Wind power equipment fault diagnosis algorithm based on equipment mechanism model and GCN fusion
By combining the equipment mechanism model and graph neural network, a multimodal global topology diagram of wind power equipment is constructed, which solves the difficulty of diagnosis of traditional wind power equipment in complex data environments, and realizes accurate fault diagnosis when information is missing in the fault location. , improves diagnostic accuracy and stability.
Patent Information
- Application Number
- CN202411060326.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-19
AI Technical Summary
Traditional wind power equipment fault diagnosis algorithms are difficult to effectively diagnose faults in complex data environments, especially when information on the fault site is missing, failure characteristics cannot be accurately extracted, resulting in diagnostic errors.
The wind power equipment fault diagnosis algorithm is adopted based on the equipment mechanism model and graph neural network (GCN) to build a single-modal topology graph and graph sampling aggregation to generate multimodal global topology through vibration data acquisition, Z-Score standardization, and K-nearest neighbor algorithm. Figure, fault diagnosis is performed based on the coupling relationship between equipment components.
It can effectively diagnose wind power equipment faults without any faulty part data, improve diagnosis accuracy and stability, reduce maintenance costs, and avoid economic losses caused by diagnostic errors.
Smart Images

Figure CN118981680A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind power equipment fault diagnosis, and specifically to a wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN. Background Art
[0002] my country's wind energy and other new energy industries are booming. As the key equipment for wind power generation, wind power equipment's diagnosis and maintenance are particularly important. Wind power equipment has a complex mechanical structure, and the wind conditions change dynamically, causing the speed and load of wind power equipment to change rapidly. In harsh environments, wind power equipment may suffer from excessive loads, causing its transmission system to fail, especially bearing-based failures. During the operation of the equipment, monitoring and fault diagnosis of each component is crucial to improving operational stability and reducing maintenance costs. Timely detection and handling of faults can avoid economic losses caused by unexpected downtime.
[0003] Although traditional wind power equipment fault diagnosis algorithms can effectively perform general fault diagnosis, the actual on-site wind power equipment operation data is complex and diverse, and it is difficult to extract fault features. There may even be cases where sensors fail to report faults, resulting in missing fault information, which makes it impossible to effectively diagnose faults. Therefore, it is very necessary to use the equipment mechanism model and GCN to effectively diagnose wind power equipment in the case of complex data and missing fault location information. Summary of the invention
[0004] Purpose of the invention: To realize a knowledge-driven wind power equipment fault diagnosis method by combining rules and learning. A wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN is proposed, which can achieve good fault diagnosis effect with or without fault location data.
[0005] To achieve the above purpose, the technical solution provided by the present invention is: a wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN, comprising the following steps:
[0006] S1: Collect vibration data of actual operation of various parts of wind power equipment;
[0007] S2: The original dataset was normalized using Z-Score standardization;
[0008] S3: Use K-nearest neighbor algorithm to construct a single-mode topological map of the fault location;
[0009] S4: merging the single-modal topological graphs of each part into nodes using the graph sampling aggregation algorithm;
[0010] S5: Generate a path diagram through the coupling relationship between equipment components, and build a multi-modal global topology diagram of wind power equipment for each component node in combination with distance measurement;
[0011] S6: Input the constructed global topology map into GCN for wind power fault diagnosis.
[0012] Furthermore, the specific method of step S2 is as follows:
[0013] S2.1: Construct a vibration data set. Filter the data of the rear bearing of the generator from the wind power equipment, divide the sample set, and construct the original data set X = {x1, x2, ..., x n};
[0014] S2.2: Calculate the mean of the data and standard deviation
[0015] S2.3: Standardize the data samples of the original data set according to the mean and standard deviation;
[0016] S2.4: Return the normalized dataset
[0017] Furthermore, the specific method of step S3 is as follows:
[0018] S3.1: Calculate the similarity matrix of the feature matrix. For the K-dimensional node feature vector obtained by the feature extractor, take N nodes as a group and record the node feature matrix as F∈R N×K , calculate the inner product of different feature vectors and obtain the similarity matrix S of the node feature matrix:
[0019] S=F×F T ,S∈R N×N ;
[0020] S3.2: Normalize the similarity matrix. Process the similarity matrix and normalize each column to the maximum value to obtain a normalized similarity matrix. The elements of the similarity matrix are denoted as S ij , i and j represent row and column indices respectively. First, get the maximum value of each row to get the vector M, which is:
[0021] M i =max j (S ij );
[0022] Then, the maximum value of each row is used for normalization to obtain the normalized similarity matrix Specifically:
[0023]
[0024] S3.3: Use the nearest neighbor to select node adjacent edges. Set the K value of the KNN algorithm. According to the original adjacency matrix, each node selects the nearest K neighbor nodes, retains the adjacent edges, deletes the remaining edges, and obtains a new adjacency matrix. The distance values in the original adjacency matrix can be used as the edge weights of the new adjacency matrix, specifically:
[0025]
[0026] For each row of the normalized similarity matrix, find the K values with the highest similarity in this row and their indexes, specifically:
[0027]
[0028] Then we get the most similar node index set T of all rows, and finally get the adjacency matrix A, which is:
[0029]
[0030] S3.4: Normalize the adjacency matrix. To avoid missing node information and prevent the information of different nodes from being too different, which may cause gradient explosion or disappearance during training, the adjacency matrix needs to be normalized. Specifically:
[0031]
[0032] in D is the degree matrix of the graph;
[0033] S3.5: Construct a single-mode topology map of the fault location. The features are measured by similarity and the KNN algorithm to generate an adjacency matrix, add features that characterize the relationship between samples, and finally form a graph structure data, namely, a single-mode topology map G of the fault location.
[0034] Furthermore, the specific method of step S4 is as follows:
[0035] S4.1: Neighbor node sampling. Process the K-order neighbors and set the depth K to determine how many orders of neighbors the graph sampling aggregation function samples.
[0036] S4.2: Feature aggregation. For each node v, the feature representations of its neighboring nodes are aggregated into The representation of each neighbor node is generated by the previous layer when it is sampled, and the original node feature x is used for the first sampling. v , aggregate neighbor information;
[0037] S4.3: Node embedding generation. The graph sampling aggregation algorithm represents the current node With neighborhood representation After concatenation and nonlinear activation function, it is transformed into a new node representation
[0038] Furthermore, the specific method of step S5 is as follows:
[0039] S5.1: Obtain multi-modal vibration data and convert each modal data into a local single-modal topology map;
[0040]
[0041] S5.2: Using the graph sampling aggregation method, aggregate each single-modal topological graph into a node representation of each modality;
[0042] Z i =([z1,z2,K,z v ],v∈V)=graphsage(G i,L (V,E));
[0043] S5.3: Based on the device model and the sensor arrangement relationship, obtain a path diagram representing the node relationship;
[0044]
[0045] S5.4: Perform distance measurement on each modal node and obtain the edge relationship between nodes;
[0046]
[0047] S5.5: The edge relationship based on distance metric and the edge relationship based on structure are combined as an adjacency matrix to obtain the global topology of each device;
[0048] BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 is a flow chart of a method in an embodiment of the present invention;
[0050] Figure 2 is a vibration signal diagram collected from various parts of the wind power equipment in the embodiment of the present invention;
[0051] Figure 3 is a comparison of vibration data before and after normalization in an embodiment of the present invention;
[0052] Figure 4 This is a process for generating a single-mode topology map of a fault location in an embodiment of the present invention;
[0053] Figure 5 is an example graph of graph sampling aggregation in an embodiment of the present invention;
[0054] Figure 6The single-mode topology diagrams of each part in the embodiment of the present invention are fused into a node schematic diagram;
[0055] Figure 7 is a path diagram generated according to the coupling relationship between device components in an embodiment of the present invention;
[0056] Figure 8 It is a fault diagnosis algorithm diagram based on a multi-modal global topology diagram in an embodiment of the present invention. DETAILED DESCRIPTION
[0057] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0058] like Figure 1 As shown, the wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN disclosed in the embodiment of the present invention includes the following steps:
[0059] S1: Collect vibration data of actual operation of various parts of wind power equipment;
[0060] Targeted vibration signal monitoring is carried out on the main shaft, gearbox and main bearings and gears of the engine of the wind turbine. A total of 7 vibration sensor measuring points are installed, which are located at the front bearing of the main shaft, the rear bearing of the main shaft, the first gear of the gearbox, the second gear of the gearbox, the high-speed shaft of the gearbox, the front bearing of the generator and the rear bearing of the generator. Each measuring point is installed in the radial direction of the bearing / gear ring (vertical direction of the bearing and gear ring). At the same time, the vibration data of each point of the whole wind turbine equipment is collected during operation. The vibration signals collected from various parts of the wind turbine equipment are referenced Figure 2 .
[0061] Among the 7 sensor measurement points, the sampling frequency at the main shaft front bearing and the main shaft rear bearing is 12800Hz, and the collection time is 10.24 seconds. The sampling frequency at the other measurement points is 25600Hz, and the collection time is 5.12 seconds. The length of the single data collection is 131072 points. A total of 8 sets of data were collected at each measurement point of each device at different times to record the vibration status of the wind turbine equipment under different operating conditions.
[0062] S2: Use Z-Score standardization to normalize the original data set. For reference, see the comparison chart before and after vibration data normalization. Figure 3 ;
[0063] S2.1: Construct a vibration data set. Select the data of the rear bearing of the generator from 33 wind turbines, divide the sample set, and construct the original data set X = {x1, x2, ..., x n};
[0064] S2.2: Calculate the mean of the data and standard deviation
[0065] S2.3: Standardize the data samples of the original data set according to the mean and standard deviation;
[0066] S2.4: Return the normalized dataset
[0067] S3: Use the K-nearest neighbor algorithm to construct a single-mode topology map of the fault location. The single-mode topology map generation process of the fault location is referenced in Figure 4 ;
[0068] S3.1: Calculate the similarity matrix of the feature matrix. For the K-dimensional node feature vector obtained by the feature extractor, take N nodes as a group and record the node feature matrix as F∈R N×K , calculate the inner product of different feature vectors and obtain the similarity matrix S of the node feature matrix:
[0069] S=F×F T ,S∈R N×N ;
[0070] S3.2: Normalize the similarity matrix. Process the similarity matrix and normalize each column to the maximum value to obtain a normalized similarity matrix. The elements of the similarity matrix are denoted as S ij , i and j represent row and column indices respectively. First, get the maximum value of each row to get the vector M, which is:
[0071] M i =max j (S ij )
[0072] Then, the maximum value of each row is used for normalization to obtain the normalized similarity matrix Specifically:
[0073]
[0074] S3.3: Use the nearest neighbor to select node adjacent edges. Set the K value of the KNN algorithm. According to the original adjacency matrix, each node selects the nearest K neighbor nodes, retains the adjacent edges, deletes the remaining edges, and obtains a new adjacency matrix. The distance values in the original adjacency matrix can be used as the edge weights of the new adjacency matrix, specifically:
[0075]
[0076] For each row of the normalized similarity matrix, find the K values with the highest similarity in this row and their indexes, specifically:
[0077]
[0078] Then we get the most similar node index set T of all rows, and finally get the adjacency matrix A, which is:
[0079]
[0080] S3.4: Normalize the adjacency matrix. To avoid missing node information and prevent the information of different nodes from being too different, which may cause gradient explosion or disappearance during training, the adjacency matrix needs to be normalized. Specifically:
[0081]
[0082] in D is the degree matrix of the graph;
[0083] S3.5: Construct a single-mode topology map of the fault location. The features are measured by similarity and the KNN algorithm to generate an adjacency matrix, add features that characterize the relationship between samples, and finally form a graph structure data, namely, a single-mode topology map G of the fault location.
[0084] S4: merging the single-modal topological graphs of each part into nodes using the graph sampling aggregation algorithm;
[0085] S4.1: Neighbor node sampling. Process the K-order neighbors and set the depth K to determine how many orders of neighbors the graph sampling aggregation function samples.
[0086] S4.2: Feature aggregation. For each node v, the feature representations of its neighboring nodes are aggregated into The representation of each neighbor node is generated by the previous layer when it is sampled, and the initial sampling is the original node feature x v , aggregate neighbor information;
[0087] S4.3: Node embedding generation. The graph sampling aggregation algorithm represents the current node With neighborhood representation After concatenation and nonlinear activation function, it is transformed into a new node representation
[0088] Graph Sampling Aggregation Example Graph Reference Figure 5 , the single-mode topology graphs of each part are fused into node reference Figure 6 .
[0089] S5: Generate a path diagram through the coupling relationship between equipment components, and build a multi-modal global topology diagram of wind power equipment for each component node in combination with distance measurement;
[0090] S5.1: Obtain multi-modal vibration data and convert each modal data into a local single-modal topology map;
[0091]
[0092] S5.2: Using the graph sampling aggregation method, aggregate each single-modal topological graph into a node representation of each modality;
[0093] Z i =([z1,z2,K,z v ],v∈V)=graphsage(G i,L (V,E));
[0094] S5.3: Based on the relationship between the device model and the sensor arrangement, obtain a path diagram representing the node relationship. The path diagram generated by the coupling relationship between device components is referenced Figure 7 ;
[0095]
[0096] S5.4: Perform distance measurement on each modal node and obtain the edge relationship between nodes;
[0097]
[0098] S5.5: The edge relationship based on distance metric and the edge relationship based on structure are combined as an adjacency matrix to obtain the global topology of each device;
[0099]
[0100] S6: Input the constructed global topology map into GCN for wind power fault diagnosis. The fault diagnosis algorithm based on multi-modal global topology map is shown in Figure 2. Figure 8 .
Claims
1. A wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN, characterized by: The following steps are involved: S1: Collect vibration data of actual operation of various parts of wind power equipment; S2: The original dataset was normalized using Z-Score standardization; S3: Use K-nearest neighbor algorithm to construct a single-mode topological map of the fault location; S4: merging the single-modal topological graphs of each part into nodes using the graph sampling aggregation algorithm; S5: Generate a path diagram through the coupling relationship between equipment components, and build a multi-modal global topology diagram of wind power equipment for each component node in combination with distance measurement; S6: Input the constructed global topology map into GCN for wind power fault diagnosis.
2. The wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN according to claim 1 is characterized in that: The step S2 comprises: S2.1: Construct a vibration data set. Filter the data of the rear bearing of the generator from the wind power equipment, divide the sample set, and construct the original data set X = {x1, x2, ..., x n }; S2.2: Calculate the mean of the data and standard deviation S2.3: Standardize the data samples of the original data set according to the mean and standard deviation; S2.4: Return the normalized dataset 3. The wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN according to claim 1 is characterized in that: The step S3 comprises: S3.1: Calculate the similarity matrix of the feature matrix. For the K-dimensional node feature vector obtained by the feature extractor, take N nodes as a group and record the node feature matrix as F∈R N×K , calculate the inner product of different feature vectors and obtain the similarity matrix S of the node feature matrix: S=F×F T ,S∈R N×N ; S3.2: Normalize the similarity matrix. Process the similarity matrix and normalize each column to the maximum value to obtain a normalized similarity matrix. The elements of the similarity matrix are denoted as S ij , i and j represent row and column indices respectively. First, get the maximum value of each row to get the vector M, which is: M i =max j (S ij ); Then, the maximum value of each row is used for normalization to obtain the normalized similarity matrix Specifically: S3.3: Use the nearest neighbor to select node adjacent edges. Set the K value of the KNN algorithm. According to the original adjacency matrix, each node selects the nearest K neighbor nodes, retains the adjacent edges, deletes the remaining edges, and obtains a new adjacency matrix. The distance values in the original adjacency matrix can be used as the edge weights of the new adjacency matrix, specifically: For each row of the normalized similarity matrix, find the K values with the highest similarity in this row and their indexes, specifically: Then we get the most similar node index set T of all rows, and finally get the adjacency matrix A, which is: S3.4: Normalize the adjacency matrix. To avoid missing node information and prevent the information of different nodes from being too different, which may cause gradient explosion or disappearance during training, the adjacency matrix needs to be normalized. Specifically: in D is the degree matrix of the graph; S3.5: Construct a single-mode topology map of the fault location. The features are measured by similarity and the KNN algorithm to generate an adjacency matrix, add features that characterize the relationship between samples, and finally form a graph structure data, namely, a single-mode topology map G of the fault location.
4. The wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN according to claim 1 is characterized in that: The step S4 comprises: S4.1: Neighbor node sampling. Process the K-order neighbors and set the depth K to determine how many orders of neighbors the graph sampling aggregation function samples. S4.2: Feature aggregation. For each node v, the feature representations of its neighboring nodes are aggregated into The representation of each neighbor node is generated by the previous layer when it is sampled, and the original node feature x is used for the first sampling. v , aggregate neighbor information; S4.3: Node embedding generation. The graph sampling aggregation algorithm represents the current node With neighborhood representation After concatenation and nonlinear activation function, it is transformed into a new node representation 5. The wind power equipment fault diagnosis algorithm based on the fusion of equipment mechanism model and GCN according to claim 1 is characterized in that: The step S5 comprises: S5.1: Obtain multi-modal vibration data and convert each modal data into a local single-modal topology map; S5.2: Using the graph sampling aggregation method, aggregate each single-modal topological graph into a node representation of each modality; Z i =([z1,z2,K,z v ],v∈V)=graphsage(G i,L (V,E)); S5.3: Based on the device model and the sensor arrangement relationship, obtain a path diagram representing the node relationship; S5.4: Perform distance measurement on each modal node and obtain the edge relationship between nodes; S5.5: The edge relationship based on distance metric and the edge relationship based on structure are combined as an adjacency matrix to obtain the global topology of each device;
Citation Information
Patent Citations
Wind turbine generator fault diagnosis method based on multi-view neural network
CN115272811A
Multi-source data anomaly monitoring system based on digital twinning
CN116049766A
Task-oriented and view-angle-invariable multi-mode fault diagnosis method and system
CN116909788A
Chemical production line state identification method and system
CN117195114A
