Wind turbine yawing reducer anomaly detection method and device

Abstract

The application provides a wind turbine yaw reducer anomaly detection method and device, comprising: collecting historical data of a target wind turbine and healthy wind turbines of the same type, and after invalid data is removed through preprocessing, jointly dividing complete yaw events through a yaw state flag and a yaw angle; extracting statistical features and mechanism features of each event to construct a feature matrix, and training an optimal double model of a single-class support vector machine and an isolation forest; repeatedly processing real-time data of the target wind turbine as described above, inputting the double model to obtain an anomaly score, and determining an anomaly after normalization and fusion, which can accurately identify early progressive failure without fault samples and at low cost, thereby improving the operation reliability of the wind turbine.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to a method and apparatus for detecting abnormalities in the yaw reducer of a wind turbine. Background Technology

[0002] The yaw reducer is the core component of the yaw system of a wind turbine. It is subjected to complex alternating loads for a long time. The failure of its gears and bearings develops gradually. If there is no timely warning, it may lead to catastrophic consequences.

[0003] Currently, condition monitoring of yaw gearboxes mainly relies on costly and demanding vibration analysis systems (CMS), making widespread adoption difficult. While the SCADA systems commonly equipped in wind turbines collect massive amounts of process parameters, their low sampling frequency has traditionally limited their use for accurate fault diagnosis. Existing SCADA data-based analysis methods have significant shortcomings: they ignore differences in operating conditions, leading to the masking of fault characteristics; the extracted statistical features are disconnected from the physical failure mechanisms; and they rely on difficult-to-obtain historical fault samples, limiting model application. Therefore, a low-cost, wide-coverage anomaly detection method capable of early warning is urgently needed. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for detecting abnormalities in the yaw reducer of a wind turbine, so as to alleviate the technical problem of lacking low-cost, wide-coverage, and accurate abnormality detection of progressive early faults in the yaw reducer in the prior art.

[0005] In a first aspect, the present invention provides a method for detecting abnormalities in the yaw reducer of a wind turbine, comprising: The historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine are preprocessed, and at least one yaw event is identified by combining the yaw status flag and yaw angle in the preprocessed data. The optimal dual model is obtained by constructing a feature matrix based on the statistical and mechanistic characteristics of each yaw event; wherein the optimal dual model is a single-class support vector machine target model and an isolated forest target model. Real-time data of the target wind turbine is acquired and anomaly detection is performed using the optimal dual model. The anomaly scores are then fused to determine whether the yaw reducer of the target wind turbine is abnormal.

[0006] In an optional implementation, the step of preprocessing historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine generator set, and jointly identifying at least one yaw event by combining the yaw status flag and yaw angle in the preprocessed data, includes: Historical data of the target wind turbine and healthy wind turbines of the same model as the target wind turbine were collected from the wind turbine generator set. The historical data is preprocessed to obtain preprocessed data with invalid data removed; By combining the state switching of the yaw status flag bit and the numerical change of the yaw angle in the preprocessed data, at least one yaw event to which the preprocessed data belongs is determined.

[0007] In an optional implementation, the historical data includes wind turbine operating parameters, yaw system parameters, yaw state parameters, and environmental parameters; the historical data collection period is no less than one year.

[0008] In an optional implementation, the step of preprocessing the historical data to obtain preprocessed data after removing invalid data includes: Invalid data from the historical data under shutdown, power limitation, and fault conditions are processed by removing abnormal values ​​such as unchanged or fluctuating sensor values ​​to obtain preprocessed data.

[0009] In an optional implementation, the step of training a dual model based on the feature matrix constructed from the statistical and mechanistic features of each yaw event includes: Based on the preprocessed data corresponding to each yaw event, relevant statistical and mechanistic features are extracted, and a feature matrix is ​​constructed. The feature matrix is ​​standardized, and the hyperparameters of the single-class support vector machine model and the isolated forest model are optimized by grid search method respectively. The optimal dual model is trained and the target hyperparameters are saved.

[0010] In an optional implementation, the statistical features include at least the mean, variance, and skewness of key variables within the yaw event; the mechanistic features include at least the yaw efficiency index, yaw stability index, and wind alignment accuracy index used to map the physical failure mechanism of the yaw reducer; the statistical features and mechanistic features of each yaw event constitute a row in the feature matrix.

[0011] In an optional implementation, the step of acquiring real-time data of the target wind turbine and performing anomaly detection using the optimal dual model, and fusing anomaly scores to determine whether the yaw reducer of the target wind turbine is abnormal, includes: The real-time data of the target wind turbine is preprocessed and segmented, and the feature matrix of each yaw event is extracted. The feature matrix is ​​standardized and then input into the optimal dual model to obtain the corresponding first anomaly score and second anomaly score. After normalizing the first and second abnormal scores using the quantile mapping method, the comprehensive abnormal score is obtained by fusing them according to a preset weight. Based on the comprehensive anomaly score and the preset threshold, the abnormal state of the yaw reducer of the target wind turbine is determined.

[0012] Secondly, the present invention provides a wind turbine yaw reducer anomaly detection device, comprising: The preprocessing module preprocesses the historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine unit, and identifies at least one yaw event by combining the yaw status flag and yaw angle in the preprocessed data. The training module, based on the feature matrix constructed from the statistical and mechanistic features of each yaw event, trains the optimal dual model to obtain the optimal dual model; wherein, the optimal dual model is a single-class support vector machine target model and an isolated forest target model; The detection module acquires real-time data of the target wind turbine and performs anomaly detection using the optimal dual model. It then fuses the anomaly scores to determine whether the yaw reducer of the target wind turbine is abnormal.

[0013] Thirdly, the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any of the foregoing embodiments.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the method described in any of the foregoing embodiments.

[0015] This invention provides a method and device for detecting anomalies in the yaw reducer of a wind turbine. First, historical data of the target wind turbine and healthy wind turbines of the same model are collected. After preprocessing by removing invalid data and processing sensor anomalies, complete yaw events are jointly segmented by the state switching of the yaw status flag and the numerical change of the yaw angle. Then, statistical features (such as the mean and variance of key variables) and mechanistic features (such as yaw efficiency indicators reflecting transmission friction resistance and wind alignment accuracy indicators characterizing positioning accuracy) are extracted for each yaw event to construct a structured feature matrix. Based on healthy data, a single-class support vector machine and an optimal dual model of isolated forest are trained using hyperparameter optimization. Finally, the real-time data of the target wind turbine is repeatedly preprocessed, segmented, and feature extracted. The standardized feature matrix is ​​input into the dual model to obtain anomaly scores, which are then normalized using quantile mapping and fused according to preset weights. Anomaly states are determined by combining these scores with a threshold. These technologies work together to fully utilize existing SCADA data from wind turbines, avoiding the high cost and limited availability of traditional vibration analysis systems. By segmenting operating conditions and dividing events, they solve the problem of fault feature overwhelming caused by the mixing of data from different operating conditions. By relying on the fusion of statistical and mechanistic features, they break through the bottleneck of the disconnect between traditional features and physical failure mechanisms. By using unsupervised dual-model training and score fusion, they get rid of the dependence on scarce fault samples. Ultimately, they achieve low-cost, wide-coverage, and highly accurate yaw reducer anomaly detection, which can identify progressive minor failures of gears and bearings at an early stage, effectively avoiding catastrophic consequences such as broken teeth and jamming, and improving the operational reliability and maintenance economy of wind turbines.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart of an abnormal detection method for a wind turbine yaw reducer provided in an embodiment of the present invention; Figure 2An abnormal score diagram provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the functional modules of a wind turbine yaw reducer anomaly detection device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware architecture of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Currently, condition monitoring of yaw gearboxes mainly relies on vibration analysis systems. However, CMS systems are expensive and have high installation requirements, making them difficult to implement on all wind turbine units. SCADA systems, commonly found on wind turbines, collect massive amounts of process parameters (such as current, temperature, and speed), but their sampling frequency is low (on the order of minutes), making it impossible to directly capture high-frequency impact signals from mechanical components. Therefore, they have traditionally been considered unsuitable for accurate mechanical fault diagnosis.

[0022] Existing SCADA data analysis methods mostly suffer from the following shortcomings: (1) Ignoring the influence of operating conditions, failing to perform detailed operating condition division of data, and mixing and analyzing data under different wind speeds and different yaw angles, resulting in fault characteristic signals being submerged by huge operating condition fluctuations, and extremely low signal-to-noise ratio.

[0023] (2) The features are disconnected from the mechanism. Most of the extracted features are statistical features and fail to establish a direct and clear mapping relationship with the physical failure mechanism of the yaw reducer (such as increased friction, decreased efficiency, and reduced positioning accuracy).

[0024] (3) Relying on historical fault data, most intelligent diagnostic models rely on supervised learning and require a large number of labeled fault samples for training, which is often difficult to obtain in engineering practice, thus limiting the promotion and application of the model.

[0025] Therefore, developing a yaw gearbox fault diagnosis method that can overcome the above-mentioned shortcomings, make full use of existing SCADA data, and achieve low cost, wide coverage, and early warning is of significant engineering value and economic benefit.

[0026] Based on this, the present invention provides a method and device for detecting yaw gearbox anomalies in wind turbines. By preprocessing SCADA data and classifying yaw events, fusing statistical features and mechanistic features, training unsupervised dual models and fusing scores, the method achieves low-cost, wide-coverage, and fault-sample-independent yaw gearbox anomaly detection. It accurately identifies progressive early failures of gears and bearings, improving monitoring reliability and early warning timeliness.

[0027] To facilitate understanding of this embodiment, a detailed description of the abnormal detection method for the yaw reducer of a wind turbine generator disclosed in this embodiment of the invention will be provided first.

[0028] Figure 1 A flowchart of an abnormal detection method for a wind turbine yaw reducer provided in an embodiment of the present invention.

[0029] Reference Figure 1 The method may include the following steps: S102, preprocess the historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine, and identify at least one yaw event by combining the yaw status flag and yaw angle in the preprocessed data. In some embodiments, step S102 may be implemented by the following steps: Step 1.1: Collect historical data of the target wind turbine and healthy wind turbines of the same model as the target wind turbine from the wind turbine unit.

[0030] Historical data includes wind turbine operating parameters, yaw system parameters, yaw state parameters, and environmental parameters; the data collection period is no less than one year. Historical data is a multi-dimensional parameter that covers all relevant factors in yaw reducer operation, ensuring rich feature extraction; a sufficiently long collection period covers different operating conditions (such as different wind speeds and seasons), providing data support for establishing a robust health benchmark model and avoiding insufficient model generalization ability due to incomplete data.

[0031] For example, first, the parameter types of historical data are determined. Wind turbine operating parameters include generator speed, output power, pitch angle, etc., which directly reflect the overall operating status of the unit; yaw system parameters include yaw motor current, yaw status flag, yaw angle, yaw speed, yaw error angle, etc., focusing on core yaw action data; yaw status parameters include yaw reducer lubricating oil temperature, nacelle lateral and longitudinal vibration data, etc., which are directly related to the reducer operating status; environmental parameters include wind speed, wind direction, ambient temperature, nacelle temperature, etc., eliminating the influence of environmental interference factors.

[0032] Secondly, ensure sufficient data collection time. Collect historical data for at least one year to ensure that the data covers different climatic conditions in spring, summer, autumn and winter, as well as different operating conditions such as low wind speed, medium wind speed and high wind speed. It also includes various working states such as normal start-up and shutdown of wind turbines and continuous operation, so that the collected data can comprehensively reflect the operating rules of yaw reducers in a healthy state and provide sufficient and diverse samples for subsequent model training.

[0033] Step 1.2: Perform preprocessing operations on the historical data to obtain preprocessed data after removing invalid data.

[0034] Here, this preprocessing operation can be understood as removing invalid data from historical data in shutdown, power-limited, and fault states, as well as outliers such as unchanged or fluctuating sensor values, to obtain preprocessed data. By accurately filtering invalid data unrelated to the normal operation of the yaw reducer, interference caused by sensor anomalies is eliminated, ensuring that the preprocessed data are all valid operational data, thus improving the accuracy of subsequent feature extraction and model training.

[0035] For example, firstly, invalid data is removed. The generator speed is used to determine shutdown and fault status. A speed threshold is set; when the speed is below this threshold, the unit is determined to be in a shutdown or fault state, and historical data for the corresponding time period is removed. For instance, generator speed can be used for shutdown and fault screening, removing data with generator speeds less than 1000 rpm. The power limit flag in the SCADA system is used to identify periods when the unit is operating at limited power, and data from those periods is removed to avoid interference from power-limited operating conditions on yaw parameters.

[0036] Secondly, abnormal sensor values ​​are processed and removed. The standard deviation of each sensor data within a set time window is calculated. If the standard deviation is less than the preset minimum value, it indicates that the sensor value is basically unchanged, and the sensor is judged to be faulty, and the abnormal data corresponding to that sensor is removed. For key parameters such as temperature and current, a reasonable effective value range is preset. When the data exceeds this range, it is judged as abnormal jump data and is removed. For example, the temperature data range is [-50, 80]. Data outside this range is abnormal data, ensuring that the data after removing abnormal data are all valid data collected by the sensor normally.

[0037] Step 1.3: By combining the state switching of the yaw status flag bit in the preprocessed data with the numerical change of the yaw angle, determine at least one yaw event to which the preprocessed data belongs.

[0038] The yaw status flag and yaw angle are from the preprocessed data obtained in the preceding steps. The preprocessed data can be jointly divided into at least one yaw event using the yaw status flag and yaw angle. This embodiment of the invention focuses on the complete yaw process through precise event segmentation, providing a high-quality data foundation for subsequent feature extraction and avoiding interference from invalid data and incomplete events in model training.

[0039] This invention divides the preprocessed data into events to obtain each complete yaw process event. As an optional embodiment, two key variables—yaw status flag and yaw angle (cabin position)—can be used for joint judgment. The yaw status flag indicates whether the crew is yawing, and the yaw angle is the cabin position after the crew performs the yaw action. When the crew begins to yaw, the yaw status flag flips from 0 to 1, at which point the yaw angle starts moving from its original position. After yaw ends, the yaw status flag flips from 1 to 0, at which point the yaw angle stops at its final yaw position. Each complete yaw event is extracted by capturing the state changes of the yaw status flag and the numerical changes of the yaw angle.

[0040] For example, using the yaw status flag and yaw angle in the preprocessed data as the core judgment criteria, when the yaw status flag changes from 0 to 1, the initial value of the yaw angle is recorded, and the yaw action is determined to begin; when the flag changes from 1 to 0, the final value of the yaw angle is recorded, and the yaw action is determined to end; the dataset between two state transitions is defined as a complete yaw event, and all independent yaw events are divided in chronological order.

[0041] S104, based on the statistical and mechanistic features of each yaw event, a feature matrix is ​​constructed, and the optimal dual model is trained to obtain the optimal dual model; among which, the optimal dual model is a single-class support vector machine target model and an isolated forest target model; In some embodiments, step S104 may be implemented by the following steps: Step 2.1: Extract the corresponding statistical and mechanistic features from the preprocessed data corresponding to each yaw event, and construct a feature matrix; The statistical features include at least the mean, variance, and skewness of key variables within the yaw event; the mechanistic features include at least the yaw efficiency index, yaw stability index, and wind accuracy index used to map the physical failure mechanism of the yaw reducer; the statistical features and mechanistic features of each yaw event constitute a row in the feature matrix, in other words, each row of the feature matrix corresponds to a yaw event.

[0042] Here, statistical features are used to capture the inherent patterns in the data, while mechanistic features are directly related to the physical failure mechanism. The combination of the two enhances the feature recognition rate. The standardized feature matrix structure ensures the orderliness of model training, enabling the model to accurately learn the feature distribution under healthy conditions.

[0043] For example, firstly, for each yaw event, key variables such as wind speed, yaw motor current, yaw angle, and yaw reducer lubricating oil temperature are selected. The mean (reflecting the average level of the variable), variance (reflecting the degree of fluctuation of the variable), and skewness (reflecting the symmetry of the variable distribution) of each key variable during the yaw event period are calculated to form a statistical feature set for the event, ensuring that the statistical features can capture the basic patterns of the data.

[0044] Secondly, core mechanism features are extracted. Based on the physical failure mechanism of the yaw reducer, yaw efficiency index (reflecting the frictional resistance of the transmission system), yaw stability index (reflecting the transmission impact), and wind alignment accuracy index (reflecting the transmission positioning accuracy) are calculated. Each index directly maps to the failure evolution process such as gear wear and bearing failure, forming a set of mechanism features, thus achieving a direct link between features and failure mechanisms.

[0045] The specific mechanistic characteristics can be calculated using the following formula: Yaw efficiency index The calculation formula is:

[0046] in, It is the average value of the yaw motor current during that event. This is the total angle of this yaw. This refers to the absolute value of the difference between the yaw termination angle and the yaw initiation angle within the specified event. This indicator directly reflects the overall frictional resistance of the transmission system. When gears or bearings wear, transmission efficiency decreases, requiring greater driving force to complete the same yaw angle. It will show an upward trend.

[0047] Yaw stability index That is, the coefficient of variation of the current, which is calculated using the following formula:

[0048] in, , This represents the average and standard deviation of the yaw motor current during the event. This indicator is exceptionally sensitive to intermittent shocks during the transmission process. When pitting or broken teeth occur in the gears, a torque shock will be generated when rotating to a specific position, manifesting as increased fluctuations in the current signal, thus leading to... Increase.

[0049] The formula for calculating the wind accuracy index F_acc is:

[0050] in, This refers to the yaw error angle over a stable period (e.g., 30 seconds) after the yaw maneuver has ended. This indicator reflects the backlash and overall rigidity of the transmission system. As gear wear intensifies, transmission clearance increases, leading to a decrease in the final positioning accuracy of the cabin. It will increase.

[0051] Thermal index The calculation formula is:

[0052] in, , This refers to the temperature of the yaw reducer lubricating oil relative to the ambient temperature during the event. This indicator eliminates the influence of ambient temperature and directly reflects the internal heat generation level of the reducer. An overall increase in internal friction losses (including gear meshing losses and bearing friction losses) directly leads to an increase in oil temperature. It then rose.

[0053] In addition, the statistical feature set and the mechanistic feature set of each yaw event can be concatenated and integrated in a fixed order, with statistical features first and mechanistic features later, to form a row of data in the feature matrix; all yaw events are processed in sequence, and all rows of data are arranged in chronological order to construct a structured feature matrix, ensuring that each row of the matrix corresponds to a complete yaw event and each column corresponds to a fixed feature.

[0054] Step 2.2: Standardize the feature matrix, and use the grid search method to optimize the hyperparameters of the single-class support vector machine model and the isolated forest model respectively. Train the optimal dual model and save the target hyperparameters.

[0055] The feature matrix in this embodiment of the invention takes into account both statistical regularities and physical mechanisms. Standardization eliminates the influence of dimensions, hyperparameters can optimize and improve model performance, and optimal dual-model training provides an accurate health benchmark for subsequent anomaly detection, ensuring that the model can effectively identify early fault characteristics. Based on the feature matrix corresponding to historical data collected under known health conditions, an optimal single-class support vector machine model and an optimal isolated forest model are trained respectively, and the model parameters are saved.

[0056] For example, the Z-score standardization method can be used to process the feature matrix to eliminate the influence of differences in the dimensions and orders of magnitude of different features, so that the mean of each feature is 0 and the variance is 1, ensuring that the weights of each feature are balanced during model training, thereby improving training efficiency and model accuracy.

[0057] The Z-score standardized calculation formula is as follows:

[0058] Hyperparameter optimization is performed on a One-Class SVM model. The selection of hyperparameters mainly includes the kernel function and boundary parameters. The candidate set of kernel functions includes radial basis function (RBF), linear kernel function, and polynomial kernel function; the search range for boundary parameters is set from 0.1 to 0.7. Using the normalized health feature matrix as input, a grid search method is employed to optimize the hyperparameters. After multiple iterations, the final One-Class SVM target model is obtained and saved.

[0059] Simultaneously, hyperparameter optimization was performed on the Isolation Forest model. The selection of hyperparameters mainly included the number of trees (n_estimators), the number of samples per tree (max_samples), and the proportion of outliers in the dataset (contamination). The search range for the number of trees (n_estimators) was set to [100, 500]; the search range for the number of samples per tree (max_samples) was set to [0.2, 0.8]; and the search range for the proportion of outliers in the dataset (contamination) was set to [0.05, 0.2]. Using the normalized health feature matrix as input, a grid search method was used to optimize the hyperparameters. After multiple iterations, the final Isolation Forest target model was obtained and saved.

[0060] Furthermore, for the single-class support vector machine model, a candidate set of kernel functions (including radial basis function kernel function, linear kernel function, and polynomial kernel function) and a search range for boundary parameters are defined. Using the standardized feature matrix as input, different hyperparameter combinations are iteratively tested through grid search, and the combination with the best performance is selected to train the single-class support vector machine target model. For the isolated forest model, a search range for the number of trees, the number of samples per tree, and the proportion of outliers is defined. The same grid search method is used to optimize the hyperparameters, and the isolated forest target model is trained. The target hyperparameters of both models are saved for subsequent use.

[0061] S106: Obtain real-time data of the target wind turbine and perform anomaly detection using the optimal dual model. Combine the anomaly scores to determine whether the yaw reducer of the target wind turbine is abnormal.

[0062] In some embodiments, step S106 achieves abnormal detection of the yaw reducer through the following steps: Step 3.1: Perform preprocessing and segmentation operations on the real-time data of the target wind turbine, and extract the feature matrix of each yaw event; Feature extraction is performed on the real-time collected yaw system operation data of the unit, using the same extraction method as for historical data. Here, the preprocessing steps, yaw event segmentation steps, and feature matrix extraction steps from the aforementioned embodiments are applied to the real-time data in the same way, and will not be repeated here.

[0063] Step 3.2: Standardize the feature matrix and input it into the optimal dual model to obtain the corresponding first anomaly score and second anomaly score; Then, the normalization method used in the training phase is applied to normalize the feature matrix corresponding to the real-time data to be detected to the same standard, so that the normalization result of the data to be detected is consistent with that of the training set. Specifically, the normalized feature matrix corresponding to the real-time data is input into the optimal dual model trained in the aforementioned embodiment steps to obtain two anomaly scores, that is, the normalized Xtest data is input into OCSVM respectively. final model and IF In the final model, anomaly detection results are obtained; Step 3.3: After normalizing the first and second anomaly scores using the quantile mapping method, the scores are then fused according to preset weights to obtain the comprehensive anomaly score. The anomaly detection results are then fused. Since the ranges and distributions of the first and second anomaly scores output by the two anomaly detection models differ, they must be standardized to the same dimension before result fusion. This embodiment of the invention uses a quantile mapping method based on offline training data scores for normalization. The calculation formula is as follows:

[0064] in, To determine the anomaly score after inputting healthy samples into the model. and The 5th and 95th percentiles of the score.

[0065] Therefore, OCSVM needs to be calculated separately. final model, IF The final model calculates the scores of all samples on the healthy training set, and then calculates the 5th and 95th percentiles. During online detection, the sample to be detected is input into the model to obtain anomaly scores, which are then normalized using the above formula. , The closer the normalized value is to 1, the more outlier the data point is.

[0066] The two normalized outlier scores are then fused to obtain the final fused outlier score. .

[0067]

[0068] in, The assigned weights can be adjusted based on the actual situation. The final output is based on the comprehensive anomaly score.

[0069] Step 3.4: Based on the comprehensive anomaly score and the preset threshold, determine the abnormal state of the yaw reducer of the target wind turbine.

[0070] Finally, based on the comprehensive anomaly score, it can be determined whether the yaw reducer corresponding to the current data is in an abnormal state, specifically as follows: Figure 2 As shown, reasonable thresholds can be set based on specific maintenance experience to classify early warnings and establish different maintenance strategies.

[0071] The embodiments of the present invention ensure data consistency by making the real-time data processing flow consistent with the training phase; avoid single-model bias by fusing dual-model scores, resolve dimensional differences by normalization processing, and achieve quantitative identification of abnormal states by threshold determination, thus ensuring the accuracy and reliability of real-time detection.

[0072] For example, firstly, real-time data is processed and a feature matrix is ​​extracted. Real-time operating data of the target wind turbine is acquired, and invalid data such as shutdowns and power limitations, as well as abnormal sensor values, are removed according to the preprocessing standards of historical data. Real-time yaw events are jointly segmented by yaw status flags and yaw angles. For each real-time event, statistical and mechanistic features consistent with those in the training phase are extracted to construct a real-time feature matrix.

[0073] Secondly, the feature matrix is ​​standardized and the original anomaly scores are obtained. The Z-score standardization method used in the training phase is adopted to standardize the real-time feature matrix to ensure that it is consistent with the standardization standard of the training data. The standardized real-time feature matrix is ​​then input into the single-class support vector machine target model and the isolated forest target model respectively to obtain the corresponding first anomaly score (single-class support vector machine output) and second anomaly score (isolated forest output).

[0074] In addition, the scores are normalized and fused to determine anomalies. The 5% and 95% quantiles of the anomaly scores of the two models on the healthy training set are calculated. The first and second anomaly scores are normalized to the same dimension using the quantile mapping method. The closer the normalized value is to 1, the more abnormal it is. The two normalized scores are weighted and summed according to preset weights (which can be adjusted according to on-site operation and maintenance experience) to obtain a comprehensive anomaly score. Multi-level warning thresholds are preset. The comprehensive anomaly score is compared with the thresholds to determine whether the yaw reducer is in a normal, slightly abnormal, or severely abnormal state, and the corresponding warning is triggered.

[0075] In a preferred embodiment for practical application, historical data of the target wind turbine and healthy wind turbines of the same model are collected. After preprocessing, yaw events are jointly classified by yaw status flags and yaw angles. A feature matrix is ​​constructed based on the statistical and mechanistic features of each yaw event, and a dual model of single-class support vector machine and isolated forest is trained. The real-time data of the target wind turbine is repeatedly preprocessed and classified into events. Anomalies are detected by the dual model and the anomaly scores are fused to determine anomalies. This embodiment of the invention can make full use of existing SCADA data, achieve unsupervised anomaly detection without fault labels, avoid the one-sidedness of a single model, and improve detection accuracy. By classifying operating conditions and extracting mechanistic features, it solves the problems of low signal-to-noise ratio and disconnect between features and mechanisms in traditional methods, realizes early fault warning of yaw reducers, and reduces detection costs.

[0076] In practical applications, the first step is to collect, preprocess, and classify yaw events. At least one year of historical data for the target wind turbine and other healthy turbines of the same model is obtained from the wind farm's SCADA system. This data covers wind turbine operating parameters, yaw system parameters, yaw status parameters, and environmental parameters. The historical data is preprocessed to remove invalid data from shutdown, power-limited, or fault states, and to process abnormal values ​​such as unchanged or fluctuating sensor readings. Then, by jointly judging the state transition of the yaw status flag bit and the change in the yaw angle value, the complete process of the flag bit flipping from 0 to 1 (yaw start) to 1 to 0 (yaw end) is captured, thus classifying at least one independent yaw event.

[0077] Secondly, a feature matrix is ​​constructed and the optimal dual model is trained. For each yaw event after segmentation, statistical and mechanistic features are extracted. Statistical features include the mean, variance, and skewness of key variables, while mechanistic features include yaw efficiency indices, yaw stability indices, and wind alignment accuracy indices. The features of all yaw events are organized into a feature matrix according to a unified dimension. The feature matrix is ​​standardized, and a grid search method is used to optimize the kernel function and boundary parameters of the single-class support vector machine model, and the number of trees and samples in the isolated forest model, among other hyperparameters. After multiple rounds of iterative training, the optimal dual model is obtained and its parameters are saved.

[0078] Furthermore, anomalies are detected and fused in real time. Real-time operating data of the target wind turbine is acquired, preprocessed and yaw event-classified according to the same standards as historical data, and corresponding feature matrices are extracted and standardized. The standardized feature matrices are then input into a single-class support vector machine target model and an isolated forest target model, respectively, to obtain two raw anomaly scores. The raw scores are normalized to the same dimension using quantile mapping, fused according to preset weights to obtain a comprehensive anomaly score, and compared with a preset threshold to determine whether the yaw reducer is abnormal.

[0079] In some embodiments, such as Figure 3As shown in the figure, this embodiment of the invention also provides a wind turbine yaw reducer abnormality detection device, comprising: The preprocessing module preprocesses the historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine unit, and identifies at least one yaw event by combining the yaw status flag and yaw angle in the preprocessed data. The training module, based on the feature matrix constructed from the statistical and mechanistic features of each yaw event, trains the optimal dual model to obtain the optimal dual model; wherein, the optimal dual model is a single-class support vector machine target model and an isolated forest target model; The detection module acquires real-time data of the target wind turbine and performs anomaly detection using the optimal dual model. It then fuses the anomaly scores to determine whether the yaw reducer of the target wind turbine is abnormal.

[0080] This invention achieves adaptive early warning of faults by establishing an independent health benchmark model for each wind turbine and monitoring the deviation of operating data from this benchmark in real time. First, data collection is performed, gathering as many yaw-related variables as possible to extract richer fault features. Then, data preprocessing and event segmentation are conducted. Preprocessing removes data related to shutdowns, power limitations, faults, and sensor anomalies to avoid affecting model detection. Event segmentation accurately extracts complete yaw events, and detailed feature extraction is performed on each yaw event to improve model detection performance. Subsequently, feature extraction is performed on each yaw event, including statistical and mechanistic features, to better reflect abnormal yaw states. Finally, a fusion of a One-Class SVM model and an Isolation Forest model is used for anomaly detection to better eliminate the limitations of individual model diagnosis and improve anomaly diagnosis.

[0081] The present invention provides an embodiment of an electronic device. In this embodiment, the electronic device may be, but is not limited to, a personal computer (PC), a laptop computer, a monitoring device, a server, or other computer device with analysis and processing capabilities.

[0082] As an exemplary embodiment, see [link to example]. Figure 4 The electronic device 110 includes a communication interface 111, a processor 112, a memory 113, and a bus 114. The processor 112, the communication interface 111, and the memory 113 are connected via the bus 114. The memory 113 is used to store a computer program that supports the processor 112 in executing the above-described method. The processor 112 is configured to execute the program stored in the memory 113.

[0083] The machine-readable storage medium mentioned in this article can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as executable instructions, data, etc. For example, machine-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0084] Non-volatile media can be non-volatile memory, flash memory, storage drives (such as hard disk drives), any type of storage disk (such as optical discs, DVDs, etc.), or similar non-volatile storage media, or combinations thereof.

[0085] It is understood that the specific operation methods of each functional module in this embodiment can be referred to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0086] The computer-readable storage medium provided in the embodiments of the present invention stores a computer program. When the computer program code is executed, it can implement the method described in any of the above embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0088] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0089] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0090] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A wind turbine generator yaw reducer abnormality detection method characterized by, include: The historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine are preprocessed, and at least one yaw event is identified by combining the yaw status flag and yaw angle in the preprocessed data. The optimal dual model is obtained by constructing a feature matrix based on the statistical and mechanistic characteristics of each yaw event; wherein the optimal dual model is a single-class support vector machine target model and an isolated forest target model. Real-time data of the target wind turbine is acquired and anomaly detection is performed using the optimal dual model. The anomaly scores are then fused to determine whether the yaw reducer of the target wind turbine is abnormal.

2. The method of claim 1, wherein, The steps of preprocessing historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine generator set, and identifying at least one yaw event by jointly using the yaw status flag and yaw angle in the preprocessed data, include: Historical data of the target wind turbine and healthy wind turbines of the same model as the target wind turbine were collected from the wind turbine generator set. The historical data is preprocessed to obtain preprocessed data with invalid data removed; By combining the state switching of the yaw status flag bit and the numerical change of the yaw angle in the preprocessed data, at least one yaw event to which the preprocessed data belongs is determined.

3. The method of claim 2, wherein, The historical data includes wind turbine operating parameters, yaw system parameters, yaw status parameters, and environmental parameters; the historical data collection period is no less than one year.

4. The method of claim 2, wherein, The step of preprocessing the historical data to obtain preprocessed data after removing invalid data includes: Invalid data from the historical data under shutdown, power limitation, and fault conditions are processed by removing abnormal values ​​such as unchanged or fluctuating sensor values ​​to obtain preprocessed data.

5. The method of claim 1, wherein, The steps for training dual models based on feature matrices constructed from the statistical and mechanistic characteristics of each yaw event include: Based on the preprocessed data corresponding to each yaw event, relevant statistical and mechanistic features are extracted, and a feature matrix is ​​constructed. The feature matrix is ​​standardized, and the hyperparameters of the single-class support vector machine model and the isolated forest model are optimized by grid search method respectively. The optimal dual model is trained and the target hyperparameters are saved.

6. The method of claim 5, wherein, The statistical features include at least the mean, variance, and skewness of key variables within the yaw event; the mechanistic features include at least the yaw efficiency index, yaw stability index, and wind accuracy index used to map the physical failure mechanism of the yaw reducer; the statistical features and mechanistic features of each yaw event constitute a row in the feature matrix.

7. The method of claim 1, wherein, The steps of acquiring real-time data of the target wind turbine and performing anomaly detection using the optimal dual model, and fusing anomaly scores to determine whether the yaw reducer of the target wind turbine is abnormal, include: The real-time data of the target wind turbine is preprocessed and segmented, and the feature matrix of each yaw event is extracted. The feature matrix is ​​standardized and then input into the optimal dual model to obtain the corresponding first anomaly score and second anomaly score. After normalizing the first and second abnormal scores using the quantile mapping method, the comprehensive abnormal score is obtained by fusing them according to a preset weight. Based on the comprehensive anomaly score and the preset threshold, the abnormal state of the yaw reducer of the target wind turbine is determined.

8. An abnormality detection device for a yaw reducer of a wind turbine generator unit, characterized by comprising: include: The preprocessing module preprocesses the historical data of the target wind turbine and healthy wind turbines of the same model collected from the wind turbine unit, and identifies at least one yaw event by combining the yaw status flag and yaw angle in the preprocessed data. The training module, based on the feature matrix constructed from the statistical and mechanistic features of each yaw event, trains the optimal dual model to obtain the optimal dual model; wherein, the optimal dual model is a single-class support vector machine target model and an isolated forest target model; The detection module acquires real-time data of the target wind turbine and performs anomaly detection using the optimal dual model. It then fuses the anomaly scores to determine whether the yaw reducer of the target wind turbine is abnormal.

9. An electronic device, comprising: It includes a memory, a processor, and a program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed, implements the method described in any one of claims 1-7.

Images