Wind turbine key equipment measurement point high-dimensional abnormal operation data online identification method

Abstract

The application provides an offshore wind turbine key equipment measuring point high-dimensional abnormal operation data online identification method, based on analysis of historical data sets, the application can identify online data, the application can identify high-dimensional abnormal operation data of offshore wind turbine key component measuring points online, after deleting or correcting abnormal data, the application can avoid affecting the conclusion of data analysis due to abnormal operation data, and through modeling of the corrected data, the application can improve model precision and improve AI decision precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine operation monitoring, in particular to a method for online identification of high-dimensional abnormal operation data of key equipment measuring points of a wind turbine. BACKGROUND

[0002] A large amount of wind turbine operation data is generated during the operation of a wind farm. Due to operation faults, data transmission, etc., data aggregation inevitably produces abnormal values. Accurate identification of abnormal values is crucial to ensure the accuracy of model calculation results.

[0003] Domestic and foreign scholars focus on the joint identification of wind speed-power curves for cleaning abnormal data of wind turbines, and have conducted a large amount of research on the identification of abnormal wind power data. However, there is little research on the identification of abnormal operation data of other key component measuring points. The main methods include statistical methods and machine learning methods. Statistical methods include quartile method, 3σ method, and absolute median deviation method, etc. The principle is relatively simple and the application is also relatively wide. However, statistical methods are heavily dependent on the distribution characteristics of data. If there are a large number of abnormal points in the data, it will affect the data distribution form, thereby reducing the accuracy of statistical identification of abnormal data. Machine learning methods have wide applicability. Common methods include clustering-based methods and distance measurement-based methods, etc. These methods require a large number of samples, especially deep learning, which requires high sample and computing resources. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a method for online identification of high-dimensional abnormal operation data of key equipment measuring points of a wind turbine.

[0005] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0006] A method for online identification of high-dimensional abnormal operation data of key equipment measuring points of an offshore wind turbine, comprising:

[0007] Establishing a set of candidate variables according to historical operation data of an offshore wind farm, and performing feature selection according to the set of candidate variables to obtain a set of selected feature variables;

[0008] Based on the set of selected feature variables, performing preliminary screening of high-dimensional operation variables in a three-dimensional space composed of wind speed, power and pitch angle according to mechanism analysis to obtain data after preliminary screening;

[0009] Selecting a clustering method with hierarchical structure, clustering the data after preliminary screening, and according to the clustering results, combining prior knowledge, dividing the wind turbine operation data into offshore wind turbine operation data in a limited power generation operation state and a maximum power tracking state;

[0010] The deep auto-encoding Gaussian mixture model is used for online identification of abnormal operation data caused by sensor errors and electromagnetic interference. First, the auto-encoder is used to reduce and reconstruct the offshore wind turbine operation data in the aforementioned limited power generation operation state and maximum power tracking state, and the reduced data and reconstruction error are combined to form a low-dimensional space representation of the aforementioned high-dimensional data.

[0011] The Gaussian mixture model is used to cluster the data reduced by the auto-encoder, and the characteristics of each mixture model are iteratively obtained to obtain the historical operation data anomaly detection result.

[0012] The deep neural network is used to replace the Gaussian mixture clustering part in the deep auto-encoding Gaussian mixture model, and the model is trained to convert the historical operation data anomaly detection result identification process into an online identification process. The network uses a deep neural network to predict whether each data point is abnormal.

[0013] Preferably, a set of candidate variables is established according to the historical operation data of the offshore wind farm, and feature selection is performed according to the set of candidate variables to obtain a set of selected feature variables, including:

[0014] The set of candidate variables is extracted from the historical operation data of the offshore wind farm; the set of candidate variables includes electrical quantities, environmental quantities, mechanical quantities, and temperature quantities.

[0015] The method of nonlinear correlation analysis using a convolutional neural network selects variables highly correlated with the main bearing temperature from the set of candidate variables; the variables highly correlated with the main bearing temperature include: main shaft front bearing temperature, main shaft rear bearing temperature, cabin temperature, gear box front bearing temperature, generator front bearing temperature, grid A-phase voltage, and grid A-phase current.

[0016] The dominant variables are added to the variables highly correlated with the main bearing temperature to form a high-dimensional space for abnormal data identification, and the high-dimensional space is determined as the set of selected feature variables; the dominant variables include: wind speed, power, pitch angle, and torque.

[0017] Preferably, the conditions for preliminary screening of the high-dimensional operation variables include:

[0018] When the wind speed is less than the preset cut-in wind speed, the data with power not equal to 0 is removed.

[0019] When the wind speed is greater than the rated wind speed, abnormal data much greater than the rated power is removed.

[0020] Preferably, a clustering method with a hierarchical structure is selected to cluster the screened data, and according to the clustering results, the wind turbine operation data is divided into offshore wind turbine operation data in a limited power generation operation state and a maximum power tracking state in combination with prior knowledge, including:

[0021] The screened data is normalized by using a normalization formula to obtain a normalized data set;

[0022] The parameter space of the clustering number k is set according to the actual situation, the normalized data set is clustered and analyzed by using the BIRCH clustering algorithm for each clustering number k, and the clustering results corresponding to each clustering number are obtained;

[0023] The contour coefficient is calculated based on the clustering results;

[0024] According to the contour coefficient, the wind turbine operation data is divided into data in a maximum power tracking working condition and data in a limited power generation working condition in combination with prior knowledge.

[0025] Preferably, the deep auto-encoding Gaussian mixture model is composed of three parts of a compression network, a Gaussian mixture model and an estimation network; the compression network is used to reduce the dimension of the input sample by the deep auto-encoder, to calculate the low-dimensional representation of the sample from the space after dimension reduction and the reconstruction error feature, and to provide the low-dimensional representation to the Gaussian mixture model; the Gaussian mixture model is used to cluster the offline historical data in the low-dimensional space to obtain the clustering results corresponding to the sample space; and the estimation network is used to learn the features of the Gaussian mixture model under the results of the Gaussian mixture model, and to set a threshold according to the output results of the network to judge whether the data is abnormal online.

[0026] Preferably, it further comprises:

[0027] Based on the cloud edge end collaborative architecture, the steps of "establishing a set of candidate variables according to the historical operation data of the offshore wind farm, and performing feature screening according to the set of candidate variables to obtain a set of selected feature variables" to "using a deep neural network to replace the Gaussian mixture clustering part in the deep auto-encoding Gaussian mixture model, performing model training, converting the historical operation data anomaly detection result identification process into an online identification process, and using a deep neural network to predict whether each data point is abnormal" are deployed to obtain a deployment model;

[0028] The deployment model is loaded into a digital twin platform for application.

[0029] According to the specific embodiments provided by the application, the following technical effects are disclosed:

[0030] The application provides an offshore wind turbine key equipment measuring point high-dimensional abnormal operation data online identification method, comprising the following steps: establishing a candidate variable set according to historical operation data of an offshore wind farm, and performing feature screening according to the candidate variable set to obtain a feature selection variable set; based on the feature selection variable set, performing high-dimensional operation variable preliminary screening according to mechanism analysis in a three-dimensional space formed by wind speed, power and pitch angle to obtain data after preliminary screening; selecting a clustering method with a hierarchical structure, clustering the data after preliminary screening, and according to the clustering result, combining prior knowledge, dividing offshore wind turbine operation data in a limited power generation operation state and a maximum power tracking state; using a deep self-encoding Gaussian mixture model to perform online identification of abnormal operation data caused by sensor errors and electromagnetic interference; first, using an autoencoder to respectively reduce dimensions and reconstruct the offshore wind turbine operation data in the limited power generation operation state and the maximum power tracking state, combining the data after dimension reduction and reconstruction errors to form a low-dimensional space representation of the aforementioned high-dimensional data; clustering the data after dimension reduction by the autoencoder using a Gaussian mixture model, and iteratively obtaining the features of each mixture model to obtain a historical operation data anomaly detection result; using a deep neural network to replace the Gaussian mixture clustering part in the deep self-encoding Gaussian mixture model, performing model training, converting the historical operation data anomaly detection result identification process into an online identification process, and estimating whether each data point is abnormal using a deep neural network. Based on the analysis of the historical data set, the online data can be accurately identified. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0032] Figure 1 The method flowchart provided for the embodiments of the present application;

[0033] Figure 2 The online identification flowchart provided for the embodiments of the present application;

[0034] Figure 3 The deep self-encoding Gaussian mixture model structure diagram provided for the embodiments of the present application;

[0035] Figure 4 The cloud edge end system architecture diagram provided for the embodiments of the present application. DETAILED DESCRIPTION

[0036] With reference to the accompanying drawings: the technical solutions in the embodiments of the present application will be described clearly and completely, obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0037] The purpose of the present application is to provide an offshore wind turbine key equipment measurement point high-dimensional abnormal operation data online identification method, which can accurately identify the abnormal data of online data based on the analysis of historical data set.

[0038] In order to make the above-mentioned purpose, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.

[0039] Figure 1 The method flow chart provided by the embodiments of the present application is shown in Figure 1 The present application provides an offshore wind turbine key equipment measurement point high-dimensional abnormal operation data online identification method, which comprises:

[0040] Step 100: according to the historical operation data of offshore wind farm, a set of candidate variables is established, and according to the set of candidate variables, feature selection is carried out to obtain a set of feature selection variables;

[0041] Step 200: based on the set of feature selection variables, through the three-dimensional space composed of wind speed, power and pitch angle, the initial screening of high-dimensional operation variables is carried out according to mechanism analysis, and the data after initial screening is obtained;

[0042] Step 300: selecting a clustering method with hierarchical structure, clustering the data after initial screening, and according to the clustering result, combining priori knowledge, dividing the offshore wind turbine operation data into offshore wind turbine operation data under the states of limited power generation operation and maximum power tracking;

[0043] Step 400: using deep auto-encoding Gaussian mixture model to identify the abnormal operation data caused by sensor error and electromagnetic interference online, first using auto-encoder to reduce dimension and reconstruct the offshore wind turbine operation data under the states of limited power generation operation and maximum power tracking respectively, using the data after dimension reduction and reconstruction error to combine the low-dimensional space representation of the aforementioned high-dimensional data;

[0044] Step 500: clustering the data after dimension reduction by auto-encoder through Gaussian mixture model, and iteratively obtaining the characteristics of each mixture model to obtain the abnormal detection result of historical operation data;

[0045] Step 600: using a deep neural network instead of the Gaussian mixture clustering part in the deep auto-encoding Gaussian mixture model, model training is performed, the historical operation data anomaly detection result identification process is converted into an online identification process, and the network uses a deep neural network to predict whether each data point is abnormal.

[0046] Figure 2 The online identification flowchart provided for the embodiment of the application is shown in FIG. 1, and the online identification flowchart of the embodiment includes the following steps. Figure 2

[0047] Step one: first, according to the historical operation data of the offshore wind farm, the environmental quantities, electrical quantities, mechanical quantities and the like related to the key component measuring points of the wind turbine are screened to construct a drive system digital twin model to establish a set of candidate variables. Using a nonlinear correlation analysis method, the highly correlated variable set is screened from the established candidate variable set through feature selection.

[0048] Step two: in the variable set of the feature selection result, according to the mechanism analysis, the high-dimensional operation variables are preliminarily screened in the three-dimensional space composed of wind speed, power and pitch angle. The purpose of this step is to eliminate obviously erroneous data in the data set.

[0049] Step three: a hierarchical clustering method is selected to cluster the data after the preliminary elimination of abnormal operation data, and according to the clustering result, the data is divided into offshore wind turbine operation data in the power limited generation state and the maximum power tracking state in combination with prior knowledge.

[0050] Step four: a deep auto-encoding Gaussian mixture model is used for online identification of abnormal operation data caused by sensor errors and electromagnetic interference. First, an auto-encoding network is constructed to complete the dimension reduction of the original data and the establishment of the reconstruction error.

[0051] Step five: the new low-dimensional space data composed of the dimension reduction and reconstruction error is clustered by the Gaussian mixture model, the characteristics of each mixture model are iteratively obtained, and the historical operation data anomaly detection result is obtained.

[0052] Step six: using a deep neural network (estimation network) instead of the Gaussian mixture clustering part in the aforementioned deep auto-encoding Gaussian mixture model, model training is performed, the aforementioned offline abnormal operation data identification process is converted into an online identification process, and the estimation network uses a deep neural network to predict whether each data point is abnormal.

[0053] Step seven: the flowchart is deployed based on a "cloud-edge-end" collaborative architecture, and the obtained model is loaded into a digital twin platform for application.

[0054] ​Preferably, the candidate variable set is established according to historical operation data of the offshore wind farm, and feature screening is performed according to the candidate variable set to obtain a feature selection variable set, which includes:

[0055] The candidate variable set is extracted from the historical operation data of the offshore wind farm; the candidate variable set includes electrical quantities, environmental quantities, mechanical quantities and temperature quantities;

[0056] The highly correlated variables of the main bearing temperature are selected from the candidate variable set by using a nonlinear correlation analysis method of a convolutional neural network; the highly correlated variables of the main bearing temperature include the front main bearing temperature, the rear main bearing temperature, the cabin temperature, the front gear box bearing temperature, the front generator bearing temperature, the grid A-phase voltage and the grid A-phase current;

[0057] The leading variables are added to the highly correlated variables of the main bearing temperature to form a high-dimensional space for abnormal data identification, and the high-dimensional space is determined as the feature selection variable set; the leading variables include the wind speed, the power, the pitch angle and the torque.

[0058] Specifically, step one in the embodiment is: first, the candidate variable set is extracted from the SCADA, including electrical quantities (voltage, current, power, etc.), environmental quantities (environmental temperature, environmental humidity, weather, etc.), mechanical quantities (generator speed, torque, etc.) and temperature quantities (cabin temperature, gear box bearing temperature, etc.). The highly correlated variables of the main bearing temperature are selected from the above candidate variable set by using a nonlinear correlation analysis method of a convolutional neural network (CNN). The CNN is composed of an input layer, a convolution layer, an activation function, a pooling layer, a full connection layer and an output layer, etc., and is a neural network algorithm containing convolution calculation and having a parameter sharing mechanism and a deep structure. The convolution calculation function and the pooling function can perform deep feature extraction on the input data.

[0059] The highly correlated variables of the main bearing are selected by feature selection, which are the front main bearing temperature, the rear main bearing temperature, the cabin temperature, the front gear box bearing temperature, the front generator bearing temperature, the grid A-phase voltage, the grid A-phase current, and the leading variables wind speed, power, pitch angle and torque are added to form a high-dimensional space for abnormal data identification.

[0060] Preferably, the conditions for the preliminary screening of the high-dimensional operation variables include:

[0061] When the wind speed is less than a preset cut-in wind speed, the data with a power of 0 is removed;

[0062] When the wind speed is greater than a rated wind speed, the abnormal data far greater than the rated power is removed.

[0063] Specifically, step two in the embodiment is: firstly, according to the wind speed, power and pitch angle in a three-dimensional space, according to the operating characteristics of the wind turbine power curve, the abnormal power is preliminarily removed, and the specific removed data is as follows: 1) when the wind speed is less than the cut-in wind speed, the data with power not equal to 0 is removed. 2) when the wind speed is greater than the rated wind speed, the abnormal data much greater than the rated power is removed.

[0064] Preferably, a clustering method with a hierarchical structure is selected to cluster the screened data, and according to the clustering results, the wind turbine operating data is divided into offshore wind turbine operating data in the maximum power tracking state and the limited power generation state, including:

[0065] The normalized data set is obtained by normalizing the screened data using a normalization formula.

[0066] According to the actual situation, the parameter space of the clustering number k is set, for each clustering number k, the BIRCH clustering algorithm is used to cluster and analyze the normalized data set, and the clustering result corresponding to each clustering number is obtained.

[0067] The contour coefficient is calculated based on the clustering result.

[0068] According to the contour coefficient, the wind turbine operating data is divided into data in the maximum power tracking state and data in the limited power generation state.

[0069] Specifically, step three in the embodiment is as follows:

[0070] After the preliminary identification of abnormal data, the high-dimensional operating data is clustered in the hypersphere space using the balanced iterative reducing and clustering using hierarchies (BIRCH) method, and the purpose of this step is to identify abnormal operating condition data such as limited power generation data in the operating data. The specific steps of BIRCH clustering are as follows:

[0071] 1) The collected high-dimensional operating data of the wind turbine is represented as X=(x1, x2, x3, L, x n ).

[0072] 2) The data set is normalized using a normalization formula, and the normalization formula is:

[0073]

[0074] Where x and x' correspond to the data before and after normalization respectively, x max and xmin These correspond to the maximum and minimum values ​​of the original data, y max and y min These correspond to the maximum and minimum values ​​of the target range of the original data, respectively. Since the activation function has a range of [-1, 1], the normalization target value is [-1, 1].

[0075] 3) Based on the actual situation, define the parameter space for the number of clusters. For each number of clusters k, use the BIRCH clustering algorithm to perform cluster analysis on the dataset, and obtain the clustering results corresponding to each number of clusters. Each node of the BIRCH clustering algorithm feature tree is represented by a cluster feature (CF) containing three elements, as shown in the following formula.

[0076]

[0077] Where N is the number of sample points within the cluster, x n Let be the representation vector of the sample points, LS be the linear sum of the sample points within the cluster, and SS be the sum of squares of the sample points within the cluster. The CF tree contains three parameters: the internal node balance factor B′, the leaf node balance factor L′, and the cluster radius critical R. The g-th feature tree node is defined as a CFT. g .

[0078] CFT g = <CF i Chilld i >1<i<H

[0079] Where H represents the maximum value of the child node, CF i It is the i-th clustering attribute of the feature node, Child i This represents the next child node of the current node. The cluster radius critical value R is expressed as:

[0080]

[0081] Typically, data clustering results are presented as spherical clusters, where a spherical cluster represents J data factors that are symmetrically distributed and spherically shaped within the data space. Here, J represents the total number of data points within each cluster, and x represents the total number of data points within each cluster. ij This represents the j-th data point in the i-th cluster.

[0082] 4) Calculate the silhouette coefficient S to measure the clustering effect and determine the optimal clustering result. The calculation method is as follows:

[0083]

[0084] Wherein, N is the total number of sample points, a(i) is used to quantify the cohesion within the cluster, i.e. the average distance of sample i to each sample point within the cluster; b(i) is used to quantify the separation between clusters, i.e. the average distance of sample i to other cluster sample points. N inside The number of sample points in the cluster where sample p i The number of sample points in the cluster where sample p j Other sample points in the cluster where sample p i Other sample points in the cluster where sample p outside The number of samples outside the cluster where sample p i The number of samples outside the cluster where sample p j The number of samples outside the cluster where sample p i The number of samples outside the cluster where sample p

[0085] 5) According to the best contour coefficient, combined with prior knowledge, the data is divided into 1) maximum power tracking working condition data. 2) power limited generation working condition data.

[0086] Preferably, the deep autoencoding Gaussian mixture model is composed of three parts of compression network, Gaussian mixture model and estimation network; the compression network is used to reduce the dimension of the input sample through the deep autoencoder, calculate the low-dimensional representation of the sample from the space after dimension reduction and reconstruction error feature, and provide the low-dimensional representation to the Gaussian mixture model; the Gaussian mixture model is used to cluster the offline historical data in the low-dimensional space, and obtain the clustering result corresponding to the sample space; the estimation network is used to learn the characteristics of the Gaussian mixture model under the result of the Gaussian mixture model, and set a threshold according to the result output by the network, to judge whether the data is abnormal online.

[0087] Specifically, step four in the embodiment is: according to the result of step three, the wind turbine operation data has been divided into two parts, 1) maximum power tracking working condition data. 2) power limited generation working condition data. This step uses deep autoencoding Gaussian mixture model (DAGMM) to further clean the abnormal operation data in the data caused by errors in the process of sensor, collection and storage under the two working conditions.

[0088] The deep autoencoding Gaussian mixture model is composed of three parts of compression network and Gaussian mixture model, and estimation network. As shown in Figure 3As shown, the working principle of DAGMM is as follows: 1) the compression network reduces the dimension of the input sample through the deep autoencoder, calculates the low-dimensional representation of the sample from the space after dimension reduction and the reconstruction error feature, and provides it to the subsequent Gaussian mixture model; 2) the Gaussian mixture model clusters the offline historical data in the low-dimensional space to obtain the clustering result corresponding to the sample space. 3) The estimation network learns the features of the Gaussian mixture model under the result of the Gaussian mixture model, replaces the GMM part with the estimation network, omits the process of obtaining the features of each mixture model in the historical data set by iteration, and at the same time can set a threshold according to the output result of the network to judge whether the data is abnormal online.

[0089] z c =h(x;θ e )

[0090] x′=g(z c ;θ d )

[0091] z r =f(x,x′)

[0092] z=[z c ,z r ]

[0093] The foregoing feature variables are taken as inputs, and the same features as the input variables are taken as output variables, and the autoencoding neural network is trained to minimize the loss function L1. The output of the autoencoding network includes two sources, one is the low-dimensional representation after re-encoding of the original high-dimensional running data, and the other part is the reconstruction error. Among them, the low-dimensional representation z c is the low-dimensional representation of the high-dimensional data, z r is the distance measure calculated from the reconstruction error, which is the relative Euclidean distance and cosine similarity. θ e , θ d are the parameters of the deep autoencoder, x' is the reconstructed corresponding item of x, h(·) is the encoding function, and g(·) is the decoding function. The relative Euclidean distance d and the cosine similarity are calculated as follows, wherein x and x' are the original data and the reconstructed data through the deep autoencoder respectively:

[0094]

[0095]

[0096]

[0097] After training the network, the low-dimensional representation z iThe relative Euclidean distance d and the cosine similarity s calculated from the reconstruction error are concatenated to form the input z = [z c of the original data.

[0098] Optionally, step five in the embodiment is to establish a Gaussian mixture model

[0099] The Gaussian mixture clustering model (GMM) is used for offline identification of point abnormal data in the limited power generation operating condition and the maximum power tracking state operating condition, and is a parameterized model, which is expressed as follows:

[0100]

[0101] wherein, α i is a coefficient, The Gaussian probability density function (PDF) parameters are different, and each Gaussian model can be regarded as a kind of category. The low-dimensional representation z c of the foregoing deep self-encoding network is concatenated with the relative Euclidean distance d and the cosine similarity s calculated according to the reconstructed data to form the input z = [z c ,d,s] of the Gaussian mixture clustering model, the number of clusters is set to 3-5, the clustering result is determined through Gaussian mixture clustering combined with prior knowledge, and the abnormal operating data is identified. The specific steps are as follows:

[0102] 1) First, the number of clusters is determined, and the parameters of each Gaussian model are initialized, and the weight of each model is calculated.

[0103] 2) The weight of each node belonging to each Gaussian model is calculated according to the following formula.

[0104]

[0105]

[0106] wherein, P(i,j) represents the proportion of each point, and represents the weight of each Gaussian model in the mixture model.

[0107] 3) The mean and variance of each Gaussian model are updated, as shown in the following formula:

[0108]

[0109]

[0110] 4) The total weight of each Gaussian model is updated.

[0111]

[0112] Step six in this embodiment is: constructing the estimation network.

[0113] A deep neural network (estimation network) is used to replace the Gaussian mixture clustering part in the aforementioned DAGMM for model training, transforming the offline anomaly data identification process into an online one. The estimation network uses the deep neural network to predict whether each data point is an anomaly.

[0114] P = MLN(z; θ) m )

[0115] γ′=softmax(P)

[0116] Where P represents the output of the neural network, and γ′ represents the K-dimensional vector used to predict the membership of data points. Given a batch of N samples and their predicted membership, Further estimates of the parameters in the GMM are as follows:

[0117]

[0118]

[0119]

[0120] Where γ′ represents the low-dimensional representation z i Membership prediction, φ k ′、μ k ′、∑ k Let '' represent the mixture probability, mean, and covariance of the k-th mixture component in the GMM. The following formula can be used to predict the degree of anomaly in a sample:

[0121]

[0122] Using the input of the aforementioned GMM model, i.e. [z] c The network takes [d,s] as input and the membership predictions for each data point obtained from the GMM as output. Using the loss function as the objective, the model is trained to obtain an online model for identifying high-dimensional abnormal operation data of key components of offshore wind turbines. This model can directly predict the abnormality of samples and set samples with high abnormality as abnormal online by setting a threshold.

[0123] Preferably, it further includes:

[0124] The cloud edge end collaborative architecture is used for deployment of steps "establishing a set of candidate variables according to historical operation data of the offshore wind farm, and performing feature screening according to the set of candidate variables to obtain a set of selected feature variables" to "using a deep neural network to replace a Gaussian mixture clustering part in the deep auto-encoder Gaussian mixture model, performing model training, converting the historical operation data anomaly detection result identification process into an online identification process, and estimating whether each data point is abnormal by using the deep neural network", to obtain a deployed model.

[0125] The deployed model is loaded into a digital twin platform for application.

[0126] Specifically, step seven in the embodiment is to deploy the process based on the "cloud edge end" collaborative architecture, and load the obtained model into the digital twin platform for application.

[0127] As shown in Figure 4 The cloud end processes the feature variable selection, BIRCH clustering, auto-encoder neural network construction, GMM based on the historical data set, and estimation network model construction processes through the central server, and obtains an online monitoring identification model for the operation data. At the same time, the cloud end can store the edge side device data, train the data, and issue data preprocessing and analysis, AI decision and real-time calculation applications to the edge side for local calculation. The edge end runs on the terminal device, is connected to the cloud platform master station through the cloud end, has data acquisition, data caching, data processing and analysis, and realizes intelligent decision making and other applications. The end side reports the collected data, and responds to the decision of the edge end. The trained model obtained by the cloud end can be loaded by the digital twin platform of the edge end for application, to realize online identification of high-dimensional abnormal operation data of key equipment of the offshore wind turbine.

[0128] The beneficial effects of the present application are as follows:

[0129] The present application can identify the high-dimensional abnormal operation data of the key components of the offshore wind turbine online. After deleting or correcting the abnormal data, the influence of the abnormal operation data on the conclusion of the data analysis can be avoided, and the model precision and the accuracy of the AI decision can be improved by modeling with the corrected data.

[0130] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other.

[0131] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. An online identification method for high-dimensional abnormal operation data of key equipment measuring points of offshore wind turbines, characterized in that, The method comprises the following steps: According to the historical operation data of the offshore wind farm, a set of candidate variables is established, and feature screening is performed according to the set of candidate variables to obtain a set of feature selection variables; Based on the set of feature selection variables, the high-dimensional operation variables are preliminarily screened in a three-dimensional space formed by wind speed, power and pitch angle according to mechanism analysis to obtain data after preliminary screening; A hierarchical clustering method is selected to cluster the data after preliminary screening, and according to the clustering results, the offshore wind turbine operation data is divided into offshore wind turbine operation data in a limited power generation operation state and a maximum power tracking state in combination with prior knowledge; A deep auto-encoding Gaussian mixture model is used for online identification of abnormal operation data caused by sensor errors and electromagnetic interference. Firstly, the auto-encoder is used to reduce and reconstruct the offshore wind turbine operation data in the aforementioned limited power generation operation state and maximum power tracking state, respectively. The reduced data and reconstruction error are combined to represent the high-dimensional data in a low-dimensional space; The data after dimension reduction by the auto-encoder is clustered by a Gaussian mixture model, and the features of each mixture model are iteratively obtained to obtain the historical operation data anomaly detection result; A deep neural network is used to replace the Gaussian mixture clustering part in the deep auto-encoding Gaussian mixture model for model training, and the historical operation data anomaly detection process is changed into an online identification process. The network uses a deep neural network to predict whether each data point is abnormal; According to the historical operation data of the offshore wind farm, a set of candidate variables is established, and feature screening is performed according to the set of candidate variables to obtain a set of feature selection variables, comprising: Extracting a set of candidate variables from the historical operation data of the offshore wind farm; the set of candidate variables includes electrical quantities, environmental quantities, mechanical quantities and temperature quantities; Using a convolutional neural network to analyze the nonlinear correlation of the set of candidate variables to select variables highly correlated with the main bearing temperature; the variables highly correlated with the main bearing temperature include: main shaft front bearing temperature, main shaft rear bearing temperature, cabin temperature, gear box front bearing temperature, generator front bearing temperature, grid A-phase voltage and grid A-phase current; Adding a dominant variable to the variables highly correlated with the main bearing temperature to form a high-dimensional space for abnormal data identification, and determining the high-dimensional space as the set of feature selection variables; the dominant variable includes: wind speed, power, pitch angle and torque.

2. The method according to claim 1, characterized in that, The conditions for the preliminary screening of the high-dimensional operation variables include: When the wind speed is less than the preset cut-in wind speed, the data with power not equal to 0 is removed; When the wind speed is greater than the rated wind speed, the abnormal data much greater than the rated power is removed.

3. The method according to claim 1, wherein, A hierarchical clustering method is selected to cluster the data after preliminary screening, and according to the clustering results, the offshore wind turbine operation data is divided into offshore wind turbine operation data in a limited power generation operation state and a maximum power tracking state in combination with prior knowledge, comprising: The normalized data set is obtained by normalizing the data after preliminary screening using a normalization formula; According to actual situation, set the number of clusters k Parameter space, for each cluster number k , using BIRCH clustering algorithm for clustering analysis of normalized data set, the cluster results corresponding to each cluster number are obtained; The contour coefficient is calculated based on the clustering results; According to the profile coefficient, in combination with prior knowledge, the wind turbine operation data is divided into data under maximum power tracking working condition and data under limited power generation working condition.

4. The method according to claim 1, wherein, The deep self-encoding Gaussian mixture model is composed of three parts of a compression network, a Gaussian mixture model and an estimation network; the compression network is used for dimension reduction of input samples by a deep self-encoder, calculation of low-dimensional representation of samples from a space after dimension reduction and reconstruction error features, and provision of the low-dimensional representation to the Gaussian mixture model; the Gaussian mixture model is used for clustering of offline historical data in a low-dimensional space to obtain a clustering result under a corresponding sample space; and the estimation network is used for learning of features of the Gaussian mixture model under a result of the Gaussian mixture model, and setting of a threshold according to a result output by the network to judge whether data is abnormal on line.

5. The method for online identification of high-dimensional abnormal operation data of offshore wind turbine key equipment measurement points according to claim 1, characterized in that, Further comprising: Based on the cloud edge-end collaborative architecture, steps "establishing a set of candidate variables according to historical operation data of an offshore wind farm, and performing feature screening according to the set of candidate variables to obtain a set of selected feature variables" to "using a deep neural network to replace a Gaussian mixture clustering part in the deep self-encoding Gaussian mixture model, performing model training, converting the historical operation data anomaly detection result identification process into an online identification process, and using the deep neural network to predict whether each data point is abnormal" are deployed to obtain a deployment model; The deployment model is carried to a digital twin platform for application.

Images