Abnormity identification method for wind driven generator

By combining the data collection and monitoring control system with the SSGC-GAT model, the problem of lack of real-time anomaly monitoring for wind turbine generators has been solved, enabling safe and reliable operation of wind turbine generators and improving the accuracy of anomaly identification.

CN118916607AInactive Publication Date: 2024-11-08WENZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410112947.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-01-26
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The lack of real-time anomaly monitoring in existing wind turbine generators increases the risk of equipment damage, energy loss, and safety accidents.

Method used

A data collection and monitoring control system is adopted, which cleans the data using the isolated forest algorithm, performs data preprocessing and feature extraction using the SSGC-GAT model, and uses graph attention network for anomaly identification. The process includes data collection, labeling, preprocessing, set partitioning, collaborative similarity graph construction and model training to achieve anomaly identification of wind turbine units.

Benefits of technology

It improves the accuracy of wind turbine anomaly identification, ensures the safe and reliable operation of wind turbines, and enables real-time monitoring and identification of abnormal situations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916607A_ABST
    Figure CN118916607A_ABST
Patent Text Reader

Abstract

The invention provides a wind driven generator anomaly recognition method. The wind driven generator anomaly recognition method comprises the following steps of data collection, data marking, data preprocessing, data set division, collaborative similar graph construction (SSGC) strategy establishment, SSGC-GAT model establishment, SSGC-GAT anomaly recognition model training and SSGC-GAT anomaly recognition model testing. Compared with the prior art, the wind turbine generator abnormity identification method based on the graph attention network is provided based on the data acquisition and monitoring control system, and the accuracy of abnormity identification is effectively improved through the provided method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the application of deep learning methods and anomaly detection technologies in the field of wind turbine anomaly identification, and in particular to a wind turbine anomaly identification method. Background Art

[0002] As an important source of clean energy, wind power generation has been widely developed and applied around the world. However, various abnormal conditions may occur during the operation of wind turbines, such as blade damage, mechanical failure, electrical failure, etc. These abnormal conditions may cause equipment damage, energy loss, and even safety accidents. Therefore, it is very important to monitor the status of wind turbines in real time to ensure the safe and reliable operation of wind turbines. Summary of the invention

[0003] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide a method for identifying abnormalities of a wind turbine generator, so as to solve the problem in the prior art that abnormalities of a wind turbine generator set during operation are not monitored in real time.

[0004] To achieve the above-mentioned purpose and other related purposes, the present invention provides a method for identifying abnormality of a wind turbine, comprising:

[0005] Data collection, collecting wind power data with several parameter time series through the data acquisition and monitoring control system to form the initial wind turbine data set;

[0006] Data labeling: cleaning the initial wind turbine data set by using an isolation forest algorithm to separate positive and abnormal data;

[0007] Data preprocessing, performing missing value filling and normalization data preprocessing on the positive and abnormal data to obtain preprocessed data;

[0008] Data set division, dividing the preprocessed data into training data and test data;

[0009] A collaborative similarity graph construction strategy generates multiple adjacency matrices respectively through four distance measurement methods: k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance, and minimum spanning tree. The optimal adjacency matrix is ​​generated using the similarity between the multiple adjacency matrices. The optimal adjacency matrix is ​​combined with the feature matrix to obtain graph data.

[0010] SSGC-GAT model, the GAT model consists of an input layer, a hidden layer and an output layer, and the graph data is used as the input of the GAT model to obtain the SSGC-GAT model;

[0011] SSGC-GAT anomaly recognition model training, inputting the training data into the SSGC-GAT model for training to obtain a trained SSGC-GAT model;

[0012] SSGC-GAT anomaly recognition model test, input the test data into the trained SSGC-GAT model, obtain the probability distribution of each label of each SSGC-GAT model, and identify the anomaly results based on the SSGC-GAT model.

[0013] Optionally, in the data collection, characteristic parameters of wind speed, power, wind direction and temperature are selected from the wind power data.

[0014] Optionally, in the data labeling, the initial fan data set is cleaned to separate normal samples and abnormal samples, the normal samples are sampled using the Prototype Generation method, and the same amount of data as the abnormal samples is selected as the normal samples, and the selected normal samples and abnormal samples are the positive abnormal data.

[0015] Optionally, in the data preprocessing, the missing values ​​are filled using the mean filling method. The data is normalized, where x′ represents the normalized data, x represents the original data, and x max and x min Represent the maximum and minimum values ​​of the original data respectively.

[0016] Optionally, in the data set division, the training data accounts for 80% of the total number of samples, and the test data accounts for 20% of the total number of samples.

[0017] Optionally, the collaborative similarity graph construction strategy includes:

[0018] Generate multiple adjacency matrices using four distance metrics: k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance, and minimum spanning tree;

[0019] The distance between the adjacency matrices is calculated using the Frobenius norm and converted into similarity and stored in the weight matrix W′;

[0020] Normalizing the weight matrix W′;

[0021] Loop through each of the adjacency matrices, multiply the adjacency matrix by the weight in the corresponding weight matrix W', and add the results to obtain a fusion matrix F;

[0022] The fusion matrix F is converted into a binary matrix using the median thresholding method.

[0023] Optionally, in calculating the distance between the adjacency matrices using the Frobenius norm, the distance d between the adjacency matrices ij =||A i -A j || F .

[0024] Optionally, the construction of the SSGC-GAT model includes:

[0025] Calculate the attention coefficient of a single node i, assuming it has a neighboring point j∈N i , after linear transformation, they are WH i and WH j , then the attention coefficient of the neighboring point to the node i is: e ij =a([WH i ||WH j ]), j∈N i , where W is the weight matrix;

[0026] The nonlinear activation function is added to the linearly changed value: LeakyReLU(·), and the final attention coefficient calculation formula is: α ij =LeakyReLU(e ij );

[0027] Perform weighted summation on the adjacent point features to obtain new node features Among them, H i is the output of the graph attention layer, σ(·) is the activation function, and the single-layer graph attention layer features are obtained;

[0028] For the multi-head attention mechanism, k independent attention mechanisms are used to obtain feature information of different subspaces; in each graph attention layer, k independent transformation matrices W are calculated (k) ; According to the previous step, the attention information of multiple subspaces is calculated in parallel and spliced ​​together, and the obtained results are spliced ​​again to obtain in is through the kth transformation matrix W (k) The calculated weight coefficient;

[0029] Aggregate multiple graph attention layers to obtain Where l represents the number of network layers;

[0030] Define the feature of each sample group as X, and define the graph embedding representation as

[0031] Optionally, the SSGC-GAT anomaly recognition model training includes the following steps:

[0032] Parameter setting, set the number of attention heads of the attention mechanism to a and b respectively, and set the batch size n batch , the maximum number of training iterations n epoch .

[0033] The output layer uses the sigmoid function to classify the results, that is, y = sigmoid (WZ), where W is a trainable model parameter;

[0034] Input the divided training data into the SSGC-GAT model for training to obtain the recognition results;

[0035] Calculate the loss function of the model Where N is the number of samples, y i represents the true label value of the i-th sample, x i Represents the output of the softmax classifier;

[0036] Update basic model parameters;

[0037] The training process is repeated until the maximum number of iterations is reached, and finally the trained SSGC-GAT model is obtained.

[0038] Optionally, during parameter setting, the Dropout method is used to avoid overfitting, and the Dropout rate is set to 0.1.

[0039] In the scheme implemented by the above-mentioned method for identifying abnormalities of wind turbines, wind power data of several parameter time series are collected in real time, and the data is analyzed to determine the operating status of the wind turbine, and abnormal situations that occur are discovered and identified in time to ensure the safe operation of the wind turbine. The present invention proposes a method for identifying abnormalities of wind turbines based on a graph attention network, which uses the good feature self-learning characteristics of the graph attention network to identify possible abnormalities of the wind turbine, thereby improving the safety and reliability of the operation of the wind turbine. Compared with the prior art, this scheme is based on a data acquisition and monitoring control system, and proposes a method for identifying abnormalities of wind turbines based on a graph attention network. The proposed method effectively improves the accuracy of abnormality identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Shown is a schematic diagram of an application of a method for identifying abnormality of a wind turbine according to an embodiment of the present invention;

[0041] Figure 2 The display is a wind power data processing result diagram of an embodiment of the present invention;

[0042] Figure 3 Shown is a confusion matrix diagram of the application result verification of a method for identifying abnormalities of a wind turbine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.

[0044] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0045] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0046] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0047] In an exemplary embodiment, Figure 1 As shown, a method for identifying abnormality of a wind turbine generator includes at least steps S210 to S260, which are described in detail as follows:

[0048] Step S210: data collection, collecting wind power data with a number of parameter time series through a data acquisition and monitoring control system to form an initial wind turbine data set;

[0049] In some embodiments, during the data collection, characteristic parameters of wind speed, power, wind direction, and temperature are selected from the wind power data.

[0050] Specifically, the data collected in this solution comes from the Penmanshiel wind farm. The total data includes 14 wind turbines, and the collection time is from January 1, 2021 to July 1, 2021; 14 wind turbines have a total of 364,910 data and 80 characteristic parameters. Through the correlation analysis of the characteristic parameters of wind power data, 34 characteristic parameters such as wind speed, power, wind direction, and temperature are selected as the sample set. Specifically, taking the characteristic parameters of wind speed and power as an example, the wind power data of wind speed and power time series are collected and Figure 2 In the display, Figure 2 The upper left graph is the power processing result graph, the lower left graph is the wind speed processing result graph, and the right graph is the relationship between power and wind speed.

[0051] Step S220: data labeling, cleaning the initial wind turbine data set by using an isolation forest algorithm to separate positive and abnormal data;

[0052] In some embodiments, in the data labeling, the initial fan data set is cleaned to separate normal samples and abnormal samples, the normal samples are sampled using the Prototype Generation method, and the same amount of data as the abnormal samples is selected as the normal samples, and the selected normal samples and abnormal samples are the positive abnormal data.

[0053] Specifically, the total samples are separated into 351,550 normal samples and 13,360 abnormal samples, and all abnormal samples are selected as experimental samples. Considering that the number of normal samples far exceeds the number of abnormal samples, this will lead to poor classification of abnormal samples. We use the Prototype Generation method to sample normal samples and select the same amount of data as abnormal samples as normal samples.

[0054] Step S230: data preprocessing, performing missing value filling and normalization data preprocessing on the positive and abnormal data to obtain preprocessed data;

[0055] In some embodiments, in the data preprocessing, the missing values ​​are filled using the mean filling method. The data is normalized, where x′ represents the normalized data, x represents the original data, and x max and x min Represent the maximum and minimum values ​​of the original data respectively.

[0056] Step S240: dividing the data set into training data and test data;

[0057] In some embodiments, in the data set division, the training data accounts for 80% of the total number of samples, and the test data accounts for 20% of the total number of samples.

[0058] Step S250: A collaborative similarity graph construction (SSGC) strategy is used to generate four adjacency matrices using four distance measurement methods, namely, k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance, and minimum spanning tree. An optimal adjacency matrix is ​​generated using the similarity between the four adjacency matrices. The optimal adjacency matrix is ​​combined with the feature matrix to obtain graph data.

[0059] It should be noted that Synergistic Similarity·Graph.Construction is the English translation of Synergistic Similarity·Graph.Construction, which is usually abbreviated as SSGC in professional terminology; the graph data generated by this method is a similarity graph.

[0060] In some embodiments, the collaborative similarity graph construction (SSGC) strategy includes the following steps:

[0061] Step S251: multiple adjacency matrices are obtained by using four distance measurement methods: k nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance, and minimum spanning tree;

[0062] Step S252: Calculate the distance d between matrices using the Frobenius norm ij =||A i -A j || F , and converted into similarity and stored in the weight matrix W′;

[0063] Step S253: normalizing the weight matrix W′;

[0064] Step S254: loop through each adjacency matrix, multiply it by the weight in the corresponding normalized weight W′, and add the results to obtain the fusion matrix F;

[0065] Step S255: using the median thresholding method, the fusion matrix F is converted into a binary matrix, and the binary matrix is ​​used to express the optimal adjacency matrix;

[0066] Step S256: The optimal adjacency matrix is ​​combined with the feature matrix to obtain graph data, and a graph structure G = (V, E) is specifically defined, where V = {v1, v2, ..., v n} is a node set, E={e1,e2,...,e m} is the edge set.

[0067] Step S260: SSGC-GAT model, the GAT model consists of an input layer, a hidden layer and an output layer, and the graph data is used as the input of the GAT model to obtain the SSGC-GAT model;

[0068] Specifically, the number of attention heads in the input layer and output layer are 15 and 1 respectively, and the number of hidden layers is 130.

[0069] In some embodiments, SSGC-GAT model construction comprises the following steps:

[0070] Step S261: Calculate the attention coefficient of a single node i, assuming it has a neighboring point j∈N i , after linear transformation, they are WH i and WH j , then the attention coefficient of the neighboring point to the node i is: e ij =a([WH i ||WH j ]), j∈N i , where W is the weight matrix;

[0071] Step S262: Add a nonlinear activation function to the linearly changed value: LeakyReLU(·), and the final attention coefficient calculation formula is: α ij =LeakyReLU(e ij );

[0072] Step S263: Perform weighted summation on neighbor node features to obtain new node features Among them, H i is the output of the graph attention layer, σ(·) is the activation function, and the single-layer graph attention layer features are obtained;

[0073] Step S264: For the multi-head attention mechanism, use k independent attention mechanisms to obtain feature information of different subspaces; in each graph attention layer, calculate k independent transformation matrices W (k) According to step 3, the attention information of multiple subspaces is calculated in parallel, and then they are spliced ​​together, and then the results are spliced ​​again to obtain in is through the kth transformation matrix W (k) The calculated weight coefficient;

[0074] Step S265: Aggregate multiple graph attention layers to obtain Where l represents the number of network layers;

[0075] Step S266: define the feature of each sample group as X, and define the graph embedding representation as Z = σ(XH I(l) ).

[0076] The above steps are mainly used to construct the hidden layer in the GAT model, wherein the input layer can be input by the graph data in step S250, and the output layer is reflected in the next step.

[0077] Step S270: SSGC-GAT anomaly recognition model training, inputting the training data into the SSGC-GAT model for training, and obtaining a trained SSGC-GAT model;

[0078] In some implementations, the SSGC-GAT anomaly recognition model training includes the following steps:

[0079] Step S271: Parameter initialization; parameter setting, set the number of attention heads of the attention mechanism to 15 and 1 respectively, and set the batch size n batch =42, maximum number of training iterations n epoch =300, and the Dropout method is used to avoid overfitting. The Dropout rate is set to 0.1.

[0080] Step S272: The output layer uses the sigmoid function to classify the results, that is, y=sigmoid(WZ), where W is a trainable model parameter.

[0081] Step S273: Input the divided training data into the SSGC-GAT model for training to obtain the recognition result.

[0082] Step S274: Calculate the loss function of the model Where N is the number of samples, y i represents the true label value of the i-th sample, x i Represents the output of the softmax classifier.

[0083] Step S275: Update basic model parameters.

[0084] Step S276: Repeat steps S273 to S275 until the maximum number of iterations is reached, and finally a trained SSGC-GAT model is obtained.

[0085] Step S280: SSGC-GAT anomaly recognition model test, input the test data into the trained SSGC-GAT model, obtain the probability distribution of each label of each SSGC-GAT model, and identify the anomaly result based on the SSGC-GAT model.

[0086] The abnormal identification of wind turbines according to the above steps is verified, and the verification is displayed using a confusion matrix, as shown in Figure 3As shown, it can be concluded that the wind turbine abnormality identification method in the scheme has a high accuracy rate in identifying abnormalities of wind turbines.

[0087] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for identifying abnormality of a wind turbine generator. It is characterized in that The following steps are involved: Data collection, collecting wind power data with several parameter time series through the data acquisition and monitoring control system to form the initial wind turbine data set; Data labeling: cleaning the initial wind turbine data set by using an isolation forest algorithm to separate positive and abnormal data; Data preprocessing, performing missing value filling and normalization data preprocessing on the positive and abnormal data to obtain preprocessed data; Data set division, dividing the preprocessed data into training data and test data; A collaborative similarity graph construction strategy generates multiple adjacency matrices respectively through four distance measurement methods: k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance, and minimum spanning tree. The optimal adjacency matrix is ​​generated using the similarity between the multiple adjacency matrices. The optimal adjacency matrix is ​​combined with the feature matrix to obtain graph data. SSGC-GAT model, the GAT model consists of an input layer, a hidden layer and an output layer, and the graph data is used as the input of the GAT model to obtain the SSGC-GAT model; SSGC-GAT anomaly recognition model training, inputting the training data into the SSGC-GAT model for training to obtain a trained SSGC-GAT model; SSGC-GAT anomaly recognition model test, input the test data into the trained SSGC-GAT model, obtain the probability distribution of each label of each SSGC-GAT model, and identify the anomaly results based on the SSGC-GAT model.

2. A method for identifying abnormality of a wind turbine according to claim 1, Features: In the data collection, characteristic parameters of wind speed, power, wind direction and temperature are selected from the wind power data.

3. A method for identifying abnormality of a wind turbine according to claim 2, Features: In the data labeling, the initial fan data set is cleaned to separate normal samples and abnormal samples, the Prototype Generation method is used to sample the normal samples, and the same amount of data as the abnormal samples is selected as the normal samples. The selected normal samples and abnormal samples are the positive abnormal data.

4. A method for identifying abnormality of a wind turbine according to claim 1, Features: In the data preprocessing, the mean filling method is used to fill the missing values. The data is normalized, where x′ represents the normalized data, x represents the original data, and x max and x min Represent the maximum and minimum values ​​of the original data respectively.

5. A method for identifying abnormality of a wind turbine according to claim 4, Features: In the data set division, the training data accounts for 80% of the total number of samples, and the test data accounts for 20% of the total number of samples.

6. A method for identifying abnormality of a wind turbine according to claim 1, Features: The collaborative similarity graph construction strategy includes: Generate multiple adjacency matrices using four distance metrics: k-nearest neighbor, threshold-based cosine similarity, threshold-based Mahalanobis distance, and minimum spanning tree; The distance between the adjacency matrices is calculated using the Frobenius norm and converted into similarity and stored in the weight matrix W'; Normalizing the weight matrix W′; Loop through each of the adjacency matrices, multiply the adjacency matrix by the weight in the corresponding weight matrix W′, and add the results to obtain a fusion matrix F; The fusion matrix F is converted into a binary matrix using the median thresholding method.

7. A method for identifying abnormality of a wind turbine according to claim 6, Features: In calculating the distance between the adjacency matrices using the Frobenius norm, the distance d between the adjacency matrices is ij =||A i -A j || F .

8. A method for identifying abnormality of a wind turbine according to claim 7, Features: The construction of the SSGC-GAT model includes: Calculate the attention coefficient of a single node i, assuming it has a neighboring point j∈N i , after linear transformation, they are WHi and WHj respectively, then the attention coefficient of the neighboring point to the node i is: e ij =a([WH i ||WH j ]), j∈N i , where W is the weight matrix; The nonlinear activation function is added to the linearly changed value: LeakyReLU(·), and the final attention coefficient calculation formula is: α ij =LeakyReLU(e ij ); Perform weighted summation on the adjacent point features to obtain new node features Among them, H i is the output of the graph attention layer, σ(·) is the activation function, and the single-layer graph attention layer features are obtained; For the multi-head attention mechanism, k independent attention mechanisms are used to obtain feature information of different subspaces; in each graph attention layer, k independent transformation matrices W are calculated (k) ; According to the previous step, the attention information of multiple subspaces is calculated in parallel and spliced ​​together, and the obtained results are spliced ​​again to obtain in is through the kth transformation matrix W (k) The calculated weight coefficient; Aggregate multiple graph attention layers to obtain Where l represents the number of network layers; Define the feature of each sample group as X, and define the graph embedding representation as 9. A method for identifying abnormality of a wind turbine according to any one of claims 1 to 8, Features: The SSGC-GAT anomaly recognition model training includes the following steps: Parameter setting, set the number of attention heads of the attention mechanism to a and b respectively, and set the batch size n batch , the maximum number of training iterations n epoch ; The output layer uses the sigmoid function to classify the results, that is, y = sigmoid (WZ), where W is a trainable model parameter; Input the divided training data into the SSGC-GAT model for training to obtain the recognition results; Calculate the loss function of the model Where N is the number of samples, y i represents the true label value of the i-th sample, x i Represents the output of the softmax classifier; Update basic model parameters; The training process is repeated until the maximum number of iterations is reached, and finally the trained SSGC-GAT model is obtained.

10. A method for identifying abnormality of a wind turbine according to claim 9, Features: During parameter setting, the Dropout method is used to avoid overfitting, and the Dropout rate is set to 0.1.