Dam seepage pressure space-time early warning method and device

Abstract

The application provides a dam seepage pressure space-time early warning method and device, the method comprises the following steps: collecting original seepage pressure time series data of multiple monitoring points in a dam monitoring process; preprocessing the time series data; using the processed time series data of the multiple monitoring points to perform seepage pressure time series characterization learning, and establishing a multi-task prediction model; and establishing an early warning model based on the multi-task prediction model. By using the application scheme, the relationship between different seepage pressure monitoring data can be effectively processed, and dam seepage pressure space-time early warning can be realized.

Description

Technical Field

[0001] This invention relates to the field of dam seepage safety monitoring in water conservancy and hydropower projects, specifically to a spatiotemporal early warning method and device for dam seepage pressure. Background Technology

[0002] The seepage safety of a dam is directly related to its structural stability. To ensure dam safety, sensor equipment is typically installed inside to monitor seepage in real time. With the increase in monitoring scope and the number of devices, efficient and accurate seepage pressure prediction becomes particularly important. However, facing a large amount of monitoring data, effectively processing the relationships between different seepage pressure monitoring data and establishing an efficient and accurate multi-monitoring point, multi-task seepage pressure prediction model has become an urgent problem to be solved. Furthermore, to enable on-site maintenance personnel to clearly understand the current status of each point, it is necessary to assess whether the prediction results exceed expectations. Once the predicted value exceeds the expected range, a timely warning must be issued to improve the accuracy and real-time nature of seepage warnings for earth-rock dams, further ensuring the safe and stable operation of earth-rock dams.

[0003] Currently, research on seepage pressure time series prediction mainly focuses on prediction at single monitoring points. However, dam seepage pressure monitoring involves multiple monitoring points, and the number of sequences requiring prediction analysis is also large. Constructing a separate prediction model for each monitoring point would lead to a significant increase in computational, storage, and time costs, making it difficult to meet the needs of practical engineering projects. Summary of the Invention

[0004] This invention provides a spatiotemporal early warning method and device for dam seepage pressure, which effectively handles the relationship between different seepage pressure monitoring data and establishes an efficient and accurate multi-monitoring point multi-task seepage pressure prediction model to realize spatiotemporal early warning of dam seepage pressure.

[0005] Therefore, the present invention provides the following technical solution:

[0006] On the one hand, the present invention provides a spatiotemporal early warning method for dam seepage pressure, the method comprising:

[0007] Collect raw seepage pressure time series data from multiple monitoring points during dam monitoring;

[0008] The time series data is preprocessed;

[0009] A multi-task prediction model was established by using time series data from multiple monitoring points after processing to characterize the seepage pressure time series.

[0010] An early warning model is established based on the aforementioned multi-task prediction model.

[0011] Optionally, the preprocessing of the time series data includes:

[0012] The time series data is preprocessed and cleaned using linear interpolation;

[0013] Z-score normalization was used to standardize the cleaned data.

[0014] Optionally, the step of using the processed time-series data from multiple monitoring points to perform characterization learning of the seepage pressure time series and establish a multi-task prediction model includes:

[0015] The time series data of the multiple monitoring points are segmented using a sliding window to obtain subsequences;

[0016] Extract the features of the subsequences and establish labels for each subsequence to obtain a training sample set;

[0017] A diffusion convolutional neural network is constructed using the training sample set to obtain a multi-task prediction model.

[0018] Optionally, different input features in the multi-task prediction model have different weights.

[0019] Optionally, the method further includes: determining the weight of each input feature in the multi-task prediction model based on an attention mechanism.

[0020] Optionally, the diffuse convolutional neural network includes graph convolution and recurrent neural network; the graph convolution is used to simulate the spatial relationship of monitoring points; the recurrent neural network simulates the time dependence.

[0021] Optionally, the method further includes: constructing an adjacency matrix through the Pearson correlation between multiple monitoring sequences to achieve graph convolution.

[0022] Optionally, establishing an early warning model based on the multi-task prediction model includes:

[0023] Determine the loss function, which includes an uncertainty trade-off loss function and / or a quantile regression loss function;

[0024] An early warning model is established based on the aforementioned loss function.

[0025] Optionally, the method further includes:

[0026] Acquire current permeability pressure time series data from multiple monitoring points;

[0027] Based on the current seepage pressure time series data and the early warning model, the dam is monitored and warned in real time.

[0028] On the other hand, the present invention also provides a dam seepage pressure spatiotemporal early warning device, the device comprising:

[0029] The data collection module is used to collect raw seepage pressure time series data from multiple monitoring points during dam monitoring.

[0030] The preprocessing module is used to preprocess the time series data;

[0031] The multi-task prediction model building module is used to perform characterization learning of seepage pressure time series using processed time series data from multiple monitoring points, and to establish a multi-task prediction model.

[0032] The early warning model construction module is used to build an early warning model based on the multi-task prediction model.

[0033] Optionally, the device further includes:

[0034] The early warning module is used to acquire the current seepage pressure time series data of multiple monitoring points, and to conduct real-time monitoring and early warning of the dam based on the current seepage pressure time series data and the early warning model.

[0035] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when run by a processor, executes the steps of the dam seepage pressure spatiotemporal early warning method.

[0036] The present invention provides a spatiotemporal early warning method and device for dam seepage pressure, which collects raw seepage pressure time series data from multiple monitoring points during dam monitoring, preprocesses the time series data, uses the processed time series data from multiple monitoring points to perform characterization learning of seepage pressure time series, establishes a multi-task prediction model, and establishes an early warning model based on the multi-task prediction model.

[0037] Compared to existing technologies, this invention, based on a multi-task learning framework, performs spatiotemporal prediction of earth-rock dams, exhibiting higher accuracy and reliability. By comprehensively considering the correlation between different tasks, the model can more fully capture the changing trends of seepage in earth-rock dams, thereby providing more accurate predictions of seepage status.

[0038] Furthermore, the present invention employs an interval prediction method, which not only provides a single predicted value but also gives a prediction interval at a certain confidence level. This approach can better assess the potential range of changes in seepage conditions, helping on-site maintenance personnel make reasonable judgments under different risk scenarios.

[0039] The prediction results not only meet the requirement of reasonable monitoring values ​​but also provide timely and accurate early warning information, which is crucial for the safe operation of earth-rock dams. By providing timely and effective early warning information to operation and maintenance personnel, the model offers important references for earth-rock dam operation and maintenance decisions, helping to take timely measures and effectively ensure the safety of seepage in earth-rock dams. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of a spatiotemporal early warning method for dam seepage pressure provided by the present invention;

[0042] Figure 2 This is a schematic diagram of a structure of the diffusion convolutional recurrent neural network provided by the present invention;

[0043] Figure 3 This is a schematic diagram of a dam seepage pressure spatiotemporal early warning device provided by the present invention. Detailed Implementation

[0044] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0046] Existing research on seepage pressure time series prediction mainly focuses on single-point prediction schemes, which are insufficient to meet the needs of practical engineering. Furthermore, research on spatiotemporal early warning of seepage in multi-task earth-rock dams is relatively limited. In multi-task prediction, the importance of each task varies, making the effective measurement of each task a pressing issue. Therefore, this invention provides a spatiotemporal early warning method and device for dam seepage pressure. Through characterization learning of flow pressure time series, a multi-task prediction model is established, and an early warning model is built upon this model. Using this early warning model, the relationships between different seepage pressure monitoring data can be effectively handled, and an efficient and accurate multi-monitoring point multi-task seepage pressure prediction model can be established, achieving spatiotemporal early warning of dam seepage pressure.

[0047] like Figure 1 The diagram shown is a flowchart of a spatiotemporal early warning method for dam seepage pressure provided by the present invention, which includes the following steps:

[0048] Step 101: Collect raw seepage pressure time series data from multiple monitoring points during dam monitoring.

[0049] For example, real-time dam monitoring technology can be used to obtain raw seepage pressure time-series data of dams, specifically relying on sensor technology, Internet of Things (IoT) technology, and computer science technology. Through these technologies, it is possible to achieve real-time acquisition, transmission, storage, and analysis of seepage pressure in earth-rock dams.

[0050] Step 102: Preprocess the time series data.

[0051] For example, linear interpolation can be used to clean the time series data, and then Z-score normalization can be used to standardize the cleaned data, transforming the original data into high-quality data that can be used for modeling and analysis.

[0052] The above preprocessing steps can effectively reduce noise, extract useful information, and ensure data consistency, thus providing a reliable foundation for subsequent time series analysis and forecasting.

[0053] Step 103: Use the processed time series data from multiple monitoring points to perform characterization learning of the seepage pressure time series and establish a multi-task prediction model.

[0054] In this embodiment of the invention, the multi-task prediction model can employ a diffusing convolutional neural network. A diffusing convolutional recurrent layer can be constructed by combining graph convolution and recurrent neural networks to learn spatiotemporal representations. Graph convolution is used to simulate the spatial relationships of monitoring points; the recurrent neural network simulates the time dependence.

[0055] Specifically, a sliding window technique can be used to segment time-series data from multiple monitoring points into fixed-length subsequences, extract features from each subsequence, and establish labels for each subsequence to obtain a training sample set. A diffusing convolutional neural network is then constructed using this training sample set to obtain a multi-task prediction model.

[0056] During training, the model learns how to predict future values ​​of stress time series based on input features; the model's performance is evaluated using a validation set or test set, and the model parameters are adjusted based on the evaluation results to optimize model performance.

[0057] To simultaneously consider the relationships between different tasks, effectively integrate information from different tasks, and improve the accuracy and comprehensiveness of feature extraction, a non-limiting embodiment can utilize an improved diffusing convolutional recurrent neural network to comprehensively consider temporal and spatial relationships, effectively learn the spatiotemporal representation of seepage pressure, and then perform multi-task prediction of dam seepage pressure. Each prediction task corresponds to a monitoring point; that is, for each monitoring point, there is a prediction result learned through spatiotemporal representation, which represents the seepage pressure at the corresponding monitoring point within a certain future time period.

[0058] It should be noted that the input of the multi-task prediction model is the permeability pressure time series data of multiple monitoring points over a certain historical period, and the output is the permeability pressure time series data of each monitoring point over a certain future period.

[0059] Multivariate time series prediction based on multi-task learning aims to learn a set of functions for m variables. Use this function to output It closely approximates the actual label y. i Since the function F(x) is embedded in the multi-task learning model, the learning tasks related to the variable m are... Execute simultaneously, utilizing each task Effective knowledge can significantly improve model performance.

[0060] Furthermore, considering the importance of different time points to the prediction results, and to enable the multi-task prediction model to focus more on the key information in the input data, in this embodiment of the invention, the weight of each input feature in the multi-task prediction model can also be determined, that is, different input features can have different weights.

[0061] For example, in a non-limiting embodiment, the weight of each input feature in the multi-task prediction model can be determined based on an attention mechanism. This attention mechanism learns the weight of each input element and dynamically allocates weights to emphasize different parts of the input data, enabling the model to selectively process and utilize information, forming an attention distribution, thereby making more accurate spatiotemporal predictions of seepage pressure.

[0062] For example, the weights of each input feature can be determined using the following time attention formula:

[0063]

[0064] Q (Query): Query matrix, used to represent the inputs that need to be focused on; K (Key): Key matrix, which can be some kind of representation of the input sequence; V (Value): Value matrix, storing the actual value information of each position in the sequence; K TIt is the transpose of the key matrix, and its purpose is to calculate the similarity score between each query and the key by multiplying it with the query matrix Q; is a scaling factor, d is the dimension of query and key; the softmax operation transforms the similarity score into a probability distribution, representing the degree of attention each position gives to all other positions in the input sequence (attention weight).

[0065] It should be noted that the above formula is only an illustrative example. In specific implementation, other calculation formulas based on attention mechanisms can also be used, and this invention does not limit the scope of the invention.

[0066] In this embodiment of the invention, a diffusing convolutional recurrent neural network can be used to model spatiotemporal relationships, such as... Figure 2 The diagram shown illustrates one structure of this diffusing convolutional recurrent neural network. Input data is processed through diffusing convolutional recurrent layers, utilizing convolution to expand the perception and capture spatiotemporal dependencies. After processing with the ReLU activation function, compressed feature representations are generated. The final features are assigned to different tasks, enabling multi-task learning and improving the performance of multiple tasks by sharing features.

[0067] In the present invention, the diffuse convolutional neural network may include graph convolution and recurrent neural network; the graph convolution is used to simulate the spatial relationship of monitoring points; and the recurrent neural network simulates the time dependence.

[0068] Specifically, graph convolution can be achieved by constructing an adjacency matrix through the Pearson correlation between multiple monitoring sequences, as follows:

[0069] First, an adjacency matrix is ​​constructed based on the Pearson correlation between the monitored sequences, as follows:

[0070]

[0071] The meanings of the parameters in the formula are as follows:

[0072] r: Pearson correlation coefficient, which reflects the degree of linear correlation between two variables.

[0073] X i : The i-th data point of variable X.

[0074] Y i : The i-th data point of variable Y.

[0075] The mean of variable X.

[0076] The mean of variable Y.

[0077] The stationary distribution of the diffusion process in a diffusing convolutional recurrent layer can be represented as a weighted combination of infinite random walks on a graph structure composed of multiple measurement points, and calculated in closed form, as shown in the following formula:

[0078]

[0079] Where α is a parameter controlling the diffusion rate, determining the decay rate of each diffusion step; k is the diffusion step; (1-α) k It is an exponentially decaying factor, and its contribution gradually decreases as k increases; is the inverse of the degree matrix of the graph, representing the normalized weights of the nodes; W is the adjacency matrix of the graph, representing the connection relationships between the nodes.

[0080] In practice, a finite k-step truncation diffusion process can be used, with a trainable weight assigned to each step.

[0081] For graph signals X∈R N×P and filter f θ The spread convolution operation is defined as:

[0082]

[0083] Where X represents the input signal matrix, containing N nodes and P features; X :,p This represents the p-th column of the input matrix X, i.e., the p-th eigenvector. Represents the spread convolution operation, θ∈R k*2 Here are the parameters of the filter; the maximum stride of K-diffusion, i.e., the maximum number of diffusion steps considered by the convolutional layer; θ k,1 θ k,2 These are the parameters of the convolution filter; and These are the transition matrices for the diffusion process and the reverse process, respectively.

[0084] Let the parameter tensor be Θ∈R Q*P*K*2 =[θ] q,p , where Θ q,p,:,: ∈R K*2 This represents a convolutional filter with parameterized P-th input and q-th output.

[0085]

[0086] Where, X∈R N×P For input, H∈R N×Q For output, Let be a filter, and 'a' be an activation function (such as ReLU).

[0087] Diffusion convolutional layers learn representations of graph structured data and can be trained using methods based on stochastic gradients.

[0088] Then, recurrent neural networks (RNNs) were used to simulate time dependence.

[0089] For example, a gated recurrent unit (GRU) can be used, which is a simple yet powerful variant of an RNN. Replacing matrix multiplication in a GRU with diffusing convolutions yields a diffusing convolution gated recurrent unit (DCGRU).

[0090]

[0091] H (t) =u (t) ⊙H (t-1) +(1-u (t) )⊙C (t)

[0092] Reset door r (t) This determines how much information from the previous hidden state needs to be discarded in order to introduce new information for the current time step; σ is the Sigmoid activation function, used to restrict the output value to the range [0,1]; Θ r ... (t) And the hidden state H from the previous moment (t-1) Perform splicing; b r To reset the door's bias.

[0093] Update Gate u (t) This determines how much information in the current hidden state needs to be retained from the previous hidden state. u b are the parameters of the spread convolution filter, used to calculate the update gate; u To update the bias term of the gate.

[0094] Candidate hidden state C (t) It is used to generate new hidden states based on the current input and candidate information from part of the previous hidden state. tanh is the hyperbolic tangent function, with output values ​​between (-1,1); Θ C Here are the parameters of the spread convolution filter, used to compute the candidate hidden states; r (t) ⊙H (t-1) To reset the door's hidden state from the previous moment, multiply it element-wise; this is used to adjust the effect of the previous hidden state. c The bias term for the candidate hidden state.

[0095] Hidden state H (t)It is the result of updating the gate control, its hidden state H in the previous moment. (t-1) and the current candidate hidden state C (t) To achieve a balance between them.

[0096] In the above formula, X (t) and H (t) Let r represent the input and output at time t, respectively. (t) and u (t) These are the reset gate and update gate at time t, respectively. Expression for diffusing convolution, Θ r Θ u ,Θ C For the corresponding parameters.

[0097] Spatial dimension features are extracted using Graph Convolutional Networks (GCNs). The extracted feature Z can be represented as:

[0098]

[0099] X is the input to the GCN module; X∈R N×C Z is the output of the GCN module; Z∈R N×F A is the adjacency matrix of the graph; A∈R N ×N N is the number of nodes; C is the input dimension hyperparameter set during network training; F is the output dimension hyperparameter set during network training; D is a diagonal matrix containing the number of connections for each node in the graph; Θ is the network parameters that need to be learned during model training; b is the network parameter for the model bias term; I N It is an identity matrix.

[0100] Similar to GRU, DCGRU (Diffusion convolution+GRU) can be used to build recurrent neural network layers and train them through backpropagation.

[0101] Finally, the encoder part of the prediction model is predicted using an improved divergent convolutional recurrent neural network, which effectively learns the representation of the sequence and uses the learned representation for multi-task prediction.

[0102] Step 104: Establish an early warning model based on the multi-task prediction model.

[0103] The early warning model is built upon a multi-task prediction model by modifying the loss function. The model's input includes the predicted results for each monitoring point and the residuals compared to the actual values; the output is an interval. For example, if the previous prediction result for a point was 1, then the early warning model's result would be an interval of [0.5, 1.5].

[0104] In some embodiments, an early warning model can be established based on an uncertainty trade-off loss function and / or a quantile regression loss function.

[0105] By introducing an uncertainty loss function, the confidence level is dynamically adjusted during the prediction process to balance prediction error and uncertainty. Quantile regression is used to capture data trends under different risk conditions and predict results at different quantiles. The model's total loss function can use any of these functions or combine both, and the model is trained by minimizing this loss function through an optimizer. The trained model not only provides accurate predictions but also provides risk warnings based on the uncertainty of the prediction results and the output at different quantiles, making it suitable for decision-making and response in various risk and uncertainty scenarios.

[0106] Since each monitoring point is treated as a prediction task, and each prediction task is evaluated using a loss function improved by the uncertainty trade-off, the prediction is performed based on the different levels of importance of each task.

[0107] Due to their different functions, each monitoring point may have different demand distribution patterns. If these differences are ignored and the predictions of all monitoring points are treated as a single-task learning process, local errors may accumulate. Conversely, if monitoring points with different distribution patterns are subjected to multi-task learning, the predictions of monitoring points with different distribution patterns can share knowledge, improving generalization performance.

[0108] In this embodiment of the invention, the total number of monitoring points is M, which represents the total number of learning tasks. For each learning task, a fully connected neural network (FNN) can be used to complete the final prediction. For any monitoring point m = 1, 2, ..., M, prediction can be performed using the following formula:

[0109]

[0110] Among them, H m This represents the spatiotemporal learning component output of the m-th monitoring point. This represents the prediction result. For ease of labeling, let W be... m and b m Let represent all learnable weight parameters and bias parameters in the network, respectively. A straightforward approach to multi-task learning is to combine the losses for each learning task using a simple weighted linear sum.

[0111] However, the performance of this naive method is affected by the weight W. m The choice of weights is highly sensitive, and adjusting these weights is both difficult and expensive.

[0112] To address the aforementioned issues, a loss function improved by weighing uncertainty losses can be used for prediction.

[0113] The prediction results for all M monitoring points are expressed using the following formula:

[0114]

[0115] The model has M outputs, corresponding to M learning tasks.

[0116] Assuming the input is x, the model output is f. w,b (X), Y is the model output f w,b (X) is the mean, σ 2 The variance follows a normal distribution.

[0117] Multitasking possibilities can be described as follows:

[0118]

[0119] Where, σ m Let be the observation noise parameter of the model, and let represent the uncertainty of the m-th task.

[0120] The maximum value of the log-likelihood of the model parameters and the observation noise parameters is taken to obtain the minimization objective, i.e., the multi-task loss function, whose expression is:

[0121]

[0122] in, This represents the loss for the m-th task. (Coefficient term) This can be viewed as the relative weight of the m-th task. As σ... m The increase, The weights decrease. Furthermore, logσ1...σ m This acts as a regularizer to prevent σ from increasing too much. In practice, networks can be trained to predict the log-variance, s:=logσ.

[0123] Furthermore, the loss function plays a crucial role in generating results for machine learning models. In this embodiment of the invention, conditional expectation or median output can be generated based on the classical squared error loss of least squares regression. The loss function for the interval prediction model is constructed using the idea of ​​quantile regression. The conditional τ-th quantile of the target distribution consists of a weighted sum of multiple individual predictions. Quantile loss. The weighting coefficients are estimated by minimizing the derived function of the absolute residuals.

[0124]

[0125] Where τ∈(0,1) is the target quantile. The larger the τ quantile, the greater the underestimation loss and the smaller the overestimation loss in terms of quantile loss. Represents the predicted value, y i Represents the actual value.

[0126] As can be seen from the definition, the loss function is asymmetrically weighted. When the quantile predicted value... Greater than the measured value y i If the penalty is multiplied by (1-τ), then the penalty will be multiplied by (1-τ); otherwise, it will be multiplied by τ.

[0127] The upper and lower bounds of the quartiles are mapped to confidence intervals, and quantile loss is used instead of the original squared error loss for interval prediction. Specifically, the quantiles of the upper and lower bounds are set to τ. u =1-α / 2 and τ l =α / 2, making it equivalent to the confidence level. However, when the prediction error is zero, the original quantile loss is non-differentiable, which means... Therefore, the Huber norm, which combines the L1 and L2 norms, is used to modify the loss function.

[0128]

[0129] In the formula, Represents the predicted value, y i Representing the true value, δ>0, it determines whether the Huber norm tends towards the L1 or L2 norm. Compared to the original quantile loss, the redefined loss function is differentiable everywhere (i.e., the new loss function is smooth over its entire domain). Combining this loss function with the proposed deep learning point prediction model allows for the construction of an interval prediction model based on ensemble deep learning. In this way, the accuracy and stability of the prediction results can be evaluated more comprehensively, providing decision-makers with more reliable information.

[0130] Accordingly, the aforementioned early warning model can be used to monitor and issue early warnings for dams. Specifically, current seepage pressure time series data from multiple monitoring points are acquired; based on the current seepage pressure time series data and the early warning model, when the predicted result exceeds the early warning range, real-time monitoring and early warning of the dam are conducted.

[0131] In this embodiment of the invention, by introducing an uncertainty trade-off loss function, not only is the accuracy of the predicted values ​​considered, but also the model's confidence in the results. By dynamically adjusting the uncertainty parameter, the model balances prediction accuracy and uncertainty, avoiding overfitting to unstable data. By introducing a quantile regression loss function, the model can predict values ​​at different quantiles, capturing data trends under different risk levels. This allows the model to generate corresponding prediction results under different risk scenarios, thus better adapting to diverse real-world situations.

[0132] When building the model, uncertainty loss and quantile loss are combined into a multi-objective loss function. This loss function is minimized by training the model, and the weight parameters are updated accordingly. After evaluation and testing, the model can provide early warnings based on prediction results in practical applications. If the model detects a quantile result with high uncertainty or high risk, it will issue an early warning. During deployment, the model is continuously iterated and optimized, constantly updated based on new data, ensuring the real-time performance and accuracy of the early warning system.

[0133] Accordingly, embodiments of the present invention also provide a dam seepage pressure spatiotemporal early warning device, such as... Figure 3 The diagram shown is a structural schematic of the device.

[0134] The dam seepage pressure spatiotemporal early warning device 300 includes the following modules:

[0135] Data collection module 301 is used to collect raw seepage pressure time series data from multiple monitoring points during dam monitoring;

[0136] Preprocessing module 302 is used to preprocess the time series data;

[0137] The multi-task prediction model building module 303 is used to perform characterization learning of seepage pressure time series using the processed time series data from multiple monitoring points, and to establish a multi-task prediction model.

[0138] The early warning model construction module 304 is used to build an early warning model based on the multi-task prediction model.

[0139] Furthermore, the dam seepage pressure spatiotemporal early warning device 300 also includes: an early warning module (not shown in the figure), used to acquire current seepage pressure time series data of multiple monitoring points, and to perform real-time monitoring and early warning of the dam based on the current seepage pressure time series data and the early warning model.

[0140] The specific implementation methods of each module in the dam seepage pressure spatiotemporal early warning device 300 can be referred to the description in the previous embodiments of the present invention, and will not be repeated here.

[0141] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0142] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0143] In the several embodiments provided by the present invention, it should be understood that the disclosed apparatus can be implemented in other ways.

[0144] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0145] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means.

[0146] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and systems of the present invention, and are merely some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatiotemporal early warning method for dam seepage pressure, characterized in that, The method includes: Collect raw seepage pressure time series data from multiple monitoring points during dam monitoring; The time series data is preprocessed; A multi-task prediction model was established by using time series data from multiple monitoring points after processing to characterize the seepage pressure time series. An early warning model is established based on the aforementioned multi-task prediction model; The step of using processed time-series data from multiple monitoring points to perform characterization learning of seepage pressure time series and establishing a multi-task prediction model includes: The time series data of the multiple monitoring points are segmented using a sliding window to obtain subsequences; Extract the features of the subsequences and establish labels for each subsequence to obtain a training sample set; A diffusing convolutional neural network is constructed using the training sample set to obtain a multi-task prediction model; The establishment of an early warning model based on the multi-task prediction model includes: Determine the loss function, which includes an uncertainty trade-off loss function; The uncertainty trade-off loss function is specifically configured as follows: The prediction results for all M monitoring points are expressed using the following formula: ； The model has M outputs, corresponding to M learning tasks; Assuming the input is X, the model output is Y is the output of the model. For the mean, The variance follows a normal distribution. Multitasking possibilities can be described as follows: in, Let be the observation noise parameter of the model, representing the uncertainty of the Mth task; The maximum value of the log-likelihood of the model parameters and the observation noise parameters is taken to obtain the minimization objective, which is the uncertainty trade-off loss function for multi-task tasks, and its expression is: in, Represents the loss of the Mth task; coefficient term It is the relative weight of the Mth task, as... The increase, The weight decreases; in addition, As a regularizer to prevent Add too much and train the network to predict the log-variance, s = log ; An early warning model is established based on the aforementioned loss function.

2. The spatiotemporal early warning method for dam seepage pressure according to claim 1, characterized in that, The preprocessing of the time series data includes: The time series data is preprocessed and cleaned using linear interpolation; Z-score normalization was used to standardize the cleaned data.

3. The spatiotemporal early warning method for dam seepage pressure according to claim 2, characterized in that, Different input features in the multi-task prediction model have different weights.

4. The spatiotemporal early warning method for dam seepage pressure according to claim 3, characterized in that, The method further includes: The weight of each input feature in the multi-task prediction model is determined based on an attention mechanism.

5. The spatiotemporal early warning method for dam seepage pressure according to claim 3, characterized in that, The diffuse convolutional neural network includes graph convolution and recurrent neural network; the graph convolution is used to simulate the spatial relationship of monitoring points; the recurrent neural network simulates the time dependence.

6. The spatiotemporal early warning method for dam seepage pressure according to claim 5, characterized in that, The method further includes: By constructing an adjacency matrix through the Pearson correlation between multiple monitoring sequences, graph convolution is achieved.

7. The method for spatiotemporal early warning of dam seepage pressure according to any one of claims 1 to 6, characterized in that, The method further includes: Acquire current permeability pressure time series data from multiple monitoring points; Based on the current seepage pressure time series data and the early warning model, the dam is monitored and warned in real time.

8. A spatiotemporal early warning device for dam seepage pressure, characterized in that, The device includes: The data collection module is used to collect raw seepage pressure time series data from multiple monitoring points during dam monitoring. The preprocessing module is used to preprocess the time series data; The multi-task prediction model building module is used to perform characterization learning of seepage pressure time series using processed time series data from multiple monitoring points, and to establish a multi-task prediction model. The early warning model construction module is used to build an early warning model based on the multi-task prediction model. The step of using processed time-series data from multiple monitoring points to perform characterization learning of seepage pressure time series and establishing a multi-task prediction model includes: The time series data of the multiple monitoring points are segmented using a sliding window to obtain subsequences; Extract the features of the subsequences and establish labels for each subsequence to obtain a training sample set; A diffusing convolutional neural network is constructed using the training sample set to obtain a multi-task prediction model; The establishment of an early warning model based on the multi-task prediction model includes: Determine the loss function, which includes an uncertainty trade-off loss function; The uncertainty trade-off loss function is specifically configured as follows: The prediction results for all M monitoring points are expressed using the following formula: ； The model has M outputs, corresponding to M learning tasks; Assuming the input is X, the model output is Y is the output of the model. The mean, The variance follows a normal distribution. Multitasking possibilities can be described as follows: in, Let be the observation noise parameter of the model, representing the uncertainty of the Mth task; The maximum value of the log-likelihood of the model parameters and the observation noise parameters is taken to obtain the minimization objective, which is the uncertainty trade-off loss function for multi-task tasks, and its expression is: in, Represents the loss of the Mth task; coefficient term It is the relative weight of the Mth task, as... The increase, The weight decreases; in addition, As a regularizer to prevent Add too much and train the network to predict the log-variance, s = log ; An early warning model is established based on the aforementioned loss function.

9. The dam seepage pressure spatiotemporal early warning device according to claim 8, characterized in that, The device further includes: The early warning module is used to acquire the current seepage pressure time series data of multiple monitoring points, and to conduct real-time monitoring and early warning of the dam based on the current seepage pressure time series data and the early warning model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the spatiotemporal early warning method for dam seepage pressure as described in any one of claims 1 to 7.

Images