Wind turbine generator fault early warning method based on machine learning

Abstract

The invention discloses a wind turbine generator fault early warning method based on machine learning, and relates to the field of wind turbine generator fault early warning. The method comprises the following steps: preprocessing data; the method comprises the following steps: preprocessing historical sample data of a healthy fan, dividing the preprocessed historical sample data of the healthy fan into a training set and a test set, respectively inputting the training set into a Stacking fusion model and a Blending fusion model for training, carrying out comprehensive testing through the test set to obtain a Stacking offline model and a Blending offline model, respectively sending verification data to the Stacking offline model and the Blending offline model, and carrying out comprehensive testing through the test set to obtain a Blending offline model and a Stacking offline model; selecting an optimal model through a self-adaptive threshold value; and inputting actual data of the wind turbine generator into the optimal model to obtain whether the wind turbine generator has a fault or not. According to the method, on the premise that the prediction effect is guaranteed, fault early warning can be carried out early, and for the conditions that the operation range of a draught fan is wide, data is diversified and missing exists, the problems that the model precision and stability are reduced due to data missing, and overfitting is caused due to the fact that the model is too complex can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine fault warning, and in particular to a wind turbine fault warning method based on machine learning. Background Art

[0002] As traditional power systems gradually develop into new power systems, the scale of wind power is becoming larger and larger, and the problems it brings are also increasing.

[0003] When the wind turbine is not faulty or has already failed, the former is to start maintenance without any fault, which will affect the normal operation of the wind turbine and waste manpower and material resources; the latter is to repair after the failure occurs, which is easy to cause great losses to the wind turbine. When we can predict in advance when the failure needs maintenance, the cost of repairing and maintaining the wind turbine will be greatly reduced, significantly avoiding expensive corrective maintenance. Therefore, fault early warning has important practical significance.

[0004] Generally speaking, the key to fault warning lies in the sample data preprocessing of the prediction model and the prediction model itself. The current research direction is also mainly focused on optimizing the prediction model and selecting input parameters, that is, analyzing and processing the SCADA data obtained from the wind turbine under offline stable state. The offline training part can be divided into three steps: SCADA data preprocessing; building a data prediction model; setting the warning threshold.

[0005] Generally speaking, wind turbines are basically equipped with SCADA systems. However, due to the wide operating range of wind turbines during normal operation and the high safety threshold set by the system, the alarm may be too late and the fault may worsen seriously. In addition, when acquiring training data, due to the remote environment and harsh conditions of wind turbines, coupled with sensor failures, data transmission noise and loss, and limited power output, data will inevitably be lost and damaged during the collection, transmission, and storage process. At the same time, when building an early warning model using general methods, overfitting is prone to occur. Therefore, how to achieve timely early warning of wind turbines is an urgent problem to be solved in this field. Summary of the invention

[0006] In view of the above-mentioned defects of the prior art, the present invention provides a wind turbine fault warning method based on machine learning to improve the accuracy and timeliness of wind turbine fault warning.

[0007] A wind turbine fault early warning method based on machine learning of the present invention comprises the following steps:

[0008] S1, data acquisition step, the data includes healthy wind turbine historical sample data and verification data, and the data is preprocessed;

[0009] S2, model training step, the model includes a Stacking fusion model and a Blending fusion model, the pre-processed healthy fan historical sample data is divided into a training set and a test set, the training set is respectively input into the Stacking fusion model and the Blending fusion model for training, and a comprehensive test is performed on the test set to obtain a Stacking offline model and a Blending offline model;

[0010] S3, model screening step, sending the verification data to the Stacking offline model and the Blending offline model respectively, obtaining the residual values ​​of the predicted values ​​and the true values ​​of the two offline models respectively, comparing the two residual values ​​with the threshold values ​​respectively, and selecting the Blending offline model as the optimal model when the residual value of the Stacking offline model is greater than or equal to the threshold value, otherwise selecting the Stacking offline model as the optimal model;

[0011] S4. Input the actual data of the wind turbine generator set into the optimal model to determine whether the wind turbine generator set fails.

[0012] Furthermore, the data preprocessing method includes:

[0013] The Isolation Forest algorithm is used to remove abnormal data from the historical sample data of healthy fans;

[0014] The L-SARIMA filling algorithm is used to fill in the data in steps according to the missing data.

[0015] Furthermore, the Stacking fusion model includes a first learning layer and a second learning layer, the first learning layer includes an XGBoost module, an SVM module, a neural network module, an RF module, an LR module and a decision tree module, and the second learning layer includes a Lasso regression module.

[0016] Furthermore, the method for setting the residual threshold includes:

[0017] The output value and the true value are smoothed by using the moving weighted exponential average method;

[0018] The KS test method is used to select the length of the minimum data subset that can reflect the characteristics of the original parent data set as the size N of the sliding window for the verification data;

[0019] Calculate the window residual based on the data in the window Re , and according to the window residual Re Set the current threshold T, then move the data window frame by frame and set a new threshold.

[0020] Furthermore, the window residual Re for:

[0021] ;

[0022] Among them, σ is the standard deviation of the residual under normal circumstances, Re t =y t- -x t ,y t is the predicted value of the window data, x t The actual value of the window data.

[0023] Furthermore, the threshold T in the window is:

[0024] T = E(Re) + 3δ;

[0025] Among them, E(Re) and δ are the expectation and deviation of the residual data set within the window, respectively.

[0026] Furthermore, the verification data is historical data in a database or historical data of a target wind turbine generator set.

[0027] Compared with the prior art, the present invention has the following technical effects:

[0028] The Stacking and Blending integrated algorithms of the present invention construct a fault warning model. The two algorithms have the same core idea and can effectively overcome the limitations brought by a single algorithm, integrate various algorithms together, make the best use of their strengths and avoid their weaknesses, and give full play to the advantages of each algorithm. The two fusion models are used to perform early warning diagnosis of fault conditions, and the calculation response speed is fast, the application range is wide, the versatility is strong, the early warning effect is stable, and fault warning can be performed earlier under the premise of ensuring the prediction effect. In addition, for situations where the fan has a wide operating range and diverse data, it can effectively avoid overfitting caused by excessive complexity of the model.

[0029] The present invention uses a method combining machine learning with traditional statistical analysis to fill in the data of wind farms that have filtered out abnormal points, effectively maintaining the trend and characteristics of time series data and solving the missing problem that occurs when saving SCADA data. For early warning models that require a large amount of healthy wind farm data for training, the accuracy and stability of the model have been greatly improved.

[0030] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a general flow chart of a wind turbine fault warning method in a specific embodiment of the present invention;

[0032] Figure 2 is a sample data diagram before abnormal data is removed in a specific embodiment of the present invention;

[0033] Figure 3 This is a sample data diagram after abnormal data is removed in a specific embodiment of the present invention;

[0034] Figure 4 is a sample data graph before missing data filling in a specific embodiment of the present invention;

[0035] Figure 5 is a sample data graph after missing data is filled in a specific embodiment of the present invention;

[0036] Figure 6 is a structural schematic diagram of a Stacking fusion model in a specific embodiment of the present invention;

[0037] Figure 7 It is a structural diagram of a Blending fusion model in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0039] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present invention in a schematic manner, and thus the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0040] While some exemplary embodiments of the present invention have been described for the purpose of illustration, it should be understood that the present invention may be implemented in other ways not specifically shown in the drawings.

[0041] like Figure 1 As shown, a wind turbine fault early warning method based on machine learning of the present invention comprises the following steps:

[0042] S1, data acquisition step, the data includes healthy wind turbine historical sample data and verification data, and the data is preprocessed, the preprocessing method includes using IsolationForest algorithm to remove abnormal data from the healthy wind turbine historical sample data, and then using L-SARIMA filling algorithm to fill the data; In this embodiment, the data is SCADA data obtained based on the offline stable state unit, including electrical data such as voltage, current, power, and physical data such as gearbox oil temperature and speed. Due to the remote geographical location, harsh environment and changeable working conditions, sensor failures, noise and loss in data transmission, and equipment power output and abnormal problems are often caused, which inevitably affect the integrity and quality of the data, and thus have a great impact on the early warning model that needs to be trained with the health data of the fan. Therefore, it is of great significance for us to remove abnormal values ​​and supplement missing data. The current target early warning fan real-time data is the SCADA data of the target early warning fan. The healthy historical sample data is used as the training set, and the normal operation data of the target fan is used as the test set. The above data is processed as follows:

[0043] S11. Eliminate abnormal data through the Isolation Forest algorithm.

[0044] Since the collected data base is large, there are many meaningless redundant data. In order to effectively eliminate some abnormal and meaningless data, clustering algorithms such as KMEANS and DBSCAN are usually used in the prior art to cluster the data. However, clustering algorithms have disadvantages such as high computational cost and high sensitivity in parameter selection. They are no longer suitable for wind turbine data with large data volume and high dimension. Therefore, this embodiment uses the Isolation Forest algorithm to eliminate abnormal data in the wind turbine SCADA data. It isolates sample points by constructing random decision trees, and uses the path length required for the sample points to be isolated to determine whether the samples are abnormal points. Sample points with short path lengths are more likely to be abnormal points. Finally, an abnormal score value is set, and data greater than this value is eliminated. The process of data processing by the Isolation Forest algorithm includes:

[0045] S111, data collection step, randomly extracting a sub-sample set from the training data set.

[0046] S112, tree structure construction step, for each sub-sample set, an Isolation Tree is constructed. During the tree construction process, a feature is randomly selected and split on the random value of the feature until each sample point is individually isolated.

[0047] S113, a path length calculation step, calculating the average path length of each sample point in all isolated trees, where the path length is the distance from the root node to the leaf node of the tree.

[0048] S114, anomaly scoring step, setting a scoring threshold, calculating anomaly scores according to the path length, the shorter the path length, the higher the anomaly score, when the anomaly score is greater than the scoring threshold, it is eliminated.

[0049] S12. Perform data filling using the L-SARIMA filling algorithm.

[0050] The Linear interpolation used in the prior art can quickly fill in short missing data segments distributed throughout the data source, while also ensuring continuity, but it is not effective for slightly longer missing data segments; for longer missing segments, the SARIMA algorithm is used, which combines seasonal, autoregressive and differential components. Although the calculation speed is slower than that of Linear interpolation, it works well for time series data with seasonal components such as wind turbines; finally, for very long data missing segments with a missing length greater than several hours, if the SARIMA model is still used, the effect will be distorted, so the data at the same time of the previous one to three days is directly used for filling. Based on the defects existing in the prior art, the improved SARIMA filling algorithm is used in this embodiment for data filling. This method combines preliminary filling and further filling, which not only ensures the continuity and rationality of the data, but also takes into account the calculation speed, especially for the complex and changeable missing situations of wind turbine data. It has a good supplementary effect. The process of data filling using the improved SARIMA filling algorithm includes:

[0051] S121. Use Linear interpolation to fill in the missing data whose length is less than 3. The Linear interpolation formula is:

[0052] ;

[0053] Where (x0, y0) and (x1, y1) are known interpolation points, x is the point whose corresponding y value we want to know, and x is between x0 and x1, and y is the interpolation result at point x.

[0054] S122. For missing segments with missing data length greater than 3 and less than 48, first use time series decomposition to identify the seasonal cycle of the data and check the trend, seasonality and residual parts. Then use the ADF (Augmented Dickey-Fuller) test to determine whether the time series is stationary. If the time series is non-stationary, perform first-order difference or seasonal difference to make it stationary.

[0055] S123. Determine SARIMA model parameters.

[0056] First, the difference order d and the seasonal difference order D are determined by analyzing the truncation characteristics of the autocorrelation function (ACF) and the partial autocorrelation function (PACF); then the order q of the moving average part and the order Q of the seasonal moving average part are determined by the ACF diagram, and the order p of the autoregressive part and the order P of the seasonal autoregressive part are determined by the PACF diagram; then the ACF diagram is searched for significant changes in the autocorrelation coefficient with the lag order. If this significant pattern is repeated every k orders, then the k value is used as the seasonal period S. Finally, the maximum likelihood estimation is used to fit the parameters p, d, q, P, D, Q and S.

[0057] S124. Substitute the determined model parameters into the SARIMA model, traverse all missing segments with the fitted SARIMA model parameters, and fill in the missing time points.

[0058] S125. Use the data from the same time in the previous one to three days to fill in the missing data segments with a length greater than 48.

[0059] Taking 17425 sets of data on gearbox oil temperature in the first five months of a wind farm in Northeast China in 24 years as an example, the results before and after the abnormal data are removed are as follows: Figure 2 and Figure 3 As shown in the figure, the comparison of the 17425 sets of data after removing the abnormal data is shown in Figure 4 and Figure 5 shown.

[0060] S2. Model training step. The model includes a Stacking fusion model and a Blending fusion model. The training sets are respectively input into the Stacking fusion model and the Blending fusion model for training. A comprehensive test is performed through the test set to determine the fault warning model under the current wind turbine operating condition. After cluster analysis and data dimensionality reduction, two fault warning models suitable for different situations are constructed through the Stacking and Blending integrated algorithm ideas. Both of them can integrate the data information obtained by training with multiple different algorithms, which can effectively overcome the limitations brought by a single algorithm, give full play to the advantages of each algorithm, and learn from each other's strengths and weaknesses.

[0061] S21. Use healthy historical sample data to train the Stacking fusion model to obtain the Stacking offline model.

[0062] like Figure 5As shown, it is a structural principle diagram of the Stacking fusion model used in this embodiment. The Stacking fusion model includes a first learning layer and a second learning layer. In this embodiment, the first learning layer is composed of 6 base learners, and the second learning layer is a meta learner, and the meta learner includes a Lasso regression model. During the training process, the original training data is first trained and learned by each base learner, and then the outputs of each base learner are stacked to form a new data set, and then the new data set is handed over to the second-layer meta learner for logistic regression, that is, the output of the first-layer base learner is used as a feature for fitting output to predict the result.

[0063] In this embodiment, the first learning layer selects XGBoost module, SVM module, neural network module, RF module, LR module and decision tree module with higher fitting degree, and the second learning layer adopts the relatively simple generalized linear model of Lasso regression. In this embodiment, for the Blending algorithm, the core idea is exactly the same as Stacking, both of which are connected in series through two layers of algorithms. The first layer is composed of multiple base learners, and the second layer has only one meta learner. Data segmentation is also consistent with the above-mentioned Stacking algorithm. After separation, the training set is input into the first layer of base learners. After training, it is verified on the verification set, and then the new verification results are trained by inputting the meta learner.

[0064] In this embodiment, the training method of the Stacking fusion model includes:

[0065] S211, after the historical sample data is trained and learned through the first learning layer, the outputs of each base learner are stacked to obtain a new training data set and a new test data set.

[0066] In order to solve the problem of too few samples in the feature matrix of the meta-learner, the method of distinguishing the training set and the validation set and separating cross-validation is used again. The preprocessed data is divided into a training set and a test set, and then the training set is divided into several parts, one of which is set as the validation set, and the remaining four are training sets. After the model is trained with a certain training set, the validation set and the test set are predicted immediately. This cycle is repeated until all combinations are traversed, and then different algorithms are used for cross-training and testing. Finally, the stacked new data set is passed to the second-layer meta-learner for regression fitting and output.

[0067] In this embodiment, the algorithms of the XGBoost module, SVM module, neural network module, RF module, LR module and decision tree module in the first learning layer are arranged in sequence as Algorithm 1, Algorithm 2, Algorithm 3, Algorithm 4, Algorithm 5 and Algorithm 6, and the data set is divided into a training set and a test set. The training set is divided into five parts, one of which is used as a validation subset and the remaining four are used as training subsets.

[0068] Taking Algorithm 1 as an example, after training and verification with four training subsets and one validation subset, Algorithm 1 obtains a single prediction model 1-1, and at the same time obtains the prediction training value 1-1; another training set is used as the validation subset, and the remaining four training sets are used as training subsets, and Algorithm 1 is continued to be trained to obtain a single prediction model 1-2 and a prediction training value 1-2; the training subsets and validation subsets are replaced until all data sets are traversed to obtain a single prediction model 1-3 and a prediction training value 1-3, a single prediction model 1-4 and a prediction training value 1-4, a single prediction model 1-5 and a prediction training value 1-5, and the test set data is input into the single prediction model 1-1 to the single prediction model 1-5 to obtain a prediction test set 1.

[0069] Algorithm 2-Algorithm 6 also follows the above steps. Algorithm 2 is verified through cross-training to obtain a single prediction model 2-1 and prediction training value 2-1, a single prediction model 2-2 and prediction training value 2-2, a single prediction model 2-3 and prediction training value 2-3, a single prediction model 2-4 and prediction training value 2-4, a single prediction model 2-5 and prediction training value 2-5. The test set data is input into the single prediction model 2-1 to the single prediction model 2-5 to obtain the prediction test set 2. Algorithm 3 is verified through cross-training to obtain a single prediction model 3-1 and prediction training value 3-1, a single prediction model 3-2 and prediction training value 3-2, a single prediction model 3-3 and prediction training value 3-3, a single prediction model 3-4 and predicted training value 3-4, a single prediction model 3-5 and predicted training value 3-5, the test set data is input into the single prediction model 3-1 to the single prediction model 3-5 to obtain the predicted test set 3; algorithm 4 is verified by cross-training to obtain a single prediction model 4-1 and predicted training value 4-1, a single prediction model 4-2 and predicted training value 4-2, a single prediction model 4-3 and predicted training value 4-3, a single prediction model 4-4 and predicted training value 4-4, a single prediction model 4-5 and predicted training value 4-5, the test set data is input into the single prediction model 4-1 to the single prediction model 4-5 to obtain the predicted test set 3 Test set 4; Algorithm 5 is verified by cross-training to obtain a single prediction model 5-1 and prediction training value 5-1, a single prediction model 5-2 and prediction training value 5-2, a single prediction model 5-3 and prediction training value 5-3, a single prediction model 5-4 and prediction training value 5-4, a single prediction model 5-5 and prediction training value 5-5, and the test set data is input into the single prediction model 5-1 to the single prediction model 5-5 to obtain the prediction test set 5; Algorithm 6 is verified by cross-training to obtain a single prediction model 6-1 and prediction training value 6-1, a single prediction model 6-2 and prediction training value 6-2, a single prediction model 6-3 and prediction training value 6-3, a single Prediction model 6-4 and prediction training value 6-4, single prediction model 6-5 and prediction training value 6-5, the test set data is input into single prediction model 6-1 to single prediction model 6-5 to obtain prediction test set 6; the above-mentioned prediction training value 1-1 to prediction training value 1-5, prediction training value 2-1 to prediction training value 2-5, prediction training value 3-1 to prediction training value 3-5, prediction training value 4-1 to prediction training value 4-5, prediction training value 5-1 to prediction training value 5-5, prediction training value 6-1 to prediction training value 6-5 are superimposed as a new training data set, and the above-mentioned prediction test set 1 to prediction test set 6 are superimposed as a new test data set.

[0070] S212: The new training data set and the new test data set are subjected to logistic regression of a meta-learner to obtain a trained final Stacking fusion model, namely, a Stacking offline model.

[0071] S22. Use healthy historical sample data to train the Blending fusion model to obtain a Blending offline model.

[0072] The Blending fusion model of this embodiment is exactly the same as Stacking, and both use two layers of algorithms in series. There are multiple strong learners in the first layer, and there is only one meta-learner in the second layer. The data segmentation method of the training set and the test set is consistent with the above-mentioned Stacking algorithm. The training set is then input into the first layer of base learners. After the training is completed, it is verified on the verification set, and then the new verification result is input into the meta-learner for training. The algorithm selection in the Blending integration model is consistent with Stacking. The algorithm selection of each layer of the Blending fusion model is consistent with that of the Stacking fusion model. The algorithm models used in the first layer are six algorithms with higher fitting degrees, including XGBoost, SVM, neural network, RF, LR and decision tree. The neural envelope selects a neural network algorithm with higher fitting degree, such as LSTM; the second layer adopts the relatively simple Lasso regression, a generalized linear model. The model in the first layer is trained by training the subset to obtain the corresponding predicted training value, and then the validation subset is input into the trained model prediction to obtain the corresponding predicted test value. The predicted training values ​​corresponding to different algorithms are superimposed to form a new training data set, and the predicted test values ​​corresponding to each algorithm are superimposed to form a new test data set. The new training data set and the new test data set are subjected to the logistic regression of the meta-learner to obtain the trained final Blending fusion model, i.e., the Blending offline model. The Blending algorithm reduces the problem of the Stacking algorithm where both the data and the algorithm are too complex, which easily leads to the problem of model overfitting, cancels the cross-validation process, and greatly reduces the amount of training data for the meta-learner.

[0073] S3, model screening step, sending the verification data to the Stacking offline model and the Blending offline model respectively, obtaining the residual values ​​of the predicted values ​​and the true values ​​of the two offline models respectively, comparing the two residual values ​​with the threshold values ​​respectively, when the residual value of the Stacking offline model is greater than or equal to the threshold value, selecting the Blending offline model as the optimal model, otherwise selecting the Stacking offline model as the optimal model.

[0074] In this embodiment, the residual is compared with the threshold value. When the parameter value reaches the threshold value, it is determined that the fan is abnormal at this time. It can be seen that the setting of the threshold value is directly related to the stability of the final fault warning model alarm. The existing threshold value is usually considered to set one or more values, which has the disadvantages of weak generalization ability of fixed thresholds, easy false alarms, and strong subjectivity of artificial settings. In order to solve this problem, this embodiment adopts a method for dynamically adaptively setting the threshold value, which specifically includes:

[0075] S31, using a moving weighted exponential average method to smooth the output value and the true value;

[0076] In order to solve the problem that the residual of the fan may have extreme values ​​when it is in complex and severe conditions, the moving weighted exponential average method is used to smooth the output value and the true value. The accumulated order residual information in the past is fully considered in the form of setting the weighted coefficient to better reflect the recent changes in the residual situation: , where Re t is the residual at time t, which is: Re t =y t- -x t ,y t is the predicted value, x t is the actual value; β is the weight coefficient. In this embodiment, β =0.9.

[0077] S32. The training set in the verification data is used as the parent data set, and a child data set is selected from the parent data set to perform a KS test with the parent data set. When the k value between the subset and the parent data set is greater than 0.05, it can be considered that the two data sets have the same distribution law, and then the length of the minimum data subset that can reflect the characteristics of the original data set is selected as the size N of the sliding window.

[0078] S33, calculate the window residual according to the data in the window Re , and set the current threshold, then move the data window frame by frame, and set a new threshold, and finally get the threshold at all times.

[0079] The window residual Re for:

[0080] ;

[0081] Wherein, σ is the standard deviation of the residual under normal conditions, and the threshold is calculated by a fixed threshold calculation method, including the following steps:

[0082] In calculating the residual Re After that, the expected E( R e) and deviation δAccording to the 3-sigma criterion, the threshold T in the window is:

[0083] T=E( Re )+3 δ ;

[0084] This embodiment takes full account of the residual variation trend of the previous period in the window when setting the dynamic threshold. Connect the thresholds at each moment and draw a fitted residual. R Adaptive threshold diagram of trends; the thresholds at various time states can be clearly obtained.

[0085] After obtaining two fault warning models and dynamic thresholds, the offline training part is completed, and then the verification data set is input for comprehensive testing steps to finally determine the warning model. After the verification data set is input into the two fusion models, the predicted value and the residual between it and the true value are obtained, and compared with the corresponding threshold to obtain the warning time. When the residual of the above model reaches or exceeds the threshold, it means that the residual is large at this time, indicating that the predicted value output by the model at this time has a large deviation from the actual operation result. Since the model of this embodiment is trained with healthy data, when the deviation between the actual operation value of the fan and the predicted value of the healthy model reaches the threshold, it indicates that the fan at this time is likely to have or is about to fail, and the time from this time point to the time point when the actual failure occurs is the warning time.

[0086] In this embodiment, if the warning model based on the Stacking algorithm has an overfitting phenomenon, the Blending fusion model with better fusion effect is selected as the warning model. If there is no overfitting phenomenon, the Stacking fusion model is selected as the warning model, because in this case, the warning model based on the Blending algorithm is prone to a rapid decline in learning ability. The overfitting means that the degree of fitting is too high, that is, it performs well when training with the original data set, but cannot be predicted well when verifying. We can think that the error is too large from the actual value, which is overfitting. The verification data is the historical data in the database; in order to further improve the prediction accuracy, the verification data can use the historical data of the target wind turbine.

[0087] S4. Input the actual data of the wind turbine generator set into the optimal model to determine whether a fault occurs in the wind turbine generator set.

[0088] The historical data of the target wind turbine are used to test the trained Stacking fusion model and Blending fusion model through steps S1 to S3. After the fusion model without overfitting is screened out, the real-time data of the target wind turbine is sent to the screened fusion model to obtain the early warning result of the target wind turbine. Even when the wind turbine has a wide operating range, the conditions between wind farms vary greatly, and the data is diverse, it can still have a better fault warning effect and has better anti-overfitting and generalization capabilities.

[0089] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A wind turbine fault early warning method based on machine learning, characterized in that: The steps include: S1, data acquisition step, the data includes healthy wind turbine historical sample data and verification data, and the data is preprocessed; S2, model training step, the model includes a Stacking fusion model and a Blending fusion model, the pre-processed healthy fan historical sample data is divided into a training set and a test set, the training set is respectively input into the Stacking fusion model and the Blending fusion model for training, and a comprehensive test is performed on the test set to obtain a Stacking offline model and a Blending offline model; S3, model screening step, sending the verification data to the Stacking offline model and the Blending offline model respectively, obtaining the residual values ​​of the predicted values ​​and the true values ​​of the two offline models respectively, comparing the two residual values ​​with the threshold values ​​respectively, and selecting the Blending offline model as the optimal model when the residual value of the Stacking offline model is greater than or equal to the threshold value, otherwise selecting the Stacking offline model as the optimal model; S4. Input the actual data of the wind turbine generator set into the optimal model to determine whether a fault occurs in the wind turbine generator set.

2. The wind turbine fault early warning method based on machine learning according to claim 1 is characterized in that: The data preprocessing method comprises: The Isolation Forest algorithm is used to remove abnormal data from the historical sample data of healthy fans; The L-SARIMA filling algorithm is used to fill in the data in steps according to the missing data.

3. The wind turbine fault early warning method based on machine learning according to claim 1 is characterized in that: The Stacking fusion model includes a first learning layer and a second learning layer, the first learning layer includes an XGBoost module, an SVM module, a neural network module, an RF module, an LR module and a decision tree module, and the second learning layer includes a Lasso regression module.

4. The wind turbine fault early warning method based on machine learning according to claim 1 is characterized in that: The method for setting the residual threshold includes: The output value and the true value are smoothed by using the moving weighted exponential average method; The KS test method is used to select the length of the minimum data subset that can reflect the characteristics of the original parent data set as the size N of the sliding window for the verification data; Calculate the window residual based on the data in the window Re , and according to the window residual Re Set the current threshold T, then move the data window frame by frame and set a new threshold.

5. A wind turbine fault early warning method based on machine learning according to claim 4, characterized in that: The window residual Re for: ； Among them, σ is the standard deviation of the residual under normal circumstances, Re t =y t- -x t ,y t is the predicted value of the window data, x t The actual value of the window data.

6. A wind turbine fault early warning method based on machine learning according to claim 5, characterized in that: The threshold T within the window is: T=E( Re )+3δ; Among them, E(Re) and δ are the expectation and deviation of the residual data set within the window, respectively.

7. The wind turbine fault early warning method based on machine learning according to claim 1 is characterized in that: The verification data is historical data in a database or historical data of a target wind turbine generator set.

Images