Wind power transmission fault early warning method and device based on single machine-cluster model interaction

Abstract

The invention relates to a wind power transmission fault early warning method and device based on single machine-cluster model interaction. The method comprises the steps that historical operation data of wind turbine generators of the same type in the same wind power plant are collected; aggregating to obtain a first data set corresponding to a single wind turbine generator set and aggregating to obtain a second data set of the same type of wind turbine generator sets; performing model training according to the first data set to obtain a single-machine early warning model, and performing model training according to the second data set to obtain a cluster early warning model; interactively correcting an output result by using the single-machine early warning model and the cluster early warning model; and adopting a voting fusion strategy to obtain a final prediction result. Through interaction and fusion of the single machine-cluster model, personalized features of the single machine model and a common rule of the cluster model are fully utilized, result misjudgment and missed judgment are avoided, and the accuracy of fault prediction of the wind turbine generator transmission system is remarkably improved. The method can better adapt to complex and changeable operation conditions of the wind turbine generator, and the reliability of fault early warning is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power generation technology, in particular to a wind power transmission fault early warning method and device based on single-machine-cluster model interaction. BACKGROUND

[0002] With the rapid development of the wind power industry, the scale and number of wind turbines are increasing. The transmission system of the wind turbine, as a core component, directly affects the power generation efficiency and economic benefits of the entire wind farm. However, the working environment of the wind turbine is complex and variable, such as high humidity and strong corrosion environment in offshore wind farms, strong wind shear and temperature sudden change environment in mountainous wind farms, etc., which makes the transmission system of the wind turbine face many fault risks, especially for the transmission system of the wind turbine, such as gear wear, bearing failure, shaft rupture, etc.

[0003] At present, the existing wind turbine transmission system fault prediction mainly analyzes the historical operation data of a specific wind turbine, such as the vibration characteristics and temperature change trend of the transmission system in a single wind turbine, to mine the potential fault hidden dangers of the transmission system of the wind turbine. However, this method is difficult to obtain common fault rules of the same type of unit in a macroscopic level because it uses single historical data to predict its own faults, and the fault prediction ability for some systemic factors (such as quality defects or design defects of a batch of transmission components) is limited. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a wind power transmission fault early warning method and device based on single-machine-cluster model interaction, which specifically adopts the following technical solutions:

[0005] A wind power transmission fault early warning method based on single-machine-cluster model interaction, the method comprising the following steps:

[0006] Collecting historical operation data of the same type of wind turbine in the same wind farm respectively;

[0007] Aggregating the historical operation data of a single wind turbine to obtain a first data set corresponding to the single wind turbine;

[0008] Aggregating the historical operation data of the same type of wind turbine in the wind farm to obtain a second data set of the same type of wind turbine;

[0009] Using the first data set to train a model to obtain a single-machine early warning model and using the second data set to train a model to obtain a cluster early warning model;

[0010] Using the single-machine early warning model and the cluster early warning model to correct each other to obtain the output results of the single-machine early warning model and the cluster early warning model;

[0011] A voting fusion strategy is used to analyze the output results of the single-machine early warning model and the cluster early warning model to obtain the final prediction result.

[0012] Optionally: The step of using the single-machine early warning model and the cluster early warning model to mutually correct the results includes:

[0013] The system collects real-time operating data using the built-in sensors of a single wind turbine and inputs it into the single-unit early warning model of the current wind turbine to obtain the first fault type.

[0014] The first fault type is transmitted to the cluster early warning model in real time, and the cluster early warning model determines whether the first fault type belongs to the fault type already identified by the cluster early warning model.

[0015] When it is determined that the first fault type belongs to the identified fault types, the cluster early warning model compares the similarity between the operating data of a single wind turbine and the data cluster of the corresponding fault type.

[0016] When it is determined that the first fault type does not belong to the identified fault types, the cluster early warning model will include the first fault type and update the second dataset.

[0017] Optionally: The step of comparing the similarity between the operating data of a single wind turbine and the data cluster of the corresponding fault type in the cluster early warning model includes:

[0018] The cluster early warning model calculates the first Mahalanobis distance between the operating data of a single wind turbine and the data cluster corresponding to the fault type.

[0019] When the first Mahalanobis distance is less than the preset threshold, the cluster early warning model determines that the first fault type has occurred in a single wind turbine.

[0020] When the first Mahalanobis distance is greater than or equal to the preset threshold, calculate the second Mahalanobis distance between the operating data of a single wind turbine and the data cluster of the remaining fault types.

[0021] When there is a data cluster in the remaining fault types whose second Mahalanobis distance is less than the preset threshold, the cluster early warning model determines the fault type corresponding to the data cluster of a single wind turbine.

[0022] When there are no data clusters in the remaining fault types whose second Mahalanobis distance is less than a preset threshold, the second Mahalanobis distance between the operating data of a single wind turbine and the data clusters of the remaining fault types is sorted, and the fault type with the smallest second Mahalanobis distance is selected as the second fault type. The cluster early warning model determines that the second fault type has occurred in a single wind turbine.

[0023] Optionally: The step of using the single-machine early warning model and the cluster early warning model to mutually correct the results further includes:

[0024] The cluster early warning model feeds back the data clusters of identified fault types to the single-machine early warning model;

[0025] The single-machine early warning model calculates the cosine similarity between the data clusters of fault types identified by the cluster early warning model and the fault types identified by itself.

[0026] When the cosine similarity between a data cluster of a certain fault type and the fault type identified in the single-machine early warning model exceeds a set threshold, it is determined that they have a high degree of similarity. At this time, the single-machine early warning model will adjust the prediction weight of the corresponding fault type.

[0027] If the cosine similarity between the data clusters of fault types identified by the cluster early warning model and the fault types identified by the single-machine early warning model does not exceed the set threshold, the single-machine early warning model maintains its current prediction weight.

[0028] Optionally: The step of analyzing the output results of the single-machine early warning model and the cluster early warning model using a voting fusion strategy includes:

[0029] When the outputs of the single-machine early warning model and the cluster early warning model are consistent, the final prediction result is output directly.

[0030] When the output results of the single-machine early warning model and the cluster early warning model are inconsistent, the accuracy rates of the single-machine early warning model and the cluster early warning model are calculated separately.

[0031] The voting weights for the single-machine early warning model and the cluster early warning model are assigned based on their accuracy.

[0032] The confidence scores of the output results of the single-machine early warning model and the cluster early warning model are obtained respectively, and the weighted confidence scores of the single-machine early warning model and the cluster early warning model are calculated based on the voting weights.

[0033] The best prediction results are obtained by weighted confidence screening based on single-machine early warning model and cluster early warning model.

[0034] Optional: After obtaining historical operating data for different transmission components, the historical operating data needs to be preprocessed.

[0035] Outliers can be replaced using median filtering or linear fitting.

[0036] Missing values ​​can be filled using linear interpolation or mean imputation.

[0037] The Z-score normalization method was used to standardize the runtime data for different attributes.

[0038] Optionally: When adjusting the prediction weights for the corresponding fault types, the single-machine early warning model adopts a Sigmoid nonlinear adjustment strategy.

[0039] Optional: When training a single-machine early warning model based on the first dataset, a convolutional neural network model is used, and the cross-entropy loss function is used during training, with the Adagrad optimizer selected as the optimizer.

[0040] Optional: When training the model to obtain the cluster early warning model based on the second dataset, a Gaussian mixture model is adopted. During the training process, the optimal number of Gaussian distributions is determined by the Bayesian information criterion, and the clustering parameters of each Gaussian distribution are calculated by the expectation-maximization algorithm.

[0041] Furthermore, this application also discloses a wind power transmission fault early warning device based on a single-machine-cluster model interaction, the device comprising:

[0042] The data acquisition module is used to collect historical operating data of the same type of wind turbines in the same wind farm.

[0043] The first data aggregation module is used to aggregate the historical operating data of a single wind turbine to obtain the first dataset corresponding to each transmission component.

[0044] The second data aggregation module is used to aggregate historical operating data of the same type of wind turbines in the wind farm to obtain a second dataset of the same type of wind turbines.

[0045] The model training module is used to train a single-machine early warning model based on the first dataset and to train a cluster early warning model based on the second dataset.

[0046] The model interaction correction module is used to correct the results of the single-machine early warning model and the cluster early warning model, and obtain the output results of the single-machine early warning model and the cluster early warning model.

[0047] The results output module is used to analyze the output results of the single-machine early warning model and the cluster early warning model using a voting fusion strategy to obtain the final prediction result.

[0048] Beneficial effects

[0049] The technical solution of this application achieves the following beneficial effects:

[0050] The wind turbine drive fault early warning method proposed in this application fully utilizes the personalized features of the single-unit model and the common patterns of the cluster model through the interaction and fusion of single-unit and cluster models. This avoids misjudgments and omissions caused by the data limitations of a single model, significantly improving the accuracy of fault prediction for wind turbine drive systems. Furthermore, through information interaction and dynamic weight adjustment between models, it can better adapt to the complex and ever-changing operating conditions of wind turbines. Whether it is a specific fault in a single unit or a systemic fault at the cluster level, it can issue timely and accurate early warnings, enhancing the reliability of fault early warning. Attached Figure Description

[0051] Figure 1 This is a flowchart of a wind power transmission fault early warning method based on a single-machine-cluster model interaction in an embodiment of this application.

[0052] Figure 2 This is a structural diagram of the wind power transmission fault early warning device based on the single-machine-cluster model interaction in the embodiments of this application.

[0053] Figure 3 This is a structural diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0054] The present application will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application and should not be construed as limiting the scope of protection of the present application. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present application.

[0055] Existing methods for wind turbine fault early warning mostly focus on individual unit data, such as vibration signals and temperature detection, analyzing only the operating data of a single wind turbine. However, due to the complex environment of wind farms, with constantly changing factors such as wind speed, temperature, and humidity, existing single-unit models lack generalization ability when facing complex and variable operating conditions, easily leading to false alarms and missed alarms. The wind turbine transmission fault early warning method based on the interaction between single-unit and cluster models proposed in this application allows the single-unit model to focus on the dynamic response of a specific unit under different operating conditions, while the cluster model provides a macro-level understanding of the overall performance of different units under similar operating conditions. The two work together; when a single unit encounters extreme operating conditions or data anomalies, the prediction results can be corrected using the experience of the cluster model, enhancing the model's resistance to interference from complex environments and abnormal data, and significantly improving the model's robustness.

[0056] Specifically, such as Figure 1 As shown in the figure, this embodiment discloses a wind power transmission fault early warning method based on a single-machine-cluster model interaction. The method includes the following steps:

[0057] First, data acquisition and preprocessing are performed:

[0058] This embodiment collects historical operating data for various transmission components (gearboxes, main shaft bearings, generator bearings, etc.) of similar wind turbines within the same wind farm. The collected historical operating data includes time-domain and frequency-domain characteristic data such as vibration amplitude, temperature, rotational speed, and load. The data sampling frequency is adjusted according to data volume and computational requirements. For example, this embodiment uses an offshore wind farm with 30 similar offshore wind turbines distributed in the same area. Given the harsh offshore environment, the focus is on collecting vibration and temperature data from the internal transmission components of the wind turbines. High-precision vibration and temperature sensors can be installed on key components such as the gearbox, main shaft, and generator of each wind turbine, collecting data at 30-minute intervals to compile historical operating data for the past year.

[0059] Subsequently, the historical operating data of different transmission components needs to be preprocessed separately. The general data preprocessing process includes:

[0060] (1) Outlier handling: Use median filtering (suitable for impulse noise) or linear fitting (suitable for gradual anomalies) to replace outliers to avoid noise interfering with model training.

[0061] (2) Missing value handling: Linear interpolation or mean filling method is used for filling. Linear interpolation is suitable for data with continuous missing values ​​in time series, while mean filling method is suitable for data with a small number of missing values ​​at random locations, thus ensuring the integrity of the running data.

[0062] (3) Standardization processing: The Z-score normalization method is used to process the running data with different attributes (such as vibration amplitude, temperature, etc.) respectively, and the running data are mapped to the standard normal distribution with a mean of 0 and a standard deviation of 1, so as to eliminate the influence of the dimension and improve the stability and accuracy of model training.

[0063] Then, a data aggregation strategy is implemented:

[0064] This embodiment aggregates historical operating data (such as temperature and vibration) of different transmission components (e.g., gearbox, generator, main shaft, etc.) in a single wind turbine unit according to time to obtain a first dataset. The purpose is to integrate various types of data collected from the same unit at different time points to form a complete dataset that reflects the unit's operating characteristics. For example, if data is collected at 30-minute intervals, 48 ​​sets of operating data can be collected daily, forming 48 datasets.

[0065] J = {D1,D2,…,D} 48};

[0066] Where D1, D2, ..., D 48These represent the operational data of a single offshore wind turbine at 48 data collection points per day. After statistically analyzing the historical operational data for the most recent year (represented as 365 days), the first dataset M is formed.

[0067]

[0068] in The data cluster representing the first data collection time of a single offshore wind turbine within one year, and similarly... This represents the operational data cluster of a single offshore wind turbine at the second data collection time of a single day within one year. This represents the operational data cluster of a single offshore wind turbine at the 48th data collection point on a single day within one year. It should be noted that during the aggregation process, due to potential slight differences in the time of data collection from different sensors, the collected data first needs to undergo time alignment processing. This is generally achieved by matching timestamps to link all relevant operational data from the same moment, ensuring data consistency and accuracy.

[0069] Secondly, this embodiment aggregates historical operating data of similar wind turbines within the wind farm to obtain a second dataset. The aggregation method is similar to that of single-unit data aggregation. First, the collected data from all similar units are time-aligned and preprocessed according to the same acquisition time to ensure data quality. Then, the data from different units at the same time are merged into a single set, forming a multidimensional dataset. This multidimensional dataset provides a rich data foundation for subsequent cluster model training, enabling the model trained on the multidimensional dataset to learn the data differences and commonalities of similar units under the same operating conditions. Finally, after aggregation processing, a second dataset of similar wind turbines is obtained, reflecting the operating status and data distribution characteristics of the entire wind power cluster.

[0070] This embodiment uses 30 identical offshore wind turbines as an example, collecting data at 30-minute intervals. Each offshore wind turbine can collect 48 sets of operational data per day. For example, J1 = {D1, D2, ..., D...} 48}, where D1, D2, ..., D 48 These represent the operational data collected at 48 different times each day for the current offshore wind turbines. Then, the daily operational data collected from 30 offshore wind turbines are aggregated according to the same collection time, resulting in 30 datasets per day for the current offshore wind farm. in This represents the set of operational data from 30 offshore wind turbines at the first collection time of the day, and so on. This represents the set of operational data from 30 offshore wind turbines at the second data collection time of the day. This represents the set of operational data from 30 offshore wind turbines at the 48th data collection point of a single day. When analyzing historical operational data over the past year (365 days), since the operational data at the same collection point is treated as a separate set, 48 second datasets can be formed. Each second dataset can be represented as... This refers to the collection of operational data from 30 offshore wind turbines within a single data collection point over a year. Since wind turbines of the same type within the same wind farm often share similar equipment characteristics, technical parameters, and operational logic, and operate in similar environments, aggregating their operational data can eliminate data fluctuations caused by individual turbine differences, highlight common operational patterns of similar equipment, expand the data sample size, improve the statistical significance and reliability of data analysis, and provide a more representative data foundation for equipment fault early warning.

[0071] Next, the early warning model is trained:

[0072] This embodiment uses the first dataset to train the model and obtain a single-machine early warning model. Specifically, in training the single-machine early warning model, this embodiment preferably uses a convolutional neural network model, and employs the cross-entropy loss function and the Adagrad optimizer during training.

[0073] Because wind turbine drive systems generate a large amount of temporally and spatially correlated operational data during operation, such as gearbox vibration signals and bearing temperature changes, this data is characterized by local feature aggregation and periodic variations. This embodiment employs a Convolutional Neural Network (CNN) with local perception and weight sharing mechanisms to efficiently extract spatial and temporal features from the operational data of the wind turbine drive system. This effectively reduces the number of model parameters, avoids overfitting, and demonstrates excellent recognition capabilities for fault feature patterns in the operational data. For example, the convolutional layers of the CNN can automatically extract fault features contained in the vibration signals, while the pooling layers perform dimensionality reduction on the features, retaining key information, thereby accurately capturing early signs of faults in the wind turbine drive system.

[0074] Specifically, in the training process of the single-machine early warning model, this embodiment divides the first dataset into a training set, a validation set, and a test set. The training set is used for updating and learning model parameters, the validation set is used to adjust hyperparameters to prevent overfitting, and the test set is used to evaluate the generalization ability of the final model. During model training, preprocessed wind turbine drive system operation data is input into the CNN model in batches. The data sequentially passes through convolutional layers, pooling layers, and fully connected layers for feature extraction and classification. In each training iteration, the model predicts the input data based on the current parameters, obtaining the prediction result. Subsequently, the cross-entropy loss function is used to measure the difference between the prediction result and the true result. Since fault categories in wind turbine drive fault early warning generally fall under classification problems, this embodiment uses the cross-entropy loss function to effectively guide the model to learn the fault patterns in the data, making the model's prediction result closer to the true result and rapidly improving the model's classification accuracy. When the model's prediction result deviates significantly from the true label, the cross-entropy loss function will give a larger loss value, prompting the model to adjust parameters more drastically during backpropagation, accelerating the model's convergence speed.

[0075] Furthermore, during the training of wind turbine transmission fault data, the update frequency and magnitude of different types of operating parameters vary. The Adagrad optimizer can adaptively adjust the learning rate based on the historical gradient of each operating parameter. A smaller learning rate is used for frequently updated parameters to avoid over-updating, while a larger learning rate is used for less frequently updated parameters to accelerate their convergence. This adaptive adjustment mechanism allows the model to quickly learn key features during training while avoiding getting trapped in local optima, thereby improving the efficiency and stability of model training and ensuring the reliability and accuracy of the single-unit early warning model.

[0076] Furthermore, this embodiment uses a second dataset for model training to obtain a cluster early warning model. When training the model using the second dataset, since it is essentially a collection of data from multiple individual units, its data distribution is more complex, exhibiting multi-peak and multi-modal characteristics. A more complex and flexible model is needed to accurately analyze its distribution patterns. Therefore, this embodiment preferably uses a Gaussian mixture model. This model describes complex data distributions through a linear combination of multiple Gaussian distributions, making it more suitable for modeling the operational data of wind power clusters. Compared to traditional single-distribution models, Gaussian mixture models can more accurately capture the potential patterns and structures in the data, and can more finely characterize the operating characteristics of different units in a wind power cluster under different operating conditions, thus providing a more reliable basis for fault early warning.

[0077] Furthermore, during the training process, the cluster early warning model in this embodiment generally determines the optimal number of Gaussian distributions using the Bayesian information criterion, achieving an optimal balance between fitting the data and avoiding overfitting. The clustering parameters of each Gaussian distribution are calculated using the expectation-maximization algorithm. The Gaussian mixture model used in the aforementioned cluster early warning model iteratively solves for the model parameters through the following steps: In the expectation calculation phase, based on the currently estimated model parameters, the probability of each data point belonging to each Gaussian distribution is calculated; in the maximization calculation phase, based on the probabilities obtained in the expectation calculation phase, the parameters of the Gaussian distributions (mean, covariance, and weights) are updated to maximize the model's log-likelihood function. By continuously repeating the above two steps, the model gradually converges to the optimal clustering parameters, thereby determining the final Gaussian mixture model.

[0078] It should be noted that the single-unit early warning model in this embodiment mainly focuses on the operating status of a single unit, emphasizing the detection and prediction of fault symptoms within the unit itself. Cluster early warning, on the other hand, needs to take a holistic approach, considering the mutual influence and correlation between units. It not only identifies faults in individual units but also predicts the propagation and spread of faults within the cluster. Using Gaussian mixture models, potential relationships and operating patterns between different units can be mined from cluster data, providing more comprehensive information support for cluster-level fault early warning. For example, when a unit experiences a fault, the Gaussian mixture model can analyze the impact of the fault on surrounding units and the overall cluster operating status, thereby allowing for proactive measures to prevent the fault from escalating.

[0079] The results were then corrected using a dual-model interaction:

[0080] This embodiment uses a single-machine early warning model and a cluster early warning model to correct each other's results, thereby obtaining the output results of the single-machine early warning model and the cluster early warning model.

[0081] Specifically, the steps for the two early warning models to correct each other in this embodiment include: inputting the data from the single-machine early warning model into the cluster early warning model for correction, and inputting the results from the cluster early warning model into the single-machine early warning model for correction.

[0082] Specifically, the process of correcting data from a standalone early warning model by inputting it into a cluster early warning model is as follows:

[0083] (1) Real-time operation data is collected using built-in sensors (such as vibration sensors, temperature sensors, speed sensors, etc.) of a single wind turbine. The operation data includes, but is not limited to, key parameters such as bearing temperature, gearbox vibration frequency, and generator speed. The collected operation data is then input into the single-unit early warning model of the current wind turbine to obtain the first fault type.

[0084] (2) The first fault type predicted by the single-unit early warning model is transmitted to the cluster early warning model in real time. Since the cluster early warning model is built based on the operating data of multiple wind turbines of the same type, it has identified a variety of common fault types and their corresponding data feature patterns (i.e., data clusters). Then the cluster early warning model determines whether the first fault type belongs to the fault type already identified by the current cluster early warning model.

[0085] (3) When it is determined that the first fault type belongs to the identified fault types, the cluster early warning model will perform the following operations:

[0086] First, compare the similarity between the operating data of a single wind turbine and the data clusters corresponding to the fault types:

[0087] This embodiment uses Mahalanobis distance to measure data similarity. Mahalanobis distance is a distance metric that considers the covariance structure of data and can more accurately reflect the similarity between data points and data clusters. Therefore, the cluster early warning model calculates the first Mahalanobis distance between the real-time operating data of a single wind turbine and the data cluster corresponding to the fault type in the cluster early warning model.

[0088] When the first Mahalanobis distance is less than the preset threshold, it indicates that the single-unit operating data is highly similar to the data cluster of the fault type. In this case, the cluster early warning model determines that the corresponding wind turbine has experienced the first fault type and may increase the early warning level or shorten the maintenance cycle.

[0089] When the first Mahalanobis distance is greater than or equal to the preset threshold, it indicates that the single-unit operating data does not match the initially judged fault type. At this time, the cluster early warning model will calculate the similarity between the real-time collected operating data of a single wind turbine and the data clusters of other fault types, i.e., the second Mahalanobis distance.

[0090] When there is a data cluster in the remaining fault types whose second Mahalanobis distance is less than the preset threshold, it indicates that there is a data cluster that is highly similar to the single-unit operating data. That is, the current single wind turbine may have a fault type corresponding to the highly similar data cluster. Then the cluster early warning model determines the fault type corresponding to the highly similar data cluster of the single wind turbine.

[0091] When no data cluster with a second Mahalanobis distance less than a preset threshold exists among the remaining fault types, the second Mahalanobis distance between the operating data of a single wind turbine and the data clusters of the remaining fault types can be sorted. The fault type with the smallest second Mahalanobis distance is selected as the second fault type. At this point, the cluster early warning model determines that a single wind turbine has experienced the second fault type. It should be noted that if the first fault type is determined to be an identified fault type, but the similarity between the data cluster and the single-unit operating data is insufficient, it indicates that the corresponding fault type may have new operating data characteristics or that the single-unit early warning model is incorrect. In this case, manual inspection results are usually used to correct the model results.

[0092] (4) When it is determined that the first fault type does not belong to the identified fault types, the cluster early warning model will include the first fault type and update the second dataset. However, it should be noted that at this stage, the first Mahalanobis distance between the real-time operating data collected by a single wind turbine and the data cluster of all fault types identified by the cluster early warning model can also be calculated first.

[0093] When there is a data cluster whose first Mahalanobis distance is less than a preset threshold, it indicates that there is a data cluster that is highly similar to the single-unit operating data. That is, the current single wind turbine may have a fault type corresponding to the highly similar data cluster. Then the cluster early warning model determines the fault type corresponding to the highly similar data cluster of the single wind turbine.

[0094] If no data cluster exists with a first Mahalanobis distance less than a preset threshold, a new fault mode may have emerged. In this case, the cluster early warning model records this first fault type and its corresponding operational data characteristics, and updates its fault type library (second dataset). This allows the cluster early warning model to continuously learn and adapt to new fault modes, improving its overall early warning capability.

[0095] More specifically, in this embodiment, the process of inputting the cluster early warning model results into the single-machine early warning model for correction is as follows:

[0096] (1) The cluster early warning model feeds back the data clusters of the identified fault types to the single-machine early warning model.

[0097] (2) The single-machine early warning model calculates the cosine similarity between the data clusters of fault types identified by the cluster early warning model and the fault types identified by itself; the cosine similarity can measure the similarity between two data sets, and the closer the value is to 1, the more similar they are.

[0098] (3) When the cosine similarity between a data cluster of a certain fault type and the fault type identified in the single-machine early warning model exceeds the set threshold (e.g., 0.7), it is determined that the two have high similarity. At this time, the single-machine early warning model will adjust the prediction weight of the corresponding fault type.

[0099] It should be noted that, in this embodiment, when adjusting the prediction weights for the corresponding fault types, a Sigmoid nonlinear adjustment strategy is generally adopted, wherein the Sigmoid function has the following form:

[0100]

[0101] Where x is the original weight, x0 is the adjustment center point, and k is the adjustment rate parameter.

[0102] (4) When the cosine similarity between the data clusters of fault types identified by the cluster early warning model and the fault types identified by the single-machine early warning model does not exceed the set threshold, the single-machine early warning model maintains the current prediction weight.

[0103] The Sigmoid nonlinear adjustment strategy has the following characteristics: for data clusters with high similarity, the corresponding prediction weights are slightly increased to avoid instability caused by over-adjustment; for data clusters with low similarity, the adjustment range is smaller to maintain model stability; when the similarity is extremely high, the corresponding weight adjustment tends to saturate to prevent the weights from increasing indefinitely. For example, this embodiment takes the fault types identified in a single-machine early warning model as an example. The identified fault types include bearing faults, gear faults, and rotor faults. The similarity between the data clusters of fault types identified by the single-machine early warning model and the fault types identified by the cluster early warning model is calculated, and the early warning weights of each fault type are adjusted based on the similarity values, as shown in Table 1.

[0104] Table 1

[0105] Fault type Cluster-single similarity Single current weight Adjusted weight Bearing wear 0.85 0.6 0.78 Gear crack 0.72 0.4 0.45 Rotor imbalance 0.65 0.5 0.5

[0106] According to the results in Table 1, when the data clusters of the cluster early warning model are significantly similar to the operational data of the single-machine early warning model (similarity > threshold + 0.1), the prediction weight of the corresponding fault can be greatly improved; when the data clusters of the cluster early warning model are moderately similar to the operational data of the single-machine early warning model (threshold < similarity < threshold + 0.1), the prediction weight of the corresponding fault can be slightly improved; and when the data clusters of the cluster early warning model are slightly similar to the operational data of the single-machine early warning model (similarity < threshold), the prediction weight of the corresponding fault remains unchanged.

[0107] This embodiment compares the similarity between single-machine data and multiple data clusters using a cluster early warning model, correcting misjudgments by the single-machine early warning model. Simultaneously, the single-machine early warning model adjusts its prediction weights based on common fault characteristics reported by the cluster, enhancing its sensitivity to common fault modes and reducing the possibility of missed alarms. Through this two-way correction mechanism, the single-machine and cluster early warning models form an organic whole, complementing and promoting each other, significantly improving the performance and reliability of the wind power equipment fault early warning system.

[0108] Finally, the output result is based on the voting fusion strategy:

[0109] A voting fusion strategy is used to analyze the output results of the single-machine early warning model and the cluster early warning model to obtain the final prediction result.

[0110] The voting fusion strategy employed in this embodiment effectively integrates the output results of the single-machine early warning model and the cluster early warning model. This strategy achieves accurate decision-making based on factors such as differences in model outputs, historical accuracy, and result confidence. Specifically, the voting fusion strategy process is as follows:

[0111] When the outputs of the single-unit early warning model and the cluster early warning model are consistent, it indicates that the two models, based on different data perspectives and algorithmic logic, have reached a consensus on the fault judgment of the current wind turbine. In this case, the same fault type can be directly output as the final prediction result. Furthermore, the confidence levels of the two models can be combined, and the final result's confidence level can be determined by averaging or using a weighted average, thereby enhancing the credibility and reliability of the decision.

[0112] When the outputs of the single-machine early warning model and the cluster early warning model are inconsistent, historical prediction data can be retrieved to statistically analyze the accuracy rates of each model in identifying similar faults. These accuracy rates are based on long-term operational monitoring and actual fault verification, reflecting the reliability of each model in identifying specific fault types.

[0113] Then, based on the historical accuracy of the single-machine early warning model and the cluster early warning model, voting weights can be assigned to the two models. Generally, the model with higher accuracy has a larger weight in the vote, meaning that the model's judgment has a greater weight in the final decision. For example, if the historical accuracy of the single-machine early warning model in judging a certain type of fault is 85%, and that of the cluster early warning model is 90%, then the cluster early warning model will receive a higher weight in the voting weight allocation.

[0114] Based on the voting weights allocated above, the weighted confidence scores of the single-machine early warning model and the cluster early warning model can be calculated separately. Generally, the original confidence score of the corresponding model (usually between 0 and 1) is multiplied by the corresponding voting weight to obtain the weighted confidence score value.

[0115] Finally, the best prediction result is obtained by weighted confidence screening based on the single-machine early warning model and the cluster early warning model. Generally, the output result of the model with the higher weighted confidence is determined as the final fault type judgment result. Through this method, when the two models diverge, a more reasonable and accurate decision can be made by combining their historical performance and the degree of certainty of the current judgment. This strategy significantly improves the fault detection rate, greatly reduces the false alarm rate, and effectively ensures the stable operation and maintenance efficiency of wind power equipment.

[0116] In addition, combined Figure 2 As shown, this application also discloses a wind turbine drive fault early warning device based on a single-machine-cluster model interaction, the device comprising:

[0117] The data acquisition module is used to collect historical operating data of different transmission components of the same type of wind turbine in the same wind farm;

[0118] The first data aggregation module is used to aggregate the historical operating data of each transmission component in a single wind turbine to obtain the first dataset corresponding to each transmission component.

[0119] The second data aggregation module is used to aggregate the historical operating data of the same transmission components of the same type of wind turbines in the wind farm to obtain the second dataset of the corresponding transmission components.

[0120] The model training module is used to train a single-machine early warning model based on the first dataset and to train a cluster early warning model based on the second dataset.

[0121] The model interaction correction module is used to correct the results of the single-machine early warning model and the cluster early warning model, and obtain the output results of the single-machine early warning model and the cluster early warning model.

[0122] The results output module is used to analyze the output results of the single-machine early warning model and the cluster early warning model using a voting fusion strategy to obtain the final prediction result.

[0123] The apparatus provided in this application embodiment can achieve... Figure 1 To avoid repetition, the various processes implemented in the method embodiments will not be described again here.

[0124] like Figure 3 As shown in the illustration, this application also provides an electronic device, including a processor and a memory, and a program or instructions stored in the memory and executable on the processor, which, when executed by the processor, implement as follows: Figure 1 The various processes of the method embodiments shown are all capable of achieving the same technical effect, and will not be described again here to avoid repetition.

[0125] This application embodiment also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the above-described functionality. Figure 1 The various processes described in the embodiments of the method described herein can achieve the same technical effect, and will not be repeated here to avoid repetition.

[0126] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 1 The various processes described in the embodiments of the method described herein can achieve the same technical effect, and will not be repeated here to avoid repetition.

[0127] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0128] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0129] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another device, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0130] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0131] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0132] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0133] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a device (which may be a terminal or platform, etc.) to execute all or part 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 mobile storage devices, ROMs, magnetic disks, or optical disks.

[0134] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A wind power transmission fault early warning method based on a single-machine-cluster model interaction, characterized in that, The method includes the following steps: Historical operating data of the same type of wind turbines in the same wind farm were collected separately. By aggregating the historical operating data of a single wind turbine, the first dataset corresponding to that single wind turbine is obtained. By aggregating historical operating data of wind turbines of the same type within the wind farm, a second dataset of wind turbines of the same type is obtained. A single-machine early warning model is obtained by training the model based on the first dataset, and a cluster early warning model is obtained by training the model based on the second dataset. The results of the single-machine early warning model and the cluster early warning model are corrected by mutual correction to obtain the output results of the single-machine early warning model and the cluster early warning model. A voting fusion strategy is used to analyze the output results of the single-machine early warning model and the cluster early warning model to obtain the final prediction result.

2. The wind power transmission fault early warning method according to claim 1, characterized in that, The steps for mutual result correction using the single-machine early warning model and the cluster early warning model include: The system collects real-time operating data using the built-in sensors of a single wind turbine and inputs it into the single-unit early warning model of the current wind turbine to obtain the first fault type. The first fault type is transmitted to the cluster early warning model in real time, and the cluster early warning model determines whether the first fault type belongs to the fault type already identified by the cluster early warning model. When it is determined that the first fault type belongs to the identified fault types, the cluster early warning model compares the similarity between the operating data of a single wind turbine and the data cluster of the corresponding fault type. When it is determined that the first fault type does not belong to the identified fault types, the cluster early warning model will include the first fault type and update the second dataset.

3. The wind power transmission fault early warning method according to claim 2, characterized in that, The steps of the cluster early warning model to compare the similarity between the operating data of a single wind turbine and the data cluster of the corresponding fault type include: The cluster early warning model calculates the first Mahalanobis distance between the operating data of a single wind turbine and the data cluster corresponding to the fault type. When the first Mahalanobis distance is less than the preset threshold, the cluster early warning model determines that the first fault type has occurred in a single wind turbine. When the first Mahalanobis distance is greater than or equal to the preset threshold, calculate the second Mahalanobis distance between the operating data of a single wind turbine and the data cluster of the remaining fault types. When there is a data cluster in the remaining fault types whose second Mahalanobis distance is less than the preset threshold, the cluster early warning model determines the fault type corresponding to the data cluster of a single wind turbine. When there are no data clusters in the remaining fault types whose second Mahalanobis distance is less than a preset threshold, the second Mahalanobis distance between the operating data of a single wind turbine and the data clusters of the remaining fault types is sorted, and the fault type with the smallest second Mahalanobis distance is selected as the second fault type. The cluster early warning model determines that the second fault type has occurred in a single wind turbine.

4. The wind power transmission fault early warning method according to claim 1, characterized in that, The step of using the single-machine early warning model and the cluster early warning model to correct the results also includes: The cluster early warning model feeds back the data clusters of identified fault types to the single-machine early warning model; The single-machine early warning model calculates the cosine similarity between the data clusters of fault types identified by the cluster early warning model and the fault types identified by itself. When the cosine similarity between a data cluster of a certain fault type and the fault type identified in the single-machine early warning model exceeds a set threshold, it is determined that they have a high degree of similarity. At this time, the single-machine early warning model will adjust the prediction weight of the corresponding fault type. If the cosine similarity between the data clusters of fault types identified by the cluster early warning model and the fault types identified by the single-machine early warning model does not exceed the set threshold, the single-machine early warning model maintains its current prediction weight.

5. The wind power transmission fault early warning method according to claim 1, characterized in that, The steps for analyzing the output results of the single-machine early warning model and the cluster early warning model using a voting fusion strategy include: When the outputs of the single-machine early warning model and the cluster early warning model are consistent, the final prediction result is output directly. When the output results of the single-machine early warning model and the cluster early warning model are inconsistent, the accuracy rates of the single-machine early warning model and the cluster early warning model are calculated separately. The voting weights for the single-machine early warning model and the cluster early warning model are assigned based on their accuracy. The confidence scores of the output results of the single-machine early warning model and the cluster early warning model are obtained respectively, and the weighted confidence scores of the single-machine early warning model and the cluster early warning model are calculated based on the voting weights. The best prediction results are obtained by weighted confidence screening based on single-machine early warning model and cluster early warning model.

6. The wind power transmission fault early warning method according to claim 1, characterized in that, After obtaining historical operating data for different transmission components, the historical operating data needs to be preprocessed: Outliers can be replaced using median filtering or linear fitting. Missing values ​​can be filled using linear interpolation or mean imputation. The Z-score normalization method was used to standardize the runtime data for different attributes.

7. The wind power transmission fault early warning method according to claim 4, characterized in that, When adjusting the prediction weights for the corresponding fault types, the single-machine early warning model adopts a Sigmoid nonlinear adjustment strategy.

8. The wind power transmission fault early warning method according to claim 1, characterized in that, When training a single-machine early warning model based on the first dataset, a convolutional neural network model is used, and the cross-entropy loss function is used during training. The Adagrad optimizer is selected as the optimizer.

9. The wind power transmission fault early warning method according to claim 1, characterized in that, When training the model to obtain the cluster early warning model based on the second dataset, a Gaussian mixture model is adopted. During the training process, the optimal number of Gaussian distributions is determined by the Bayesian information criterion, and the clustering parameters of each Gaussian distribution are calculated by the expectation-maximization algorithm.

10. A wind power transmission fault early warning device based on a single-machine-cluster model interaction, characterized in that, The device includes: The data acquisition module is used to collect historical operating data of the same type of wind turbines in the same wind farm. The first data aggregation module is used to aggregate the historical operating data of a single wind turbine to obtain the first dataset corresponding to the single wind turbine. The second data aggregation module is used to aggregate historical operating data of the same type of wind turbines in the wind farm to obtain a second dataset of the same type of wind turbines. The model training module is used to train a single-machine early warning model based on the first dataset and to train a cluster early warning model based on the second dataset. The model interaction correction module is used to correct the results of the single-machine early warning model and the cluster early warning model, and obtain the output results of the single-machine early warning model and the cluster early warning model. The results output module is used to analyze the output results of the single-machine early warning model and the cluster early warning model using a voting fusion strategy to obtain the final prediction result.

Images