Wind turbine generator main bearing fault diagnosis method based on machine learning algorithm

Abstract

The invention discloses a wind turbine generator main bearing fault diagnosis method based on a machine learning algorithm, and the method comprises the following steps: 1, obtaining the operation data of a wind turbine generator, carrying out the preprocessing of the operation data, obtaining a data set, and dividing the data set into a training set and a test set; 2, constructing an XGBoost model, and inputting the training set into the XGBoost model for training to obtain an optimal XGBoost model; the optimal hyper-parameter combination is input into an XGBoost model for training, and the XGBoost model is obtained; step 3, constructing an evaluation index to test and optimize the XGBoost model; step 4, carrying out adaptive threshold monitoring on the difference value between the predicted value and the true value of the main bearing temperature; and 5, migrating the XGBoost model to other wind turbine generators by using migration learning. The fault diagnosis accuracy and the fault early warning capability of the main bearing of the wind turbine generator can be improved, the universality of the fault diagnosis model is improved, and the overall fault diagnosis cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of offshore wind turbine fault diagnosis, and specifically to a wind turbine main bearing fault diagnosis method based on a machine learning algorithm. Background Art

[0002] In recent years, with the deterioration of the global environment and the depletion of fossil fuels, wind energy, as a simple, clean and reliable energy source obtained from nature, has become a rapidly developing renewable energy source and has received more and more attention. Offshore wind farms are located in offshore areas. The locations of wind turbines are relatively scattered and far away from the control center. Sometimes there will be extremely bad weather. Once a fault occurs, it is difficult to repair the wind turbines in time. Wind farms have a low tolerance for downtime due to faults of wind turbines. Bearings are also important components widely used in wind turbines. The main shaft of the wind turbine is directly connected to the blades and belongs to the low speed area. The operating condition of the main shaft bearing is low speed and heavy load. The main shaft of the wind turbine is the core rotating component of the wind turbine. Once the main shaft bearing fails, it may cause uneven force on the main shaft and deflection of the shaft center, which will cause the wind turbine to shut down for a long time and generate high maintenance costs. With the widespread application of wind turbines, the requirements for their reliability and maintainability are getting higher and higher. Traditional fault diagnosis methods have problems such as low accuracy and untimely diagnosis. Summary of the invention

[0003] The purpose of the present invention is to provide a wind turbine main bearing fault diagnosis method based on a machine learning algorithm. The present invention can improve the wind turbine main bearing fault diagnosis accuracy and fault warning capability, improve the versatility of main bearing fault warning, and reduce the overall fault diagnosis cost.

[0004] The technical solution of the present invention is a method for diagnosing main bearing faults of a wind turbine generator set based on a machine learning algorithm, comprising the following steps:

[0005] Step 1: Obtain the operating data of the wind turbine, perform preprocessing, normalization and feature extraction on the operating data to obtain a data set;

[0006] Step 2: Build an XGBoost model, input the data set into the XGBoost model for training, and use a genetic algorithm to optimize the hyperparameters of the XGBoost model until the model prediction effect reaches the optimal value, thus obtaining the optimal XGBoost model;

[0007] Step 3: Predict the main bearing temperature of the wind turbine using the optimal XGBoost model, perform adaptive threshold monitoring on the difference between the predicted value and the true value of the main bearing temperature, and complete the main bearing fault diagnosis of the wind turbine;

[0008] Step 4: Use transfer learning to migrate the optimal XGBoost model to other wind turbines, predict the main bearing temperature of other wind turbines based on their operating data, and then realize the main bearing fault diagnosis of other wind turbines.

[0009] In the aforementioned wind turbine main bearing fault diagnosis method based on machine learning algorithm, in step 1, the operating data of the wind turbine is collected by the SCADA system.

[0010] In the aforementioned wind turbine main bearing fault diagnosis method based on machine learning algorithm, the data preprocessing includes eliminating missing values, abnormal values ​​and data when not working properly in the operation data of the wind turbine, and then processing the abnormal values ​​of the parameters by the 3σ rule.

[0011] In the above-mentioned wind turbine main bearing fault diagnosis method based on machine learning algorithm, the normalization process is maximum and minimum value normalization, and the formula is as follows:

[0012] The process of data normalization is as follows:

[0013]

[0014] In the formula, x max and x min They represent the maximum and minimum values ​​of a characteristic parameter in the running data respectively; x represents the current processing value of the characteristic parameter; x * Indicates the normalized value.

[0015] In the aforementioned wind turbine main bearing fault diagnosis method based on machine learning algorithm, the feature extraction process includes first using the Pearson correlation coefficient method to calculate the correlation coefficient between feature parameters, and then using the feature importance index in the XGBoost model to screen features, and the calculation method uses the "gain" weight form.

[0016] In the aforementioned wind turbine main bearing fault diagnosis method based on machine learning algorithm, in step 2, the construction process of the XGBoost model is as follows:

[0017] Given a training set of n samples and m features:

[0018] D=(x i ,y i )(|D|=n,x i ∈R m ,y i ∈R);

[0019] The prediction function is defined as follows:

[0020]

[0021] in, is the i-th sample x i The predicted value; K represents all decision trees in the XGBoos model; f k is the kth regression decision tree; f k (x i ) represents the score of the i-th sample in the k-th decision tree;

[0022] The XGBoost model uses a forward addition model, using multiple decision trees to make decisions together, and the result of each decision tree is the difference between the target value and the prediction results of all previous decision trees, and all the results are accumulated to get the final result; the loss function of the XGBoost model is as follows:

[0023]

[0024] Among them, L(θ) represents the loss function of the XGBoos model, n is the number of samples, T is the number of leaf nodes, ω j is the score of each leaf node, γ and λ are regularization parameters; the loss function of the XGBoost model is composed of empirical loss and the regularization term The loss function is minimized within the given training parameter range to obtain the optimal model.

[0025] In the above-mentioned wind turbine main bearing fault diagnosis method based on machine learning algorithm, in step 3, the prediction of the wind turbine main bearing temperature and the adaptive threshold monitoring of the difference between the predicted value and the true value of the main bearing temperature are specifically:

[0026] After data preprocessing, normalization and feature extraction, the XGBoost model hyperparameters are adjusted, the maximum depth, learning rate and weak estimator are selected to adjust the XGBoost model, and the grid search is used to optimize the model hyperparameter adjustment to obtain the XGBoost model with the best prediction effect to predict the main bearing temperature; then the root mean square error and mean absolute percentage error are used to evaluate the model prediction effect.

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

[0028] The present invention uses the XGBoost model to construct an XGBoost model to accurately predict the main bearing temperature of the wind turbine. By performing adaptive threshold calculation on its error, it is possible to accurately detect hidden faults at an early stage. The present invention migrates the XGBoost model of the offshore wind turbine to the same wind farm and other onshore wind farms, and performs main bearing fault warning for other types of wind turbines. The warning effect is good, which can reduce the cost of the fault warning model of the wind turbine, and is of great significance for improving the power generation efficiency of the wind turbine and ensuring the stable operation of the wind turbine. In addition, the present invention uses data obtained by the SCADA system. The SCADA system is deployed in most wind farms, which can improve the generalization of the XGBoost model in different wind farms, avoid the use of vibration signals, and reduce costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the process of the present invention;

[0030] Figure 2 is the Pearson correlation coefficient heat map of the SCADA state variables in the present invention;

[0031] Figure 3 is a curve graph after data normalization in the present invention;

[0032] Figure 4 It is a comparison of the prediction results of different models for the main bearing temperature in the present invention;

[0033] Figure 5 It is a main bearing temperature detection diagram based on an adaptive threshold value in the present invention;

[0034] Figure 6 It is the main bearing temperature detection diagram based on the 3σ criterion in the present invention;

[0035] Figure 7 This is the verification effect of the present invention migrating the XGBoost model to the wind turbines in the same wind farm;

[0036] Figure 8 This is the verification effect of the present invention migrating the XGBoost model to the wind turbines in the onshore wind farm. DETAILED DESCRIPTION

[0037] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0038] Embodiment 1: A method for diagnosing main bearing faults of a wind turbine generator system based on a machine learning algorithm, such as Figure 1 As shown, the following steps are included:

[0039] Step 1: Obtain the operating data of the wind turbine, perform preprocessing, normalization and feature extraction on the operating data to obtain a data set;

[0040] The operating data of 20 state variables including gearbox main bearing temperature, gearbox temperature, and cabin temperature are extracted from the SCADA data acquisition system. Taking Unit 3 as an example, the operating data of these 20 state variables are analyzed. The sampling period of the operating data is from January 1, 2022 to October 1, 2023, and the time interval of the sample points is 10 minutes. A total of 80,317 pieces of original data are obtained. The SCADA data state variables are shown in Table 1.

[0041] Table 1 SCADA data status variables

[0042] serial number feature serial number feature V1 Average gear main bearing temperature V11 Average cabin air cooler outlet temperature V2 Average wind speed V12 Average gearbox water pump inlet temperature V3 Average generator nacelle side bearing temperature V13 Average hydraulic pump operation times V4 Average gearbox generator side bearing temperature V14 Water content in gearbox lubricating oil V5 Average hub temperature V15 Average gearbox secondary oil temperature V6 Average cabin temperature V16 Average tower base second level platform temperature V7 Average generator gearbox side bearing temperature V17 Corrected wind speed V8 Average gearbox oil temperature V18 Average transmission coolant temperature V9 Average outdoor temperature V19 Average generator winding temperature V10 Average tower base first floor platform temperature V20 Average inverter 2 water cooling inlet temperature

[0043] Data preprocessing includes removing missing values, outliers, and data from abnormal operation of wind turbines. First, the original data is checked, and most of the data with zero values, missing characteristic parameters, and "set values" are deleted, leaving only the measured values. Then, according to the 3σ rule, outliers are treated as gross errors to remove data outliers.

[0044] The 3σ rule is a probability statistics theory based on the normal distribution. In the normal distribution, about 68.27% of the data should be within the range of the mean plus or minus one standard deviation, more than 95.45% of the data should be within the range of the mean plus or minus two standard deviations, and about 99.73% of the data should be within the range of the mean plus or minus three standard deviations. According to this theory, a value should be within the range of the mean plus or minus three standard deviations. When a data point exceeds the range of the mean plus or minus three standard deviations, it means that the data has a large deviation and the data has problems and should be deleted.

[0045] During the model training process, some feature parameters cannot significantly improve the prediction effect of the model, and may even have a negative impact. Therefore, it is necessary to perform feature selection on the data, which is a crucial step in machine learning. Feature selection is to select the feature parameters that contribute most to the model prediction effect, optimize the model structure, reduce the required data dimension, and speed up the learning and training of the model. Feature selection is required for these pre-processed data.

[0046] Then, the data is normalized as follows:

[0047]

[0048] In the formula, x max and x minThey represent the maximum and minimum values ​​of a characteristic parameter in the running data respectively; x represents the current processing value of the characteristic parameter; x * Represents the normalized value; the normalized characteristic curve obtained by screening is as follows Figure 3 shown.

[0049] In order to reduce the complexity of the XGBoost model and improve the warning accuracy of the XGBoost model, the Pearson correlation coefficient is used to analyze the correlation with the target parameter gearbox main bearing temperature, and the characteristic parameters adapted to it are screened out. The calculation method adopts the "gain" weight form. Figure 2 This is the Pearson correlation coefficient heat map of the SCADA state variables of Unit 3. In order to highlight the correlation between each data and the target parameter, the values ​​with a correlation coefficient greater than 0.5 are marked and used as the input of the XGBoost model; the final selected parameters are V1, V3, V4, V5, V6, V7, V8, V10, V11, V12, V15, V16, V18, and V19. Figure 2 The analysis shows that the cabin temperature, gearbox oil temperature, and cabin air cooling inner circulation outlet temperature have the highest correlation with the main bearing temperature. This shows that the cabin temperature environment can directly reflect the temperature level of the main bearing. At the same time, the gearbox oil temperature and the cabin air cooling inner circulation outlet temperature represent the close relationship between the lubrication system, air cooling system, and gearbox working conditions.

[0050] After data processing, 80,384 data were obtained. 70% of these data were divided into training sets and input into the model for training, and the remaining 30% were used as test sets to test the accuracy of the main bearing temperature warning model.

[0051] Step 2: Build an XGBoost model, input the data set into the XGBoost model for training, and use a genetic algorithm to optimize the hyperparameters of the XGBoost model until the model prediction effect reaches the optimal value, thus obtaining the optimal XGBoost model;

[0052] Given a training set of n samples and m features:

[0053] D=(x i ,y i )(|D|=n,x i ∈R m ,y i ∈R);

[0054] The prediction function is defined as follows:

[0055]

[0056] in, is the i-th sample x i The predicted value; K represents all decision trees in the XGBoos model; f k is the kth regression decision tree; f k (x i ) represents the score of the i-th sample in the k-th decision tree;

[0057] The XGBoost model uses a forward addition model, using multiple decision trees to make decisions together. The result of each decision tree is the difference between the target value and the prediction results of all previous decision trees. All the results are accumulated to get the final result, so as to improve the effect of the entire XGBoost model.

[0058] The loss function of the XGBoos model is as follows:

[0059]

[0060] Among them, L(θ) represents the loss function of the XGBoos model, n is the number of samples, T is the number of leaf nodes, ω j is the score of each leaf node, γ and λ are regularization parameters; the loss function of the XGBoost model is composed of empirical loss and the regularization term The loss function is minimized within the given training parameter range to obtain the optimal model.

[0061] Hyperparameters include maximum depth, learning rate, and weak estimator. After determining the parameter adjustment range, a genetic algorithm is used to find the optimal value. The overall process is as follows:

[0062] (1) Set the hyperparameter range of the XGBoost model.

[0063] (2) The genetic algorithm (GA) is embedded in the XGBoost model to optimize hyperparameters. Specifically, the objective function of the XGBoost model is defined, the input of which is the hyperparameter combination, and the output is the evaluation index of the model on the test set.

[0064] (3) Iterate through the genetic algorithm, adjust the hyperparameter combination, and update the individuals in the population to find the global optimal hyperparameter combination, as shown in Table 2.

[0065] (4) Input the global optimal parameter combination into the XGBoost model for training and calculate the XGBoost model.

[0066] Table 2 Parameters of Genetic Algorithm Optimization of XGBoost

[0067] Parameter name Optimal parameters Maximum Depth 7 Learning Rate 0.1 Weak Learner 298 Number of leaf nodes 1.2

[0068] Step 3: Predict the main bearing temperature of the wind turbine using the optimal XGBoost model, perform adaptive threshold monitoring on the difference between the predicted value and the true value of the main bearing temperature, and complete the main bearing fault diagnosis of the wind turbine;

[0069] After data preprocessing, maximum and minimum normalization and feature extraction, the data in the dataset are adjusted to adjust the XGBoost model hyperparameters, select the maximum depth (max_depth), learning rate (learning_rate) and weak estimators (n_estimators) to adjust XGBoost, and use grid search (GridSearchCV) to optimize the XGBoost model hyperparameter adjustment to obtain the XGBoost model with the best prediction effect, and predict the average cabin temperature, average outdoor temperature, average cabin humidity and average atmospheric humidity. Then the root mean square error (Root Mean SquaredError, RMSE) and mean absolute percentage error (Mean Absolute Percentage Error, MAPE) are used to evaluate the prediction effect of the XGBoost model.

[0070] Due to the non-stationary operation of wind turbines and the impact of environmental noise on the data acquisition system, the residual of the main bearing temperature prediction and the actual value presents dynamic changes, and the calculated detection index always has an extreme distribution, which may lead to false alarms. Therefore, the present invention sets a dynamic threshold according to a specific time point to reduce the effectiveness of the false alarm rate. First, in order to strengthen the correlation between the previous and subsequent temperature values, a weighted residual value is introduced. This value is an adaptive residual value obtained by weighted fusion of the original residual value at the current moment and the residual value at the previous moment. The purpose of this step is to make full use of past information to reduce errors caused by data fluctuations.

[0071] In order to verify the effectiveness of the proposed offshore wind turbine main bearing temperature early warning method, 2000 data of unit 3 in a certain period of time were selected for verification. The prediction results of the XGBoost model were compared with the true value to obtain the corresponding residual, and then an adaptive threshold was introduced to dynamically monitor the fluctuation of the main bearing temperature.

[0072] The calculation formula of the adaptive threshold monitoring is as follows:

[0073]

[0074] in, represents the adaptive residual value at time t, α represents the weighting factor with a set value of 0.1, and R t represents the residual at time t;

[0075] In order to enhance the anti-noise capability, the mean value of the residual is set as the noise factor used to construct the adaptive threshold. The specific calculation is shown in the following formula:

[0076]

[0077] Among them, n represents the number of samples, and N represents the residual mean of the noise factor;

[0078] Finally, based on the adaptive residual value The residual mean value N of the noise factor and the empirical adjustment parameter λ are used to construct the adaptive threshold U(t).

[0079]

[0080] Among them, U(t) represents the adaptive threshold at time t, and λ is the knowledge-based adjustment factor, which is set to 0.025.

[0081] When the threshold at that time point exceeds the adaptive threshold, the system will issue an alarm, such as Figure 5 As shown. Figure 5 It can be seen that within the 4 light blue circles before the 1250th data point, although the residual value of the temperature is very close to the set threshold, it is still within the safe range. In the circle after 1500, the residual value at this time has exceeded the adaptive threshold, and the system issues an alarm. The specific sample point is 1557. However, in order to eliminate the influence of accidental factors, most SCADA systems usually issue an alarm signal only after the error continues for a period of time. Therefore, there is no main bearing overtemperature alarm in the selected time period. By checking the input data, it is found that the maximum temperature value in this time period is 45.29℃, which is still within the normal bearing temperature range. The main bearing temperature monitoring chart with the 3σ criterion as the standard for setting the threshold value, because the residual value is always in a fluctuating state, the constant threshold cannot achieve effective monitoring of the main bearing temperature. In the actual operation of the wind turbine, various factors may cause temperature fluctuations, but it does not necessarily mean that the main bearing is faulty. Although the SCADA system itself gives a certain tolerance to the time point exceeding the threshold, a single constant threshold may lead to false alarms or omissions. Figure 6 There are two false alarms, which will waste extra energy of maintenance personnel. In order to improve the accuracy and reliability of the monitoring system, it is necessary to consider the use of dynamically adjusted thresholds to adapt to changes under different working conditions. By introducing adaptive thresholds, it is possible to effectively capture the dynamic changes of the main bearing temperature in the actual operating environment, thereby greatly reducing the false alarm rate of the system.

[0082] The evaluation index uses the root mean square error RMSE as the main evaluation index, and the calculation formula is as follows:

[0083]

[0084] In the formula, is the i-th sample x i The predicted value of y is n, where n is the number of samples; i is the true value of the main bearing temperature; then, the mean absolute error MAE is introduced as an auxiliary evaluation index, and the calculation formula is as follows:

[0085]

[0086] By analyzing the determination coefficient R 2 , we can get the fitting degree of XGBoost model to the true value. The calculation formula is as follows:

[0087]

[0088] In the formula, is the average of the true values.

[0089] The accuracy of the XGBoost model is tested and optimized using evaluation indicators to obtain the XGBoost model.

[0090] In order to verify the performance of the XGBoost model after parameter optimization, LSTM, GRU, LightGBM, and Random forest algorithms were selected as analog models for comparative experiments on the same data set. The output results of the five models were denormalized and then the indicators were calculated. The evaluation results are shown in the table.

[0091] Table 3 Evaluation indicators of experimental results of different models

[0092] Model RMSE MAE <![CDATA[R 2 ]]> XGBoost 0.432 0.271 0.995 LSTM 0.830 0.563 0.981 GRU 1.079 0.806 0.968 LightGBM 0.853 0.575 0.980 Randomforest 0.527 0.348 0.992

[0093] As shown in Table 3, compared with other models, the XGBoost model with optimized parameters achieved the best performance in the main bearing temperature prediction task. First, the XGBoost model achieved the lowest values ​​in RMSE and MAE indicators, which is better than other analog models. 2 It reached 0.995, further confirming the excellent fitting ability of the XGBoost model to the input data. Figure 4 The fitting curves of the prediction results of each model and the true value are shown.

[0094] Step 4: Use transfer learning to migrate the optimal XGBoost model to other wind turbines, predict the main bearing temperature of other wind turbines based on their operating data, and then realize the main bearing fault diagnosis of other wind turbines.

[0095] In view of the fact that it is difficult to train a good XGBoost model for a newly built wind farm due to insufficient data, we consider using the existing wind farm data and models for transfer learning. This method uses the XGBoost model of the existing wind farm as a pre-trained model, and then fine-tunes it on the limited data of the newly built wind farm to better adapt to the new environment. Because the pre-trained model has learned some common features, the model can still achieve satisfactory prediction results even when the amount of data for the newly built wind farm is small.

[0096] The core idea of ​​transfer learning is to use the knowledge learned in the source domain (existing wind farms) to improve the learning performance in the target domain (new wind farms). Through transfer learning, the features and patterns learned from existing data can be effectively used to accelerate the training process of the new wind farm model and improve the performance of the XGBoost model.

[0097] In step 5, the transfer learning is:

[0098] The source domain consists of the feature space χ and the marginal probability density distribution P(X), where X = {x 1 、x 2 …x n}∈χ. If the two domains are different, their feature spaces and marginal probability distributions are also different. In a given domain, a task consists of a label space Y and a target prediction function f(.); the source domain and task are represented as:

[0099] D = {χ, P(X)};

[0100] H = {Y, f(.)};

[0101] The target prediction function f(.) is learned through training samples; given the source domain D S and source domain learning task H S , target domain D T and the target task H T In the case, and D S Not equal to D T and H S Not equal to H T In the case of transfer learning, the source domain D S and H S to improve and optimize the target domain D T The target prediction function f T (.) learning effect.

[0102] Embodiment 2: Step 1: Obtaining the operating data of the wind turbine generator set, performing preprocessing, normalization processing and feature extraction processing on the operating data to obtain a data set;

[0103] Step 2: Build an XGBoost model, input the data set into the XGBoost model for training, and use a genetic algorithm to optimize the hyperparameters of the XGBoost model until the model prediction effect reaches the optimal value, thus obtaining the optimal XGBoost model;

[0104] Step 3: Predict the main bearing temperature of the wind turbine using the optimal XGBoost model, perform adaptive threshold monitoring on the difference between the predicted value and the true value of the main bearing temperature, and complete the main bearing fault diagnosis of the wind turbine;

[0105] Step 4: Use transfer learning to migrate the optimal XGBoost model to other wind turbines, predict the main bearing temperature of other wind turbines based on their operating data, and then realize the main bearing fault diagnosis of other wind turbines.

[0106] The 15#2MW wind turbine generator set of the Daheishan Wind Farm in Yunnan was found to have a main shaft bearing rolling element crushing fault at 7:10 on July 27, 2021. The No. 15 wind turbine was in a shutdown maintenance state from the time of fault discovery to 17:30 on August 7, 2021. The present invention uses the operating data taken from the SCADA data acquisition system of the Daheishan Wind Farm in Yunnan, and extracts 32 characteristic values ​​of the 15#2MW wind turbine generator set including wind speed, power, gearbox oil temperature, and main bearing temperature. The sampling time distribution of the operating data is from January 1, 2021 to December 31, 2021, with a sample interval of 5 minutes, and a total of 91,038 original data are obtained.

[0107] First, simply process the operating data to remove missing values, outliers, and data from abnormal operations. The actual time point of the main bearing failure must be before the time point when the failure is manually discovered. Therefore, the wind turbine maintenance time period is used as the node to divide the data into two operating conditions before the wind turbine failure is discovered and after the maintenance. The operating data of the wind turbine under normal working conditions is used to fine-tune the migrated XGBoost model. At the same time, since the wind turbine needs a period of time to restart and adapt after maintenance, the data after the restart needs to be removed. According to the wind turbine maintenance records of the wind farm, this example deletes the data 48 hours after the wind turbine restarts.

[0108] The data exported from the SCADA system of Yunnan Daheishan Wind Farm has a total of 32 characteristic parameters. The same data preprocessing method as offshore wind turbines is first used to remove outliers and exclude characteristic variables with a large number of missing values ​​and identical values ​​to improve the quality of the data set.

[0109] Since the data comes from the SCADA systems of different engineering projects, the models of wind turbines are also different, the names of many characteristic variables are different, and there is a lack of documentation to explain them. It is not known whether many characteristic variables correspond to each other. Finally, the characteristic variables of the system are determined to be wind speed, cabin temperature, ambient temperature, gearbox oil temperature and main bearing temperature. The time domain analysis comparison of the 15#2MW unit of Yunnan Daheishan Wind Farm and the 61#8.3MW unit of Guangdong Bodhi Sea Wind Farm is shown in Table 4.

[0110] Table 4 Evaluation indicators of experimental results of different models

[0111]

[0112] As can be seen from Table 4, the mean and standard deviation of the data of the 15#2MW unit of Yunnan Daheishan Wind Farm and the 3#6.8MW unit and 61#8.3MW unit of Guangdong Bodhihai Wind Farm are quite different, which makes migration difficult.

[0113] Within the given range of training parameters, the final result of model training must minimize the loss function and obtain the optimal model.

[0114] After determining the parameter adjustment range, the genetic algorithm is used for optimization. The overall process is as follows:

[0115] (1) Set the hyperparameter range of XGBoost.

[0116] (2) The genetic algorithm (GA) is embedded in the XGBoost algorithm to optimize hyperparameters. Specifically, the objective function of the XGBoost model is defined, the input of which is the hyperparameter combination, and the output is the evaluation index of the model on the test set.

[0117] (3) Iterate through the genetic algorithm, adjust the hyperparameter combination, and update the individuals in the population to find the global optimal hyperparameter combination, as shown in Table 5.

[0118] (4) Input the global optimal parameter combination into the XGBoost model for training and the accuracy of the XGBoost model.

[0119] Table 5. Parameters of Genetic Algorithm Optimization of XGBoost

[0120] Parameter name Optimal parameters Maximum Depth 12 Learning Rate 0.1 Weak Learner 360 Number of leaf nodes 1.6

[0121] After completing the model migration, we tested the data set before the fault was discovered. We used 100 data points (about three days) to predict the fine-tuned XGBoost model. We found that the prediction accuracy of migrating the XGBoost model of the Guangdong Bodhi Sea Offshore Wind Farm to the Yunnan Daheishan Onshore Wind Farm was lower than the prediction accuracy of migrating between different types of units in the Guangdong Bodhi Sea Offshore Wind Farm. However, this prediction accuracy is sufficient to achieve early warning of the main bearing fault of the wind turbine. We conducted a fault warning experiment on the historical data before the fault was discovered. The experimental results are as follows: Figure 8 shown.

[0122] Comparative example: In order to verify the effectiveness of transfer learning, the following experimental scheme is designed to verify it between different units:

[0123] 1. Model settings:

[0124] Model 1: The basic model trained using 21 months of data from Unit 3.

[0125] Model 2-Model 5: Transfer learning models based on Model 1 were fine-tuned using 10 days of data from units 24, 40, 58, and 61 (a total of 1,440 records).

[0126] 2. Unit characteristics:

[0127] Units 3, 24, and 40 are 6.8MW, and units 58 and 61 are 8.3MW.

[0128] 3. Comparison model:

[0129] For comparison, the RF (Random Forest) model was selected, which performed closest to the XGBoost model in the analogy model experiment.

[0130] Through the above experimental design, the goal is to verify the effectiveness of transfer learning between different units in the case of few samples, and compare it with the RF model to evaluate the effect of transfer learning on the performance of the main bearing temperature warning model. The transfer learning model will be fine-tuned on the limited data of each unit to adapt to the new environment, so as to obtain better prediction performance when the newly built wind farm data is insufficient. The results are shown in Table 6.

[0131] Table 6 Comparison test between XGBoost model and Random forest

[0132]

[0133] According to the experimental results, the transfer learning effect of the XGBoost model on different units is better than that of the RandomForest model. Among them, the transfer performance of the XGBoost model on units 24 and 40 is better than that on units 58 and 61. This is because units 24 and 40 are geographically closer to unit 3 and have the same specifications, which means that the pre-trained model contains more common features. When the data sample is relatively limited, when the main bearing temperature fluctuates greatly, the R 2 This means that in this scenario R 2 The reference value is low. As shown in the table, Unit 61 also achieved a good fit index with a lower RMSE. In contrast, the performance of the Random Forest model is relatively poor. The verification of each unit after migration is as follows Figure 7 shown.

[0134] In summary, the present invention can improve the accuracy of fault diagnosis and fault warning capability, improve the versatility of the fault diagnosis model, and reduce the overall fault diagnosis cost.

Claims

1. A wind turbine main bearing fault diagnosis method based on machine learning algorithm, characterized in that: The following steps are involved: Step 1: Obtain the operating data of the wind turbine, perform preprocessing, normalization and feature extraction on the operating data to obtain a data set; Step 2: Build an XGBoost model, input the data set into the XGBoost model for training, and use a genetic algorithm to optimize the hyperparameters of the XGBoost model until the model prediction effect reaches the optimal value, thus obtaining the optimal XGBoost model; Step 3: Predict the main bearing temperature of the wind turbine using the optimal XGBoost model, perform adaptive threshold monitoring on the difference between the predicted value and the true value of the main bearing temperature, and complete the main bearing fault diagnosis of the wind turbine; Step 4: Use transfer learning to migrate the optimal XGBoost model to other wind turbines, predict the main bearing temperature of other wind turbines based on their operating data, and then realize the main bearing fault diagnosis of other wind turbines.

2. The wind turbine main bearing fault diagnosis method based on machine learning algorithm according to claim 1 is characterized in that: In step 1, the operation data of the wind turbine is collected by a SCADA system.

3. The wind turbine main bearing fault diagnosis method based on machine learning algorithm according to claim 1 is characterized in that: The data preprocessing includes eliminating missing values, abnormal values ​​and data when the wind turbine generator set is not working properly in the operation data, and then processing the abnormal values ​​of the parameters by using the 3σ rule.

4. The wind turbine main bearing fault diagnosis method based on machine learning algorithm according to claim 1 is characterized in that: The normalization process is maximum and minimum value normalization, and the formula is as follows: The process of data normalization is as follows: In the formula, x max and x min They represent the maximum and minimum values ​​of a characteristic parameter in the running data respectively; x represents the current processing value of the characteristic parameter; x * Indicates the normalized value.

5. The wind turbine main bearing fault diagnosis method based on machine learning algorithm according to claim 1 is characterized in that: The feature extraction process includes first using the Pearson correlation coefficient method to calculate the correlation coefficient between feature parameters, and then using the feature importance index in the XGBoost model to screen features, and the calculation method uses the "gain" weight form.

6. The wind turbine main bearing fault diagnosis method based on machine learning algorithm according to claim 1, characterized in that: In step 2, the construction process of the XGBoost model is as follows: Given a training set of n samples and m features: D=(x i ,y i )(|D|=n,x i ∈R m ,y i ∈R); The prediction function is defined as follows: in, is the i-th sample x i The predicted value; K represents all decision trees in the XGBoos model; f k is the kth regression decision tree; f k (x i ) represents the score of the i-th sample in the k-th decision tree; The XGBoost model uses a forward addition model, using multiple decision trees to make decisions together, and the result of each decision tree is the difference between the target value and the prediction results of all previous decision trees, and all the results are accumulated to get the final result; the loss function of the XGBoost model is as follows: Among them, L(θ) represents the loss function of the XGBoos model, n is the number of samples, T is the number of leaf nodes, ω j is the score of each leaf node, γ and λ are regularization parameters; the loss function of the XGBoost model is composed of empirical loss and the regularization term The loss function is minimized within the given training parameter range to obtain the optimal model.

7. The wind turbine main bearing fault diagnosis method based on machine learning algorithm according to claim 1 is characterized in that: In step 3, the method of predicting the main bearing temperature of the wind turbine generator set and performing adaptive threshold monitoring on the difference between the predicted value and the actual value of the main bearing temperature is as follows: After data preprocessing, normalization and feature extraction, the hyperparameters of the XGBoost model are adjusted, the maximum depth, learning rate and weak estimator are selected to adjust the XGBoost model, and the grid search is used to optimize the XGBoost model hyperparameter adjustment to obtain the XGBoost model with the best prediction effect, and the main bearing temperature is predicted; then the prediction effect of the XGBoost model is evaluated by the root mean square error and mean absolute percentage error.

Images