Offshore wind power generation operation and maintenance anomaly detection and positioning method based on graph neural network

Abstract

The invention discloses an offshore wind power generation operation and maintenance anomaly detection and positioning method based on a graph neural network, and the method comprises the following steps: S1, collecting operation monitoring data of an offshore wind power generation system, and carrying out the preprocessing of the operation monitoring data; s2, constructing a global graph of the wind power generation system; s3, designing an abnormal disturbance simulation mechanism, and generating disturbance diagram sample data; s4, constructing a graph comparison learning module based on the improved Grapher network, and training the Grapher network through comparison learning; s5, extracting a sub-graph region associated with the potential anomaly from the global graph; s6, executing graph embedding propagation on each sub-graph; and S7, constructing a dual-task module, and generating an anomaly recognition report. According to the method, the graph comparison learning and the improved Grapher network are fused, offshore wind power anomaly detection and positioning are achieved, and the method has the advantages of being high in structural perception, high in positioning precision and rapid in response.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power system operation and maintenance technology, and in particular to a method for detecting and locating anomalies in offshore wind power operation and maintenance based on graph neural networks. Background Technology

[0002] With the continuous growth of global demand for clean energy, offshore wind power, as an important component of renewable energy, has become a key area of ​​green energy development for many countries. Compared with onshore wind power, offshore wind power has advantages such as abundant wind energy resources, stable wind speed, and less conflict with land use, and has the potential for large-scale development and centralized grid connection. However, due to the complex construction and operation environment, such as high humidity, high salinity, and high corrosion in the ocean, offshore wind power equipment is prone to failure, degradation, and other abnormal operation during actual operation, which puts forward higher requirements for equipment condition monitoring, fault diagnosis, and anomaly location.

[0003] Currently, a large number of operation monitoring devices are widely deployed in wind power operation and maintenance systems to collect multi-dimensional sensor data such as wind speed, current, voltage, temperature, and vibration in real time. These monitoring data can provide data support for the status analysis of wind power equipment. However, in actual operation and maintenance, traditional data processing methods mostly rely on single-device, single-variable anomaly detection methods, such as threshold discrimination, statistical models, and isolated forest algorithms. These methods usually treat the equipment as an independent node and ignore the physical connection relationship, control logic dependency, and collaborative operation mode between the equipment in the wind power system. As a result, it is impossible to understand the propagation characteristics and structural impact of anomalies from the global system level.

[0004] In existing research on intelligent wind power monitoring, some methods have attempted to introduce machine learning and deep learning models, such as convolutional neural networks, recurrent neural networks, or autoencoders, to process time-series data or model equipment status. These methods have made improvements in local feature extraction and trend prediction, but they still lack the ability to model the topology of wind power systems and struggle to effectively capture the complex relationships between equipment in terms of spatial structure and information transmission. Furthermore, these models suffer from poor scalability, weak generalization ability, and insufficient localization capability when dealing with multiple equipment and multiple data source inputs at the wind farm level.

[0005] In recent years, graph neural networks have been widely used in social networks, power systems, traffic flow analysis and other fields due to their outstanding ability in structured data modeling. In wind power systems, the equipment such as wind turbines, pitch systems, transformer substations and collector lines have natural graph structure characteristics. Therefore, graph neural networks have a natural advantage in modeling wind power systems. Some studies have made preliminary attempts to apply graph neural networks to wind power systems, such as wind turbine group power prediction and wind farm state assessment. However, most of these methods are still at the level of basic graph convolution operations and have failed to combine the control dependency characteristics and multi-source data attributes between wind power equipment. At the same time, they are difficult to accurately output the location or propagation path of anomalies in anomaly detection.

[0006] On the other hand, in actual operation and maintenance scenarios, simply detecting abnormal events is no longer sufficient to meet engineering requirements. Accurately tracking the location and propagation direction of the abnormal source node has become the key to improving operation and maintenance efficiency and accuracy. Therefore, current graph-based models have not yet achieved the dual goals of abnormal detection and abnormal location. They lack the ability to model "abnormal causal paths" in multi-device collaborative systems. In addition, most existing graph learning methods are based on static graph training, which lacks the ability to adapt to the real-time and dynamic changes of offshore wind farms and is also difficult to form a generalizable comparative expression ability from historical disturbances.

[0007] In summary, existing offshore wind power system operation and maintenance anomaly detection technologies generally suffer from the following shortcomings: First, they lack a system-wide perspective and fail to model the structural relationships and operational dependencies between equipment, resulting in low accuracy of detection results; second, most traditional models can only determine whether an anomaly is present, but cannot further accurately track the anomaly propagation path or locate specific abnormal equipment nodes; third, they lack structural comparison enhancement mechanisms in data utilization, making it difficult to train deep models with generalization capabilities; and fourth, while graph neural network models have strong generality, they lack specificity and have not been specifically optimized for wind farm application scenarios. Summary of the Invention

[0008] One objective of this invention is to propose a method for detecting and locating anomalies in offshore wind power generation based on graph neural networks. This invention integrates graph contrastive learning mechanism and an improved Graphormer graph neural network model to construct a structured anomaly detection and source node location method for offshore wind power systems. It can accurately characterize the structural relationships and anomaly propagation paths between equipment, realize intelligent identification and tracing of the operating status of offshore wind power equipment, and has the advantages of strong structural perception, high positioning accuracy, fast response efficiency and strong model generalization ability.

[0009] A method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to an embodiment of the present invention includes the following steps: S1. Collect operational monitoring data of offshore wind power generation systems and preprocess the operational monitoring data; S2. Construct a global graph of the wind power generation system based on operation monitoring data to form original graph data with node attributes and edge attributes; S3. Design an abnormal disturbance simulation mechanism to generate a set of disturbance map sample data based on the original map data; S4. Construct a graph contrastive learning module based on the improved Graphormer network. Input the original graph data and perturbed graph sample data as positive and negative samples into the Graphormer network, train the Graphormer network through contrastive learning, and generate node prior scores. S5. Based on the node prior scores and attention weights generated by the Graphormer network, subgraph regions associated with potential anomalies are dynamically extracted from the global graph to form candidate subgraphs. S6. Perform graph embedding propagation based on the Graphormer network on each subgraph, extract local embedding representations, and generate global embedding representations by weighting attention between subgraphs; S7. Construct a dual-task module. The first task outputs the anomaly detection results for each node, and the second task outputs the location of the source node where the anomaly propagates and generates an anomaly identification report.

[0010] Optionally, the operational monitoring data includes wind speed, current, voltage, vibration, and temperature data, and the preprocessing includes outlier removal, missing value filling, noise removal, data normalization, and data standardization.

[0011] Optionally, the global graph represents wind power equipment as nodes in the global graph, the connections between equipment as edges in the global graph, the operation monitoring data as node feature vectors, and the physical connection relationships and control dependencies as edge weights.

[0012] Optionally, the original graph data combines the node set, edge set, node attribute matrix, and edge attribute matrix to form graph structure input data. The node attributes are operation monitoring data, and the edge attributes include control command frequency, communication bandwidth, and number of historical collaborative actions.

[0013] Optionally, the abnormal disturbance simulation mechanism includes node feature occlusion, edge connection variation, and topology disturbance.

[0014] Optionally, S3 specifically includes: S31. Perturb the node attributes in the original graph data, randomly select some nodes, and perform masking processing on the feature vectors of the selected nodes. Feature masking is achieved by setting a mask or adding random perturbation. S32. Perturb the edge connection relationships in the original graph structure, randomly select some edges to modify the connection state, including three operation methods: deleting edges, adding new edges, or modifying edge weights. The operation of deleting edges is to remove the current edge connection relationship, the operation of adding edges is to establish a new edge connection between unconnected nodes, and the operation of modifying edge weights is to perturb the attribute feature values ​​of existing edges. S33. Disrupt the topology of the graph by adjusting the connection order between nodes, exchanging local adjacency relationships, or rearranging edge combinations to simulate structural anomalies or connection failures between devices, thus forming a structurally disrupted graph. S34. Based on three types of perturbation methods—node feature perturbation, edge connection mutation, and topological structure disturbance—perturbation graph sample data is generated and used as input data for the graph comparison learning training stage.

[0015] Optionally, S4 specifically includes: S41. Construct a graph contrast learning module based on an improved Graphormer network, wherein the improved Graphormer network includes a structure centrality encoding layer, a spatial bias layer, a multi-head attention mechanism, a feedforward network layer, and an output layer. S42. Construct positive and negative samples from the original graph data and the perturbation graph sample data respectively; S43. In the structural centrality coding layer, calculate the shortest path distance for each pair of nodes and calculate the structural centrality score of the nodes to generate a structural position coding vector. The structural centrality adopts the normalized betweenness centrality. ; in, Represents the structural position encoding vector. Represents a node With nodes The shortest path distance between them Represents a node Structural centrality score, Represents a node Structural centrality score; The structural location encoding vector is then injected into the node feature vector to form an enhanced node representation: ; in, Represents a node The enhanced node representation, Represents the node feature vector. Represents the structural position encoding vector. Represents a node The set of adjacent nodes; S44. In the spatial bias layer guided by edge features, attention bias terms are generated based on edge attribute vectors: ; in, Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first The edge bias mapping weight matrix of each attention head. Represents the edge attribute vector. Indicates the number of attention heads; S45. In multi-head self-attention mechanisms, augmented node representations are used. Construct query vectors and key vectors, and combine them with bias terms. Calculate attention weights: ; in, Indicates the first Nodes in each attention head With nodes Attention weights This represents the natural exponential function. Indicates the first Each attention node The query vector, Indicates the first Each attention node The key vector, This represents the dimension of the embedding vector. Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first Each attention node The key vector, Represents a node The set of adjacent nodes, Indicates the transpose operation; S46. The results of each attention head are concatenated and input into the feedforward network. Nonlinear mapping, residual connection, and normalization are performed to generate a set of node embedding representations. ; in, The node embedding represents a set. Indicates the first Nodes in each attention head The node embedding representation, Indicates the first Nodes in each attention head With nodes Attention weights Indicates the first Nodes in each attention head value vector, Indicates the number of nodes; S47, Embedding Sets of Nodes Perform global pooling, then map it to a graph comparison embedding representation using a projection head. A contrastive loss function is constructed, and the Graphormer network is trained by minimizing the contrastive loss function to optimize the node embedding representation. The projection head is a fully connected neural network. ; in, This represents the contrastive loss function. Graph representation versus embedded representation This indicates an enhanced perturbation version of the embedding representation. The graph compares the cosine similarity between the embedded representation and the enhanced perturbation version of the embedded representation. Indicates the number of negative samples. Indicates the temperature coefficient. Indicates the first Graph embedding representation of each negative sample Representation of graphs compared to embedded representations and the first Cosine similarity between the graph embedding representations of negative samples; S48. Based on the trained Graphormer network, calculate the Euclidean distance between the center vector of each node embedding representation and the normal node embedding, and generate the node prior score.

[0016] Optionally, S5 specifically includes: S51. Based on the trained Graphormer network, obtain the prior scores of nodes; S52. Obtain the attention weight of each edge in the global graph, construct a node-edge joint anomaly index matrix based on the node prior score and attention weight, and filter the set of nodes whose node prior score is greater than a preset threshold. S53. Using the filtered set of nodes as the center in the global graph, extract the first-order and second-order adjacency structures associated with the abnormal nodes by combining the attention weight of each edge, and form an initial candidate subgraph set. S54. Aggregate the mean of the prior scores of all nodes in each initial candidate subgraph. S55. Sort all subgraphs by average score and attention weight density, and select subgraphs with scores higher than the subgraph threshold to form candidate subgraphs.

[0017] Optionally, S6 specifically includes: S61. Perform graph embedding propagation processing based on Graphormer network on the candidate subgraphs respectively; S62. Input the node feature vector and structural position encoding vector into the shared Graphormer network, perform the graph embedding propagation process, and output the local embedding representation of each node; S63. Perform pooling operation on the local embedding representation of nodes in each subgraph; S64. Based on the attention calculation results between the local embedding representations of each subgraph and the query vector, assign attention weights to each subgraph in the final fusion, and aggregate them using an attention weighting method to form a global embedding representation.

[0018] Optionally, S7 specifically includes: S71. Based on the global embedding representation and the local embedding representation of each node, a dual-task module is constructed. The dual-task module includes two output branches: a node-level anomaly detection subtask and an anomaly source localization subtask. S72. In the node-level anomaly detection subtask, the local embedding representation of each node is input into the fully connected network for anomaly probability prediction. S73. Based on the prediction results of all nodes, an anomaly detection loss function is constructed using binary cross-entropy loss. S74. In the anomaly source localization subtask, the global embedding representation is input to the source node prediction branch network, and the source node index distribution is calculated. The source node prediction branch network is a fully connected network. S75. Construct the cross-entropy loss function for anomaly source localization and define the total loss function, which is the weighted sum of the losses of the two sub-tasks; S76. The objective optimization dual-task module based on minimizing the total loss function outputs the anomaly detection results and anomaly source localization probability results for each node, and generates an anomaly identification report.

[0019] The beneficial effects of this invention are: First, this invention introduces a graph neural network to perform structured modeling of offshore wind power systems. Combined with key modules such as graph comparison learning, node prior score generation, candidate subgraph extraction and embedding fusion, and multi-task output, it achieves accurate detection of abnormal states and spatial location of source nodes during the operation and maintenance of offshore wind power systems, thus improving the intelligent operation and maintenance capabilities of offshore wind power systems. In existing technologies, wind power equipment is often modeled as an independent entity, failing to effectively utilize the physical connections and control dependencies between devices. This invention proposes mapping operational monitoring data to a global graph structure, constructing a graph data model with node and edge attributes. This allows the structural collaborative relationships between devices to be fully expressed in the graph network, overcoming the limitations of point-to-point isolated analysis in traditional methods.

[0020] Secondly, by designing an anomaly perturbation simulation mechanism involving node feature perturbation, edge connection mutation, and topological structure disruption, and combining it with an improved Graphormer network to construct a graph comparison learning module, this invention enables the model to identify structural changes and feature deviations during the training phase, thereby enhancing the model's ability to perceive real anomaly samples in complex maritime environments. Furthermore, this invention introduces a joint node-edge anomaly index using node prior scores and attention weights to dynamically extract subgraph regions associated with potential anomalies, effectively reducing the scope of embedding propagation computation and improving the overall processing efficiency and real-time response capability of the system. Simultaneously, by performing graph embedding propagation on each candidate subgraph and constructing an attention fusion mechanism between subgraphs, this invention not only extracts local anomaly structural features but also generates a full-graph-level fusion representation, providing strong expressive support for subsequent dual-task detection and localization.

[0021] Finally, in the final anomaly identification stage, this invention designs a dual-task module that simultaneously includes node-level anomaly detection and anomaly source localization. This module can output the location of potential fault propagation source nodes while identifying anomaly states, improving the efficiency and accuracy of anomaly response and overcoming the practical operational challenge of "only detecting but not locating" in traditional methods. Furthermore, by jointly minimizing the dual-task loss function, the model can balance global discrimination capability and spatial source tracing capability during the optimization process, further enhancing its adaptability to complex anomaly patterns. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for detecting and locating anomalies in offshore wind power generation based on graph neural networks, as proposed in this invention. Figure 2This is a schematic diagram of the graph contrast learning module structure based on the improved Graphormer network in the proposed method for detecting and locating anomalies in offshore wind power generation based on graph neural networks. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figure 1-2 A method for detecting and locating anomalies in offshore wind power operation and maintenance based on graph neural networks includes the following steps: S1. Collect operational monitoring data of offshore wind power generation systems and preprocess the operational monitoring data; S2. Construct a global graph of the wind power generation system based on operation monitoring data to form original graph data with node attributes and edge attributes; S3. Design an abnormal disturbance simulation mechanism to generate a set of disturbance map sample data based on the original map data; S4. Construct a graph contrastive learning module based on the improved Graphormer network. Input the original graph data and perturbed graph sample data as positive and negative samples into the Graphormer network, train the Graphormer network through contrastive learning, and generate node prior scores. S5. Based on the node prior scores and attention weights generated by the Graphormer network, subgraph regions associated with potential anomalies are dynamically extracted from the global graph to form candidate subgraphs. S6. Perform graph embedding propagation based on the Graphormer network on each subgraph, extract local embedding representations, and generate global embedding representations by weighting attention between subgraphs; S7. Construct a dual-task module. The first task outputs the anomaly detection results for each node, and the second task outputs the location of the source node where the anomaly propagates and generates an anomaly identification report.

[0025] In this embodiment, the operational monitoring data includes wind speed, current, voltage, vibration, and temperature data, and the preprocessing includes outlier removal, missing value filling, noise removal, data normalization, and data standardization.

[0026] In this embodiment, the global graph represents wind power equipment as nodes, the connections between equipment as edges, the operation monitoring data as node feature vectors, and the physical connection relationships and control dependencies as edge weights.

[0027] In this embodiment, the original graph data combines the node set, edge set, node attribute matrix, and edge attribute matrix to form graph structure input data. The node attributes are operation monitoring data, and the edge attributes include control command frequency, communication bandwidth, and number of historical collaborative actions.

[0028] In this embodiment, the abnormal disturbance simulation mechanism includes node feature occlusion, edge connection variation, and topology disturbance.

[0029] In this embodiment, S3 specifically includes: S31. Perturb the node attributes in the original graph data, randomly select some nodes, and perform masking processing on the feature vectors of the selected nodes. Feature masking is achieved by setting a mask or adding random perturbation. S32. Perturb the edge connection relationships in the original graph structure, randomly select some edges to modify the connection state, including three operation methods: deleting edges, adding new edges, or modifying edge weights. The operation of deleting edges is to remove the current edge connection relationship, the operation of adding edges is to establish a new edge connection between unconnected nodes, and the operation of modifying edge weights is to perturb the attribute feature values ​​of existing edges. S33. Disrupt the topology of the graph by adjusting the connection order between nodes, exchanging local adjacency relationships, or rearranging edge combinations to simulate structural anomalies or connection failures between devices, thus forming a structurally disrupted graph. S34. Based on three types of perturbation methods—node feature perturbation, edge connection mutation, and topological structure disturbance—perturbation graph sample data is generated and used as input data for the graph comparison learning training stage.

[0030] In this embodiment, S4 specifically includes: S41. Construct a graph contrast learning module based on an improved Graphormer network, wherein the improved Graphormer network includes a structure centrality encoding layer, a spatial bias layer, a multi-head attention mechanism, a feedforward network layer, and an output layer. S42. Construct positive and negative samples from the original graph data and the perturbation graph sample data respectively; S43. In the structural centrality coding layer, calculate the shortest path distance for each pair of nodes and calculate the structural centrality score of the nodes to generate a structural position coding vector. The structural centrality adopts the normalized betweenness centrality. ; in, Represents the structural position encoding vector. Represents a node With nodes The shortest path distance between them Represents a node Structural centrality score, Represents a node Structural centrality score; The structural location encoding vector is then injected into the node feature vector to form an enhanced node representation: ; in, Represents a node The enhanced node representation, Represents the node feature vector. Represents the structural position encoding vector. Represents a node The set of adjacent nodes; S44. In the spatial bias layer guided by edge features, attention bias terms are generated based on edge attribute vectors: ; in, Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first The edge bias mapping weight matrix of each attention head. Represents the edge attribute vector. Indicates the number of attention heads; S45. In multi-head self-attention mechanisms, augmented node representations are used. Construct query vectors and key vectors, and combine them with bias terms. Calculate attention weights: ; in, Indicates the first Nodes in each attention head With nodes Attention weights This represents the natural exponential function. Indicates the first Each attention node The query vector, Indicates the first Each attention node The key vector, This represents the dimension of the embedding vector. Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first Each attention node The key vector, Represents a node The set of adjacent nodes, Indicates the transpose operation; S46. The results of each attention head are concatenated and input into the feedforward network. Nonlinear mapping, residual connection, and normalization are performed to generate a set of node embedding representations. ; in, The node embedding represents a set. Indicates the first Nodes in each attention head The node embedding representation, Indicates the first Nodes in each attention head With nodes Attention weights Indicates the first Nodes in each attention head value vector, Indicates the number of nodes; S47, Embedding Sets of Nodes Perform global pooling, then map it to a graph comparison embedding representation using a projection head. A contrastive loss function is constructed, and the Graphormer network is trained by minimizing the contrastive loss function to optimize the node embedding representation. The projection head is a fully connected neural network. ; in, This represents the contrastive loss function. Graph representation versus embedded representation This indicates an enhanced perturbation version of the embedding representation. The graph compares the cosine similarity between the embedded representation and the enhanced perturbation version of the embedded representation. Indicates the number of negative samples. Indicates the temperature coefficient. Indicates the first Graph embedding representation of each negative sample Representation of graphs compared to embedded representations and the first Cosine similarity between the graph embedding representations of negative samples; S48. Based on the trained Graphormer network, calculate the Euclidean distance between the center vector of each node embedding representation and the normal node embedding, and generate the node prior score.

[0031] In this embodiment, S5 specifically includes: S51. Based on the trained Graphormer network, obtain the prior scores of nodes; S52. Obtain the attention weight of each edge in the global graph, construct a node-edge joint anomaly index matrix based on the node prior score and attention weight, and filter the set of nodes whose node prior score is greater than a preset threshold. S53. Using the filtered set of nodes as the center in the global graph, extract the first-order and second-order adjacency structures associated with the abnormal nodes by combining the attention weight of each edge, and form an initial candidate subgraph set. S54. Aggregate the mean of the prior scores of all nodes in each initial candidate subgraph. S55. Sort all subgraphs by average score and attention weight density, and select subgraphs with scores higher than the subgraph threshold to form candidate subgraphs.

[0032] In this embodiment, S6 specifically includes: S61. Perform graph embedding propagation processing based on Graphormer network on the candidate subgraphs respectively; S62. Input the node feature vector and structural position encoding vector into the shared Graphormer network, perform the graph embedding propagation process, and output the local embedding representation of each node; S63. Perform pooling operation on the local embedding representation of nodes in each subgraph; S64. Based on the attention calculation results between the local embedding representations of each subgraph and the query vector, assign attention weights to each subgraph in the final fusion, and aggregate them using an attention weighting method to form a global embedding representation.

[0033] In this embodiment, S7 specifically includes: S71. Based on the global embedding representation and the local embedding representation of each node, a dual-task module is constructed. The dual-task module includes two output branches: a node-level anomaly detection subtask and an anomaly source localization subtask. S72. In the node-level anomaly detection subtask, the local embedding representation of each node is input into the fully connected network for anomaly probability prediction. S73. Based on the prediction results of all nodes, an anomaly detection loss function is constructed using binary cross-entropy loss. S74. In the anomaly source localization subtask, the global embedding representation is input to the source node prediction branch network, and the source node index distribution is calculated. The source node prediction branch network is a fully connected network. S75. Construct the cross-entropy loss function for anomaly source localization and define the total loss function, which is the weighted sum of the losses of the two sub-tasks; S76. The objective optimization dual-task module based on minimizing the total loss function outputs the anomaly detection results and anomaly source localization probability results for each node, and generates an anomaly identification report.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to the operation and maintenance data of an offshore wind farm in a coastal area and fully verified in the scenario of wind power system operation anomaly detection and source node location. This wind farm was put into operation in 2020 with an installed capacity of 500MW, consisting of 125 4MW wind turbine generators. The equipment is widely distributed, and the operating environment has high humidity and large wind speed variations. It has long suffered from typical operation and maintenance problems such as equipment communication interruption, abnormal vibration, power fluctuation, and blade pitch failure. It is a typical example of a highly complex offshore wind power system.

[0035] The wind farm deploys various types of sensor equipment during actual operation, including but not limited to anemometers, voltage and current sensors, temperature probes, vibration and acceleration sensors, and control status recording devices. Sampling is performed every 30 seconds, collecting wind speed, current, voltage, vibration, and temperature data from each wind turbine node around the clock. This also includes control signal interaction logs and communication status data between each turbine. In actual operation and maintenance, issues arise such as unstable monitoring data quality, diverse fault types, and difficulty in predicting abnormal patterns, severely impacting the efficiency and accuracy of maintenance responses.

[0036] In this scenario, the present invention first performs preprocessing operations on the collected raw operation monitoring data through a preprocessing module deployed on an edge computing server. This includes outlier removal (removing data points with mutations greater than 3 times the standard deviation), missing value filling (filling based on the mean of a sliding window), noise removal (using a five-point median filtering method), and normalization to the range of 0 to 1, thus preparing data for subsequent graph modeling.

[0037] Next, each wind turbine, electrical control cabinet, converter, etc. in the wind power system is regarded as a node in the graph. The node features are composed of five types of sensor data vectors. The physical connections (such as feeders) and logical dependencies (such as linkage control) between devices are regarded as edges. Each edge is assigned edge attributes, including control signal frequency, communication bandwidth and historical number of coordinated actions, to construct a global graph data with node attributes and edge attributes.

[0038] For the constructed original graph data, this invention designs a perturbation simulation mechanism, which performs masking on the features of some nodes, deletes and adds edges, and simultaneously perturbs the local topology to simulate real-world abnormal perturbation scenarios, generating structurally perturbed graph samples. The original graph and the perturbed graph serve as positive and negative sample pairs, which are input into the improved Graphormer network. The network internally includes modules for structure centrality encoding, edge bias modeling, multi-head attention propagation, and node representation generation. After projection head mapping, it is trained using a graph-level contrastive loss function.

[0039] After training, this invention extracts the node prior scores and attention weight information generated by Graphormer from the whole graph, dynamically extracts the local subgraph structure with potential anomaly propagation, performs local embedding propagation on each subgraph, generates the final global embedding vector through the attention weighting mechanism, and further uses it to construct a dual-task module to realize the detection of abnormal states of each node and the location and identification of anomaly propagation sources.

[0040] The experiment was conducted using 12 months of historical operation and maintenance data from the "Haidong No. 1" wind farm, from May 2022 to April 2023. For comparison and verification, three typical methods were selected as the control group: traditional isolated forest (IF), graph convolutional network (GCN), and graph attention network (GAT).

[0041] Table 1 Performance Comparison Table ; In terms of anomaly detection accuracy, this invention achieved 94.6%, which is more than 13 percentage points higher than the traditional isolated forest method and about 5.5 percentage points higher than the currently advanced Graph Attention Network (GAT). This indicates that the invention has a stronger perception ability in identifying anomalies in complex structures. At the same time, the F1 score also reached 92.7%, which is much higher than the 77.5% of isolated forest, 84.9% of GCN, and 87.4% of GAT. This shows that it has achieved a better balance between precision and recall and has stable and reliable detection performance.

[0042] In terms of anomaly source node localization, this invention achieves a localization accuracy of 91.2%, significantly higher than GCN's 74.5% and GAT's 78.9%. Meanwhile, its average localization error is only 0.9 nodes, which is nearly half the error range of GAT. This indicates that this invention has higher accuracy in identifying anomaly propagation paths and tracing source points. This performance improvement is due to the improvements made in candidate subgraph extraction, graph embedding propagation, and attention fusion mechanisms in this invention, enabling the model to more accurately model the graph structure features of anomaly propagation.

[0043] In terms of response efficiency, this invention demonstrates a significant lightweight advantage in the overall computational process. The average response time is 45.0 milliseconds, a reduction of 31.5 milliseconds compared to GCN and 18.7 milliseconds compared to GAT, indicating that it maintains high precision while also possessing excellent real-time processing capabilities. This performance is of great significance for applications requiring high reliability and rapid feedback, such as offshore wind power.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks, characterized in that, Includes the following steps: S1. Collect operational monitoring data of offshore wind power generation systems and preprocess the operational monitoring data; S2. Construct a global graph of the wind power generation system based on operation monitoring data to form original graph data with node attributes and edge attributes; S3. Design an abnormal disturbance simulation mechanism to generate a set of disturbance map sample data based on the original map data; S4. Construct a graph contrastive learning module based on the improved Graphormer network. Input the original graph data and perturbed graph sample data as positive and negative samples into the Graphormer network, train the Graphormer network through contrastive learning, and generate node prior scores. S5. Based on the node prior scores and attention weights generated by the Graphormer network, subgraph regions associated with potential anomalies are dynamically extracted from the global graph to form candidate subgraphs. S6. Perform graph embedding propagation based on the Graphormer network on each subgraph, extract local embedding representations, and generate global embedding representations by weighting attention between subgraphs; S7. Construct a dual-task module. The first task outputs the anomaly detection results for each node, and the second task outputs the location of the source node where the anomaly propagates and generates an anomaly identification report.

2. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, The operational monitoring data includes wind speed, current, voltage, vibration, and temperature data. The preprocessing includes outlier removal, missing value filling, noise removal, data normalization, and data standardization.

3. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, The global graph represents wind power equipment as nodes, connections between equipment as edges, operation monitoring data as node feature vectors, and physical connection relationships and control dependencies as edge weights.

4. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, The original graph data combines the node set, edge set, node attribute matrix, and edge attribute matrix to form graph structure input data. The node attributes are operation monitoring data, and the edge attributes include control command frequency, communication bandwidth, and number of historical collaborative actions.

5. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, The abnormal disturbance simulation mechanism includes node feature occlusion, edge connection mutation, and topology disturbance.

6. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, S3 specifically includes: S31. Perturb the node attributes in the original graph data, randomly select some nodes, and perform masking processing on the feature vectors of the selected nodes. Feature masking is achieved by setting a mask or adding random perturbation. S32. Perturb the edge connection relationships in the original graph structure, randomly select some edges to modify the connection state, including three operation methods: deleting edges, adding new edges, or modifying edge weights. The operation of deleting edges is to remove the current edge connection relationship, the operation of adding edges is to establish a new edge connection between unconnected nodes, and the operation of modifying edge weights is to perturb the attribute feature values ​​of existing edges. S33. Disrupt the topology of the graph by adjusting the connection order between nodes, exchanging local adjacency relationships, or rearranging edge combinations to simulate structural anomalies or connection failures between devices, thus forming a structurally disrupted graph. S34. Based on three types of perturbation methods—node feature perturbation, edge connection mutation, and topological structure disturbance—perturbation graph sample data is generated and used as input data for the graph comparison learning training stage.

7. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, S4 specifically includes: S41. Construct a graph contrast learning module based on an improved Graphormer network, wherein the improved Graphormer network includes a structure centrality encoding layer, a spatial bias layer, a multi-head attention mechanism, a feedforward network layer, and an output layer. S42. Construct positive and negative samples from the original graph data and the perturbation graph sample data respectively; S43. In the structural centrality coding layer, calculate the shortest path distance for each pair of nodes and calculate the structural centrality score of the nodes to generate a structural position coding vector. The structural centrality adopts the normalized betweenness centrality. ； in, Represents the structural position encoding vector. Represents a node With nodes The shortest path distance between them Represents a node Structural centrality score, Represents a node Structural centrality score; The structural location encoding vector is then injected into the node feature vector to form an enhanced node representation: ； in, Represents a node The enhanced node representation, Represents the node feature vector. Represents the structural position encoding vector. Represents a node The set of adjacent nodes; S44. In the spatial bias layer guided by edge features, attention bias terms are generated based on edge attribute vectors: ； in, Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first The edge bias mapping weight matrix of each attention head. Represents the edge attribute vector. Indicates the number of attention heads; S45. In multi-head self-attention mechanisms, augmented node representations are used. Construct query vectors and key vectors, and combine them with bias terms. Calculate attention weights: ； in, Indicates the first Nodes in each attention head With nodes Attention weights This represents the natural exponential function. Indicates the first Each attention node The query vector, Indicates the first Each attention node The key vector, This represents the dimension of the embedding vector. Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first Nodes in each attention head With nodes The bias term between them Indicates the first Each attention node The key vector, Represents a node The set of adjacent nodes, Indicates the transpose operation; S46. The results of each attention head are concatenated and input into the feedforward network. Nonlinear mapping, residual connection, and normalization are performed to generate a set of node embedding representations. ； in, The node embedding represents a set. Indicates the first Nodes in each attention head The node embedding representation, Indicates the first Nodes in each attention head With nodes Attention weights Indicates the first Nodes in each attention head value vector, Indicates the number of nodes; S47, Embedding Sets of Nodes Perform global pooling, then map it to a graph comparison embedding representation using a projection head. A contrastive loss function is constructed, and the Graphormer network is trained by minimizing the contrastive loss function to optimize the node embedding representation. The projection head is a fully connected neural network. ； in, This represents the contrastive loss function. Graph representation versus embedded representation This indicates an enhanced perturbation version of the embedding representation. The graph compares the cosine similarity between the embedded representation and the enhanced perturbation version of the embedded representation. Indicates the number of negative samples. Indicates the temperature coefficient. Indicates the first Graph embedding representation of each negative sample Representation of graphs compared to embedded representations and the first Cosine similarity between the graph embedding representations of negative samples; S48. Based on the trained Graphormer network, calculate the Euclidean distance between the center vector of each node embedding representation and the normal node embedding, and generate node prior scores.

8. The method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, S5 specifically includes: S51. Based on the trained Graphormer network, obtain the prior scores of nodes; S52. Obtain the attention weight of each edge in the global graph, construct a node-edge joint anomaly index matrix based on the node prior score and attention weight, and filter the set of nodes whose node prior score is greater than a preset threshold. S53. Using the filtered set of nodes as the center in the global graph, extract the first-order and second-order adjacency structures associated with the abnormal nodes by combining the attention weight of each edge, and form an initial candidate subgraph set. S54. Aggregate the mean of the prior scores of all nodes in each initial candidate subgraph. S55. Sort all subgraphs by average score and attention weight density, and select subgraphs with scores higher than the subgraph threshold to form candidate subgraphs.

9. A method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, S6 specifically includes: S61. Perform graph embedding propagation processing based on Graphormer network on the candidate subgraphs respectively; S62. Input the node feature vector and structural position encoding vector into the shared Graphormer network, perform the graph embedding propagation process, and output the local embedding representation of each node; S63. Perform pooling operation on the local embedding representation of nodes in each subgraph; S64. Based on the attention calculation results between the local embedding representations of each subgraph and the query vector, assign attention weights to each subgraph in the final fusion, and aggregate them using an attention weighting method to form a global embedding representation.

10. A method for detecting and locating anomalies in offshore wind power generation operation and maintenance based on graph neural networks according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the global embedding representation and the local embedding representation of each node, a dual-task module is constructed. The dual-task module includes two output branches: a node-level anomaly detection subtask and an anomaly source localization subtask. S72. In the node-level anomaly detection subtask, the local embedding representation of each node is input into the fully connected network for anomaly probability prediction. S73. Based on the prediction results of all nodes, an anomaly detection loss function is constructed using binary cross-entropy loss. S74. In the anomaly source localization subtask, the global embedding representation is input to the source node prediction branch network, and the source node index distribution is calculated. The source node prediction branch network is a fully connected network. S75. Construct the cross-entropy loss function for anomaly source localization and define the total loss function, which is the weighted sum of the losses of the two sub-tasks; S76. The objective optimization dual-task module based on minimizing the total loss function outputs the anomaly detection results and anomaly source localization probability results for each node, and generates an anomaly identification report.

Images