Wind turbine generator blade icing fault detection method
By using the Relief algorithm to filter key features and constructing a cross-LSTM model, combined with the specific characteristics of wind turbines, the problem of difficulty in capturing time dependencies and high computational resource consumption in wind turbine blade icing detection by traditional methods is solved. This achieves efficient and accurate icing detection and reduces operation and maintenance costs.
Patent Information
- Application Number
- CN202511105170.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-07
- Publication Date
- 2025-10-28
AI Technical Summary
Traditional deep neural networks and recurrent neural networks struggle to capture time-dependent relationships in wind turbine blade icing detection, consume significant computational resources, and are prone to overfitting. Existing LSTM algorithms also have limitations when dealing with complex systems composed of multiple sensor parameters, resulting in insufficient detection accuracy.
The Relief algorithm is used for feature selection, and a cross-LSTM prediction model and a sequence classification model are constructed. Combining wind speed-power and ambient temperature-cabin temperature features, residual signals are generated through cross-LSTM. The classification model is trained using regularization terms and stochastic momentum gradient descent optimization algorithms, and time-domain features are extracted. Cross-entropy is used as the loss function for icing detection.
It improves the accuracy and efficiency of blade icing detection, reduces model complexity, decreases operation and maintenance costs, and enhances the operational stability and economic benefits of wind turbine units.
Smart Images
Figure CN120845285A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine fault detection technology, specifically to a method for detecting icing faults on wind turbine blades. Background Technology
[0002] With the booming development of the wind power industry, the operation and maintenance of key components of wind turbines has become a research focus. In northern and coastal areas, the severe cold weather in autumn and winter frequently causes icing problems on wind turbine blades. Blade icing not only reduces the power generation efficiency of wind turbines, but in severe cases, it can also lead to turbine shutdown, accelerate component aging, shorten the service life of the turbine, and in extreme cases, even cause serious accidents such as blade breakage or turbine damage.
[0003] Traditional deep neural networks (DNNs) suffer from numerous problems when processing time-series data, such as difficulty in capturing temporal dependencies, high computational resource consumption, and susceptibility to overfitting, which limits their application in wind turbine icing detection. Recurrent neural networks (RNNs) and their variants, especially Long Short-Term Memory (LSTM) networks, while effectively processing time-series data and capturing temporal variations in the blade icing process to improve fault detection accuracy due to their unique architecture, still have limitations when handling complex systems composed of multiple sensor parameters. Therefore, there is an urgent need for an efficient and accurate method for detecting icing faults in wind turbine blades to improve the operational stability, safety, and economic efficiency of wind turbines and reduce maintenance costs. Summary of the Invention
[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0005] In view of the above-mentioned problems, the present invention is proposed.
[0006] Therefore, the purpose of this invention is to provide a method for detecting icing faults in wind turbine blades.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for detecting icing faults in wind turbine blades, comprising:
[0008] The Relief algorithm was used to select features from the multi-dimensional feature data collected by the wind turbine SCADA system. The feature weights were evaluated based on the contribution of the features to distinguishing between blade icing and normal conditions, and key features were selected.
[0009] Based on the selected feature data and its visualization, and combined with the icing pattern, a new feature X1 of wind speed-power and a new feature X2 of ambient temperature-cabin temperature were constructed.
[0010] Construct a cross-LSTM prediction model, predict one set of data based on the characteristics of multiple sets of data, and generate residual signals;
[0011] Construct a sequence classification model, using the product of the number of feature variables and the residual feature signal as input data, to determine whether icing occurs at a specific moment in the time series;
[0012] The trained cross-LSTM network is used to predict the test set data under normal conditions, and the temporal features of the residuals are extracted to construct a new feature matrix.
[0013] Cross-entropy was selected as the loss function, and the classification model was trained using the newly constructed feature matrix. The confusion matrix for ice accretion classification was obtained through the test set data.
[0014] As a preferred embodiment of the wind turbine blade icing fault detection method of the present invention, the feature weight vector update formula in the Relief algorithm is as follows: Where W(A) is the weight of feature A, m is the number of repetitions, and diff(A,R,H) is the weight of feature A. i ) represents sample R and similar samples H i Distance on feature A, diff(A,R,M) i Let R be a sample and M be a different sample. i Distance on feature A.
[0015] As a preferred embodiment of the wind turbine blade icing fault detection method of the present invention, the steps of constructing the cross-LSTM prediction model include:
[0016] The dataset containing t characteristic variable measurements and n observation points is filtered to remove high-frequency noise, and after normalization, a new dataset x with dimension t*n is formed.
[0017] The sliding window algorithm is introduced, with a sliding window length of 10. The feature data from time t-10 to time t are input to obtain the prediction data at time t.
[0018] The test set is input into the prediction model to obtain the predicted value of the corresponding sensor. The test value is then denormalized to obtain the residual signal between the predicted value and the actual value.
[0019] As a preferred embodiment of the wind turbine blade icing fault detection method of the present invention, the sequence classification model is iteratively performed using a stochastic momentum gradient descent optimization algorithm, the algorithm formula of which is: Where θ represents the weight vector, k represents the number of updates, and (η) represents the update rate. The gradient represents the loss function, and r represents the weighted effect of the previous gradient on the current gradient.
[0020] As a preferred embodiment of the wind turbine blade icing fault detection method described in this invention, a regularization term E is added to the original loss function to reduce the risk of overfitting during model training. R The loss function formula is: E(θ) = E M (θ)+λE R (θ); Regularization is calculated as follows: E R (θ)=∑ w∈W Ω(w); where λ represents the regularization factor and W represents the weight vector.
[0021] As a preferred embodiment of the wind turbine blade icing fault detection method of the present invention, the time-domain features of the extracted residuals include mean, peak value, peak-to-peak value, sum of squares, and root mean square amplitude, and the calculation formulas are as follows:
[0022] Mean:
[0023] Peak value: x peak =max(X);
[0024] Peak-to-peak value: x p-p =max(X) - min(X);
[0025] Sum of squares:
[0026] Root mean square value: Where, x i Let m be the number of residual signal data points, and X be the set of residual signal data.
[0027] As a preferred embodiment of the wind turbine blade icing fault detection method of the present invention, an icing classification label is introduced in the icing fault detection, where 0 represents the normal operating state of the blade and 1 represents the icing fault state of the blade.
[0028] As a preferred embodiment of the wind turbine blade icing fault detection method of the present invention, the sequence classification model structure includes an input layer, an LSTM layer for capturing long-term and short-term dependencies in the time series, a fully connected layer, and a classification layer, with the number of loop steps set to 6000.
[0029] In summary, the present invention has at least one of the following beneficial effects:
[0030] This invention, through feature selection and an improved LSTM algorithm, can more accurately capture the temporal variation features during blade icing, improving fault detection accuracy and significantly enhancing detection precision. Feature selection reduces unnecessary feature inputs, lowers model complexity, and improves training efficiency and generalization ability. The Relief algorithm filters key features, avoiding the impact of excessive redundant features on model performance. Improved fault detection accuracy and operational efficiency help reduce fault frequency and maintenance costs, enhancing the overall economic benefits of wind turbines. Timely and accurate detection of blade icing faults allows for proactive measures, reducing downtime and maintenance costs caused by faults. Attached Figure Description
[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0032] Figure 1 This is a diagram of the cross-LSTM prediction model of the present invention;
[0033] Figure 2 This is a diagram of the sequence LSTM classification model of the present invention;
[0034] Figure 3 This is a flowchart of the icing detection process of the present invention. Detailed Implementation
[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0036] This invention discloses a method for detecting icing faults in wind turbine blades.
[0037] Please refer to the following: Figure 1-Figure 3 As shown, a method for detecting icing faults in wind turbine blades is characterized by comprising the following steps:
[0038] The Relief algorithm is used to select features from the multi-dimensional feature data collected by the wind turbine SCADA system. Feature weights are evaluated based on their contribution to distinguishing between blade icing and normal conditions, and key features are selected. The feature weight vector update formula in the Relief algorithm is as follows: Where W(A) is the weight of feature A, m is the number of repetitions, and diff(A,R,H) is the weight of feature A. i ) represents sample R and similar samples H i Distance on feature A, diff(A,R,M) i Let R be a sample and M be a different sample. i Distance on feature A.
[0039] Based on the selected feature data and its visualization, and combined with the icing pattern, a new feature X1 of wind speed-power and a new feature X2 of ambient temperature-cabin temperature were constructed.
[0040] A cross-LSTM prediction model is constructed, which predicts one set of data based on the characteristics of multiple sets of data, generating a residual signal; the steps for constructing the cross-LSTM prediction model include:
[0041] The dataset containing t characteristic variable measurements and n observation points is filtered to remove high-frequency noise, and after normalization, a new dataset x with dimension t*n is formed.
[0042] The sliding window algorithm is introduced, with a sliding window length of 10. The feature data from time t-10 to time t are input to obtain the prediction data at time t.
[0043] The test set is input into the prediction model to obtain the predicted value of the corresponding sensor. The test value is then denormalized to obtain the residual signal between the predicted value and the actual value.
[0044] A sequence classification model is constructed, using the product of the number of feature variables and the residual feature signal as input data to determine whether icing occurs at a specific moment in the time series. The sequence classification model is iteratively optimized using a stochastic momentum gradient descent algorithm, the formula of which is: Where θ represents the weight vector, k represents the number of updates, and (η) represents the update rate. The gradient represents the loss function, and r represents the weighted effect of the previous gradient on the current gradient.
[0045] A trained cross-LSTM network is used to predict test set data under normal conditions, and the temporal features of the residuals are extracted to construct a new feature matrix. To reduce the risk of overfitting during model training, a regularization term E is added to the original loss function. R The loss function formula is: E(θ) = E M (θ)+λE R (θ);
[0046] Regularization is calculated as follows: E R (θ)=∑ w∈WΩ(w); where λ represents the regularization factor and W represents the weight vector. The time-domain features of the extracted residuals include the mean, peak value, peak-to-peak value, sum of squares, and root mean square magnitude, calculated using the following formulas:
[0047] Mean:
[0048] Peak value: x peak =max(X);
[0049] Peak-to-peak value: x p-p =max(X) - min(X);
[0050] Sum of squares:
[0051] Root mean square value: Where, x i Let m be the number of residual signal data points, and X be the set of residual signal data.
[0052] Cross-entropy was selected as the loss function, and a classification model was trained using a newly constructed feature matrix. The confusion matrix for icing classification was obtained using test set data. In icing fault detection, icing classification labels were introduced, with 0 representing normal blade operation and 1 representing icing fault. The sequence classification model structure includes an input layer, an LSTM layer for capturing long-term and short-term dependencies in the time series, a fully connected layer, and a classification layer, with a loop count set to 6000.
[0053] The specific implementation method is as follows:
[0054] Data Acquisition and Preprocessing: Multi-dimensional feature data, including wind speed, ambient temperature, motor power, and nacelle temperature, are collected using the wind turbine SCADA system. The collected data is filtered to remove high-frequency noise, and then normalized to map the data to the [0,1] interval, resulting in a new dataset x of dimension t*n.
[0055] Feature selection and reconstruction: The Relief algorithm is used to select features from the preprocessed data, calculate the weight of each feature, and filter out key features with higher weights. Based on the feature data and its visualization, and combined with icing patterns, new features X1 (wind speed-power) and X2 (ambient temperature-cabin temperature) are constructed.
[0056] Model Training and Optimization: A cross-LSTM prediction model and a sequence LSTM classification model are constructed. The cross-prediction model is trained using training data, and its parameters are adjusted to ensure accurate prediction. The residual signal output by the cross-prediction model is processed to extract five time-domain features: mean, peak value, peak-to-peak value, sum of squares, and root mean square magnitude, constructing a new feature matrix. This feature matrix is then used to train the sequence LSTM classification model. Iterative training is performed using the stochastic momentum gradient descent optimization algorithm, and a regularization term is added to the loss function to reduce the risk of overfitting.
[0057] Fault detection and assessment: The test set data is input into a trained cross-LSTM prediction model to obtain the residuals between the predicted and actual values. Temporal features of the residuals are extracted, a feature matrix is constructed, and this matrix is input into a trained sequence LSTM classification model to obtain the icing classification results. The detection accuracy of the model is evaluated using a confusion matrix, and the model is further optimized based on the evaluation results.
[0058] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for detecting icing faults in wind turbine blades, characterized in that, The following steps are involved: The Relief algorithm was used to select features from the multi-dimensional feature data collected by the wind turbine SCADA system. The feature weights were evaluated based on the contribution of the features to distinguishing between blade icing and normal conditions, and key features were selected. Based on the selected feature data and its visualization, and combined with the icing pattern, a new feature X1 of wind speed-power and a new feature X2 of ambient temperature-cabin temperature were constructed. Construct a cross-LSTM prediction model, predict one set of data based on the characteristics of multiple sets of data, and generate residual signals; Construct a sequence classification model, using the product of the number of feature variables and the residual feature signal as input data, to determine whether icing occurs at a specific moment in the time series; The trained cross-LSTM network is used to predict the test set data under normal conditions, and the temporal features of the residuals are extracted to construct a new feature matrix. Cross-entropy was selected as the loss function, and the classification model was trained using the newly constructed feature matrix. The confusion matrix for ice accretion classification was obtained through the test set data.
2. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, The feature weight vector update formula in the Relief algorithm is as follows: Where W(A) is the weight of feature A, m is the number of repetitions, and diff(A,R,H) is the weight of feature A. i ) represents sample R and similar samples H i Distance on feature A, diff(A,R,M) i Let R be a sample and M be a different sample. i Distance on feature A.
3. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, The steps for constructing the cross-LSTM prediction model include: The dataset containing t characteristic variable measurements and n observation points is filtered to remove high-frequency noise, and after normalization, a new dataset x with dimension t*n is formed. The sliding window algorithm is introduced, with a sliding window length of 10. The feature data from time t-10 to time t are input to obtain the prediction data at time t. The test set is input into the prediction model to obtain the predicted value of the corresponding sensor. The test value is then denormalized to obtain the residual signal between the predicted value and the actual value.
4. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, The sequence classification model is iteratively optimized using the stochastic momentum gradient descent algorithm, the formula of which is: Where θ represents the weight vector, k represents the number of updates, and (η) represents the update rate. The gradient represents the loss function, and r represents the weighted effect of the previous gradient on the current gradient.
5. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, To reduce the risk of overfitting during model training, a regularization term E is added to the original loss function. R The loss function formula is: E(θ) = E M (θ)+λE R (θ); Regularization is calculated as follows: E R (θ)=∑ w∈W Ω(w); where λ represents the regularization factor and W represents the weight vector.
6. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, The time-domain features of the extracted residuals include the mean, peak value, peak-to-peak value, sum of squares, and root mean square magnitude, calculated using the following formulas: Mean: Peak value: x peak =max(X); Peak-to-peak value: x p-p =max(X) - min(X); sum of squares: Root mean square value: Among them, x i Let m be the number of residual signal data points, and X be the set of residual signal data.
7. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, In the detection of icing faults, icing classification labels are introduced, with 0 representing the normal operating state of the blade and 1 representing the icing fault state of the blade.
8. The method for detecting icing faults in wind turbine blades according to claim 1, characterized in that, The sequence classification model structure includes an input layer, an LSTM layer for capturing long-term and short-term dependencies in the time series, a fully connected layer, and a classification layer, with the number of loop steps set to 6000.