Wind turbine abnormality detection and positioning method based on interpretable graph neural network

Abstract

The application discloses a wind turbine unit abnormality detection and positioning method based on an interpretable graph neural network, and belongs to the field of wind turbine unit state monitoring.The method comprises the following steps: S1, screening multivariate time series SCADA data; S2, covariate preprocessing and spectral domain graph convolution network model initialization; S3, inputting the health data after the covariate preprocessing into the spectral domain graph convolution network to extract features; S4, setting a threshold according to a training set; S5, performing abnormality detection and positioning through a detection and positioning module; S6, performing abnormality detection and positioning through a post hoc interpretability module; and S7, comprehensively evaluating the detection and positioning module and the decision interpretability module to scientifically evaluate and guide the health state of the unit.The application realizes early warning of unit faults, realizes accurate positioning of unit faults, is beneficial to timely processing and maintenance of the unit in a sub-health state, and avoids deep damage to the unit and key components.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind turbine condition monitoring, in particular to a wind turbine anomaly detection and positioning method based on an interpretable graph neural network. BACKGROUND

[0002] In recent years, deep learning technology has developed rapidly. As a new technology in the field of deep learning, graph neural network (GNN) has attracted much attention due to its ability to capture and learn non-Euclidean data and their correlations. GNN has made a lot of important achievements in image detection and recognition, classification and other fields. However, GNN also has many unsolved problems. One of the important problems is the interpretability of deep learning models. The purpose of studying interpretability is to explain the operation mechanism of deep learning and explore the explanation of the reliability of deep learning models, and then find a way to guide the modification of the network, detect and repair the defects of the trained network, and present the results with strong interpretability.

[0003] Traditional non-graph network interpretability methods include two kinds: the first is to develop a complete neural network with a simple substitute model. This can be done in a model-agnostic way, usually by learning a local confidence approximation about the predictions, such as representing sufficient conditions for predictions by linear models or rule sets; the second method determines important aspects of the computation, such as feature gradients, backpropagation of neurons on input features, and counterfactual reasoning. However, these methods generate saliency maps that have been shown to be misleading in some cases and are prone to problems such as gradient saturation. These problems are more serious in discrete inputs (such as graph adjacency matrices), because gradient values can be very large within very small intervals. For this reason, this method is not suitable for explaining the predictions made by graph neural networks.

[0004] Currently, GNN research on interpretability is mainly divided into two categories: pre-interpretability and post-decision interpretability. Pre-interpretability techniques directly create a new, inherently interpretable model to explain the model, while post-decision interpretability techniques do not create a new, fixed interpretable model, but rather treat the network model as a black box and infer the interpretability of the network model by the features of the model.

[0005] To solve the above problems, the present application provides a wind turbine anomaly detection and positioning method based on an interpretable graph neural network. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a wind turbine abnormality detection and positioning method based on an interpretable graph neural network, to use the interpretability technology in the graph network, to obtain an interpretable health evaluation method, to use the graph network health evaluation method with interpretability in the wind turbine SCADA data, to make decision interpretability research on the correlation between different sensors of the SCADA data, and to detect and repair possible defects of the graph network model according to the health evaluation method.

[0007] To solve the above technical problems, the technical solution adopted by the present application is:

[0008] A wind turbine abnormality detection and positioning method based on an interpretable graph neural network comprises the following steps:

[0009] S1, obtaining and screening multivariate time series SCADA data: obtaining wind turbine multivariate time series historical data from a wind farm SCADA system, removing missing values and abnormal data through data cleaning, and screening experimental variables and covariates;

[0010] S2, eliminating the influence of environmental variables and working condition changes on experimental variables, and obtaining temperature sensor data after covariate preprocessing, constructing a random forest multi-input-multi-output regression model based on a Bayesian optimization algorithm, and initializing a spectral domain graph convolution network model;

[0011] S3, inputting the health data after covariate preprocessing into the spectral domain graph convolution network to learn the correlation features: constructing a prediction model based on the spectral domain graph convolution network, first extracting the spatial correlation features between different sensor variables in the state variable time series data after covariate preprocessing by using the self-attention mechanism, and calculating the adjacency matrix representing the spatial correlation of the sensors; at the same time, inputting the state variable time series data after covariate preprocessing into the spectral domain graph convolution module to extract the time dimension features, and inputting the time dimension features and the adjacency matrix into the graph convolution module to extract the spatial dimension features; training the prediction model based on the spectral domain graph convolution network using the training set;

[0012] S4, setting a threshold according to the training set: inputting the training set into the prediction model based on the spectral domain graph convolution network to be trained, calculating the root mean square error of the predicted value and the true value of all variables in the training set, and designing a threshold according to the kernel density estimation method;

[0013] S5, inputting online SCADA test data into the model, and judging whether an abnormality occurs according to the threshold through the detection and positioning module, and collecting false alarm points and locating faults:

[0014] Obtaining online real-time monitoring multivariate time data from a wind turbine, inputting the data into a random forest multi-input-multi-output regression model first, calculating the covariate pretreated temperature state data, then inputting the data into a trained prediction model based on a spectral domain graph convolution network, comparing the root mean square error between the predicted value and the true value with a pre-set threshold through a detection and positioning module; when the root mean square error is greater than the threshold, a fault warning is issued for the wind turbine component; at the same time, the root mean square error of the test data is down-sampled by day and by hour, and positioning is performed by the detection and positioning module, and finally, false alarm conditions are judged according to the validation set and the threshold setting;

[0015] S6, inputting online SCADA test data into a prediction model based on a spectral domain graph convolution network, adaptively setting a threshold according to data correlation, visualizing the result model through a decision interpretability module, realizing fault warning and fault positioning, and optimizing model parameter structure according to a result optimization method;

[0016] S7, comprehensively evaluating the detection and positioning module and the decision interpretability module to scientifically evaluate and guide the health status of the wind turbine.

[0017] Further improvement of the technical scheme of the application is that in S1, the following steps are specifically included:

[0018] S11, using a density-based outlier factor detection algorithm to detect outliers in original SCADA data, and removing data not conforming to physical meaning and missing values;

[0019] S12, screening temperature state variables related to experiments and covariates affecting temperature for next step modeling.

[0020] Further improvement of the technical scheme of the application is that in S2, the following steps are specifically included:

[0021] S21, dividing the entire experimental data into a training set, a validation set and a test set, wherein the test set data is experimental data; in the training set, taking the covariates as independent variables and the temperature state variables as dependent variables, constructing a random forest multi-input-multi-output regression model based on a Bayesian optimization algorithm according to the number of sensors, and taking the regression model residual as the temperature state variable processed by the covariates;

[0022] S22, using a sliding window to establish time series data of a fixed window size for the temperature state variable processed by the covariates, selecting appropriate data according to different faults of different units, dividing the data according to the ratio of training set: validation set: test set = 7:2:1, and initializing the spectral domain graph convolution network model.

[0023] Further improvement of the technical scheme of the application is that in S3, the following steps are specifically included:

[0024] S31, training set data X ∈ R S×N Input gated linear unit GLU, where S is a sliding window size, N is the number of temperature state variables sensors, the layer calculates the hidden state corresponding to each window S in order, uses the last hidden state R as the representation of the entire time sequence, and calculates the weight matrix W through the self-attention mechanism, and calculates the adjacency matrix A of the sensor space coupling graph according to the weight matrix W, where A = 0.5 * W * W T , that is, a fixed graph structure G based on each fixed time window is formed, G = (X, W);

[0025] S32, the fixed graph structure G generated in S31 is input into the graph Fourier layer to capture the internal inter-sequence relationship, and the following is obtained:

[0026] G F (X) = U T X (1)

[0027] Where, U ∈ R N×S represents a feature vector matrix, X ∈ R S×N represents input data;

[0028] S33, the result in S32 is input into the discrete Fourier layer and the gated linear unit GLU to learn the frequency domain representation of the input time sequence, so as to capture the repetitive patterns in the periodic data or the autocorrelation features between different time stamps;

[0029] S34, the frequency representation in S33 is input into two one-dimensional convolution layers and the gated linear unit GLU to capture feature patterns, and the feature h s (X) is obtained:

[0030]

[0031] Where, b and c are bias parameters, and W and V are convolution kernel parameters;

[0032] S35, the learned features in S34 are learned through the graph convolution layer to represent the spectral matrix, and then through the inverse graph Fourier representation, and the result Z is obtained:

[0033]

[0034] Z = Z1||Z2…Z n (4)

[0035] Where, gθ ij is the i-th input and j-th output channel corresponding to the graph convolution kernel, Λ i is a normalized Laplacian eigenvalue matrix, S represents the above model layer, and each output channel Z jSpliced together is the final result Z;

[0036] S36, the result of S35 is passed through a full connection layer to be the final result.

[0037] The further improvement of the technical scheme of the application is that in S4, the following steps are specifically included:

[0038] S41, the root mean square error (RMSE) of the result of S36 and the original result is calculated to obtain:

[0039]

[0040] Wherein, The root mean square error of the i-th data is represented by X train_pre,i The predicted value of the training set of the i-th data is represented by X train_tru,i The true value of the training set of the i-th data is represented by X

[0041] S42, the root mean square error obtained in S41 is estimated by kernel density estimation to obtain the probability density:

[0042]

[0043] Wherein, k represents the data length, K(·) represents the kernel function, and the Gaussian kernel is used, that is σ represents the bandwidth;

[0044] S43, set the confidence level to integrate formula (6) to obtain the threshold value Threshold, as shown in formula (7):

[0045]

[0046] Wherein, alpha is the confidence level, and p(x) is the probability density function of the data, that is, the threshold value Threshold is obtained.

[0047] The further improvement of the technical scheme of the application is that in S5, the following steps are specifically included:

[0048] S51, compare the threshold value set in S43 with the verification set data to obtain the false positive rate, as follows:

[0049]

[0050] The false positive rate is:

[0051]

[0052] S52, according to formula (5), the root mean square error (RMSE) of the test set prediction result and the original data is obtained, as follows:

[0053]

[0054] Wherein, S is the length of the test set, N is 24 (the number of temperature state variables);

[0055] S53, the value in S52 is down-sampled by day and hour to obtain the Rmse based on each day and the Rme based on each hour, denoted as:

[0056] S531, the result in S53 is compared with a threshold value, and the comparison result is denoted as:

[0057]

[0058] S532, each temperature state variable in S52 is down-sampled by day and by hour, and the comparison result is denoted as:

[0059]

[0060] Wherein, N is 24 (the number of temperature state variables), and S is the length of the test set;

[0061] S533, the result in S532 is visualized by a heat map, and the positioning condition is obtained.

[0062] The further improvement of the technical scheme is that in S6, the following steps are specifically included:

[0063] S61, the obtained result in S53 is divided into individual matrices according to respective units, and the correlation is calculated, as follows:

[0064] ρ xy ∈ (0, 1) and Or

[0065] Wherein, N is 24, which is the number of temperature state variables; ρ represents the correlation to be solved; and S represents the length of the time series data;

[0066] S62, according to the correlation kernel density distribution curve drawn in S61, the correlation at T time is divided into three gradients, as follows:

[0067]

[0068] S63, according to the kernel density distribution curve and the gradient in S62, a different time change threshold value is set as 0.4, and the comparison result is denoted as:

[0069]

[0070] ​​S64, input the experimental result data of S61 into the explainability module according to different time scales, and map the correlation as the distance of the edge between different nodes Vertex in the module, that is, the stronger the correlation, the shorter the distance between nodes, and the weaker the correlation, the farther the distance between nodes;

[0071] S641, draw the undirected weighted graph of G=(V, E, T) according to the distance features learned by S64, wherein V represents the node, that is, the temperature state variable, E represents the edge directly between different nodes, that is, the correlation is represented by distance, and T represents time;

[0072] S65, according to the result of S641, and the correlation between the real unit and different components, the result is evaluated, and if there is deviation, the model framework parameters are modified in reverse feedback.

[0073] Thanks to the above technical solutions, the technical progress achieved by the present application is:

[0074] 1, by establishing a nonlinear model between the covariate and the temperature state variable, the present application can effectively eliminate the influence of the covariate on the temperature state variable.

[0075] 2, by constructing a self-attention mechanism, the present application can adaptively construct a fixed graph structure, and fully extract the spatiotemporal features through a spectral domain graph convolution network, thereby establishing a global state monitoring model based on the temperature state variable.

[0076] 3, the present application proposes an explainability analysis module, through which the correlation gradient and threshold value can be adaptively set, and the correlation between nodes can be mapped as edge length whether from single variable analysis or from global variable analysis, and a three-dimensional graph based on nodes, edge length and time can be clearly and intuitively drawn. Compared with the result presentation in the detection and positioning module, the change in correlation can be captured sensitively through the decision explainability module, so that accurate sub-health early warning can be carried out more in advance, and the abnormal points and the trend of the causes afterwards can be clearly and intuitively reflected. Through the decision explainability module and the detection and positioning module, early warning of unit failure is realized, accurate positioning of unit failure is realized, a clear and intuitive global health evaluation guidance method is provided, which is conducive to timely processing and maintaining the unit when the unit is in sub-health state, and deep damage to the unit and key components is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0077] 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 or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0078] Figure 1 is a flowchart of an embodiment of the present application;

[0079] Figure 2 is a schematic diagram of a method module of an embodiment of the present application;

[0080] Figure 3 is a schematic diagram of a covariate preprocessing model of an embodiment of the present application;

[0081] Figure 4 is a schematic diagram of a raw twenty-four-dimensional temperature state variable of an embodiment of the present application;

[0082] Figure 5 is a schematic diagram of a temperature state variable after the influence of covariates is removed by a random forest multi-input-multi-output regression model based on a Bayesian optimization algorithm of an embodiment of the present application;

[0083] Figure 6 is a schematic diagram of a spectral domain graph convolution network of an embodiment of the present application;

[0084] Figure 7 is a schematic diagram of test set anomaly detection of a traditional method analysis module of an embodiment of the present application;

[0085] Figure 8 is a schematic diagram of test set anomaly localization of a traditional method analysis module of an embodiment of the present application;

[0086] Figure 9(a) is a schematic diagram of detection and localization of a faulty sensor by an explainability module of an embodiment of the present application Figure 1 ;

[0087] Figure 9(b) is a schematic diagram of detection and localization of a faulty sensor by an explainability module of an embodiment of the present application Figure 2 ;

[0088] Figure 9(c) is a schematic diagram of detection and localization of a faulty sensor by an explainability module of an embodiment of the present application Figure 3 . DETAILED DESCRIPTION

[0089] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and in the claims and the above-described accompanying drawings are intended to cover not only the inclusion of the stated steps or elements but also the inclusion of other steps or elements not clearly listed or inherent to such processes, methods, products, or devices.

[0090] The present application will be further described below in conjunction with the accompanying drawings and embodiments:

[0091] AsFigure 1 、 2 The wind turbine abnormality detection and positioning method based on the interpretable graph neural network includes the following steps:

[0092] S1, obtaining and screening multivariate time series SCADA data:

[0093] Obtain the multivariate time series historical data of the wind turbine from the SCADA system of the wind farm, remove missing values and abnormal data through data cleaning, and screen the experimental variables and covariates;

[0094] Specifically, the following steps are included:

[0095] S11, use the density-based outlier factor detection algorithm to detect outliers of the original SCADA data, and remove data that does not conform to the physical meaning and missing values;

[0096] S12, screen the temperature state variables related to the experiment and the covariates affecting the temperature for the next modeling.

[0097] S2, eliminate the influence of environmental variables and working condition changes on the experimental variables, and obtain the temperature sensor data after covariate preprocessing, construct a random forest multi-input-multi-output regression model and initialize a spectral domain graph convolution network model;

[0098] Specifically, the following steps are included:

[0099] S21, divide the entire experimental data into a training set, a validation set and a test set, wherein the test set data is the experimental data; in the training set, the covariates are used as independent variables, and the temperature state variables are used as dependent variables; according to the number of sensors, a random forest multi-input-multi-output regression model based on the Bayesian optimization algorithm is constructed, and the residual error of the calculated regression model is used as the temperature state variable after covariate processing;

[0100] S22, use a sliding window to establish time series data with a fixed window size for the temperature state variable after covariate preprocessing; according to different faults of different units, select appropriate data and divide the data according to the ratio of training set: validation set: test set = 7:2:1, and initialize a spectral domain graph convolution network model;

[0101] S3. Input the health data after covariate preprocessing into the spectral domain graph convolutional network to learn correlation features: Construct a prediction model based on the spectral domain graph convolutional network. First, use the self-attention mechanism to extract the spatial correlation features between different sensor variables in the time series data of the state variables after covariate preprocessing, and calculate the adjacency matrix representing the spatial correlation of sensors. At the same time, input the time series data of the state variables after covariate preprocessing into the spectral domain graph convolutional module to extract the features of the time dimension, and input it together with the adjacency matrix into the graph convolutional module to extract the features of the spatial dimension. Use the training set to train the prediction model based on the spectral domain graph convolutional network.

[0102] Specifically, the following steps are included:

[0103] S31. Transfer the training set data X∈R S×N The input is a gated linear unit (GLU), where S is the sliding window size and N is the number of sensors for the temperature state variable. This layer sequentially calculates the hidden state corresponding to each window S, and uses the last hidden state R as the representation of the entire time series. A weight matrix W is calculated using a self-attention mechanism, and the adjoining matrix A of the sensor spatial coupling graph is calculated based on the weight matrix W, where A = 0.5 * W * W. T That is, to form a fixed graph structure G, G=(X,W) based on each fixed time window;

[0104] S32. Input the fixed graph structure G generated in S31 into the graph Fourier layer to capture the relationships between internal sequences, and obtain:

[0105] G F (X)=U T X (1)

[0106] Where, U∈R N×S Let X ∈ R be the eigenvector matrix. S×N Indicates input data;

[0107] S33. Input the results from S32 into the Discrete Fourier layer and the gated linear unit (GLU) to learn the frequency domain representation of the input time series, thereby capturing the repeating patterns in periodic data or the autocorrelation features between different timestamps.

[0108] S34. Input the frequency representation from S33 into two one-dimensional convolutional layers and a gated linear unit (GLU) to capture the feature pattern, thus obtaining the feature h. s (X):

[0109]

[0110] Where b and c are bias parameters, and W and V are convolution kernel parameters;

[0111] S35, the features learned in S34 are learned through a graph convolution layer to obtain a spectral matrix representation, and then through an inverse graph Fourier representation to obtain a result Z:

[0112]

[0113] Z = Z1||Z2…Z n (4)

[0114] where gθ ij is the i-th input and j-th output channel corresponding to the graph convolution kernel, Λ i is a normalized Laplacian eigenvalue matrix, S represents the above model layer, and each output channel Z j is spliced to obtain the final result Z.

[0115] S36, the result of S35 is input into a full connection layer to obtain the final result.

[0116] S4, set the threshold according to the training set: input the training set into the prediction model based on the spectral domain graph convolution network to be trained, calculate the root mean square error of the predicted value and the true value of all variables in the training set, and design a threshold according to the kernel density estimation method; Specifically, the following steps are included:

[0117] S41, calculate the root mean square error RMSE between the result of S36 and the original result, and obtain:

[0118]

[0119] where, represents the root mean square error of the i-th data, X train_pre,i represents the predicted value of the i-th data in the training set, X train_tru,i represents the true value of the i-th data in the training set.

[0120] S42, the root mean square error obtained in S41 is estimated by the kernel density estimation method to obtain the probability density:

[0121]

[0122] where k represents the data length, K(·) represents the kernel function, and the Gaussian kernel is used in the present application, that is, σ represents the bandwidth;

[0123] S43, set the confidence level to integrate formula (6) to obtain the threshold Threshold, as shown in formula (7):

[0124]

[0125] where α is the confidence level, and p(x) is the probability density function of the data, that is, the threshold Threshold can be obtained.

[0126] S5, input the online SCADA test data into the model, and judge whether an anomaly is generated according to a threshold through a detection and positioning module, and perform false alarm point collection and fault positioning:

[0127] Obtain online real-time monitoring multivariate time data from a wind turbine, input the data into a random forest multi-input-multi-output regression model first, calculate temperature state data after covariate preprocessing, then input the data into a trained prediction model based on spectral domain graph convolution, compare the root mean square error between the predicted value and the true value with a pre-set threshold through a detection and positioning module; when the root mean square error is greater than the threshold, a fault warning is issued for the wind turbine component; meanwhile, the root mean square error of the test data is down-sampled by day and by hour, positioned by the detection and positioning module, and finally the false alarm condition is judged according to the validation set and the threshold setting;

[0128] Specifically, the following steps are included:

[0129] S51, compare the validation set data and the threshold set in S43 to obtain a false alarm rate, as follows:

[0130]

[0131] Then the false alarm rate can be obtained:

[0132]

[0133] S52, calculate the root mean square error (RMSE) of the test set prediction result and the original data according to formula (5), and obtain:

[0134]

[0135] Wherein, S is the length of the test set, and N is 24 (the number of temperature state variables);

[0136] S53, down-sample the value in S52 by day and by hour to obtain the Rmse based on each day and the Rme based on each hour, denoted as:

[0137] S531, compare the value in S53 with the threshold, and obtain:

[0138]

[0139] S532, down-sample each temperature state variable in S52 by day and by hour, and obtain:

[0140]

[0141] Wherein N is 24 (temperature state variable number), S is the length of the test set;

[0142] S533, the result in S532 is visualized by a heat map to obtain the positioning situation;

[0143] S6, input the online SCADA test data into the prediction model based on the spectral domain graph convolution network, adaptively set the threshold according to the data correlation, visualize the result model through the decision explainability module, realize fault early warning and fault positioning, and optimize the model parameter structure according to the result optimization method;

[0144] Specifically includes the following steps:

[0145] S61, the correlation between the temperature state variables in S53 is calculated as follows: Cut into individual matrices with their respective units and calculate the correlation as follows:

[0146] ρ xy ∈(0,1) and Or

[0147] Wherein, N is 24, the number of temperature state variables; ρ represents the correlation to be solved; S represents the length of the time series data;

[0148] S62, draw the correlation kernel density distribution curve according to S61, and divide the correlation at T time into three gradients as follows:

[0149]

[0150] S63, according to the kernel density distribution curve and the gradient of S62, set different time change thresholds to 0.4, and obtain:

[0151]

[0152] S64, input the experimental result data of S61 into the explainability module according to different time scales, and map the correlation to the distance between different nodes Vertex in the module, that is, the stronger the correlation between nodes, the shorter the distance between nodes, and the weaker the correlation between nodes, the farther the distance between nodes;

[0153] S641, the distance feature learned by S64 is drawn into a undirected weighted graph G=(V,E,T), wherein V represents nodes, i.e. temperature state variables, E represents edges between different nodes, i.e. distance representing correlation strength, and T represents time;

[0154] S65, according to the results of S641, and the correlation between different components and the actual unit situation, the results are evaluated, and if there is deviation, the model framework parameters are modified in reverse.

[0155] S7, the detection and positioning module and the decision explainability module are comprehensively evaluated to give clearer and more scientific guidance and opinions on the fault condition of the wind turbine, and to give timely feedback on the possible sub-health state.

[0156] Embodiment

[0157] As shown in Figure 1 , 2 , the wind turbine anomaly detection and positioning method based on the explainable graph neural network includes multivariate time series acquisition and data preprocessing, covariate preprocessing, dividing the temperature state data after covariate preprocessing into training set, test set and validation set, and inputting the data into the spectral domain graph network to capture the associated features, and setting a threshold according to the training set, inputting the test set data into the model to input the detection and positioning module and perform root mean square error calculation, anomaly detection and anomaly positioning; the data form is reconstructed, the data is regrouped, the threshold is reset, and anomaly detection and anomaly positioning are performed through the explainability module. Finally, the two analysis methods are comprehensively evaluated to give clearer and more scientific guidance and opinions on the fault condition of the wind turbine, and to give timely feedback on the possible sub-health state of the five parts;

[0158] As shown in Figure 3 , the covariate preprocessing module uses four covariates as independent variables and all temperature state variables as dependent variables to establish a multiple-input-multiple-output regression model using a random forest network to preprocess the covariates to eliminate the influence of environmental variables on temperature state variables. The covariate parameters are four-dimensional, including power, environmental wind speed, environmental temperature and rotor speed; the temperature variables in the regression model are all temperature state variables, which are twenty-four-dimensional, including: average temperature of generator bearing 2 (drive end), average temperature of stator winding 1 phase inside the generator, average temperature of stator winding 2 phase inside the generator, average temperature of stator winding 3 phase inside the generator, average temperature of ring chamber, average temperature of generator bearing 2 (drive end), average temperature of transformer L1 phase, average temperature of transformer L2 phase, average temperature of transformer L3 phase, average temperature inside the bus section, average temperature measured by the IGBT driver on the grid side inverter, average temperature measured by the IGBT driver on the rotor side inverter phase 1, average temperature measured by the IGBT driver on the rotor side inverter phase 2, average temperature measured by the IGBT driver on the rotor side inverter phase 3, average temperature of the gearbox bearing on the high-speed shaft, average temperature of the oil inside the gearbox, average temperature of the VCS throttle coil, average temperature in the hub controller, average temperature on the VCP circuit board, average temperature in the top nacelle controller, average temperature of the VCS cooling water, average temperature of the hydraulic group oil, average temperature of the nacelle, average temperature inside the nose cone;

[0159] Figure 4 and Figure 5 Figure 2 shows a comparison of results of a random forest multi-input-multi-output regression model based on Bayesian optimization algorithm, Figure 4 is the original 24-dimensional temperature state variable, Figure 5 is the temperature state variable after removing the influence of covariates by a random forest multi-input-multi-output regression model based on Bayesian optimization algorithm, and it can be clearly seen that the non-stationarity has been largely eliminated;

[0160] Figure 6 is a schematic diagram of a spectral graph convolution network. The training set data X e R S×N Input gated linear unit (GLU), where S is the sliding window size, N is the number of temperature state variable sensors, this layer calculates the hidden state corresponding to each window S in order, uses the last hidden state R as the representation of the entire time series, and calculates the weight matrix W through the self-attention mechanism, and calculates the adjacency matrix A of the sensor space coupling graph according to the weight matrix W, where A = 0.5 * W * W T , that is, a fixed graph structure G based on each fixed time window is formed, G = (X, W); the fixed graph structure G is input into the graph Fourier layer to capture the internal sequence relationship, and G F (X) = U T X; where U e R N×S represents the feature vector matrix, X e R S×N The above results are input into the discrete Fourier layer and the gated linear unit (GLU) to learn the frequency domain representation of the input time series, thereby capturing the repeating patterns in the periodic data or the autocorrelation features between different timestamps;

[0161] The frequency representation of the above formula is input into two one-dimensional convolution layers and the gated linear unit GLU to capture feature patterns, and the feature h s (X) is obtained: where b and c are bias parameters, and W and V are convolution kernel parameters;

[0162] h s (X) is learned by the graph convolution layer to represent the spectral matrix, and then the inverse graph Fourier representation is used to obtain the result Z:

[0163] where θ ij is the i-th input and j-th output channel corresponding to the graph convolution kernel, Λ i is a normalized Laplacian eigenvalue matrix, K represents the above model layer, and Z j represents the j-th output channel;

[0164] Z = Z1||Z2…Z nwherein Z represents splicing different output channels;

[0165] Figure 7 and Figure 8 The detection and positioning module is shown for anomaly detection and anomaly positioning, and a threshold is set according to the kernel density estimation. By using the kernel density estimation method to obtain the probability density through the root mean square error, the following can be obtained:

[0166]

[0167] wherein k represents the data length, K(·) represents the kernel function, and the Gaussian kernel is used in the present application, that is σ represents the bandwidth;

[0168] The threshold Threshold can be obtained by setting the confidence level and integrating the probability density function, that is:

[0169]

[0170] wherein a is the confidence level, a = 0.95 in the present application, and p(x) is the probability density function of the data, so that the threshold Threshold can be obtained;

[0171] The selected time period is 2017-08-03 to 2017-08-18, a total of 15 days, and the anomaly occurs on August 18. Through the detection and positioning module, it can be clearly seen that on 2017-08-15, the threshold is exceeded and continuously rises, so that the transformer anomaly can be effectively warned 3 days in advance.

[0172] Fig. 9(a), Fig. 9(b), Fig. 9(c) show the single abnormal variable and global variable anomaly detection and anomaly positioning of the decision explainability analysis module.

[0173] The Threshold obtained in the detection and positioning module is Cut into individual matrices with their respective units and calculate the correlation as follows: ρ xj ∈(0,1) and or wherein N is 24 (the number of temperature state variables), ρ represents the correlation to be solved, and S represents the length of the time series data.

[0174] According to the correlation value calculated according to the above formula, the correlation kernel density distribution curve is drawn, and the correlation at T time is divided into three gradients, as follows:

[0175] According to the kernel density distribution curve and the gradient of , set the different time change threshold to 0.4, and the following can be obtained:

[0176] The experimental result data is input into the explainability module according to different time scales, and the correlation is mapped as the distance between different nodes (Vertex) in the module, that is, the stronger the correlation between nodes, the shorter the distance between nodes, and the weaker the correlation between nodes, the farther the distance between nodes.

[0177] The learned distance features are drawn into the undirected weighted graph of G=(V, E, T), where V represents the node, that is, the twenty-four temperature state variables, E represents the edge between different nodes, that is, the correlation between the distance, and T represents the time.

[0178] It can be seen from the single variable three-dimensional graphs of FIG. 9(a), FIG. 9(b), and FIG. 9(c) that the explainability module can not only implement more accurate anomaly detection and anomaly point positioning, but also can capture the correlation changes between temperature state variables one day in advance compared with the detection and positioning module. Thus, a more explainable health evaluation guidance method can be provided, which can more subtly and scientifically evaluate and guide the system health.

[0179] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for anomaly detection and localization of wind turbine units based on interpretable graph neural networks, characterized in that, Includes the following steps: S1. Obtain and filter multivariate time series SCADA data: Obtain historical multivariate time series data of wind turbine units from the wind farm SCADA system, perform data cleaning to remove missing values ​​and outliers, and filter experimental variables and covariates. S2. Eliminate the influence of environmental variables and changes in operating conditions on experimental variables, obtain temperature sensor data after covariate preprocessing, construct a random forest multi-input multi-output regression model based on Bayesian optimization algorithm, and initialize the spectral domain graph convolutional network model. S3. Input the health data after covariate preprocessing into the spectral domain graph convolutional network to learn correlation features: Construct a prediction model based on the spectral domain graph convolutional network. First, use the self-attention mechanism to extract the spatial correlation features between different sensor variables in the time series data of the state variables after covariate preprocessing, and calculate the adjacency matrix representing the spatial correlation of sensors. At the same time, input the time series data of the state variables after covariate preprocessing into the spectral domain graph convolutional module to extract the features of the time dimension, and input it together with the adjacency matrix into the graph convolutional module to extract the features of the spatial dimension. Use the training set to train the prediction model based on the spectral domain graph convolutional network. S4. Set thresholds based on the training set: Input the training set into the prediction model based on the spectral domain graph convolutional network to be trained, calculate the root mean square error between the predicted and true values ​​of all variables in the training set, and design a threshold based on the kernel density estimation method. S5. Input the online SCADA test data into the model, and use the detection and localization module to determine whether an anomaly has occurred based on the threshold, and collect false alarm points and locate the fault: Online real-time monitoring multivariate time data is obtained from wind turbine units. This data is first input into a random forest multi-input multi-output regression model to calculate preprocessed temperature status data for covariates. Then, this data is input into a trained prediction model based on a spectral domain graph convolutional network. The detection and localization module compares the root mean square error (RMSE) between the predicted and actual values ​​with a pre-set threshold. When the RMSE exceeds the threshold, a fault warning is issued for the wind turbine components. Simultaneously, the RMSE of the test data is downsampled by day and by hour, and the detection and localization module locates the fault. Finally, the false alarm situation is determined based on the validation set and the threshold setting. S6. Input the online SCADA test data into the prediction model based on the spectral domain graph convolutional network, adaptively set the threshold according to the data correlation, visualize the result model through the decision interpretability module, realize fault warning and fault location, and optimize the model parameter structure according to the result optimization method. S7. Integrate the detection and positioning module and the decision interpretability module to conduct a scientific assessment and guidance on the health status of wind turbine units.

2. The method for anomaly detection and localization of wind turbine units based on interpretable graph neural networks according to claim 1, characterized in that: S1 specifically includes the following steps: S11. Use a density-based outlier detection algorithm to detect outliers in the original SCADA data and remove data that does not conform to physical meaning and missing values. S12. Select the temperature state variables and covariates that affect temperature relevant to the experiment for the next step of modeling.

3. The method for anomaly detection and localization of wind turbine units based on interpretable graph neural networks according to claim 1, characterized in that: S2 specifically includes the following steps: S21. Divide the entire experimental data into a training set, a validation set, and a test set, where the test set data is the experimental data; in the training set, use covariates as independent variables and temperature state variables as dependent variables, construct a random forest multi-input multi-output regression model based on the Bayesian optimization algorithm according to the number of sensors, and use the calculated regression model residuals as temperature state variables treated as covariates. S22. Use a sliding window to establish time series data with a fixed window size for the temperature state variables after covariate preprocessing. Select appropriate data according to different faults of different units and divide the data according to the ratio of training set: validation set: test set = 7:2:

1. Then initialize the spectral domain graph convolutional network model.

4. The method for anomaly detection and localization of wind turbine units based on interpretable graph neural networks according to claim 1, characterized in that: S3 specifically includes the following steps: S31. Transfer the training set data X∈R S×N Input a gated linear unit (GLU), where S is the sliding window size and N is the number of sensors for the temperature state variable. The hidden states corresponding to each window S are calculated sequentially, and the last hidden state R is used as the representation of the entire time series. A weight matrix W is calculated using a self-attention mechanism, and the adjoining matrix A of the sensor spatial coupling graph is calculated based on the weight matrix W, where A = 0.5 * W * W. T That is, to form a fixed graph structure G, G=(X,W) based on each fixed time window; S32. Input the fixed graph structure G generated in S31 into the graph Fourier layer to capture the relationships between internal sequences, resulting in: G F (X)=U T X (1) Where, U∈R N×S Let X ∈ R be the eigenvector matrix. S×N Indicates input data; S33. Input the results from S32 into the Discrete Fourier layer and the gated linear unit (GLU) to learn the frequency domain representation of the input time series, thereby capturing the repeating patterns in periodic data or the autocorrelation features between different timestamps. S34. Input the frequency representation from S33 into two one-dimensional convolutional layers and a gated linear unit (GLU) to capture the feature pattern, and obtain the feature h. s (X): Where b and c are bias parameters, and W and V are convolution kernel parameters; S35. The features learned in S34 are represented by a spectral matrix through a graph convolutional layer and then expressed by an inverse graph Fourier transform to obtain the result Z: Z=Z1||Z2…Z n (4) Wherein, gθ ij Let Λ be the i-th input and j-th output channels corresponding to the graph convolution kernel. i It is the normalized Laplacian eigenvalue matrix, where S represents the model layer mentioned above, and Z represents the output channel of each... j The final result Z is obtained by piecing them together. S36. Passing the result of S35 through a fully connected layer yields the final result.

5. The wind turbine anomaly detection and localization method based on interpretable graph neural networks according to claim 4, characterized in that: S4 specifically includes the following steps: S41. Calculate the root mean square error (RMSE) between the result of S36 and the original result, and obtain: in, X represents the root mean square error of the i-th data point. train_pre,i Let X represent the predicted value of the i-th data point in the training set. train_tru,i This represents the pre-true value of the i-th data point in the training set; S42. Using the kernel density estimation method, calculate the probability density of the root mean square error obtained in S41, and obtain: Where k represents the data length, and K(·) represents the kernel function, using a Gaussian kernel. σ represents bandwidth; S43. Set the confidence level. Integrate equation (6) to obtain the threshold, as shown in equation (7): Where α is the confidence level, p(x) is the probability density function of the data, and the threshold is obtained.

6. The method for anomaly detection and localization of wind turbine units based on interpretable graph neural networks according to claim 5, characterized in that: S5 specifically includes the following steps: S51. Compare the validation set data with the threshold set in S43 to obtain the false alarm rate, as follows: The false alarm rate is: S52. According to equation (5), the root mean square error (RMSE) is obtained by calculating the prediction results of the test set and the original data. Where S is the test set length and N is 24, representing the number of temperature state variables; S53. Downsample the values ​​in S52 by day and hour to obtain Rmse based on daily and hourly rates, denoted as: S531, in S53 Compared with the threshold, we get: S532. Downsample each temperature state variable in S52 by day and by hour to obtain: Where N is 24, representing the number of temperature state variables, and S is the test set length; S533. Visualize the results from S532 using a heatmap to obtain the location information.

7. The method for anomaly detection and localization of wind turbine units based on interpretable graph neural networks according to claim 6, characterized in that: S6 specifically includes the following steps: S61, the result obtained in S53 The matrix is ​​divided into separate matrices according to their respective units, and the correlation is calculated as follows: Where N is 24, representing the number of temperature state variables; ρ represents the correlation to be calculated; and S represents the length of the time series data. S62. Based on S61, draw the correlation kernel density distribution curve and divide the correlation at time T into three gradients, as follows: S63. Based on the kernel density distribution curve and gradient in S62, set the threshold for different time changes to 0.4, and obtain: S64. Input the experimental results data of S61 into the interpretability module according to different time scales. In the module, the correlation is mapped to the distance between the edges of different nodes Vertex. That is, the stronger the correlation, the shorter the distance between nodes, and the weaker the correlation, the farther the distance between nodes. S641. Using the distance features learned in S64, draw an undirected weighted graph G = (V, E, T), where V represents a node, i.e., the temperature state variable, E represents the edge between different nodes, i.e., the correlation strength is represented by distance, and T represents time. S65. Based on the results of S641, as well as the actual unit conditions and the correlation between different components, evaluate the results. If deviations occur, provide feedback to modify the model framework parameters.

Images