Health state model construction method, wind turbine generator yaw friction plate anomaly detection method, device and equipment

Abstract

The invention provides a health state model construction method, and a wind turbine generator yaw friction plate anomaly detection method, device and equipment. The method comprises the steps of obtaining yaw event data collected in real time; performing preprocessing and feature extraction on the yaw event data to obtain multi-dimensional input features; predicting the multi-dimensional input features by using a pre-constructed health state model, and outputting a predicted yaw motor current mean value in a normal state; and detecting an abnormal state of the yaw friction plate based on a residual value between the predicted average current value of the yaw motor and the actually observed average current value of the yaw motor. According to the method, the current mean value prediction is performed on the real-time yaw event by using the pre-constructed health state model, and the residual error is generated by comparing the predicted value with the actual observation value, so that the residual error can effectively peel off interference brought by normal working condition fluctuation, and normal phenomena such as large current at a high wind speed are prevented from being misjudged as faults. The false alarm rate caused by working condition changes is greatly reduced, and the detection accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of wind turbine operation and maintenance technology, and more specifically, to a method for constructing a health status model, a method for detecting abnormalities in the yaw friction pads of a wind turbine, and related devices and equipment. Background Technology

[0002] The yaw system of a wind turbine is used to adjust the nacelle's orientation in real time to align it with the prevailing wind direction, maximizing wind energy capture efficiency. This yaw system uses friction pads (or brake pads) to achieve yaw braking and positioning. During long-term operation, the yaw friction pads are susceptible to wear and contamination, leading to a decrease in braking torque, increased yaw positioning deviation, and even yaw failure or structural damage, seriously threatening the unit's safety and power generation performance.

[0003] Currently, wind farms typically monitor the condition of yaw friction plates using simple threshold alarms based on current or vibration monitoring. This method is susceptible to interference from changes in operating conditions and has a high false alarm rate. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method for constructing a health status model, a method, device and equipment for detecting abnormalities in yaw friction pads of wind turbine units, so as to improve the accuracy of yaw friction pad abnormality detection.

[0005] Firstly, a method for constructing a health state model is provided. This health state model is used to predict the mean current of a yaw motor, and it employs a random forest model. This includes: Acquire time-series data of historical yaw events of wind turbines under normal operating conditions, and extract multidimensional features associated with yaw events to form an original sample set using the multidimensional features of several historical yaw events; The original sample set is divided into a training set and a validation set; Several initial random forest models are initially constructed based on the training set and several preset hyperparameter combinations; the performance of each initial random forest model is verified based on the validation set, and the optimal hyperparameter combination is determined based on the validation results. The random forest model is reconstructed using the optimal combination of hyperparameters and the full original sample set to obtain the health status model.

[0006] Optionally, time-series data of historical yaw events of the wind turbine under normal operating conditions can be obtained, and multi-dimensional features associated with the yaw events can be extracted, including: Obtain raw data of historical yaw events under normal operating conditions; the raw data should include at least the yaw status flag. The raw data is filtered by status to obtain data under normal operating conditions of the unit; Based on the rising and falling edges of the yaw status flag, yaw event slices are made on the data under normal operating conditions, and the time sequence segment corresponding to each complete yaw event is extracted. Based on the time sequence of each yaw event, multidimensional features corresponding to each yaw event are extracted.

[0007] Optional, multidimensional features include current dynamic features, operating condition background features, operating mode features, and performance index features; among which: The current dynamic characteristics include at least one of the following: peak starting current, current standard deviation, and peak factor; The background characteristics of the operating conditions include at least one of the following: average wind speed, turbulence intensity, and ambient temperature. The action pattern features include at least one of yaw angle, yaw duration, and maximum yaw speed; Performance characteristics include energy consumption per unit angle of yaw or current fluctuation coefficient.

[0008] Optionally, several initial random forest models are initially constructed based on the training set and several preset hyperparameter combinations; and the performance of each initial random forest model is verified based on the validation set, and the optimal hyperparameter combination is determined based on the validation results, including: For each set of preset hyperparameters, a new sample set is generated on the training set by sampling with replacement. And based on the new sample set and the corresponding hyperparameter combination, the corresponding initial random forest model is constructed; The performance of each initial random forest model was validated using a validation set. The performance was measured as the error between the predicted mean yaw motor current and the actual mean current. Repeat the above process multiple times to complete cross-validation. Based on the cross-validation results, obtain the average performance index of the random forest model corresponding to each hyperparameter combination. Then, take the hyperparameter combination with the best average performance index as the optimal hyperparameter combination.

[0009] Optionally, optimal hyperparameters include the number of decision trees, the maximum number of features per tree split, the maximum depth of the decision tree, the minimum number of samples required for node splits, and the minimum number of samples for leaf nodes. Using the optimal hyperparameter combination and the full original sample set, the random forest model is reconstructed to obtain the healthy state model, which includes: Based on the number of decision trees, construct a corresponding number of decision trees; For each decision tree, when splitting at each internal node of the decision tree, candidate features with a number not exceeding the maximum number of features are randomly selected from the multidimensional features; and the thresholds of each candidate feature are traversed; the feature and threshold that maximize the reduction in variance of the mean yaw motor current are selected for splitting. Splitting stops when any of the following conditions are met: the number of samples in the current node is less than the minimum number of samples required for the node to split, the number of samples in the leaf node is less than the minimum number of samples in the leaf node, or the depth of the current tree reaches the maximum depth of the decision tree. Once all decision trees are constructed, a health status model is formed.

[0010] Secondly, a method for detecting abnormalities in the yaw friction pads of a wind turbine is provided, including: Acquire real-time yaw event data; Preprocessing and feature extraction of yaw event data yields multidimensional input features; Using a pre-built health status model, multi-dimensional input features are predicted, and the average yaw motor current predicted under normal conditions is output. The abnormal state of the yaw friction plate is detected based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current.

[0011] Optionally, based on the residual between the predicted mean yaw motor current and the actual observed mean yaw motor current, the abnormal state of the yaw friction pads can be detected, including: Based on multiple residual values ​​obtained within a preset window, the residual change trend and health index are determined. Based on the sign of each residual value, the residual change trend, and the accompanying current dynamic characteristics, the failure modes of the yaw friction pads are identified. Early warning levels are determined based on the range of health index values, and corresponding early warning actions are triggered based on the results of the early warning levels.

[0012] Optionally, based on the sign of each residual value, the residual variation trend, and the accompanying current dynamic characteristics, the failure modes of the yaw friction pads can be identified as follows: If the sign of the residual is consistently positive and shows a slow upward trend, it is determined to be a wear failure. If the sign of the residual changes to a negative value, accompanied by increased current fluctuations and intermittent slippage, it is determined to be a pollution fault. If the residual increases significantly after a long period of yaw and partially recovers after cooling, it is determined to be a thermal decay failure. If the residual performance is asymmetrical under different yaw directions, it is judged as an uneven wear fault.

[0013] Thirdly, a device for detecting abnormal yaw friction pads in wind turbine generators is provided, comprising: The acquisition unit is used to acquire real-time yaw event data; The feature extraction unit is used to preprocess and extract features from yaw event data to obtain multidimensional input features; The prediction unit is used to predict multi-dimensional input features using a pre-built health status model and output the predicted mean yaw motor current under normal conditions. The detection unit is used to detect abnormal conditions of the yaw friction plate based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current.

[0014] Fourthly, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements either the first aspect or the second aspect of the method.

[0015] This invention provides a method for constructing a health state model, a method for detecting anomalies in the yaw friction pads of wind turbines, an apparatus, and equipment. The method involves acquiring real-time yaw event data; preprocessing and extracting features from the yaw event data to obtain multi-dimensional input features; using a pre-constructed health state model to predict the multi-dimensional input features, outputting the predicted mean yaw motor current under normal conditions; and detecting abnormal states of the yaw friction pads based on the residual value between the predicted mean yaw motor current and the actual observed mean yaw motor current. This invention utilizes a pre-constructed health state model to predict the mean current of real-time yaw events and generates a residual by comparing the predicted value with the actual observed value, thereby determining whether an anomaly exists. Since the predicted value fully considers the influence of operating conditions such as current wind speed, turbulence intensity, and yaw angle, the residual can effectively eliminate interference caused by fluctuations in normal operating conditions, avoiding misjudging normal phenomena such as high current under high wind speeds as faults. Compared with traditional alarm methods based on fixed current thresholds, this method significantly reduces the false alarm rate caused by changes in operating conditions, achieving highly reliable detection of the true abnormal state of the yaw friction pads.

[0016] 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

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1A flowchart illustrating a method for constructing a health status model according to an embodiment of the present invention is shown; Figure 2 The flowchart of a method for detecting abnormal yaw friction pads in a wind turbine provided by an embodiment of the present invention is shown. Figure 3 This diagram illustrates the variation trend of the mean yaw motor current j under different fault modes provided in the embodiments of the present invention. Figure 4 This diagram illustrates the trend of health index changes provided in an embodiment of the present invention. Figure 5 This diagram illustrates the structure of a wind turbine yaw friction plate anomaly detection device provided in an embodiment of the present invention. Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments 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, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] This invention provides a method for constructing a health state model, which is used to predict the mean yaw motor current. The health state model employs a random forest model. Figure 1 As shown, the method includes: Step S101: Obtain time-series data of historical yaw events of wind turbines under normal operating conditions, and extract multi-dimensional features associated with yaw events to form an original sample set using the multi-dimensional features of several historical yaw events.

[0021] In this embodiment of the invention, the time-series data of historical yaw events are derived from the long-term operational data collected by the wind turbine SCADA system, with a time resolution typically of 1 second or higher.

[0022] Normal operating condition refers to the unit not being in abnormal operating conditions such as shutdown, power limitation, scheduled maintenance, or fault alarm, and only includes yaw action during stable power generation.

[0023] Multidimensional features are a set of key indicators obtained by feature engineering the original signal. Each yaw event corresponds to a feature vector, which serves as a training sample for the model. By constructing a large number of such samples, a benchmark dataset reflecting the health status of the friction plates can be formed.

[0024] For example, if a wind farm has recorded 500 confirmed yaw events without any abnormalities, and each yaw event has 25 dimensions of features extracted, then a raw sample set of 500 rows × 25 columns can be generated, providing a data foundation for subsequent modeling.

[0025] Step S102: Divide the original sample set into a training set and a validation set.

[0026] Specifically, the partitioning method employs a cross-validation strategy, such as five-fold cross-validation, which randomly divides the original sample set into five mutually exclusive subsets. In each round of validation, four subsets are selected as the training set, and the remaining subset is used as the validation set, ensuring the unbiasedness of the model evaluation.

[0027] It should be noted that this partitioning is only used in the hyperparameter optimization stage; all samples will be used during the final model training. Continuing the previous example, if the original sample set contains 500 samples, then each fold contains 100 samples, resulting in a training set of 400 samples and a validation set of 100 samples. This partitioning ensures that the hyperparameter selection process can assess the generalization ability of the data not yet seen.

[0028] Step S103: Construct several initial random forest models based on the training set and several preset hyperparameter combinations; verify the performance of each initial random forest model based on the validation set, and determine the optimal hyperparameter combination based on the validation results.

[0029] In this embodiment of the invention, the hyperparameter combination is predefined by grid search, and each set of parameters corresponds to an initial random forest model, the performance of which is measured by the prediction error on the validation set. Through multiple rounds of cross-validation, the average performance index of each set of parameters can be obtained, thereby selecting the optimal combination. This process avoids the risk of overfitting to a specific subset of data and ensures the robustness of the selected hyperparameters.

[0030] Step S104: Reconstruct the random forest model using the optimal hyperparameter combination and the full original sample set to obtain the health status model.

[0031] The final model makes full use of all healthy sample information and has stronger representation ability and predictive stability compared to the temporary model in the cross-validation stage.

[0032] Once deployed, this health status model can output the average current value that should be present under normal conditions for newly occurring yaw events, providing a benchmark for anomaly detection.

[0033] This invention, through the construction of a random forest health state model based on multi-dimensional operating condition features, effectively overcomes the shortcomings of traditional fixed threshold methods that cannot adapt to changes in operating conditions such as wind speed, turbulence, and yaw angle. The model uses historical normal yaw events as its learning foundation and explicitly integrates multi-source features such as current dynamics, environmental conditions, and action patterns during training. It can dynamically predict the "normal average current under the current operating condition" rather than using a static threshold. The resulting health benchmark exhibits strong operating condition adaptability, significantly improving the accuracy of representing the normal behavior of the yaw system and providing a reliable and robust reference for subsequent anomaly detection.

[0034] Based on the above embodiments, time-series data of historical yaw events of wind turbines under normal operating conditions are obtained, and multi-dimensional features associated with yaw events are extracted, including: Step S101A: Obtain raw data of historical yaw events under normal operation; the raw data shall include at least the yaw status flag bit.

[0035] Specifically, the raw data is continuously recorded by the SCADA system, including the yaw motor three-phase current, yaw angle, wind speed, wind direction, ambient temperature, unit power, and yaw command signal. Among them, the yaw status flag is a Boolean variable (0 or 1), which is generated by the main control system based on the yaw enable signal and the rate of change of angle, and is used to accurately identify the start and end times of the yaw action.

[0036] Step S101B: Filter the raw data by status to obtain data under normal operating conditions of the unit.

[0037] It should be noted that the status screening aims to eliminate interfering data. For example, when the unit is operating with limited power (due to grid dispatch), undergoing scheduled maintenance, experiencing pitch failure, or in an emergency shutdown, the yaw behavior is affected by abnormal torque and cannot represent the true health status of the friction plates.

[0038] Therefore, only data from periods when the main control system is marked as generating power normally and without fault codes are retained to ensure the purity of the sample.

[0039] Step S101C: Based on the rising and falling edges of the yaw status flag, slice the data under normal operating conditions into yaw events and extract the time sequence segment corresponding to each complete yaw event.

[0040] For example, when the yaw status flag changes from 0 to 1, it is considered the start of yaw; when it changes back from 1 to 0, it is considered the end of yaw. The system then extracts all timing data from the start to the end to form a complete yaw event segment.

[0041] To improve robustness, the yaw angle change rate (e.g., >0.1° / s) can be used for joint judgment to avoid erroneous slicing due to signal jitter.

[0042] Step S101D: Based on the time sequence segment of each yaw event, extract the multidimensional features corresponding to each yaw event.

[0043] In this embodiment of the invention, the multidimensional features include current dynamic features, operating condition background features, action mode features, and performance index features. Current dynamic features reflect the current response pattern, operating condition background features characterize the external environment, action mode features describe the yaw process, and performance index features quantify execution efficiency. Through this structured feature system, the model can comprehensively perceive the working state of the friction pad. Specifically: The current dynamic characteristics include at least one of the following: peak starting current, current standard deviation, and peak factor; The background characteristics of the operating conditions include at least one of the following: average wind speed, turbulence intensity, and ambient temperature. The action pattern features include at least one of yaw angle, yaw duration, and maximum yaw speed; Performance characteristics include energy consumption per unit angle of yaw or current fluctuation coefficient.

[0044] For example, a single yaw event may extract 25 features, such as peak starting current of 68.2A, turbulence intensity of 12.5%, yaw angle of 45°, and energy consumption per unit angle of 0.017kWh / °, to form a high-dimensional input vector.

[0045] This can be explained by the fact that the peak starting current reflects the resistance at the moment of yaw initiation, and wear will cause this value to increase; turbulence intensity characterizes wind disturbance and affects yaw load; yaw duration is directly related to friction torque; and energy consumption per unit angle comprehensively reflects yaw efficiency. These features together constitute a multi-angle observation of the friction plate state, significantly improving the model's discriminative ability. In practical applications, feature subsets can be flexibly selected according to data availability without affecting the method's universality.

[0046] Based on the above embodiments, several initial random forest models are initially constructed using a training set and several preset hyperparameter combinations; the performance of each initial random forest model is verified using a validation set, and the optimal hyperparameter combination is determined based on the verification results, including: Step S103A: For each set of preset hyperparameter combinations, generate a new sample set by sampling with replacement on the training set.

[0047] Specifically, there is sampling with replacement, also known as bootstrap sampling, which randomly selects the same number of samples from the training set (e.g., 400 samples), allowing for repetition. For example, a sampling might include sample #5 three times, sample #12 once, while sample #3 is not selected. This mechanism introduces randomness at the data level, enhancing the diversity between trees.

[0048] Step S103B: Construct the corresponding initial random forest model based on the new sample set and the corresponding hyperparameter combination.

[0049] In this embodiment of the invention, each decision tree is constructed independently. During splitting, only a randomly selected subset of features (e.g., 5) is considered, and the optimal split point is selected by minimizing the mean variance of the current. The structure of the entire tree is constrained by hyperparameters (e.g., maximum depth = 5) to prevent overfitting.

[0050] Step S103C: Validate the performance of each initial random forest model using the validation set. The performance is the error between the predicted mean yaw motor current and the actual mean current.

[0051] For example, the performance metric can be the mean squared error (MSE):

[0052] in, This represents the average actual yaw motor current. The model predicts the mean current of the d-yaw motor. The smaller the MSE, the more accurate the model.

[0053] Step S103D: Repeat the above process multiple times to complete cross-validation, obtain the average performance index of the random forest model corresponding to each hyperparameter combination based on the cross-validation results, and take the hyperparameter combination with the best average performance index as the optimal hyperparameter combination.

[0054] Multiple rounds of validation eliminate the randomness of a single partition, ensuring that the selected hyperparameter combinations are robust across different data subsets.

[0055] The optimal hyperparameters include the number of decision trees, the maximum number of features when each tree splits, the maximum depth of the decision tree, the minimum number of samples required for node splits, and the minimum number of samples for leaf nodes.

[0056] These parameters collectively control the model's complexity and generalization ability, and are key to achieving high-precision health modeling.

[0057] Based on the above embodiments, the random forest model is reconstructed using the optimal hyperparameter combination and the full original sample set, resulting in a health state model including: Step S104A: Based on the number of decision trees, construct the corresponding number of decision trees.

[0058] Specifically, if the optimal number of decision trees is 100, then 100 independent decision trees are constructed sequentially, and each tree is based on the full original sample set (e.g., 500 samples) for Bootstrap sampling.

[0059] Step S104B: For each decision tree, when splitting at each internal node of the decision tree, randomly select candidate features from the multidimensional features with a number not exceeding the maximum number of features; and traverse the threshold of each candidate feature; select the feature and threshold that maximizes the reduction in variance of the yaw motor current mean for splitting.

[0060] In this embodiment of the invention, the process of constructing a decision tree is a process of continuously splitting nodes. When each decision tree splits internal nodes, its core objective is to make the average yaw motor current in the child nodes as pure as possible, that is, to make the predicted values ​​more concentrated and the error smaller.

[0061] To this end, the algorithm iterates through all candidate features and their possible splitting thresholds, calculates the reduction in the mean variance of the current brought about by each splitting scheme, and selects the feature and threshold that maximizes the reduction as the splitting rule for the current node.

[0062] In a specific example, the total number of input features is 25 (including: peak starting current, standard deviation of current, turbulence intensity, average wind speed, yaw angle, energy consumption per unit angle, etc.).

[0063] Current node to be split: contains 80 yaw event samples (all from Bootstrap sampling); Hyperparameter setting: max_features=5 (that is, only 5 random features are considered in each split).

[0064] The steps of node splitting are explained in detail using the current internal node as an example: Step 1: Randomly select 5 features from all 25 features as candidates; Suppose that the five randomly selected features are: peak starting current (A), turbulence intensity (%), yaw duration (s), energy consumption per unit angle (kWh / °) and ambient temperature (°C).

[0065] These 5 are the only candidates for this split; the remaining 20 features are not considered at all.

[0066] Step 2: For each candidate feature, iterate through all possible splitting thresholds; For each feature, collect the values ​​of the 80 samples in the current node for that feature, and try multiple potential split points (usually the midpoint between adjacent sample values).

[0067] Taking the processing of peak starting current characteristics as an example The peak values ​​of the starting current of the 80 samples, after being sorted, are: [62.1, 63.5, ..., 78.9].

[0068] Possible splitting thresholds include: 62.8, 64.2, ..., 78.0 (79 candidate thresholds in total).

[0069] For each threshold t, the samples are divided into two groups: left subset: peak starting current ≤ t; right subset: peak starting current > t.

[0070] Step 3: Calculate the splitting quality for each (feature + threshold) combination; Decrease in variance of mean yaw motor current:

[0071] in: : The variance of the mean label of the yaw motor current in the parent node (current node); , : Variance of the mean labels of the yaw motor currents in the left and right child nodes; , : The number of samples in the left and right child nodes; .

[0072] The goal of node splitting is to maximize improvement (even if the child nodes are purer).

[0073] The threshold and improvement calculated for each candidate feature are shown in the table below:

[0074] Step 4: Select the feature with the greatest improvement (feature + threshold) for splitting.

[0075] Based on the table above, the selected feature is turbulence intensity; the selected threshold is 14.2%.

[0076] Splitting is performed based on selected features and thresholds: if the turbulence intensity is ≤14.2%, the sample enters the left child node; otherwise, it enters the right child node.

[0077] At this point, the internal node has split, generating two child nodes. The above process is then repeated for each child node (with 5 more features randomly selected, etc.) until the stopping condition is met.

[0078] The greater the reduction in variance, the more concentrated the mean yaw motor current of the child nodes after splitting and the lower the prediction uncertainty. Therefore, this embodiment can automatically learn which operating conditions (such as high turbulence and long yaw time) will cause the current response to deviate from the normal range through this mechanism, thereby constructing a health status benchmark that is sensitive to the degradation of the friction plate.

[0079] Step S104C: Stop splitting when any of the following conditions are met.

[0080] The conditions are: the number of samples in the current node is less than the minimum number of samples required for the node to split, the number of samples in the leaf node is less than the minimum number of samples in the leaf node, or the depth of the current tree reaches the maximum depth of the decision tree.

[0081] It should be noted that the minimum number of samples required for node splitting prevents splitting on extremely small samples (such as noise points), the minimum number of samples required for leaf nodes ensures that the predicted values ​​of leaf nodes are statistically significant (e.g., averaging at least two samples), and the maximum depth of the decision tree limits the model complexity. These three factors work together to avoid overfitting.

[0082] Step S104D: After all decision trees are constructed, a health status model is formed.

[0083] In a specific example, suppose we build a random forest model with 100 decision trees and 16 samples. The optimal combination of hyperparameters is such that the minimum number of samples required for node splits is 4, and the minimum number of samples required for leaf nodes is 2.

[0084] Step 1: Bootstrap sampling: Draw 16 samples with replacement from the 16 samples to obtain a subset of samples.

[0085] Suppose that the only sample drawn is: #5, #6, #7, #8, #9, #10, #11, #12, #13, #14 (a total of 10, the rest are duplicates or not drawn).

[0086] Step 2: Begin building the tree: The root node has 10 samples.

[0087] Attempt to split: Assume "turbulence intensity" is selected as the best feature, and the threshold is 15%.

[0088] After splitting, the left child node contains 5 samples (#5–#9).

[0089] Right child node: 5 samples (#10–#14) The minimum number of samples required to check for node splitting is 4: if both the left and right nodes have ≥4 samples, splitting is allowed.

[0090] Continue splitting the left child node: Suppose that after splitting, we get two child nodes: 3 samples + 2 samples.

[0091] At this point, the number of samples in these two new nodes is less than 4. They cannot split further and become leaf nodes.

[0092] Similarly, the right child node is processed, and finally the first decision tree is constructed.

[0093] By constructing the remaining decision trees using the same steps, a complete health status model is finally obtained.

[0094] In this model, each of the 100 decision trees outputs a current mean prediction, and the final result is the arithmetic mean. This ensemble mechanism significantly reduces the variance of individual decision trees, improves model robustness, and provides a reliable benchmark for subsequent anomaly diagnosis.

[0095] Another embodiment of the present invention provides a method for detecting abnormalities in the yaw friction pads of a wind turbine, such as... Figure 2 As shown, it includes the following steps: Step S201: Obtain real-time yaw event data.

[0096] In this embodiment of the invention, the real-time data also comes from the SCADA system, including signals such as current, angle, and wind speed collected during the yaw process, and the time window covers the complete yaw action.

[0097] Step S202: Preprocess and extract features from the yaw event data to obtain multidimensional input features.

[0098] Specifically, preprocessing includes synchronous slicing (based on the yaw flag) and outlier filtering; the feature extraction process is consistent with the training phase to ensure input consistency, and will not be elaborated here. For example, 25-dimensional features such as peak startup current and yaw duration are extracted.

[0099] Step S203: Using a pre-built health status model, predict the multi-dimensional input features and output the predicted average yaw motor current under normal conditions.

[0100] The health status model is pre-built based on the method described in the above embodiments.

[0101] Step S204: Based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current, detect the abnormal state of the yaw friction plate.

[0102] The residual calculation section calculates the residual between the predicted mean value of the yaw motor current and the actual mean value of the yaw motor current. The calculation formula is as follows:

[0103] in, This is the actual average value of the yaw motor current. It is a prediction of the average current of the yaw motor.

[0104] This invention utilizes a pre-constructed health status model to predict the mean current of real-time yaw events. By comparing the predicted values ​​with actual observed values ​​to generate residuals, it determines whether an anomaly exists. Since the predicted values ​​fully consider the influence of operating conditions such as current wind speed, turbulence intensity, and yaw angle, the residuals effectively eliminate interference from fluctuations in normal operating conditions, avoiding misjudging normal phenomena such as high current under high wind speeds as faults. Compared to traditional alarm methods based on fixed current thresholds, this method significantly reduces the false alarm rate caused by changes in operating conditions, achieving highly reliable detection of the true abnormal state of the yaw friction pads.

[0105] Based on the above embodiments, the abnormal state of the yaw friction pads is detected by using the residual value between the predicted average yaw motor current and the actual observed average yaw motor current. Step S204A: Based on multiple residual values ​​obtained within a preset window, determine the residual change trend and health index.

[0106] For example, multiple residual values ​​can be processed using a sliding window method. The health index is the average value of the residual values ​​over the sliding window, and the calculation formula is as follows:

[0107] in, The window size for East China is typically set to 10-20, meaning that 10 to 20 historical monitoring data points constitute a sliding window.

[0108] Calculating the Health Index can help avoid false alarms caused by single points of failure; a persistently high HI value indicates a systemic failure.

[0109] Step S204B: Identify the fault modes of the yaw friction pads based on the sign of each residual value, the residual change trend, and the accompanying current dynamic characteristics.

[0110] In one feasible implementation, based on the sign of each residual value, the residual variation trend, and the accompanying current dynamic characteristics, the failure modes of the yaw friction pads are identified as follows: If the residual sign remains positive and shows a slow upward trend, it is determined to be a wear failure.

[0111] If the sign of the residual changes to a negative value, accompanied by increased current fluctuations and intermittent slippage, it is determined to be a pollution fault.

[0112] If the residual increases significantly after a long period of yaw and partially recovers after cooling, it is determined to be a thermal decay failure.

[0113] If the residual performance is asymmetrical under different yaw directions, it is judged as an uneven wear fault.

[0114] This can be explained by the following: wear increases friction, requiring a larger current to drive the system, hence the residual is positive; contamination or lubrication reduces friction, lowering the current, resulting in a negative residual; thermal decay is temperature-dependent; and uneven wear exhibits direction sensitivity. This mechanism enables physically explainable identification of failure modes.

[0115] like Figure 3 The figure shows the current variation under different fault modes.

[0116] Step S204C: Analyze the warning levels based on the numerical range of the health index, and trigger the corresponding warning actions based on the results of the warning levels.

[0117] Figure 4 The chart provides a trend monitoring graph for the health index.

[0118] The degree of abnormality can be further determined based on the calculated health index, and the judgment thresholds are shown in the table below:

[0119] This tiered strategy supports differentiated operation and maintenance decisions, improving the reliability and efficiency of wind turbines.

[0120] This method is based on the residual between the predicted mean current output by the health status model and the actual observed value. It can not only determine whether there is an anomaly, but also distinguish typical fault modes such as wear, pollution, thermal decay and uneven wear by combining the sign of the residual, the time trend and the dynamic characteristics of the current. Furthermore, it introduces a health index and sets multi-level thresholds to realize graded early warning from normal to emergency, and supports differentiated operation and maintenance strategies.

[0121] Based on the same inventive concept, a device for detecting abnormal yaw friction plates in wind turbine generators is provided, such as... Figure 5 As shown, it includes: Acquisition unit 501 is used to acquire real-time yaw event data.

[0122] The feature extraction unit 502 is used to preprocess and extract features from the yaw event data to obtain multidimensional input features.

[0123] The prediction unit 503 is used to predict multi-dimensional input features using a pre-built health state model and output the predicted average yaw motor current under normal conditions.

[0124] The detection unit 504 is used to detect abnormal conditions of the yaw friction plate based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current.

[0125] Based on the same technical concept, embodiments of the present invention also provide an electronic device, such as... Figure 6As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604.

[0126] Memory 603 is used to store computer programs; When the processor 601 executes the program stored in the memory 603, it implements the steps of the above method embodiments.

[0127] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0128] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0129] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0130] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0131] The device for detecting abnormal yaw friction plates in wind turbines provided in this embodiment of the invention can be specific hardware on the equipment or software or firmware installed on the equipment. The implementation principle and technical effects of the device provided in this embodiment of the invention are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiments can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0132] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, the functional units in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0135] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0137] 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 it. 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 technical scope 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. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a health status model, characterized in that, The health state model is used to predict the mean yaw motor current, and the health state model adopts a random forest model; the method includes: Time series data of historical yaw events of wind turbines under normal operating conditions are obtained, and multidimensional features associated with yaw events are extracted. The multidimensional features of several historical yaw events constitute the original sample set. The original sample set is divided into a training set and a validation set; Several initial random forest models are initially constructed based on the training set and several preset hyperparameter combinations; the performance of each initial random forest model is verified based on the validation set, and the optimal hyperparameter combination is determined based on the validation results. The random forest model is reconstructed using the optimal hyperparameter combination and the full original sample set to obtain the health status model.

2. The method according to claim 1, characterized in that, The process of acquiring time-series data of historical yaw events of wind turbines under normal operating conditions and extracting multidimensional features associated with yaw events includes: Obtain raw data of historical yaw events under normal operation; the raw data includes at least the yaw status flag bit; The raw data is filtered by status to obtain data under normal operating conditions of the unit; Based on the rising and falling edges of the yaw status flag, the data under normal operating conditions is sliced ​​into yaw events, and the time sequence segment corresponding to each complete yaw event is extracted. Based on the time sequence of each yaw event, multidimensional features corresponding to each yaw event are extracted.

3. The method according to claim 2, characterized in that, The multidimensional features include current dynamic features, operating condition background features, operating mode features, and performance index features; among which: The current dynamic characteristics include at least one of the following: peak starting current, current standard deviation, and peak factor; The background characteristics of the operating conditions include at least one of the following: average wind speed, turbulence intensity, and ambient temperature. The action pattern features include at least one of yaw angle, yaw duration, and maximum yaw speed; Performance characteristics include energy consumption per unit angle of yaw or current fluctuation coefficient.

4. The method according to claim 1, characterized in that, The process of initially constructing several initial random forest models based on the training set and a preset combination of hyperparameters, and verifying the performance of each initial random forest model based on the validation set, and determining the optimal hyperparameter combination based on the validation results includes: For each set of preset hyperparameter combinations, a new sample set is generated on the training set by sampling with replacement. And based on the new sample set and the corresponding hyperparameter combinations, an initial random forest model is constructed. The performance of each initial random forest model is verified using the validation set, where the performance is the error between the predicted mean yaw motor current and the actual mean current. Repeat the above process multiple times to complete cross-validation. Based on the cross-validation results, obtain the average performance index of the random forest model corresponding to each hyperparameter combination. Then, take the hyperparameter combination with the best average performance index as the optimal hyperparameter combination.

5. The method according to claim 1, characterized in that, The optimal hyperparameters include the number of decision trees, the maximum number of features when each tree splits, the maximum depth of the decision tree, the minimum number of samples required for node splitting, and the minimum number of samples for leaf nodes. The process of reconstructing the random forest model using the optimal hyperparameter combination and the full original sample set to obtain the healthy state model includes: Based on the number of decision trees, construct a corresponding number of decision trees; For each decision tree, when splitting at each internal node of the decision tree, candidate features with a number not exceeding the maximum number of features are randomly selected from the multidimensional features; and the thresholds of each candidate feature are traversed; the feature and threshold that maximize the reduction in variance of the mean yaw motor current are selected for splitting. Splitting stops when any of the following conditions are met: the number of samples in the current node is less than the minimum number of samples required for the node to split, the number of samples in the leaf node is less than the minimum number of samples in the leaf node, or the depth of the current tree reaches the maximum depth of the decision tree. Once all decision trees are constructed, the health status model is formed.

6. A method for detecting abnormalities in the yaw friction pads of a wind turbine, characterized in that, include: Acquire real-time yaw event data; The yaw event data is preprocessed and features are extracted to obtain multidimensional input features; Using the health status model pre-constructed by any of the methods in claims 1-5, the multidimensional input features are predicted, and the average yaw motor current predicted under normal conditions is output. The abnormal state of the yaw friction plate is detected based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current.

7. The method according to claim 6, characterized in that, The method of detecting abnormal states of the yaw friction plates based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current includes: Based on multiple residual values ​​obtained within a preset window, the residual change trend and health index are determined. Based on the sign of each residual value, the residual change trend, and the accompanying current dynamic characteristics, the failure mode of the yaw friction plate is identified. Early warning levels are determined based on the numerical range of the health index, and corresponding early warning actions are triggered based on the results of the early warning levels.

8. The method according to claim 7, characterized in that, The method of identifying the fault modes of the yaw friction pad based on the sign of each residual value, the residual change trend, and the accompanying current dynamic characteristics includes: If the sign of the residual is consistently positive and shows a slow upward trend, it is determined to be a wear failure. If the sign of the residual changes to a negative value, accompanied by increased current fluctuations and intermittent slippage, it is determined to be a pollution fault. If the residual increases significantly after a long period of yaw and partially recovers after cooling, it is determined to be a thermal decay failure. If the residual performance is asymmetrical under different yaw directions, it is judged as an uneven wear fault.

9. A device for detecting abnormal yaw friction plates in wind turbine generators, characterized in that, include: The acquisition unit is used to acquire real-time yaw event data; The feature extraction unit is used to preprocess and extract features from the yaw event data to obtain multidimensional input features; The prediction unit is used to predict the multidimensional input features using a health state model pre-constructed by any of the methods in claims 1-5, and output the predicted average yaw motor current under normal conditions. The detection unit is used to detect abnormal conditions of the yaw friction plate based on the residual value between the predicted average yaw motor current and the actual observed average yaw motor current.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the method described in any one of claims 1-8.

Images