Wind turbine fault diagnosis method based on feature selection and stacking model fusion

Abstract

The application provides a wind turbine fault diagnosis method based on feature selection and stacking model fusion, comprising the following steps: S1, collecting SCADA data, wherein the SCADA data comprises wind speed, wind direction angle, gearbox temperature, blade angle and power generation; S2, performing SMOTE oversampling on small category data, increasing the proportion of the small category data and balancing the proportion of fault categories; S3, performing feature selection by using RFECV and dividing the data into a training set and a test set; S4, training a stacking model by using the training set, and obtaining a trained stacking model; and S5, inputting the SCADA data to be tested into the trained stacking model, and obtaining a wind turbine fault diagnosis result. Compared with other traditional methods, the application can obtain a more accurate wind turbine fault diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault diagnosis technology, and in particular to a wind turbine fault diagnosis method based on feature selection and stacking model fusion. Background Technology

[0002] In real-world data collection scenarios, data

[0003] The demand for data is not always met. For example, historical SCADA data is quite limited for new wind turbines that have been in operation for less than a year. Furthermore, data loss is very common for wind turbines located in remote areas with poor communication conditions. In these cases, wind turbines can only provide a small amount of data. Within this limited data, data imbalance is highly likely, such as a significant difference between the proportion of fault types and normal types, which seriously affects the accuracy of research results.

[0004] In addition, wind turbine fault information is often complex and has a large number of features. However, existing technologies often use subjective feature selection and traditional machine learning methods, resulting in inaccurate wind turbine fault diagnosis results. Summary of the Invention

[0005] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a wind turbine fault diagnosis method based on feature selection and stacking model fusion.

[0006] To achieve the above-mentioned objectives of this invention, this invention provides a wind turbine fault diagnosis method based on feature selection and stacking model fusion, comprising the following steps:

[0007] S1, Collect SCADA data, including wind speed, wind direction angle, gearbox temperature, blade angle, and power generation;

[0008] S2, perform SMOTE oversampling on the small category data to increase its proportion and balance the proportion of fault categories;

[0009] S3 uses RFECV for feature selection and divides the data into training and test sets;

[0010] S4, use the training set to train the stacking model to obtain the trained stacking model;

[0011] S5. Input the SCADA data to be tested into the trained stacking model to obtain the wind turbine fault diagnosis results.

[0012] Furthermore, S1 also includes:

[0013] If SCADA data is in an abnormal operating state, it will be marked as fault data and removed. At the same time, invalid data caused by communication signal or transmission equipment failure will be excluded to reduce the interference of such invalid data on the model prediction results.

[0014] Determining whether SCADA data is faulty involves the following steps:

[0015] Determine if the WEC status data value is greater than 0; if it is, the data is fault data.

[0016] Determine whether the RTU status data is greater than 100% or less than 50%. If the condition is met, the data is fault data.

[0017] The SCADA data is stored in the WEC status dataset and the RTU status dataset.

[0018] If the SCADA data is warning data, and the same warning message persists for a certain period of time without being cleared, then the data is fault data.

[0019] Data on normal and abnormal operating conditions are stored in WEC status data and RTU status data. WEC status data corresponds to status information directly related to the wind turbine itself, while RTU status data corresponds to power control data at the grid connection point. A WEC status data value greater than 0 indicates a fault. In the RTU, status data represents the active and reactive power setpoints; for example, status 100:82 corresponds to limiting active power output to 82% of the actual current output. An RTU status data value greater than 100% or less than 50% indicates a fault.

[0020] Warning data corresponds to general information about the wind turbine and is not directly related to its operation or safety. Sometimes, warning messages correspond to potential malfunctions on the wind turbine, because if the same warning message persists for a certain period of time without being cleared quickly, a malfunction may occur, generating new status information.

[0021] Based on this information, fault and warning data are filtered to determine periods of normal and faulty operation.

[0022] Furthermore, the feature selection using RFECV includes the following steps:

[0023] S3-1, Build a random forest classification model on the SCADA dataset, where each feature is assigned a weight;

[0024] S3-2, Remove the least important features from the feature set: Obtain the importance of each feature through the coef_ attribute and feature_importances_ attribute, and remove the poorly performing features;

[0025] S3-3, Repeat steps S3-1 to S3-2 on the remaining signal dataset features until all features have been traversed, obtaining the feature importance dataset.

[0026] S3-4: Based on the feature importance dataset obtained by recursive feature elimination, each time a dataset with a different number of features is extracted;

[0027] S3-5, perform 5-fold cross-validation on these samples with different numbers of features respectively;

[0028] S3-6, finally select the feature with the highest average score, that is, the feature with the best performance.

[0029] Furthermore, the base learner of the first layer of the stacking model includes at least two strong learners, and the second layer of the stacking model uses logistic regression to fuse the results obtained from the base learners to obtain the final result.

[0030] The logistic regression is used to improve the stability and generalization ability of the model, thereby enhancing stability.

[0031] Furthermore, the strong learner is extreme gradient boosting, or LIGHTGBM.

[0032] Furthermore, S4 also includes:

[0033] Use 6-fold cross-validation during model training to prevent overfitting.

[0034] Furthermore, it also includes:

[0035] S6. Input the test set into the trained stacking model to obtain the prediction results, verify the prediction results, and verify them according to the model performance metrics; the model performance metrics include accuracy, AUC, recall and balance score F1.

[0036] In summary, by adopting the above technical solution, this invention can obtain more accurate wind turbine fault diagnosis results compared with other traditional methods.

[0037] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0038] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0039] Figure 1 This is a flowchart of the RFECV of the present invention.

[0040] Figure 2 This is a schematic diagram of the stacking model superposition process of the present invention.

[0041] Figure 3 This is a flowchart of the wind turbine generator set fault detection process of the present invention.

[0042] Figure 4 This is a schematic diagram of RFECV, which shows how cross-validation automatically adjusts the number of selected features in an embodiment of the present invention.

[0043] Figure 5 This is a schematic diagram of the confusion matrix of normal and faulty samples in an embodiment of the present invention. Detailed Implementation

[0044] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0045] This paper uses SCADA data from a foreign wind farm to address the issue of poor results in fault diagnosis methods based on small-class data, which may be caused by subjective feature selection and the use of traditional machine learning methods. Therefore, a fault diagnosis method for wind turbines based on the fusion of feature selection and stacking models is proposed, including the following steps:

[0046] First, SMOTE oversampling is performed on the smaller category data to balance the proportion of fault categories.

[0047] Then, correlation analysis was performed on the data features, and the importance of the features was ranked by combining 5-fold cross-validation of RFECV to extract the effective features that affect the health status of wind turbine units.

[0048] Finally, a stacking model fusion method is used to combine two strong learners, and logistic regression is applied to improve the model's stability and generalization ability. Results show that this method outperforms traditional machine learning methods.

[0049] 1. SCADA Data Preprocessing

[0050] SCADA (Supervisory Control and Data Acquisition) is a system in wind farms responsible for real-time monitoring of wind turbine operation data and its stable transmission to terminals. Data is collected every 10 minutes, including wind speed, wind direction angle, gearbox temperature, blade angle, and power generation.

[0051] SCADA data is labeled based on warning and status data, marking them as normal and fault data respectively. Invalid data caused by communication signal or transmission equipment malfunctions is excluded to reduce the interference of such invalid data on model prediction results. When there are significant differences in the proportion between data categories, it can severely affect the model's predictions, rendering the classification results meaningless, because prediction models built on imbalanced data tend to favor labels from multiple categories.

[0052] This paper employs the SMOTE (Synthetic Minority Oversampling Technique) to synthesize minority class oversampling. Traditional random oversampling techniques directly copy samples with minority class labels, which easily leads to overfitting of the classification results. The SMOTE algorithm improves upon this by directly copying samples from each minority class sample x. i Randomly select a sample from the nearest neighbor minority class samples Then, connect the two samples and randomly select one sample from the line to classify it as a newly synthesized minority class sample.

[0053] The basic idea of ​​the SMOTE algorithm is to analyze the minority class samples and artificially synthesize new samples based on the minority class samples to add to the dataset. The algorithm flow is as follows:

[0054] (1) For each minority class sample x i Calculate the Euclidean distance of each of them to the remaining minority class samples;

[0055] (2) Determine the sampling multiplier N based on the set imbalanced sampling ratio, and sample from each minority class sample x i Randomly select samples from the nearest neighbor minority class samples

[0056] (3) For each randomly selected Construct new samples from the original samples using the formula below.

[0057]

[0058] Here, rand(0,1) represents generating a random number between 0 and 1;

[0059] x new This represents a final synthesized sample; the synthesized sample is x. i and A point that appears randomly on the line.

[0060] x i x represents i This represents the i-th sample, which is an input minority class sample;

[0061] This represents a neighboring sample of the selected x;

[0062] 2 Feature Selection

[0063] The data collected by SCADA in wind farms contains a wide variety of features. These include features related to normal and abnormal categories, as well as irrelevant features and redundant features related to the category labels. We can perform feature selection based on turbine characteristics and correlation analysis between features. Alternatively, we can use machine learning methods for feature selection, which is more efficient and better able to analyze the relationship between features and category labels to retain the best-performing features. Therefore, this paper uses RFECV, a method combining recursive feature elimination and cross-validation, to obtain the optimal number of features.

[0064] RFECV is divided into two stages, such as Figure 1 As shown, the first stage is RFE: recursive feature elimination, which sorts the features by importance; the second stage is CV: cross-validation, which selects the number of features with the best performance after sorting the importance of the features.

[0065] The specific steps of recursive feature elimination are:

[0066] Recursive feature elimination:

[0067] ① Use all feature variables in the SCADA dataset of wind farms as the initial features for recursive feature elimination;

[0068] ② A random forest classification model is built on the wind farm signal dataset, and each feature is assigned a weight.

[0069] ③ Remove the least important features from the feature set: Obtain the importance of each feature through the coef_ attribute and feature_importances_ attribute, and remove the poorly performing features.

[0070] ④ Then repeat the process on the remaining signal dataset features until all features have been traversed.

[0071] The specific steps of cross-validation:

[0072] ① Based on the feature importance dataset obtained from recursive feature elimination, each time a dataset with a different number of features is extracted.

[0073] ② Perform 5-fold cross-validation on these samples containing different numbers of features respectively.

[0074] ③ Finally, select the number of features with the highest average score, i.e., the features with optimal performance. In practice, the top 48 features out of 61 feature variables are ultimately determined to be the most important.

[0075] 3 Fault Diagnosis Model

[0076] 3.1 Stacking Model Fusion

[0077] Stacking models are hierarchical structures. The first layer is the base layer, containing the predictions of several base learners. The second layer is a model that integrates the results of the base layer, typically a weak learner. Stacking integrates the results of the base layer into sample features, while the original samples are treated as a new dataset, creating a new training set. This new training set is then used to train the new model and make predictions on the samples.

[0078] Since each learner in the base layer predicts both the training and test sets, directly combining these predictions into the training and test sets for the weak learner in the second layer can easily lead to overfitting. Therefore, K-fold cross-validation is typically used. Cross-validation is employed at the base layer level. The training and test sets generated by a base learner through cross-validation correspond to a series of features in the training and test sets of the second-layer model, respectively.

[0079] The stacking framework consists of two layers. The first layer comprises two base learners, XGBoost and LightGBM; the second layer uses logistic regression as the meta-learner. The steps for fault diagnosis using the Stacking model are as follows:

[0080] First, the SCADA dataset is divided into a training set and a test set. The training set is divided into 6 parts: train1, train2, train3, train4, train5, and train6.

[0081] Then, the divided datasets are used as inputs to the base learners, and the model is trained using a 6-fold cross-validation method.

[0082] Finally, the fault classification results of the wind turbine were obtained by the xgboost and lightgbm classifiers respectively, and the fault types included faulty and faultless.

[0083] The fault classification results on the training set obtained after training the base learner are combined with the training set labels and input into the meta-learner logistic regression for training. The meta-learner then outputs the wind turbine fault detection and prediction results.

[0084] 3.1.1 Extreme Gradient Boosting XGBoost

[0085] XGBoost is a tool for massively parallel boosting trees. It is currently the fastest and best open-source boosting tree toolkit, more than 10 times faster than common toolkits.

[0086] XGBoost is an additive expression consisting of k base models:

[0087]

[0088] Where f k It is the k-th base model. It is the predicted value of the i-th sample.

[0089] Taking the model at step t as an example, the model applies to the i-th sample x. i The prediction is:

[0090]

[0091] in The predicted value given by the model at step t-1 is a known constant, f. t (x i Let be the predicted value of the new model we need to add this time, and let represent the predicted value of the i-th sample in step t. At this point, the objective function can be written as:

[0092]

[0093] Obj (t) Denotes the objective function at step t;

[0094] n represents the total number of samples;

[0095] Indicates the relationship between y i , Loss function;

[0096] f i Represents the i-th base model;

[0097] t represents the t-th step;

[0098] y i Let x represent the i-th sample. i The true value;

[0099] Ω(f i ) represents the regularization term;

[0100] This represents the predicted value of the i-th sample in the model at step t;

[0101] According to Taylor's formula, perform a second-order Taylor expansion of the function f(x+Δx) at point x:

[0102]

[0103] Where f'(x) and f''(x) are the first and second derivatives of f(x) respectively, and Δx represents the model at the t-th step of training.

[0104] Bundle Consider x, f t (x i Since Δx is considered as Δx, the objective function can be written as:

[0105]

[0106] The loss function represents the loss from the first t-1 steps.

[0107] This represents the predicted value of the i-th sample in the model at step t-1;

[0108] g i h is the first derivative of the loss function. i Let l be the second derivative of the loss function. The loss function is l(·,·) in formula (4).

[0109] 3.1.2 LightGBM

[0110] LightGBM is a new member of the boosting ensemble model family. Like XGBoost, it is an efficient implementation of GBDT, and in many aspects, it outperforms XGBoost. In principle, it is similar to GBDT and XGBoost, using the negative gradient of the loss function as an approximation of the residual of the current decision tree to fit a new decision tree.

[0111] The basic idea of ​​the histogram algorithm is to first discretize the continuous floating-point feature values ​​into k integers, and simultaneously construct a histogram with a width of k. When traversing the data, the discretized values ​​are used as indices to accumulate statistics in the histogram. After traversing the data once, the histogram has accumulated the required statistics. Then, based on the discrete values ​​of the histogram, the optimal split point is found through traversal.

[0112] When the base learners are XGBoost and LightGBM, the stacking model stacking process is as follows: Figure 2 As shown:

[0113] First, the training set is divided into 6 parts: train1, train2, train3, train4, train5, and train6.

[0114] Then, XGBoost and LightGBM were chosen as the base models. For the XGBoost model: train1, train2, train3, train4, train5, and train6 were used as the validation set, and the remaining 5 sets were used as the training set. Six-fold cross-validation was performed to train the model; then predictions were made on the test set. This resulted in 6 predictions trained on the XGBoost model on the training set and 1 prediction B1 on the test set. These 6 predictions were then vertically overlapped and merged to obtain A1. The LightGBM model was trained similarly.

[0115] Finally, after the two base models are trained, the predicted values ​​of the two models on the training set are used as two "features" A1 and A2 respectively, and the LR model is used to train and build the LR model.

[0116] Using a trained logistic regression model, predictions are made based on the values ​​of two "features" (B1, B2) constructed from the predictions on the test set before the two base models, to arrive at the final predicted category.

[0117] 3.2 Modeling Process

[0118] This paper implements wind turbine fault diagnosis based on the fusion of SMOTE, RFECV, and stacking models. The specific steps are as follows: Figure 3 As shown, it includes the following steps:

[0119] ① In the data preprocessing stage, SCADA data is screened to remove invalid data that interferes with the prediction results. SMOTE oversampling is performed on small category data to prevent the classification results from being meaningless due to data imbalance.

[0120] ② After preprocessing, the RFECV algorithm based on recursive feature elimination and cross-validation using a random forest classifier is used. This paper uses 6-fold cross-validation.

[0121] ③ After feature selection, the data is divided into training and test sets in a ratio of 0.67:0.33;

[0122] ④ The training set is modeled using the stacking model fusion method. The base learners are extreme gradient boosting and LIGHTGBM, and the second layer uses logistic regression to fuse the results obtained from the base learners together to improve stability. Six-fold cross-validation is used during model training to prevent overfitting.

[0123] ⑤ Validate the prediction results using the test set, examine the classification results through the confusion matrix, and evaluate them based on model performance metrics such as accuracy, AUC, recall, and F1 balance score. Compare the experimental results with those of Gradient Boosting Tree, XGBoost, and LightGBM.

[0124] 4. Case Analysis

[0125] 4.1 Feature Selection

[0126] This article uses status data, warning data, and SCADA data from a foreign wind farm from May 1, 2014 to April 9, 2015. The SCADA data contains 49,026 records and 61 features. After feature preprocessing, 6,869 normal samples and 298 fault samples are retained. After SMOTE, the number of normal samples remains unchanged, while the number of fault samples becomes 6,869.

[0127] Table 1. Counters that change sample shape via Smote

[0128]

[0129] Through RFECV feature selection, it can be seen that the performance is best when the number of selected features is 48. Furthermore, by combining the correlation analysis between SCADA data features and removing features with high correlation, the classification accuracy can be improved.

[0130] After performing correlation analysis using heatmaps, this paper will select 46 features for subsequent model training, such as... Figure 4 As shown.

[0131] A heatmap, also known as a correlation coefficient plot, determines the strength of the correlation between variables by analyzing the correlation coefficients corresponding to different colored squares in the heatmap. The formula for calculating the correlation coefficient between two variables is:

[0132]

[0133] Where ρ represents the correlation coefficient, Cov represents the covariance, and E represents the expected value / mean.

[0134] The higher the correlation coefficient, the stronger the linear correlation between the variables. Therefore, we choose to retain only one of the variables with a high correlation coefficient between them.

[0135] The 46 characteristics are as follows: average wind speed, maximum wind speed, minimum wind speed, average rotational speed, maximum rotational speed, minimum rotational speed, average power, maximum power, minimum power, nacelle location, operating time, kilowatt-hour output, minute output, average reactive power, maximum reactive power, minimum reactive power, average active power with wind, average active power due to technical reasons, average active power due to force majeure, average active power due to external reasons, average blade A angle, system 1 converter 3 cabinet temperature, system 1 converter 6 cabinet temperature, system 2 converter 1 Cabinet temperature, System 2 converter 2 cabinet temperature, System 2 converter 3 cabinet temperature, System 2 converter 4 cabinet temperature, front bearing temperature, rear bearing temperature, blade A pitch control cabinet temperature, blade B pitch control cabinet temperature, blade C pitch control cabinet temperature, rotor temperature 1, rotor temperature 2, stator temperature 1, stator temperature 2, nacelle ambient temperature 1, nacelle ambient temperature 2, nacelle temperature, nacelle cabinet temperature, spindle temperature, rectifier cabinet temperature, yaw inverter cabinet temperature, converter cabinet fan temperature, ambient temperature, tower temperature, control cabinet temperature, transformer temperature.

[0136] 4.2 Model Training Results

[0137] Stacking models are used to train the feature-selected data and classify normal and faulty data. The confusion matrix is ​​as follows: Figure 5 As shown in the confusion matrix, 0 represents normal data and 1 represents faulty data. Recall indicates the correctness of the category. The higher the recall, the better the recognition performance of that category. f1_score, also known as balanced f1_score, is defined as the harmonic mean of precision and recall. The f1_score metric combines the outputs of precision and recall. f1_score is a metric for classification problems. The maximum value is 1, and the minimum value is 0.

[0138] Support represents the actual number of samples belonging to a class. The micro-mean does not distinguish between classes and calculates overall accuracy, recall, and F1 score. Weighted average, as the name suggests, adds weights to the macro-mean. Its weight refers to the ratio of the number of samples belonging to a certain class to the total number of samples. This ratio can be seen in the accuracy, recall, or F1 score, indicating that the classification results have reached a good level.

[0139] Where 0 represents normal data and 1 represents faulty data, it can be seen that the classification results are quite good. Looking at precision, recall, or the f1 score (the average of precision and recall), the classification results have all reached a good level.

[0140] The confusion matrix shows that 0 represents normal data and 1 represents faulty data. Recall indicates the correctness of the classification. The higher the recall, the better the classification performance. f1_score, also known as balanced f1_score, is defined as the harmonic mean of accuracy and recall. The f1_score combines the outputs of precision and recall. f1_score is a metric for classification problems. The maximum value is 1, and the minimum is 0. Support represents the actual number of samples belonging to a class. The micro-mean does not distinguish between classes and calculates the overall accuracy, recall, and f1_score. Weighted avg, as the name suggests, adds weights to the macro-mean. Its weight refers to the ratio of the number of samples belonging to a certain class to the total number of samples. It can be seen from accuracy, recall, or f1_score that the classification results have reached a good level.

[0141] Table 2 Classification Report of Confusion Matrix

[0142]

[0143] Finally, the accuracy of the stacking model fusion experiment results is compared with XGBOOST, LightGBM, and GBDT, as follows:

[0144] Table 3 Comparison of evaluation indicators for each model

[0145]

[0146] As can be seen, the classification score of the stacked model fusion is superior to the other three classifiers in terms of AUC, precision, recall, and F1 score. Although LightGBM scores higher in recall score than the stacked model fusion, after comprehensive consideration, the AUC value of stacked is significantly higher than that of XGBoost, LightGBM, and GBDT, which best represents classifier performance. Therefore, in this paper, the classifier with stacked model fusion performs better.

[0147] 5. Conclusion

[0148] This paper proposes a stacking-based wind turbine fault diagnosis algorithm based on feature selection, using actual SCADA detection data from wind farms for fault diagnosis. To address the problem of imbalanced sample labels, the SMOTE oversampling algorithm is employed to balance the data. Recursive feature elimination and the 5-fold cross-validation (RFECV) algorithm are used to select high-performance features based on feature importance. Finally, a stacking model fusion method is used. First, two base learners are trained separately on the training set, and 6-fold cross-validation is used. Then, a logistic regression (LR) algorithm is used in the second layer to fuse these training and test sets. Experiments show that the accuracy after stacking reaches 0.99044, which is higher than other single learners. This stacking machine learning-based wind turbine fault diagnosis, through data mining and big data analysis, provides a new approach to wind turbine blade fault diagnosis.

[0149] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A wind turbine generator fault diagnosis method based on feature selection and stacking model fusion, characterized in that, Comprising the following steps: S1, collecting SCADA data, the SCADA data comprising wind speed, wind direction angle, gearbox temperature, blade angle and power generation; S2, SMOTE oversampling small category data to increase its proportion and balance the proportion of fault categories; S3, using RFECV for feature selection and dividing the data into a training set and a test set; the use of RFECV for feature selection comprises the following steps: S3-1, establishing a random forest classification model on the SCADA data set, each feature being assigned a weight; S3-2, deleting the least important features in the feature set: obtaining the importance of each feature through the coef_ attribute and the feature_importances_ attribute, and removing poorly performing features; S3-3, repeating steps S3-1-S3-2 on the remaining signal data set features until all features have been traversed to obtain a feature importance data set S3-4, according to the feature importance data set obtained by recursive feature elimination, taking out data sets containing different numbers of features each time; S3-5, performing 5-fold cross-validation on these samples containing different numbers of features respectively; S3-6, finally selecting the number of features with the highest average score, i.e. the optimal performance feature; S4, training the stacking model using the training set to obtain a trained stacking model; the base learners of the first layer of the stacking model include at least two strong learners, and the second layer of the stacking model uses logistic regression to fuse the results obtained by the base learners to obtain the final result; the strong learners are extreme gradient boosting and LIGHTGBM; Further comprising inputting the test set into the trained stacking model to obtain a prediction result, verifying the prediction result, and verifying according to a model performance measurement index; the model performance measurement index includes accuracy, AUC, recall rate and balanced score F1; S5, inputting the SCADA data to be tested into the trained stacking model to obtain a wind turbine fault diagnosis result.

2. The wind turbine fault diagnosis method based on feature selection and stacking model fusion according to claim 1, characterized in that, The S1 further comprises: If the SCADA data is in an abnormal operating state, mark the data as fault data and exclude it, and exclude invalid data caused by communication signals and transmission equipment failure to reduce the interference of such invalid data on the prediction result of the model.

3. The wind turbine fault diagnosis method based on feature selection and stacking model fusion according to claim 1, characterized in that, S4 further comprises: Using 6-fold cross-validation during model training to prevent overfitting.

Images