A wind turbine blade icing diagnosis method based on synthetic minority samples

By combining the instantaneous and timing characteristics of blade icing, a few samples were synthesized using KBS2 to solve the accuracy of identification of the icing state of the wind turbine blades, and the accurate diagnosis of the blades of the wind turbine is achieved, which is a wind turbine. Operational maintenance and failure prediction provide a reliable approach.

CN115450864BActive Publication Date: 2024-11-15NORTH CHINA ELECTRIC POWER UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211341472.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-10-26
Publication Date
2024-11-15
Estimated Expiration
2042-10-26

AI Technical Summary

Technical Problem

When processing unbalanced data of blade icing in the wind turbine assembly, the use of synthetic a few samples oversampling methods leads to noise in the fitted data, affecting the model's accurate identification of the blade icing state, making it difficult for wind farms to achieve accurate diagnosis.

Method used

A method for icing blades of wind turbines based on the synthesis of a few samples is proposed. By monitoring and recording the freezing state of wind turbines, combining the instantaneous characteristics of blade icing, the original SCADA data and time sequence characteristics, the quartile method and bin are used. The method fits the theoretical power curve, uses KBS2 to synthesize a few samples to form a new data set, and uses machine learning methods to train the prediction model of the blade icy state of the wind turbine set.

Benefits of technology

Improve the model's learning ability of icing data, realize accurate prediction of the icing state of the wind turbine blades, solve the problem of icing feature extraction and data proportion too low, and provide the operation and maintenance of wind turbines and fault prediction. A reliable method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115450864B_ABST
    Figure CN115450864B_ABST
Patent Text Reader

Abstract

The present invention discloses a wind turbine blade icing diagnosis method based on synthetic minority class samples, which belongs to the technical field of wind turbine status warning and fault diagnosis. The method includes step 1: monitoring and recording the icing state of the wind turbine, adding icing labels to the corresponding SCADA data, and combining the instantaneous features, features extracted from the original SCADA data, and time series features into the final features; step 2: fitting the theoretical power curve, using the KBS2 synthetic minority class sample oversampling method to form a new data set; step 3: performing multiple iterations on the training set, and obtaining the accuracy index of the test set; step 4: offline training, online deployment and application of the wind turbine blade icing state prediction model. The present invention can solve the problem of wind turbine blade icing feature extraction and the low proportion of icing data, and provide a reliable methodological basis for wind turbine operation and maintenance, status warning and fault prediction research.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine status warning and fault diagnosis, and in particular to a wind turbine blade icing diagnosis method based on synthesized minority class samples. Background Art

[0002] Wind energy is a kind of renewable energy with no pollution to the environment and a wide range of sources. It is one of the most promising renewable energy sources. Vigorously developing wind power generation will help solve problems such as energy shortage and climate deterioration. As an important component for absorbing and converting wind energy, wind turbine blades play a vital role in the actual power output of wind turbines. However, wind farms are generally located in high-latitude areas. Although wind resources are abundant in high-latitude areas, the temperature is low and the humidity is high, which causes wind turbine blades to often freeze in winter. This greatly reduces the efficiency and output of wind energy utilization by wind turbines, and in severe cases, it causes the blades to lose control and break, posing serious safety hazards. There are still some problems with the existing technology for diagnosing the icing state of wind turbine blades. In particular, when processing unbalanced data of wind turbine blade icing, a synthetic minority sample oversampling method is usually used to balance the sample data ratio, but the data fitted by this method is noisy, which affects the model's accurate identification of the blade icing state. Therefore, wind farms are in urgent need of accurately diagnosing the state of the unit blades. Summary of the invention

[0003] The purpose of the present invention is to propose a wind turbine blade icing diagnosis method based on synthetic minority class samples, characterized in that the method comprises:

[0004] Step 1: Monitor and record the icing status of the wind turbine, add icing tags to the corresponding SCADA data, and combine the instantaneous features of blade icing, the features extracted from the original SCADA data, and the blade icing time series features into the final features;

[0005] Step 2: Use the quartile method and bin method to fit the theoretical power curve, construct new data based on the final features, and use KBS2 (BorderlineSMOTE-KNN-BorderlineSMOTE) to synthesize minority class sample oversampling method to form a new data set;

[0006] Step 3: Divide 80% of the new data set formed in step 2 into a training set and 20% into a test set. Use machine learning methods to iterate the training set multiple times to train the wind turbine blade icing state prediction model, and obtain the accuracy index of the test set.

[0007] Step 4: Offline training of the wind turbine blade icing state prediction model, and online deployment and application of the wind turbine blade icing state prediction model.

[0008] The step 1 specifically includes the following sub-steps:

[0009] Step 11: Select SCADA data of n wind turbines that are in good operating condition during the icing period;

[0010] Step 12: Record the freezing period and normal period of the wind turbine during the observation period;

[0011] Step 13: Add an icing status tag to the SCADA data during the monitoring period;

[0012] Step 14: Construct the instantaneous characteristics of blade icing;

[0013] Step 15: Using the recursive feature elimination method, extract the features that are highly correlated with blade icing, including generator speed, ambient temperature, cabin temperature, yaw angle, yaw position, and pitch motor temperature;

[0014] Step 16: Set the sliding window length to h, width to j, and sliding step to d, and use the sliding window method to calculate the standard deviation of each feature extracted in step 14 and step 15 in a continuous time period to obtain the time series feature;

[0015] Step 17: Combine the features extracted in step 14 and step 15 as short-term features, and combine the short-term features with the time series features to obtain the final features.

[0016] The instantaneous characteristics of blade icing in step 14 include:

[0017] Theoretical power P pred :

[0018] P pred =f(P real ) (1)

[0019] In formula (1), P real is the actual power, f is the fitted polynomial function, P pred is the theoretical power obtained according to the fitted polynomial function;

[0020] The square of wind speed V 2 :

[0021]

[0022] In formula (2), V in is the wind speed;

[0023] The cube of wind speed V 3 :

[0024]

[0025] Tip speed ratio λ:

[0026]

[0027] In formula (4), ω is the angular velocity of the wind turbine blade, and R is the blade radius;

[0028] Icing severity ξ:

[0029]

[0030] The short-term characteristics include wind speed, wind speed square, wind speed cube, grid-side active power, theoretical power, blade tip speed ratio, icing severity, generator speed, ambient temperature, cabin temperature, yaw angle, yaw position, and pitch motor temperature;

[0031] The time series characteristics include the standard deviation of wind speed, the standard deviation of the square of wind speed, the standard deviation of the cube of wind speed, the standard deviation of grid-side active power, the standard deviation of theoretical power, the standard deviation of tip speed ratio, the standard deviation of icing severity, the standard deviation of generator speed, the standard deviation of ambient temperature, the standard deviation of cabin temperature, the standard deviation of yaw angle, the standard deviation of yaw position, and the standard deviation of pitch motor temperature.

[0032] The step 2 specifically includes the following sub-steps:

[0033] Step 21: Divide the wind speed into n equally spaced intervals, use the quartile method to clean the data, and remove abnormal values ​​of wind speed and power;

[0034] Step 22: Using the bin method, calculate the average wind speed and power of all data points in the n intervals;

[0035] Step 23: Fit the power curve and the polynomial function according to the average values ​​of wind speed and power in each interval, and substitute the actual power into the polynomial function to obtain the theoretical power;

[0036] Step 24: Based on the final features obtained in step 17, a new data set is formed;

[0037] Step 25: Perform maximum and minimum normalization on the data of the new data set according to formula (6).

[0038]

[0039] In formula (6): x i For each feature, x min is the minimum value of each feature, x max is the maximum value of each feature;

[0040] Step 26: Divide the normalized data set in step 25 into feature data x and icing label data y;

[0041] Step 27: For the minority class sample in the data of step 26, calculate the Euclidean distance to the b points closest to it according to formula (7);

[0042]

[0043] In formula (7), d represents the Euclidean distance between sample x'(x1,y1) and sample x"(x2,y2) in two-dimensional space;

[0044] According to the calculated Euclidean distance d, determine the number of majority class samples and minority class samples among the b points closest to the minority class samples;

[0045] For the dangerous minority class sample, use formula (7) to calculate the c points near the sample, randomly select one of the points, and use formula (8) to synthesize a new minority class sample until the number of frozen samples in the data set is balanced with the number of normal samples;

[0046] x new =x+rand(0,1)*(x neighbor -x) (8)

[0047] In formula (8): x new is a new minority class sample synthesized, x is a dangerous minority class sample, rand(0,1) is a random number between 0 and 1, x neighbor is a neighbor sample of the selected x;

[0048] Step 28: Substitute the resampled data set of dangerous minority samples into the KNN model for repeated training, and delete the data samples diagnosed by the KNN model as not icing but actually icing;

[0049] Step 29: Repeat the process of step 27 for the data set obtained in step 28 until the number of frozen samples and normal samples in the data set is balanced.

[0050] The minority class samples in step 28 are divided into the following categories:

[0051] If the number of minority class samples among b points is greater than or equal to the number of majority class samples, then the minority class samples belong to safe minority class samples;

[0052] If among the b points, the number of minority class samples is less than the number of majority class samples and the number of minority class samples is not zero, then the minority class samples belong to dangerous minority class samples;

[0053] If the number of minority class samples among b points is zero, then the minority class samples belong to the noise minority class samples.

[0054] The accuracy index of the test set in step 3 is:

[0055]

[0056] In formula (9), TN is the total number of times the model predicts the actual icing data as icing state; FN is the total number of times the model predicts the actual non-icing data as icing state.

[0057] The step 4 specifically includes the following sub-steps:

[0058] Step 41: Set the time resolution to m minutes. When the duration of the updated SCADA data exceeds m minutes, add the updated data to the historical data to form a new wind turbine data set.

[0059] Step 42: Repeat the operation of step 2 for the updated SCADA data to form a new data set for training;

[0060] Step 43: Perform the operation of step 3 on the new data set obtained in step 42;

[0061] Step 44: Predict the blade icing state according to the trained model obtained in step 43, and issue a warning based on the predicted blade icing state.

[0062] The beneficial effects of the present invention are:

[0063] The method provided by the present invention improves the model's ability to learn icing data, can solve the problems of icing feature extraction of wind turbine blades and too low a proportion of icing data, and thus achieve accurate icing prediction of the icing state of wind turbine blades, providing a reliable methodological basis for wind turbine operation and maintenance, status warning and fault prediction research. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The flowchart of the wind turbine blade icing diagnosis method based on synthetic minority samples;

[0065] Figure 2 Flowchart of the method for synthesizing minority class sample oversampling based on mixed features and KBS2;

[0066] Figure 3 It is the relationship diagram of fitting theoretical power curve over time;

[0067] Figure 4 This is a comparison chart of the diagnostic accuracy of different algorithms using different synthetic minority sample methods. DETAILED DESCRIPTION

[0068] The present invention proposes a wind turbine blade icing diagnosis method based on synthesized minority class samples, which is further described below in conjunction with the accompanying drawings and specific embodiments.

[0069] Figure 1The flowchart of the wind turbine blade icing diagnosis method based on synthetic minority samples is as follows:

[0070] Step 1: Monitor and record the icing status of wind turbines, add icing labels to the corresponding SCADA data, and combine the instantaneous features of blade icing, the features extracted from the original SCADA data using the recursive feature elimination method, and the blade icing time series features constructed using the sliding window method into the final features. The specific steps are as follows:

[0071] Step 11: Select the SCADA data of a wind turbine that is in good operating condition during the icing period; the data source is the 2017 China Industrial Big Data Innovation Competition;

[0072] Step 12: Record the freezing period and normal period of the wind turbine during the observation period;

[0073] Step 13: Add an icing status tag to the SCADA data during the monitoring period. The icing status tag is shown in Table 1 below:

[0074] Table 1

[0075]

[0076] Step 14: Construct the instantaneous feature of blade icing. The newly constructed instantaneous feature and formula are as follows:

[0077] Theoretical power P pred :

[0078] P pred =f(P real ) (1)

[0079] In formula (1): P real is the actual power, f is the fitted polynomial function, P pred is the theoretical power obtained based on the fitted polynomial function.

[0080] Wind speed square V 2 :

[0081]

[0082] In formula (2): V in is the wind speed.

[0083] Wind speed cubed V 3 :

[0084]

[0085] Tip speed ratio λ:

[0086]

[0087] In formula (4), ω is the angular velocity of the wind turbine blade, and R is the blade radius.

[0088] Icing severity ξ:

[0089]

[0090] The features extracted from the instantaneous features of blade icing include: wind speed, wind speed square, wind speed cube, grid-side active power, theoretical power, blade tip speed ratio, and icing severity;

[0091] Step 15: Use the recursive feature elimination method to extract features that are highly correlated with blade icing, including: generator speed, ambient temperature, cabin temperature, yaw angle, yaw position, pitch motor 1 temperature, pitch motor 2 temperature, pitch motor 3 temperature and wind speed;

[0092] Step 16: Using the sliding window method, for the feature variables extracted in steps 14 and 15, calculate the standard deviation of each feature variable in the continuous time period to obtain the time series features, set the sliding window length to 6, the width to 1; and the sliding step length to 6;

[0093] Step 17: Combine the features extracted by mining the blade icing process in step 14 with the features extracted by the artificial intelligence method in step 15 as short-term features. The short-term features include: wind speed, wind speed square, wind speed cube, grid-side active power, theoretical power, blade tip speed ratio, icing severity, generator speed, ambient temperature, cabin temperature, yaw angle, yaw position, pitch motor 1 temperature, pitch motor 2 temperature, pitch motor 3 temperature wind speed; use the features obtained by step 16 as time series features, which include: standard deviation of wind speed within 1 hour, standard deviation of wind speed square within 1 hour, standard deviation of wind speed cube within 1 hour The standard deviation of the grid-side active power within 1 hour, the standard deviation of the theoretical power within 1 hour, the standard deviation of the tip speed ratio within 1 hour, the standard deviation of the icing severity within 1 hour, the standard deviation of the generator speed within 1 hour, the standard deviation of the ambient temperature within 1 hour, the standard deviation of the cabin temperature within 1 hour, the standard deviation of the yaw angle within 1 hour, the standard deviation of the yaw position within 1 hour, the standard deviation of the temperature of the pitch motor 1 within 1 hour, the standard deviation of the temperature of the pitch motor 2 within 1 hour, and the standard deviation of the temperature of the pitch motor 3 within 1 hour; the short-term features extracted and constructed from the original SCADA data of the wind turbine are combined with the time series features to obtain the final features;

[0094] Step 2: Use the quartile method and bin method to fit the theoretical power curve. Based on the final features obtained in step 17, construct new data and use the KBS2 synthetic minority class sample oversampling method to form a new data set. Figure 2This is a flowchart of the method for synthesizing minority class sample oversampling based on mixed features and KBS2. The specific steps are as follows:

[0095] Step 21: Divide the wind speed into 100 equally spaced intervals, use the quartile method to clean the data, and remove outliers in wind speed and power;

[0096] Step 22: Using the bin method, calculate the average of wind speed and power for all data points in 100 intervals;

[0097] Step 23: Fitting a power curve and a polynomial function according to the average values ​​of wind speed and power in each interval, substituting the wind speed into the polynomial function to obtain the theoretical power; Figure 3 It is the relationship diagram of fitting theoretical power curve over time;

[0098] Step 24: Based on the final features obtained in step 17, a new data set is formed;

[0099] Step 25: Perform maximum and minimum normalization on the data of the deicing tag in the new SCADA data obtained in step 24. The maximum and minimum normalization formula is as follows:

[0100]

[0101] In formula (6): x i For each feature variable, x min is the minimum value of each feature variable, x max is the maximum value of each feature variable;

[0102] Step 26: Divide the data set normalized in step 25 into x and y, where x is the data containing the feature variables and y is the ice label data;

[0103] Step 27: Calculate the Euclidean distance between the minority class sample in the step 26 data and the 7 points closest to it. The formula is as follows:

[0104]

[0105] In formula (7), d represents the Euclidean distance between sample x'(x1,y1) and sample x"(x2,y2) in two-dimensional space;

[0106] According to the calculated Euclidean distance d, the number of majority class samples and minority class samples among the 7 points closest to the minority class samples is determined, and the minority class samples are divided into three categories:

[0107] (1) If the number of minority class samples among the 7 points is greater than or equal to the number of majority class samples, then the minority class samples belong to the safe minority class samples;

[0108] (2) If among the 7 points, the number of minority class samples is less than the number of majority class samples and the number of minority class samples is not zero, then the minority class samples belong to dangerous minority class samples;

[0109] (3) If the number of minority class samples among the 7 points is zero, then the minority class samples belong to the noise minority class samples;

[0110] For the dangerous minority class sample, use formula (7) to calculate the five points near the sample, randomly select one of the points, and use formula (8) to synthesize a new minority class sample until the number of frozen samples in the data set is balanced with the number of normal samples. Formula (8) is as follows:

[0111] x new =x+rand(0,1)*(x neighbor -x) (8)

[0112] In formula (8): x new is a new minority class sample synthesized, x is a dangerous minority class sample, rand(0,1) is a random number between 0 and 1, x neighbor is a neighbor sample of the selected x;

[0113] Step 28: Substitute the data set after resampling the dangerous minority class samples in step 27 into the KNN model for repeated training, and delete the data samples diagnosed by the KNN model as not icing but actually icing;

[0114] Step 29: Repeat the process of step 27 for the data set obtained in step 28 until the number of frozen samples and normal samples in the data set is balanced;

[0115] Step 3: Divide 80% of the data set into a training set and 20% into a test set. Use machine learning methods to iterate the training set multiple times to train the wind turbine blade icing diagnosis model. Use the test set to evaluate the model accuracy. The specific steps are as follows:

[0116] Step 31: Divide 80% of the data set obtained in step 2 into a training set and 20% into a test set;

[0117] Step 32: Substitute the training set obtained in step 31 into the LSTM, GRU, and BiLSTM models, and train the wind turbine blade icing diagnosis model after multiple iterations;

[0118] Step 33: Substitute the test set obtained in step 31 into the trained wind turbine blade icing diagnosis model to obtain the accuracy index of the test set. The formula is as follows:

[0119]

[0120] In formula (9), TN is the total number of times the model diagnoses the actual icing data as icing state; FN is the total number of times the model diagnoses the actual non-icing data as icing state;

[0121] Step 4: Offline training of the blade icing status diagnosis model, online deployment and application of the blade icing status diagnosis model. The specific steps are as follows:

[0122] Step 41: Set the time resolution to 10 minutes. When the duration of the updated SCADA data exceeds 10 minutes, add the updated data to the historical data to form a new wind turbine data set.

[0123] Step 42: Repeat the operation described in step 2 for the new SCADA data to form a new data set that can be used for training;

[0124] Step 43: Perform the operation described in step 3 on the new data set obtained in step 42;

[0125] Step 44: diagnose the blade icing state according to the trained model obtained in step 43, and issue an early warning according to the blade icing state diagnosed by the model;

[0126] Step 5: Compare the diagnostic accuracy of different machine learning models using the SMOTE synthetic minority sample oversampling method, the BorderlineSMOTE synthetic minority sample oversampling method, and the KBS2 synthetic minority sample oversampling method proposed in this paper. Figure 4 This is a comparison chart of the diagnostic accuracy of different algorithms using different synthetic minority sample methods. The comparison results are shown in Table 2 below:

[0127] Table 2

[0128]

[0129] The present invention first extracts and constructs short-term features with great correlation with blade icing from original SCADA data based on the physical process of wind turbine blade icing and the recursive feature elimination method, and uses the sliding window method to construct time series features with great correlation with blade icing. The short-term features are combined with the time series features to obtain the final features; in view of the problem that the ratio of icing data to normal data is too low, a KBS2 synthetic minority class sample oversampling method is proposed to increase the ratio of icing data to normal data; and a machine learning method is used to train and diagnose the model. This method can solve the problem of icing feature extraction and low proportion of icing data of wind turbine blades, and then realize accurate icing diagnosis of the icing state of wind turbine blades, providing a reliable methodological basis for wind turbine operation and maintenance, status warning and fault diagnosis research.

Claims

1. A wind turbine blade icing diagnosis method based on synthetic minority samples, It is characterized in that The method comprises: Step 1: Monitor and record the icing status of the wind turbine, add icing tags to the corresponding SCADA data, and combine the instantaneous features of blade icing, the features extracted from the original SCADA data, and the blade icing time series features into the final features; The step 1 specifically includes the following sub-steps: Step 11: Select SCADA data of n wind turbines that are in good operating condition during the icing period; Step 12: Record the freezing period and normal period of the wind turbine during the observation period; Step 13: Add an icing status tag to the SCADA data during the monitoring period; Step 14: Construct the instantaneous characteristics of blade icing; The instantaneous characteristics of blade icing in step 14 include: Theoretical power P pred : P pred =f(P real ) (1) In formula (1), P real is the actual power, f is the fitted polynomial function, P pred is the theoretical power obtained according to the fitted polynomial function; The square of wind speed V 2 : In formula (2), V in is the wind speed; The cube of wind speed V 3 : Tip speed ratio λ: In formula (4), ω is the angular velocity of the wind turbine blade, and R is the blade radius; Icing severity ξ: Step 15: Using the recursive feature elimination method, extract the features that are highly correlated with blade icing, including generator speed, ambient temperature, cabin temperature, yaw angle, yaw position, and pitch motor temperature; Step 16: Set the sliding window length to h, width to j, and sliding step to d, and use the sliding window method to calculate the standard deviation of each feature extracted in step 14 and step 15 in a continuous time period to obtain the time series feature; Step 17: Combine the features extracted in step 14 and step 15 as short-term features, and combine the short-term features with the time series features to obtain the final features; The short-term characteristics include wind speed, wind speed square, wind speed cube, grid-side active power, theoretical power, blade tip speed ratio, icing severity, generator speed, ambient temperature, cabin temperature, yaw angle, yaw position, and pitch motor temperature; The time series characteristics include the standard deviation of wind speed, the standard deviation of the square of wind speed, the standard deviation of the cube of wind speed, the standard deviation of grid-side active power, the standard deviation of theoretical power, the standard deviation of tip speed ratio, the standard deviation of icing severity, the standard deviation of generator speed, the standard deviation of ambient temperature, the standard deviation of cabin temperature, the standard deviation of yaw angle, the standard deviation of yaw position, and the standard deviation of pitch motor temperature; Step 2: Use the quartile method and bin method to fit the theoretical power curve, construct new data based on the final features, and use the BorderlineSMOTE-KNN-BorderlineSMOTE method to synthesize minority class sample oversampling to form a new data set; The step 2 specifically includes the following sub-steps: Step 21: Divide the wind speed into n equally spaced intervals, use the quartile method to clean the data, and remove abnormal values ​​of wind speed and power; Step 22: Using the bin method, calculate the average wind speed and power of all data points in the n intervals; Step 23: Fit the power curve and the polynomial function according to the average values ​​of wind speed and power in each interval, and substitute the actual power into the polynomial function to obtain the theoretical power; Step 24: Based on the final features obtained in step 17, a new data set is formed; Step 25: Perform maximum and minimum normalization on the data of the new data set according to formula (6). In formula (6): x i For each feature, x min is the minimum value of each feature, x max is the maximum value of each feature; Step 26: Divide the normalized data set in step 25 into feature data x and icing label data y; Step 27: For the minority class sample in the data of step 26, calculate the Euclidean distance to the b points closest to it according to formula (7); In formula (7): d represents the sample x'(x 1 ,y 1 ) and sample x”(x 2 ,y 2 ) in two-dimensional space; According to the calculated Euclidean distance d, determine the number of majority class samples and minority class samples among the b points closest to the minority class samples; For the dangerous minority class sample, use formula (7) to calculate the c points near the sample, randomly select one of the points, and use formula (8) to synthesize a new minority class sample until the number of frozen samples in the data set is balanced with the number of normal samples; x new =x+rand(0,1)*(x neighbor -x) (8) In formula (8): x new is a new minority class sample synthesized, x is a dangerous minority class sample, rand(0,1) is a random number between 0 and 1, x neighbor is a neighbor sample of the selected x; Step 28: Substitute the resampled data set of dangerous minority samples into the KNN model for repeated training, and delete the data samples diagnosed by the KNN model as not icing but actually icing; The minority class samples in step 28 are divided into the following categories: If the number of minority class samples among b points is greater than or equal to the number of majority class samples, then the minority class samples belong to safe minority class samples; If among the b points, the number of minority class samples is less than the number of majority class samples and the number of minority class samples is not zero, then the minority class samples belong to dangerous minority class samples; If the number of minority class samples among b points is zero, then the minority class samples belong to the noise minority class samples; Step 29: Repeat the process of step 27 for the data set obtained in step 28 until the number of frozen samples and normal samples in the data set is balanced; Step 3: Divide 80% of the new data set formed in step 2 into a training set and 20% into a test set. Use machine learning methods to iterate the training set multiple times to train the wind turbine blade icing state prediction model, and obtain the accuracy index of the test set. Step 4: Offline training of the wind turbine blade icing state prediction model, online deployment and application of the wind turbine blade icing state prediction model; The step 4 specifically includes the following sub-steps: Step 41: Set the time resolution to m minutes. When the duration of the updated SCADA data exceeds m minutes, add the updated data to the historical data to form a new wind turbine data set. Step 42: Repeat the operation of step 2 for the updated SCADA data to form a new data set for training; Step 43: Perform the operation of step 3 on the new data set obtained in step 42; Step 44: Predict the blade icing state according to the trained model obtained in step 43, and issue a warning based on the predicted blade icing state.

2. According to the wind turbine blade icing diagnosis method based on synthetic minority samples according to claim 1, It is characterized in that The accuracy index of the test set in step 3 is: In formula (9), TN is the total number of times the model predicts the actual icing data as icing state; FN is the total number of times the model predicts the actual non-icing data as icing state.