and clear mathematical relationships to reflect the non-stationary and nonlinear time series of concrete dam deformation, which affects the accuracy of deformation monitoring data and prediction accuracy.
A combined model of convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM) and attention mechanism was adopted, and hyperparameters were optimized by the Northern Eagle optimization algorithm to establish a gravity dam deformation monitoring model and improve prediction accuracy.
This improves the prediction accuracy and long-term prediction performance of gravity dam deformation monitoring data, and provides a high-precision deformation prediction method.
Smart Images
Figure CN121960088A_ABST
Abstract
Description
A Gravity Dam Deformation Monitoring Method Based on Intelligent Analysis Technical Field
[0001] This invention belongs to the field of gravity dam deformation monitoring technology, specifically relating to a gravity dam deformation monitoring method based on intelligent analysis. Background Technology
[0002] The most significant form of failure for concrete dams is foundation sliding; therefore, dam deformation is the most direct reflection of their safety status and is recognized by experts both domestically and internationally as the most reliable indicator. Based on different theories, dam deformation prediction models can be categorized into statistical models, deterministic models, and hybrid models. Among these, statistical models use environmental variables as independent variables and monitoring effect sizes as dependent variables, allowing for the explanation of dam deformation mechanisms by uncovering the inherent patterns in monitoring data.
[0003] During actual dam service, deformation is affected by various factors such as water pressure and temperature. Deformation monitoring data is typically a complex, non-stationary, and nonlinear time series, making it difficult to establish a simple and clear mathematical relationship. To better reflect the time-varying characteristics of the concrete dam structure response and meet the accuracy requirements of dam monitoring data analysis, deformation monitoring models combined with artificial intelligence algorithms, such as Support Vector Regression (SVR), Random Forest (RF), and Long Short-Term Memory (LSTM), have developed rapidly. Compared to traditional statistical models, the selection of hyperparameters in artificial intelligence models affects their prediction accuracy and increases model complexity. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, this invention provides a gravity dam deformation monitoring method based on intelligent analysis. It utilizes the Northern Eagle optimization algorithm to optimize hyperparameters, including the initial learning rate, the number of BiLSTM neurons, the number of key values in the self-attention mechanism, and regularization parameters, in order to improve prediction accuracy.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A gravity dam deformation monitoring method based on intelligent analysis, characterized by the following steps:
[0007] S1: Select influencing factors and establish the model input factor set;
[0008] S2: Data partitioning: Divide the original dataset into training and test sets according to a random proportion or chronological order;
[0009] S3: Model Training: Train the CNN-BiLSTM-Attention model on the training set to enable the model to learn the features of the measured deformed data and their correlation with various factors.
[0010] S4: Parameter optimization: The hyperparameters of the CNN-BiLSTM-Attention model are optimized using the Northern Eagle optimization algorithm. The optimal hyperparameter combination is applied to the CNN-BiLSTM-Attention model to form a concrete dam deformation prediction model.
[0011] S5: Model Evaluation: Using root mean square error, mean relative percentage error, and coefficient of determination R0 2 And the multiple correlation coefficient R is used as an error evaluation index for the model;
[0012] S6: Substitute the explanatory variables from the prediction sample set into the optimized CNN-BiLSTM-Attention concrete dam deformation prediction model to achieve the prediction of concrete dam deformation.
[0013] Furthermore, in step S1, the model input factor set includes: water pressure factor Hup, Hup 2 Hup 3 Hup 10d Hup 15d Hup 30d Hup 60d Hup 90d Hup 120d ,Hd,Hd 2 ,Hd 3 ,Hd 10d ,Hd 15d ,Hd 30d ,Hd 60d ,Hd 90d ,Hd 120d This refers to the 1st, 2nd, and 3rd powers of the upstream and downstream water levels, as well as the average upstream water level over the preceding 2, 3, 5, 7, 10, 15, 30, 60, 90, and 120 days; temperature factor T, T 5d ,T 10d ,T 15d ,T 30d ,T 60d ,T 90d ,T 120d This refers to the daily air temperature and its average temperature over the preceding 5, 10, 15, 30, 60, 90, and 120 days; the time factor s, s 2 ,s 3 , where s is the number of monitoring days / 100.
[0014] Furthermore, in step S2, the data partitioning can be carried out by randomly dividing the input factor set and the corresponding deformation monitoring value into training set and test set according to a ratio or according to time order.
[0015] Furthermore, in step S3, the model training process involves using the deformation explanatory variable factors in the training set as input and the measured deformation values of the concrete dam as output to train the CNN-BiLSTM-Attention model.
[0016] Furthermore, the parameter optimization described in step S4 uses the Northern Eagle Optimization Algorithm (NGO) to optimize the hyperparameters of the CNN-BiLSTM-Attention model based on the data in the training set. The specific method is as follows:
[0017] First, it is necessary to initialize the population size and iteration count of the Northern Eagle optimization algorithm, and set the search range of the model parameters to be optimized. With the goal of minimizing the mean squared error, optimize the initial learning rate, the number of BiLSTM neurons, the number of key values of the self-attention mechanism, and the regularization parameter in the model. Set the search range [LB,UB] of the above four parameters to [0.001,0.01], [10,50], [2,50], and [0.0001,0.001], respectively.
[0018] Then, the Northern Eagle optimization algorithm is used to set initial values for each parameter (four in this case):
[0019] x ij =LB ij +r1×(UB ij -LB ij )
[0020] i = 1, 2, ..., n, j = 1, 2, 3, 4
[0021] Where, x ij It is the initial value of the i-th population in the j-th parameter, where n is the population size; UB i LB i These are the upper and lower boundaries of the search domain, respectively; r1 is a random number between (0,1);
[0022] In the first stage of prey identification, a global search is performed on the search space to determine the optimal region. The population then randomly selects an individual from the search space and launches an attack. The mathematical model for this is as follows:
[0023]
[0024] p i =X k k = 1, 2, ..., i-1, i+1, ..., n
[0025] Where j is the parameter type number, i are the population type numbers, and x ij F represents the position of the population at the current iteration. iLet it be the value of its objective function; The new position of the population after iterative update; p ij For a random individual that is different from the current individual; F pi Let r1 and r2 be random numbers between (0,1); and I be a random integer of 1 or 2.
[0026] After the population position is updated, the decision on whether to adopt the updated population position is based on the objective function values before and after the update, as shown in the following expression:
[0027]
[0028] In the second phase, the northern goshawk relentlessly pursues and captures its prey at a relatively high speed. Assuming this hunting behavior approximates an attack pattern with a radius of R, the mathematical expression for this phase is as follows:
[0029]
[0030] Where, x ij This represents the position of the population at the current iteration. The new position of the population after iterative update; r4 is a random number between (0,1); t is the current iteration number; T is the maximum iteration number. Similar to the first phase, the expression for determining whether to update the population position is as follows:
[0031]
[0032] By repeatedly going through the first and second stages until the maximum number of iterations is reached, the optimal value of the parameter to be optimized is obtained.
[0033] Furthermore, in step S5, the root mean square error, mean relative percentage error (MAPE), coefficient of determination R², and multiple correlation coefficient R are used as error evaluation indicators for the model.
[0034] The beneficial effects of this invention are:
[0035] 1) This invention targets gravity dam deformation monitoring data and employs a CNN-BiLSTM-Attention combination method with excellent nonlinear fitting and prediction capabilities to establish a model for predicting the deformation behavior of concrete dams. Considering that the parameters of CNN-BiLSTM-Attention have a certain impact on its training accuracy and generalization performance, the Northern Eagle Optimization Algorithm (NGO) is introduced to optimize its hyperparameters. Through iterative optimization calculation, an optimized CNN-BiLSTM-Attention model is obtained, and gravity dam deformation prediction is performed based on this optimal model.
[0036] 2) This invention has excellent prediction accuracy and long-term prediction performance, providing a new method for high-precision prediction of dam deformation. Attached Figure Description
[0037] Figure 1 is a flowchart of the modeling process of this invention. Detailed Implementation
[0038] The present invention will now be described in detail with reference to specific embodiments.
[0039] The operating environment of dams is complex, and there is often a significant nonlinear relationship between their deformation and external loads. This invention, targeting gravity dam deformation monitoring data, combines a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism to establish a gravity dam deformation monitoring model. Furthermore, the Northern Eagle Optimization Algorithm (NGO) is introduced to optimize hyperparameters for gravity dam deformation prediction. This invention can accurately predict dam deformation, showing a significant improvement in prediction accuracy compared to unoptimized CNN-BiLSTM-Attention models, BiLSTM models, and support vector regression (SVR) models.
[0040] As shown in Figure 1, the present invention includes the following steps:
[0041] S1: Select influencing factors and establish the model input factor set;
[0042] In step S1, establishing the model input factor set requires creating the dependent variable (measured value y = {y1, y2, ..., y3}) based on the deformation theory of concrete dams, and according to the deformation measurements, upstream and downstream water levels, temperature, and service days. n The model consists of a set of explanatory variables (model factors) and a corresponding set of explanatory variables. The model factors are all relative values between the monitoring date and the initial date of the modeling sequence.
[0043] For gravity dams, the explanatory variables for deformation include: water pressure factor Hup. 2 Hup 3 Hup 10d Hup 15d Hup 30d Hup 60d Hup 90d Hup 120d ,Hd,Hd 2 ,Hd 3 ,Hd 10d ,Hd 15d ,Hd 30d ,Hd 60d ,Hd 90d ,Hd 120dThis refers to the 1st, 2nd, and 3rd powers of the upstream and downstream water levels, as well as the average upstream water level over the preceding 2, 3, 5, 7, 10, 15, 30, 60, 90, and 120 days; temperature factor T, T 5d ,T 10d ,T 15d ,T 30d ,T 60d ,T 90d ,T 120d The average temperature of the monitoring day and the preceding 5, 10, 30, ... 120 days; the time factor s, s 2 ,s 3 , where s is the number of monitoring days / 100.
[0044] To avoid multicollinearity in the model factor set, the Bayesian Model Selection (BMS) algorithm is used to determine whether the above factors should be used as model inputs based on the BIC criterion.
[0045] S2: Data partitioning: In order to effectively train and evaluate the model, the original dataset is randomly divided into training set and test set according to proportion or time order;
[0046] In step S2, the data can be partitioned proportionally, randomly dividing the input factor set and its corresponding deformation monitoring values into training and testing sets, or according to time sequence. For example, the first 90% (or other proportions, such as 75%) of the data can be used for training, and the last 10% can be used for validation; or a certain proportion (such as 90%, 75%, etc.) of the data can be randomly selected as the training set, and the remaining data can be used as the validation set.
[0047] S3: Model Training: Train the CNN-BiLSTM-Attention model on the training set to enable the model to learn the features of the measured deformed data and their correlation with various factors; then apply the trained model to the test set to evaluate its performance and prediction accuracy.
[0048] In step S3, the model training process involves using the deformation explanatory variable factors in the training set as input and the measured deformation values of the concrete dam as output to train the CNN-BiLSTM-Attention model.
[0049] Among them, CNN is an effective feedforward neural network, mainly used for data feature extraction. It mainly consists of input layers, convolutional layers, and pooling layers, with neurons in each layer interconnected, but neurons in the same layer being unconnected.
[0050] The convolutional layer performs the convolution operation between the input and the weights, that is:
[0051] a 1 =σ1(b1+ξ * a0 )
[0052] In the formula: a 1 — Convolutional layer output; σ1 — Activation function; b1 — Bias; ξ — Filter matrix; * — Inner product symbol; a 0 --enter.
[0053] Pooling layers optimize the output of convolutional layers and reduce feature dimensionality, primarily involving two processes: max pooling and square root pooling. Fully connected layers map the output of convolutional layers to a classification label space for use by the output layer. This invention utilizes a CNN to extract features from concrete dam deformation monitoring data and employs max pooling in the pooling layer, using a fixed-size sliding window to reduce the dimensionality of the extracted data features. This paper uses a convolutional neural network structure with two convolutional layers, one pooling layer, and one fully connected layer to achieve secondary feature extraction, inputting the data after two convolutional pooling operations into a BiLSTM layer.
[0054] BiLSTM processes sequence data by running two LSTM networks simultaneously in the time dimension: one from front to back (forward) and the other from back to front (backward). By concatenating or merging information from the two directions, BiLSTM can more comprehensively understand and represent sequence data. The update process of the BiLSTM network involves the following formula:
[0055]
[0056] In the formula: —Output of the forward layer; — Output of the reverse layer; Y t —The merged output of the two layers; W y —The weights of the output layer; b y — Output layer bias; [;] — Connect operation.
[0057] The attention layer assigns different weights to different input features to enhance important features and prevent irrelevant information from affecting the final result. Specifically, a scoring function is first used, the expression of which is:
[0058] s t =tanh(W h h t +b h )
[0059] Normalization is performed using the softmax function to obtain the weights of each input vector:
[0060]
[0061] In the formula: W y b represents the weights of Attention; y The bias of the attention; v is the attention value; Let α be the output of the Attention layer at time t; t These are the weighting coefficients.
[0062] S4: Parameter optimization: The hyperparameters of the CNN-BiLSTM-Attention model are optimized using the Northern Eagle Optimization Algorithm (NGO). The optimal hyperparameter combination is applied to the CNN-BiLSTM-Attention model to form a concrete dam deformation prediction model.
[0063] The optimization method is as follows:
[0064] The Northern Eagle Optimization Algorithm (NGO) is mainly divided into two stages: the exploration stage, which involves identifying and attacking the prey; and the development stage, which involves chasing and the prey attempting to escape. The main adjustable parameters involved in the CNN-BiLSTM-Attention model include the initial learning rate, the number of BiLSTM neurons, the number of keys in the self-attention mechanism, and the regularization parameter. Since the interaction between these parameters is minimal, a random search method is chosen for hyperparameter optimization to achieve the optimal model configuration.
[0065] First, it is necessary to initialize the two parameters of the Northern Eagle Optimization Algorithm: population size and number of iterations, and set the search range of the model parameters to be optimized. For example, set the population size of the NGO algorithm to n (e.g., 10 or 20), the maximum number of iterations to 50 or more, and optimize the initial learning rate, the number of BiLSTM neurons, the number of key values of the self-attention mechanism, and the regularization parameter in the model with the goal of minimizing the mean squared error. Set the search range [LB, UB] of the above four parameters to [0.001, 0.01], [10, 50], [2, 50], and [0.0001, 0.001], respectively.
[0066] Then, the Northern Eagle optimization algorithm is used to set initial values for each parameter (four in this case):
[0067] x ij =LB ij +r1×(UB ij -LB ij )
[0068] i = 1, 2, ..., n, j = 1, 2, 3, 4
[0069] Where, x ij It is the initial value of the i-th population in the j-th parameter, where n is the population size; UB i LBi These are the upper and lower boundaries of the search domain, respectively; r1 is a random number between (0,1).
[0070] In the first stage, prey identification (exploration stage), since the selection of prey in the search space is random, this stage enhances the exploration capability of the NGO algorithm. This stage performs a global search of the search space to determine the optimal region. The population randomly selects an individual in the search space and launches an attack; the mathematical model is as follows:
[0071]
[0072] p i =X k k = 1, 2, ..., i-1, i+1, ..., n
[0073] Where j is the parameter type number, i are the population type numbers, and x ij F represents the position of the population at the current iteration. i Let it be the value of its objective function; The new position of the population after iterative update; p ij For a random individual that is different from the current individual; F pi Let r1 and r2 be the values of the objective function; r1 and r2 are random numbers between (0,1); I is a random integer of 1 or 2.
[0074] After the population position is updated, the decision on whether to adopt the updated population position is based on the objective function values before and after the update, as shown in the following expression:
[0075]
[0076] In the second phase, the northern goshawk relentlessly pursues and captures its prey at a relatively high speed. Assuming this hunting behavior approximates an attack pattern with a radius of R, the mathematical expression for this phase is as follows:
[0077]
[0078] Where, x ij This represents the position of the population at the current iteration. The new position of the population after iterative update; r4 is a random number between (0,1); t is the current iteration number; T is the maximum iteration number. Similar to the first phase, the expression for determining whether to update the population position is as follows:
[0079]
[0080] By repeatedly going through the first and second stages until the maximum number of iterations is reached, the optimal value of the parameter to be optimized is obtained.
[0081] S5: Model Evaluation: Using root mean square error (RMSE), mean relative percentage error (MAPE), and coefficient of determination R0 2 The multiple correlation coefficient R is used as an error evaluation index for the model; the smaller the root mean square error and the average relative percentage error, the larger the coefficient of determination and the multiple correlation coefficient R, indicating higher prediction accuracy.
[0082] Suppose the measured value at a certain measuring point is y = {y1, y2, ..., y...} n The predicted value is} The above indicators are then calculated using the following formula:
[0083]
[0084] S6: Substitute the explanatory variables from the prediction sample set into the optimized CNN-BiLSTM-Attention concrete dam deformation prediction model to achieve the prediction of concrete dam deformation.
[0085] The content of this invention is not limited to the embodiments listed. Any equivalent modifications made by those skilled in the art to the technical solutions of this invention by reading this specification are covered by the claims of this invention.
Claims
1. A method for monitoring the deformation of a gravity dam based on intelligent analysis, characterized in that: The steps include: S1: Select influencing factors and establish the model input factor set; S2: Data partitioning: Divide the original dataset into training and test sets according to a random proportion or chronological order; S3: Model Training: Train the CNN-BiLSTM-Attention model on the training set to learn the features of measured deformation data and their correlation with various factors. S4: Parameter Optimization: Optimize the hyperparameters of the CNN-BiLSTM-Attention model using the Northern Eagle optimization algorithm. Apply the optimal hyperparameter combination to the CNN-BiLSTM-Attention model to form a concrete dam deformation prediction model. S5: Model Evaluation: Use root mean square error, mean relative percentage error, and coefficient of determination R². 2 And the multiple correlation coefficient R is used as an error evaluation index for the model; S6: Substitute the explanatory variables from the prediction sample set into the optimized CNN-BiLSTM-Attention concrete dam deformation prediction model to achieve the prediction of concrete dam deformation.
2. The gravity dam deformation monitoring method based on intelligent analysis according to claim 1, characterized in that: In step S1, the model input factor set includes: water pressure factor Hup, Hup 2 Hup 3 Hup 10d Hup 15d Hup 30d Hup 60d Hup 90d Hup 120d ,Hd,Hd 2 ,Hd 3 ,Hd 10d ,Hd 15d ,Hd 30d ,Hd 60d ,Hd 90d ,Hd 120d This refers to the 1st, 2nd, and 3rd powers of the upstream and downstream water levels, as well as the average upstream water level over the preceding 2, 3, 5, 7, 10, 15, 30, 60, 90, and 120 days; temperature factor T, T 5d ,T 10d ,T 15d ,T 30d ,T 60d ,T 90d ,T 120d This refers to the daily air temperature and its average temperature over the preceding 5, 10, 15, 30, 60, 90, and 120 days; the time factor s, s 2 ,s 3 , where s is the number of monitoring days / 100.
3. The gravity dam deformation monitoring method based on intelligent analysis according to claim 2, characterized in that: In step S2, the data can be partitioned into training set and test set by randomly or according to time order, based on the input factor set and the corresponding deformation monitoring values.
4. The gravity dam deformation monitoring method based on intelligent analysis according to claim 3, characterized in that: In step S3, the model training process involves using the deformation explanatory variable factors in the training set as input and the measured deformation values of the concrete dam as output to train the CNN-BiLSTM-Attention model.
5. The gravity dam deformation monitoring method based on intelligent analysis according to claim 4, characterized in that: Step S4 describes parameter optimization using the Northern Eagle Optimization Algorithm (NGO) on the training data to optimize the hyperparameters of the CNN-BiLSTM-Attention model. The specific method is as follows: First, initialize the population size and iteration count of the NGO, and set the search range for the model parameters to be optimized. With the goal of minimizing the mean squared error, optimize the initial learning rate, the number of BiLSTM neurons, the number of key values in the self-attention mechanism, and the regularization parameter. Set the search ranges [LB, UB] for these four parameters to [0.001, 0.01], [10, 50], [2, 50], and [0.0001, 0.001], respectively. Then, use the NGO to set initial values for each parameter (four in this case): x ij =LB ij +r1×(UB ij -LB ij i = 1, 2, ..., n, j = 1, 2, 3, 4, where x ij It is the initial value of the i-th population in the j-th parameter, where n is the population size; UB i LB i These are the upper and lower boundaries of the search domain, respectively; r1 is a random number between (0,1); in the first stage of prey identification, a global search is performed on the search space to determine the optimal region; the population randomly selects an individual in the search space and launches an attack, the mathematical model of which is as follows: p i =X k k = 1, 2, ..., i-1, i+1, ..., n where j is the parameter type number, i is the population type number, and x ij F represents the position of the population at the current iteration. i Let it be the value of its objective function; The new position of the population after iterative update; p ij For a random individual that is different from the current individual; Let r1 and r2 be random numbers between (0,1); let I be a random integer of 1 or 2; after the population position is updated, determine whether to adopt the updated population position based on the objective function values before and after the update, as shown in the following expression: In the second phase, the northern goshawk relentlessly pursues and captures its prey at a relatively high speed. Assuming this hunting behavior approximates an attack pattern with a radius of R, the mathematical expression for this phase is as follows: Where, x ij This represents the position of the population at the current iteration. The new position of the population after iterative update; r4 is a random number between (0,1); t is the current iteration number; T is the maximum iteration number; similar to the first stage, the expression for determining whether to update the population position is as follows: By repeatedly going through the first and second stages until the maximum number of iterations is reached, the optimal value of the parameter to be optimized is obtained.
6. The gravity dam deformation monitoring method based on intelligent analysis according to claim 5, characterized in that: In step S5, the root mean square error, mean relative percentage error (MAPE), coefficient of determination R², and multiple correlation coefficient R are used as error evaluation indicators for the model.