Partition monitoring method and model for key parts of concrete dam operation

Abstract

The present application relates to the technical field of concrete dam structure safety monitoring, in particular to a concrete dam operation key position partition monitoring method, model and storage medium. The concrete dam operation key position partition monitoring method provided by the present application first divides the key positions of the concrete dam by using the extracted monitoring data time-frequency vector, and on this basis, obtains the time sequence measurement data of different types of monitoring instruments with high space-time correlation, thereby establishing a graph structure, capturing the time dimension and variable dimension dependency relationship of the multivariate time sequence data, providing a graph attention network to further learn and represent this relationship, obtaining the final feature representation of the time sequence measurement data, and finally calculating an anomaly score by using the final feature representation to detect anomalies; the complementary mutual verification of multiple types of monitoring instruments and multiple measuring points is realized, and the structural integrity and spatial distribution law of the concrete dam are fully reflected.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for concrete dam structures, and in particular to a method and model for zoned monitoring of key operational components of a concrete dam. Background Technology

[0002] The deployment of safety monitoring instruments for concrete dam structures is based on the principle of balancing structural safety and engineering economy. According to the calculation results of the arch-beam load distribution method, and in accordance with the technical requirements of controlling key parts, paying attention to the spatiotemporal relationship, and using multiple types of monitoring instruments for mutual backup and verification in key parts, the horizontal arch ring is used as the arch-direction monitoring base surface and the vertical dam section is used as the beam-direction monitoring section. Various monitoring instruments are deployed on the monitoring base surface and the monitoring section to form a spatial grid system for monitoring the arch-beam of concrete dams.

[0003] By setting up measuring points in the monitoring spatial grid of the arch beam, monitoring data reflecting the operating status of the concrete dam can be collected. Under the influence of external factors, there is a certain correlation between the same or multiple types of monitoring data in different parts. This correlation is mainly reflected in the similarity of the time series trend of monitoring data in the same or similar parts.

[0004] Because the impact and even damage of internal and external loads on concrete dam structures are random, there is a lack of methods to dynamically determine key components through structural monitoring data analysis. Currently, monitoring models for specific monitoring components generally use multi-point correlation models constructed from monitoring data of similar instruments with high correlation to conduct structural component diagnostic analysis. However, there is a lack of research on establishing zonal monitoring models using monitoring data from multiple types of instruments, failing to meet the technical requirement that structural monitoring effect quantities must be cross-verified and comprehensively analyzed through monitoring data from multiple types of instruments. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problem that in the prior art, it is difficult to dynamically divide the key parts of concrete dams under various loads, and at the same time, it is impossible to accurately diagnose the operational status of key parts by relying solely on the measurement data of a single type of monitoring instrument.

[0006] To address the aforementioned technical problems, this invention provides a method for zoned monitoring of key operational components of a concrete dam, comprising:

[0007] The key operational components of the concrete dam are divided into regions, and time-series measurement data of different types of monitoring instruments in a certain region are obtained.

[0008] Graph structures are constructed for the time-series measurement data in both time and variable dimensions to obtain time feature maps and variable feature maps.

[0009] The temporal feature map and the variable feature map are respectively input into the temporal graph attention network and the variable graph attention network to obtain the temporal attention matrix and the variable attention matrix;

[0010] The time-series measurement data, the time attention matrix, and the variable attention matrix are concatenated and input into a gated convolutional network to obtain the target features.

[0011] An anomaly score is calculated based on the target features. If the anomaly score exceeds a preset threshold, it is determined to be an operational anomaly.

[0012] Preferably, the zoning of key operational components of the concrete dam includes:

[0013] Based on the time-frequency vectors and spatial vectors of the measuring points at key operational parts of the concrete dam, a spatiotemporal data matrix of the time-frequency vectors of the measuring points is constructed.

[0014] Gaussian mixture clustering is applied to the spatiotemporal data matrix of the time-frequency vector of the measuring point, and the spatial information of the safety measuring point of the concrete dam is used as prior knowledge of the number of components to construct a partitioning model of key parts of the concrete dam under spatial constraints.

[0015] The parameters of the key operational component partitioning model of the concrete dam are iteratively optimized and solved using the expectation-maximization algorithm.

[0016] The key operational components of a concrete dam are divided using an iteratively optimized model.

[0017] Preferably, the process of obtaining the time-frequency vector of the measurement point includes:

[0018] Historical concrete dam structure monitoring data were decomposed using wavelet packet transform, and the wavelet packet coefficients of the Mth layer were calculated.

[0019] Extract the time-domain vector for each low-frequency coefficient in the M-th layer wavelet packet coefficients;

[0020] Calculate the wavelet energy spectrum of each wavelet packet coefficient in the Mth layer and extract the frequency domain vector;

[0021] The time-domain vectors corresponding to multiple low-frequency coefficients and the frequency-domain vectors corresponding to multiple wavelet packet coefficients are normalized respectively, and the time-frequency vector of the measurement point is calculated based on the normalized time-domain vector and frequency-domain vector.

[0022] Preferably, the step of establishing graph structures for the time-series measurement data in both time and variable dimensions to obtain time feature graphs and variable feature graphs includes:

[0023] An embedding vector is set for each variable in the time series measurement data;

[0024] Calculate the correlation between variables based on the embedding vectors corresponding to any two variables;

[0025] For any variable, in the spatial graph, connect its K neighboring variables that are most correlated with it with edges to obtain the variable feature graph;

[0026] For the time-series measurement data, an embedding vector and position code are set for each time point in the sliding time window;

[0027] Calculate the temporal correlation based on the embedding vectors corresponding to any two time points;

[0028] For any point in time, in the spatial graph, connect the data of its K neighboring points with the highest temporal correlation with the data with edges to obtain the temporal feature graph.

[0029] Preferably, the step of inputting the temporal feature map and the variable feature map into the temporal graph attention network and the variable graph attention network, respectively, to obtain the temporal attention matrix and the variable attention matrix includes:

[0030] The variable feature maps are respectively input into the multi-head attention module, intra-index attention module and inter-index attention module in the variable graph attention network to capture the variable dependencies between multivariate time measurement data, the correlation of all measurement points under the same type of monitoring instrument and the correlation of all measurement points under different types of monitoring instruments.

[0031] The outputs of the multi-head attention module, the intra-index attention module, and the inter-index attention module are concatenated to obtain the variable attention matrix;

[0032] The temporal feature maps are input into the temporal map attention network, and combined with positional encoding, the feature representation of each time point is updated by aggregating the data of neighboring time points using a multi-head attention module, thus obtaining the temporal attention matrix.

[0033] Preferably, calculating the anomaly score based on the target features includes:

[0034] The target features are input into the prediction module and the reconstruction module to obtain the predicted value and the reconstruction probability;

[0035] The anomaly score is calculated based on the predicted value and the reconstruction probability.

[0036] Preferably, the prediction module is a multilayer perceptron.

[0037] Preferably, the reconstruction module includes a discriminator and an autoencoder.

[0038] Preferably, the formula for calculating the anomaly score based on the predicted value and the reconstruction probability is as follows:

[0039]

[0040] in, For the predicted value, x i y is the measured value, y2 is the hyperparameter of the balanced prediction module and the reconstruction module, and p i Let be the reconstruction probability.

[0041] This invention also provides a zoned monitoring model for key operational components of concrete dams, including:

[0042] The data acquisition module is used to divide the key parts of the concrete dam operation into regions and acquire time-series measurement data of different types of monitoring instruments in a certain region.

[0043] The feature map construction module is used to build graph structures for the time series measurement data in the time and variable dimensions respectively, to obtain the time feature map and the variable feature map;

[0044] The attention mechanism module is used to input the temporal feature map and the variable feature map into the temporal graph attention network and the variable graph attention network, respectively, to obtain the temporal attention matrix and the variable attention matrix;

[0045] The target feature acquisition module is used to concatenate the time-series measurement data, the time attention matrix, and the variable attention matrix and input them into a gated convolutional network to obtain the target features;

[0046] An anomaly detection module is used to calculate an anomaly score based on the target features. If the anomaly score exceeds a preset threshold, it is determined to be an operational anomaly.

[0047] The technical solution of the present invention has the following advantages over the prior art:

[0048] The method for zoning monitoring of key components of a concrete dam as described in this invention, based on zoning of key components of the concrete dam, acquires time-series measurement data from different types of monitoring instruments with high spatiotemporal correlation, thereby establishing a graph structure to capture the time and variable dimension dependencies of multivariate time-series data. A graph attention network is then provided to further learn and represent these relationships, obtaining the final feature representation of the time-series measurement data. Finally, anomaly scores are calculated using the final feature representation to detect anomalies. This method achieves complementary and mutual verification among multiple types of monitoring instruments and multiple measurement points, fully reflecting the structural integrity and spatial distribution patterns of the concrete dam. Attached Figure Description

[0049] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0050] Figure 1 The flowchart illustrates the implementation of a method for zoned monitoring of key operational components of a concrete dam, as provided by this invention.

[0051] Figure 2 This is an F-GAT network architecture diagram;

[0052] Figure 3 This is a diagram of the Generative Adversarial Network framework;

[0053] Figure 4 A framework for a multi-index graphical attention network model of key components of a concrete dam;

[0054] Figure 5 A schematic diagram of horizontal radial displacement monitoring data from a vertical measuring point;

[0055] Figure 6 This is a schematic diagram of horizontal tangential displacement monitoring data at a typical vertical measuring point.

[0056] Figure 7 A schematic diagram of the monitoring data for the opening and closing degree of the horizontal seam using a seam gauge.

[0057] Figure 8 This is a schematic diagram of the abnormal detection results of horizontal tangential displacement monitoring data at a typical vertical measuring point.

[0058] Figure 9 This is a schematic diagram showing the abnormal detection results of the horizontal seam opening and closing degree monitoring data of the seam measuring instrument.

[0059] Figure 10 This is a schematic diagram of the similarity matrix of the monitoring model of the arched crown beam of the upper part of the dam. Detailed Implementation

[0060] The core of this invention is to provide a method and model for zoned monitoring of key parts of a concrete dam during operation, which realizes the complementary and mutual verification of multiple types of monitoring instruments and multiple measuring points, and fully reflects the structural integrity and spatial distribution law of the concrete dam.

[0061] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely 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 are within the scope of protection of the present invention.

[0062] In accordance with design principles and technical requirements, multiple types of monitoring instruments are deployed at key structural locations of concrete dams for the same monitoring effect, achieving mutual backup and verification. Based on this, this invention, by dividing the key operational components of concrete dams, establishes a zoned monitoring method for key components using multiple types of monitoring instruments and multiple measuring points for complementary and mutual verification, for online evaluation of the operational performance of concrete dams.

[0063] Please refer to Figure 1 , Figure 1 The flowchart illustrates the implementation of a method for zoned monitoring of key components in the operation of a concrete dam, as provided by this invention. The specific operation steps are as follows:

[0064] S101: Divide the key operating parts of the concrete dam into regions and obtain time-series measurement data of different types of monitoring instruments in a certain region;

[0065] Based on the spatiotemporal characteristic matrix of concrete dam monitoring data, the monitoring data of multiple types and multiple measuring points involved in key parts are usually multivariate time series data. Multivariate time series data consists of a set of univariate time series data. Each univariate time series data represents an indicator with unique attributes, and these indicators are interconnected through linear and nonlinear relationships.

[0066] S102: Establish graph structures for the time series measurement data in the time and variable dimensions respectively to obtain the time feature graph and the variable feature graph;

[0067] S103: Input the time feature map and the variable feature map into the time graph attention network and the variable graph attention network respectively to obtain the time attention matrix and the variable attention matrix;

[0068] S104: Concatenate the time-series measurement data, the time attention matrix, and the variable attention matrix, and input them into a gated convolutional network to obtain the target features;

[0069] S105: Calculate the anomaly score based on the target features. If the anomaly score exceeds a preset threshold, it is determined to be an operational anomaly.

[0070] Based on the above embodiments, this embodiment will provide a detailed description of step S101:

[0071] The specific methods for dividing key operational components of concrete dams into zones include:

[0072] Step a: Based on the time-frequency vectors and spatial vectors of the measuring points at key operational parts of the concrete dam, construct the spatiotemporal data matrix of the measuring point time-frequency vectors; the construction approach for the spatiotemporal feature matrix of the concrete dam measuring point values ​​is as follows:

[0073] (1) Time-frequency vector of measured value at a measuring point: The time-frequency vector of the monitoring data of a specific measuring point within a certain time period is extracted and represented by an n-dimensional vector:

[0074] x i =(x i1 ,x i2 ,…,x in )

[0075] In the formula, x i Let x be the time-frequency vector of the measured data at a certain point during a certain period. in Let n be a time-frequency vector of the measured value at this measurement point. Based on the time-frequency vector extracted in Chapter 3, n = 24.

[0076] The extraction of the time-frequency vector includes:

[0077] Historical concrete dam structure monitoring data were decomposed using wavelet packet transform, and the wavelet packet coefficients of the Mth layer were calculated.

[0078] Extract the time-domain vector for each low-frequency coefficient in the M-th layer wavelet packet coefficients;

[0079] Calculate the wavelet energy spectrum of each wavelet packet coefficient in the Mth layer and extract the frequency domain vector;

[0080] The time-domain vectors corresponding to multiple low-frequency coefficients and the frequency-domain vectors corresponding to multiple wavelet packet coefficients are normalized respectively, and the time-frequency vector is calculated based on the normalized time-domain vectors and frequency-domain vectors.

[0081] (2) Measurement point spatial vector: Each measurement point has spatial attributes, represented by a three-dimensional spatial vector s(i) = (n i ,e i ,h i ,d i In the formula, n i e i These represent the horizontal coordinates of the locations where the measuring points were set up; h i The vertical coordinate representing the location of the measuring point; d i This indicates the type of monitoring instrument to which the measuring point belongs.

[0082] (3) Spatiotemporal vector of the measuring point: The spatiotemporal vector of the measuring point is formed by combining the time-frequency vector and the spatial vector of the measuring point, and is expressed as follows:

[0083] x i =[x i1 x i2 ,...,x in |s(i)] i＝1,2,...,n

[0084] (4) Spatiotemporal matrix of monitoring data: used to represent the set of various types of measuring points deployed at a certain structural part of the concrete dam for monitoring items such as deformation, stress-strain, and seepage.

[0085]

[0086] In the formula, Rn represents the time-frequency vector spatiotemporal matrix of monitoring data from all measuring points in a certain structural part.

[0087] Step b: Gaussian mixture clustering is applied to the spatiotemporal data matrix of the time-frequency vector of the measuring points, and the spatial information of the concrete dam safety measuring points is used as prior knowledge of the number of components to construct a spatially constrained model for classifying key parts of the concrete dam operation. Based on the spatiotemporal correlation analysis of the measured values ​​of the concrete dam measuring points, there are significant differences in the similarity characteristics of measured values ​​for the same or different monitoring projects under various conditions such as identical spatial locations, symmetrical locations, and different elevations. However, overall, combined with the first law of geography, "everything is related to other things, but things that are close are more closely related," and relevant technical specifications for the structural parts of the concrete dam and the layout of the monitoring system, the characteristic similarity of measured values ​​at the same or close locations is higher. For the Gaussian mixture model, the closer the spatial locations of the measuring points, the greater the probability that the measured values ​​belong to the same Gaussian distribution component. Therefore, using the spatial information of the concrete dam safety measuring points as prior knowledge for the model for classifying key parts of the concrete dam operation means loading the product between the distance feature and the component weight into the latent feature vector.

[0088] Obtain the measurement dataset X = (x1, x2, ..., x) of the safety monitoring points of the concrete dam. N Spatial nearest measurement point data of the safety measurement point measurement dataset. Obtain the spatial distance q = (q1, q2, ..., q) between the safe measuring points. N ),in, n k e k h k These represent the three-dimensional spatial coordinates of the measuring points, k = 1...N; the spatial distance feature Q = (Q1, Q2, ..., QN) is calculated based on the spatial distance of the safe measuring points. N ),in, q max q min These represent the maximum and minimum values ​​of q, respectively; the component weights {w1, w2, ..., w...} are obtained based on the monitoring instrument type of the safety measurement point. C The latent feature vector Z = (Z1, Z2, ..., Zn) is calculated based on the component weights and the spatial distance features. n ), where Z i =(z i1 ,zi2 ,...z iC ), i = 1, ..., N, z ij The value of z depends on the spatial distance characteristics of the measuring points and the component weights. In one embodiment, z ij =w j *Q i C represents the number of components in the model for key operational parts of the concrete dam; calculation The joint probability density is: The log-likelihood function for all data is calculated as follows:

[0089] Step c: The parameters of the key operational component partitioning model of the concrete dam are iteratively optimized and solved using the expectation-maximization algorithm;

[0090] Initialize the number of model components C and the mean and covariance matrix of each component, and set the component weights. The parameters of the key component partitioning model of the concrete dam operation are calculated using the expectation-maximization algorithm; after the k-th iteration, the components with a weight of 0 in the key component partitioning model of the concrete dam operation are removed, and k = k + 1 is set; steps 2-4 are repeated until the model converges, and the parameters and number of components of the iteratively optimized key component partitioning model of the concrete dam operation are obtained.

[0091] The parameters for calculating the critical operational component partitioning model of the concrete dam using the expectation-maximization algorithm include:

[0092] Calculate x for each sample based on the current parameters. j The posterior distribution γ belonging to each Gaussian mixture component ji ;

[0093] Update the parameter {(α) based on the posterior distribution. i ,μ i ,∑ i |1≤i≤k}:

[0094]

[0095]

[0096]

[0097] Step d: Use the iteratively optimized critical component division model of the concrete dam to divide the critical components of the concrete dam into components.

[0098] Based on the above embodiments, this embodiment will provide a detailed description of step S102:

[0099] Anomalies in the operation of concrete dams manifest as changes in the monitoring data of various indicators over time. Therefore, for monitoring data within the same time window, two graph structures are used to explicitly model the dependencies between time and variable dimensions. The details are as follows:

[0100] (1) Variable Feature Map. Before constructing the variable feature map, a representation vector needs to be randomly initialized for each variable to reduce the decrease in model accuracy caused by different data types and value ranges. In the latent representation space composed of embedded vectors, vectors that are closer in distance are more similar, and the correlation between their corresponding variables (i.e., the correlation between different types of concrete dam safety monitoring instruments) is stronger. Graph attention networks require data with an explicit graph structure as input, so the N variables within the input time window are used as N nodes to construct the graph structure. First, the similarity relationship between nodes is calculated based on the initial representation vectors of the nodes, and then the largest number of node pairs are selected and connected with edges to obtain a sparse directed graph structure. The specific steps are as follows:

[0101] First, set an embedding vector for each variable. Where d represents the dimension of the vector, and different values ​​are set according to the actual situation. The initial value of the vector is given randomly, and the specific value is continuously adjusted through backpropagation during the training process of the model. The correlation of variables can be calculated from the embedding vector. For variables i and j, the correlation is calculated by (1):

[0102]

[0103] In the formula, f(v) i ,v j To construct a directed graph, an asymmetric similarity calculation method is needed, using cosine similarity. After calculating the similarity of all variables pairwise, for any variable i, select the K neighbor variables j with the highest similarity. Connect i and j with an edge in the spatial graph. The column of i in the spatial adjacency matrix A is represented by equation (2):

[0104]

[0105] The value of K can be given by the user to adjust the sparsity of the graph. With prior information about the graph structure, the adjacency matrix A can be directly given by the user. In certain special scenarios, the adjacency matrix can be set to a matrix with all values ​​of 1, thus constructing a fully connected graph.

[0106] (2) Temporal Feature Map. To explicitly model the temporal information between data points, a temporal feature map within the input time window is constructed using steps similar to those described for constructing the variable feature map. The temporal feature map differs from the variable feature map in its embedding vector settings and construction method. In the variable feature map, the embedding vector is used to reduce the impact of different variable data types and value ranges on model accuracy. However, in the temporal feature map, in addition to assuming a corresponding embedding vector for each time point within the sliding window, positional encoding is also needed to represent the temporal position differences of the data. The embedding vector for each node is... u i The initial value is also given randomly, and then the final representation is obtained through training.

[0107] For a time series within an input sliding window, a positional encoding is given for the vector at any timestamp j. Calculated using equation (3):

[0108]

[0109]

[0110] In the formula, d is the dimension of the position encoding, which needs to be the same as the dimension of the variable at the current position, i.e., d = N, and d also needs to be a multiple of 2. Unlike the variable feature graph above, where each variable is treated as a node, to construct an explicit time graph structure, the data at each time step needs to be treated as a node in the graph. The specific form of the position encoding is calculated using equation (.4):

[0111]

[0112] After assigning a location code to each time point, a time map is constructed by calculating the similarity between the embedding vectors of different time points. This is calculated using equation (5):

[0113]

[0114] The pairwise similarity between different time points represents the degree of similarity between different timestamps. When constructing the time graph, similar to the construction of the variable graph above, for any data point i at timestamp, select the K neighboring time points j with the highest similarity and connect them with edges. The constructed time adjacency matrix is ​​calculated using equation (6):

[0115]

[0116] In the formula, The value of K is also specified by the user, and it is generally the same as the sparsity of the spatial adjacency matrix.

[0117] Based on the above embodiments, this embodiment will provide a detailed description of step S103:

[0118] After constructing the graph structure data, a graph attention network is used to learn information from the time and variable dimensions. Through learning, each node in the graph structure receives a final representation vector containing information about the current node and its neighboring nodes. The graph attention layer consists of parallel variable graph attention networks and a time graph attention network, used to simultaneously capture dependencies across different dimensions of the data. Furthermore, by improving the attention mechanism in the variable graph attention layer, the correlation of metrics in the data is explicitly captured. Specifically, as shown below:

[0119] (1) Variable Graph Attention Network (F-GAT). The constructed variable feature map is used as the input of F-GAT, and the information in the map is further mined through the attention mechanism. First, assume that the feature representation of F-GAT in the l-th layer is H. l Initial input formula (7):

[0120]

[0121] In the formula, Given an input sequence of length ω at timestamp t, Let be the learnable transformation matrix of the input data, || be the concatenation operation, and V be the matrix composed of node representation vectors.

[0122] The architecture of F-GAT is as follows Figure 2 As shown, the system consists of three modules: multi-head attention, intra-index attention, and inter-index attention. The multi-head attention module primarily models the variable dependencies between multivariate time series, while intra-index and inter-index attention are used to capture the correlations between different time series. Specifically, intra-index correlation refers to the correlation among all measuring points under the same type of monitoring instrument. For example, in a concrete dam deformation monitoring project, which includes multiple types of monitoring instruments such as vertical lines, surface deformation observation, static leveling, and joint gauges, the scope of intra-index refers to all measuring points included by the vertical line monitoring instrument. Inter-index correlation refers to the correlation between different types of monitoring instruments. For example, in a concrete dam deformation monitoring project, which includes multiple types of monitoring instruments such as vertical lines, surface deformation observation, static leveling, and joint gauges, the scope of inter-index correlation includes vertical line measuring points, surface deformation observation points, static leveling points, and joint gauges.

[0123] The multi-head attention module updates the feature representation of each node by aggregating the neighbor node information of the target node, and calculates it using equation (8):

[0124]

[0125] In the formula, Let be the feature representation of node i at layer (l+1), || be the concatenation operation, and S be the number of attention heads. This represents the attention score of node i and node j in the s-th attention head of the l-th layer. It is the learnable weight matrix of the s-th attention head in the l-th layer. It is the feature representation of node j at layer l. It is the set of neighboring nodes of node i in the adjacency matrix A representing the feature map of the variables. The attention score is calculated using equations (8), (9), and (10):

[0126]

[0127]

[0128]

[0129] In the formula, a T Let be a learnable bias vector, || be the concatenation operation, and LeakyReLU be the non-linear activation function.

[0130] Traditional graph attention networks fail to consider the correlation between indicators in multivariate time series, thus losing important information from some variable dimensions. Neighbor nodes with different dependencies have different effects on the central node. This section improves the effectiveness of the model in modeling the dependencies between variables in the series by adding two relational attention modules: intra-indicator and inter-indicator attention. The adjacency matrices of the intra-indicator attention graph and the inter-indicator attention graph are defined by equations (12) and (13):

[0131]

[0132]

[0133] In the formula, and It is the candidate set, that is This indicates a node that shares the same index as node i. This indicates a node whose monitoring metric differs from that of node i. It's important to note that when... or When doing this, the TopK operation needs to be used to select the indices of the top K largest cosine similarities to construct the adjacency matrix.

[0134] Then, the multi-indicator correlations between different time series are explicitly captured through two relational attention modules. The features of the in-indicator attention modules are calculated using equations (14), (15), and (16):

[0135]

[0136]

[0137]

[0138] In the formula, It is the feature representation of node i at layer l+1. Let i be the set of its indexed neighbor nodes. Let i and j be the attention scores at layer l. and Let be the weight matrix of the l-th layer. Let be the bias vector of the l-th layer. Similarly, the feature representation of the attention module between metrics can be calculated. Let i be the set of neighboring nodes among the indices of node i. This is the final output of the variable graph attention layer. It is obtained by concatenating the inputs of the three attention modules and calculated using equations (17) and (18):

[0139]

[0140]

[0141] In the formula, This represents the final representation of node i at level l+1. This represents the weight matrix of the (l+1)th layer. This represents the bias vector of the (l+1)th layer, where || is the concatenation operation. By using the intermediate features of the (l+1)th layer and It was pieced together.

[0142] (2) Temporal Graph Attention Network (T-GAT). Assume Z... l For the feature representation of T-GAT at layer l, the initial input of T-GAT is in Let U be the learnable transformation matrix of the input data, and U be the matrix composed of node representation vectors. The temporal graph attention layer takes the constructed temporal graph structure as input, combines it with position encoding, and uses a multi-head attention module to aggregate the information of neighboring nodes to update the feature representation of each time point, calculated by equation (19):

[0143]

[0144] In the formula, Let be the feature representation of node i at layer (l+1), || be the concatenation operation, and S be the number of attention heads. It is the set of neighboring nodes of node i in the adjacency matrix A′ of the time graph mentioned above. These are the attention scores of node i and node s at the s-th attention head in layer l. It is the weight matrix of the s-th attention head in the l-th layer. It is the feature representation of node j at layer l. The calculation steps are the same as above. The calculation steps are similar, specifically obtained from equations (20), (21), and (22):

[0145]

[0146]

[0147]

[0148] In the formula, This represents the attention score of node i and node j in the l-th layer of the time graph, which is the attention score of the s-th attention head. It is the learnable weight matrix of the s-th attention head in the l-th layer of the time-graph attention module, a T Let be a learnable bias vector, || be the concatenation operation, and LeakyReLU be the non-linear activation function.

[0149] Based on the above embodiments, this embodiment will provide a detailed description of step S104:

[0150] The output of the variable graph attention network is an N×ω matrix, where each row represents the relationship between a node in the variable feature map and its neighboring nodes, captured by the graph attention network. Similarly, the output of the temporal graph attention network is an ω×N matrix. Concatenating the outputs of the two graph attention layers with the original time series data forms an ω×3N matrix, where each row represents a 3N-dimensional feature vector of a timestamp within the input time window. Finally, this ω×3N matrix is ​​used as input to a gated convolutional network (GRU). As a variant of recurrent convolutional networks, GRU effectively captures sequence pattern information in the data to obtain the target features.

[0151] Based on the above embodiments, this embodiment will provide a detailed description of step S105:

[0152] The target features are input into the prediction module and the reconstruction module to obtain the predicted value and the reconstruction probability. The anomaly score is then calculated based on the predicted value and the reconstruction probability, as follows:

[0153] To leverage the advantages of both reconstruction-based and prediction-based models, the loss function of MTS-GAT has two objectives: capturing the distribution of the entire input data in the reconstruction module and accurately predicting the value at the next time stamp in the prediction module. The inputs to the reconstruction and prediction modules at time t... This is the output of XχХGRU. The loss function for joint optimization is defined by equation (23):

[0154]

[0155] in, The loss function for the reconstruction module, Let γ be the loss function of the prediction module, and γ1 be the hyperparameter that balances the weights of the two modules.

[0156] Prediction module uses To predict the observation at the next timestamp, this section uses a multilayer perceptron (MLP) as the prediction module, and the loss function is defined by equation (24):

[0157]

[0158] In the formula, x i，t+1 It is the measured value of the i-th time series at t+1. It is the predicted value of the i-th sequence at time t+1.

[0159] The reconstruction module learns the reconstruction probabilities of the input data. To enhance the robustness of the model, two discriminators D are used. E (·) and D D (·) For the autoencoder G A Adversarial training is performed as a reconstruction-based model. G A encoder G E (·) and decoder G D (·) can then be considered as two generators. The model is as follows: Figure 3 As shown, for a given input This is the latent representation of the autoencoder, where p(z) represents the prior distribution of z, and q(z) is the posterior distribution generated by the autoencoder in the latent space, calculated by equation (4):

[0160]

[0161] In the formula, Represents the coding distribution. This represents the distribution of the input data. Adversarial Network D E The · operator is used to adjust the posterior distribution q(z) to satisfy the prior distribution p(z), i.e., to maximize the loss function. Calculate using equation (26):

[0162]

[0163] The corresponding generator G E (·) Mixed D E (·), that is, minimizing the loss function Calculate using equation (27):

[0164]

[0165] Similarly, adversarial networks D D (·) Overfitting can be avoided by increasing the difference between the input data and the reconstructed data, i.e., maximizing the loss function. Calculated using equation (28):

[0166]

[0167] The corresponding generator G D (·) then we need to minimize the loss function. Calculate using equation (29):

[0168]

[0169] Reconstructed representation of input data Finally, use and As an adversarial regularization to ensure the robustness of the reconstructed model, the loss function of the reconstructed model is defined by equations (30) and (31):

[0170]

[0171]

[0172] In the formula, This represents the reconstruction loss.

[0173] For the i-th univariate time series, at any timestamp t, the prediction module generates a predicted value. The refactoring module generates the refactoring probability p. i The final anomaly score for each timestamp is balanced by equation (32) to account for the weights of the two modules:

[0174]

[0175] In the formula, x i γ0 represents the measured value, and γ2 is a hyperparameter balancing the two modules. During the anomaly detection phase, when the anomaly score of a certain timestamp exceeds a given anomaly threshold, that timestamp is marked as an "abnormal timestamp"; otherwise, it is marked as a "normal timestamp." The anomaly threshold is selected using the peaks-over-threshold (POT) algorithm on the validation set.

[0176] Based on the above embodiments, Figure 4 This invention provides a multi-index graphical attention network model framework for key components of concrete dams, which may specifically include:

[0177] The data acquisition module is used to acquire time-series measurement data from different types of monitoring instruments in the monitoring of concrete dam structures;

[0178] The feature map construction module is used to build graph structures for the time series measurement data in the time and variable dimensions, respectively, to obtain time feature maps and variable feature maps;

[0179] The attention mechanism module is used to input the temporal feature map and the variable feature map into the temporal graph attention network and the variable graph attention network, respectively, to obtain the temporal attention matrix and the variable attention matrix;

[0180] The target feature acquisition module is used to concatenate the time-series measurement data, the time attention matrix, and the variable attention matrix and input them into a gated convolutional network to obtain the target features;

[0181] An anomaly detection module is used to calculate an anomaly score based on the target features. If the anomaly score exceeds a preset threshold, it is determined to be an operational anomaly.

[0182] The zoning monitoring model for key components of concrete dam operation in this embodiment is used to implement the aforementioned zoning monitoring method for key components of concrete dam operation. Therefore, the specific implementation of the zoning monitoring model for key components of concrete dam operation can be found in the previous embodiment section of the zoning monitoring method for key components of concrete dam operation. For example, the data acquisition module, feature map construction module, attention mechanism module, target feature acquisition module, and anomaly detection module are used to implement steps S101, S102, S103, S104, and S105 in the aforementioned zoning monitoring method for key components of concrete dam operation. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0183] This invention first extracts the time-frequency vector of concrete dam structural monitoring data and establishes a spatiotemporal matrix of the time-frequency vector. Then, combining the typical time-frequency vectors of the real-time identified structural monitoring data, a key component partitioning method based on the data's time-frequency vector is proposed to dynamically divide the key components of the concrete dam. Based on this, the upper arched beam area of ​​the dam body is selected, including the dam body from 1190m to 1245m elevation range between dam sections 19# and 25#. Four types of monitoring instruments are deployed in this key component area to monitor deformation effects: vertical lines, surface deformation observation points, static levels, and joint gauges. The vertical lines and surface deformation observation points include C4-A19-PL-01~02, C4-A22-PL-01~02, C4-A25-PL-01~02, C4-A19~A25-TP-01, and C4-A19~A25-TP-02, totaling 2. There are 0 measuring points used to monitor the horizontal radial and tangential displacement of this part; the static leveling includes 7 measuring points from 1245-SL4-03 to 09, used to monitor the vertical displacement of this part; the joint gauge includes 10 measuring points from C4-A19-J-33, C4-A25-J-31, and C4-A22-KZ-J-04 to 09, used to monitor the opening and closing degree of the transverse joints of this part, as well as the opening and closing degree of the cross joint area of ​​the seismic reinforcement on the upstream and downstream sides.

[0184] Based on the preprocessing of the original measured data, the data from the four types of monitoring instruments used in the experiment were first fitted (vertical line, radial displacement of surface deformation observation points, and tangential displacement were each considered as one type of data). The data time series spanned from January 1 to December 31, 2021, therefore a total of 365 timestamps were set. A typical monitoring data process line is shown below. Figures 5 to 7 As shown.

[0185] Through model training, it was found that the horizontal tangential displacement monitoring data of two measuring points C4-A22-PL-01 and 02 on the vertical line of section 22 of the dam, and the transverse joint opening and closing degree monitoring data of seven measuring points including C4-A22-KZ-J-05 of the joint measuring instrument, showed a total of 13 instances of abnormally high scores. Figure 8 and Figure 9As shown in the figure, the red vertical lines mark 13 abnormal timestamps, representing monitoring data from March 26, 2021, March 27, March 28, March 29, March 30, March 31, April 2, April 3, May 27, May 28, August 25, and August 26, 2021. After cross-checking with the on-site monitoring system's operation logs, it was found that the first eight abnormal timestamps (i.e., from March 26 to April 3) were due to a malfunction in the automated monitoring and acquisition module prior to March 2021, resulting in missing monitoring data. This caused the measured value changes to be identified and scored as abnormal, thus indicating anomalies. The subsequent five anomalous time stamps (May 27-28 and August 24-26), after ruling out data anomalies due to monitoring system malfunctions, were analyzed to be caused by two earthquakes in Yangbi County, Yunnan Province, with a maximum magnitude of 6.4 and a maximum magnitude of 3.6, occurring between May 21 and June 1, and between August 20 and August 30, 2021, respectively. Multiple aftershocks followed in the ten days that followed. These unusual conditions led to excessively high anomaly scores in the monitoring data related to the concrete dam, resulting in its identification as anomaly. Figure 8 and Figure 9 The images show two relatively obvious changes on May 27th and May 28th.

[0186] After identifying the time points of abnormal monitoring data, a similarity matrix was established using all monitoring data from the upper area of ​​the arched beam from May 27th to May 28th to further analyze the correlation between the abnormal monitoring points. For example... Figure 10 The horizontal and vertical axes represent measurement points, characterizing the pairwise correlation between them, expressed as similarity, with values ​​ranging from [0,1], where 0 represents complete dissimilarity and 1 represents complete similarity. The color intensity indicates the degree of similarity. It can be seen that the proportion of darker colored squares in the graph is relatively small, indicating that most deformation measurement points did not show abnormalities between May 27th and May 28th, thus confirming that the deformation behavior of this key component is normal.

[0187] A specific embodiment of the present invention also provides a zoned monitoring device for key operational components of a concrete dam, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the aforementioned zoned monitoring method for key operational components of a concrete dam.

[0188] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for zoning monitoring of key components of a concrete dam.

[0189] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0191] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0193] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for zoned monitoring of key operational components of a concrete dam, characterized in that, include: The key operational components of the concrete dam are divided into regions, and time-series measurement data of different types of monitoring instruments in a certain region are obtained. Graph structures are constructed for the time-series measurement data in both time and variable dimensions to obtain time feature maps and variable feature maps; The temporal feature map and the variable feature map are input into the temporal graph attention network and the variable graph attention network, respectively, to obtain the temporal attention matrix and the variable attention matrix, including: The variable feature maps are respectively input into the multi-head attention module, intra-index attention module and inter-index attention module in the variable graph attention network to capture the variable dependencies between multivariate time measurement data, the correlation of all measurement points under the same type of monitoring instrument and the correlation of all measurement points under different types of monitoring instruments. The outputs of the multi-head attention module, the intra-index attention module, and the inter-index attention module are concatenated to obtain the variable attention matrix; The time feature maps are respectively input into the time map attention network. Combined with position encoding, the multi-head attention module is used to aggregate the data of neighboring time points to update the feature representation of each time point, thereby obtaining the time attention matrix. The time-series measurement data, the time attention matrix, and the variable attention matrix are concatenated and input into a gated convolutional network to obtain the target features. An anomaly score is calculated based on the target features. If the anomaly score exceeds a preset threshold, it is determined to be an operational anomaly.

2. The method for zoned monitoring of key operational components of a concrete dam according to claim 1, characterized in that, The regional division of key operational components of concrete dams includes: Based on the time-frequency vectors and spatial vectors of the measuring points at key operational parts of the concrete dam, a spatiotemporal data matrix of the time-frequency vectors of the measuring points is constructed. Gaussian mixture clustering is applied to the spatiotemporal data matrix of the time-frequency vector of the measuring point, and the spatial information of the safety measuring point of the concrete dam is used as prior knowledge of the number of components to construct a partitioning model of key parts of the concrete dam under spatial constraints. The parameters of the key operational component partitioning model of the concrete dam are iteratively optimized and solved using the expectation-maximization algorithm. The key operational components of a concrete dam are divided using an iteratively optimized model for key operational components.

3. The method for zoned monitoring of key operational components of a concrete dam according to claim 2, characterized in that, The process of obtaining the time-frequency vector of the measurement point includes: Historical concrete dam structure monitoring data were decomposed using wavelet packet transform, and the wavelet packet coefficients of the Mth layer were calculated. Extract the time-domain vector for each low-frequency coefficient in the M-th layer wavelet packet coefficients; Calculate the wavelet energy spectrum of each wavelet packet coefficient in the Mth layer and extract the frequency domain vector; The time-domain vectors corresponding to multiple low-frequency coefficients and the frequency-domain vectors corresponding to multiple wavelet packet coefficients are normalized respectively, and the time-frequency vector of the measurement point is calculated based on the normalized time-domain vector and frequency-domain vector.

4. The method for zoned monitoring of key operational components of a concrete dam according to claim 1, characterized in that, The step of establishing graph structures for the time-series measurement data in both time and variable dimensions to obtain time feature graphs and variable feature graphs includes: An embedding vector is set for each variable in the time series measurement data; Calculate the correlation between variables based on the embedding vectors corresponding to any two variables; For any variable, in the space graph, select the variable with the highest correlation to the previous variable. K Connecting the neighbor variables with edges yields the variable feature map; For the time-series measurement data, an embedding vector and position code are set for each time point in the sliding time window; Calculate the temporal correlation based on the embedding vectors corresponding to any two time points; For any given point in time, in the spatial graph, select the data that has the highest temporal correlation with that point. K The data of each neighbor's time point are connected by edges to obtain the time feature map.

5. The method for zoned monitoring of key operational components of a concrete dam according to claim 1, characterized in that, The calculation of the anomaly score based on the target features includes: The target features are input into the prediction module and the reconstruction module to obtain the predicted value and the reconstruction probability; The anomaly score is calculated based on the predicted value and the reconstruction probability.

6. The method for zoned monitoring of key operational components of a concrete dam according to claim 5, characterized in that, The prediction module is a multilayer perceptron.

7. The method for zoned monitoring of key operational components of a concrete dam according to claim 5, characterized in that, The reconstruction module includes a discriminator and an autoencoder.

8. The method for zoned monitoring of key operational components of a concrete dam according to claim 5, characterized in that, The formula for calculating the anomaly score based on the predicted value and the reconstruction probability is as follows: in, For predicted values, These are measured values. To balance the hyperparameters of the prediction and reconstruction modules, Let be the reconstruction probability.

9. A zoned monitoring model for key operational components of a concrete dam, characterized in that, include: The data acquisition module is used to divide the key parts of the concrete dam operation into regions and acquire time-series measurement data of different types of monitoring instruments in a certain region. The feature map construction module is used to build graph structures for the time series measurement data in the time and variable dimensions, respectively, to obtain time feature maps and variable feature maps; The attention mechanism module is used to input the temporal feature map and the variable feature map into the temporal graph attention network and the variable graph attention network, respectively, to obtain the temporal attention matrix and the variable attention matrix; The target feature acquisition module is used to concatenate the time-series measurement data, the time attention matrix, and the variable attention matrix and input them into a gated convolutional network to obtain the target features; An anomaly detection module is used to calculate an anomaly score based on the target features. If the anomaly score exceeds a preset threshold, it is determined to be an operational anomaly. The attention mechanism module is also used to input the variable feature map into the multi-head attention module, intra-index attention module and inter-index attention module in the variable graph attention network respectively, to capture the variable dependency between multivariate time measurement data, the correlation of all measurement points under the same type of monitoring instrument and the correlation of all measurement points under different types of monitoring instruments; The outputs of the multi-head attention module, the intra-index attention module, and the inter-index attention module are concatenated to obtain the variable attention matrix; The temporal feature maps are input into the temporal map attention network, and combined with positional encoding, the feature representation of each time point is updated by aggregating the data of neighboring time points using a multi-head attention module, thus obtaining the temporal attention matrix.

Images