Multi-source observation data dam automatic monitoring and early warning method fused with deep learning

Abstract

The invention discloses a multi-source observation data dam automatic monitoring and early warning method fused with deep learning, and relates to the technical field of hydraulic engineering, a grey wolf optimization algorithm is used to optimize variational mode decomposition parameters, and an envelope entropy is used as a fitness function to obtain an optimal parameter combination; introducing a multi-scale permutation entropy as a standard for screening signals, determining an effective modal component, and reconstructing the effective modal component to obtain an effective signal; and according to the optimal parameter combination and the effective signal, fusing a deep learning algorithm long-short term neural network to carry out multi-source observation data dam deformation feature learning, taking a mean absolute error and a root-mean-square error as evaluation indexes of GWO-VMD-LSTM prediction accuracy, introducing a decision coefficient to judge the performance of a prediction model, and carrying out multi-source observation data dam deformation feature prediction. And accurate prediction of dam deformation displacement is realized. The prediction result has higher accuracy and precision, and a reliable basis is provided for automatic monitoring and early warning of the dam.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy projects, and particularly discloses a multi-source observation data dam automatic monitoring and early warning method integrating deep learning. Background Art

[0002] High-precision prediction of the surface displacement and deformation of reservoir dams is of great significance for humans to prevent geological disasters. According to different theoretical methods, dam deformation prediction models can be divided into deterministic models, statistical regression models, artificial intelligence models and other methods. However, such methods all have limitations in the prediction process. For example, common single prediction models such as neural networks (BackPropagation, BP), grey models (Grey Models, GM), and support vector machines (Support Vector Machines, SVM) have problems such as complex structures, multicollinearity, being affected by parameters or noise, and poor prediction effects. Therefore, it is imperative to integrate new intelligent deep learning methods.

[0003] In recent years, intelligent deep learning methods have been applied to displacement and deformation predictions such as settlement, landslide, and dam deformation due to their powerful ability to process non-linear time series and features. Existing research has proposed a prediction model integrating variational mode decomposition (VMD) and long short-term memory neural network (LSTM) by converting the dam deformation time series decomposition process into a variational solution problem. After decomposing the complex dam deformation time series into multiple relatively simple subsequences with different frequency bands, predictions are made. However, the selection of the key parameter mode number K and the penalty factor in VMD is difficult, and in practical applications, it is often determined according to experience. Improper selection of K and may cause over-decomposition or under-decomposition of signals.

[0004] Therefore, how to effectively remove the noise in the original sequence, obtain effective signals, and better retain the characteristics of the original signals is an urgent problem to be solved for the multi-source observation data dam automatic monitoring and early warning integrating deep learning. Summary of the Invention

[0005] The present invention provides a multi-source observation data dam automatic monitoring and early warning method integrating deep learning, aiming to solve at least one defect existing in the above-mentioned prior art.

[0006] The present invention relates to a multi-source observation data dam automatic monitoring and early warning method integrating deep learning, including the following steps: Effective decomposition: Using the grey wolf optimization algorithm to optimize the variational mode decomposition parameters, with the envelope entropy as the fitness function, to obtain the optimal parameter combination; Signal recombination: Introduce multi-scale permutation entropy as the standard for screening signals, determine the effective modal components and reconstruct the effective modal components to obtain the effective signals; Feature learning: According to the optimal parameter combination and the effective signals, fuse the long short-term neural network of the deep learning algorithm to perform feature learning on the multi-source observation data of dam deformation. Use the mean absolute error and the root mean square error as the evaluation indexes for the prediction accuracy of GWO-VMD-LSTM, and introduce the coefficient of determination to judge the performance of the prediction model, so as to achieve accurate prediction of the dam deformation displacement.

[0007] Furthermore, the steps of effective decomposition include: Use the grey wolf optimization algorithm to optimize the variational mode decomposition parameters; Use the grey wolf optimization algorithm to optimize the variational mode decomposition parameters. The variational mode decomposition parameter is a swarm intelligence optimization algorithm based on the social hierarchy and hunting behavior of the grey wolf population. During the optimization process, the wolf pack surrounds the prey to search for the optimal hunting route; after surrounding the prey, alpha wolf, beta wolf and delta wolf guide the wolf pack to hunt. In each iteration, the positions of the alpha wolf, beta wolf and delta wolf are updated to gradually approach the prey, realizing the optimization of the prey position. Among them, the grey wolf population alpha wolf, beta wolf, delta wolf and omega wolf in the social hierarchy, alpha wolf is the optimal solution; beta wolf and delta wolf are the first two wolves with sub-optimal fitness; and the remaining wolves are omega wolf, which are candidate solutions; After obtaining the final position of the prey, the grey wolves attack the prey to complete the hunting process; By setting the modal parameter K and the penalty parameter in the variational mode decomposition parameters, the signal is decomposed into k modal functions u k with the central frequency of ω k ; Select the envelope entropy as the fitness function optimized by the grey wolf optimization algorithm, and use the grey wolf optimization algorithm to optimize the modal parameter and the penalty parameter in the variational mode decomposition parameters; Use the grey wolf optimization algorithm to obtain the optimal parameters , , and introduce multi-scale permutation entropy as the standard for judging noise and signals; Set the modal parameters according to the judgment results of multi-scale permutation entropy and the penalty parameter value range, and randomly generate the positions of gray wolves; According to the optimal parameters , , after variational mode decomposition, calculate the IMF fitness value through envelope entropy, and update wolf, wolf and wolf positions, and then complete the optimization of the variational mode decomposition parameters by the gray wolf optimization algorithm.

[0008] Furthermore, in the step of using the gray wolf optimization algorithm to optimize the variational mode decomposition parameters, the mathematical model in the optimization process is: Among them, is the current iteration number, is the distance between the gray wolf and the prey, is the prey position vector, is the gray wolf position vector, and are coefficient vectors. By adjusting and values, the positions of points around the optimal solution can be searched to ensure the local search ability of the algorithm; the convergence factor , is the maximum number of iterations, that is linearly decreases from 2 to 0 during iteration; and are both random vectors in the range of [0, 1].

[0009] Furthermore, in the step of using the gray wolf optimization algorithm to optimize the variational mode decomposition parameters, the updated mathematical expression after surrounding the prey is: Among them, , , are respectively wolf, wolf and the position vectors between the wolf and the prey, , , are respectively wolf, wolf and the coefficient vectors for updating the wolf position, , , are respectively wolf, wolf and the updated position vectors of the wolf; The position vector updated for the grey wolf.

[0010] Furthermore, by setting the mode parameter K and the penalty parameter in the variational mode decomposition parameters , the signal is decomposed into k mode functions with the central frequencies of In the step of , the algorithm of the variational mode decomposition parameters constrains the variational problem, which is expressed as:

[0011] where is time, is the original signal, is the impulse function, is the mode function, is the actual central frequency of each mode, is the estimated central frequency of each analytic signal, is to find the L2 norm, represents the constraint condition, is to sum over all the mode numbers.

[0012] Furthermore, by setting the mode parameter K and the penalty parameter in the variational mode decomposition parameters , the signal is decomposed into k mode functions with the central frequencies of In the step of , by introducing the second-order penalty factor and the Lagrange multiplier , it is transformed into an unconstrained variational problem, and the obtained extended Lagrangian expression is as follows:

[0013] where is the second-order penalty factor; is the Lagrange multiplier, is the Hilbert transform signal parameter.

[0014] Furthermore, in the step of selecting the envelope entropy as the fitness function optimized by the grey wolf optimization algorithm and using the grey wolf optimization algorithm to optimize the mode parameter and the penalty parameter in the variational mode decomposition parameters, the envelope entropy is selected as the fitness function optimized by GWO, and the principle of the envelope entropy is shown in the following formula:

[0015] where is the number of sampling points of the signal, is in the normalized form, is the signal The envelope signal obtained after Hilbert demodulation, is the envelope entropy.

[0016] Further, in the signal reconstruction step, the effective IMF components are determined according to the threshold of MPE and reconstructed into a signal, and the remaining components are reconstructed into noise. Let be the threshold of IMF, and the judgment instruction is set as:

[0017] where represents the threshold of IMF i , represents the threshold of IMF i+1 . Further, in the feature learning step, the LSTM network structure is called a cell, including an input layer, a hidden layer, and an output layer. Each hidden layer controls the storage and access of data through an input gate, a forget gate, and an output gate. The multi-source data sequence of the dam is decomposed by VMD to obtain k subsequences, and the original sequence is defined as :

[0018] where is the original sequence, is the kth subsequence of the nth term; Let the first m groups of sequences be the training set and the validation set, denoted as , and its mathematical expression is:

[0019] where is the training set and the validation set of the first m groups of sequences, and the remaining n - m groups of sequences are used as the test set, denoted as , and its mathematical expression is:

[0020] where represents the test set of n - m groups of sequences, represents the number of the mth term of the first k subsequences, represents the number of the nth term sequence of the k subsequences.

[0021] Let have a length of L, the input of , the output is , , and the expression is:

[0022] where represents Input item;

[0023] Among them, represents Output item; Let The input is , The output is , and The expression is:

[0024] Among them, represents Input item;

[0025] Among them, represents Output item.

[0026] Furthermore, in the steps of feature learning, after feature learning through LSTM, the prediction results are evaluated for accuracy through , and :

[0027] Among them, represents the root mean square error, represents the mean absolute error, represents the coefficient of determination, is the original time series, is the prediction result of each model, is the average value of the original time series data, and n is the number of original time series data.

[0028] The beneficial effects achieved by the present invention are: The present invention provides a method for automatic monitoring and early warning of dam deformation using multi-source observation data integrated with deep learning. The grey wolf optimization algorithm is used to optimize the parameters of variational mode decomposition, and the envelope entropy is used as the fitness function to obtain the optimal parameter combination. The multi-scale permutation entropy is introduced as the standard for screening signals, and the effective modal components are determined and reconstructed to obtain the effective signals. According to the optimal parameter combination and the effective signals, the long short-term neural network of the deep learning algorithm is used to learn the deformation characteristics of the dam from multi-source observation data. The mean absolute error and the root mean square error are used as the evaluation indexes for the prediction accuracy of GWO-VMD-LSTM, and the coefficient of determination is introduced to judge the performance of the prediction model, so as to achieve accurate prediction of the dam deformation displacement. The present invention provides a method for automatic monitoring and early warning of dam deformation using multi-source observation data integrated with deep learning. On the one hand, the parameters of VMD are optimized by GWO. After obtaining the optimal parameters, MPE is introduced as the standard for screening signals, and the envelope entropy is used as the fitness function for GWO optimization to determine the effective modal components and reconstruct them to obtain the effective signals. On the other hand, the effective signals are used as features, and the deep learning algorithm LSTM is used to learn the deformation characteristics of the dam from multi-source observation data. Finally, with MAE, RMSE and R 2 to judge the performance of the prediction model, so as to achieve accurate prediction of the dam deformation displacement and improve the prediction accuracy. The present invention takes into account the characteristics of the original sequence, and the constructed GWO-VMD-LSTM prediction method has higher accuracy and precision in the prediction results after the original sequence is decomposed and reconstructed, and has obvious advantages in the prediction of dam deformation time series, providing a reliable basis for the automatic monitoring and early warning of dams. Description of the Drawings

[0029] Figure 1 is a schematic flow chart of an embodiment of a method for automatic monitoring and early warning of dam deformation using multi-source observation data integrated with deep learning according to the present invention; Figure 2(a) is a schematic diagram of the fitness value in the N direction of the present invention; Figure 2(b) is a schematic diagram of the fitness value in the E direction of the present invention; Figure 2(c) is a schematic diagram of the fitness value in the U direction of the present invention; Figure 3(a) is a schematic diagram of the distribution of the optimal parameter K value at each site of the present invention; Figure 3(b) is a schematic diagram of the distribution of the optimal parameter α value at each site of the present invention; Figure 4 is a schematic diagram of the signal after GWO-VMD decomposition in an embodiment of a method for automatic monitoring and early warning of dam deformation using multi-source observation data integrated with deep learning according to the present invention; Figure 5 is a schematic diagram of the distribution of the MPE values of the IMF components at each site in an embodiment of a method for automatic monitoring and early warning of dam deformation using multi-source observation data integrated with deep learning according to the present invention; Figure 6(a) is a comparison diagram of the embodiment of the method of the present invention and the prediction result of VMD-LSTM in the N direction; Figure 6(b) is a comparison diagram of the embodiment of the method of the present invention and the prediction result of VMD-LSTM in the E direction; Figure 6(c) is a comparison diagram of the embodiment of the method of the present invention and the prediction result of VMD-LSTM in the U direction; Figure 7(a) is a schematic diagram of the prediction evaluation results of GWO-VMD-LSTM and VMD-LSTM in the N direction of the present invention; Figure 7(b) is a schematic diagram of the prediction evaluation results of GWO-VMD-LSTM and VMD-LSTM in the E direction of the present invention; Figure 7(c) is a schematic diagram of the prediction evaluation results of GWO-VMD-LSTM and VMD-LSTM in the U direction of the present invention; Figure 8(a) is the result schematic diagram of GWO-VMD-LSTM and other combined models in the N direction ; Figure 8(b) is the result schematic diagram of GWO-VMD-LSTM and other combined models in the E direction ; Figure 8(c) is the result schematic diagram of GWO-VMD-LSTM and other combined models in the U direction ; Detailed implementation manners

[0030] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the accompanying drawings of the specification and specific implementation manners.

[0031] As Figure 1 shown in FIGS. 3 to 8, a multi-source observation data dam automatic monitoring and early warning method integrating deep learning is proposed in the first embodiment of the present invention. Aiming at the problems in the nonlinear and non-stationary dam displacement time series prediction method, such as difficult acquisition of model parameters, difficult extraction of useful signals, and inappropriate signal decomposition, a new prediction model of parameter-optimized variational mode decomposition long short-term memory neural network (grey wolf optimization and variational mode decomposition and long short-term memory, GWO-VMD-LSTM) is constructed, including the following steps: Step S100, effective decomposition: Use the grey wolf optimization algorithm to optimize the parameters of variational mode decomposition, and use the envelope entropy as the fitness function to obtain the optimal parameter combination.

[0032] Optimize the parameters of Variational Mode Decomposition (VMD) using the Grey Wolf Optimization (GWO) algorithm , , and use the envelope entropy as the fitness function to obtain the optimal parameter combination.

[0033] Step S200, Signal recombination: Introduce multiscale permutation entropy as the criterion for screening signals, determine the effective modal components, and reconstruct the effective modal components to obtain the effective signal.

[0034] Introduce multiscale permutation entropy (MPE) as the criterion for screening signals, determine the effective modal components and reconstruct them to obtain the effective signal.

[0035] Step S300, Feature learning: According to the optimal parameter combination and the effective signal, fuse the deep learning algorithm of long short-term neural network for learning the deformation characteristics of the dam from multi-source observation data. Use the mean absolute error and root mean square error as the evaluation indicators for the prediction accuracy of GWO-VMD-LSTM. Introduce the coefficient of determination to judge the performance of the prediction model, and realize the accurate prediction of the dam deformation displacement.

[0036] Fuse the long short-term neural network (LSTM) of the deep learning algorithm for learning the deformation characteristics of the dam from multi-source observation data. Use the mean absolute error (MAE) and root mean square error (RMSE) as the evaluation indicators for the prediction accuracy of GWO-VMD-LSTM. In order to better reflect the quality of the prediction model, the coefficient of determination (Coefficient of determination, R 2 ) is introduced to judge the performance of the prediction model, so as to realize the accurate prediction of the dam deformation displacement and improve the prediction accuracy.

[0037] Furthermore, see Figure 1 to Figure 8. The method for automatic monitoring and early warning of dam with multi-source observation data integrating deep learning provided in this embodiment, step S100 includes: Step S110, Optimize the parameters of variational mode decomposition using the grey wolf optimization algorithm.

[0038] The grey wolf optimization algorithm GWO optimizes the parameters of variational mode decomposition VMD , . Use the multi-source observation data of the dam for VMD decomposition.

[0039] GWO-optimized variational mode decomposition (VMD). The VMD is performed using multi-source observation data of the dam. The number of modes and the penalty factor settings have a significant impact on the VMD decomposition results. The remaining parameters are set to default values. When is too small, the decomposed signal may be under-decomposed completely. If the value of is set too large, it may cause over-decomposition of the signal and result in mode mixing. To solve this problem, in this embodiment, GWO is used to optimize the VMD parameters and . When using the GWO algorithm to optimize VMD, it is very important to select a suitable fitness function as the optimization judgment criterion. The envelope entropy is selected as the fitness function for GWO optimization. The envelope entropy can better reflect the sparse characteristics and uncertainty of the original signal. FIG. 2 is the fitness convergence graph of the embodiment of the method of the present invention, and FIG. 3 is the distribution of the optimal parameters of the stations in the embodiment of the method of the present invention, Figure 4 and this is the signal after GWO-VMD decomposition in the embodiment of the method of the present invention.

[0040] Step S120: Use the grey wolf optimization algorithm to optimize the variational mode decomposition parameters. The variational mode decomposition parameters are a swarm intelligence optimization algorithm based on the social hierarchy and hunting behavior of the grey wolf population. During the optimization process, the wolf pack surrounds the prey to search for the optimal hunting route; after surrounding the prey, the alpha wolf, the beta wolf and the delta wolf guide the wolf pack to hunt. In each iteration, the positions of the alpha wolf, the beta wolf, the delta wolf and the omega wolf with the optimal fitness are used to update the position of the omega wolf, gradually approaching the prey, and achieving the optimization of the prey position. Among them, in the social hierarchy of the grey wolf population the alpha wolf, the beta wolf, the delta wolf and the omega wolf, the alpha wolf is the optimal solution; the beta wolf and the delta wolf are the first two wolves with the sub-optimal fitness; and the remaining wolves are the omega wolf, which are candidate solutions.

[0041] GWO obtains the optimal parameters , . The intrinsic mode components obtained by VMD decomposition contain more noise and the signal is more complex. GWO is used to optimize the VMD parameters and After optimization, the multi-scale permutation entropy (MPE) is introduced as a criterion for judging noise and signals. After calculating the MPE of each IMF component by VMD decomposition, the low-frequency signals and high-frequency noises are judged through the set MPE threshold. When there is less noise in the IMF component, the signal is more regular, and the MPE value is smaller; conversely, the MPE value is larger. After multiple experiments, in this embodiment, through multiple tests on multi-source data, the MPE value is set to 0.6, and the low-frequency IMF components smaller than the MPE threshold are reconstructed into a new signal. Figure 5 The MPE value distribution of each site's IMF component in the method embodiment of the present invention.

[0042] The mathematical model in the optimization process is:

[0043] In formulas (1) to (4), is the current iteration number, is the distance between the grey wolf and the prey, is the prey position vector, is the grey wolf position vector, and are coefficient vectors. By adjusting the values of and , the positions of points around the optimal solution can be searched to ensure the local search ability of the algorithm; the convergence factor , is the maximum iteration number, that is, linearly decreases from 2 to 0 during iteration; and are both random vectors within the range of [0, 1].

[0044] After surrounding the prey, , and wolves guide the wolf pack to hunt. In each iteration, the positions of , and wolves with the best fitness are updated to approach the prey step by step, so as to optimize the position of the prey. The updated mathematical expression is:

[0045] In formulas (5) to (7), , , are respectively wolf, wolf and the position vectors between the , , wolves and the prey, Wolf, wolves and the coefficient vector for updating the position of wolves, , , are respectively wolf, wolves and the position vectors updated by wolves; is the position vector updated by the gray wolf. Equations (5) and (6) define the direction of the gray wolf towards , and the step length and direction of the wolf's forward movement. The location of the prey can be determined by Equation (7).

[0046] Step S130, after obtaining the final position of the prey, the gray wolf attacks the prey to complete the hunting process.

[0047] After obtaining the final position of the prey, the gray wolf attacks the prey to complete the hunting process. To simulate the gray wolf approaching the prey, the value of decreases linearly, and the fluctuation range of also decreases accordingly. That is, when decreases from 2 to 0, the corresponding also varies within the range of . The wolf pack takes actions according to the value of . When

[0048] Step S140, by setting the mode parameter K and the penalty parameter in the variational mode decomposition parameters, the signal is decomposed into k mode functions u k with central frequencies of ω k .

[0049] The VMD algorithm decomposes the signal into k mode functions u with central frequencies of ω k by setting the mode parameter k, the penalty parameter k and the step size τ, etc. VMD has strong robustness to noise and sampling errors. The algorithmic constrained variational problem can be expressed as:

[0050] In Equation (8), is time, is the original signal, is the impulse function, is the mode function, is the actual central frequency of each mode, is the estimated center frequency of each parsed signal, is to calculate the L2 norm, represents the constraint condition, is to sum over all modal numbers.

[0051] By introducing the second-order penalty factor and the Lagrange multiplier it is transformed into an unconstrained variational problem, and the resulting extended Lagrangian expression is as follows:

[0052] In formula (9), is the second-order penalty factor; is the Lagrange multiplier, is the Hilbert transform signal parameter, is time. It is iteratively updated through the multiplier direction alternating algorithm , , to find the saddle point of formula (9), which is the optimal solution of formula (8).

[0053] Step S150: Select the envelope entropy as the fitness function optimized by the grey wolf optimization algorithm, and use the grey wolf optimization algorithm to optimize the modal parameter and the penalty parameter in the variational mode decomposition parameters.

[0054] Select the envelope entropy as the fitness function optimized by GWO. The number of modes and the penalty factor settings have a significant impact on the VMD decomposition result. The remaining parameters are set to default values. When is too small, the decomposed signal may be under-decomposed completely. An overly large value setting may cause over-decomposition of the signal and result in modal aliasing. To solve this problem, this software uses GWO to optimize the VMD parameters and . When optimizing VMD using the GWO algorithm, it is very important to select a suitable fitness function as the optimization judgment criterion. Select the envelope entropy as the fitness function optimized by GWO. The envelope entropy can better reflect the sparse characteristics and uncertainty of the original signal. The principle of the envelope entropy is shown in formula (10):

[0055] In formula (10), is the number of sampling points of the signal, is the normalized form of is the signal The envelope signal obtained after Hilbert demodulation, is the envelope entropy.

[0056] Step S160: Use the Grey Wolf Optimization (GWO) algorithm to obtain the optimal parameters , , and introduce the multi-scale permutation entropy as the criterion for judging noise and signals.

[0057] GWO obtains the optimal parameters , . The intrinsic mode components obtained by variational mode decomposition (VMD) contain more noise and the signals are more complex. After GWO optimizes the VMD parameters and , the multi-scale permutation entropy (MPE) is introduced as the criterion for judging noise and signals. After calculating the multi-scale permutation entropy of each intrinsic mode function (IMF) component by VMD decomposition, the low-frequency signals and high-frequency noise are judged through the set MPE threshold. When there is less noise in the IMF component, the signal is more regular, and the MPE value is smaller; conversely, the MPE value is larger. After multiple experiments, this software has passed multiple tests on multi-source data and sets the MPE value to 0.6, and reconstructs the low-frequency IMF components smaller than the MPE threshold into new signals.

[0058] Step S170: According to the judgment result of the multi-scale permutation entropy, set the value ranges of the modal parameter and the penalty parameter , and randomly generate the positions of grey wolves.

[0059] According to the judgment of MPE in step S160, initialize the parameters of the GWO algorithm, set the number of wolf packs to 30, and the maximum number of iterations to 10. Considering the computational efficiency and algorithm accuracy, this software sets the value range of to [3, 12], and sets the value range of

[0060] to [100, 4000], and randomly generates the positions of grey wolves. , , calculate the fitness value of the IMF through the envelope entropy after variational mode decomposition, and update wolf, wolf and wolf positions, thereby completing the optimization of the variational mode decomposition parameters by the Grey Wolf Optimization algorithm.

[0061] According to the optimal parameter combination , obtained in step S170, after VMD decomposition, calculate the fitness value of the IMF through the envelope entropy in step S150, and update , and the position of the wolf, thereby completing the optimization of VMD by GWO.

[0062] Furthermore, step S200 includes: Step S210, the signal recombination process, through the threshold judgment of multi-scale permutation entropy, determines the effective IMF components according to the threshold size and reconstructs them into a signal, and the remaining components are reconstructed into noise.

[0063] Determine the effective IMF components according to the threshold size of MPE and reconstruct them into a signal, and the remaining components are reconstructed into noise. Let be the threshold of IMF

[0064] In formula (11), represents the threshold of IMF i , represents the threshold of IMF i+1 .

[0065] Step S220, update the position of the grey wolf, iterate until the variational mode decomposition parameters of the optimal solution are obtained.

[0066] Update the position of the grey wolf, iterate, and return to step S210 until the optimal solution of , is obtained.

[0067] Step S230, calculate the MPE value of the IMF components, reconstruct the sequence into a denoised signal, and end the optimization of VMD by GWO.

[0068] According to the optimal parameter combination obtained in step S220 , , calculate the IMF fitness value through the envelope entropy after VMD decomposition, and update , and the position of the wolf, thereby completing the optimization of VMD by GWO.

[0069] Furthermore, in step S300, the LSTM network structure is called a cell, including an input layer, a hidden layer, and an output layer. Each hidden layer controls the storage and access of data through an input gate, a forget gate, and an output gate. Decompose the multi-source data sequence of the dam by VMD to obtain k subsequences, and the original sequence is defined as :

[0070] In formula (12), is the original sequence, The k subsequence of the nth term.

[0071] Let the first m groups of sequences be the training set and the validation set, denoted as , and its mathematical expression is:

[0072] In formula (13), The first m groups of sequences are the training set and the validation set.

[0073] The remaining n - m groups of sequences are used as the test set, denoted as , and its mathematical expression is:

[0074] Formula (14) represents the test set of n - m groups of sequences, represents the number of the mth term of the first k subsequence, represents the number of the nth term sequence of the k subsequence; Let have a length of L, The input of , The output is , , and the expression is:

[0075] In formula (15), represents the input term of

[0076] In formula (16), represents the output term of Let The input be , The output is , and The expression is:

[0077] In formula (17), represents the input term of

[0078] In formula (18), represents the output term of

[0079] After training is completed, this embodiment is compared with the existing method VMD-LSTM. Figure 6 is a comparison chart of the prediction results of the method embodiment of the present invention and VMD-LSTM.

[0080] Preferably, in step S300, after feature learning by LSTM, the prediction result is , and are used for accuracy evaluation:

[0081] In formulas (19) to (21), represents the root mean square error, represents the mean absolute error, represents the coefficient of determination, is the original time series, is the prediction result of each model, is the average value of the original time series data, and n is the number of original time series data. , and The values of

[0082] and can better reflect the accuracy of the prediction result, while can better reflect the quality of the prediction model. The value range of is [0,1]. The closer the value of

[0083] is to 1, the better the prediction model. Figure 8 is the evaluation of the prediction result of GWO-VMD-LSTM relative to the combined model. Table 1 shows the improvement in accuracy index evaluation of GWO-VMD-LSTM in different directions compared to VMD-LSTM. The RMSE and MAE accuracies of the GWO-VMD-LSTM model for predicting each site are significantly improved compared to the VMD-LSTM model. In the N direction, the maximum improvement in RMSE accuracy is 52.17%, and the minimum improvement is 7.41%. The maximum improvement in MAE accuracy is 36.36%, and the minimum improvement is 10.00%. In the E direction, the maximum improvement in RMSE accuracy is 40.00%, and the minimum improvement is 0.56%. The maximum improvement in MAE accuracy is 42.11%, and the minimum improvement is 10.05%. In the U direction, the maximum improvement in RMSE accuracy is 52.78%, and the minimum improvement is 10.53%. The maximum improvement in MAE accuracy is 48.00%, and the minimum improvement is 18.00%. In summary, it is proved that the RMSE and MAE accuracies of the GWO-VMD-LSTM model in different directions are improved by 0.56% - 52.78% and 10.00% - 48.00% compared to the VMD-LSTM model. The modeling and prediction results of the GWO-VMD-LSTM prediction model are closer to the accurate prediction of the measured dam deformation displacement, improving the prediction accuracy.

[0084] Table 1 Improvement in accuracy index evaluation

[0085] The dam automation monitoring and early warning method for multi-source observation data integrating deep learning provided in this embodiment, compared with the prior art, uses the grey wolf optimization algorithm to optimize the variational mode decomposition parameters, takes the envelope entropy as the fitness function to obtain the optimal parameter combination; introduces multi-scale permutation entropy as the standard for screening signals, determines the effective modal components and reconstructs the effective modal components to obtain the effective signal; according to the optimal parameter combination and the effective signal, integrates the deep learning algorithm long short-term neural network to learn the dam deformation characteristics of multi-source observation data, uses the mean absolute error and root mean square error as the evaluation indexes for the prediction accuracy of GWO-VMD-LSTM, and introduces the coefficient of determination to judge the performance of the prediction model, so as to achieve the accurate prediction of the dam deformation displacement. The dam automation monitoring and early warning method for multi-source observation data integrating deep learning provided in this embodiment, on the one hand, optimizes the VMD parameters by GWO, after obtaining the optimal parameters, introduces MPE as the standard for screening signals, takes the envelope entropy as the fitness function for GWO optimization to determine the effective modal components and reconstructs them to obtain the effective signal. On the other hand, the effective signal is used as a feature, and the deep learning algorithm LSTM is used to learn the dam deformation characteristics of multi-source observation data. Finally, with MAE, RMSE and R 2Judge the performance of the prediction model, so as to achieve accurate prediction of the deformation displacement of the dam and improve the prediction accuracy. This embodiment takes into account the characteristics of the original sequence, and the constructed GWO-VMD-LSTM prediction method has higher accuracy and precision in the prediction results after the original sequence decomposition and reconstruction, and has obvious advantages in the prediction of the dam deformation time series, providing a reliable basis for the dam automatic monitoring and early warning.

[0086] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-source observation data dam automatic monitoring and early warning method integrating deep learning, characterized in that: The following steps are involved: Effective decomposition: Use the Grey Wolf Optimization Algorithm to optimize the variational mode decomposition parameters, and use envelope entropy as the fitness function to obtain the optimal parameter combination; Signal reconstruction: Introduce multi-scale permutation entropy as a criterion for screening signals, determine effective modal components and reconstruct the effective modal components to obtain effective signals; Feature learning: Based on the optimal parameter combination and the effective signal, the deep learning algorithm and long-term short-term neural network are integrated to perform dam deformation feature learning of multi-source observation data. The mean absolute error and root mean square error are used as evaluation indicators of the GWO-VMD-LSTM prediction accuracy, and the determination coefficient is introduced to judge the performance of the prediction model to achieve accurate prediction of dam deformation displacement.

2. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 1 is characterized in that: The steps of effectively decomposing include: The variational mode decomposition parameters are optimized using the Grey Wolf Optimization Algorithm; The gray wolf optimization algorithm is used to optimize the variational mode decomposition parameters. The variational mode decomposition parameters are a group intelligence optimization algorithm based on the social hierarchy and hunting behavior of the gray wolf population. During the optimization process, the wolf pack surrounds the prey to search for the optimal hunting route; after surrounding the prey, Wolf, Wolf and The wolf guides the pack to hunt, and in each iteration, the wolf with the best fitness Wolf, Wolf and Wolf location updates The wolf's position gradually approaches the prey to optimize the prey's position. Among them, the gray wolf population Wolf, Wolf, Wolf and In the social hierarchy of wolves, The wolf is the best solution; Wolf and The wolves are the first two wolves with the second best fitness; the rest are The wolf is a candidate solution; After obtaining the final location of the prey, the gray wolf attacks the prey to complete the hunting process; By setting the modal parameter K and penalty parameter in the variational mode decomposition parameters , decompose the signal into k components with center frequency ω k The modal function u k ; The envelope entropy is selected as the fitness function of the gray wolf optimization algorithm, and the modal parameters in the variational mode decomposition parameters are optimized by the gray wolf optimization algorithm. and penalty parameters Optimize Use the Grey Wolf Optimization Algorithm to obtain the optimal parameters[ , ], introduced multi-scale permutation entropy as a criterion for judging noise and signal; According to the judgment results of multi-scale permutation entropy, set the modal parameters and penalty parameters The value range of , randomly generates the position of the gray wolf; According to the optimal parameters [ , ], after variational mode decomposition, the IMF fitness value is calculated by the envelope entropy, and updated Wolf, Wolf and The position of the wolf is determined, and then the gray wolf optimization algorithm is used to optimize the variational mode decomposition parameters.

3. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 2 is characterized in that: In the step of optimizing the variational modal decomposition parameters using the Grey Wolf Optimization Algorithm, the mathematical model in the optimization process is: ； in, is the current iteration number, is the distance between the gray wolf and its prey, is the prey position vector, is the gray wolf position vector, and is the coefficient vector, by adjusting and The value of can search for the location of points around the optimal solution to ensure the local search ability of the algorithm; the convergence factor , is the maximum number of iterations, that is Decreases linearly from 2 to 0 in iterations; and are all random vectors in the range [0, 1].

4. The multi-source observation data dam automatic monitoring and early warning method integrating deep learning as claimed in claim 3 is characterized in that: In the step of using the gray wolf optimization algorithm to optimize the variational mode decomposition parameters, the updated mathematical expression after surrounding the prey is: ； in, , , They are Wolf, Wolf and The position vector between the wolf and the prey, , , They are Wolf, Wolf and The coefficient vector for the wolf's position update, , , They are Wolf, Wolf and The wolf's updated position vector; Updated position vector for the gray wolf.

5. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 4 is characterized in that: By setting the modal parameter K and penalty parameter in the variational mode decomposition parameters , decompose the signal into k center frequencies The modal function In the steps of , the algorithmic constrained variational problem of variational mode decomposition parameters is expressed as: ； in, For time, is the original signal, is the pulse function, is the modal function, is the actual center frequency of each mode, is the estimated center frequency of each analytical signal, To find the L2 norm, represents the constraints, is the sum of all the mode numbers.

6. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 5 is characterized in that: By setting the modal parameter K and penalty parameter in the variational mode decomposition parameters , decompose the signal into k center frequencies The modal function In the step of With Lagrange multipliers Converted to an unconstrained variational problem, the obtained extended Lagrangian expression is as follows: ； in, is the second-order penalty factor; is the Lagrange multiplier, is the Hilbert transform signal parameter, For time.

7. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 4 is characterized in that: The envelope entropy is selected as the fitness function optimized by the gray wolf optimization algorithm, and the modal parameters in the variational modal decomposition parameters are optimized by the gray wolf optimization algorithm. and penalty parameters In the optimization step, envelope entropy is selected as the fitness function of GWO optimization. The principle of the envelope entropy is shown in the following formula: ； in, is the number of sampling points of the signal, yes The normalized form of For signal The envelope signal obtained after Hilbert demodulation, is the envelope entropy.

8. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 4 is characterized in that: In the signal reconstructing step, the effective IMF component is determined according to the threshold value of MPE and reconstructed into the signal, and the remaining component is reconstructed into noise. is the threshold of IMF, and the judgment instruction is: in, IMF i The threshold value, express The threshold value.

9. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 1 is characterized in that: In the feature learning step, the LSTM network structure is called a cell, which includes an input layer, a hidden layer, and an output layer. Each hidden layer controls the storage and access of data through an input gate, a forget gate, and an output gate. The multi-source data sequence of the dam is decomposed by VMD to obtain k subsequences. The original sequence is defined as : in, is the original sequence, is the k-th subsequence of the n-th item; Assume the first m groups of sequences as training set and validation set, expressed as , its mathematical expression is: in, The first m groups of sequences are training sets and validation sets; The remaining nm group sequences are used as the test set, expressed as , its mathematical expression is: in, represents the test set of nm group sequences, represents the number of the mth item in the first k subsequence, The number representing the nth item of the k-subsequence; set up The length is L, The input is , The output is , , the expression is: ； in, express Input items; ； in, express Output items; set up Input is , The output is , and The expression is: ； in, express Input items; ； in, express The output item of .

10. The multi-source observation data automated dam monitoring and early warning method integrating deep learning as claimed in claim 1, characterized in that: In the feature learning step, after feature learning through LSTM, the prediction result is , and To evaluate the accuracy: ； in, represents the root mean square error, represents the mean absolute error, represents the coefficient of determination, is the original time series, is the prediction result of each model, is the average value of the original time series data, and n is the number of original time series data.

Images