A Yaw Stuck Diagnosis Method and System for Wind Turbine Based on SVM Algorithm

Abstract

The present invention discloses a yaw jamming diagnosis method and system for wind turbine based on SVM algorithm. The method comprises the following steps: S1, obtaining original data; S2, preprocessing the original data; S3, performing feature extraction; S4, dividing the training data into a training set and a test set, and annotating the training set; S5, inputting the annotated training set into the SVM algorithm for training to obtain a yaw jamming diagnosis model; S6, calculating the accuracy of the test set and constructing an ROC curve; S7, evaluating the classification effect of the yaw jamming diagnosis model; S8, manually verifying the accuracy of the actual recognition result of the yaw jamming diagnosis model. If the verification is qualified, the yaw jamming diagnosis model is released and the analysis is started. By obtaining the sensor data of the wind turbine and improving and optimizing the yaw control algorithm, the present invention can quickly and effectively identify the yaw jamming during the operation of the wind turbine, give a timely warning feedback to the site, reduce the component failure rate and operation and maintenance costs, and save costs.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine yaw diagnosis, and in particular to a wind turbine yaw jamming diagnosis method and system based on a SVM algorithm. Background Art

[0002] Wind turbines have been operating in harsh natural environments for a long time. Although the extreme loads caused by extreme wind conditions have been considered during the research and development of wind turbines, the deviation between the actual wind resources and the feasibility study data leads to inconsistencies between the actual and simulated conditions. Extreme working conditions still occur during operation, which leads to exceeding the design limit and causing force majeure damage to the various components of the wind turbines. Eventually, wind turbines frequently fail. The yaw system of a wind turbine plays a vital role in power generation. Early detection and maintenance of frequent yaw jams or targeted adjustments and optimization of the control system will result in less economic losses than the development of failures to component damage and shutdown for replacement. High-frequency yaw jams will cause damage to the yaw drive and then cause yaw system failures, causing the wind turbine to need unplanned shutdowns for maintenance, which will reduce power generation performance at the least and cause serious damage to key components such as the yaw motor and yaw brake in the yaw system at the worst. In addition, the vibration caused by yaw jams will indirectly increase the fatigue load of the unit. These effects not only increase the operation and maintenance cost pressure on wind power manufacturers, but also deteriorate the operating performance of wind turbines, resulting in serious economic losses. Therefore, rapid detection of wind turbine yaw jam and early warning have become issues of keen concern in the wind power industry, but there is still little research on yaw jam detection at this stage.

[0003] At present, the main fault detection methods for wind turbines include data-based fault detection and model-based fault detection. Although the former avoids the complex and time-consuming modeling process and can quickly detect faults, it is extremely limited to identify faults by manually setting relevant thresholds or manual observations, and it is easy to cause misjudgment when the system is disturbed or the state changes during operation. When faced with massive amounts of fault data, the efficiency of mining patterns through manual data analysis is far from expected. The latter can make full use of the internal deep information of the entire unit system and effectively reflect the essential characteristics of physical system faults. It not only solves the complicated manual calculations, but also can mine hidden patterns that cannot be manually identified, and is driven by data for self-learning to continuously improve the accuracy of the early warning model.

[0004] With the rapid development of artificial intelligence technology, machine learning and deep learning methods have been widely used in the field of wind power, such as using neural networks to detect faults in the variable pitch system of wind turbines; based on the improved random forest algorithm, fault detection and early warning of large components of wind turbines, etc. However, such methods have not yet been applied to the yaw jamming problem of wind turbines. Summary of the invention

[0005] The first object of the present invention is to solve the deficiencies in the prior art, and provides a yaw jamming diagnosis method for a wind turbine based on the SVM algorithm. By acquiring the sensor data of the wind turbine and improving and optimizing the yaw control algorithm, the yaw jamming problem can be quickly diagnosed and early warning feedback can be carried out, which can quickly and effectively identify the yaw jamming during the operation of the unit, timely give early warning feedback to the site, reduce the component failure rate and operation and maintenance costs, and save costs.

[0006] The second object of the present invention is to provide a yaw jamming diagnosis system for a wind turbine based on the SVM algorithm.

[0007] The first object of the present invention is achieved by the following technical solutions: A yaw jamming diagnosis method for a wind turbine based on the SVM algorithm, comprising the following steps:

[0008] S1. Acquire the sensor data recording the fault actions of the wind turbine as the original data;

[0009] S2. Perform preprocessing of normalization on the original data;

[0010] S3. Extract features from the preprocessed original data to obtain model data;

[0011] S4. Perform upsampling processing on the model data, then divide it into a training set and a test set, and label the training set according to the manual annotation and weak supervision learning method;

[0012] S5. Input the labeled training set into the SVM algorithm for training to obtain a yaw jamming diagnosis model;

[0013] S6. Verify and calculate the accuracy of the test set, and construct an ROC curve based on the data of normal yaw and yaw jamming of the wind turbine;

[0014] S7. Evaluate the effect of the yaw jamming diagnosis model according to the accuracy of the test set and the ROC curve. If the effect of the evaluated yaw jamming diagnosis model meets the preset requirements, execute step S8; if the effect of the evaluated yaw jamming diagnosis model does not meet the preset requirements, repeat steps S1 to S6 until the effect of the evaluated yaw jamming diagnosis model meets the preset requirements, and then execute step S8;

[0015] S8. Manually verify the accuracy of the actual recognition result of the yaw jamming diagnosis model. If the verification is unqualified, add new data samples to the training set, and repeat steps S5 to S7 until the verification is qualified; if the verification is qualified, release the yaw jamming diagnosis model and start the analysis to give early warning of the yaw fault problem of the wind turbine.

[0016] Furthermore, the step S1 includes the following steps:

[0017] The sensor data is Tracelog data, and each Tracelog data is millisecond-level data within a one-minute time period before and after a fault action of a wind turbine generator.

[0018] Screen the obtained Tracelog data, and extract the data of the time period that records the yaw jamming action of the wind turbine generator: Compare the change rate of the nacelle azimuth angle. When the wind turbine generator is in normal yaw, the change rate of the nacelle azimuth angle is uniform and unchanged, that is, the change rates of the nacelle azimuth angle at adjacent moments are equal. When the wind turbine generator is in abnormal yaw, the change rates of the nacelle azimuth angle at adjacent moments are not equal, and at this time it is regarded as yaw jamming.

[0019] Further, step S2 includes the following steps:

[0020] Perform a linear transformation on the original data using linear function normalization:

[0021] The original data includes continuous value variables and categorical variables. The continuous value variables include the maximum value of the nacelle azimuth angle change amount, the mean value of the nacelle azimuth angle change amount, the quartiles of the nacelle azimuth angle change amount, the average wind speed, the wind direction standard deviation, and the type of nacelle azimuth angle change amount. The categorical variables include the yaw direction, the electromagnetic brake, and the grid connection mode.

[0022] Map the continuous value variables to the range of [0, 1] to achieve the equal ratio scaling of each continuous value variable. The normalization formula is as follows:

[0023]

[0024] Among them, X is the value of each continuous value variable, X min is the minimum value in the values of the corresponding continuous value variable, and X max is the maximum value in the values of the corresponding continuous value variable;

[0025] At the same time, encode the yaw direction. Denote the clockwise yaw CW as 0, the counterclockwise yaw CCW as 1, and the non-yaw state as 2.

[0026] Further, step S3 includes the following steps:

[0027] In the preprocessed original data, extract the type of the nacelle azimuth angle change amount during yaw, the mean value of the nacelle azimuth angle change amount during yaw, the standard deviation of the nacelle azimuth angle change amount during yaw, the maximum value of the nacelle azimuth angle change amount during yaw, the quartiles of the nacelle azimuth angle change amount, the yaw direction, the average wind speed, and the wind direction standard deviation as feature values.

[0028] Further, in step S6, the construction of the ROC curve according to the data of the normal yaw and yaw jamming of the wind turbine generator includes the following steps:

[0029] The classification ability of the yaw jamming diagnosis model is evaluated by the area AUC value calculated by the ROC curve. The abscissa of the ROC curve is the false positive rate FPR, and the ordinate is the true positive rate TPR. The formulas for FPR and TPR are as follows:

[0030]

[0031]

[0032] Among them, P is the number of yaw jamming data, N is the number of normal yaw data, TP is the number of yaw jamming data among the P yaw jamming data predicted as yaw jamming by the yaw jamming diagnosis model, and FP is the number of normal yaw data among the N normal yaw data predicted as normal yaw by the yaw jamming diagnosis model.

[0033] Further, the step S7 includes the following steps:

[0034] Evaluate the effect of the yaw jamming diagnosis model according to the test set accuracy and the ROC curve: the test set accuracy is not less than 0.8, and at the same time calculate the area AUC under the ROC curve, which reflects the model classification performance of the yaw jamming diagnosis model measured by the ROC curve. If AUC is not less than 80%, it proves that the model classification performance of the yaw jamming diagnosis model is good, and execute step S8; if the test set accuracy is less than 0.8 and AUC is less than 80%, repeat steps S1 to S6 until the model classification performance of the yaw jamming diagnosis model is good, and execute step S8.

[0035] Further, the step S8 includes the following steps:

[0036] Perform preliminary yaw jamming identification on the yaw jamming diagnosis model, input actual wind turbine data for yaw jamming identification and perform manual verification. If the accuracy of the actual identification result does not reach the preset accuracy, perform manual correction, add the actual wind turbine data to the training set, and retrain the yaw jamming diagnosis model until the training set accuracy, test set accuracy, and actual application identification accuracy of the yaw jamming diagnosis model are approximately the same or the same, then release the yaw jamming diagnosis model and start analysis to warn of the yaw fault problem of the wind turbine.

[0037] The second object of the present invention is achieved by the following technical solution: A wind turbine yaw jamming diagnosis system based on the SVM algorithm, including:

[0038] An original data acquisition module, used to acquire sensor data recording the fault actions of the wind turbine as the original data;

[0039] A data preprocessing module, used to perform normalization preprocessing on the original data;

[0040] A feature extraction module for extracting features from the preprocessed original data;

[0041] A data annotation module that divides the training data into a training set and a test set, and annotates the training set according to manual annotation and weak supervision learning methods;

[0042] A yaw jitter diagnosis model training module for inputting the annotated training set into the SVM algorithm for training to obtain a yaw jitter diagnosis model;

[0043] A model online module for releasing a stable yaw jitter diagnosis model and starting analysis to warn of yaw faults in wind turbines.

[0044] Furthermore, the sensor data is Tracelog data, and each Tracelog data is millisecond-level data within a one-minute time period before and after a wind turbine fault action;

[0045] Screen the obtained Tracelog data and extract the data in the time period recording the yaw jitter action of the wind turbine: Compare the change rate of the nacelle azimuth angle. When the wind turbine is in normal yaw, the change rate of the nacelle azimuth angle is uniform and unchanged, that is, the change rates of the nacelle azimuth angle at adjacent moments are equal, while when the wind turbine is in abnormal yaw, the change rates of the nacelle azimuth angle at adjacent moments are not equal, and at this time it is regarded as yaw jitter.

[0046] Furthermore, the feature extraction module includes the following steps:

[0047] In the preprocessed original data, extract the type of the change amount of the nacelle azimuth angle during yaw, the mean value of the change amount of the nacelle azimuth angle during yaw, the standard deviation of the change amount of the nacelle azimuth angle during yaw, the maximum value of the change amount of the nacelle azimuth angle, the quartile of the change amount of the nacelle azimuth angle, the yaw direction, the average wind speed, and the standard deviation of the wind direction as feature values.

[0048] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0049] 1. Scientific and effective. Compared with traditional data analysis and detection methods, the present invention avoids the tediousness and subjectivity of relying on a large number of manual observations to determine relevant thresholds;

[0050] 2. High accuracy. When encountering outliers or unobserved information based on data analysis methods, misjudgment is likely to occur, while the SVM algorithm can avoid this defect by learning the distribution and rules of the data; Comparing the artificial data with the diagnostic results based on the algorithm and issuing a warning investigation form to the wind farm for actual verification, both show that the SVM algorithm has good yaw jitter diagnosis effect through experiments;

[0051] 3. Reduce costs and increase economic benefits. Through early warning of the yaw system and even the drive train system of wind turbines, preventive maintenance is realized to ensure the healthy and efficient operation of wind turbines, reduce the failure rate, and save costs and improve the power generation performance of wind farms;

[0052] 4. Promote the digital operation and maintenance of wind turbines. Promote the replacement of regular manual inspections with remote monitoring, and contribute to the digital operation and maintenance era based on unattended, minimally attended, and predictive operation and maintenance. Brief Description of the Drawings

[0053] Figure 1 It is a flowchart of the yaw jamming diagnosis method for wind turbines.

[0054] Figure 2 It is a graph of the accuracy rate results of the SVM algorithm of the present invention in the training set and the test set.

[0055] Figure 3 It is the ROC curve graph of the present invention. Detailed Embodiments

[0056] The present invention will be further described below in conjunction with specific embodiments.

[0057] Embodiment 1

[0058] See Figure 1 As shown, the yaw jamming diagnosis method for wind turbines based on the SVM algorithm provided in this embodiment includes the following steps:

[0059] S1. Obtain the sensor data recording the fault actions of the wind turbine as the original data;

[0060] The sensor data is Tracelog data. Each Tracelog data is millisecond-level data within one minute before and after a fault action of the wind turbine. The data volume of each Tracelog file is 12,000; for the data of each obtained Tracelog file, there are two possibilities. One is that there is no yaw action in the file, and the data is generated by the fault actions of other components; the other is the Tracelog file with a yaw action, which may be triggered by the fault actions of other components or by the yaw fault action. Therefore, when identifying whether there is yaw jamming during the yaw action, it is necessary to extract the data of the time period when the yaw action occurs. The identification of the yaw jamming action is mainly to compare the change rate of the nacelle azimuth angle. When the wind turbine is in the yaw action, the change rate of the nacelle azimuth angle during normal yaw is uniform and unchanged, that is, the change rate of the nacelle azimuth angle at adjacent moments is equal, while the change rate of the nacelle azimuth angle at adjacent moments is not equal during abnormal yaw, and at this time it is regarded as yaw jamming.

[0061] S2. Preprocess the original data by normalization, including the following steps:

[0062] When yawing, it may be affected by the vibration of other large components of the wind turbine, resulting in the vibration of the sensors on the wind turbine itself, causing slight oscillations in the data of the relevant components fixed on the nacelle, which are regarded as noise. To avoid misjudgment, filter variables such as the nacelle azimuth angle to eliminate noise;

[0063] After performing noise filtering on the original data, linearly transform the filtered data using linear function normalization. The original data includes continuous value variables and categorical variables. The continuous value variables include the maximum value of the nacelle azimuth angle change, the mean value of the nacelle azimuth angle change, the quartiles of the nacelle azimuth angle change, the average wind speed, the standard deviation of the wind direction, and the type of nacelle azimuth angle change. The categorical variables include the yaw direction, electromagnetic brake, and grid connection mode; Map the continuous value variables to the range of [0,1] to achieve the equal ratio scaling of each continuous value variable. The normalization formula is as follows:

[0064]

[0065] Among them, X is the value of each continuous value variable, X min is the minimum value in the corresponding continuous value variable, X max is the maximum value in the corresponding continuous value variable;

[0066] At the same time, encode the yaw direction, record clockwise yaw CW as 0, counterclockwise yaw CCW as 1, and non-yaw state as 2.

[0067] S3. Extract features from the preprocessed original data to obtain training data, including the following steps:

[0068] In the preprocessed original data, extract the type of nacelle azimuth angle change during yaw, the mean value of nacelle azimuth angle change during yaw, the standard deviation of nacelle azimuth angle change during yaw, the maximum value of nacelle azimuth angle change during yaw, the quartiles of nacelle azimuth angle change, yaw direction, average wind speed, and standard deviation of wind direction as feature values.

[0069] S4. Divide the training data into a training set and a test set, and label the training set according to manual annotation and weak supervision learning methods, including the following steps:

[0070] Divide the training data into a training set and a test set at a ratio of 7:3, and label the training set according to manual annotation and weak supervision learning methods; set the label of the sample cases with yaw jitter in the training set to 1, and the label of the samples with normal yaw to 0. In addition, if it is found that there is a sample imbalance problem between the number of yaw jitter samples and the number of normal yaw samples in the training set, the SMOTE algorithm is used to augment the yaw jitter sample data so that the number of normal samples and the number of fault samples are the same.

[0071] S5. Input the labeled training set into the SVM algorithm for training to obtain a yaw jitter diagnosis model. See Figure 2 As shown, the SVM algorithm performs excellently on the training set;

[0072] S6. Perform cross-validation testing on the yaw jitter diagnosis model according to the test set and the ROC curve, including the following steps:

[0073] See Figure 3 As shown, evaluate the classification ability of the yaw jitter diagnosis model through the area AUC value calculated by the ROC curve. The abscissa of the ROC curve is the false positive rate FPR, and the ordinate is the true positive rate TPR; the formulas for FPR and TPR are as follows:

[0074]

[0075]

[0076] Among them, P is the number of yaw jitter data, N is the number of normal yaw data, TP is the number of yaw jitter data among the P yaw jitter data predicted as yaw jitter by the yaw jitter diagnosis model, and FP is the number of normal yaw data among the N normal yaw data predicted as normal yaw by the yaw jitter diagnosis model;

[0077] S7. Evaluate the effect of the yaw jitter diagnosis model according to the test set accuracy rate and the ROC curve. If the effect of the evaluated yaw jitter diagnosis model meets the preset requirements, execute step S8; if the effect of the evaluated yaw jitter diagnosis model does not meet the preset requirements, repeat steps S1 to S6 until the effect of the evaluated yaw jitter diagnosis model meets the preset requirements and execute step S8, including the following steps:

[0078] Evaluate the effectiveness of the yaw jitter diagnosis model based on the test set accuracy and the ROC curve: the test set accuracy is not less than 0.8. At the same time, calculate the area under the ROC curve, AUC, to reflect the model classification performance of the yaw jitter diagnosis model measured by the ROC curve. If AUC is not less than 80%, it proves that the model classification performance of the yaw jitter diagnosis model is good, and execute step S8; if the test set accuracy is less than 0.8 and AUC is less than 80%, repeat steps S1 to S6 until the model classification performance of the yaw jitter diagnosis model is good, and then execute step S8. By calculating the area under the ROC curve, AUC = 0.973, which reflects the model performance measured by the ROC curve, proving that the yaw jitter diagnosis model in this embodiment has good classification performance.

[0079] S8. Manually verify the accuracy of the actual recognition results of the yaw jitter diagnosis model. If the verification is unqualified, add new data samples to the training set and repeat steps S5 to S7 until the verification is qualified; if the verification is qualified, release the yaw jitter diagnosis model and start the analysis to warn about the yaw faults of the wind turbine, including the following steps:

[0080] When both the model training set accuracy and the test set accuracy are greater than 0.8 and the difference between them is not large, perform a preliminary yaw jitter recognition on the yaw jitter diagnosis model, input the actual wind turbine data for yaw jitter recognition and perform manual verification. If the accuracy of the actual recognition results does not reach 80%, perform manual correction, add the actual wind turbine data to the training set, and retrain the yaw jitter diagnosis model until the training set accuracy, test set accuracy, and actual application recognition accuracy of the yaw jitter diagnosis model are approximately the same or the same, then release the yaw jitter diagnosis model and start the analysis to warn about the yaw faults of the wind turbine.

[0081] Take the Tracelog file data of a wind farm's turbines in the recent month as the practical data, and use the SVM model to identify the yaw jitter problems of the actual turbines. First, obtain the Tracelog file path corresponding to each wind turbine in the wind farm in the database, and at the same time set the analysis period for the program. Here, the data in June 2022 is used as the analysis period. To better illustrate the accuracy and timeliness of the SVM algorithm, this embodiment also performs fault detection of yaw jitter based on data analysis and compares the prediction accuracies of the two. The results are shown in Table 1.

[0082] Table 1

[0083]

[0084] As can be seen from Table 1, yaw jamming occurs frequently in units 006#, 013#, and 022#. Therefore, manual data verification was carried out on these three units to check for misjudgment. The verification results show that most of the yaw jamming detected based on the SVM model is correctly judged, while the diagnosis of yaw jamming based on the data analysis method may be misjudged due to limited observed data features.

[0085] Subsequently, a warning work order was issued to the site, and the wind farm operation and maintenance personnel inspected the status of each component of the yaw system. The on-site feedback found that the yaw brake of unit 013# was severely worn, and the yaw brake pressure parameters of units 006# and 022# were set incorrectly. The operation and maintenance personnel carried out targeted treatment, which timely avoided more serious damage to the components, not only reduced the failure rate and the committed pressure, but also effectively avoided the power generation loss caused by the unit's shutdown due to failure. In summary, the reliability of identifying yaw jamming problems based on the SVM algorithm is relatively strong.

[0086] Embodiment 2

[0087] This embodiment discloses a yaw jamming diagnosis system for a wind turbine based on the SVM algorithm, including:

[0088] An original data acquisition module, used to acquire sensor data recording the fault actions of the wind turbine as the original data, including the following steps: The sensor data is Tracelog data, and each Tracelog data is millisecond-level data within a one-minute time period before and after a wind turbine fault action;

[0089] Screen the obtained Tracelog data and extract the data of the time period recording the yaw jamming action of the wind turbine: Compare the change rate of the nacelle azimuth angle. When the wind turbine is in normal yaw, the change rate of the nacelle azimuth angle is uniform and unchanged, that is, the change rates of the nacelle azimuth angle at adjacent times are equal, while when the wind turbine is in abnormal yaw, the change rates of the nacelle azimuth angle at adjacent times are not equal, and at this time it is regarded as yaw jamming.

[0090] A data preprocessing module, used to perform normalization preprocessing on the original data, including the following steps:

[0091] Perform a linear transformation on the original data using linear function normalization:

[0092] The original data includes continuous value variables and categorical variables. The continuous value variables include the maximum value of the nacelle azimuth angle change, the mean value of the nacelle azimuth angle change, the quartiles of the nacelle azimuth angle change, the average wind speed, the wind direction standard deviation, and the type of nacelle azimuth angle change. The categorical variables include the yaw direction, electromagnetic brake, and grid connection mode;

[0093] Map the continuous value variables to the range of [0, 1] to achieve the equal ratio scaling of each continuous value variable. The normalization formula is as follows:

[0094]

[0095] where X is the value of each continuous value variable, X min is the minimum value in the values of the corresponding continuous value variable, and X max is the maximum value in the values of the corresponding continuous value variable.

[0096] The feature extraction module is used to extract features from the preprocessed original data, including the following steps:

[0097] In the preprocessed original data, extract the type of the yaw nacelle azimuth change amount, the mean value of the yaw nacelle azimuth change amount, the standard deviation of the yaw nacelle azimuth change amount, the maximum value of the yaw nacelle azimuth change amount, the quartile of the nacelle azimuth change amount, the yaw direction, the average wind speed, and the wind direction standard deviation as feature values;

[0098] The data annotation module divides the training data into a training set and a test set, and annotates the training set according to manual annotation and weak supervision learning methods;

[0099] The yaw jamming diagnosis model training module is used to input the annotated training set into the SVM algorithm for training to obtain a yaw jamming diagnosis model;

[0100] The model online module is used to release a stable yaw jamming diagnosis model and start the analysis to warn of the yaw fault problem of the wind turbine; among them, when the yaw jamming diagnosis model reaches the preset accuracy rate and is verified by actual unit data, and the accuracy rate of the actual unit data is approximately the same as that of the training set, the yaw jamming diagnosis model is considered stable.

[0101] The above-described embodiments are only the preferred embodiments of the present invention, and do not limit the scope of implementation of the present invention. Therefore, all changes made according to the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A yaw jamming diagnosis method for wind turbines based on the SVM algorithm, characterized in that, Including the following steps: S1. Obtain the sensor data recording the fault actions of the wind turbine as the original data; the sensor data is Tracelog data, and each Tracelog data is millisecond-level data within one minute before and after a fault action of the wind turbine; screen the obtained Tracelog data to extract the data of the time period recording the yaw jamming action of the wind turbine: compare the change rate of the nacelle azimuth angle. When the wind turbine is in normal yaw, the change rate of the nacelle azimuth angle is uniform and unchanged, that is, the change rates of the nacelle azimuth angle at adjacent moments are equal. When the wind turbine is in abnormal yaw, the change rates of the nacelle azimuth angle at adjacent moments are not equal, and at this time it is regarded as yaw jamming; S2. Perform preprocessing of normalizing the original data; S3. Extract features from the preprocessed original data to obtain model data; S4. Perform upsampling processing on the model data, then divide it into a training set and a test set, and label the training set according to the manual annotation and weak supervision learning method; S5. Input the labeled training set into the SVM algorithm for training to obtain a yaw jamming diagnosis model; S6. Verify and calculate the accuracy of the test set, and construct an ROC curve based on the data of normal yaw and yaw jamming of the wind turbine; S7. Evaluate the effect of the yaw jamming diagnosis model according to the accuracy of the test set and the ROC curve. If the effect of the evaluated yaw jamming diagnosis model meets the preset requirements, execute step S8; if the effect of the evaluated yaw jamming diagnosis model does not meet the preset requirements, repeat steps S1 to S6 until the effect of the evaluated yaw jamming diagnosis model meets the preset requirements, and then execute step S8; S8. Manually verify the accuracy of the actual recognition result of the yaw jamming diagnosis model. If the verification is unqualified, add new data samples to the training set, and repeat steps S5 to S7 until the verification is qualified; If the verification is qualified, release the yaw jamming diagnosis model and start the analysis to warn of the yaw fault problem of the wind turbine.

2. The yaw jamming diagnosis method for a wind turbine based on the SVM algorithm according to claim 1, wherein The step S2 includes the following steps: Perform linear transformation on the original data using linear function normalization; The original data includes continuous value variables and categorical variables. The continuous value variables include the maximum value of the nacelle azimuth angle change amount, the mean value of the nacelle azimuth angle change amount, the quartiles of the nacelle azimuth angle change amount, the average wind speed, the wind direction standard deviation, and the type of the nacelle azimuth angle change amount. The categorical variables include the yaw direction, electromagnetic brake, and grid connection mode; Map the continuous value variables to the range of [0, 1] to achieve the equal ratio scaling of each continuous value variable; At the same time, encode the yaw direction, record the clockwise yaw CW as 0, the counterclockwise yaw CCW as 1, and the non-yaw state as 2.

3. The yaw jamming diagnosis method for wind turbine based on SVM algorithm according to claim 1, characterized in that The step S3 includes the following steps: In the preprocessed original data, extract the type of the nacelle azimuth angle change amount during yaw, the mean value of the nacelle azimuth angle change amount during yaw, the standard deviation of the nacelle azimuth angle change amount during yaw, the maximum value of the nacelle azimuth angle change amount during yaw, the quartiles of the nacelle azimuth angle change amount, the yaw direction, the average wind speed, and the wind direction standard deviation as feature values.

4. A yaw jamming diagnosis method for wind turbines based on the SVM algorithm according to claim 1, characterized in that, In step S6, the construction of the ROC curve based on the data of normal yaw and yaw jamming of the wind turbine generator set includes the following steps: Evaluate the classification ability of the yaw jamming diagnosis model through the area AUC value calculated by the ROC curve. The abscissa of the ROC curve is the false positive rate FPR, and the ordinate is the true positive rate TPR. The formulas for FPR and TPR are as follows: ； ； Where P is the number of yaw jamming data, N is the number of normal yaw data, TP is the number of yaw jamming data among the P yaw jamming data predicted as yaw jamming by the yaw jamming diagnosis model, and FP is the number of normal yaw data among the N normal yaw data predicted as normal yaw by the yaw jamming diagnosis model.

5. A yaw jamming diagnosis method for wind turbines based on the SVM algorithm according to claim 1, characterized in that The step S7 includes the following steps: Evaluate the effect of the yaw jamming diagnosis model according to the test set accuracy and the ROC curve: the test set accuracy is not less than 0.8, and at the same time calculate the area size AUC under the ROC curve, which reflects the model classification performance of the yaw jamming diagnosis model measured by the ROC curve. If the AUC is not less than 80%, it proves that the model classification performance of the yaw jamming diagnosis model is good, and execute step S8; if the test set accuracy is less than 0.8 and the AUC is less than 80%, repeat steps S1 to S6 until the model classification performance of the yaw jamming diagnosis model is good, and execute step S8.

6. A yaw jamming diagnosis method for a wind turbine based on the SVM algorithm according to claim 1, characterized in that, The step S8 includes the following steps: Perform preliminary yaw jamming identification on the yaw jamming diagnosis model, input the actual wind turbine generator set data for yaw jamming identification and perform manual verification. If the accuracy of the actual identification result does not reach the preset accuracy, perform manual correction, add the actual wind turbine generator set data to the training set, and retrain the yaw jamming diagnosis model until the training set accuracy, test set accuracy, and actual application identification accuracy of the yaw jamming diagnosis model are approximately the same or the same, then release the yaw jamming diagnosis model and start the analysis to warn of the yaw fault problem of the wind turbine generator set.

7. A yaw jamming diagnosis system for wind turbines based on the SVM algorithm, characterized in that, Include: An original data acquisition module, which is used to acquire the sensor data recording the fault actions of the wind turbine generator set as the original data; the sensor data is Tracelog data, and each Tracelog data is millisecond-level data within one minute before and after a wind turbine generator set fault action; screen the obtained Tracelog data, and extract the data of the time period recording the yaw jamming action of the wind turbine generator set: compare the change rate of the nacelle azimuth angle. When the wind turbine generator set is in normal yaw, the change rate of the nacelle azimuth angle is uniform and unchanged, that is, the change rates of the nacelle azimuth angle at adjacent moments are equal, while when the wind turbine generator set is in abnormal yaw, the change rates of the nacelle azimuth angle at adjacent moments are not equal, and at this time it is regarded as yaw jamming; A data preprocessing module, which is used to perform normalization preprocessing on the original data; A feature extraction module, which is used to extract features from the preprocessed original data; A data annotation module, which divides the training data into a training set and a test set, and annotates the training set according to manual annotation and weak supervision learning methods; A yaw jamming diagnosis model training module, which is used to input the annotated training set into the SVM algorithm for training to obtain a yaw jamming diagnosis model; The model online module is used to release a stable yaw jitter diagnosis model and start the analysis to warn of yaw faults in wind turbines.

8. The yaw jamming diagnosis system for a wind turbine based on the SVM algorithm according to claim 7, characterized in that, The feature extraction module includes the following steps: In the preprocessed original data, extract the type of the change in the nacelle azimuth angle during yaw, the mean value of the change in the nacelle azimuth angle during yaw, the standard deviation of the change in the nacelle azimuth angle during yaw, the maximum value of the change in the nacelle azimuth angle during yaw, the quartile of the change in the nacelle azimuth angle, the yaw direction, the average wind speed, and the standard deviation of the wind direction as feature values.

Images