Anomaly detection method and system for hydro-turbine unit monitoring data based on graph neural network

Abstract

The present invention discloses a method and system for detecting anomalies in monitoring data of a turbine unit based on a graph neural network, belonging to the field of equipment detection technology. The method comprises the following steps: S1: collecting a historical monitoring data set of the turbine unit and preprocessing the historical monitoring data set to generate a monitoring data sample set; S2: using the monitoring data sample set, constructing and optimizing a turbine unit monitoring data anomaly detection model; S3: collecting the latest monitoring data of the turbine unit, and using the optimized turbine unit monitoring data anomaly detection model to perform monitoring data anomaly detection. The turbine unit monitoring data sample set of the present invention reflects the operating status of the turbine unit in real time. By conducting anomaly detection research on the turbine unit monitoring data and timely discovering abnormal data, the safety performance of the turbine unit can be effectively improved, the reliability of the turbine unit can be effectively enhanced, and the safe and stable operation of the turbine unit can be ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment detection, and in particular relates to a method and system for detecting anomalies in monitoring data of a hydro-turbine unit based on a graph neural network. Background Art

[0002] A hydraulic turbine is a power machine that converts the energy of water flow into rotational mechanical energy and is a key component of a hydropower station. In recent years, significant breakthroughs in hydraulic turbine structural design, process production, and materials science have led to the gradual development of hydraulic turbines towards greater sophistication and intelligence, while also placing increasing demands on their maintenance.

[0003] Hydro-turbine monitoring data is multivariate time series data, characterized by high dimensionality and massive volume. It reflects the operating status of the hydro-turbine in real time and exhibits complex spatiotemporal correlations. When a hydro-turbine anomaly occurs, regional anomaly data features often emerge. These regional anomaly features imply complex spatiotemporal relationships. Traditional machine learning or deep learning methods are poor at capturing temporal dependencies. This paper proposes a hydro-turbine anomaly detection method based on graph neural networks, combining the powerful spatial feature extraction capabilities of graph neural networks with the powerful temporal feature extraction capabilities of LSTM. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems of difficulty in extracting spatiotemporal features of multivariate time series monitoring data of turbine units and low accuracy of anomaly monitoring, and proposes a method and method for anomaly detection in turbine unit monitoring data based on graph neural network.

[0005] The technical solution of the present invention is: a method for detecting anomalies in monitoring data of a hydro-turbine unit based on a graph neural network comprises the following steps:

[0006] S1: Collect historical monitoring data sets of the turbine group, preprocess the historical monitoring data sets, and generate monitoring data sample sets;

[0007] S2: Using the monitoring data sample set, build and optimize the turbine unit monitoring data anomaly detection model;

[0008] S3: Collect the latest monitoring data of the turbine unit and use the optimized turbine unit monitoring data anomaly detection model to perform monitoring data anomaly detection.

[0009] The beneficial effects of the present invention are:

[0010] (1) The turbine monitoring data sample set of the present invention reflects the operating status of the turbine in real time. By conducting anomaly detection research on the turbine monitoring data and timely discovering abnormal data, the safety performance of the turbine can be effectively improved, the reliability of the turbine can be effectively enhanced, and the safe and stable operation of the turbine can be guaranteed.

[0011] (2) Using the embedding vector method and the attention mechanism, a model for detecting anomalies in turbine monitoring data was constructed, which clarified the implicit relationships in the data and solved the problem of insufficient feature extraction when traditional methods were used to process multivariate time series data.

[0012] (3) The anomaly detection model for the turbine unit monitoring data of the present invention designs a multi-port prediction model, which can simultaneously output the predicted values ​​of the next timestamp of all monitoring parameters, effectively improving the accuracy of anomaly detection.

[0013] Furthermore, step S1 includes the following sub-steps:

[0014] S11: collecting a historical monitoring data set of the hydro turbine group, removing monitoring data with a variance of 0 in the historical monitoring data set, and obtaining a monitoring data set;

[0015] S12: Perform Z-score normalization and sliding window processing on the monitoring data set in sequence to obtain a monitoring data sample set.

[0016] The beneficial effect of the above further solution is that step S1 converts the original data into sample data that conforms to deep learning, which is beneficial to deep learning network training and parameter optimization.

[0017] Furthermore, step S2 includes the following sub-steps:

[0018] S21: Generate a corresponding node embedding vector for each monitoring data in the monitoring data sample set;

[0019] S22: Calculate the correlation between the embedding vectors of each node, and use several nodes with correlations greater than a set correlation threshold as the graph adjacency matrix to obtain a turbine unit monitoring data anomaly detection model based on a graph neural network;

[0020] S23: Extracting spatial features of each node in the anomaly detection model of turbine monitoring data;

[0021] S24: extracting the temporal features of each node based on the spatial features of each node;

[0022] S25: converting the time features of each node into the output features of each node;

[0023] S26: Based on the output characteristics of each node, a loss function is constructed, and the loss function is used to optimize the anomaly detection model of the turbine unit monitoring data.

[0024] The beneficial effects of the above further scheme are: treating each monitoring data as a node, learning the association relationship between nodes (i.e., graph structure) through monitoring data, using graph neural attention and LSTM network to learn the spatiotemporal correlation relationship in the data, and finally outputting the predicted value of the next timestamp of each node, and using the loss function composed of residual size to optimize the turbine unit monitoring data anomaly detection model.

[0025] Furthermore, in step S21, the specific method for generating a corresponding node embedding vector for each monitoring data in the monitoring data sample set is: generating a corresponding feature vector based on each monitoring data; converting the feature vector corresponding to each monitoring data into a latent factor of the feature; and splicing each latent factor to obtain a node embedding vector.

[0026] The beneficial effect of this further solution is that this step randomly generates a node embedding vector for each piece of monitoring data, achieving the transition from digital text to vector representation. Gradient descent is then used to update the embedding vector, ensuring that it truly reflects the relationship between the actual parameters in the vector space.

[0027] Furthermore, in step S21, the feature vector l corresponding to each monitoring data i The expression is:

[0028]

[0029] Where x i Indicates the status of monitoring data.

[0030] The beneficial effect of the above further solution is that the monitoring data is represented by vectors, which facilitates the construction of the model in the subsequent steps.

[0031] Furthermore, in step S22, the expression of the hydro-turbine unit monitoring data anomaly detection model D is:

[0032] D=(V,E)=({v i |v i ∈{P∪Q}},{(v i ,w,v j )|w∈W,v i ∈V,v j ∈V})

[0033] In the formula, V represents the node set of the graph neural network, E represents the graph adjacency matrix set of the graph neural network, and v i represents the node embedding vector corresponding to node i, v jrepresents the node embedding vector corresponding to the neighbor node j, P represents the monitoring data sample set, Q represents the monitoring data sample set corresponding to the central node, w represents the graph adjacency matrix corresponding to the central node, and W represents the graph adjacency matrix formed by several nodes with the highest correlation.

[0034] The beneficial effect of the above further scheme is that by constructing a turbine unit monitoring data anomaly detection model based on graph neural network, the global and local information of the monitoring data can be learned to ensure that the prediction structure is not affected by data noise.

[0035] Furthermore, in step S23, the spatial features of each node The calculation formula is:

[0036]

[0037] Where RELU(·) represents the activation function, W1 represents the first parameter of the trainable adjacency matrix, W2 represents the second parameter of the trainable adjacency matrix, and α i,i represents the attention coefficient of node i, α i,j represents the attention coefficient between node i and its neighbor node j, represents the characteristics of node i, Represents the characteristics of neighbor node j.

[0038] Furthermore, the attention coefficient α i,j The calculation formula is:

[0039]

[0040]

[0041]

[0042] Where, For splicing operations, is the embedding v of node i i and mapped The concatenated features, W3∈R 4d ×4d is a trainable matrix, and θ(i,j) is the attention coefficient after activation using LeakyReLU. The beneficial effect of the above further scheme is that the spatial feature function aggregates the relevant node information and realizes the feature extraction of the spatial dimension of the turbine group.

[0043] Furthermore, in step S24, the time characteristics of each node The calculation formula is:

[0044]

[0045] In the formula, RELU(·) represents the activation function, LSTM(·) represents the full connection function, Represents spatial features.

[0046] The beneficial effect of the above further solution is that this step uses shared weight LSTM to extract the time dimension features in the data based on the spatial features in the previous step, thereby realizing spatiotemporal feature extraction.

[0047] Furthermore, in step S25, the output features of each node The calculation formula is:

[0048]

[0049] Where, represents the time feature, and W4 represents the weight of the fully connected layer.

[0050] The beneficial effect of the above further solution is that this step uses a fully connected neural network to reduce the high-dimensional node vector to a single point output, so that the deep learning network meets the dimension of the predicted target value.

[0051] Furthermore, in step S26, the expression of the loss function Loss is:

[0052]

[0053] Where I represents the total number of nodes, represents the output feature of node i, represents the output feature of neighbor node j, b i represents the potential factor of node i, b j represents the potential factor of neighbor node j, l i represents the feature vector of node i, l j represents the feature vector of neighbor node j, λ represents the penalty factor, and ||·||2 represents the L2 norm.

[0054] The beneficial effect of the above further scheme is: by optimizing the anomaly detection model of turbine unit monitoring data based on graph neural network, the predicted value of the next timestamp of all monitoring data is output, which effectively improves the accuracy of anomaly detection.

[0055] The turbine unit monitoring data anomaly detection system includes:

[0056] Data preprocessing unit: used to obtain historical monitoring data and preprocess it to construct a monitoring data sample set;

[0057] Anomaly detection model construction and optimization unit: used to construct and optimize the anomaly detection model based on the monitoring data sample set;

[0058] Anomaly detection unit: used to perform anomaly detection on monitoring data using the optimized anomaly detection model.

[0059] The beneficial effects of the present invention are as follows: the present invention conducts anomaly detection research on the monitoring data of the turbine unit through the anomaly detection model, and timely discovers abnormal data, which can effectively improve the safety performance of the turbine unit, effectively improve the reliability of the turbine unit, and ensure the safe and stable operation of the turbine unit.

[0060] Furthermore, the anomaly detection model includes a graph structure learning module, a spatiotemporal feature extraction module, and an anomaly determination module connected in sequence;

[0061] The graph structure learning module is used to learn the monitoring parameters in the monitoring data sample set and learn to generate a graph structure that represents the correlation between the nodes of each monitoring parameter;

[0062] The spatiotemporal feature extraction module is used to extract the spatial and temporal features of each node in the graph structure;

[0063] The anomaly determination module is used to perform anomaly detection based on the extracted spatial and temporal features.

[0064] The beneficial effect of the above further solution is that the present invention clarifies the implicit relationship in the data by constructing an anomaly detection model for detecting turbine unit monitoring data, and solves the problem of insufficient feature extraction when traditional methods process multivariate time series data.

[0065] Furthermore, the spatial features of each node extracted by the spatiotemporal feature extraction module The expression is:

[0066]

[0067] Where RELU(·) represents the activation function, W1 represents the first parameter of the trainable adjacency matrix, W2 represents the second parameter of the trainable adjacency matrix, and α i,i represents the attention coefficient of node i, α i,j represents the attention coefficient between node i and its neighbor node j, represents the characteristics of node i, Represents the characteristics of neighbor node j.

[0068] The beneficial effect of the above further solution is that the above spatial feature function aggregates relevant node information and realizes the extraction of spatial dimension feature of the turbine unit.

[0069] Furthermore, the abnormality determination module performs abnormality detection according to the following determination rules:

[0070] if l i >r i :oi =1

[0071] else o i =0

[0072] Where, l i is the feature vector of spatial and temporal features of node i extracted by the spatiotemporal feature extraction module, r i is the abnormality judgment vector of the turbine unit, o i is the input monitoring data corresponding to node i, when o i =1, the input monitoring data corresponding to node i is abnormal. i = 0, the input monitoring data corresponding to node i is normal; i ∈r∈R N , r is the constructed abnormality judgment vector set.

[0073] The beneficial effect of the above further solution is that, based on the pre-established hydro-turbine unit abnormality determination vector, a rapid determination of abnormal monitoring data is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 The flowchart of the method for detecting anomaly in monitoring data of a hydro turbine unit is shown in FIG.

[0075] Figure 2 Schematic diagram of anomaly detection model. DETAILED DESCRIPTION

[0076] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0077] Before describing the specific embodiments of the present invention, in order to make the solutions of the present invention clearer and more complete, the abbreviations and key terms used in the present invention are first defined as follows:

[0078] Z-score normalization: Z-score normalization is a common method for data processing. It converts data of varying magnitudes into a uniform Z-score for comparison. This improves data comparability while reducing data interpretability.

[0079] Example 1:

[0080] The embodiment of the present invention provides a method for detecting anomalies in monitoring data of a hydro-turbine unit based on a graph neural network. Figure 1 As shown, the following steps S1-S3 are included:

[0081] S1: Collect historical monitoring data sets of the turbine group, preprocess the historical monitoring data sets, and generate monitoring data sample sets;

[0082] S2: Using the monitoring data sample set, build and optimize the turbine unit monitoring data anomaly detection model;

[0083] S3: Collect the latest monitoring data of the turbine unit and use the optimized turbine unit monitoring data anomaly detection model to perform monitoring data anomaly detection.

[0084] The hydro-turbine monitoring data sample set of the present invention reflects the operating status of the hydro-turbine unit in real time. By conducting anomaly detection research on the hydro-turbine monitoring data and timely discovering abnormal data, the safety performance of the hydro-turbine unit can be effectively improved, the reliability of the hydro-turbine unit can be effectively enhanced, and the safe and stable operation of the hydro-turbine unit can be ensured. The hydro-turbine monitoring data anomaly detection model is constructed using the embedding vector method and the attention mechanism, clarifying the implicit relationships in the data and solving the problem of insufficient feature extraction when traditional methods process multivariate time series data. Furthermore, the hydro-turbine monitoring data anomaly detection model of the present invention designs a multi-port prediction model that can simultaneously output the predicted values ​​of the next timestamp of all monitoring parameters, effectively improving the accuracy of anomaly detection.

[0085] Example 2:

[0086] Step S1 in Example 1 includes the following sub-steps S11-S12:

[0087] S11: collecting a historical monitoring data set of the hydro turbine group, removing monitoring data with a variance of 0 from the historical monitoring data set, and obtaining a monitoring data set;

[0088] S12: Perform Z-score normalization and sliding window processing on the monitoring data set in sequence to obtain a monitoring data sample set.

[0089] Step S1 of the embodiment of the present invention converts the original data into sample data that conforms to deep learning, which is beneficial to deep learning network training and parameter optimization.

[0090] In the embodiments of the present invention, turbine monitoring data refers to data collected by the turbine's own sensors during operation, such as speed, vibration, voltage, current, and pressure. Turbine monitoring data is time series data, characterized by high dimensionality, large sample sizes, and multimodality. Traditional machine learning or deep learning methods are poor at extracting its internal spatiotemporal correlation features, and the intermediate computational processes are often unexplainable.

[0091] After removing the monitoring data with a variance of 0 in the historical monitoring data set, the timestamp data containing missing values ​​in the data set can also be removed; then the data is Z-score standardized to obtain a clean data set {s (1) ,s (2) ,…,s (T)},s (t) ∈RN ,0≤t≤T, where s (t) is the data collected at time stamp t, N is the remaining monitoring system sensor data; a window with a window width of w is used to perform sliding window processing on each data set to generate samples. The sample input is x (t) =[s (t-w) ,s (t -w+1) ,…,s (t-1) ],x (t) ∈R w×N , the sample output target value is y (t) =s (t) .

[0092] Example 3:

[0093] Step S2 in Example 1 includes the following sub-steps S21-S26:

[0094] S21: Generate a corresponding node embedding vector for each monitoring data in the monitoring data sample set;

[0095] S22: Calculate the correlation between the embedding vectors of each node, and use several nodes with correlations greater than a set correlation threshold as the graph adjacency matrix to obtain a turbine unit monitoring data anomaly detection model based on a graph neural network;

[0096] In this embodiment of the present invention, the node embedding vector is used to calculate the correlation between nodes. The calculation formula is:

[0097]

[0098] The TopK operation is used to select the top K nodes with the highest correlation with the central node as the graph adjacency matrix. The graph adjacency matrix describes the spatial relationship between nodes, that is, A ji =1{j∈TopK(e ki :k∈{1,2,...N})}. This step uses cosine correlation to calculate the relationship between each embedded vector in the vector space, and uses the TopK method to construct a graph structure, clarifying the correlation between parameters. This achieves the transition from Euclidean data to non-Euclidean data, allowing the graph neural network to be implemented on hydroturbine unit data.

[0099] S23: Extracting spatial features of each node in the anomaly detection model of turbine monitoring data;

[0100] S24: extracting the temporal features of each node based on the spatial features of each node;

[0101] S25: converting the time features of each node into the output features of each node;

[0102] S26: Based on the output characteristics of each node, a loss function is constructed, and the loss function is used to optimize the anomaly detection model of the turbine unit monitoring data.

[0103] The embodiment of the present invention regards each monitoring data as a node, learns the association relationship between nodes (i.e., graph structure) through the monitoring data, uses graph neural attention and LSTM network to learn the spatiotemporal correlation relationship in the data, and finally outputs the predicted value of the next timestamp of each node. The loss function composed of residual size is used to optimize the anomaly detection model of turbine group monitoring data.

[0104] Example 4:

[0105] For step S21 in Example 3, the specific method for generating a corresponding node embedding vector for each monitoring data in the monitoring data sample set is: generating a corresponding feature vector based on each monitoring data; converting the feature vector corresponding to each monitoring data into a latent factor of the feature; and splicing each latent factor to obtain a node embedding vector.

[0106] This step in the embodiment of the present invention randomly generates a node embedding vector for each piece of monitoring data, achieving the transition from digital text to vector representation. Gradient descent is then used to update the embedding vector, ensuring that it truly reflects the relationship between actual parameters in the vector space. The embodiment of the present invention can use the Embedding method to randomly generate a node embedding vector for each piece of monitoring data.

[0107] Example 5:

[0108] For step S21 in embodiment 3, the feature vector l corresponding to each monitoring data i The expression is:

[0109]

[0110] Where x i Indicates the status of monitoring data.

[0111] The embodiment of the present invention represents the monitoring data using vectors to facilitate model building in subsequent steps.

[0112] Example 6:

[0113] With respect to step S22 in Example 3, the expression of the hydro-turbine unit monitoring data abnormality detection model D is:

[0114] D=(V,E)=({v i |v i ∈{P∪Q}},{(v i ,w,v j )|w∈W,v i∈V,v j ∈V})

[0115] In the formula, V represents the node set of the graph neural network, E represents the graph adjacency matrix set of the graph neural network, and v i represents the node embedding vector corresponding to node i, v j represents the node embedding vector corresponding to the neighbor node j, P represents the monitoring data sample set, Q represents the monitoring data sample set corresponding to the central node, w represents the graph adjacency matrix corresponding to the central node, and W represents the graph adjacency matrix formed by several nodes with the highest correlation.

[0116] By constructing a hydro-turbine unit monitoring data anomaly detection model based on a graph neural network, the embodiment of the present invention can learn the global and local information of the monitoring data, ensuring that the prediction structure is not affected by data noise.

[0117] Example 7:

[0118] Regarding step S23 in embodiment 3, the spatial features of each node The calculation formula is:

[0119]

[0120] Where RELU(·) represents the activation function, W1 represents the first parameter of the trainable adjacency matrix, W2 represents the second parameter of the trainable adjacency matrix, and α i,i represents the attention coefficient of node i, α i,j represents the attention coefficient between node i and its neighbor node j, represents the characteristics of node i, Represents the characteristics of neighbor node j.

[0121] The spatial feature function of the embodiment of the present invention aggregates relevant node information, thereby realizing the extraction of spatial dimension features of the turbine group.

[0122] In the embodiment of the present invention, α i,j The calculation formula is:

[0123]

[0124]

[0125]

[0126] Where, For splicing operations, is the embedding v of node i i and mapped The concatenated features, W3∈R 4d ×4dis a trainable matrix, and θ(i,j) is the attention coefficient after activation with LeakyReLU. This step calculates the real-time attention coefficient of each edge in the graph network using the node embedding vector and the node input data.

[0127] Example 8:

[0128] Regarding step S24 in embodiment 3, the time characteristics of each node The calculation formula is:

[0129]

[0130] In the formula, RELU(·) represents the activation function, LSTM(·) represents the full connection function, Represents spatial features.

[0131] This step of the embodiment of the present invention uses shared weight LSTM to extract time dimension features in the data based on the spatial features in the previous step, thereby realizing spatiotemporal feature extraction.

[0132] Example 9:

[0133] Regarding step S25 in embodiment 3, the output features of each node The calculation formula is:

[0134]

[0135] Where, represents the time feature, and W4 represents the weight of the fully connected layer.

[0136] This step of the embodiment of the present invention uses a fully connected neural network to reduce high-dimensional node vectors to single-point outputs, so that the deep learning network meets the dimension of the predicted target value.

[0137] Example 10:

[0138] For step S26 in Example 3, the expression of the loss function Loss is:

[0139]

[0140] Where I represents the total number of nodes, represents the output feature of node i, represents the output feature of neighbor node j, b i represents the potential factor of node i, b j represents the potential factor of neighbor node j, l i represents the feature vector of node i, l j represents the feature vector of neighbor node j, λ represents the penalty factor, and ||·||2 represents the L2 norm.

[0141] The embodiment of the present invention optimizes the anomaly detection model of turbine group monitoring data based on graph neural network, outputs the predicted value of the next timestamp of all monitoring data, and effectively improves the accuracy of anomaly detection.

[0142] In this embodiment of the present invention, historical monitoring data is extracted from the turbine data recording computer and divided into two categories: normal data set and abnormal data set according to the maintenance records in the data recording computer. The normal data set is divided into a training set, a validation set, and a normal test set according to the ratio of 8:2:3, and the normal test set and the abnormal data set are combined into a test set. The model is trained using the training set and validation set, and the model performance is evaluated using the test set. Expert knowledge is used to establish the turbine abnormality judgment vector r∈R N , this vector describes the abnormal deviation value allowed for each node, and the abnormality judgment rule is:

[0143] if l i >r i :o i =1

[0144] else o i =0

[0145] Where o∈R N , records the abnormal nodes in the input vector, such as o i =1, indicating that node i in the input data is abnormal, o i =0 means node i is normal.

[0146] Example 11:

[0147] With respect to the hydro-turbine unit monitoring data anomaly detection method in Examples 1 to 10, this embodiment provides a hydro-turbine unit monitoring data anomaly detection system that implements the method, including:

[0148] Data preprocessing unit: used to obtain historical monitoring data and preprocess it to construct a monitoring data sample set;

[0149] Anomaly detection model construction and optimization unit: used to construct and optimize the anomaly detection model based on the monitoring data sample set;

[0150] Anomaly detection unit: used to perform anomaly detection on monitoring data using the optimized anomaly detection model.

[0151] like Figure 2 As shown, the anomaly detection model in this embodiment includes a graph structure learning module, a spatiotemporal feature extraction module, and an anomaly determination module connected in sequence;

[0152] The graph structure learning module is used to learn the monitoring parameters in the monitoring data sample set and learn to generate a graph structure that represents the correlation between the nodes of each monitoring parameter;

[0153] The spatiotemporal feature extraction module is used to extract the spatial and temporal features of each node in the graph structure;

[0154] The anomaly determination module is used to perform anomaly detection based on the extracted spatial and temporal features.

[0155] In this embodiment, the spatiotemporal feature extraction module is specifically an LSTM model, which extracts the spatial features of each node. The expression is:

[0156]

[0157] Where RELU(·) represents the activation function, W1 represents the first parameter of the trainable adjacency matrix, W2 represents the second parameter of the trainable adjacency matrix, and α i,i represents the attention coefficient of node i, α i,j represents the attention coefficient between node i and its neighbor node j, represents the characteristics of node i, Represents the characteristics of neighbor node j.

[0158] The temporal features of each node extracted by the spatiotemporal feature extraction module for:

[0159]

[0160] In the formula, RELU(·) represents the activation function, LSTM(·) represents the full connection function, Represents spatial features.

[0161] In this embodiment, the time dimension features in the data are extracted using the shared weight LSTM, thereby realizing spatiotemporal feature extraction.

[0162] In this embodiment, the abnormality determination module performs abnormality detection according to the following rules:

[0163] if l i >r i :o i =1

[0164] else o i =0

[0165] Where, l i is the feature vector of spatial and temporal features of node i extracted by the spatiotemporal feature extraction module, r i is the abnormality judgment vector of the turbine unit, oi is the input monitoring data corresponding to node i, when o i =1, the input monitoring data corresponding to node i is abnormal. i = 0, the input monitoring data corresponding to node i is normal; i ∈r∈R N , r is the constructed abnormality judgment vector set. Among them, r∈R N The abnormal deviation value allowed for each node is described. Specific embodiments are used in this invention to illustrate the principles and implementation methods of the invention. The description of the above embodiments is only used to help understand the method and core concept of the invention. At the same time, for those skilled in the art, according to the concept of the invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the invention.

[0166] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in monitoring data of a hydro-turbine unit based on a graph neural network, characterized in that: The following steps are involved: S1: Collect historical monitoring data sets of the turbine group, preprocess the historical monitoring data sets, and generate monitoring data sample sets; S2: Using the monitoring data sample set, build and optimize the turbine unit monitoring data anomaly detection model; The following sub-steps are included: S21: Generate a corresponding node embedding vector for each monitoring data in the monitoring data sample set; S22: Calculate the correlation between the embedding vectors of each node, and use several nodes with correlations greater than a set correlation threshold as the graph adjacency matrix to obtain a turbine unit monitoring data anomaly detection model based on a graph neural network; Anomaly detection model for hydro turbine monitoring data D The expression is: Where, V represents the set of nodes of the graph neural network, E The set of graph adjacency matrices representing graph neural networks, v i Representation node i The corresponding node embedding vector, v j Represents neighbor nodes j The corresponding node embedding vector, P represents the monitoring data sample set, Q represents the monitoring data sample set corresponding to the central node, w represents the graph adjacency matrix corresponding to the central node, W Represents the graph adjacency matrix formed by several nodes with the highest correlation; S23: Extracting spatial features of each node in the anomaly detection model of turbine monitoring data; Spatial characteristics of each node The calculation formula is: Where RELU(·) represents the activation function, W 1 represents the first parameter of the trainable adjacency matrix, W 2 represents the second parameter of the trainable adjacency matrix, α i,i Representation node i The attention coefficient, α i,j Representation node i With neighboring nodes j The attention coefficient between Representation node i Features, Represents neighbor nodes j characteristics; S24: extracting the temporal features of each node based on the spatial features of each node; S25: converting the time features of each node into the output features of each node; S26: constructing a loss function based on the output characteristics of each node, and using the loss function to optimize the anomaly detection model for the turbine unit monitoring data; Loss Function Loss The expression is: Where, I Indicates the total number of nodes, Representation node i The output features of Represents neighbor nodes j The output features of b i Representation node i The potential factors, b j Represents neighbor nodes j The potential factors, l i Representation node i The eigenvector of l j Represents neighbor nodes j The eigenvector of λ represents the penalty factor, ||·||2 represents the L2 norm; S3: Collect the latest monitoring data of the turbine unit and use the optimized turbine unit monitoring data anomaly detection model to perform monitoring data anomaly detection.

2. The method for detecting anomalies in monitoring data of a hydro-turbine unit based on graph neural network according to claim 1 is characterized in that: The step S1 includes the following sub-steps: S11: collecting a historical monitoring data set of the hydro turbine group, removing monitoring data with a variance of 0 from the historical monitoring data set, and obtaining a monitoring data set; S12: Perform Z-score normalization and sliding window processing on the monitoring data set in sequence to obtain a monitoring data sample set.

3. The method for detecting anomalies in monitoring data of a hydro-turbine unit based on graph neural network according to claim 1 is characterized in that: In step S21, the specific method for generating a corresponding node embedding vector for each monitoring data in the monitoring data sample set is: generating a corresponding feature vector according to each monitoring data; converting the feature vector corresponding to each monitoring data into a latent factor of the feature; and splicing each latent factor to obtain a node embedding vector.

4. The method for detecting anomalies in monitoring data of a hydro-turbine unit based on graph neural network according to claim 1, characterized in that: In step S21, the feature vector corresponding to each monitoring data l i The expression is: Where, x i Indicates the status of monitoring data.

5. The method for monitoring abnormality of hydro-turbine unit monitoring data based on graph neural network according to claim 1 is characterized in that: The attention coefficient α i,j The calculation formula is: Where, For splicing operations, For nodes i Embed v i and mapped The features after splicing, W 3∈ R 4d×4d is a trainable matrix, θ ( i , j ) is the use of Attention coefficient after activation.

6. The method for detecting anomalies in monitoring data of a hydro-turbine unit based on graph neural network according to claim 1, characterized in that: In step S24, the time characteristics of each node The calculation formula is: In the formula, RELU(·) represents the activation function, LSTM(·) represents the full connection function, Represents spatial features.

7. The method for detecting anomalies in monitoring data of a hydro-turbine unit based on graph neural network according to claim 1, characterized in that: In step S25, the output features of each node The calculation formula is: Where, Represents time characteristics, W 4 represents the weight of the fully connected layer.

8. A system for detecting anomalies in monitoring data of a hydraulic turbine unit based on the method for detecting anomalies in monitoring data of a hydraulic turbine unit based on a graph neural network according to any one of claims 1 to 7, characterized in that: include: Data preprocessing unit: used to obtain historical monitoring data and preprocess it to construct a monitoring data sample set; Anomaly detection model construction and optimization unit: used to construct and optimize the anomaly detection model based on the monitoring data sample set; Anomaly detection unit: used to perform anomaly detection on monitoring data using the optimized anomaly detection model.

9. The hydro turbine unit monitoring data anomaly detection system according to claim 8, characterized in that: The anomaly detection model includes a graph structure learning module, a spatiotemporal feature extraction module, and an anomaly determination module connected in sequence; The graph structure learning module is used to learn the monitoring parameters in the monitoring data sample set and learn to generate a graph structure that represents the correlation between the nodes of each monitoring parameter; The spatiotemporal feature extraction module is used to extract the spatial and temporal features of each node in the graph structure; The anomaly determination module is used to perform anomaly detection based on the extracted spatial and temporal features.

10. The hydro turbine unit monitoring data anomaly detection system according to claim 9, characterized in that: The abnormality determination module performs abnormality detection according to the following rules: Where, Nodes extracted by the spatiotemporal feature extraction module i The feature vectors of spatial and temporal features, is the abnormality determination vector of the turbine unit, For the node i The corresponding input monitoring data, when When i The corresponding input monitoring data is abnormal. When i The corresponding input monitoring data is normal; among them, , r is the constructed abnormality judgment vector set.

Images