Deformation prediction method and system considering complex correlation of multiple measuring points of rockfill dam in time and space fusion

Abstract

The application provides a deformation prediction method and system considering the space-time fusion of the complex correlation of multiple measuring points of a rock-fill dam, can fully consider the complex correlation between multiple measuring points for space-time fusion, and realizes accurate and real-time prediction of overall deformation in view of the characteristics of high noise and drift of rock-fill dam deformation monitoring data. The deformation prediction method comprises the following steps: step 1, monitoring data preprocessing: preprocessing the monitored dam deformation and environmental factor time series data; step 2, constructing a sample based on the preprocessed data, adopting a sliding window mode to build a data set, and dividing the sample into a training set and a validation set in proportion; step 3, determining the prior information to be injected in the multi-measuring point fusion process under the complex relationship and calculating an adjacency matrix; step 4, constructing a space-time fusion deformation prediction model; step 5, training the deformation prediction model by using the training set, obtaining the optimal model through the validation set, and completing the quasi-real-time deformation prediction according to the latest monitoring data.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic structure engineering technology, specifically involving a deformation prediction method and system that considers the spatiotemporal fusion of complex correlations among multiple measuring points in rockfill dams. Background Technology

[0002] Models for predicting the deformation of rockfill dams are categorized into deterministic models, hybrid models, and statistical models. The first two require solving physical equations and belong to the model-driven category; the latter establishes regression relationships between influencing factors and deformation effect sizes based on monitoring data, primarily fitting multivariate linear equations. With the development of data science and artificial intelligence, machine learning and intelligent optimization algorithms are also being used for nonlinear regression problems. Existing rockfill dam deformation prediction methods utilize deep learning models such as Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), temporal convolutional networks (TRCs), and Transformers to consider the temporal dependencies between variables.

[0003] However, monitoring data for rockfill dams exhibits characteristics of drift and high noise, issues that existing models have not yet addressed, resulting in low generalization ability. Dam engineering theory posits that a rockfill dam is a deformation-coordinated whole, and currently, no model has incorporated prior information to account for the complex relationships among multiple measuring points, nor has it achieved spatiotemporal information fusion or completed overall prediction of the dam body across multiple measuring points. Summary of the Invention

[0004] This invention is made to solve the above-mentioned problems. Its purpose is to provide a deformation prediction method and system that takes into account the complex spatiotemporal fusion of multiple measuring points in rockfill dams. It can fully consider the complex correlation between multiple measuring points for spatiotemporal fusion and achieve near real-time prediction of overall deformation, taking into account the characteristics of high noise and drift in rockfill dam deformation monitoring data.

[0005] To achieve the above objectives, the present invention adopts the following solution:

[0006] <Method>

[0007] like Figure 1 As shown, this invention provides a deformation prediction method that considers the complex spatiotemporal fusion of multiple measuring points in a rockfill dam, characterized by comprising the following steps:

[0008] Step 1, Data Preprocessing: Preprocess the time series data of dam deformation and environmental factors.

[0009] Step 2: Construct samples based on the preprocessed data, and use a sliding window method to construct the dataset, dividing the samples into training set and validation set according to the proportion;

[0010] Step 3: Determine the prior information to be injected during the multi-measurement point fusion process under complex relationships, and calculate the adjacency matrix;

[0011] Step 4: Construct a spatiotemporal fusion deformation prediction model;

[0012] Step 4.1: Feature space fusion under complex relationships with multiple measurement points based on graph convolutional networks;

[0013] In the dam monitoring system, measuring points are treated as nodes with corresponding feature vectors. During each graph convolution process, the current measuring point is considered the central measuring point, and neighboring measuring points connected to it transfer their features to the central measuring point. The graph convolution process can be represented by the following equation:

[0014]

[0015] In the formula, X (l) Let W be the feature matrix of the l-th layer. (l) Let be the trainable weight matrix of the l-th layer, σ(·) be the nonlinear transformation function, and A be the weighted adjacency matrix used as prior information. In graph convolution, self-connections of center nodes need to be considered; therefore, the actual calculation uses... Where I is the identity matrix; for The degree matrix, with diagonal elements taking...

[0016] Multi-point feature fusion is performed to consider the complex relationships among multiple measurement points.

[0017]

[0018] In the formula, GCN (·) A represents a graph convolution method using adjacency matrices under various prior information. (·) Let X represent the adjacency matrix under each relation, and let X represent the input multi-point feature matrix. (·) X represents the fusion characteristics under each relation. fusion The final fusion feature considering m complex relationships;

[0019] Step 4.2: Use Long Short-Term Memory (LSTM) units to perform temporal dependency mining;

[0020] Step 4.3 proposes two spatiotemporal fusion unit methods based on the spatial feature fusion process of graph convolutional networks and the temporal dependency mining process based on long short-term memory units: discrete spatiotemporal fusion unit and coupled spatiotemporal fusion unit. The discrete spatiotemporal fusion unit designs the spatial fusion process and temporal dependency mining in a serial manner, where the output of spatial fusion serves as the input to the long short-term memory unit. The coupled spatiotemporal fusion unit combines graph convolution with long short-term memory units, using graph convolution to calculate information such as the forget gate, update weights, and output gate, replacing the trainable matrix W in the traditional long short-term memory unit.

[0021] In the discrete-time fusion unit, steps 4.1 and 4.2 are connected in series, and the input X of step 4.2 is... t The final fused feature X, which takes into account various complex relationships, is the output of step 4.1. fusion ;

[0022] In the coupled spatiotemporal fusion unit, steps 4.1 and 4.2 are coupled. The forget gate, update weights and output gate in the long short-term memory unit are all calculated using graph convolution, replacing the trainable matrix W in the long short-term memory unit.

[0023] Step 4.4: Assuming that the noise distribution has temporal differences, the spatiotemporal fusion unit outputs hidden layer memory H. t Gaussian distribution parameters for deformation probability prediction are obtained through linear layers;

[0024] Step 5: Train the deformation prediction model using the training set, obtain the optimal model using the validation set, and complete near real-time deformation prediction based on the latest monitoring data.

[0025] Preferably, in the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points, provided by the present invention, the calculation formulas for the forget gate, update weights, and output gate information of the coupled spatiotemporal fusion unit in step 4.3 are as follows:

[0026]

[0027] In the formula, X t The input features for the current time step; H t-1 C t-1 H t C t F represents the hidden memory H and cell state C of the previous time step t-1 and the current time step t, respectively; t Forget gates represent the filtering of important information from the previous time step. GCN xf (·), GCN hf (·) represents the trainable graph convolution process with the forget gate, b f The bias of the forget gate; I t To update the weights, GCN xi (·), GCN hi (·) represents the convolution process of the trainable graph for updating the gate, b i To update the gate bias; Information pending updates, GCN xc (·), GCN hc (·) represents the convolution process of the trainable graph containing the information to be updated, b c The bias of the information to be updated; O t For output gates, GCN xo(·), GCN ho (·) represents the trainable graph convolution process of the output gate, b o To update the gate bias.

[0028] Preferably, in the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points, provided by the present invention, the probability prediction output in step 4.4 is expressed by the following formula:

[0029] μ t =W μ H t +b μ

[0030] σ t =softplus(W σ H t +b σ )

[0031] y t ~N(μ) t ,σ t )

[0032] In the formula, W μ b μ W σ b σ For trainable parameters, Softplus(·) is the activation function type, and y t This represents the final deformation prediction value of the deformation prediction model.

[0033] Preferably, the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points provided by the present invention includes the following sub-steps in step 5:

[0034] Step 5.1 describes the inference task of the model trained throughout the entire process as follows: Given a multi-point deformation hysteresis sequence. With multi-point environmental factor (e.g., time factor, water level factor) series Predicting future deformation sequences at multiple measurement points T0 represents the historical lag length, and T1 represents the prediction length; environmental factors and effect sizes are combined into a matrix form X. t =concat(y t-1 ,x t ), which serves as the initial input feature X for the corresponding time step spatiotemporal fusion unit;

[0035] The model is trained using a training set throughout the entire process, with the output and label sequence being processed simultaneously. They are used together to calculate the loss function and trained using the Adam optimizer;

[0036] Step 5.2 involves substituting the validation set into the model for sample inference in each training round, and setting an early stopping mechanism to obtain the final model that optimizes the validation set metrics. During inference, the feature matrix... Substituting these parameters into the model sequentially, we obtain the Gaussian distribution parameters at time T0+1, and then sample from them to obtain... Then, this predicted value is treated as the known data and iterated into the input feature matrix at time T0+2. Similarly, the inferred sequence is finally obtained after T1-1 iterations.

[0037] Step 5.3: After obtaining the optimal prediction model, rolling prediction is performed based on the last T0 data points of the real-time deformation monitoring sequence. The prediction results of the previous round are added to the deformation sequence, and the data of the last T0 length are treated as known values ​​to participate in the next prediction. Finally, a real-time deformation prediction sequence that meets the prediction length is obtained.

[0038] Preferably, in the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points, provided by the present invention, the probability prediction loss function is calculated using the following formula in step 5.1:

[0039]

[0040] In the formula, the subscripts s, n and t represent the sample number, measurement point number and time step, respectively, p(X) is the sample distribution and S is the total number of training samples.

[0041] Preferably, the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points provided by the present invention includes the following sub-steps in step 3:

[0042] Step 3.1: Based on the characteristics of the deformation monitoring data of the rockfill dam and combined with specific engineering experience, determine the prior information that needs to be injected under the complex correlation of multiple measuring points;

[0043] Step 3.2: Determine the calculation method of the adjacency matrix based on prior information.

[0044] Preferably, the deformation prediction method for considering the complex spatiotemporal fusion of multiple measuring points in a rockfill dam provided by the present invention, in step 3.2, needs to consider prior information under complex relationships, including: spatial distance between measuring points, arrangement of measuring points, partition where measuring points are located, and similarity of measuring point sequences, etc.

[0045] The following are some methods for calculating the adjacency matrix to address the complex relationships described above:

[0046] Spatial distance Gaussian similarity function, considering the spatial distance relationship between multiple measurement points:

[0047]

[0048] In the formula, dist(i,j) represents the distance between the two measuring points in Euclidean space, and σ is the statistical standard deviation of the distance;

[0049] A 0-1 step function, considering the relationship between multiple measurement points, such as measurement point layout and zoning attributes:

[0050]

[0051] In the formula, P i P j It represents a certain reference attribute between two measuring points;

[0052] Sequence similarity function, taking into account the similarity of actual monitored sequences:

[0053]

[0054] In the formula, DTW(i,j) represents the similarity distance between two measuring points under dynamic time warping.

[0055] Depending on the specific project requirements, any one or more of the above three methods can be selected to calculate the adjacency matrix; alternatively, other calculation methods can be designed in step 3 based on the required injected association information to calculate the adjacency matrix. These methods do not affect the calculation process in subsequent steps, and the present invention can be applied to all of them.

[0056] Preferably, in step 4.2, the deformation prediction method of the present invention, which considers the complex spatiotemporal fusion of multiple measuring points of rockfill dams, uses long short-term memory units as a tool for temporal dependency mining. Historical information is transmitted to the current moment as a carrier of cell state and hidden memory in temporal fusion, and temporal dependency information is mined in combination with the current feature input.

[0057] Preferably, in the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points, the present invention, in step 2, treats the preprocessed multi-measuring point deformation monitoring data from step 1 as the following sequence: {y1, y2, ..., y t-1 ,y t},y∈ N The total length is t, and the total number of measurement points is N. A window of fixed length T0+T1 is gradually slid across the multi-measurement point sequence. The data in each window is regarded as a training sample, thereby obtaining a total number of training samples of t-T0-T1+1, as follows:

[0058]

[0059] In the formula, T0 represents the lag length of the model input data, and T1 represents the prediction length of the model output data.

[0060] All the above samples are divided into training set and validation set according to a certain ratio. Generally, 20%-30% of the tail sample data is taken as the validation set.

[0061] Preferably, the deformation prediction method for rockfill dams that considers the complex spatiotemporal fusion of multiple measuring points provided by the present invention includes, in step 1, quality inspection, interpolation resampling, and normalization. Specifically, the time series of dam deformation and environmental factors are collected by the monitoring device in the dam site area and quality inspection is performed. The main process is as follows: removing measuring points with long missing segments and few sampling numbers; using the Grubbs test to remove outliers with obvious abnormalities in the sequence; using the linear interpolation method to resample the sequence at the target frequency; and normalizing the sequence (the normalization method can be 0-1 normalization, -1-1 normalization, Z-Score normalization, etc.).

[0062] In addition, all the methods described above in this invention only require defining a small number of model hyperparameters and providing raw data in tabular form in a text document. The remaining steps and techniques are designed into a complete program using, for example, the Python language and the PyTorch framework.

[0063] <System>

[0064] Furthermore, this invention also provides a deformation prediction system that considers the complex spatiotemporal fusion of multiple measuring points in a rockfill dam, capable of automatically implementing the above-mentioned <method>, characterized by comprising:

[0065] The monitoring data preprocessing department performs preprocessing on the time series data of dam deformation and environmental factors.

[0066] The sample construction unit uses the preprocessed data as samples, constructs the dataset using a sliding window method, and divides the samples into training and validation sets proportionally.

[0067] The prior information determination unit determines the prior information to be injected during the fusion of multiple measurement points under complex relationships and calculates the adjacency matrix.

[0068] The prediction model construction department constructs a spatiotemporal fusion deformation prediction model according to the following steps 4.1 to 4.4;

[0069] Step 4.1: Feature space fusion under complex relationships with multiple measurement points based on graph convolutional networks;

[0070] In the dam monitoring system, each measuring point is treated as a node with a corresponding feature vector. During each graph convolution process, the current measuring point is considered the central measuring point, and neighboring measuring points connected to it transfer their features to the central measuring point. Multi-measuring point feature fusion is performed considering the complex relationships between multiple measuring points.

[0071]

[0072] In the formula, GCN (·) A represents a graph convolution method using adjacency matrices under various prior information. (·) Let X represent the adjacency matrix under each relation, and let X represent the input multi-point feature matrix. (·) X represents the fusion characteristics under each relation. fusion The final fusion feature considering m complex relationships;

[0073] Step 4.2: Use Long Short-Term Memory (LSTM) units to perform temporal dependency mining;

[0074] Step 4.3: Construct two types of spatiotemporal fusion units: discrete spatiotemporal fusion unit and coupled spatiotemporal fusion unit;

[0075] In the discrete-time fusion unit, steps 4.1 and 4.2 are connected in series, and the input X of step 4.2 is... t The final fused feature X, which takes into account various complex relationships, is the output of step 4.1. fusion ;

[0076] In the coupled spatiotemporal fusion unit, steps 4.1 and 4.2 are coupled. The forget gate, update weights and output gate information in the long short-term memory unit are all calculated using graph convolution, replacing the trainable matrix W in the long short-term memory unit.

[0077] Step 4.4: Assuming that the noise distribution has temporal differences, the spatiotemporal fusion unit outputs hidden layer memory H. t Gaussian distribution parameters of deformation are obtained through linear layers;

[0078] The deformation prediction unit trains the deformation prediction model using the training set, obtains the optimal model through the validation set, and completes near real-time deformation prediction based on the latest monitoring data.

[0079] The input display unit is used to allow users to input operation commands and to display the corresponding commands.

[0080] The control unit is communicatively connected to the monitoring data preprocessing unit, sample construction unit, prior information determination unit, prediction model construction unit, deformation prediction unit, and input display unit, and controls their operation.

[0081] The role and effect of invention

[0082] (1) This invention takes into account the complex correlation between multiple measuring points in a rockfill dam and proposes two spatiotemporal fusion units, so that the spatial characteristics and temporal dependence information under the complex relationship of multiple measuring points can be integrated and progressively advanced in the time axis, and the overall deformation prediction can be realized based on the spatiotemporal fusion characteristics.

[0083] (2) In view of the characteristics of high noise and drift of rockfill dam data, this invention proposes probabilistic prediction and full-process training methods, which can solve the long-term prediction problem of drift data and improve the robustness of the model for modeling various types of data.

[0084] (3) This invention has low requirements for model hyperparameter formulation and can achieve overall prediction of rockfill dams at multiple measurement points through a single model. It has a small number of model parameters, fast training and inference speed, and can be deployed to a digital twin platform to achieve overall, accurate, efficient and near real-time prediction of rockfill dam deformation.

[0085] In summary, this invention can achieve comprehensiveness, accuracy, and efficiency in predicting the deformation of rockfill dams. Attached Figure Description

[0086] Figure 1 The flowchart of the spatiotemporal fusion deformation prediction method for rockfill dams, which considers the complex correlation of multiple measuring points, is the subject of this invention.

[0087] Figure 2 This is a time history diagram before and after preprocessing according to an embodiment of the present invention;

[0088] Figure 3 This is a schematic diagram illustrating the calculation of the weighted adjacency matrix of the Gaussian distance similarity function based on the layout of measuring points in a rockfill dam, according to an embodiment of the present invention.

[0089] Figure 4 This is a schematic diagram of the structure of the shallow series spatiotemporal fusion unit involved in an embodiment of the present invention;

[0090] Figure 5 This is a schematic diagram illustrating the model training and inference process involved in an embodiment of the present invention;

[0091] Figure 6 This is a graph showing the changes in model training loss and validation metrics in an embodiment of the present invention.

[0092] Figure 7 The above are comparison charts of prediction results of various models involved in the embodiments of the present invention; wherein, (a) is a training sequence diagram of four test points, and (b) to (e) are comparison charts of prediction results of test points 4, 13, 30 and 37 respectively;

[0093] Figure 8 This is a comparison chart of the model parameters and training time involved in the embodiments of the present invention;

[0094] Figure 9 This is a schematic diagram illustrating the near real-time prediction results of the latest monitoring data involved in an embodiment of the present invention. Detailed Implementation

[0095] The following describes in detail, with reference to the accompanying drawings, the specific implementation scheme of the deformation prediction method and system that considers the complex spatiotemporal fusion of multiple measuring points in rockfill dams, which is involved in this invention.

[0096] <Example>

[0097] like Figure 1 As shown, this embodiment takes a rockfill dam project with a concrete panel as an example and proposes a spatiotemporal fusion deformation prediction method considering the complex correlation of multiple measuring points in a rockfill dam, including the following steps:

[0098] Step 1 involves quality checks, interpolation, resampling, and normalization of the time series data on dam deformation and environmental factors. This project considers the largest monitoring section; after removing 31 monitoring points with significant data gaps, a total of 37 monitoring point sequences were used as the research object. Figure 2 This is the time history curve of a certain measurement point sequence before and after the above preprocessing steps.

[0099] Step 2: The dataset is constructed using a sliding window method. In the experiment, the total sequence length t is 158, the lag length T0 is 12, the prediction length T1 is 4, and the total number of samples is 143. The samples are divided in an 8:2 ratio, resulting in 114 training samples and 29 validation samples.

[0100] Step 3, taking the prior information of spatial distance between measuring points as an example, calculate the weighted adjacency matrix based on the Gaussian similarity function of spatial distance. The coordinate information between each measuring point can be obtained from the design drawings, and the following can be calculated: Figure 3 The matrix shown.

[0101] Step 4: Construct a spatiotemporal fusion deformation prediction model. Taking a shallow spatiotemporal fusion unit as an example, Figure 4 The graph convolution operation and long short-term memory are cascaded, considering 37 measurement points. The input features include historical settlement deformation, upstream water level, and time-related factors, as shown in the following formula:

[0102] {y t-1 ,Δ t ,(θ / 100) t ln(θ / 100) t}

[0103] Therefore, the input feature shape of the graph convolutional network is 37×4. The features obtained through the graph convolutional network are directly used as the input of the long short-term memory, and the hidden layer dimension is determined to be 20. The output hidden memory shape after spatiotemporal fusion is calculated by the network and is 37×20. Finally, the Gaussian distribution parameters of the linear layer output are used, with a mean and standard deviation shape of 37×1 and 37×1, respectively.

[0104] Step 5: Train the model using a full-process approach, and perform inference on the validation set to obtain the optimal model. Model training and inference are as follows: Figure 5 As shown, the optimal model is obtained through the early stopping mechanism. Figure 6 The curves show the changes in the loss function and validation metrics during model training.

[0105] To highlight the advantages of this model in terms of prediction accuracy and training efficiency compared to other models in predicting the overall deformation of rockfill dams, a comparative experiment was designed. In this model, hyperparameters such as lag length, prediction length, and number of neurons were initially determined based on experience, while other models used Bayesian hyperparameter optimization to select their optimal hyperparameters. Four representative measurement points were selected for deformation prediction performance testing, such as... Figure 7 As shown, the settlement deformation data at measuring point 4 exhibits obvious periodic changes, the data at measuring point 13 shows a mixture of trend and periodicity, the data at measuring point 30 is mainly trend-based, and the data at measuring point 37 shows a typical step-like increase. The above-mentioned types of changes at measuring points basically cover several modes of deformation of rockfill dams. Figure 8 The comparison of the number of model parameters and training time shows that, compared with other methods, this method not only has high prediction accuracy, but also high model efficiency and low requirements for hyperparameter formulation, which can meet the requirements for near real-time prediction of the overall deformation of rockfill dams.

[0106] Subsequently, based on the latest existing monitoring data, a one-year rolling long-term forecast was conducted, and the forecast results are as follows: Figure 9 As shown in the figure, the prediction results indicate that the deformation sequence at the measuring points conforms to the cyclic loading and unloading deformation pattern of the dam body caused by reservoir water level fluctuations. Dam deformation will continue to accumulate, and the magnitude of deformation growth has not shown a significant convergence trend. Therefore, the prediction results of this model are reasonable and can be applied to the deformation prediction of rockfill dams.

[0107] <Example 2>

[0108] In this second embodiment, a deformation prediction system considering the complex spatiotemporal fusion of multiple measuring points in a rockfill dam is provided, which can automatically implement the above-mentioned method of the present invention. The system includes a preprocessing unit, a sample construction unit, a priori information determination unit, a prediction model construction unit, a deformation prediction unit, an input display unit, and a control unit.

[0109] The preprocessing department performs the steps described in step 1 above to preprocess the time series data of the monitored dam deformation and environmental factors.

[0110] The sample construction section follows the steps described in step 2 above, using the preprocessed data as samples, constructing the dataset using a sliding window approach, and dividing the samples into training and validation sets proportionally.

[0111] The prior information determination section follows the steps described in step 3 above, determines the prior information to be injected during the multi-measurement point fusion process under complex relationships, and calculates the adjacency matrix.

[0112] The prediction model construction section follows the steps described in step 4 above to construct a spatiotemporal fusion deformation prediction model.

[0113] The deformation prediction department follows the steps described in step 5 above, trains the deformation prediction model using the training set, obtains the optimal model using the validation set, and completes near real-time deformation prediction based on the latest monitoring data.

[0114] The input display unit is used to allow users to input operation commands and displays the corresponding information. For example, the input and output data and processing procedures of each unit can be displayed in the form of text, tables, static or dynamic diagrams, or two-dimensional or three-dimensional model diagrams.

[0115] The control unit is communicatively connected to the preprocessing unit, sample construction unit, prior information determination unit, prediction model construction unit, deformation prediction unit, and input display unit, and controls their operation.

[0116] The above embodiments are merely illustrative examples of the technical solutions of the present invention. The deformation prediction method and system for considering the complex spatiotemporal fusion of multiple measuring points in rockfill dams, as described in the present invention, are not limited to the content described in the above embodiments, but are defined by the scope of the claims. Any modifications, additions, or equivalent substitutions made by those skilled in the art based on these embodiments are within the scope of protection claimed by the claims of the present invention.

Claims

1. A deformation prediction method considering the complex spatiotemporal fusion of multiple measuring points in a rockfill dam, characterized in that, Includes the following steps: Step 1, Data Preprocessing: Preprocess the time series data of dam deformation and environmental factors. Step 2: Construct samples based on the preprocessed data, and use a sliding window method to construct the dataset, dividing the samples into training set and validation set according to the proportion; Step 3: Determine the prior information to be injected during the multi-measurement point fusion process under complex relationships, and calculate the adjacency matrix; Step 4: Construct a spatiotemporal fusion deformation prediction model; Step 4.1: Feature space fusion under complex relationships with multiple measurement points based on graph convolutional networks; In the dam monitoring system, measuring points are treated as nodes with corresponding feature vectors. During each graph convolution process, the current measuring point is considered the central measuring point, and neighboring measuring points connected to it pass their features to the central measuring point. Multi-measuring point feature fusion is performed considering the complex relationships among multiple measuring points. In the formula, GCN (·) A represents a graph convolution method using adjacency matrices under various prior information. (·) Let X represent the adjacency matrix under each relation, and let X represent the input multi-point feature matrix. (·) X represents the fusion characteristics under each relation. fusion The final fusion feature considering m complex relationships; Step 4.2: Use Long Short-Term Memory (LSTM) units to perform temporal dependency mining; Step 4.3: Construct two types of spatiotemporal fusion units: discrete spatiotemporal fusion unit and coupled spatiotemporal fusion unit; In the discrete-time fusion unit, steps 4.1 and 4.2 are connected in series, and the input X of step 4.2 is... t The final fused feature X, which takes into account various complex relationships, is the output of step 4.

1. fusion ; In the coupled spatiotemporal fusion unit, steps 4.1 and 4.2 are coupled. The forget gate, update weights and output gate information in the long short-term memory unit are all calculated using graph convolution, replacing the trainable matrix W in the long short-term memory unit. Step 4.4: Assuming that the noise distribution has temporal differences, the spatiotemporal fusion unit outputs hidden layer memory H. t Gaussian distribution parameters of deformation are obtained through linear layers; Step 5: Train the deformation prediction model using the training set, obtain the optimal model using the validation set, and complete near real-time deformation prediction based on the latest monitoring data.

2. The deformation prediction method for rockfill dams that considers the complex spatiotemporal correlations of multiple measuring points, as described in claim 1. Its features are: In step 4.3, the formulas for calculating the forget gate, update weights, and output gate information of the coupled spatiotemporal fusion unit are as follows: In the formula, X t The input features are for the current time step; H t-1 C t-1 H t C t F represents the hidden memory H and cell state C of the previous time step t-1 and the current time step t, respectively; t Forget gates represent the filtering of important information from the previous time step. GCN xf (·), GCN hf (·) represents the trainable graph convolution process with the forget gate, b f The bias of the forget gate; I t To update the weights, GCN xi (·), GCN hi (·) represents the convolution process of the trainable graph for updating the gate, b i To update the gate bias; Information pending updates, GCN xc (·), GCN hc (·) represents the convolution process of the trainable graph containing the information to be updated, b c The bias of the information to be updated; O t For output gates, GCN xo (·), GCN ho (·) represents the trainable graph convolution process of the output gate, b o To update the gate bias.

3. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points as described in claim 1, characterized in that: in, In step 4.4, the probability prediction output is represented by the following formula: m t =W μ H t +b μ σ t *softplus(W σ H t +b σ ) y t ~N(μ t ,s t ) In the formula, W μ b μ W σ b σ For trainable parameters, Softplus(·) is the activation function type, and y t This represents the final predicted deformation value of the deformation prediction model.

4. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points, as described in claim 1. Its features are: Step 5 includes the following sub-steps: Step 5.1 describes the inference task of the model trained throughout the entire process as follows: Given a multi-point deformation hysteresis sequence. With multi-point environmental factor sequences Predicting future deformation sequences at multiple measurement points T0 represents the historical lag length, and T1 represents the prediction length; environmental factors and effect sizes are combined into a matrix form X. t =concat(y t-1 ,x t ), which serves as the initial input feature X for the corresponding time step spatiotemporal fusion unit; The model is trained using a training set throughout the entire process, with the output and label sequence being processed simultaneously. They are used together to calculate the loss function and trained using the Adam optimizer; Step 5.2 involves substituting the validation set into the model for sample inference in each training round, and setting an early stopping mechanism to obtain the final model that optimizes the validation set metrics. During inference, the feature matrix... Substituting these parameters into the model sequentially, we obtain the Gaussian distribution parameters at time T0+1, and then sample from them to obtain... Then, this predicted value is treated as the known data and iterated into the input feature matrix at time T0+2. Similarly, the inferred sequence is finally obtained after T1-1 iterations. Step 5.3: After obtaining the optimal prediction model, rolling prediction is performed based on the last T0 data points of the real-time deformation monitoring sequence. The prediction results of the previous round are added to the deformation sequence, and the data of the last T0 length are treated as known values ​​to participate in the next prediction. Finally, a real-time deformation prediction sequence that meets the prediction length is obtained.

5. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points as described in claim 4, characterized in that: in, In step 5.1, the formula for calculating the probabilistic prediction loss function is as follows: In the formula, the subscripts s, n and t represent the sample number, measurement point number and time step, respectively, p(X) is the sample distribution and S is the total number of training samples.

6. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points as described in claim 1, characterized in that: in, Step 3 includes the following sub-steps: Step 3.1: Based on the characteristics of the deformation monitoring data of the rockfill dam and combined with specific engineering experience, determine the prior information that needs to be injected under the complex correlation of multiple measuring points; Step 3.2: Determine the calculation method of the adjacency matrix based on prior information.

7. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points as described in claim 1, characterized in that: in, In step 3.2, prior information under complex relationships needs to be considered, including: spatial distance between measurement points, arrangement of measurement points, partition where measurement points are located, and similarity of measurement point sequences; The following are some methods for calculating the adjacency matrix to address the complex relationships described above: Spatial distance Gaussian similarity function, considering the spatial distance relationship between multiple measurement points: In the formula, dist(i,j) represents the distance between the two measuring points in Euclidean space, and σ is the statistical standard deviation of the distance; 0-1 step function, considering the arrangement of measurement points and the relationship of zoning attributes among multiple measurement points: In the formula, P i P j It represents a certain reference attribute between two measuring points; Sequence similarity function, taking into account the similarity of actual monitored sequences: In the formula, DTW(i,j) represents the similarity distance between two measuring points under dynamic time warping.

8. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points as described in claim 1, characterized in that: in, In step 4.2, long short-term memory units are used as tools for temporal dependency mining. Historical information is transmitted to the current moment through cell states and hidden memory in temporal fusion, and temporal dependency information is mined by combining the current feature input.

9. The deformation prediction method for rockfill dams considering the complex spatiotemporal correlation of multiple measuring points as described in claim 1, characterized in that: in, In step 2, the preprocessed multi-point deformation monitoring data from step 1 is considered as the following sequence: {y1, y2, ..., y t-1 ,y t }, y∈N, with a total length of t and a total number of measurement points of N; a window of fixed length T0+T1 is used to slide gradually over the multi-measurement point sequence, and the data in each window is regarded as a training sample, thereby obtaining a total number of training samples of t-T0-T1+1, in the following form: In the formula, T0 represents the lag length of the model input data, and T1 represents the prediction length of the model output data. All the above samples are divided into training set and validation set according to a certain ratio.

10. A deformation prediction system considering the complex spatiotemporal fusion of multiple measuring points in a rockfill dam, characterized in that, include: The preprocessing department performs preprocessing on the time series data of dam deformation and environmental factors under monitoring. The sample construction unit uses the preprocessed data as samples, constructs the dataset using a sliding window method, and divides the samples into training and validation sets proportionally. The prior information determination unit determines the prior information to be injected during the fusion of multiple measurement points under complex relationships and calculates the adjacency matrix. The prediction model construction department constructs a spatiotemporal fusion deformation prediction model according to the following steps 4.1 to 4.4; Step 4.1: Feature space fusion under complex relationships with multiple measurement points based on graph convolutional networks; In the dam monitoring system, each measuring point is treated as a node with a corresponding feature vector. During each graph convolution process, the current measuring point is considered the central measuring point, and neighboring measuring points connected to it transfer their features to the central measuring point. Multi-measuring point feature fusion is performed considering the complex relationships between multiple measuring points. In the formula, GCN (·) A represents a graph convolution method using adjacency matrices under various prior information. (·) Let X represent the adjacency matrix under each relation, and let X represent the input multi-point feature matrix. (·) X represents the fusion characteristics under each relation. fusion The final fusion feature considering m complex relationships; Step 4.2: Use Long Short-Term Memory (LSTM) units to perform temporal dependency mining; Step 4.3: Construct two types of spatiotemporal fusion units: discrete spatiotemporal fusion unit and coupled spatiotemporal fusion unit; In the discrete-time fusion unit, steps 4.1 and 4.2 are connected in series, and the input X of step 4.2 is... t The final fused feature X, which takes into account various complex relationships, is the output of step 4.

1. fusion ; In the coupled spatiotemporal fusion unit, steps 4.1 and 4.2 are coupled. The forget gate, update weights and output gate information in the long short-term memory unit are all calculated using graph convolution, replacing the trainable matrix W in the long short-term memory unit. Step 4.4: Assuming that the noise distribution has temporal differences, the spatiotemporal fusion unit outputs hidden layer memory H. t Gaussian distribution parameters of deformation are obtained through linear layers; The deformation prediction unit trains the deformation prediction model using the training set, obtains the optimal model through the validation set, and completes near real-time deformation prediction based on the latest monitoring data. The input display unit is used to allow users to input operation commands and to display the corresponding commands. The control unit is communicatively connected to the preprocessing unit, sample construction unit, prior information determination unit, prediction model construction unit, deformation prediction unit, and input display unit, and controls their operation.

Images