A method and apparatus for detecting and jointly interpreting abnormal states of wind turbine generator sets

Abstract

This invention discloses a method and apparatus for detecting and jointly interpreting abnormal states of wind turbine generator sets, comprising: acquiring multi-dimensional operational data from the SCADA system of the wind turbine generator set and preprocessing it; clustering using the k-means clustering algorithm, selecting the central features of the calculated central clusters to construct a new feature dataset; training a wind turbine generator set abnormal state detection model using a classification decision boosting algorithm to detect and identify abnormal states; establishing a simplified model output by the detection model, using the SHAP algorithm to calculate the global contribution of input features to the abnormal state detection results and provide a global interpretation; establishing local models of detection model instances, using the LIME algorithm to generate local interpretations for specific abnormal instances, and analyzing the contribution of key features in specific abnormal instances. This invention improves the accuracy and interpretability of wind turbine generator set anomaly detection, thereby providing solid data support and theoretical foundation for the maintenance and optimization of complex systems.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, particularly to the field of wind turbine generator condition monitoring and anomaly detection technology, and more specifically to a method and apparatus for detecting and interpreting abnormal conditions of wind turbine generators. Background Technology

[0002] As wind power accounts for an increasing proportion of the global energy structure, wind farms are expanding in scale, and the number and complexity of wind turbine generators are also increasing significantly. In large wind farms, wind turbine generators often operate in clusters, and their operating status is directly related to the power generation efficiency and safety of the entire wind farm. However, due to the complex operating environment and diverse fault types of wind turbine generators, traditional monitoring and anomaly detection methods face many challenges in practical applications.

[0003] Traditional methods mainly rely on the construction of physical models and rule-based anomaly detection. Physical models usually require accurate equipment parameters and operating environment information, which are often difficult to obtain in practice or limited by the complexity and nonlinear characteristics of the model, thus limiting the accuracy and timeliness of anomaly detection. Rule-based detection methods rely on predefined thresholds and empirical rules, which makes it difficult to cope with the detection needs of new or complex faults.

[0004] With the development of data-driven technologies, machine learning-based anomaly detection methods can automatically learn the normal operating modes and abnormal characteristics of wind turbine generators by utilizing a large amount of historical operating data, thus achieving real-time detection of abnormal states. However, traditional machine learning methods often face two main problems in application: First, when dealing with high-dimensional data and complex features, they are prone to falling into a "black box" state, lacking sufficient interpretability, making it difficult for operation and maintenance personnel to understand and trust the detection results; Second, missing data and noise problems can significantly affect the performance of the detection model, leading to a decrease in the accuracy and reliability of anomaly detection.

[0005] Therefore, how to provide a method and device for detecting and jointly interpreting the abnormal state of wind turbine generator sets that can improve the accuracy of anomaly detection, overcome the lack of interpretability caused by the "black box" effect of traditional machine learning methods, and deal with missing values ​​and noise in high-dimensional data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and apparatus for detecting and jointly interpreting abnormal states of wind turbine generator sets to solve some of the technical problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for detecting and jointly interpreting abnormal states of wind turbine generator sets includes the following steps:

[0009] S1. Obtain multi-dimensional operational data from the SCADA system of the wind turbine generator set and perform preprocessing;

[0010] S2. The preprocessed wind turbine SCADA data is clustered using the k-means clustering algorithm, and the central features of the calculated central clusters are selected to construct a new feature dataset;

[0011] S3. A new feature dataset is constructed based on the extracted central feature data. A classification decision enhancement algorithm is used to train a wind turbine abnormal state detection model to detect and identify abnormal states.

[0012] S4. Establish a simplified model for the output of the wind turbine abnormal state detection model, use the SHAP algorithm to calculate the global contribution of the input features to the abnormal state detection results, and provide a global interpretation;

[0013] S5. Establish a local model for the wind turbine abnormal state detection model instance, use the LIME algorithm to fit the local linear model, generate a local interpretation for the specific abnormal instance, and analyze the contribution of key features to the specific abnormal instance.

[0014] Preferably, the preprocessing in step S1 includes the following:

[0015] S11. Based on the SCADA data of the wind turbine generator set, the mode imputation method is used to fill in the missing classification features in the data;

[0016] S12. Based on the SCADA data of the wind turbine generator set, the mean imputation method is used to fill in the missing numerical features in the data;

[0017] S13. Use the completed SCADA data of the wind turbine generator set, i.e., the preprocessed SCADA data of the wind turbine generator set, as the pre-training dataset.

[0018] Preferably, the specific content of step S2 is as follows:

[0019] S21. The elbow method is used to determine the optimal number of clusters. By calculating the total sum of squared deviations under different k values, the k value with the significantly reduced total sum of squared deviations is selected as the optimal number of clusters.

[0020] S22. Based on the determined optimal number of clusters, the k-means clustering algorithm is applied to cluster the wind turbine features in the pre-training dataset of the pre-processed wind turbine SCADA data.

[0021] S23. Based on the completed wind turbine feature clustering results, calculate the distance from the feature to the cluster center, and select the feature with the smallest distance as the central feature;

[0022] S24. Use the central features selected by clustering to construct a new feature set to reduce collinearity and redundancy among features.

[0023] Preferably, the specific content of step S3 is as follows:

[0024] S31. A classification decision boosting algorithm is used to establish an abnormal state detection model for wind turbine generators. Starting from a constant model, the classification prediction value is initialized to be the mean of the target value.

[0025] S32. Calculate the gradient and diagonal elements of the Hessian matrix for each sample in each iteration based on the initialized classification prediction values;

[0026] S33. Construct a histogram for each split point of each feature, calculate the gain of each split point using the histogram values, select the split point with the largest gain for splitting, and then construct a decision tree based on the selected best split point;

[0027] S34. Use the currently constructed tree to update the model's predictions. After multiple iterations, the final model is a weighted sum of all base learners.

[0028] Preferably, the specific content of step S4 is as follows:

[0029] S41. Construct a simplified model of the abnormal state detection model for wind turbine generator sets, and approximate the output of the original model through a linear combination.

[0030] S42. Calculate the differences between the simplified model and the original model to satisfy local accuracy, missing data, and consistency;

[0031] S43. Calculate the SHAP value of each feature in the simplified model, perform global visualization based on the SHAP value of each feature, and display the ranking of the features that affect the abnormal state of the wind turbine generator and the correlation between the features in the figure.

[0032] Preferably, the specific content of step S5 is as follows:

[0033] S51 generates a new dataset by locally perturbing the original data in order to simulate the behavior of the model in the local space.

[0034] S52. The Euclidean distance is used to measure the similarity between the generated new dataset samples and the original data samples. It is used to assign weights to each perturbation sample. The higher the weight, the more similar the perturbation sample is to the original sample.

[0035] S53. A weighted linear regression model is used to fit the perturbed samples and the corresponding model output. By minimizing the weighted loss function, a local explanatory model is obtained.

[0036] S54. Based on the constructed local interpretation model, analyze a single instance in the SCADA data of the wind turbine and calculate the probability of abnormal state of a single instance.

[0037] Preferably, the method for detecting and interpreting abnormal states of wind turbine generator sets further includes: S6. Visualizing the abnormal state detection results and interpretation results, and triggering an alarm mechanism based on preset rules.

[0038] A wind turbine generator abnormal state detection and joint interpretation system, based on the aforementioned wind turbine generator abnormal state detection and joint interpretation method, includes a data acquisition module, a preprocessing module, a feature selection module, an abnormal state detection module, a global interpretation module, a local interpretation module, and a visualization module;

[0039] The data acquisition module is used to obtain multi-dimensional operational data from the SCADA system of the wind turbine generator set;

[0040] The preprocessing module is used to preprocess the acquired multi-dimensional runtime data;

[0041] The feature selection module is used to cluster the preprocessed wind turbine SCADA data using the k-means clustering algorithm, and select the central features of the calculated central clusters to construct a new feature dataset.

[0042] The abnormal state detection module is used to construct a new feature dataset based on the extracted central feature data, and to train the wind turbine abnormal state detection model using a classification decision enhancement algorithm to detect and identify abnormal states.

[0043] The global interpretation module is used to establish a simplified model of the output of the wind turbine abnormal state detection model, calculate the global contribution of the input features to the abnormal state detection results using the SHAP algorithm, and provide a global interpretation.

[0044] The local interpretation module is used to establish a local model of the wind turbine abnormal state detection model instance. It uses the LIME algorithm to generate a local interpretation for a specific abnormal instance by fitting a local linear model and analyzes the contribution of key features in the specific abnormal instance.

[0045] The visualization module is used to visualize the detection and interpretation results and trigger alarm mechanisms.

[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for detecting and jointly interpreting abnormal states of a wind turbine generator set.

[0047] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the aforementioned method for detecting and jointly interpreting abnormal states of a wind turbine generator set.

[0048] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method and device for detecting and jointly interpreting abnormal states of wind turbine generator sets. Based on data-driven machine learning methods and interpretability models, by integrating KMeans clustering algorithm and classification decision boosting algorithm, it can effectively extract the central features in the operating data of wind turbine generator sets, construct an efficient anomaly detection model, and at the same time use SHAP and LIME algorithms to provide global and local result interpretations, thereby overcoming the problems of "black box" models in traditional methods being difficult to understand and lacking transparency.

[0049] To effectively address different types of missing data and noise, mode imputation and mean imputation methods are used, which improves data quality and enhances the robustness and applicability of the model.

[0050] In the complex and multidimensional operating environment of wind turbine generators, this invention can not only accurately detect abnormal conditions, but also provide detailed explanations and analyses for operation and maintenance personnel, quickly respond to abnormal situations, reduce the risk of failure, and improve the overall operating efficiency and stability of wind farms. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of an abnormal state detection and joint interpretation method for wind turbine generator sets provided by the present invention;

[0053] Figure 2 A schematic diagram of the framework for the wind turbine generator abnormal state detection and joint interpretation method provided by the present invention;

[0054] Figure 3 This is a schematic diagram illustrating the calculation of central features using the k-means clustering algorithm provided by this invention.

[0055] Figure 4This is a schematic diagram of feature selection for the k-means clustering algorithm provided by the present invention;

[0056] Figure 5 A schematic diagram illustrating the accuracy of the wind turbine abnormal state detection model based on the classification decision enhancement algorithm provided by this invention;

[0057] Figure 6 A schematic diagram of the confusion matrix of the wind turbine abnormal state detection model based on the classification decision enhancement algorithm provided by the present invention;

[0058] Figure 7 This is a schematic diagram illustrating the predicted recall rate of the wind turbine abnormal state detection model based on the classification decision enhancement algorithm provided by this invention.

[0059] Figure 8 A schematic diagram illustrating the global feature interpretation of SHAP provided by this invention;

[0060] Figure 9 This is a schematic diagram illustrating the correlation between SHAP features provided by the present invention.

[0061] Figure 10 This is a schematic diagram of LIME single instance abnormal state prediction provided by the present invention;

[0062] Figure 11 This is a schematic diagram illustrating the influence characteristics of a single LIME instance provided by the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] This invention discloses a method for detecting and jointly interpreting abnormal states of wind turbine generator sets, such as... Figure 1 and Figure 2 This includes the following steps:

[0065] S1. Obtain multi-dimensional operational data from the SCADA system of the wind turbine generator set and perform preprocessing;

[0066] S2. The preprocessed wind turbine SCADA data is clustered using the k-means clustering algorithm, and the central features of the calculated central clusters are selected to construct a new feature dataset;

[0067] S3. A new feature dataset is constructed based on the extracted central feature data. A classification decision enhancement algorithm is used to train a wind turbine abnormal state detection model to detect and identify abnormal states.

[0068] S4. Establish a simplified model for the output of the wind turbine abnormal state detection model, use the SHAP algorithm to calculate the global contribution of the input features to the abnormal state detection results, and provide a global interpretation;

[0069] S5. Establish a local model for the wind turbine abnormal state detection model instance, use the LIME algorithm to fit the local linear model, generate a local interpretation for the specific abnormal instance, and analyze the contribution of key features to the specific abnormal instance.

[0070] In this embodiment, multi-dimensional operating data includes, but is not limited to, parameters such as wind speed, temperature, current, and voltage.

[0071] To further implement the above technical solution, the specific content of the preprocessing in step S1 includes:

[0072] S11. Based on the SCADA data of the wind turbine generator set, the mode imputation method is used to fill in the missing classification features in the data;

[0073] S12. Based on the SCADA data of the wind turbine generator set, the mean imputation method is used to fill in the missing numerical features in the data;

[0074] S13. Use the completed SCADA data of the wind turbine generator set, i.e., the preprocessed SCADA data of the wind turbine generator set, as the pre-training dataset.

[0075] In this embodiment, the mode imputation method is specifically as follows:

[0076]

[0077] Among them, Mode X Let I(x) be the mode of feature X, and c be a specific category of the categorical feature. i =c) is an indicator function, when x i If the category is c, the indicator function has a value of 1; otherwise, it has a value of 0.

[0078] The mean-filling method is as follows:

[0079]

[0080] Where X is the set of numerical features, n is the number of all values ​​in the set of numerical features, m is the number of missing values ​​in the set of numerical features, and x i For the i-th non-missing value, μ X Let X be the mean of the feature X.

[0081] To further implement the above technical solution, the specific content of step S2 is as follows:

[0082] S21. The elbow method is used to determine the optimal number of clusters. By calculating the total sum of squared deviations under different k values, the k value with the significantly reduced total sum of squared deviations is selected as the optimal number of clusters.

[0083] S22. Based on the determined optimal number of clusters, apply the k-means clustering algorithm to cluster the wind turbine features in the pre-training dataset of the pre-processed wind turbine SCADA data, such as... Figure 3 ;

[0084] S23. Based on the completed wind turbine feature clustering results, calculate the distance from each feature to the cluster center, and select the feature with the smallest distance as the cluster center feature, such as... Figure 4 ;

[0085] S24. Use the central features selected by clustering to construct a new feature set to reduce collinearity and redundancy among features.

[0086] In this embodiment, in step S21, the total sum of squared deviations is:

[0087]

[0088] Among them, C i Let μ be the i-th cluster, and μ be the center point of the i-th cluster.

[0089] In step S23, the selection equation for the central feature is:

[0090]

[0091] Where, x * Let i be the central feature, i be the i-th cluster, μ be the center point of the i-th cluster, and the intra-cluster features be {x1, x2, ..., x...}. m}

[0092] like Figure 3 This shows the clustering results after reducing high-dimensional data to a three-dimensional space through principal component analysis (PCA). Different colors represent different cluster labels, and the coordinate axes represent the three principal components. The position of the data points is determined by the scores of these principal components. The colors from purple to yellow indicate different cluster labels, showing the distribution and clustering of data points in the reduced space.

[0093] like Figure 4The graph shows the variance variation of different features. The horizontal axis represents the feature index, and the vertical axis represents the variance of each feature. The blue line in the graph represents the variance of all original features, and the red dots mark the selected features. The selected features are distributed in the region with larger variance, indicating that these features may have high representativeness or discriminative power in the overall data.

[0094] To further implement the above technical solution, the specific content of step S3 is as follows:

[0095] S31. A classification decision boosting algorithm is used to establish an abnormal state detection model for wind turbine generators. Starting from a constant model, the classification prediction value is initialized to be the mean of the target value.

[0096] The equation for initializing the mean of the classification predictions as the target value is:

[0097]

[0098] in, For the predicted value, y i For a single target value, i is the number of the specific target value, and n is the total number of target values;

[0099] S32. Calculate the gradient and diagonal elements of the Hessian matrix for each sample in each iteration based on the initialized classification prediction values;

[0100] The equations for calculating the gradient of the sample and the diagonal elements of the Hessian matrix are as follows:

[0101]

[0102] Among them, g j For gradient, Let y be the loss function. j The actual label value of the sample. h is the predicted value of the model. j These are the diagonal elements of the Hessian matrix;

[0103] S33. Construct a histogram for each split point of each feature, calculate the gain of each split point using the histogram values, select the split point with the largest gain for splitting, and then construct a decision tree based on the selected best split point;

[0104] The equation for constructing histogram values ​​is:

[0105]

[0106] Where b represents a candidate value for the split point;

[0107] The equation for calculating the gain at each split point using histogram values ​​is:

[0108]

[0109] Where b is a candidate value for the split point, and Gain is the gain function. The sum of gradients for all samples in the left child node. The sum of gradients for all samples in the right child node. The sum of the Hessian matrices for all samples in the left child node. Let be the sum of the Hessian matrices of all samples in the right child node, where λ is the regularization parameter and γ is the splitting penalty parameter;

[0110] S34. Use the currently constructed tree to update the model's predictions. After multiple iterations, the final model is a weighted sum of all base learners.

[0111] The iterative training equation is:

[0112]

[0113] The weighted sum equation for the base learners is:

[0114]

[0115] Where η is the learning rate, f m is the weak learner trained in the m-th round.

[0116] like Figure 5 As shown, the ROC curves of the constructed wind turbine generator abnormal state detection model are displayed under different categories. The horizontal axis represents the false positive rate, and the vertical axis represents the true positive rate. The figure includes the ROC curves for categories 0 and 1, both with an AUC of 0.99, indicating that the model has excellent classification performance in these two categories. In addition, the AUC of the micro-average ROC curve is 1.00, and the AUC of the macro-average ROC curve is 0.99. Overall, the constructed wind turbine generator abnormal state detection model has a near-perfect classification performance in all categories, indicating that it has extremely high discrimination ability.

[0117] like Figure 6As shown, this is the classification report of the constructed wind turbine generator abnormal state detection model under different categories, displaying precision, recall, F1 score, and number of supports. The horizontal axis represents these four evaluation indicators, and the vertical axis represents the classification results of category 0 and category 1: For category 1, the precision is 0.997, the recall is 0.959, the F1 score is 0.978, and the number of supports is 413; for category 0, the precision is 0.988, the recall is 0.999, the F1 score is 0.993, and the number of supports is 1369. Overall, the model performs very well in classifying both types of data, especially in category 0, where the F1 score is close to perfect, indicating that the constructed wind turbine generator abnormal state detection model has a relatively balanced classification effect in most categories, and the number of supports also shows that the sample size of category 0 is large.

[0118] like Figure 7 As shown, the precision-recall curves of the constructed wind turbine generator abnormal state detection model are displayed. The horizontal axis represents recall, and the vertical axis represents precision. The solid line represents the binary precision-recall curve of the model, and the dashed line represents the average precision of 0.99. It can be seen that the precision of the model remains close to 1.0 in a high recall range, indicating that the constructed wind turbine generator abnormal state detection model can maintain high precision and recall in most cases. The rapid drop in the curve only occurs in the region close to the recall of 1.0, indicating that in very few cases, the precision of the model will decrease slightly, but the overall performance is still very good.

[0119] To further implement the above technical solution, the specific content of step S4 is as follows:

[0120] S41. Construct a simplified model of the abnormal state detection model for wind turbine generator sets, and approximate the output of the original model through a linear combination.

[0121] The equation for the linear combination is:

[0122]

[0123] Where g(z′) is the output of the global interpretation model, φ0 is a constant term representing the model output when all features are missing, and φ i z is the importance value of feature i, representing the contribution of the feature to the model output; z′ is the simplified input variable, usually binary 0 or 1, indicating whether the feature exists; M is the total number of features.

[0124] S42. Calculate the differences between the simplified model and the original model to satisfy local accuracy, missing data, and consistency;

[0125] The local accuracy equation is:

[0126]

[0127] The missing value equation is:

[0128]

[0129] The consistency equation is:

[0130]

[0131] Where f(z′) is the output of the original model with simplified input z′, f′(z′) is the output of the improved model with simplified input z′, and z′\i is the feature i removed from the simplified input z′;

[0132] S43. Calculate the SHAP value of each feature in the simplified model, and perform global visualization based on the SHAP value of each feature. Display the ranking of feature importance and the correlation between features in relation to the abnormal state of the wind turbine generator in a graph, such as... Figure 8 and Figure 9 ;

[0133] The equation for calculating the SHAP value is:

[0134]

[0135] Where, φ i (f, x) represents the SHAP value of feature i in model f, indicating its contribution to the model output, φ i (f′, x) represents the SHAP value of feature i in the improved model f′.

[0136] Figure 8 This graph illustrates feature importance analysis based on SHAP (Shapley Additive Explanations) values. The horizontal axis represents the SHAP values ​​of the model output, indicating the contribution of each feature to the model's prediction results. The vertical axis lists different features, such as "Gearbox_Oil-2_Temperature" and "Tower Acceleration Lateral". Each point in the graph represents the SHAP value of a data point, and the color indicates the strength of the feature value, from blue (low) to red (high). The shape corresponding to each feature represents the distribution of its SHAP value; a wider area indicates a larger number of samples within that range. This graph shows the positive and negative impact of certain features on the model output within different value ranges. Redder points indicate a greater impact of high feature values ​​on positive SHAP values, while bluer points indicate a greater impact of low feature values ​​on negative SHAP values. This graph allows observation of the contribution of each feature to the model's predictions and how these contributions change across different feature value ranges.

[0137] Figure 9The graph shows the SHAP value distribution of the feature "TemperatureBatteryBox-1" to explain its contribution to the model's prediction results. The horizontal axis represents the value range of "TemperatureBatteryBox-1", the vertical axis represents the corresponding SHAP value, and the color bar represents the value of another feature, "Gearbox_Oil-2_Temperature", gradually changing from blue (low temperature) to red (high temperature). As can be seen from the graph, the vast majority of data points are concentrated in the lower value range of "TemperatureBatteryBox-1", and the SHAP values ​​of these points are close to 0, indicating that within this range, the feature contributes little to the model's prediction. In the high value region (close to 100,000), the SHAP values ​​of the points increase or decrease slightly, indicating that "TemperatureBatteryBox-1" has a greater impact on the model's prediction in these cases. This graph reflects the influence of this feature on the model output in different value ranges. The color bar, combined with the temperature value of another feature, shows the influence of the interaction between multiple features on the model's prediction.

[0138] To further implement the above technical solution, the specific content of step S5 is as follows:

[0139] S51 generates a new dataset by locally perturbing the original data in order to simulate the behavior of the model in the local space.

[0140] The equation for generating the new dataset is:

[0141] z=x+ò

[0142] Where z is the new data, ò is the disturbance value, and x is the original data;

[0143] S52. The Euclidean distance is used to measure the similarity between the generated new dataset samples and the original data samples. It is used to assign weights to each perturbation sample. The higher the weight, the more similar the perturbation sample is to the original sample.

[0144] The similarity calculation equation is:

[0145]

[0146] Where, π x (z) is the weight of sample z, distance(x,z) is the distance between samples x and z, and σ is an adjustment parameter used to control the rate of similarity decay;

[0147] S53. A weighted linear regression model is used to fit the perturbed samples and the corresponding model output. By minimizing the weighted loss function, a local explanatory model is obtained.

[0148] The equation for minimizing the weighted loss function is:

[0149]

[0150] Where ξ(g, f, π) x ) represents the weighted squared loss function, g represents the local explanation model, f represents the constructed complex prediction model, and π represents the total squared loss function. x (z) is the weight used to measure the similarity between the perturbation sample z and the explanation sample x, and Ω(g) is the complexity penalty term used to avoid overfitting;

[0151] S54. Based on the constructed local interpretation model, analyze individual instances in the SCADA data of wind turbine units, and calculate the probability of abnormal states for each individual instance, such as... Figure 10 and 11 .

[0152] Figure 10 This chart illustrates the key features and their contributions for two different categories (0 and 1). The left side of the chart represents category 0, and the right side represents category 1. Each bar represents the contribution of a feature to that category; the larger the value, the longer the bar, indicating a greater influence of the feature on that category. In category 0, the "State and Fault" feature has a value of 2, the highest contribution (0.85), followed by the "Gearbox_T1_High_Speed_Shaft_Temperature" feature (contribution of 0.06). Below the bars for category 0, the classification probabilities are displayed; a blue "0" indicates a probability of 1.00 for category 0 and 0 for category 1. In category 1, the "Gearbox_Oil_Temperature" feature has a value of 46, a relatively low contribution (0.03). Other features, such as "Tower Acceleration," have a lower contribution. The contributions of "Lateral" and "Gearbox_Oil-1_Temperature" are also relatively small; this indicates that the influence of features in category 1 is more dispersed. This figure is used to explain the model's prediction process for the two categories, highlighting the main features of each category and their contribution to the prediction results.

[0153] Figure 11This table illustrates the correspondence between features and values. The left side lists the feature names and their values, while the right side indicates whether these feature values ​​are true. Feature names in the figure include "State and Fault" and "Gearbox_T1_High_Speed_Shaft_Temperature," with each feature followed by a specific numerical value, such as "State and Fault = 2" and "Gearbox_Oil_Temperature = 46." All feature values ​​are marked as "True," indicating that these features and their corresponding values ​​are true under the current circumstances. The alternating blue and orange background in the figure helps to distinguish different features and enhances readability. This tabular format is used to display the specific feature values ​​of the model input or to explain the role of features in a particular prediction.

[0154] To further implement the above technical solution, a method for detecting and interpreting abnormal states of wind turbine generator sets also includes: S6. Visualizing the abnormal state detection results and interpretation results, and triggering an alarm mechanism based on preset rules.

[0155] A wind turbine generator abnormal state detection and joint interpretation system, based on a wind turbine generator abnormal state detection and joint interpretation method, includes a data acquisition module, a preprocessing module, a feature selection module, an abnormal state detection module, a global interpretation module, a local interpretation module, and a visualization module;

[0156] The data acquisition module is used to obtain multi-dimensional operational data from the SCADA system of the wind turbine generator set;

[0157] The preprocessing module is used to preprocess the acquired multi-dimensional runtime data;

[0158] The feature selection module is used to cluster the preprocessed wind turbine SCADA data using the k-means clustering algorithm, and select the central features of the calculated central clusters to construct a new feature dataset.

[0159] The abnormal state detection module is used to construct a new feature dataset based on the extracted central feature data, and to train the wind turbine abnormal state detection model using a classification decision enhancement algorithm to detect and identify abnormal states.

[0160] The global interpretation module is used to establish a simplified model of the output of the wind turbine abnormal state detection model, calculate the global contribution of the input features to the abnormal state detection results using the SHAP algorithm, and provide a global interpretation.

[0161] The local interpretation module is used to establish a local model of the wind turbine abnormal state detection model instance. It uses the LIME algorithm to generate a local interpretation for a specific abnormal instance by fitting a local linear model and analyzes the contribution of key features in the specific abnormal instance.

[0162] The visualization module is used to visualize the detection and interpretation results and trigger alarm mechanisms.

[0163] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for detecting and jointly interpreting abnormal states of a wind turbine generator set.

[0164] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements a method for detecting and jointly interpreting abnormal states of a wind turbine generator set.

[0165] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0166] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting and jointly interpreting abnormal states of wind turbine generator sets, characterized in that, Includes the following steps: S1. Obtain multi-dimensional operational data from the SCADA system of the wind turbine generator set and perform preprocessing; S2. The preprocessed wind turbine SCADA data is clustered using the k-means clustering algorithm, and the central features of the calculated central clusters are selected to construct a new feature dataset; S3. A new feature dataset is constructed based on the extracted central feature data. A classification decision enhancement algorithm is used to train a wind turbine abnormal state detection model to detect and identify abnormal states. S4. Establish a simplified model for the output of the wind turbine abnormal state detection model, use the SHAP algorithm to calculate the global contribution of the input features to the abnormal state detection results, and provide a global interpretation; S5. Establish a local model for the wind turbine abnormal state detection model instance, use the LIME algorithm to fit the local linear model, generate a local interpretation for the specific abnormal instance, and analyze the contribution of key features to the specific abnormal instance. The specific content of step S3 is as follows: S31. A classification decision boosting algorithm is used to establish an abnormal state detection model for wind turbine generators. Starting from a constant model, the classification prediction value is initialized to be the mean of the target value. S32. Calculate the gradient and diagonal elements of the Hessian matrix for each sample in each iteration based on the initialized classification prediction values; S33. Construct a histogram for each split point of each feature, calculate the gain of each split point using the histogram values, select the split point with the largest gain for splitting, and then construct a decision tree based on the selected best split point; S34. Use the currently constructed tree to update the model's predictions. After multiple iterations, the final model is a weighted sum of all base learners. The specific content of step S4 is as follows: S41. Construct a simplified model of the abnormal state detection model for wind turbine generator sets, and approximate the output of the original model through a linear combination. S42. Calculate the differences between the simplified model and the original model to satisfy local accuracy, missing data, and consistency; S43. Calculate the SHAP value of each feature in the simplified model, perform global visualization based on the SHAP value of each feature, and display the ranking of the features that affect the abnormal state of the wind turbine generator and the correlation between the features in the figure. The specific content of step S5 is as follows: S51 generates a new dataset by locally perturbing the original data in order to simulate the behavior of the model in the local space. The equation for generating the new dataset is: ； Where z represents new data. Here is the perturbation value, and x is the original data; S52. The Euclidean distance is used to measure the similarity between the generated new dataset samples and the original data samples. It is used to assign weights to each perturbation sample. The higher the weight, the more similar the perturbation sample is to the original sample. The similarity calculation equation is: ； in, The weights of sample z, The distance between samples x and z. This is an adjustable parameter used to control the rate at which similarity decays; S53. A weighted linear regression model is used to fit the perturbed samples and the corresponding model output. By minimizing the weighted loss function, a local explanatory model is obtained. The equation for minimizing the weighted loss function is: ； in, For weighted squared loss function, For local interpretation models, For the complex prediction model that has been constructed, , where is the weight used to measure the similarity between the perturbation sample z and the explanation sample x. This is a complexity penalty term used to avoid overfitting; S54. Based on the constructed local interpretation model, analyze a single instance in the SCADA data of the wind turbine and calculate the probability of abnormal state of a single instance.

2. The method for detecting and jointly interpreting abnormal states of a wind turbine generator set according to claim 1, characterized in that, The specific preprocessing steps in step S1 include: S11. Based on the SCADA data of the wind turbine generator set, the mode imputation method is used to fill in the missing classification features in the data; S12. Based on the SCADA data of the wind turbine generator set, the mean imputation method is used to fill in the missing numerical features in the data; S13. Use the completed SCADA data of the wind turbine generator set, i.e., the preprocessed SCADA data of the wind turbine generator set, as the pre-training dataset.

3. The method for detecting and jointly interpreting abnormal states of a wind turbine generator set according to claim 1, characterized in that, The specific content of step S2 is as follows: S21. The elbow method is used to determine the optimal number of clusters. By calculating the total sum of squared deviations under different k values, the k value with the significantly reduced total sum of squared deviations is selected as the optimal number of clusters. S22. Based on the determined optimal number of clusters, the k-means clustering algorithm is applied to cluster the wind turbine features in the pre-training dataset of the pre-processed wind turbine SCADA data. S23. Based on the completed wind turbine feature clustering results, calculate the distance from the feature to the cluster center, and select the feature with the smallest distance as the central feature; S24. Use the central features selected by clustering to construct a new feature set to reduce collinearity and redundancy among features.

4. The method for detecting and jointly interpreting abnormal states of a wind turbine generator set according to claim 1, characterized in that, Also includes: S6. Visualize the abnormal state detection results and interpretation results, and trigger the alarm mechanism based on preset rules.

5. A system for detecting and jointly interpreting abnormal states of wind turbine generator sets, characterized in that, A method for detecting and jointly interpreting abnormal states of a wind turbine generator set according to any one of claims 1-4 includes a data acquisition module, a preprocessing module, a feature selection module, an abnormal state detection module, a global interpretation module, a local interpretation module, and a visualization module. The data acquisition module is used to obtain multi-dimensional operational data from the SCADA system of the wind turbine generator set; The preprocessing module is used to preprocess the acquired multi-dimensional runtime data; The feature selection module is used to cluster the preprocessed wind turbine SCADA data using the k-means clustering algorithm, and select the central features of the calculated central clusters to construct a new feature dataset. The abnormal state detection module is used to construct a new feature dataset based on the extracted central feature data, and to train the wind turbine abnormal state detection model using a classification decision enhancement algorithm to detect and identify abnormal states. The global interpretation module is used to establish a simplified model of the output of the wind turbine abnormal state detection model, calculate the global contribution of the input features to the abnormal state detection results using the SHAP algorithm, and provide a global interpretation. The local interpretation module is used to establish a local model of the wind turbine abnormal state detection model instance. It uses the LIME algorithm to generate a local interpretation for a specific abnormal instance by fitting a local linear model and analyzes the contribution of key features in the specific abnormal instance. The visualization module is used to visualize the detection and interpretation results and trigger alarm mechanisms.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for detecting and jointly interpreting abnormal states of a wind turbine generator set as described in any one of claims 1-4.

7. A processing terminal, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for detecting and jointly interpreting abnormal states of a wind turbine generator set as described in any one of claims 1-4.

Images