Abnormity monitoring method and system for wind power generation yaw system based on deep learning
By fusing multimodal data using deep learning methods and utilizing the Transformer model, the problem of insufficient utilization of multimodal data in traditional wind power yaw system monitoring methods is solved, achieving accurate monitoring and adaptive improvement of yaw system anomalies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG YUEDIAN ZHUHAI OFFSHORE WIND POWER CO LTD
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional wind power yaw system monitoring methods rely on fixed thresholds, making it difficult to integrate multimodal operating data, unable to identify periodic or gradual anomalies, and have weak time-series feature capture capabilities, making them unable to cope with dynamic operating conditions such as wind speed fluctuations and equipment aging.
A deep learning-based approach is adopted, which involves fusion of multimodal operating data into an improved pre-trained Transformer model to extract effective features and train the model, thereby constructing an anomaly detection model for yaw systems. This includes data cleaning, feature extraction, and model training, and the Transformer autoencoder and generative adversarial network are used for feature capture and anomaly identification.
It achieves accurate detection of yaw system anomalies, has strong generalization and scalability, can adapt to real-time monitoring under different operating conditions, and improves the ability to resist noise interference.
Smart Images
Figure CN121952809A_ABST
Abstract
Description
A Deep Learning-Based Anomaly Monitoring Method and System for Wind Power Yaw Systems Technical Field
[0001] This invention belongs to the field of wind power generation technology, specifically relating to an anomaly monitoring method and system for wind power generation yaw systems based on deep learning. Background Technology
[0002] The yaw system of a wind turbine, also known as a wind alignment device, is part of the wind turbine nacelle. Its function is to quickly and smoothly align with the wind direction when the wind speed vector changes, so that the wind turbine can obtain maximum wind energy. Abnormalities in the yaw system can lead to power generation loss, shortened equipment lifespan, or even safety accidents. Therefore, monitoring yaw system anomalies is particularly important. Traditional monitoring methods rely on fixed thresholds to detect single parameters such as vibration and temperature, failing to fully integrate multi-modal operating data such as wind speed, wind direction, and motor current. This makes it difficult to identify periodic or gradual anomalies, and suffers from insufficient utilization of multi-source data, weak ability to capture time-series features, and inability to handle various dynamic operating conditions such as wind speed fluctuations and equipment aging using fixed thresholds. Summary of the Invention
[0003] Based on the aforementioned problems in the existing technology, the purpose of this invention is to provide an anomaly monitoring method and system for wind power generation yaw systems based on deep learning. By fusing multimodal operating data and inputting it into an improved pre-trained Transformer model, the system can accurately capture abnormal situations in the yaw system, and the model has strong generalization ability and scalability.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: an anomaly monitoring method for a wind power generation yaw system based on deep learning, comprising: step S1, acquiring multimodal operation data of the yaw system of a wind turbine, preprocessing the multimodal operation data to obtain an initial dataset; step S2, extracting features that effectively distinguish the normal and abnormal states of the yaw system from the initial dataset, comparing the features, and obtaining the optimal features to obtain a target dataset; wherein, the features include: vibration features, temperature features, current and voltage features of the core components of the yaw system, and wind speed and direction correlation features with yaw action; step S3, dividing the target dataset into a training set, a validation set, and a test set according to a ratio; step S4, inputting the training set into a pre-trained Transformer model for model training, and improving the training parameters of the Transformer model based on the training results to obtain a yaw system anomaly monitoring model; step S5, inputting the validation set into the yaw system anomaly monitoring model for model evaluation; step S6, inputting the test set into the yaw system anomaly monitoring model to obtain yaw system anomaly monitoring results.
[0005] A further improvement of the present invention is that step S1 specifically includes: collecting the multimodal operation data of the yaw system through various sensors in the yaw system via the distributed acquisition system of the wind turbine generator; cleaning the multimodal operation data, the data cleaning including: noise processing, missing value processing and outlier processing; and normalizing and standardizing the cleaned multimodal operation data.
[0006] A further improvement of the present invention is that step S2 includes: the vibration characteristics include time-domain characteristics, frequency-domain characteristics, and time-series characteristics; the time-domain characteristics are mean, variance, peak value, peak-to-peak value, kurtosis, and skewness; the frequency-domain characteristics are power spectral density, peak frequency, and centroid frequency; the time-series characteristics are extracted through autocorrelation function and cross-correlation function; the temperature characteristics include temperature value, temperature change rate, and temperature difference between the motor and the bearing; the current and voltage characteristics include peak current, current fluctuation amplitude, voltage value, three-phase voltage difference, and current waveform distortion rate; the wind speed, wind direction, and yaw action correlation characteristics include yaw frequency, yaw angle, yaw response lag time, and the matching degree between yaw frequency and wind speed.
[0007] A further improvement of the present invention is that the feature comparison in step S2 includes: determining whether the yaw system is in an abnormal state by statistical threshold comparison, frequency domain energy distribution comparison, temporal correlation comparison, multi-feature fusion comparison, and physical rule verification comparison. When the feature exceeds the normal threshold range, the difference from the normal frequency domain energy distribution exceeds a set threshold, the autocorrelation coefficient is lower than the lower limit of the normal threshold range, or it does not conform to the physical operation rules, the yaw system is determined to be in an abnormal state.
[0008] A further improvement of the present invention is that step S4 includes: constructing a Transformer model based on a Transformer autoencoder and a generative adversarial network, and performing pre-training to obtain the pre-trained Transformer model; the pre-trained Transformer model includes an input layer, a feature extraction layer, a classification layer, an adversarial training layer, and a dynamic thresholding layer; the feature extraction layer embeds a modality adaptive embedding layer and a cross-modality self-attention mechanism, and the adversarial training layer introduces a generator and a discriminator.
[0009] A further improvement of this invention is that the input layer concatenates the data in the target dataset according to a preset time window; the feature extraction layer uses a Transformer autoencoder for feature extraction and captures the long-range dependencies between different sensor data through a self-attention mechanism to obtain representative global features; the classification layer constructs a multi-task fully connected layer to output the probabilities of different anomaly types in the yaw system; the adversarial training layer simulates abnormal data under extreme operating conditions of the yaw system through the generator, distinguishes between real data and generated data through the discriminator, and improves the ability of the pre-trained Transformer model to identify abnormal data by alternately optimizing the generator and the discriminator; the dynamic threshold layer adaptively adjusts the anomaly judgment criteria according to the real-time operating parameters of the yaw system to achieve accurate differentiation between normal and abnormal states of the yaw system under different operating scenarios.
[0010] A further improvement of the present invention is that the model training step in step S4 includes: using the PyTorch framework for model training, setting the learning rate to 0.001, the batch size to 32, and the number of training epochs to 50; using the cross-entropy loss function as the loss function for model training, and using the Adam optimizer to optimize the model parameters; during the model training process, monitoring the model performance in real time through the validation set, and stopping training when the model performance on the validation set no longer improves, to prevent model overfitting.
[0011] A further improvement of this invention is that the step of inputting the validation set into the yaw system anomaly monitoring model for model evaluation includes: inputting the validation set into the yaw system anomaly monitoring model, and evaluating it based on the precision P, recall R, and mean precision mAP of the yaw system anomaly monitoring model; the formula for evaluating model performance is as follows:
[0012]
[0013]
[0014] Where TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, and FP is the number of incorrectly predicted positive samples.
[0015] A deep learning-based anomaly monitoring system for wind power yaw systems includes: a data acquisition module for acquiring multimodal operational data of the wind power yaw system, preprocessing the multimodal operational data to obtain an initial dataset; a feature extraction module for extracting features from the initial dataset that effectively distinguish between the normal and abnormal states of the wind power yaw system, comparing the features, and obtaining the optimal features to obtain a target dataset; a dataset partitioning module for proportionally dividing the target dataset into a training set, a validation set, and a test set; a model training module for inputting the training set into a pre-trained Transformer model for model training, and improving the training parameters of the Transformer model based on the training results to obtain a yaw system anomaly monitoring model; a model validation module for inputting the validation set into the yaw system anomaly monitoring model for model evaluation; and a result generation and testing module for inputting the test set into the yaw system anomaly monitoring model to obtain yaw system anomaly monitoring results.
[0016] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned anomaly monitoring method for a deep learning-based wind power generation yaw system.
[0017] Compared with existing technologies, the present invention has at least the following beneficial technical effects: The present invention provides a method and system for anomaly monitoring of wind power generation yaw systems based on deep learning. The present invention includes acquiring multimodal operating data of the yaw system of a wind turbine and preprocessing it to obtain an initial dataset; extracting features from the initial dataset that effectively distinguish between normal and abnormal states of the yaw system, comparing these features, and obtaining the optimal features to obtain a target dataset; dividing the target dataset into a training set, a validation set, and a test set; inputting the training set into a pre-trained Transformer model for model training, and improving the training parameters of the Transformer model based on the training results to obtain a yaw system anomaly monitoring model; inputting the validation set into the yaw system anomaly monitoring model for model evaluation; and inputting the test set into the yaw system anomaly monitoring model to obtain the yaw system anomaly monitoring results. The anomaly monitoring method of the present invention, by fusing multimodal operating data and inputting it into an improved pre-trained Transformer model, can accurately capture abnormal situations of the yaw system, and the model has strong generalization ability and scalability. 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 is a flowchart illustrating the anomaly monitoring method for a wind power generation yaw system based on deep learning according to Embodiment 1 of the present invention; Figure 2 is a module diagram illustrating the anomaly monitoring system for a wind power generation yaw system based on deep learning according to Embodiment 2 of the present invention. Detailed Implementation
[0020] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.
[0021] In the description of this invention, it should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0022] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0023] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0024] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.
[0025] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0026] Example 1, as shown in Figure 1, provides an anomaly monitoring method for a wind turbine yaw system based on deep learning. The anomaly monitoring method includes: Step S1, acquiring multimodal operating data of the wind turbine's yaw system, and preprocessing the multimodal operating data to obtain an initial dataset. In this example, the distributed acquisition system of the wind turbine collects multimodal operating data of the yaw system through various sensors in the yaw system. The multimodal operating data includes vibration data, temperature data, wind speed and direction data, and current and voltage data. Vibration data refers to the vibration of components such as bearings and gearboxes during the operation of the yaw system, reflecting their health status. Vibration data can be collected by accelerometers installed on key parts such as the yaw bearings and gearboxes, with a sampling frequency of 100–500 Hz to capture rich vibration characteristics. Temperature data refers to the abnormal temperature rise that occurs in components such as motors and bearings during abnormal operation. Temperature sensors installed on heat-generating components such as motors and bearings can be used to measure temperature changes in real time. Wind speed and direction data refer to the impact of changes in wind speed and direction on the operating load and attitude of the yaw system, and are correlated with yaw system anomalies. This data can be obtained from existing weather stations at the wind farm, with sampling intervals of 1–10 seconds. Current and voltage data refer to the changes in current and voltage of the yaw system motor, reflecting the motor's operating status, such as the presence of overload or other anomalies. This data is collected from the yaw system motor's power supply circuit using current transformers and voltage sensors, with a sampling frequency consistent with that of the vibration data.
[0027] As an optional embodiment, the preprocessing of multimodal operating data in step S1 specifically includes: data cleaning of the multimodal operating data, which includes noise processing, missing value processing, and outlier processing; and normalization and standardization of the cleaned multimodal operating data. Specifically, noise processing refers to using wavelet transform denoising for high-frequency noise in vibration signals. This involves first performing wavelet decomposition on the vibration signal, then selecting an appropriate threshold to process the wavelet coefficients based on the noise distribution characteristics, and finally obtaining the denoised signal through wavelet reconstruction. For random noise in data such as temperature, current, and voltage, a moving average method is used for smoothing. The window size is determined based on the data sampling frequency and variation characteristics, generally ranging from 5 to 20 sampling points. Missing value processing refers to using mean imputation for continuous data when the missing data ratio is small (e.g., less than 5%), filling missing values with the mean of the parameter before and after the missing time period; and using linear interpolation or spline interpolation for data with strong time-series characteristics. When the proportion of missing data is large, a prediction model based on a long short-term memory network is used to fill in the missing values, combining the equipment's operating status and historical data. Outlier handling refers to identifying data values that exceed the mean plus or minus three standard deviations as outliers. For identified outliers, if they are obvious measurement errors, they are replaced with valid values from adjacent time points; if they may be genuine abnormal states, the data is retained as the basis for subsequent anomaly analysis. In addition, to eliminate the dimensional differences between different data types and improve the training efficiency and accuracy of the model, step S1 of this embodiment normalizes or standardizes the preprocessed dataset; specifically, for data such as vibration, current, and voltage that follow a normal distribution, standardization is used to convert the data into normally distributed data with a mean of 0 and a standard deviation of 1; for data such as temperature with a relatively fixed value range, normalization is used to map the data to the interval [0, 1].
[0028] Step S2: Extract features from the initial dataset that effectively distinguish between the normal and abnormal states of the yaw system, compare the features, and obtain the optimal features to obtain the target dataset. In this embodiment, under normal conditions, the components of the wind power yaw system operate in coordination, and the parameter changes show a stable regularity. Under abnormal conditions, the parameters of the yaw system deviate from the normal range and exhibit characteristics that are significantly different from those under normal conditions. The features that can effectively distinguish between the normal and abnormal states include the vibration characteristics, temperature characteristics, current and voltage characteristics of the core components of the yaw system, as well as the correlation characteristics between wind speed and wind direction and yaw action. Vibration characteristics include time-domain characteristics, frequency-domain characteristics, and time-series characteristics. Time-domain characteristics include mean, variance, peak value, peak-to-peak value, kurtosis, and skewness. Frequency-domain characteristics include power spectral density, peak frequency, and centroid frequency. Time-series characteristics are obtained by extracting the autocorrelation coefficient through autocorrelation and cross-correlation functions. Temperature characteristics include temperature value, temperature change rate, and temperature difference between the motor and bearing. Current and voltage characteristics include peak current, current fluctuation amplitude, voltage value, three-phase voltage difference, and current waveform distortion rate. Wind speed, wind direction, and yaw action correlation characteristics include yaw frequency, yaw angle, yaw response lag time, and the matching degree between yaw frequency and wind speed. Among the vibration characteristics, peak value, kurtosis, frequency peak value, and autocorrelation coefficient are particularly crucial. Under normal conditions, the yaw system exhibits low and stable peak and kurtosis values, no obvious peaks in the frequency domain, and a high autocorrelation coefficient. However, under abnormal conditions, the peak and kurtosis values increase significantly, peaks appear at the characteristic frequencies corresponding to the faulty components, and the autocorrelation coefficient decreases. These distinguishing characteristics of vibration can effectively reflect impact faults in the core components of the yaw system, such as gearbox jamming or bearing ball wear, accurately identifying specific fault types and indicating whether the periodicity of the vibration signal has been disrupted. Temperature characteristics, such as the rate of temperature change and the temperature difference between the motor and bearing, are also critical. Under normal conditions, the rate of temperature change is small, and the temperature difference between the motor and bearing is stable. However, under abnormal conditions, the rate of temperature change increases significantly, and the temperature difference between the motor and bearing becomes abnormal. These distinguishing characteristics of temperature can promptly detect overheating faults in the motor or bearing and reflect abnormal operation of the bearing or motor. The peak current and fluctuation amplitude, as well as the current waveform distortion rate, are crucial characteristics of the current and voltage characteristics. When the yaw system is in a normal state, the peak current and fluctuation amplitude are within a certain range, and the current waveform distortion rate is low. However, when it is in an abnormal state, the peak current and fluctuation amplitude exceed the range, and the current waveform distortion rate increases. These distinguishing characteristics of the current and voltage characteristics can reflect whether the motor is overloaded or has a circuit fault, as well as whether the motor's electrical performance is abnormal.The correlation characteristics between wind speed and direction and yaw action, particularly the matching degree between yaw frequency and wind speed, and the yaw response lag time, are crucial. When the yaw system is in a normal state, the matching degree between yaw frequency and wind speed follows a certain pattern, and the yaw response lag time is short. However, when in an abnormal state, the matching degree between yaw frequency and wind speed is low, and the yaw response lag time is prolonged. These distinguishing characteristics of the correlation characteristics between wind speed and direction and yaw action can reflect whether the yaw system's response is normal and whether there are any control problems.
[0029] As an optional embodiment, the feature comparison in step S2 includes: determining whether the yaw system is in an abnormal state by comparing statistical thresholds, frequency domain energy distribution, temporal correlation, multi-feature fusion, and physical rule verification. When the features exceed the normal threshold range, the difference from the normal frequency domain energy distribution exceeds the set threshold, the autocorrelation coefficient is lower than the lower limit of the normal threshold range, or the physical operation rules are not met, the yaw system is determined to be in an abnormal state.
[0030] In this embodiment, the statistical threshold comparison is as follows: For features such as the mean, peak value, and kurtosis of vibration, the value and rate of change of temperature, and the peak value of current, the range and distribution pattern of values under normal conditions are first statistically analyzed to determine the normal threshold range; when the feature value extracted in real time exceeds the normal threshold range, it is judged as abnormal. For example, the normal range of vibration peak value is 0-5 m / s², and if the real-time peak value is 8 m / s², it is judged as abnormal. The frequency domain energy distribution comparison is as follows: The energy proportion of each frequency band under normal conditions is calculated to establish a normal frequency domain energy distribution model. The real-time vibration signal is subjected to Fourier transform to obtain its frequency domain energy distribution, which is compared with the normal model to calculate the difference in energy distribution; when the difference exceeds a set threshold, it is judged as abnormal. For example, under normal conditions, the energy proportion of the 50-200 Hz frequency band is 10%, while the real-time proportion is 30%, which is a large difference and is judged as abnormal. The temporal correlation comparison involves calculating the autocorrelation coefficient of adjacent periods of the vibration signal under normal conditions to determine the normal correlation range; calculating the autocorrelation coefficient of adjacent periods of the real-time vibration signal and comparing it with the normal range; if the autocorrelation coefficient is lower than the lower limit of the normal range, it is judged as abnormal. For example, if the normal autocorrelation coefficient is 0.8-1.0 and the real-time value is 0.4, it is judged as abnormal. The multi-feature fusion comparison involves using machine learning algorithms to establish a classification model for normal and abnormal states, using multiple key features as input; extracting real-time multi-feature data, inputting it into the classification model, obtaining the classification result, and determining whether it is normal or abnormal. For example, inputting features such as vibration peak value, temperature change rate, and current fluctuation amplitude into the model, the model outputs an abnormal probability; if the probability exceeds 0.8, it is judged as abnormal. The physical rule verification comparison involves establishing the correlation between features according to the physical operating rules of the yaw system, such as the correspondence between wind speed and yaw frequency, and the coordinated change relationship between motor current and temperature; monitoring in real time whether the correlation between features conforms to the physical rules, and if not, it is judged as abnormal. For example, when the wind speed is 10 m / s, the normal yaw frequency is 3 to 4 times every 10 minutes. If the real-time yaw frequency is 10 times, it does not conform to the physical rules and is judged as abnormal.
[0031] Step S3: Divide the target dataset into a training set, a validation set, and a test set proportionally. In this embodiment, the target dataset is randomly divided into the training set, the validation set, and the test set in a 7:2:1 ratio.
[0032] Step S4: Input the training set into the pre-trained Transformer model for model training, and improve the training parameters of the Transformer model based on the training results to obtain the yaw system anomaly monitoring model.
[0033] As an optional embodiment, step S4 includes: constructing a Transformer model based on Transformer autoencoder and Generative Adversarial Networks (GAN), and pre-training it to obtain a pre-trained Transformer model; the pre-trained Transformer model includes an input layer, a feature extraction layer, a classification layer, an adversarial training layer, and a dynamic thresholding layer; the feature extraction layer embeds a modality adaptive embedding layer and a cross-modality self-attention mechanism, and the adversarial training layer introduces a generator and a discriminator.
[0034] In this embodiment, the input layer concatenates the data in the target dataset according to a preset time window; the feature extraction layer uses a Transformer autoencoder to extract features and captures the long-range dependencies between different sensor data through a self-attention mechanism to obtain representative global features; the classification layer constructs a multi-task fully connected layer to output the probability of different anomaly types in the yaw system; the adversarial training layer simulates abnormal data under extreme operating conditions of the yaw system through a generator, distinguishes between real data and generated data through a discriminator, and improves the ability of the pre-trained Transformer model to identify abnormal data by alternately optimizing the generator and discriminator; the dynamic threshold layer adaptively adjusts the anomaly judgment criteria according to the real-time operating parameters of the yaw system to achieve accurate differentiation between normal and abnormal states of the yaw system under different operating scenarios.
[0035] As an optional embodiment, the model training step in step S4 includes: training the model using the PyTorch framework, setting the learning rate to 0.001, the batch size to 32, and the number of training epochs to 50; using the cross-entropy loss function as the loss function for model training, and using the Adam optimizer to optimize the model parameters.
[0036] Step S5: Input the validation set into the yaw system anomaly monitoring model for model evaluation.
[0037] Step S6: Input the test set into the yaw system anomaly monitoring model to obtain the yaw system anomaly monitoring results.
[0038] In this embodiment, during model training, the model's performance is monitored in real time using a validation set. Training is stopped when the model's performance on the validation set no longer improves to prevent overfitting. As an optional embodiment, step S5, which involves inputting the validation set into the yaw system anomaly monitoring model for model evaluation, includes: inputting the validation set into the yaw system anomaly monitoring model and evaluating it based on the model's precision P, recall R, and mean precision mAP. The formulas for evaluating model performance are as follows:
[0039]
[0040]
[0041] Where TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, and FP is the number of incorrectly predicted positive samples.
[0042] This embodiment provides a deep learning-based anomaly monitoring method for wind power generation yaw systems. It fuses multimodal data and inputs it into a yaw system anomaly monitoring model trained collaboratively by an improved Transformer autoencoder and a GAN. A dynamic threshold layer is then used to adaptively adapt to different operating conditions of the yaw system, enabling precise capture of subtle anomalies. Furthermore, considering the complex environment of wind farms with fluctuating wind speeds and directions, a physical constraint generator simulates extreme conditions, and the dynamic threshold layer adjusts the judgment criteria in real time. This ensures the yaw system anomaly monitoring model can operate stably even in special scenarios such as low wind speeds and high turbulence, improving its resistance to noise interference.
[0043] As shown in Figure 2, this embodiment provides an anomaly monitoring system 10 for a wind power generation yaw system based on deep learning. The anomaly monitoring system includes a data acquisition module 11, a feature extraction module 12, a dataset partitioning module 13, a model training module 14, a model verification module 15, and a result generation and testing module 16.
[0044] The data acquisition module 11 is used to acquire multimodal operation data of the wind power generation yaw system, and preprocess the multimodal operation data to obtain an initial dataset.
[0045] The feature extraction module 12 is used to extract features that effectively distinguish the normal and abnormal states of the wind power generation yaw system from the initial dataset, and to compare the features to obtain the optimal features to obtain the target dataset.
[0046] The dataset partitioning module 13 is used to divide the target dataset into a training set, a validation set, and a test set according to a certain ratio.
[0047] The model training module 14 is used to input the training set into the pre-trained Transformer model for model training, and improve the training parameters of the Transformer model based on the training results to obtain the yaw system anomaly monitoring model.
[0048] The model validation module 15 is used to input the validation set into the yaw system anomaly monitoring model for model evaluation.
[0049] The result generation test module 16 is used to input the test set into the yaw system anomaly monitoring model to obtain the yaw system anomaly monitoring results.
[0050] This embodiment provides an anomaly monitoring system for wind power generation yaw systems based on deep learning. This anomaly monitoring system is based on the deep learning-based anomaly monitoring method for wind power generation yaw systems proposed in Embodiment 1 above; specific details are not repeated here. This embodiment's deep learning-based anomaly monitoring system for wind power generation yaw systems fuses multimodal data and inputs it into a yaw system anomaly monitoring model obtained by co-training an improved Transformer autoencoder and GAN. It also incorporates a dynamic threshold layer to adaptively adapt to different operating conditions of the yaw system, achieving accurate capture of subtle anomalies. Simultaneously, considering the complex environment of wind farms with fluctuating wind speed and direction, it simulates extreme conditions through a physical constraint generator and adjusts the judgment criteria in real time using a dynamic threshold layer. This ensures the yaw system anomaly monitoring model can still operate stably in special scenarios such as low wind speed and high turbulence, improving its resistance to noise interference.
[0051] Example 3 This example relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the deep learning-based anomaly monitoring method for wind power generation yaw systems described in Example 1 above.
[0052] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0053] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0054] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.
Claims
1. A deep learning-based anomaly monitoring method for wind power generation yaw systems, characterized in that, include: Step S1: Obtain multimodal operation data of the yaw system of the wind turbine, and preprocess the multimodal operation data to obtain an initial dataset; Step S2: Extract features from the initial dataset that effectively distinguish the normal and abnormal states of the yaw system, compare the features, and obtain the optimal features to obtain the target dataset; wherein, the features include: vibration characteristics, temperature characteristics, current and voltage characteristics of the core components of the yaw system, and wind speed and direction correlation characteristics with yaw action; Step S3: Divide the target dataset into a training set, a validation set, and a test set proportionally; Step S4: Input the training set into a pre-trained Transformer model for model training, and improve the training parameters of the Transformer model based on the training results to obtain a yaw system anomaly monitoring model; Step S5: Input the validation set into the yaw system anomaly monitoring model for model evaluation; Step S6: Input the test set into the yaw system anomaly monitoring model to obtain the yaw system anomaly monitoring results.
2. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 1, characterized in that, Step S1 further includes: collecting the multimodal operation data of the yaw system through various sensors in the yaw system via the distributed acquisition system of the wind turbine; cleaning the multimodal operation data, including noise processing, missing value processing and outlier processing; and normalizing and standardizing the cleaned multimodal operation data.
3. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 1, characterized in that, Step S2 includes: the vibration characteristics include time-domain characteristics, frequency-domain characteristics, and time-series characteristics; the time-domain characteristics are mean, variance, peak value, peak-to-peak value, kurtosis, and skewness; the frequency-domain characteristics are power spectral density, peak frequency, and centroid frequency; the time-series characteristics are extracted using autocorrelation and cross-correlation functions; the temperature characteristics include temperature value, temperature change rate, and temperature difference between the motor and bearing; the current and voltage characteristics include peak current, current fluctuation amplitude, voltage value, three-phase voltage difference, and current waveform distortion rate; the wind speed, wind direction, and yaw action correlation characteristics include yaw frequency, yaw angle, yaw response lag time, and the matching degree between yaw frequency and wind speed.
4. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 3, characterized in that, The feature comparison in step S2 includes: determining whether the yaw system is in an abnormal state by comparing statistical thresholds, frequency domain energy distribution, temporal correlation, multi-feature fusion, and physical rule verification. When the feature exceeds the normal threshold range, the difference from the normal frequency domain energy distribution exceeds the set threshold, the autocorrelation coefficient is lower than the lower limit of the normal threshold range, or it does not conform to the physical operation rules, the yaw system is determined to be in an abnormal state.
5. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 1, characterized in that, Step S4 includes: constructing a Transformer model based on a Transformer autoencoder and a generative adversarial network, and pre-training it to obtain the pre-trained Transformer model; the pre-trained Transformer model includes an input layer, a feature extraction layer, a classification layer, an adversarial training layer, and a dynamic thresholding layer; the feature extraction layer embeds a modality adaptive embedding layer and a cross-modality self-attention mechanism, and the adversarial training layer introduces a generator and a discriminator.
6. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 4, characterized in that, The input layer concatenates the data in the target dataset according to a preset time window; the feature extraction layer uses a Transformer autoencoder to extract features and captures the long-range dependencies between different sensor data through a self-attention mechanism to obtain representative global features. The classification layer constructs a multi-task fully connected layer, outputting the probabilities of different anomaly types in the yaw system; the adversarial training layer simulates abnormal data under extreme operating conditions of the yaw system through the generator, distinguishes between real data and generated data through the discriminator, and improves the ability of the pre-trained Transformer model to identify abnormal data by alternately optimizing the generator and the discriminator; the dynamic threshold layer adaptively adjusts the anomaly judgment criteria according to the real-time operating parameters of the yaw system to achieve accurate differentiation between normal and abnormal states of the yaw system under different operating scenarios.
7. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 6, characterized in that, The model training steps in step S4 include: using the PyTorch framework for model training, setting the learning rate to 0.001, the batch size to 32, and the number of training epochs to 50; using the cross-entropy loss function as the loss function for model training, and using the Adam optimizer to optimize the model parameters; during the model training process, monitoring the model performance in real time through the validation set, and stopping training when the model performance on the validation set no longer improves to prevent overfitting.
8. The anomaly monitoring method for a wind power generation yaw system based on deep learning according to claim 1, characterized in that, The step of inputting the validation set into the yaw system anomaly monitoring model for model evaluation includes: inputting the validation set into the yaw system anomaly monitoring model, and evaluating it based on the precision P, recall R, and mean precision mAP of the yaw system anomaly monitoring model; the formula for evaluating model performance is as follows: Where TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, and FP is the number of incorrectly predicted positive samples.
9. An anomaly monitoring system for a wind power generation yaw system based on deep learning, characterized in that, include: The data acquisition module is used to acquire multimodal operation data of the wind power generation yaw system, and preprocess the multimodal operation data to obtain an initial dataset; The feature extraction module is used to extract features from the initial dataset that effectively distinguish the normal and abnormal states of the wind power yaw system, and to compare the features to obtain the optimal features to obtain the target dataset. The dataset partitioning module is used to divide the target dataset into training set, validation set and test set according to a certain ratio; The model training module is used to input the training set into the pre-trained Transformer model for model training, and improve the training parameters of the Transformer model based on the training results to obtain the yaw system anomaly monitoring model. The model validation module is used to input the validation set into the yaw system anomaly monitoring model for model evaluation; the result generation test module is used to input the test set into the yaw system anomaly monitoring model to obtain the yaw system anomaly monitoring results.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the anomaly monitoring method for a wind power generation yaw system based on deep learning, as described in any one of claims 1 to 8.