Fault State Detection Method for Wind Turbine Based on Spatiotemporal Neural Network

Abstract

The present invention relates to a method for detecting the fault state of a wind turbine based on a spatio-temporal neural network, an electronic device, and a storage medium. The method includes acquiring real-time detection data of sensors of the wind turbine to be measured and performing data preprocessing; inputting the detection data into a trained spatio-temporal feature extraction model to obtain the multi-dimensional residual data; calculating the multi-dimensional residual data using the Mahalanobis distance to obtain the real-time performance index of the spatio-temporal feature extraction model; inputting the performance index and the real-time detection data into a trained vector regression algorithm model to obtain a real-time dynamic threshold; and detecting the fault state of the wind turbine based on the real-time dynamic threshold and the real-time performance index. The beneficial effect of the present invention is that it can solve the technical problems of long time-consuming fault detection, lagging fault feedback, low efficiency, large error, and waste of human and material resources in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine fault detection, and in particular to a wind turbine fault state detection method based on spatiotemporal neural network. Background Art

[0002] Wind energy has many advantages such as cleanliness, large reserves, and easy use. It is a renewable energy source with great development potential. With the development of modern society, the wind power industry has risen rapidly, and wind power technology has also continued to develop and become mature and perfect. However, since many wind turbines are installed in relatively harsh working environments such as islands and mountains, they are exposed to dust, rain, high temperature, snow and other environments for a long time. With the influence of wind force and impact with uncertain direction and load, wind turbine failures occur frequently, which seriously affects the reliability and safety of wind turbine operation.

[0003] In the prior art, wind farms widely use Supervisory Control and Data Acquisition (SCADA) systems. However, due to the huge amount of data for fault detection based on SCADA data, manual analysis takes a long time, and there are defects such as untimely fault feedback, low efficiency, large errors, and waste of manpower and material resources. In addition, there are also fault detection methods in the prior art that study single components. Due to the small detection range, there are also defects such as high omissions and false detection rates. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a wind turbine fault status detection method, electronic device and storage medium based on spatiotemporal neural network, which solves the technical problems in the prior art that the fault detection technology based on SCADA data is time-consuming, has delayed fault feedback, low efficiency, large errors, and wastes manpower and material resources.

[0006] (II) Technical solution

[0007] In order to achieve the above object, in a first aspect, the present invention provides a method for detecting a fault state of a wind turbine generator set based on a spatiotemporal neural network, characterized in that it comprises:

[0008] S1, obtaining real-time detection data of the sensor of the wind turbine to be tested and performing data preprocessing; the detection data includes: wind speed, power, rotation speed, pitch angle, temperature and detection time;

[0009] S2, inputting the detection data into the trained spatiotemporal feature extraction model to obtain the multidimensional residual data; the multidimensional residual data is the residual of the real-time detection data and the output of the spatiotemporal feature extraction model;

[0010] The spatiotemporal feature extraction model includes a deep autoencoder and a gated recursive unit, wherein the deep autoencoder is used to extract the spatial features of the detection data; the gated recursive unit is used to extract the temporal features of the detection data;

[0011] S3, using Mahalanobis distance to calculate the multidimensional residual data to obtain a real-time performance index of the spatiotemporal feature extraction model;

[0012] S4, inputting the performance index and the real-time detection data into the trained vector regression algorithm model to obtain a real-time dynamic threshold;

[0013] S5. Detecting a fault state of the wind turbine generator system based on the real-time dynamic threshold and the real-time performance index.

[0014] Optionally, the data preprocessing in S1 mainly includes:

[0015] Adopting a fusion strategy based on ideal wind speed and power, analyzing the detection data, setting unnecessary data points as outliers and removing them;

[0016] The local mean filling strategy is used to fill the detection data after removing outliers;

[0017] Normalize the padded detection data.

[0018] Optionally, the S3 further includes smoothing the performance index by exponentially weighted moving average to obtain an optimized performance index;

[0019] The calculation formula for obtaining the performance index using the Mahalanobis distance is:

[0020]

[0021] The e is the reconstruction error of the spatiotemporal feature extraction model, and μ is the dimensional mean of the detection data;

[0022] The calculation formula for smoothing the performance index is:

[0023] RE t =λE t +(1-λ)RE t-1 ;

[0024] The λ is a smoothing coefficient.

[0025] Optionally, the S5 is specifically:

[0026] comparing the dynamic threshold to the performance index;

[0027] If the dynamic threshold < the performance index, it is determined that the wind turbine is in a fault state;

[0028] If the dynamic threshold ≥ the performance index, it is determined that the wind turbine is in a normal operating state.

[0029] Optionally, before the S1, it further includes S0. Based on the historical detection data of the wind turbine to be measured, train the spatio-temporal feature extraction model and the support vector regression model;

[0030] The S0 includes:

[0031] S01. Based on the historical detection data of the wind turbine to be measured, train the spatio-temporal feature extraction model; specifically:

[0032] S011. Obtain the historical detection data of the wind turbine to be measured and perform data preprocessing, and use the historical detection data without fault state as the training data set;

[0033] S012. With the help of a pre-constructed regularization method, input the training data set into the pre-constructed spatio-temporal feature extraction model for iteration, and output the spatio-temporal feature reconstruction data at each iteration stage;

[0034] S013. Based on the spatio-temporal feature reconstruction data, use the pre-constructed gradient descent rule and loss function to calculate the target update weight at each iteration stage until the loss function converges, and complete the training of the spatio-temporal feature extraction model;

[0035] The loss function is:

[0036]

[0037] The is the training data, is the spatio-temporal feature reconstruction data output by the spatio-temporal feature extraction model data.

[0038] Optionally, the S0 further includes:

[0039] S02. Based on the historical detection data of the wind turbine to be measured, train the support vector regression model, specifically:

[0040] S021. Obtain the training data input to the pre-constructed support autoregressive model;

[0041] The training data includes the historical detection data of the wind turbine to be measured after preprocessing and the performance index of the wind turbine with the performance index to be measured;

[0042] The performance index is obtained by inputting the historical detection data into the trained spatio-temporal feature extraction model, calculating the residual between the historical detection data and the output value of the spatio-temporal feature extraction model using the Mahalanobis distance, and obtaining the performance index of the spatio-temporal feature extraction model.

[0043] S022. Input the training data into a pre-constructed support vector regression model and train the support vector regression model. Among them, the parameters to be selected for the support vector regression model are obtained through the grey wolf optimization algorithm.

[0044] Optionally, the parameters to be selected for the vector regression model include: penalty factor C and kernel function parameter σ.

[0045] In S022, the parameters to be selected for the support vector regression model obtained based on the grey wolf optimization algorithm specifically include:

[0046] S0221. Initialize the parameters of the grey wolf algorithm and the support vector regression model. The parameters are the grey wolf population size N, the maximum allowable number of iterations tmax, and the value ranges of the penalty factor C and the kernel function parameter σ.

[0047] S0222. Initialize the population using the pre-constructed good point set rule to determine the initial values of the positions (C, σ) of each grey wolf.

[0048] S0223. Input the training data into the support vector regression model and calculate the fitness of each grey wolf at this initial value.

[0049] S0224. Select the top 3 grey wolves with the best fitness and update the positions (C, σ) of the 3 grey wolves.

[0050] S0225. Calculate the fitness of all individual grey wolves, compare the fitness before and after the position update. If the current value is better than the fitness obtained in the previous iteration, update the positions of the top three grey wolves with the best fitness, otherwise the position is not updated.

[0051] S0226: Compare the current iteration number with the maximum allowable number of iterations. If it has not reached N, continue the optimization, otherwise the training ends. The position (C, σ) value of the lead wolf is the optimal solution, and the corresponding support vector regression model training is completed.

[0052] Optionally, the deep autoencoder structure is 15-100-50-25-50-100-15, and the number of neurons in the gated recurrent unit is 100.

[0053] In a second aspect, the present invention provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor executes the computer program stored in the memory to implement the steps of the wind turbine fault state detection method according to any one of the above first aspects.

[0054] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the wind turbine fault state detection method according to any one of the above first aspects are implemented.

[0055] (III) Advantageous Effects

[0056] The present invention provides a wind turbine fault state detection method, an electronic device, and a storage medium based on a spatio-temporal neural network. The method, due to the strong coupling of the wind turbine, makes the SCADA data have spatial correlation and time dependence. It comprehensively considers the high-dimensional spatial correlation between different variables of the SCADA data and the time dependence of the same variable at different times. By training a spatio-temporal feature extraction model to extract the spatio-temporal features of the SCADA data, and through Mahalanobis distance, the data residuals input to and output from the spatio-temporal feature extraction model are transformed, converting the multi-dimensional output residuals into a one-dimensional performance index, reducing the detection difficulty. Also, based on the variable operating state characteristics of the wind turbine, a dynamic threshold is calculated based on the Grey Wolf Optimizer and the pre-constructed Support Vector Regression (GWO-SVR). Compared with the prior art, it can simultaneously consider the time dependence and spatial correlation of the wind turbine SCADA data, improving the accuracy of wind turbine fault detection. By reducing the dimension of the data through Mahalanobis distance, the detection difficulty is reduced, and the dynamic threshold is adaptively set, improving the final detection accuracy of the wind turbine, achieving the purposes of a wide detection range, timely detection and feedback, high efficiency, small error, and high accuracy. Description of the Drawings

[0057] Figure 1 is a flowchart of a wind turbine fault state detection method based on a spatio-temporal neural network proposed in an embodiment of the present invention;

[0058] Figure 2 (a) and Figure 2 (b) are comparative diagrams of the wind speed-power curve distributions before and after data cleaning provided in an embodiment of the present invention;

[0059] Figure 3 is a logical flowchart of training a spatio-temporal feature extraction model and a support vector regression model provided in an embodiment of the present invention;

[0060] Figure 4 Schematic diagram of the structure of the spatio-temporal feature extraction model provided by an embodiment of the present invention;

[0061] Figure 5 Schematic diagram of the flow of the grey wolf optimization algorithm provided by an embodiment of the present invention;

[0062] Figure 6 Schematic diagram of the confusion matrix of the wind turbine fault detection results provided by an embodiment of the present invention. Detailed implementation manners

[0063] For better explaining the present invention and facilitating understanding, the present invention will be described in detail below with reference to the accompanying drawings through specific implementation manners. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more clear and thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0064] Wind energy has many advantages such as cleanliness, large reserves, and convenient utilization, and is a renewable energy source with great development potential. The large-scale development and utilization of wind energy have promoted the rapid rise of the wind power generation industry. Due to the harsh placement environment, the reliability and safety of wind turbine operation cannot be well guaranteed. For this reason, a method for detecting the fault state of a wind turbine based on a spatio-temporal neural network proposed in an embodiment of the present invention, based on the Supervisory Control and Data Acquisition (SCADA) system widely used in wind farms, does not require additional sensors and other devices to be installed on the wind turbine, and collects various low-frequency signals such as power, wind speed, temperature, etc., for detecting the fault state of the wind turbine, reducing the fault detection cost.

[0065] As Figure 1 shown, the Figure 1 is a schematic diagram of the flow of the method for detecting the fault state of a wind turbine based on a spatio-temporal neural network provided by an embodiment of the present invention, and the method may include:

[0066] S1. Obtain the real-time detection data of the sensors of the wind turbine to be tested and perform data preprocessing; the detection data may include: wind speed, power, rotation speed, pitch angle, temperature, detection time, etc.

[0067] The data preprocessing may include data cleaning, data filling, and / or normalization processing, etc.

[0068] S2. Input the detection data into the trained spatiotemporal feature extraction model to obtain the multidimensional residual data; the multidimensional residual data is the residual between the real-time detection data and the output of the spatiotemporal feature extraction model.

[0069] The spatiotemporal feature extraction model includes a deep autoencoder and a gated recursive unit. The deep autoencoder is used to extract the spatial features of the detection data; the gated recursive unit is used to extract the temporal features of the detection data. The output of the spatiotemporal feature extraction model is the spatiotemporal feature reconstruction data of the input detection data.

[0070] S3, using Mahalanobis distance to calculate the multidimensional residual data to obtain a real-time performance index of the spatiotemporal feature extraction model;

[0071] S4, inputting the performance index and the real-time detection data into the trained vector regression algorithm model to obtain a real-time dynamic threshold;

[0072] S5. Detecting a fault state of the wind turbine generator system based on the real-time dynamic threshold and the real-time performance index.

[0073] Specifically, in one embodiment, the S5 is specifically:

[0074] comparing the dynamic threshold to the performance index;

[0075] If the dynamic threshold value is less than the performance index, it is determined that the wind turbine generator set is in a fault state;

[0076] If the dynamic threshold value is greater than or equal to the performance index, it is determined that the wind turbine generator set is in a normal operating state.

[0077] The present invention provides a wind turbine fault status detection method based on spatiotemporal neural network, which makes full use of SCADA detection data collected in real time, inputs them into a spatiotemporal feature extraction model and a support vector regression model respectively, calculates a real-time performance index and a real-time dynamic threshold, can quickly respond to the fault status of the wind turbine, improves feedback efficiency, reduces operation and maintenance costs, and is beneficial to the maintenance and repair of the wind turbine.

[0078] In a specific embodiment, in the wind turbine fault state detection method based on spatiotemporal neural network, the data preprocessing in S1 may include:

[0079] S11. Adopting a fusion strategy based on ideal wind speed and power, the detection data is analyzed, and unnecessary data points are set as outliers and removed; the data can be cleaned to remove abnormal real-time data detected by the SCADA system under the influence of interference factors such as wind abandonment, sensor failure, etc. of normally operating wind turbines.

[0080] In practical applications, the fusion strategy can be a hybrid model that takes into account the high-dimensional characteristics of the SCADA data of wind turbines and the relationships between known low-dimensional variables (such as wind speed-power, etc.), such as Figure 2 (a) and Figure 2 (b) are the comparison diagrams of the wind speed-power curve distributions before and after data cleaning provided by an embodiment of the present invention.

[0081] S12. Adopt the local mean filling strategy to fill the detection data after removing outliers; this can avoid the discontinuity of the detection data in time caused by data cleaning.

[0082] In one embodiment, in order to eliminate the influence of the data measurement scale on the accuracy of the final detection model and accelerate the convergence of the model, S13 can also be performed, that is, normalize the filled detection data. To normalize the detection data, methods such as Max-min normalization extreme value normalization can be used.

[0083] Specifically, in some embodiments, the formula of the local mean filling strategy is:

[0084]

[0085] X m represents the values of each sensor measured at time m to be filled, X i is the normal detection data, and K is the number of existing data around the missing data.

[0086] The formula for the normalization process can be:

[0087]

[0088] Furthermore, in this embodiment, for S2, the input detection data is a multi-dimensional data sequence. After being input into the trained spatio-temporal feature extraction model, spatio-temporal feature reconstruction data is output; based on the detection data and the spatio-temporal feature reconstruction data, a multi-dimensional residual sequence between the input and the output is obtained.

[0089] Furthermore, in one embodiment, considering the correlation relationships between various variables, S3 can be performed, that is, calculate the multi-dimensional residual sequence through the Mahalanobis distance and convert the multi-dimensional residual sequence into a one-dimensional performance index. In some other embodiments, due to the influence of SCADA data noise and the uncertainty in the modeling process, the performance index can also be smoothed through the exponentially weighted moving average EWMA to obtain an optimized performance index.

[0090] The calculation formula for obtaining the performance index by using the Mahalanobis distance is:

[0091]

[0092] where \(e\) is the reconstruction error of the spatio-temporal feature extraction model, and \(\mu\) is the dimensional mean of the detection data;

[0093] The calculation formula for smoothing the performance index is:

[0094] RE t =\(\lambda E\) t +(1 - \(\lambda\))RE t-1 ;

[0095] where \(\lambda\) is the smoothing coefficient.

[0096] In this embodiment, the preferred value of \(\lambda\) is 0.01. In some other embodiments, there may be other values of the smoothing coefficient, which are not limited here.

[0097] The fan resistance fault detection method provided in this embodiment can effectively eliminate abnormal data caused by wind abandonment or sensor failures, etc. by performing data cleaning, filling, and normalization on the real-time acquired detection data, increasing the rationality of the detection data and reducing the error rate; by converting multi-dimensional residual data into one-dimensional performance data through the Mahalanobis distance, the complexity of data monitoring is simplified, and the detection efficiency is effectively improved.

[0098] As Figure 3 shown, Figure 3 is the logical flow chart for training the spatio-temporal feature extraction model and the support vector regression model provided in an embodiment of the present invention. In this embodiment, specifically:

[0099] Before S1, perform S0, and train the spatio-temporal feature extraction model DAE-GRU and the support vector regression model based on the historical detection data of the wind turbine to be measured.

[0100] S0 includes:

[0101] S01. Train the spatio-temporal feature extraction model based on the historical detection data of the wind turbine to be measured; specifically:

[0102] S011. Obtain the historical detection data of the wind turbine to be measured and perform data preprocessing, and use the historical detection data in the non-fault state as the training data set.

[0103] In this embodiment, Represent the data collected by each sensor of the wind turbine within a preset sampling time period, which may include: wind speed, power, rotational speed, gearbox oil temperature, etc. l is the length of time. M is the dimension of the SCADA data, that is, the number of variables collected. The process of data preprocessing includes: data cleaning, data filling, and / or normalization processing, etc.

[0104] In practical applications, the data preprocessing process can be the same as the above-mentioned data preprocessing process for real-time monitoring.

[0105] Further, implement S012. By means of a pre-constructed regularization method, input the training data set into a pre-constructed spatio-temporal feature extraction model for iteration, and output the spatio-temporal feature reconstruction data at each iteration stage.

[0106] Specifically, as Figure 4 is a schematic structural diagram of the spatio-temporal feature extraction model provided by an embodiment of the present application. In Figure 4 shown in an embodiment, the spatio-temporal feature extraction model, also known as the spatio-temporal model normal behavior model based on DAE-GRU, includes a deep autoencoder and a gated recurrent unit.

[0107] Input the training data into the spatio-temporal feature extraction model, and the deep autoencoder extracts the spatial features of the training data ; f E represents the network structure of the deep autoencoder, W E , B E are the parameters of the encoder, W E is the weight of the internal network of the deep autoencoder, B E is the bias of the internal network of the deep autoencoder; the gated recurrent unit is used to extract the time features of the training data; the output of the gated recurrent unit GRU at time t is: The output of the spatio-temporal feature extraction model is the spatio-temporal feature reconstruction data of the input detection data

[0108] Train the model by continuously reducing the root mean square error (Mean Square Error, MSE) between the model input data and the output data until the model training reaches the set number of iterations. The pre-constructed regularization strategy can be the regularization method of Dropout, which is used to prevent the model from overfitting and can improve the accuracy of the model.

[0109] In one embodiment, the gradient descent algorithm in the Adam optimizer can be used to train the loss function of the model to minimize its loss. The Adam optimizer replaces the first-order optimization algorithm of the traditional gradient descent process and updates the weights of the network iteratively using the gradient descent method based on the training data.

[0110] In this embodiment, M is preferably 15, the input of the gated recurrent unit is preferably a time window with a fixed length l = 12, the deep autoencoder structure of the deep autoencoder is preferably 15-100-50-25-50-100-15, the number of neurons in the gated recurrent unit is preferably 100, the amount of training data input in each batch is 100, the learning rate is preferably 0.001, and the dropout rate of the GRU neurons in the gated recurrent unit is preferably 0.2. In some other embodiments, the above various data are confirmed according to the actual situation and are not limited here.

[0111] Further, implement S013. Based on the spatio-temporal features to reconstruct the data, use the pre-constructed gradient descent rule and loss function to calculate the target update weights at each iteration stage until the loss function converges, and complete the training of the spatio-temporal feature extraction model.

[0112] In one embodiment, the gradient descent rule can be the gradient descent rule of the Adam optimizer, and the activation functions used for non-linear transformation of the data are all sigmoid to enhance the expression ability of the model.

[0113] Specifically, the loss function is a function using the root mean square error:

[0114]

[0115] The is the training data, is the spatio-temporal feature reconstruction data output by the spatio-temporal feature extraction model.

[0116] Further, the S0 further includes:

[0117] S02. Based on the historical detection data of the wind turbine to be measured and the trained spatio-temporal feature extraction model, train the support vector regression model. Specifically:

[0118] S021. Obtain the training data input to the pre-constructed support autoregressive model.

[0119] The training data includes the pre-processed historical detection data of the wind turbine to be measured and the performance index of the wind turbine to be measured.

[0120] The performance index is a spatio-temporal feature extraction model performance index obtained by inputting the historical detection data into the trained spatio-temporal feature extraction model and calculating the residual between the historical detection data and the output value of the spatio-temporal feature extraction model using the Mahalanobis distance. Specifically, the training data are all normal data collected from the wind turbine to be measured under normal operating conditions.

[0121] S022. Input the training data into a pre-constructed support vector regression model and train the support vector regression model. Among them, the parameters to be selected for the support vector regression model are obtained through the grey wolf optimization algorithm.

[0122] In one embodiment, the specific steps of training the support vector regression model may include:

[0123] Given the training data Input it into the support vector regression model to fit a hyperplane f(X)=ω(X)+b, so that the difference between f(X) and RE is minimized.

[0124] The loss function of this model can be formalized as:

[0125]

[0126] where ω is the weight, b is the corresponding bias, C is the penalty factor, ε is the maximum error allowed for regression, is the insensitive loss function.

[0127] The constraint can be expressed as:

[0128] By introducing slack variables, the loss function of the model is transformed into:

[0129]

[0130] The corresponding constraint is transformed into:

[0131] Introduce Lagrange multipliers to transform the loss function of the model into:

[0132]

[0133] The corresponding constraint is transformed into:

[0134] Finally, introduce the kernel function to convert the expression of the model into:

[0135]

[0136] where, is the Lagrange multiplier, K(X i,X) is a kernel function used to convert the inner product operation in the low-dimensional space into a kernel function operation in the high-dimensional space.

[0137] In this embodiment, the kernel function used is the Gaussian kernel function:

[0138]

[0139] where σ is the kernel function parameter.

[0140] Optimizing the penalty factor C and the kernel function parameter σ using the Grey Wolf algorithm can improve the accuracy of the support vector regression model. As Figure 5 shown, Figure 5 is a schematic flowchart of the Grey Wolf optimization algorithm provided in an embodiment of the present invention. Specifically, the parameters that need to be selected for the vector regression model (GWO-SVR regressor model) include: the penalty factor C and the kernel function parameter σ.

[0141] In S022, the parameters required to obtain the support vector regression model based on the Grey Wolf optimization algorithm specifically may include:

[0142] S0221. Initialize the parameters of the Grey Wolf algorithm and the support vector regression model. The parameters are the Grey Wolf population size N, the maximum allowable number of iterations tmax, and the value ranges of the penalty factor C and the kernel function parameter σ.

[0143] S0222. Initialize the population using the pre-constructed good point set rule to determine the initial values of the positions (C, σ) of each Grey Wolf.

[0144] S0223. Input the training data into the support vector regression model and calculate the fitness of each Grey Wolf at this initial value.

[0145] S0224. Select the top 3 Grey Wolves with the best fitness and update the positions (C, σ) of the 3 Grey Wolves.

[0146] S0225. Calculate the fitness of all individual Grey Wolves, compare the fitness before and after the position update. If the current value is better than the fitness obtained in the previous iteration, update the positions of the top three Grey Wolves with the best fitness, otherwise the position is not updated.

[0147] S0226: Compare the current iteration number with the maximum allowable number of iterations. If it has not reached N, continue the optimization, otherwise the training ends, and the position (C, σ) value of the leading wolf is the optimal solution, and the corresponding support vector regression model training is completed.

[0148] In some other embodiments, it further includes obtaining the test data sets of the spatio-temporal feature extraction model and the support vector regression model, and calculating the fault detection accuracy rate of the wind turbine, specifically including:

[0149] A1. Obtain the historical detection data of the wind turbine to be measured and perform data preprocessing. The historical detection data includes the detection data under normal operating conditions and the detection data under fault conditions.

[0150] A2. Input the historical detection data into the trained spatio-temporal feature extraction model and support vector regression model respectively to obtain the performance index and the dynamic threshold.

[0151] A3. Compare the magnitudes of the performance index and the dynamic threshold to determine the type (normal operation or fault) of the input historical detection data.

[0152] A4. Based on the judgment result, calculate the detection accuracy rate for detecting the faults of the wind turbine.

[0153] For example, in one embodiment, after calculation, the detection accuracy rate is obtained as 95% or above. As Figure 6 shown, it is a schematic diagram of the confusion matrix of the wind turbine fault detection result in one embodiment of the present invention.

[0154] As Figure 6 shown, in one embodiment, the detection accuracy rate for detecting the faults of the wind turbine is calculated. When the true label is 0, 1089 out of 1142 input test detection data are correctly identified, and 53 are misidentified; when the true label is 1, 498 out of 516 input test detection data are correctly identified, and 18 are misidentified.

[0155] In the above embodiment, first, the historical SCADA detection data of the wind turbine is collected and divided into a training data set and a test data set. The training data set is subjected to data cleaning to obtain the normal data of the wind turbine operation. A spatio-temporal neural network model based on DAE-GRU is established, and the normal data is input into the spatio-temporal model. Through training, the reconstruction error of the spatio-temporal neural network model is continuously reduced. Then, the output of the spatio-temporal neural network model is converted into a performance index for monitoring. By fitting the input normal data and the corresponding performance index, the GWO-SVR regressor model is trained to obtain an ideal regressor model and the dynamic threshold.

[0156] Since a too large sample size of the training dataset can affect the training of the regression model, and the choice of the penalty factor and kernel function of SVR itself can also affect the final training effect, in some embodiments, a grey wolf optimization algorithm is introduced to adaptively optimize the penalty factor C and kernel function σ in SVR, which can avoid the influence of manual setting on the regression result of the model. The test dataset is input into the trained spatio-temporal model to obtain the corresponding performance index. At the same time, the test dataset is also input into the trained regressor to obtain the corresponding dynamic threshold. Finally, the performance index and the dynamic threshold are compared to judge the detection accuracy of detecting the faults of the wind turbine by finally obtaining the operating state of the wind turbine.

[0157] The wind turbine is a complex system with strong coupling and nonlinearity. SCADA data is multi-dimensional time series. Due to the dependence and interaction between different subsystems in the wind turbine, each sensor variable has strong time dependence, and there is spatial correlation between different sensor variables. A method for detecting the fault state of a wind turbine based on a spatio-temporal neural network provided by the present invention takes into account the time dependence and spatial correlation of the SCADA data of the wind turbine, and detects the entire wind turbine, which can effectively avoid false detection and missed detection that may be caused by studying from a single component; adopts the symmetric structure of the autoencoder, fuses the spatial feature extraction ability of the autoencoder and the time feature extraction ability of the gated recurrent unit, and establishes a spatio-temporal feature extraction model based on the deep autoencoder and the gated recurrent unit, considering the correlation of variables in time and space. Comprehensively consider the correlation relationship between variables; use the Mahalanobis distance to convert the multi-dimensional residuals obtained through the spatio-temporal model into a one-dimensional performance index for monitoring. Adopt an algorithm based on the grey wolf optimization algorithm and support vector regression to realize the adaptive setting of the dynamic threshold, consider the situation of the variable operating state of the wind turbine, improve the final detection accuracy of the wind turbine, and can detect faults more timely, avoid the further deterioration of the faults, and improve the economic benefits.

[0158] In addition, the present invention also provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the computer program stored in the memory to implement the steps of the method for detecting the fault state of a wind turbine according to any one of the above embodiments.

[0159] In addition, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method for detecting the fault state of a wind turbine according to any one of the above embodiments are implemented.

[0160] The method for detecting the fault state of a wind turbine based on a spatio-temporal neural network provided in each of the above embodiments of the present invention does not require additional sensors or other devices to be installed on the wind turbine. It directly makes judgments based on the detection data of the existing widely used Supervisory Control and Data Acquisition (SCADA) system, with low implementation difficulty and wide applicability. Considering the high-dimensional spatial correlation between different variables of SCADA data, the time dependence of the same variable at different times, and the problem that fixed thresholds ignore the variable operating states of wind turbines, by training a spatio-temporal feature extraction model and a support vector regression model, this method has high detection accuracy and small errors, can achieve the effect of real-time detecting the fault state of the wind turbine, improve the operation reliability of the wind turbine and reduce the operation and maintenance costs, which is conducive to promoting the development of wind power, and further reducing environmental pollution caused by fossil energy consumption.

[0161] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more, unless otherwise specifically defined.

[0162] In the description of this specification, the description of terms such as "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0163] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A fault state detection method for wind turbine units based on spatio-temporal neural network, characterized in that, include: S1, obtaining real-time detection data of sensors of the wind turbine to be tested and performing data preprocessing; The detection data include: wind speed, power, rotation speed, pitch angle, temperature and detection time; S2, inputting the detection data into the trained spatiotemporal feature extraction model to obtain multidimensional residual data; the multidimensional residual data is the residual of the real-time detection data and the output of the spatiotemporal feature extraction model; The spatiotemporal feature extraction model includes a deep autoencoder and a gated recursive unit, wherein the deep autoencoder is used to extract the spatial features of the detection data; the gated recursive unit is used to extract the temporal features of the detection data; S3, using Mahalanobis distance to calculate the multidimensional residual data to obtain a real-time performance index of the spatiotemporal feature extraction model; S4, inputting the performance index and the real-time detection data into the trained vector regression algorithm model to obtain a real-time dynamic threshold; S5. Detecting a fault state of the wind turbine generator system based on the real-time dynamic threshold and the real-time performance index.

2. The method for detecting a fault state of a wind turbine generator set according to claim 1, characterized in that: The data preprocessing in S1 mainly includes: Adopting a fusion strategy based on ideal wind speed and power, analyzing the detection data, setting unnecessary data points as outliers and removing them; The local mean filling strategy is used to fill the detection data after removing outliers; Normalize the padded detection data.

3. The method for detecting a fault state of a wind turbine generator set according to claim 1, characterized in that: The S3 further includes smoothing the performance index by using an exponentially weighted moving average to obtain an optimized performance index; The calculation formula for obtaining the performance index using the Mahalanobis distance is: The e is the reconstruction error of the spatiotemporal feature extraction model, and μ is the dimensional mean of the detection data; The calculation formula for smoothing the performance index is: RE t = λE t +(1 - λ)RE t-1 ; The λ is a smoothing coefficient.

4. The method for detecting a fault state of a wind turbine generator set according to claim 1, characterized in that: The S5 is specifically: comparing the dynamic threshold to the performance index; If the dynamic threshold value is less than the performance index, it is determined that the wind turbine generator set is in a fault state; If the dynamic threshold value is greater than or equal to the performance index, it is determined that the wind turbine generator set is in a normal operating state.

5. The method for detecting a fault state of a wind turbine generator set according to claim 1, characterized in that: Before S1, the method further includes: S0, training a spatiotemporal feature extraction model and a support vector regression model based on historical detection data of the wind turbine to be tested; The S0 includes: S01. Based on the historical detection data of the wind turbine to be tested, the spatiotemporal feature extraction model is trained; specifically: S011, obtaining historical detection data of the wind turbine to be tested and performing data preprocessing, and using the historical detection data without fault status as a training data set; S012, by means of a pre-constructed regularization method, inputting the training data set into a pre-constructed spatiotemporal feature extraction model for iteration, and outputting spatiotemporal feature reconstruction data of each iteration stage; S013. Reconstruct the data based on the spatio-temporal features, and use the pre-constructed gradient descent rule and loss function to calculate the target update weights at each iteration stage until the loss function converges, thus completing the training of the spatio-temporal feature extraction model; The loss function is: The said is training data, is the spatio-temporal feature reconstruction data output by the spatio-temporal feature extraction model.

6. The wind turbine fault state detection method according to claim 5, characterized in that The S0 further includes: S02. Train the support vector regression model based on the historical detection data of the wind turbine to be measured, specifically: S021. Obtain the training data input to the pre-constructed autoregressive model; The training data includes the historical detection data of the wind turbine to be measured after preprocessing and the performance index of the wind turbine to be measured; The performance index is obtained by calculating the residual between the historical detection data and the output value of the spatio-temporal feature extraction model using the Mahalanobis distance after inputting the historical detection data into the trained spatio-temporal feature extraction model, and the performance index of the spatio-temporal feature extraction model is obtained; S022. Input the training data into the pre-constructed support vector regression model to train the support vector regression model, wherein the parameters to be selected for the support vector regression model are obtained through the grey wolf optimization algorithm.

7. The wind turbine fault state detection method according to claim 6, characterized in that The parameters to be selected for the vector regression model include: penalty factor C and kernel function parameter σ; In S022, obtaining the parameters to be selected for the support vector regression model based on the grey wolf optimization algorithm specifically includes: S0221. Initialize the parameters of the grey wolf algorithm and the support vector regression model, and the parameters are the grey wolf population size N, the maximum allowable number of iterations tmax, and the value ranges of the penalty factor C and the kernel function parameter σ; S0222. Initialize the population using the pre-constructed good point set rule to determine the initial values of the positions (C, σ) of each grey wolf; S0223. Input the training data into the support vector regression model to calculate the fitness of each grey wolf at this initial value; S0224. Select the top 3 grey wolves with the best fitness and update the positions (C, σ) of the 3 grey wolves; S0225. Calculate the fitness of all grey wolf individuals, compare the fitness before and after the position update. If the current value is better than the fitness obtained in the previous iteration, update the positions of the top three grey wolves with the best fitness, otherwise the position is not updated; S0226: Compare the current iteration number with the maximum allowable number of iterations. If it has not reached N, continue the optimization, otherwise the training ends, and the position (C, σ) value of the leading wolf is the optimal solution, and the corresponding support vector regression model training is completed.

8. The wind turbine fault state detection method according to claim 1, characterized in that The structure of the deep autoencoder is 15-100-50-25-50-100-15, and the number of neurons in the gated recurrent unit is 100.

9. An electronic device, characterized in that, It includes a memory and a processor. A computer program is stored in the memory, and the processor executes the computer program stored in the memory to implement the steps of the wind turbine fault state detection method according to any one of claims 1 to 8 above.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium. When the computer program is executed by the processor, it implements the steps of the wind turbine fault state detection method according to any one of claims 1 to 8 above.

Images