Wind driven generator fault prediction method and system and storage medium

Abstract

The invention discloses a wind driven generator fault prediction method and system, and a storage medium. The method comprises the following steps: obtaining a sample historical data set obtained by preprocessing multi-source historical data of a target wind driven generator; obtaining a sample expansion data set obtained by performing data expansion processing through the wind driven generator physical model; a dynamic time sequence attention mechanism is introduced into a feature extraction module of the pre-trained neural network, and a confidence coefficient weighting mechanism is introduced into a prediction module of the neural network; obtaining a first feature and a second feature corresponding to the sample historical data set and the sample expansion data set; weighting the first feature and the second feature through a confidence weighting mechanism to output a corresponding first prediction result and a corresponding second prediction result; and inputting the real-time sample data into the fault prediction model, and outputting a fault prediction result of the target wind driven generator. Therefore, the problems of few training receipts, lack of real fault data, single data source and the like in the prior art are solved, and the fault prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a method, system and storage medium for predicting wind turbine faults. Background Technology

[0002] As a crucial component of clean energy, the stable operation of wind power equipment is vital to the reliability of the power system. Wind turbines operate in complex natural environments, susceptible to fluctuations in wind speed, temperature changes, mechanical wear, and other factors, making them prone to malfunctions and shutdowns, resulting in significant economic losses. Therefore, predicting wind turbine failures is essential.

[0003] Currently, deep learning technology has begun to be applied to wind turbine fault prediction, and it has shown great potential in the field of fault prediction. However, in practical applications, existing technologies have limited training data, relying only on historical operating data and historical fault data of wind turbines for research and prediction. Real fault data of wind turbines is scarce, and training prediction models based solely on limited historical data leads to low prediction accuracy in complex operating conditions or rare fault scenarios. Moreover, the single data source cannot comprehensively capture the operating status characteristics of wind turbines, affecting the accuracy of fault prediction. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a method for predicting wind turbine faults.

[0005] One of the objectives of this invention is achieved through the following technical solution: a method for predicting faults in a wind turbine generator, the method comprising the following steps:

[0006] Step S1: Obtain multi-source historical data of the target wind turbine, and preprocess the multi-source historical data to obtain a sample historical dataset;

[0007] Step S2: Perform data augmentation processing on the historical sample dataset using a wind turbine physical model to obtain an augmented sample dataset;

[0008] Step S3: Introduce a dynamic temporal attention mechanism into the feature extraction module of the pre-trained neural network, and introduce a confidence weighting mechanism into the prediction module of the neural network.

[0009] Step S4: Input the historical sample dataset and the expanded sample dataset into the feature extraction module, and extract the data features of the historical sample dataset and the expanded sample dataset by allocating the weights of features in different time windows through the dynamic temporal attention mechanism to obtain the first feature and the second feature;

[0010] Step S5: Input the first feature and the second feature into the prediction module respectively, and perform weighted processing on the first feature and the second feature through the confidence weighting mechanism to output the corresponding first prediction result and second prediction result;

[0011] Step S6: Calculate the loss value between the first prediction result and the second prediction result and the corresponding actual result, and adjust the parameters of the neural network according to the loss value to obtain the trained wind turbine fault prediction model.

[0012] Step S7: Obtain real-time multi-source data of the target wind turbine, preprocess the real-time multi-source data to obtain real-time sample data, input the real-time sample data into the trained wind turbine fault prediction model, and output the fault prediction result of the target wind turbine.

[0013] Furthermore, the multi-source historical data in step S1 includes at least the deformation data, vibration data, operation data, environmental dynamic data, and grid interaction data of each blade of the target wind turbine within a preset time period;

[0014] It also includes data cleaning and data standardization processing of the operational data;

[0015] The data cleaning method uses a combination of the 3σ criterion and the isolated forest algorithm to remove noise and outliers;

[0016] The data standardization adopts the z-score standardization method, and the formula is:

[0017] ;

[0018] in, This is the original data. The mean of the dimensional data. denoted as the standard deviation of the data in that dimension.

[0019] Furthermore, step S2 also includes:

[0020] Based on the blade aerodynamic load model, transmission chain dynamics model, and material fatigue damage model, and combined with the deformation data and vibration data of each blade, the theoretical stress distribution under different wind speeds and rotational speeds is calculated; the failure evolution process of blade fatigue crack propagation and bearing wear is simulated, and sample augmentation data that conforms to physical laws is generated to form the sample augmentation dataset.

[0021] Furthermore, the blade aerodynamic load model is based on fluid mechanics principles and combines the wind speed and rotational speed in the sample historical dataset to calculate the aerodynamic load distribution of each section of the blade under different operating conditions, thereby deriving the blade deformation and vibration characteristics.

[0022] The transmission chain dynamics model is based on the principle of multibody dynamics, simulating the meshing clearance, friction torque and vibration transmission path of gearbox and bearings at different speeds. Combined with the vibration characteristics in the historical data of the sample, it generates extended data that conforms to the correlation between vibration, vibration frequency and component characteristics.

[0023] The material fatigue damage model is based on the fatigue accumulation theory. It combines the deformation data and vibration data in the historical data of the samples to simulate the entire life cycle damage process of the material from no damage to minor damage and then to obvious failure, and generates the extended data containing the complete failure evolution stages.

[0024] Furthermore, step S4 also includes:

[0025] Step S41: Align the historical sample dataset and the expanded sample dataset in chronological order to form two parallel time-series datasets;

[0026] Step S42: Divide the two sets of time series datasets into continuous fixed-length windows, each fixed-length window containing multi-source data within the corresponding time period; the fixed-length windows are dynamically adjusted according to the fault types in the time series datasets;

[0027] Step S43: Calculate the weight of each fixed-length window using the dynamic temporal attention mechanism, and divide the weight values ​​accordingly;

[0028] Step S44: Extract basic features from the multi-source data within each fixed-length window, and sum the extracted basic features according to their corresponding weight values ​​to obtain the global features of each time series dataset, thereby generating the corresponding first feature and second feature.

[0029] Furthermore, step S5 also includes:

[0030] Step S51: Assign a credibility evaluation index to each data feature in the first feature and the second feature;

[0031] Step S52: Based on the credibility assessment index, assign fault allocation weights to the first feature and the second feature according to preset weights;

[0032] Step S53: Input the weighted first feature and second feature into the prediction module, and output the first prediction result and the second prediction result of the fault probability distribution corresponding to the real data.

[0033] Furthermore, step S6 also includes:

[0034] The loss value is backpropagated to each layer of the neural network using the gradient descent algorithm. The weights of the time window features in the feature extraction module, the neural network convolution parameters, and the weighting coefficients of the confidence weighting mechanism in the prediction module are adjusted to optimize the loss value in a decreasing direction.

[0035] Furthermore, step S6 also includes:

[0036] The trained wind turbine fault prediction model is validated using a validation set. The validation set is then input into the trained wind turbine fault prediction model for model performance evaluation. If the test performance fails to meet expectations, the time window allocation of the dynamic temporal attention mechanism, the initial coefficients of the confidence weighting mechanism, or the parameters of the loss function are adjusted, and the model is retrained until the model converges. Otherwise, the relevant parameters of the model training are saved.

[0037] Compared with existing technologies, the beneficial effects of this invention are as follows: This embodiment provides a method for predicting the faults of wind turbines. It expands upon historical sample data of wind turbines using a physical model of the wind turbine, comprehensively capturing the operating state characteristics of the wind turbine, ensuring the authenticity and completeness of the wind turbine fault training data, and overcoming the problems of limited training data, lack of real fault data, and single data source in existing technologies. By improving the neural network, a dynamic temporal attention mechanism is introduced into the feature extraction module. By automatically assigning weights to different time windows, focusing on key temporal features strongly correlated with faults, the effectiveness of temporal features can be significantly improved. A confidence weighting mechanism is introduced into the prediction module, dynamically adjusting weights based on the real-time credibility of the data source, which can weaken the interference of low-quality data and improve prediction accuracy and reliability. Training the wind turbine fault prediction model using the improved neural network results in a model with good reliability, interpretability, transferability, and generalization ability.

[0038] To overcome the shortcomings of the prior art, the second objective of this invention is to provide a wind turbine generator fault prediction system.

[0039] The second objective of this invention is achieved by the following technical solution: a wind turbine generator fault prediction system, the fault prediction system comprising:

[0040] The historical data acquisition module is used to acquire multi-source historical data of the target wind turbine, and preprocess the multi-source historical data to obtain a sample historical dataset.

[0041] An expanded data acquisition module is used to perform data expansion processing on the sample historical dataset using a wind turbine physical model to obtain an expanded sample dataset.

[0042] A neural network construction module is used to introduce a dynamic temporal attention mechanism into the feature extraction module of a pre-trained neural network and a confidence weighting mechanism into the prediction module of the neural network.

[0043] The data feature extraction module is used to input the historical sample dataset and the expanded sample dataset into the feature extraction module, and to extract the data features of the historical sample dataset and the expanded sample dataset by allocating the weights of features in different time windows through the dynamic temporal attention mechanism, so as to obtain the first feature and the second feature.

[0044] The feature weighting processing module is used to input the first feature and the second feature into the prediction module respectively, and to perform weighting processing on the first feature and the second feature through the confidence weighting mechanism to output the corresponding first prediction result and second prediction result.

[0045] The prediction model training module is used to calculate the loss value between the first prediction result and the second prediction result and the corresponding actual result through a loss function, and adjust the parameters of the neural network according to the loss value through a backpropagation algorithm to complete the training of the neural network and obtain the trained wind turbine fault prediction model.

[0046] The prediction result generation module is used to acquire real-time multi-source data of the target wind turbine, preprocess the real-time multi-source data to obtain real-time sample data, input the real-time sample data into the trained wind turbine fault prediction model, and output the state prediction result of the target wind turbine.

[0047] To overcome the shortcomings of the prior art, the third objective of this invention is to provide a storage medium.

[0048] The third objective of this invention is achieved by the following technical solution: a storage medium storing a computer program, which, when executed by a processor, implements the wind turbine generator fault prediction method. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the wind turbine fault prediction method in Embodiment 1 of the present invention.

[0050] Figure 2 This is a flowchart illustrating step S4 of the wind turbine fault prediction method in Embodiment 1 of the present invention.

[0051] Figure 3 This is a flowchart illustrating step S5 of the wind turbine generator fault prediction method in Embodiment 1 of the present invention.

[0052] Figure 4 This is a block diagram of the control module of the wind turbine fault prediction system in Embodiment 2 of the present invention.

[0053] In the picture:

[0054] 10. Fault prediction system; 11. Historical data acquisition module; 12. Extended data acquisition module; 13. Neural network component module; 14. Data feature extraction module; 15. Feature weighting processing module; 16. Prediction model training module; 17. Prediction result generation module. Detailed Implementation

[0055] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0056] Example 1

[0057] like Figure 1-3 As shown, a method for predicting faults in a wind turbine includes the following steps:

[0058] Step S1: Obtain multi-source historical data of the target wind turbine, preprocess the multi-source historical data to obtain a sample historical dataset; wherein, the multi-source historical data includes deformation data, vibration data, operation data, environmental dynamic data and grid interaction data of each blade of the target wind turbine within a preset time period.

[0059] In this embodiment, the historical data of the prior art is mainly focused on blade deformation data and vibration data, which greatly limits the training data. In contrast, this embodiment adds wind power generation operation data, environmental dynamic data and grid interaction data to the blade deformation data and vibration data.

[0060] The operating data of wind turbines includes blade speed, output power, gearbox oil temperature, generator stator temperature, and hydraulic system pressure. This data directly reflects the working status of the core components of the wind turbine equipment. For example, an abnormally high gearbox oil temperature may indicate lubrication failure or gear wear, complementing existing blade deformation and vibration data. For instance, combining blade deformation data with speed can determine whether the deformation is caused by overspeed load, while combining vibration data with oil temperature can distinguish whether mechanical vibration originates from component wear or overheating deformation.

[0061] Environmental dynamic data includes real-time wind speed, wind direction, air humidity, and atmospheric pressure. Wind turbine failures are strongly correlated with the environment. For example, high wind speeds can cause a surge in blade load, easily leading to fatigue cracks, and humid environments accelerate the corrosion of electrical components. Existing technologies do not consider these factors, potentially causing failure predictions to overlook external triggers. In this embodiment, by spatiotemporally correlating environmental data with equipment data such as wind turbine blade deformation and vibration data, and coupling the wind speed change rate with the blade deformation rate, the ability to predict failures under extreme operating conditions can be improved.

[0062] Grid interaction data includes data such as grid-connected voltage, current fluctuations, and changes in active / reactive power. Abnormal interactions between wind turbines and the grid, such as a sudden voltage surge causing overvoltage in the wind turbine, can lead to equipment failure. Current technologies do not address this type of data, making it difficult to detect coupled faults between the grid and wind turbine equipment, such as converter overload.

[0063] Therefore, by integrating grid interaction data into the input parameters of the model training, targeted prediction of grid-related faults can be achieved. This not only focuses on the fault factors of the equipment itself but also addresses the potential grid interaction faults of wind turbines as part of the grid, which is often overlooked in existing technologies, thus improving the accuracy of fault prediction. Through multi-dimensional data fusion and supplementing with state data during the interaction between the wind turbine and the grid, the fault prediction model can more accurately identify faults and more comprehensively reflect the causes of wind turbine equipment failures under complex operating conditions.

[0064] Furthermore, the step of preprocessing multi-source historical data in step S1 to obtain the sample historical dataset also includes: data cleaning and data standardization of the running data; wherein, the data cleaning adopts a method based on the combination of the 3σ criterion (also known as the Laida criterion) and the isolated forest algorithm to remove noise and outliers.

[0065] More specifically, for normally distributed data such as blade deformation, vibration, and environmental dynamics, the 3σ criterion is used to identify and remove data exceeding the mean ± 3 times the standard deviation. For non-normally distributed data such as operational data and power grid interaction data, the Isolation Forest algorithm is used to detect and remove outliers. Data standardization employs the Z-score standardization method, which accurately reflects the relative standard distance of a score from the mean; it is the process of dividing the difference between a number and the mean by the standard deviation.

[0066] Specifically, the Z-score is calculated by measuring the distance of a raw score from the mean using the standard deviation. The Z-score is determined by the number of standard deviations within this distance. The Z-score is a very important indicator. When the raw scores are normally distributed, transforming all raw scores into Z-scores creates a standard normal distribution, thus determining the position of the data point within the entire dataset. This process is called standardization. The formula for Z-score transformation is:

[0067] ;

[0068] in, This is the original data. The mean of the dimensional data. denoted as the standard deviation of the data in that dimension.

[0069] Step S2: The historical sample dataset is augmented using a wind turbine physical model to obtain an augmented dataset. Existing technologies obtain relatively scarce historical data, relying solely on limited historical data to train prediction models, resulting in low prediction accuracy under complex operating conditions or rare fault scenarios. In this embodiment, to address the scarcity of real fault data for wind turbines, effective data augmentation is performed by introducing a wind turbine physical model to augment the historical sample dataset.

[0070] Furthermore, step S2, which involves augmenting the historical sample dataset using a wind turbine physical model to obtain an augmented sample dataset, also includes:

[0071] Based on the blade aerodynamic load model, transmission chain dynamics model, and material fatigue damage model, combined with the deformation and vibration data of each blade, the theoretical stress distribution under different wind speeds and rotational speeds is calculated. The failure evolution process of blade fatigue crack propagation and bearing wear is simulated, generating sample augmented data that conforms to physical laws, forming a sample augmented dataset.

[0072] Furthermore, the deformation and vibration of wind turbine blades are essentially the result of aerodynamic loads. Changes in wind speed and direction directly alter the lift, drag, and torque experienced by the blades, thus affecting their bending deformation. In this embodiment, a blade aerodynamic load model accurately correlates the causal relationships between the external environment, blade forces and deformation, and vibration. Specifically, based on fluid mechanics principles and combined with wind speed and rotational speed from historical sample datasets, the blade aerodynamic load model calculates the aerodynamic load distribution across various blade sections under different operating conditions, thereby deriving the blade deformation and vibration characteristics. This ensures that the generated extended data accurately reflects the intrinsic correlation between wind speed, load, deformation, and vibration, ensuring that blade-related fault characteristics, such as precursors to blade cracks caused by excessive aerodynamic loads, conform to physical logic.

[0073] A large proportion of the vibration data of wind turbines comes from the operation of components such as gearboxes, main shafts, and bearings. Gear meshing and the contact between bearing rolling elements and raceways generate periodic vibrations, the frequency of which is directly related to the rotational speed. That is, the gear meshing frequency is the product of the number of teeth and the rotational speed. In this embodiment, a transmission chain dynamics model is used to ensure the mechanical rationality of the vibration characteristics of the transmission system. Specifically, based on the principle of multibody dynamics, the meshing clearance, frictional torque, and vibration transmission path of the gearbox and bearings at different rotational speeds are simulated. Combined with the vibration characteristics in the historical sample data, extended data that conforms to the correlation between "vibration-vibration frequency-component characteristics" is generated.

[0074] Core failures of wind turbines, such as blade fatigue cracks and bearing wear, are mostly the result of cumulative damage to component materials under cyclic loading. Their evolution exhibits a clear temporal sequence; for example, blades progress from minor deformation to crack propagation, and bearings from slight wear to increased vibration. In this embodiment, a material fatigue damage model is used to recreate the material evolution process and its patterns. Specifically, based on fatigue accumulation theory and incorporating deformation and vibration data from historical sample data, the model simulates the entire lifecycle damage process of materials, from no damage to minor damage and then to significant failure, generating expanded data encompassing all stages of the failure evolution.

[0075] Therefore, the sample augmentation data generated by these three wind turbine physical models can not only be consistent with the blade deformation and vibration characteristics in the historical sample data, but also meet the mechanical stress balance and kinematic speed and vibration frequency matching constraints of wind turbine equipment operation, thereby improving the effectiveness of the model training dataset and making the fault characteristics learned by the neural network closer to the actual operating rules.

[0076] Step S3: Introduce a dynamic temporal attention mechanism into the feature extraction module of the pre-trained neural network, and introduce a confidence weighting mechanism into the prediction module of the neural network. In this embodiment, wind turbine faults, such as blade fatigue cracks and bearing wear, mostly have clear temporal evolution patterns: either they accumulate slowly, for example, blade deformation increases from small to significant; or they occur suddenly, for example, vibration peaks suddenly surge. The feature extraction module of existing neural networks processes all time window features equally, which may lead to the submergence of key temporal information about the fault. For example, if the deformation rate suddenly increases 24 hours before a blade crack, the blade deformation features are likely to be diluted by a large amount of stable data, and the network will be unable to capture early warning signals. Therefore, in this embodiment, a dynamic temporal attention mechanism is introduced. By automatically assigning weights to different time windows, it focuses on key temporal features that are strongly correlated with the fault, which can significantly improve the effectiveness of temporal features.

[0077] Furthermore, wind turbines operate in complex environments, such as outdoor and high-humidity conditions. The reliability of multi-source data often fluctuates; blade deformation data may be affected by heavy rain and dust, and vibration sensors may drift due to aging. These are all low-reliability data that can mislead prediction results. In contrast, grid voltage data is relatively stable and has higher reliability. Therefore, in this embodiment, a confidence-weighted mechanism is introduced. The weights are dynamically adjusted based on the real-time reliability of the data sources, which can reduce the interference of low-quality data and improve the reliability of prediction results.

[0078] Step S4: Input the historical sample dataset and the expanded sample dataset into the feature extraction module. Extract data features from the historical sample dataset and the expanded sample dataset by assigning weights to features in different time windows using a dynamic temporal attention mechanism, thus obtaining the first feature and the second feature. In this embodiment, by dividing multi-source data in the historical sample dataset and the expanded sample dataset, such as vibration time series and operating parameter time series, into time windows, calculate the contribution weight of features at different times to fault prediction. For example, the vibration peak value in the 24 hours before the fault occurs has a higher weight than the earlier stable value. Then, divide the multi-source data into time windows, such as 10 minutes per window, and calculate the contribution weight of features in different time windows to fault prediction using an attention mechanism; and learn the temporal coupling relationship between different data types using the attention mechanism.

[0079] Furthermore, such as Figure 2 As shown, step S4, which involves using a dynamic temporal attention mechanism to assign weights to features in different time windows to extract data features from the historical dataset and the augmented dataset, and obtaining the first and second features, further includes:

[0080] Step S41: Align the historical sample dataset and the expanded sample dataset sequentially to form two parallel time-series datasets. Step S42: Divide each time-series dataset into consecutive fixed-length windows, each containing multi-source data within its corresponding time period; the fixed-length windows are dynamically adjusted according to the fault types in the time-series datasets. Step S43: Calculate the weight of each fixed-length window using a dynamic temporal attention mechanism and assign weight values. Step S4: Extract basic features from the multi-source data within each fixed-length window, and sum the extracted basic features according to their corresponding weight values ​​to obtain the global features of each time-series dataset, thereby generating the corresponding first and second features.

[0081] Step S5: Input the first feature and the second feature into the prediction module respectively. Weight the first feature and the second feature using a confidence-weighted mechanism to output the corresponding first prediction result and second prediction result. In this embodiment, a real-time confidence evaluation index is set for each data source in the first feature and the second feature, such as vibration, operating parameters, and environmental data. Examples include the signal-to-noise ratio of the vibration sensor and the calibration deviation of the anemometer. Weights are assigned to features from different data sources based on confidence. For example, the weight of high-confidence vibration data is increased from 0.5 to 0.8, while the weight of blade deformation data affected by factors such as heavy rain is decreased from 0.5 to 0.2. Finally, the fault prediction probability is recalculated by combining the weighted multi-source features to avoid the influence of errors from a single data source on the results.

[0082] Furthermore, such as Figure 3 As shown, step S5, which uses a confidence-weighted mechanism to weight the first and second features to output corresponding first and second prediction results, includes: Step S51, assigning a confidence assessment index to each data feature in the first and second features; Step S52, assigning fault allocation weights to the first and second features according to preset weights based on the confidence assessment index; and Step S53, inputting the weighted first and second features into the prediction module and outputting the first and second prediction results corresponding to the fault probability distribution of the real data.

[0083] Step S6 further includes: calculating the loss value between the first prediction result and the second prediction result and the corresponding actual result through the loss function, adjusting the parameters of the neural network according to the loss value through the backpropagation algorithm to complete the training of the neural network and obtain the trained wind turbine fault prediction model.

[0084] Furthermore, the loss value is backpropagated to each layer of the neural network using the gradient descent algorithm. The weights of the time window features in the feature extraction module, the convolution parameters of the neural network, and the weighting coefficients of the confidence weighting mechanism in the prediction module are adjusted to optimize the loss value in the direction of reduction.

[0085] Step S6 further includes: verifying the trained wind turbine fault prediction model with a validation set, inputting the validation set into the trained wind turbine fault prediction model for model performance evaluation, and if the test performance fails to meet expectations, adjusting the time window allocation of the dynamic temporal attention mechanism, the initial coefficients of the confidence weighting mechanism, or the parameters of the loss function, and retraining the model until the model converges; otherwise, saving the relevant parameters of the model training.

[0086] The validation set is a part of the historical sample dataset and the expanded sample dataset mentioned above. Before training, the historical sample dataset and the expanded sample dataset are randomly divided into a training set and a validation set in a 7:3 ratio. The training set is used to train the trained wind turbine fault prediction model, and the validation set is used to validate the trained wind turbine fault prediction model.

[0087] Step S7: Obtain real-time multi-source data of the target wind turbine, preprocess the real-time multi-source data to obtain real-time sample data, input the real-time sample data into the trained wind turbine fault prediction model, and output the fault prediction result of the target wind turbine.

[0088] Thus, this embodiment provides a method for predicting wind turbine faults. It expands upon historical sample data of the wind turbine's physical model to comprehensively capture the operating state characteristics of the wind turbine, ensuring the authenticity and completeness of the training data for wind turbine faults. This overcomes the problems of limited training data, lack of real fault data, and single data source in existing technologies. By improving the neural network, a dynamic temporal attention mechanism is introduced into the feature extraction module. By automatically assigning weights to different time windows, focusing on key temporal features strongly correlated with faults, the effectiveness of temporal features can be significantly improved. A confidence-weighted mechanism is introduced into the prediction module, dynamically adjusting weights based on the real-time credibility of the data source, which weakens the interference of low-quality data and improves prediction accuracy and reliability. The improved neural network is used to train the wind turbine fault prediction model, giving it good reliability, interpretability, transferability, and generalization ability.

[0089] Example 2

[0090] like Figure 4As shown, this embodiment provides a fault prediction system 10 for wind turbines. This fault prediction system is implemented based on the fault prediction method for wind turbines in Embodiment 1 above. Specific details will not be repeated here.

[0091] The fault prediction system mainly includes: a historical data acquisition module 11, an expanded data acquisition module 12, a neural network construction module 13, a data feature extraction module 14, a feature weighting processing module 15, a prediction model training module 16, and a prediction result generation module 17.

[0092] The historical data acquisition module 11 is used to acquire multi-source historical data of the target wind turbine, preprocess the multi-source historical data, and obtain a sample historical dataset.

[0093] The data acquisition module 12 is used to perform data augmentation processing on the historical dataset of the sample through the physical model of the wind turbine to obtain the sample augmented dataset.

[0094] The neural network construction module 13 is used to introduce a dynamic temporal attention mechanism in the feature extraction module of the pre-trained neural network and a confidence weighting mechanism in the prediction module of the neural network.

[0095] The data feature extraction module 14 is used to input the historical sample dataset and the expanded sample dataset into the feature extraction module, and to extract the data features of the historical sample dataset and the expanded sample dataset by allocating the weights of features in different time windows through a dynamic temporal attention mechanism, so as to obtain the first feature and the second feature.

[0096] The feature weighting processing module 15 is used to input the first feature and the second feature into the prediction module respectively, and to perform weighting processing on the first feature and the second feature through a confidence weighting mechanism to output the corresponding first prediction result and second prediction result.

[0097] The prediction model training module 16 is used to calculate the loss value between the first prediction result and the second prediction result and the corresponding actual result through the loss function, and adjust the parameters of the neural network according to the loss value through the backpropagation algorithm to complete the training of the neural network and obtain the trained wind turbine fault prediction model.

[0098] The prediction result generation module 17 is used to acquire real-time multi-source data of the target wind turbine, preprocess the real-time multi-source data to obtain real-time sample data, input the real-time sample data into the trained wind turbine fault prediction model, and output the target wind turbine state prediction result.

[0099] Example 3

[0100] This embodiment provides a storage medium storing a computer program. When executed by a processor, this computer program implements the fault prediction method for the wind turbine generator described in Embodiment 1. Thus, by providing a deployable, replicable, and scalable software implementation, it supports the rapid promotion and mass deployment of the entire system.

[0101] Those skilled in the art will understand that all or part of the steps in the above-described wind turbine fault prediction method 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 wind turbine fault prediction method 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, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0102] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A wind power generator failure prediction method characterized by, The fault prediction method comprises the following steps: Step S1, obtaining multi-source historical data of a target wind turbine, preprocessing the multi-source historical data to obtain a sample historical data set; Step S2, performing data augmentation processing on the sample historical data set through a wind turbine physical model to obtain a sample augmented data set; Step S3, introducing a dynamic time sequence attention mechanism into a feature extraction module of a pre-trained neural network, and introducing a confidence weighting mechanism into a prediction module of the neural network; Step S4, inputting the sample historical data set and the sample augmented data set into the feature extraction module, assigning weights to different time window features through the dynamic time sequence attention mechanism to extract data features of the sample historical data set and the sample augmented data set, and obtaining first features and second features; Step S5, inputting the first features and the second features into the prediction module respectively, performing weighting processing on the first features and the second features through the confidence weighting mechanism, and outputting corresponding first prediction results and second prediction results; Step S6, calculating loss values between the first prediction results and the second prediction results and corresponding actual results, adjusting parameters of the neural network according to the loss values, and thus obtaining a trained wind turbine fault prediction model; Step S7, obtaining real-time multi-source data of the target wind turbine, preprocessing the real-time multi-source data to obtain real-time sample data, inputting the real-time sample data into the trained wind turbine fault prediction model, and outputting a fault prediction result of the target wind turbine.

2. The wind generator failure prediction method of claim 1 wherein, The multi-source historical data in step S1 at least includes deformation data, vibration data, running data, environmental dynamic data and power grid interaction data of each blade of the target wind turbine within a preset time period; And further comprising data cleaning and data standardization processing on the running data; The data cleaning adopts a method combining 3σ criterion and isolated forest algorithm to remove noise and outliers; The data standardization adopts z-score standardization method, and the formula is: ； wherein, is the original data, is the mean of the dimension data, is the standard deviation of the dimension data.

3. The wind generator failure prediction method of claim 1 wherein, The step S2 further comprises: Based on a blade aerodynamic load model, a transmission chain dynamics model and a material fatigue damage model, the theoretical stress distribution under different wind speeds and different rotating speeds is calculated combined with the deformation data of each blade and the vibration data; the fault evolution process of blade fatigue crack propagation and bearing wear is simulated to generate sample augmented data conforming to physical laws, forming the sample augmented data set.

4. The wind generator failure prediction method of claim 3 wherein, The blade aerodynamic load model is based on the principle of fluid mechanics, combined with wind speed and rotating speed in the sample historical data set, to calculate the aerodynamic load distribution of each section of the blade under different working conditions, and then deduce the blade deformation and vibration characteristics; The transmission chain dynamics model is based on the principle of multi-body dynamics, simulates the meshing gap, friction torque and vibration transmission path of the gear box and bearing under different rotating speeds, and generates augmented data conforming to the correlation of vibration, vibration frequency and component characteristics combined with the vibration characteristics in the sample historical data; The material fatigue damage model is based on fatigue accumulation theory, combines the deformation data and the vibration data in the sample historical data, simulates the whole life cycle damage process of material from no damage to slight damage and then to obvious failure, and generates the extended data containing complete failure evolution stages.

5. The wind generator failure prediction method of claim 1 wherein, The step S4 further includes: In step S41, the sample historical data set and the sample extended data set are aligned in time sequence to form two groups of parallel time sequence data sets; In step S42, the two groups of time sequence data sets are cut into continuous fixed length windows, each of which contains multi-source data in a corresponding time sequence period; and the fixed length windows are dynamically adjusted according to the failure types in the time sequence data sets; In step S43, the weight of each fixed length window is calculated by the dynamic time sequence attention mechanism, and the weight value is divided; In step S44, the basic features of the multi-source data in each fixed length window are extracted, and the extracted basic features are weighted and summed according to the corresponding weight values to obtain the global features of each time sequence data set, thereby generating the first feature and the second feature.

6. The wind generator failure prediction method of claim 1 wherein, The step S5 further includes: In step S51, a credibility evaluation index is assigned to each data feature in the first feature and the second feature; In step S52, a failure distribution weight is assigned to the first feature and the second feature according to a preset weight based on the credibility evaluation index; In step S53, the first feature and the second feature after the weight assignment are input into the prediction module to output the first prediction result and the second prediction result of the failure probability distribution corresponding to the true data.

7. The wind generator failure prediction method of claim 1 wherein, The step S6 further includes: The loss value is back propagated to each layer of the neural network by a gradient descent algorithm to adjust the weight of the time window feature of the feature extraction module, the neural network convolution parameter and the weighting coefficient of the confidence weighting mechanism of the prediction module, so that the loss value is optimized in the direction of decreasing.

8. The wind turbine fault prediction method of claim 1, wherein, The step S6 further includes: The trained wind turbine failure prediction model is verified by a verification set, the verification set is input into the trained wind turbine failure prediction model for model performance evaluation, if the test performance fails to meet the expectation, the time window distribution of the dynamic time sequence attention mechanism, the initial coefficient of the confidence weighting mechanism or the loss function parameter is adjusted, the model training is performed again until the model converges, otherwise, the model training related parameters are saved.

9. A wind generator fault prediction system characterized by, The failure prediction system includes: A historical data acquisition module is configured to acquire multi-source historical data of a target wind turbine, and preprocess the multi-source historical data to obtain a sample historical data set; An extended data acquisition module is configured to perform data extension processing on the sample historical data set by a wind turbine physical model to obtain a sample extended data set; A neural network construction module is configured to introduce a dynamic time sequence attention mechanism into a feature extraction module of a pre-trained neural network, and introduce a confidence weighting mechanism into a prediction module of the neural network. The data feature extraction module is configured to input the sample historical data set and the sample expansion data set into the feature extraction module, extract data features of the sample historical data set and the sample expansion data set by assigning weights of different time window features through the dynamic time sequence attention mechanism, and obtain first features and second features. The feature weighting processing module is configured to input the first features and the second features into the prediction module respectively, perform weighting processing on the first features and the second features through the confidence weighting mechanism, and output corresponding first prediction results and second prediction results. The prediction model training module is configured to calculate loss values between the first prediction results and the second prediction results and corresponding actual results through a loss function, adjust parameters of the neural network according to the loss values through a back propagation algorithm, complete training of the neural network, and obtain a trained wind turbine fault prediction model. The prediction result generation module is configured to acquire real-time multi-source data of the target wind turbine, pre-process the real-time multi-source data to obtain real-time sample data, input the real-time sample data into the trained wind turbine fault prediction model, and output a state prediction result of the target wind turbine.

10. A storage medium storing a computer program, characterized in that The computer program, when executed by a processor, implements the wind turbine fault prediction method in any one of claims 1 to 8.

Images