Knowledge graph and space-time diagram network fusion driven wind turbine generator state monitoring method

Abstract

The invention discloses a wind turbine generator state monitoring method driven by network fusion of a knowledge graph and a space-time diagram. The method comprises the following steps: S1, collecting time sequence SCADA data and wind power text data of a wind turbine generator; s2, constructing the wind power text data into a wind power operation and maintenance knowledge graph; s3, generating time sequence diagram data by querying a wind power operation and maintenance knowledge graph and combining the preprocessed SCADA data; s4, carrying out normal behavior modeling on the time sequence diagram data by utilizing the time-space diagram neural network model, predicting each node in the diagram data in the healthy operation time period of the unit in the normal behavior modeling process, and calculating the maximum absolute value of the difference between the predicted value and the actual value of each node as the maximum absolute value of the difference between the predicted value and the actual value; the maximum absolute error is used as a health threshold value of each node to perform state monitoring; in the state monitoring stage, the matching condition of the mapping information of each node in the graph structure in the monitoring result of the real-time data and the fault transmission chain is compared, the health state of the unit is judged, and the fault early warning time is determined. According to the invention, the accuracy and robustness of the monitoring result are enhanced.

Description

Technical Field

[0001] This invention relates to the field of wind turbine condition assessment technology, specifically to a wind turbine condition monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network. Background Technology

[0002] Because wind-rich areas are mostly located in remote regions with harsh environments, poor communication, and inconvenient transportation, as well as near-shore areas along the coastline, and because the key components of the wind turbines (nacelles) are generally installed tens or even hundreds of meters above the ground, and offshore wind turbines are particularly susceptible to severe weather (such as typhoons, heavy rains, and waves), the daily operation and maintenance of wind turbines is quite difficult. Therefore, in order to improve the early warning capability of wind turbine faults and ensure their stable operation and efficient utilization of wind energy resources, it is necessary to conduct research on condition monitoring and fault early warning technologies for wind turbines.

[0003] Current condition monitoring methods primarily rely on time-series data analysis of wind turbines, failing to effectively utilize the value of textual data and ignoring the causal relationships of fault propagation, ultimately affecting the accuracy of diagnosis. Therefore, this invention proposes a wind turbine condition monitoring method driven by the fusion of knowledge graphs and spatiotemporal graph networks. This method first constructs a wind power operation and maintenance knowledge graph using textual data combined with fault propagation information, forming a directed graph structure for wind turbines. Then, it embeds supervisory control and data acquisition (SCADA) system data into the graph structure to generate wind power time-series graph data; it utilizes high-order graph attention networks and Transformer models to mine the spatial and temporal characteristics of the graph data; it constructs a normal behavior model of the turbine based on historical health data; and it determines the turbine's operating status based on the information mapped to nodes in the graph during turbine operation, constructing a monitoring strategy to determine fault warning times and interpreting the condition monitoring results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the above-mentioned problems and to provide a wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A wind turbine condition monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network includes the following steps:

[0007] S1: Collect wind turbine timing SCADA data and wind turbine text data. The wind turbine text data mainly includes: papers and patents, fault record tables and turbine operation and maintenance manuals, etc.

[0008] S2: Construct a wind power operation and maintenance knowledge graph from wind turbine text data. The main steps include text data preprocessing, entity extraction, and graph visualization to obtain an operation and maintenance knowledge graph that can be added, deleted, modified, and queried.

[0009] S3: Perform fault propagation causal relationship query on the operation and maintenance knowledge graph, obtain the wind turbine group graph structure for condition monitoring through filtering, embed the preprocessed SCADA data into the corresponding nodes in the graph structure, and generate time series graph data.

[0010] S4: The spatiotemporal graph neural network model is used to model the normal behavior of time series graph data. In the normal behavior modeling process, the nodes in the graph data of the healthy operation period of the unit are predicted. The maximum absolute value of the difference between the predicted value and the actual value of each node, i.e., the maximum absolute error, is used as the health threshold of each node for status monitoring. In the status monitoring stage, the mapping information of each node in the graph structure and the matching of the fault transmission chain in the monitoring results are compared with the real-time data to determine the health status of the unit and the fault warning time. At the same time, the interpretation and analysis of the monitoring results are realized.

[0011] As a preferred option, in step S2, the wind turbine text data is constructed into a wind power operation and maintenance knowledge graph, which mainly includes the following steps:

[0012] S2.1 Delete blank lines in the unstructured wind turbine text data, and use the jieba tool to segment the text data into words, and then use Word2ve to convert the segmented data into word vectors.

[0013] S2.2. The main entity categories in unstructured wind turbine text data include faulty system, faulty component, fault time, handler, fault cause, and fault handling method. The BiLSTM-CRF named entity recognition model is used to extract each entity from the text data, and triples of type {head entity, relation, tail entity} are constructed based on predefined entity relationships. For example, the relationship between "faulty system" and "faulty sensor" is set to "fault transmission," indicating that when a sensor fails, the fault phenomenon will be reflected in the subsystem. For structured tabular text data, a knowledge graph is constructed by setting rows as nodes, columns as relationships between nodes in the graph, and corresponding cells as attributes of that node.

[0014] S2.3. Use Neo4j software to visualize the triples to obtain a wind power operation and maintenance knowledge graph that can be added, deleted, modified, and queried.

[0015] As a preferred option, the process of generating time series data in step S3 is as follows:

[0016] S3.1 A knowledge graph of wind power operation and maintenance is queried, and subsystems and sensor entity nodes that have experienced anomalies more than 10 times in the wind turbine units are selected. This yields a wind turbine unit graph structure containing the causal relationships of fault propagation. The wind turbine unit graph structure includes three layers: the top layer is the wind turbine unit layer, the next layer is the subsystem layer, and the bottom layer is the sensor layer. Through this graph structure, after the monitoring results are generated, not only can the source of the fault be effectively located, but the fault propagation information contained within can also be used to verify false alarms.

[0017] S3.2 Abnormal data cleaning and dimensionless transformation of SCADA data are required. Dimensionless transformation mainly uses the maximum-minimum value normalization method, as shown in the following formula:

[0018]

[0019] In the formula, x represents the original data, x′ is the normalized value, and x... max With x min These are the maximum and minimum values ​​within the original data range, respectively.

[0020] S3.3. Represent the obtained wind power graph structure using G = (V, E), where V represents the set of all entity nodes in the graph structure, E represents the set of edges for information transmission between nodes, and v i ∈V represents the i-th node, e ij =(v i ,v j ) represents a sub-node v i Point to v j Directed edges; preprocessed time-series SCADA data is embedded into corresponding nodes in the wind turbine group diagram structure to generate time-series graph data, which is then analyzed using the adjacency matrix A∈R. N×N With characteristic matrix X∈R N×F This represents the connections between nodes and the data features inside nodes in a time series graph, where N is the number of nodes and F is the dimension of node features.

[0021] As a preferred option, the specific process of predicting each node in step S4 is as follows:

[0022] S4.1 The spatiotemporal graph neural network consists of a high-order graph attention network, a Transformer layer, and a linear regression prediction layer. The high-order graph attention network learns weights from neighboring nodes of different orders in the temporal graph data, obtaining more information about higher-order nodes to update nodes. This allows each node to fully acquire the spatial information rich in the graph structure. The attention coefficient is calculated during the node feature update process as follows:

[0023]

[0024] In the formula, matrix U∈R n×nLet u represent the parameter matrix of the attention network. i m is the i-th row in matrix U; i,l ∈R L+1 The x represents the l-th order neighborhood weight of node i containing a self-loop, where L represents the order; i,l Represents the node features after the l-th order aggregation of node i; x i,k In this context, k is the neighborhood order traversal index, and L is the highest order of the neighborhood.

[0025] S4.2. Using the multi-head attention mechanism in a high-order graph attention network, the higher-order neighborhoods of nodes are aggregated to obtain the updated node representation:

[0026]

[0027] In the formula, M represents the number of attention heads, and σ represents the nonlinear activation function;

[0028] S4.3. Concatenate the features of each node in the graph data at the current moment to obtain the features of the entire graph data as follows:

[0029] H = [h1, h2, ..., h N (4)

[0030] In the formula, N is the number of nodes in the current time-mapping data, and H∈R NF′ This represents the spatial features constructed from the current time-map data through the HGAT layer, where F′ represents the dimension after node updates;

[0031] S4.4 After capturing spatial feature information, the Transformer layer is used to calculate the contribution of each node's features to the entire window, capturing the global temporal feature information of each node in the sliding window. The output of the HGAT layer is then used to construct a graph sequence with a certain window length S. t ∈H 1:S (t = 1, 2, ..., S) represents the graph feature matrix at time t in the HGAT layer. When the Transformer updates global node information in the window, the graph data is transformed through three independent fully connected layers to generate the corresponding query matrix. Key matrix Value matrix d k With d v To convert the dimensions of matrix K and V to those of a single attention head, the transformation process is as follows:

[0032]

[0033] In the formula, Let b1 be a learnable linear matrix, and b2, b3 be the corresponding deviations. Let i be the query matrix for node i at time k. and Let i be the value matrix and key matrix at time t; and These are the features of node i at time k and the features of node i at time t, respectively.

[0034] Then the attention weight of node i at time k to time t is determined by... and The calculation yielded:

[0035]

[0036] In the formula, Let represent the attention weight of node i at time k to node i at time t. Then, after updating the global temporal information of the node through the Transformer at time t, the features of node i are:

[0037]

[0038] In the formula, Let represent the updated features of node i at time t, which integrates information from other nodes i in the window. Then, the features of the graph data at time t after being updated by the Transformer with a single attention head are represented as follows:

[0039]

[0040] In the formula,

[0041] To capture diverse feature patterns, the Transformer layer employs a multi-head attention mechanism, concatenating the graph data output by each attention head and performing a linear transformation to obtain the final output:

[0042]

[0043] In the formula, Z represents the number of attention heads. Let d be the global linear projection matrix. model This represents the dimension of the hidden layer. Finally, the feature vectors of each node output by the Transformer layer are calculated through the linear output layer to predict the future state of each node in the graph structure.

[0044] As a preferred approach, in step S4, the normal behavior modeling method involves analyzing historical health data to extract features that characterize the normal operating status of the unit, thus revealing the unit's inherent health pattern through modeling. During unit monitoring, when the deviation between the predicted and actual values ​​of SCADA features exceeds the normal behavior threshold, the wind turbine's operating status can be classified as abnormal. Therefore, the state monitoring method combining graph neural networks can be described as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.]1:T The input is fed into function f, which extracts the normal behavioral features of the wind turbine through the time series and graph structure within a window, predicts the next graph after the window, and finally obtains the absolute error between the predicted graph and the actual graph for each window. The calculation formula is as follows:

[0045]

[0046]

[0047] In the formula, T represents the window length, and G... T+1 The next actual image after the window represents the view. G represents the corresponding prediction graph. e This is an absolute error plot reflecting graph anomalies. Because the graph structure contains multiple nodes, the prediction residual of each node is used to represent its own anomaly. The parameters (absolute error) derived from the normal behavior model are used as the anomaly index for each node, calculated as follows:

[0048] e N =|N predcit -N real | (13)

[0049] In the formula, N predcit N represents the model prediction value for the node. real e represents the actual value of the node. N This represents the absolute error at the node.

[0050] As a preferred option, in step S4, after modeling normal behavior, the maximum absolute error of each node in the health data is used as the health threshold. When the absolute error of a node exceeds the threshold, it is considered that the node has an abnormal situation.

[0051] As a preferred approach, in step S4, considering the challenges of variable operating conditions, rapid environmental changes, and a wide variety of sensors in actual wind turbine operation, monitoring results for a single node may be insufficient to accurately reflect the overall health status of the wind turbine, leading to erroneous warnings. Therefore, the status of wind turbine nodes that aggregate the characteristic information of each node is first checked. If an anomaly warning is detected, an anomaly query is initiated to the subsystem layer nodes, following the wind turbine graph structure until it reaches the bottom-level sensor nodes, forming a top-down fault query link. If the reverse fault query link matches the directed wind turbine graph structure, it can be considered as the establishment of a fault transmission chain, triggering a fault warning. After establishing the fault transmission chain, the starting sensor node in the link can be considered the starting point of this anomaly, i.e., the fault source of this wind turbine anomaly or a node close to the fault source.

[0052] The design concept of this invention is as follows:

[0053] This paper proposes a wind turbine condition monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network, which fully utilizes time-series and text-based data generated during wind turbine operation. First, text-based data is preprocessed and knowledge extraction is performed using entity recognition methods to construct wind power operation and maintenance triples. These triples are then visualized using Neo4j to generate an operation and maintenance knowledge graph. Based on fault propagation causal relationship queries, a graph structure rich in fault propagation information is constructed. Next, time-series SCADA data is preprocessed, and relevant features are embedded into corresponding graph structure nodes to generate time-series graph data. A spatiotemporal graph neural network composed of a high-order graph attention network and a Transformer is used to predict the condition of each node. Simultaneously, based on the idea of ​​normal behavior modeling, thresholds for the health status of each node in the graph structure are set to achieve condition monitoring. Finally, combining the fault propagation information in the graph structure, a fault query link and fault propagation chain monitoring strategy are constructed. By analyzing the mapping relationship of each node's monitoring results in the graph structure, the fault warning time is determined, and interpretability support is provided for the condition monitoring results and fault source location.

[0054] This invention makes full use of wind turbine text data to generate time series graph data, and uses a spatiotemporal graph neural network to extract the spatiotemporal features between each node, thereby enhancing the accuracy and robustness of the monitoring results.

[0055] The beneficial effects of this invention are:

[0056] By making full use of textual data from wind turbines, the impact of missing key information on the accuracy and reliability of monitoring results can be reduced. At the same time, SCADA data is embedded into the graph structure generated by the operation and maintenance knowledge graph to achieve multi-source data fusion monitoring. Spatiotemporal graph neural networks are used to extract the spatiotemporal features between nodes, which effectively improves the reliability of wind turbine condition monitoring and the interpretability of monitoring results. Attached Figure Description

[0057] Figure 1 This is a roadmap of the technical implementation scheme of the present invention;

[0058] Figure 2 It is an operation and maintenance knowledge graph constructed from text data of wind turbine units;

[0059] Figure 3 It is a knowledge graph-driven state monitoring wind turbine group diagram structure;

[0060] Figure 4 This is a schematic diagram of each level in the diagram structure from top to bottom;

[0061] Figure 5 These are comparison images of SCADA data before and after cleaning;

[0062] Figure 6 This is a time sequence diagram of wind turbine condition monitoring;

[0063] Figure 7 These are the results of wind turbine condition monitoring;

[0064] Figure 8 It is the timeline of early warning at the node level of wind turbine units and the transmission path of node faults. Detailed Implementation

[0065] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content of this description. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0066] See attached document Figure 1 A method for wind turbine condition monitoring driven by the fusion of knowledge graph and spatiotemporal graph network, the specific steps of which are as follows:

[0067] S1: Acquisition of SCADA data and text data of wind turbine units. Text data mainly includes: papers and patents, operation and maintenance manuals and fault record forms, etc.

[0068] S2: Construct a wind power operation and maintenance knowledge graph from wind power text data. The main steps include text data preprocessing, entity extraction, and graph visualization, resulting in an operation and maintenance knowledge graph that can be added, deleted, modified, and queried (see attached). Figure 2 (as shown);

[0069] In the text data preprocessing stage, blank lines in the unstructured wind power text data are deleted, and the text data is segmented using the jieba tool. The segmented data is then converted into word vectors using Word2ve.

[0070] In the entity extraction stage, the main entity categories in unstructured text data—faulty system, faulty component, fault time, handler, fault cause, and fault handling method—are extracted using the BiLSTM-CRF named entity recognition model. Triples of type {head entity, relation, tail entity} are constructed based on predefined entity relationships. For example, the relationship between "faulty system" and "faulty sensor" is set to "fault transmission," indicating that when a sensor malfunctions, the fault phenomenon will be reflected in the subsystem. For structured tabular text data, a knowledge graph is constructed by setting rows as nodes, columns as relationships between nodes in the graph, and corresponding cells as attributes of that node.

[0071] In the graph visualization stage, Neo4j software is used to visualize the triples, resulting in a wind power operation and maintenance knowledge graph that can be added, deleted, modified, and queried.

[0072] S3: Perform fault propagation causal relationship query on the operation and maintenance knowledge graph, obtain the wind turbine group graph structure for condition monitoring through filtering, embed the preprocessed SCADA data into the corresponding nodes in the graph structure, and generate time series graph data.

[0073] At the wind turbine assembly diagram construction node, by querying the wind power operation and maintenance knowledge graph, subsystems and sensor entity nodes that have experienced more than 10 anomalies in the wind turbine assembly are selected. This yields a wind turbine assembly diagram structure containing the causal relationship of fault propagation. The diagram structure includes three layers: the top layer is the wind turbine assembly layer, the next layer is the subsystem layer, and the bottom layer is the sensor layer (as shown in the attached diagram). Figure 3 , 4 (As shown in the diagram). Through this diagram structure, after the monitoring results are generated, not only can the source of the fault be effectively located, but the fault transmission information contained therein can also be used to verify false alarms.

[0074] During the time-series data preprocessing stage, SCADA needs to perform abnormal data cleaning (as shown in the attached document). Figure 5 (As shown) and dimensionless transformation, which mainly uses the maximum-minimum normalization method, the formula is as follows:

[0075]

[0076] In the formula, x represents the original data, x′ is the normalized value, and x... max With x min These are the maximum and minimum values ​​within the original data range, respectively.

[0077] In the time series graph data generation stage, the resulting wind power graph structure is represented by G = (V, E), where V represents the set of entity nodes in the graph structure, E represents the set of edges for information transmission between nodes, and v i ∈V represents the i-th node, e ij =(v i ,v j ) represents a sub-node v i Point to v j Directed edges; embed the preprocessed time-series SCADA data into the corresponding nodes in the wind turbine group diagram structure to generate time-series diagram data (as shown in the attached diagram). Figure 6 As shown), through the adjacency matrix A∈R N×N With characteristic matrix X∈R N ×F This represents the connections between nodes and the data features inside nodes in a time series graph, where N is the number of nodes and F is the dimension of node features.

[0078] S4: Use a spatiotemporal graph neural network model to model the normal behavior of time series graph data. In the process of normal behavior modeling, predict each node in the graph data during the healthy operation period of the unit, and use the maximum absolute value of the difference between the predicted value and the actual value of each node, i.e., the maximum absolute error, as the health threshold for status monitoring. By comparing the mapping information of each node in the graph structure with the fault transmission chain in the monitoring results, the health status of the unit is judged and the fault warning time is determined. At the same time, the interpretation and analysis of the monitoring results are realized.

[0079] The spatiotemporal graph neural network consists of a high-order graph attention network, a Transformer layer, and a linear regression prediction layer.

[0080] The high-order graph attention network learns weights from neighboring nodes of different orders in the time-series graph data, thus obtaining more information about higher-order nodes to update the nodes. This allows each node to fully acquire the spatial information rich in the graph structure. The attention coefficient is calculated as follows during the node feature update process:

[0081]

[0082] In the formula, matrix U∈R n×n Let u represent the parameter matrix of the attention network. i m is the i-th row in matrix U; i,l ∈R L+1 The x represents the l-th order neighborhood weight of node i containing a self-loop, where L represents the order; i,l Represents the node features after the l-th order aggregation of node i; x i,k In this context, k is the neighborhood order traversal index, and L is the highest order of the neighborhood.

[0083] By using the multi-head attention mechanism in high-order graph attention networks to aggregate the high-order neighborhoods of nodes, the updated node representation is obtained as follows:

[0084]

[0085] In the formula, M represents the number of attention heads, and σ represents the nonlinear activation function;

[0086] By concatenating the features of each node in the graph data at the current moment, the features of the entire graph data are obtained as follows:

[0087] H = [h1, h2, ..., h N (4)

[0088] In the formula, N is the number of nodes in the current time-mapping data, and H∈R NF′ This represents the spatial features constructed from the current time-map data through the HGAT layer, where F′ represents the dimension after node updates;

[0089] After capturing spatial feature information, the Transformer layer is used to calculate the contribution of each node's features to the entire window, capturing the global temporal feature information of each node in the sliding window. The output of the HGAT layer is then used to construct a graph sequence with a certain window length S. t ∈H 1:S (t = 1, 2, ..., S) represents the graph feature matrix at time t in the HGAT layer. When the Transformer updates global node information in the window, the graph data is transformed through three independent fully connected layers to generate the corresponding query matrix. Key matrix Value matrix d k With d v To convert the dimensions of matrix K and V to those of a single attention head, the transformation process is as follows:

[0090]

[0091] In the formula, Let b1 be a learnable linear matrix, and b2, b3 be the corresponding deviations. Let i be the query matrix for node i at time k. and Let i be the value matrix and key matrix at time t; and These are the features of node i at time k and the features of node i at time t, respectively.

[0092] Then the attention weight of node i at time k to time t is determined by... and The calculation yielded:

[0093]

[0094] In the formula, Let represent the attention weight of node i at time k to node i at time t. Then, after updating the global temporal information of the node through the Transformer at time t, the features of node i are:

[0095]

[0096] In the formula, Let represent the updated features of node i at time t, which integrates information from other nodes i in the window. Then, the features of the graph data at time t after being updated by the Transformer with a single attention head are represented as follows:

[0097]

[0098] In the formula,

[0099] To capture diverse feature patterns, the Transformer layer employs a multi-head attention mechanism, concatenating the graph data output by each attention head and performing a linear transformation to obtain the final output:

[0100]

[0101] In the formula, Z represents the number of attention heads. Let d be the global linear projection matrix. model This represents the dimension of the hidden layer. Finally, the feature vectors of each node output by the Transformer layer are calculated through the linear output layer to predict the future state of each node in the graph structure.

[0102] A spatiotemporal graph neural network framework is constructed to model the normal behavior of wind turbines. This method is based on data prediction-driven condition monitoring technology. By training on historical health data, features that characterize the normal operating status of the turbine are extracted, revealing the turbine's inherent health patterns through modeling. During turbine monitoring, when the deviation between the predicted and actual values ​​of SCADA features exceeds the normal behavior threshold, the wind turbine's operating status can be identified as abnormal. Therefore, the condition monitoring method combined with the spatiotemporal graph neural network framework can be described as follows: A graph G within a time window... 1:T The input is fed into function f, which extracts the normal behavioral features of the wind turbine through the time series and graph structure within a window, predicts the next graph after the window, and finally obtains the absolute error between the predicted graph and the actual graph for each window. The calculation formula is as follows:

[0103]

[0104] In the formula, T represents the window length, and G... T+1 The next actual image after the window represents the view. G represents the corresponding prediction graph. e This is an absolute error plot reflecting graph anomalies. Because the graph structure contains multiple nodes, the prediction residual of each node is used to represent its own anomaly. The parameters (absolute error) derived from the normal behavior model are used as the anomaly index for each node, calculated as follows:

[0105]

[0106] In the formula, N predcit N represents the model prediction value for the node. real e represents the actual value of the node. N The absolute error of the node;

[0107] The maximum absolute error (MAE) is the maximum absolute difference between the model's predicted and actual values ​​for each node in the graph data during the healthy operation period of the wind turbine, used as the health threshold for status monitoring. Considering the challenges of variable operating conditions, rapid environmental changes, and a wide variety of sensors in actual wind turbine operation, monitoring results for a single node may be insufficient to accurately reflect the overall health status of the wind turbine, leading to erroneous warnings. Therefore, status detection is first performed on wind turbine nodes that aggregate feature information. If an anomaly warning is detected, an anomaly query is initiated to the subsystem layer nodes, following the wind turbine graph structure until reaching the bottom-level sensor nodes, forming a top-down fault query chain. If the reverse fault query chain matches the directed wind turbine graph structure, it can be considered the establishment of a fault propagation chain, triggering a fault warning. After establishing the fault propagation chain, the starting sensor node in the chain can be considered the starting point of the current anomaly, i.e., the fault source of the current wind turbine anomaly or a node close to the fault source (as shown in the attached diagram). Figure 7 , 8 (As shown).

[0108] The method proposed in this invention was verified using time-series SCADA data and text data generated during the actual operation of doubly-fed induction generators in a wind farm. The overall flowchart is as follows. Figure 1 As shown, the details are as follows:

[0109] (1) Select text data related to unit operation and maintenance from wind power text data, perform data preprocessing, and extract knowledge using named entity recognition to obtain wind power operation and maintenance triples. Visualize these triples using Neo4j to obtain a wind power operation and maintenance knowledge graph that can be added, deleted, modified, and queried (as shown in the appendix). Figure 2 (As shown). By querying the fault propagation causal relationships and adjusting the wind power operation and maintenance knowledge graph using expert experience, a wind turbine group diagram structure for condition monitoring was obtained (as shown in the attached diagram). Figure 3 , 4 As shown in the table below, the information for each node is as follows.

[0110] node Node information node Node information 1 wind turbine 12 Tower bottom control cabinet temperature 2 dynamo 13 External environment sensors 3 generator speed 14 Average wind speed 30s 4 Generator non-drive end bearing temperature 15 Ambient temperature 5 Hydraulic system 16 spindle 6 Hydraulic system pressure 17 Spindle speed 7 Yaw system 18 gearbox 8 Yaw angle 19 Gearbox pump outlet pressure 9 cabin 20 Gearbox pump inlet pressure 10 Cabin temperature 21 Gearbox oil temperature 11 tower

[0111] (2) Perform data preprocessing on the SCADA data (as shown in the attached document). Figure 5 As shown in the figure, the corresponding features are embedded into the wind turbine group diagram structure nodes to generate time series diagram data (as shown in the attached figure). Figure 6 (As shown).

[0112] (3) Combining the normal behavior modeling approach, the spatiotemporal graph neural network framework is used to train the health graph data to obtain the threshold of the health state.

[0113] (4) Based on the knowledge graph-driven graph structure and the spatiotemporal graph neural network framework, if the absolute error of a node exceeds a predetermined threshold, it becomes an abnormal node. Over a period of time, this node will transmit the abnormal information to adjacent nodes in the graph structure, leading to abnormal conditions in those adjacent nodes. Therefore, initially, only the overall status of the wind turbine unit represented by node 1 is observed. If a condition exceeds the health threshold, the next subsystem layer is monitored. If a node in the subsystem layer becomes abnormal at that moment, the query continues down to the next layer. When a fault occurs, the monitoring results of each abnormal node in the graph structure are as follows: Figure 7 As shown in the table below, the first time when each abnormal node continuously exceeds the threshold is as follows. The mapping of the abnormal node in the graph structure at each time point is as follows. Figure 8 As shown in the diagram, node 1 was below the threshold before time t4, then rose rapidly. Therefore, the graph data at time t4 was analyzed using a top-down fault query method to check for node anomalies. It was found that node 2 at the subsystem level and node 4 at the sensor level also exhibited anomaly warnings at this time. Subsequent node-level state monitoring results showed that nodes 1, 2, and 4 all generated anomaly warnings at this time, forming a fault query chain. By observing the reverse of the fault query chain, it was found that at time t4, it matched the wind turbine graph structure, forming a fault transmission chain, which was considered an anomaly warning issued by the wind turbine. The fault source or the sensor near the fault source was the temperature of the non-drive end bearing of the generator. The subsequent anomaly warnings at nodes 20, 18, and 3 also demonstrated that fault information propagates along the knowledge graph-driven graph structure, effectively verifying the wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network.

[0114]

[0115]

[0116] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope of the present invention.

Claims

1. A wind turbine condition monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network, characterized in that, Includes the following steps: S1: Collect wind turbine timing SCADA data and wind turbine text data; S2: Construct a wind power operation and maintenance knowledge graph from the text data of wind turbine units. The construction process includes text data preprocessing, entity extraction and graph visualization to obtain an operation and maintenance knowledge graph that can be added, deleted, modified and queried. S3: Perform fault propagation causal relationship query on the operation and maintenance knowledge graph, obtain the wind turbine group diagram structure for condition monitoring through filtering, embed the preprocessed SCADA data into the corresponding nodes in the wind turbine group diagram structure, and generate time series diagram data. S4: The spatiotemporal graph neural network model is used to model the normal behavior of time series graph data. In the normal behavior modeling process, the nodes in the graph data of the healthy operation period of the unit are predicted. The maximum absolute value of the difference between the predicted value and the actual value of each node, i.e., the maximum absolute error, is used as the health threshold of each node for status monitoring. In the status monitoring stage, the mapping information of each node in the graph structure and the matching of the fault transmission chain in the monitoring results are compared with the real-time data to determine the health status of the unit and determine the fault warning time. At the same time, the interpretation and analysis of the monitoring results are realized. The healthy operation period of the unit is the period when the operating parameters are continuously within the design threshold range and the unit is not triggered by the fault of the unit itself.

2. The wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network according to claim 1, characterized in that, Step S2 specifically includes: S2.1 Text data preprocessing: Remove blank lines from unstructured wind turbine text data and convert the text data into word vectors using a dictionary; S2.2 Entity Extraction: Entity categories in unstructured wind turbine text data include faulty system, faulty component, fault time, handler, fault cause, and fault handling method. The BiLSTM-CRF named entity recognition model is used to extract each entity from the wind turbine text data, and triples of type {head entity, relation, tail entity} are constructed based on predefined entity relationships. For structured tabular wind turbine text data, a knowledge graph is constructed by setting rows as nodes, columns as relationships between nodes in the graph, and corresponding cells as node attributes. S2.3, Graph Visualization: Using Neo4j software, the triples are visualized to obtain a wind power operation and maintenance knowledge graph that can be added, deleted, modified, and queried.

3. The wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network according to claim 1, characterized in that, Step S3 is as follows: S3.1 Query the wind power operation and maintenance knowledge graph, filter the subsystems and sensor entity nodes that have anomalies more than 10 times in the wind turbine, and obtain the wind turbine diagram structure containing the causal relationship of fault transmission. The wind turbine diagram structure includes a three-layer structure: the top layer is the wind turbine layer, the next layer is the subsystem layer, and the bottom layer is the sensor layer. S3.2 Perform abnormal data cleaning and dimensionless transformation on SCADA. Dimensionless transformation adopts the maximum-minimum value normalization method, and the formula is as follows: In the formula, x represents the original data, x′ is the normalized value, and x... max With x min These are the maximum and minimum values ​​within the original data range, respectively. S3.

3. Represent the obtained wind turbine group graph structure using G = (V, E), where V represents the set of all entity nodes in the graph structure, E represents the set of edges for information transmission between nodes, and v i ∈V represents the i-th node, e ij =(v i ,v j ) represents a sub-node v i Point to v j Directed edges; preprocessed time-series SCADA data is embedded into corresponding nodes in the wind turbine group diagram structure to generate time-series graph data, which is then analyzed using the adjacency matrix A∈R. N×N With characteristic matrix X∈R N×F This represents the connections between nodes and the data features inside nodes in a time series graph, where N is the number of nodes and F is the dimension of node features.

4. The wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network according to claim 1, the specific process of predicting each node in step S4 is as follows: S4.1 The spatiotemporal graph neural network model includes a high-order graph attention network, a Transformer layer, and a linear regression prediction layer. The high-order graph attention network learns weights from neighboring nodes of different orders in the time-series graph data to obtain more information about higher-order nodes and update them, ensuring that each node fully acquires the spatial information rich in the graph structure. The attention coefficient is calculated during the node feature update process as follows: In the formula, matrix U∈R n×n Let u represent the parameter matrix of the attention network. i m is the i-th row in matrix U; i,l ∈R L+1 The x represents the l-th order neighborhood weight of node i containing a self-loop, where L represents the order; i,l Represents the node features after the l-th order aggregation of node i; x i,k In this context, k is the neighborhood order traversal index, and L is the highest order of the neighborhood. S4.

2. Using the multi-head attention mechanism in a high-order graph attention network, the higher-order neighborhoods of nodes are aggregated to obtain the updated node representation: In the formula, M represents the number of attention heads, and σ represents the nonlinear activation function; S4.

3. Concatenate the features of each node in the graph data at the current moment to obtain the features of the entire graph data as follows: H=[h1,h2,···,h N ] (4) In the formula, N is the number of nodes in the current time-mapping data, and H∈R NF′ This represents the spatial features constructed from the current time-map data through the HGAT layer, where F′ represents the dimension after node updates; S4.4 After capturing spatial feature information, the Transformer layer is used to calculate the contribution of each node's features to the entire window, capturing the global temporal feature information of each node in the sliding window. The output of the HGAT layer is then used to construct a graph sequence with a certain window length S. t ∈H 1:S (t = 1, 2, ..., S) represents the graph feature matrix at time t in the HGAT layer. When the Transformer updates global node information in the window, the graph data is transformed through three independent fully connected layers to generate the corresponding query matrix. Key matrix Value matrix d k With d v To convert the dimensions of matrix K and V to those of a single attention head, the transformation process is as follows: In the formula, Let b1 be a learnable linear matrix, and b2, b3 be the corresponding deviations. Let i be the query matrix for node i at time k. and Let i be the value matrix and key matrix at time t; and These are the features of node i at time k and the features of node i at time t, respectively. Then the attention weight of node i at time k to time t is determined by... and The calculation yielded: In the formula, Let represent the attention weight of node i at time k to node i at time t. Then, after updating the global temporal information of the node through the Transformer at time t, the features of node i are: In the formula, Let represent the updated features of node i at time t, which integrates information from other nodes i in the window. Then, the features of the graph data at time t after being updated by the Transformer with a single attention head are represented as follows: In the formula, To capture diverse feature patterns, the Transformer layer employs a multi-head attention mechanism, concatenating the graph data output by each attention head and performing a linear transformation to obtain the final output: In the formula, Z represents the number of attention heads. Let d be the global linear projection matrix. model The hidden layer dimension is represented by the feature vector of each node output by the Transformer layer through the linear output layer, which is used to predict the future state of each node in the graph structure.

5. The wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network according to claim 1, the specific operation of normal behavior modeling in step S4 is as follows: The normal behavior modeling method is to analyze historical health data and extract features that can characterize the normal operating status of the unit, that is, to reveal the inherent health pattern of the unit through modeling. During the monitoring of the unit, when the deviation between the predicted value and the actual value of the SCADA feature exceeds the normal behavior threshold, the operating status of the wind turbine is judged as abnormal. The state monitoring method combining a temporal graphical neural network model is described as follows: G is a graph within a time window... 1:T The input is fed into function f, which extracts the normal behavioral features of the wind turbine through the time series and graph structure within a window, predicts the next graph after the window, and finally obtains the absolute error between the predicted graph and the actual graph for each window. The calculation formula is as follows: In the formula, T represents the window length, and G... T+1 The next actual image after the window. G represents the corresponding prediction graph. e An absolute error graph to reflect anomalies in the graph; The anomaly of each node is represented by its own prediction residual. The anomaly index for each node is obtained by using the absolute error parameter derived from the normal behavior model. The calculation formula is as follows: e N =|N predcit -N real | (13) In the formula, N predcit N represents the model prediction value for the node. real e represents the actual value of the node. N This represents the absolute error at the node.

6. According to claim 1, a wind turbine state monitoring method driven by the fusion of knowledge graph and spatiotemporal graph network, in step S4, the maximum absolute error of each node in the graph data under normal conditions is used as the health threshold for state monitoring; by comparing the mapping information of each node in the graph structure with the fault transmission chain in the monitoring results, the health status of the turbine is determined and the fault warning time is determined. The specific process is as follows: First, the status of the wind turbine nodes that aggregate the feature information of each node is detected. If an anomaly warning is detected, an anomaly query is initiated to the subsystem layer nodes. The query is then carried out along the wind turbine diagram structure until the sensor nodes at the bottom layer are reached, forming a top-down fault query link. If the reverse fault query link matches the directed wind turbine group diagram structure, it can be regarded as the establishment of a fault transmission chain, and fault early warning can be carried out. After establishing the fault transmission chain, the starting sensor node in the chain is regarded as the starting point of this anomaly, that is, the fault source of this wind turbine abnormality or the node close to the fault source.

Images