A method, system, medium and device for detecting icing of fan blades

Abstract

The present invention belongs to the technical field of wind power generation, and discloses a method, system, medium and equipment for detecting icing of wind turbine blades. The method for detecting icing of wind turbine blades includes combining a tri-training three-body training method semi-supervised learning algorithm on the basis of an XGBoost algorithm, and reasonably using data with inaccurate labels; constructing new features based on the original features according to expert knowledge, using the Pearson correlation coefficient as a model for feature screening, and selecting Focal Loss as a loss function for iterative updates of the algorithm. The present invention improves the accuracy and generalization performance of the model by eliminating sample points where the wind turbine is operating at limited power due to non-icing factors. The present invention incorporates new features, uses the Pearson correlation coefficient to reduce the dimensionality of the features, reduces the complexity of the model, and improves the efficiency of blade icing recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and in particular relates to a method, system, medium and equipment for detecting icing of wind turbine blades. Background Art

[0002] At present, for altitudes below 1,000 meters, the wind speed increases by 0.1m / s for every 100m increase in altitude. Areas with abundant wind resources are often in cold areas and high altitude areas. Since cold air is denser than warm air, the air density in high-altitude cold areas is relatively high, and the potential wind resources in cold areas are 10% higher than in other areas. Therefore, the site selection of wind farms is usually also in these areas. In winter or when the temperature is low, these areas are prone to icing of wind turbine blades. The icing of wind turbine blades is a slow process. The occurrence of the problem of wind turbine icing is related to many factors in the external environment, such as ambient temperature, wind speed and air humidity. When the wind turbine blades are in an icing state, the lift of the blades will be reduced and the resistance will be increased. At the same time, the blades will be heavier and prone to blade breakage, which poses a great safety hazard. In the most common case of blade icing, the actual output of wind turbine power generation will be affected. The annual power generation efficiency in areas with severe climate will be reduced by 20% to 50%. Early diagnosis of blade icing is conducive to reducing power generation losses and improving the safety of unit operation. The existing technology is mainly divided into mechanism model-based methods and data-driven methods. The mechanism model-based method relies on the accumulation of expert experience. Due to the large number of moving parts in the wind turbine, the high complexity of the wind turbine system, and the rapid update and iteration of complete machine products in the wind power industry, it is very difficult for experts to establish a mechanism model. The data-driven model does not rely on physical mechanisms, but it requires a large amount of data, and the distribution of data must be reasonable, and the quality requirements for data and data labels are high. The wind turbine SCADA system solves the problem of data volume, but the number of icing samples is small, resulting in the problem of unbalanced sample categories. In addition, the blade icing sample labels need to be manually marked. Due to the relatively scattered geographical locations and poor accessibility of wind turbine units, the blade icing samples have the problem of inaccurate labels.

[0003] Through the above analysis, the problems and defects of the prior art are as follows:

[0004] (1) The existing method based on mechanism model relies on the accumulation of expert experience. Since the fan has many moving parts and the fan system is very complex, it is very difficult for experts to establish a mechanism model.

[0005] (2) The existing technology has a small number of icing samples in the wind turbine SCADA system, resulting in the problem of unbalanced sample categories. In addition, the blade icing sample labels need to be manually labeled. Due to the relatively scattered geographical locations and poor accessibility of wind turbine units, the blade icing samples have the problem of inaccurate labels. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention provides a method, system, medium and equipment for detecting icing of fan blades.

[0007] The present invention is implemented as follows: a method for detecting icing of fan blades, the method comprising:

[0008] Based on the XGBoost algorithm, the tri-training semi-supervised learning algorithm is combined to reasonably use the data with inaccurate labels; new features are constructed based on the original features according to expert knowledge, the Pearson correlation coefficient is used as the model for feature screening, and Focal Loss is selected as the loss function for iterative updating of the algorithm.

[0009] Furthermore, the specific process of the fan blade icing detection method is as follows:

[0010] Step 1: Training set construction and data preprocessing. Obtain data from the SCADA system and divide the raw data into labeled samples and unlabeled samples. Labeled samples include icing samples marked as 1 and normal samples marked as 0. Based on the tri-XGBoost algorithm, a corresponding mechanism is established; in the data preprocessing stage, sample points that cause wind turbine power limitation due to non-icing are removed; this step removes noise data, which is conducive to classifier adaptive learning;

[0011] Step 2: Feature engineering and feature screening. Combine the existing features of the original data and select three new features to input into the model. Use the Pearson correlation coefficient to reduce the dimension of the features, screen out the features with too high correlation and remove them. This step improves the learning efficiency of the classifier, prevents the classifier from learning redundant features, and combines domain expertise to mine new features to improve the accuracy of the model.

[0012] Step 3: Category imbalance processing and model training: select cost-sensitive learning method, introduce Focal Loss loss function, and replace the original loss function of XGBoost algorithm; input labeled data and unlabeled data into tri-XGBoost model for collaborative training; this step enhances the model's ability to correctly classify minority class samples, effectively improving the detection accuracy of wind turbine blade icing while maximizing the retention of original data;

[0013] Step 4: Parameter optimization and model evaluation. Select the grid search algorithm to determine the optimal hyperparameters; use the processed test set to test and verify the model classification effect. This step traverses the given parameter combination and finds the parameter with the highest accuracy on the test set from all parameters, optimizes the model parameters, and further improves the detection accuracy of wind turbine blade icing.

[0014] Furthermore, in step 1, the specific process of training set construction and data preprocessing is as follows:

[0015] Based on the tri-XGBoost algorithm, a mechanism is established. In the original SCADA data, the samples in the period before the freezing data are taken as unlabeled samples, and the remaining samples are taken as labeled samples. In the data preprocessing stage, the sample points where the fan is running at limited power due to non-freezing are eliminated.

[0016] Furthermore, in step 2, the specific process of selecting three new features and inputting them into the model is as follows:

[0017] The ratio of short-filtered wind speed to active power, the ratio of short-filtered wind speed to generator speed, and the ratio of short-filtered wind speed to active power and generator speed are input into the model as three new fused features.

[0018] Furthermore, in step 3, the specific process of class imbalance processing and model training is as follows:

[0019] Select the cost-sensitive learning method, introduce the Focal Loss loss function, and replace the original loss function of the XGBoost algorithm; input the labeled data and unlabeled data into the tri-XGBoost model for collaborative training, perform Bootstrap sampling on the original labeled data to obtain three labeled training sets, use the three labeled training sets to train three base classifiers XGB1, XGB2, and XGB3 respectively, input the unlabeled samples into the trained base classifiers, and the single classifier XGB i The new labeled samples obtained are provided by the collaboration of the other two classifiers. If the other two classifiers make the same prediction for the same unlabeled sample, the sample is considered to have a higher confidence level, and the predicted value is used as the pseudo label of the sample and added to XGB. i The classifier is continuously updated iteratively, and finally the classification model is obtained according to the voting method.

[0020] Furthermore, in step 4, the specific process of parameter optimization and model evaluation is as follows:

[0021] Select the grid search algorithm to determine the optimal hyperparameters, perform the same data preprocessing and feature engineering on the test set to make the feature dimensions the same, and use the processed test set for testing to verify the model classification effect.

[0022] Another object of the present invention is to provide a fan blade icing detection system for implementing the fan blade icing detection method, the fan blade icing detection system comprising:

[0023] The training set construction and data preprocessing module obtains data from the SCADA system and divides the raw data into labeled samples and unlabeled samples, where the labeled samples include icing samples marked as 1 and normal samples marked as 0. Based on the tri-XGBoost algorithm, a corresponding mechanism is established; in the data preprocessing stage, sample points that cause wind turbine power limitation due to non-icing are eliminated;

[0024] The feature engineering and feature screening module combines the existing features of the original data and selects three new features to input into the model; the Pearson correlation coefficient is used to reduce the dimension of the features, and the features with too high correlation are screened out and removed;

[0025] In the category imbalance processing and model training module, the cost-sensitive learning method is selected, the Focal Loss loss function is introduced, and the original loss function of the XGBoost algorithm is replaced; the labeled data and the unlabeled data are input into the tri-XGBoost model for collaborative training;

[0026] The parameter optimization and model evaluation module selects the grid search algorithm to determine the optimal hyperparameters; the processed test set is used for testing to verify the model classification effect.

[0027] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0028] Step 1: Training set construction and data preprocessing. Obtain data from the SCADA system and divide the raw data into labeled samples and unlabeled samples. Labeled samples include icing samples marked as 1 and normal samples marked as 0. Based on the tri-XGBoost algorithm, a corresponding mechanism is established; in the data preprocessing stage, sample points that cause wind turbine power limitation due to non-icing are removed;

[0029] Step 2: Feature engineering and feature screening: combine the existing features of the original data and select three new features to input into the model; use the Pearson correlation coefficient to reduce the dimension of the features, screen out the features with too high correlation and remove them;

[0030] Step 3: Category imbalance processing and model training: select cost-sensitive learning method, introduce Focal Loss loss function, and replace the original loss function of XGBoost algorithm; input labeled data and unlabeled data into tri-XGBoost model for collaborative training;

[0031] Step 4: Parameter optimization and model evaluation: select the grid search algorithm to determine the optimal hyperparameters; use the processed test set to test and verify the model classification effect.

[0032] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:

[0033] Step 1: Training set construction and data preprocessing. Obtain data from the SCADA system and divide the raw data into labeled samples and unlabeled samples. Labeled samples include icing samples marked as 1 and normal samples marked as 0. Based on the tri-XGBoost algorithm, a corresponding mechanism is established; in the data preprocessing stage, sample points that cause wind turbine power limitation due to non-icing are removed;

[0034] Step 2: Feature engineering and feature screening: combine the existing features of the original data and select three new features to input into the model; use the Pearson correlation coefficient to reduce the dimension of the features, screen out the features with too high correlation and remove them;

[0035] Step 3: Category imbalance processing and model training: select cost-sensitive learning method, introduce Focal Loss loss function, and replace the original loss function of XGBoost algorithm; input labeled data and unlabeled data into tri-XGBoost model for collaborative training;

[0036] Step 4: Parameter optimization and model evaluation: select the grid search algorithm to determine the optimal hyperparameters; use the processed test set to test and verify the model classification effect.

[0037] Another object of the present invention is to provide an information data processing terminal, which is used to implement the wind turbine blade icing detection method.

[0038] In combination with the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solutions to be protected by the present invention from the following aspects:

[0039] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving the problems, the technical solutions to be protected by the present invention and the results and data during the research and development process are closely combined to analyze in detail and deeply how the technical solutions of the present invention solve the technical problems, and some creative technical effects brought about after solving the problems. The specific description is as follows:

[0040] The present invention improves the accuracy and generalization performance of the model by eliminating sample points where the fan is operating at limited power due to non-icing factors. The present invention incorporates new features and uses the Pearson correlation coefficient to reduce the dimension of the features, thereby reducing the complexity of the model and improving the efficiency of blade icing recognition. The present invention introduces the Focal Loss function to retain the original data as much as possible while solving the problem of class imbalance and improving the classification performance of the model; the three-body training method is introduced to reasonably use data with inaccurate labels, so that the model still has good classification performance when the data quality is low.

[0041] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by the present invention are described in detail as follows:

[0042] The present invention uses semi-supervised learning combined with the powerful generalization ability of ensemble learning methods to achieve the best classification performance. It solves the problems of missing labels, inaccurate labels and low data quality in engineering practice in a low-cost way, enhances the model's ability to diagnose early icing phenomena, ensures the safety of wind turbine operation, and greatly reduces the loss of wind turbine power generation.

[0043] Third, as auxiliary evidence of the inventiveness of the present invention, it is also reflected in the following important aspects:

[0044] The technical solution of the present invention solves the technical problems that people have been eager to solve but have never been able to solve successfully:

[0045] The present invention solves the problem of inaccurate labels of wind turbine SCADA data due to network communication disconnection and difficulty in manual labeling, introduces unlabeled samples, and effectively improves the accuracy of blade icing detection. The algorithm focuses on minority samples. Under the condition of data imbalance, a semi-supervised algorithm is used to learn a small number of labeled samples for pseudo-labeling of unlabeled samples. Combining unlabeled samples with pseudo-labels and adding them to model training effectively reduces the difficulty of wind turbine operation and maintenance, reduces the loss of wind turbine power generation, and avoids the risk of wind turbine blade breakage during cold periods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of a method for detecting icing of fan blades provided by an embodiment of the present invention;

[0047] Figure 2 It is a schematic diagram of the structure of a fan blade icing detection system provided by an embodiment of the present invention;

[0048] Figure 3 is a schematic diagram of a blade icing detection modeling process provided by an embodiment of the present invention;

[0049] Figure 4is a waveform diagram of a new fusion feature provided by an embodiment of the present invention;

[0050] In the figure: 1. Training set construction and data preprocessing module; 2. Feature engineering and feature screening module; 3. Category imbalance processing and model training module; 4. Parameter optimization and model evaluation module. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] 1. Explanatory Examples In order to enable those skilled in the art to fully understand how to implement the present invention, this section provides an illustrative example that describes the technical solution.

[0053] like Figure 1 As shown, the fan blade icing detection method provided by the embodiment of the present invention includes:

[0054] S101: Training set construction and data preprocessing, obtain data from the SCADA system, divide the original data into labeled samples and unlabeled samples, where the labeled samples include icing samples marked as 1 and normal samples marked as 0. Based on the tri-XGBoost algorithm, establish a corresponding mechanism; in the data preprocessing stage, remove the sample points that cause the fan to operate at limited power due to non-icing.

[0055] S102: Feature engineering and feature screening: combine the existing features of the original data, select three new features and input them into the model; use the Pearson correlation coefficient to reduce the dimension of the features, screen out features with too high correlation and remove them.

[0056] S103: Category imbalance processing and model training, select cost-sensitive learning method, introduce Focal Loss loss function, replace the original loss function of XGBoost algorithm; input labeled data and unlabeled data into tri-XGBoost model for collaborative training.

[0057] S104: Parameter optimization and model evaluation, select the grid search algorithm to determine the optimal hyperparameters; use the processed test set to test and verify the model classification effect.

[0058] In S101 provided in the embodiment of the present invention, the specific process of training set construction and data preprocessing is as follows:

[0059] Based on the tri-XGBoost algorithm, a mechanism is established: in the original SCADA data, the samples in the period before the freezing data are used as unlabeled samples, and the remaining samples are used as labeled samples. In the data preprocessing stage, the sample points where the fan is limited in power due to non-freezing are removed.

[0060] In S102 provided in the embodiment of the present invention, the specific process of selecting three new features and inputting them into the model is as follows:

[0061] The ratio of short-filtered wind speed to active power, the ratio of short-filtered wind speed to generator speed, and the ratio of short-filtered wind speed to active power and generator speed are input into the model as three new fused features.

[0062] In S103 provided in the embodiment of the present invention, the specific process of class imbalance processing and model training is as follows:

[0063] The cost-sensitive learning method is selected, and the Focal Loss loss function is introduced to replace the original loss function of the XGBoost algorithm. The labeled data and unlabeled data are input into the tri-XGBoost model for collaborative training. After the base classifier is initialized with the labeled data, the pseudo labels of the unlabeled data are generated according to the trained base classifier, and the classifier is continuously updated iteratively, and finally the classification model is obtained according to the voting method.

[0064] In S104 provided in the embodiment of the present invention, the specific process of parameter optimization and model evaluation is as follows:

[0065] Select the grid search algorithm to determine the optimal hyperparameters, perform the same data preprocessing and feature engineering on the test set to ensure the same feature dimensions, and use the processed test set for testing to verify the model classification effect.

[0066] like Figure 2 As shown, the fan blade icing detection system provided by the embodiment of the present invention includes:

[0067] Training set construction and data preprocessing module 1, based on the tri-XGBoost algorithm, establishes a corresponding mechanism; in the data preprocessing stage, removes sample points where the fan is running at limited power due to non-icing.

[0068] Feature engineering and feature screening module 2 combines the existing features of the original data and selects three new features to input into the model; the Pearson correlation coefficient is used to reduce the dimension of the features, and the features with too high correlation are screened out and removed.

[0069] In the category imbalance processing and model training module 3, a cost-sensitive learning method is selected, and the Focal Loss loss function is introduced to replace the original loss function of the XGBoost algorithm; labeled data and unlabeled data are input into the tri-XGBoost model for collaborative training.

[0070] Parameter optimization and model evaluation module 4 selects the grid search algorithm to determine the optimal hyperparameters; uses the processed test set to test and verify the model classification effect.

[0071] The technical solution of the present invention is described in detail below in conjunction with specific embodiments.

[0072] Aiming at the problems that wind turbine blades have a short freezing time, data set categories are unbalanced, and manually annotated data have inaccurate labels, an embodiment of the present invention proposes a tri-XGBoost method, which combines the tri-training (three-body training method) semi-supervised learning algorithm on the basis of the XGBoost algorithm, rationally utilizes the data with inaccurate labels, constructs new features based on the original features according to expert knowledge, uses the Pearson correlation coefficient as a model for feature screening, and selects FocalLoss as the loss function for iterative updating of the algorithm, which provides a solution to the problems of inaccurate labels and unbalanced categories in blade icing detection.

[0073] The fan blade icing detection process provided by the embodiment of the present invention is as follows:

[0074] Training set construction and data preprocessing. The freezing of wind turbine blades is a slowly evolving process. Usually, the annotation information of blade freezing is obtained by professionals through observation. However, due to the untimely observation, the early frozen samples may be mislabeled as normal samples. Introducing these samples into model training will affect the accuracy and generalization performance of the model. The present invention is based on the tri-XGBoost algorithm and establishes the following mechanism: in the original SCADA data, the samples in the period before the freezing data are used as unlabeled samples, and the remaining samples are used as labeled samples. Blade freezing can cause the wind turbine to operate with limited power. According to expert experience, the reasons for the limited power operation of the wind turbine are: power grid scheduling power limitation. When the wind speed is too high, the power sent by the wind farm to the power grid is difficult to store; problems with large parts of the wind turbine require artificial power limitation to avoid accidents; the wind farm is close to residential areas, and power limitation is required to avoid excessive noise during the night. Therefore, in order to improve the accuracy and generalization performance of the model, these sample points that cause the limited power operation of the wind turbine due to non-freezing are eliminated in the data preprocessing stage.

[0075] Feature engineering and feature screening. The embodiment of the present invention combines the existing features of the original data, selects the ratio of short-filtered wind speed to active power, the ratio of short-filtered wind speed to generator speed, and the ratio of short-filtered wind speed to active power and generator speed as three new features to be fused and input into the model. The Pearson correlation coefficient is used to reduce the dimension of the features, and the features with too high correlation are screened out and removed.

[0076] Class imbalance processing and model training. Methods for processing class imbalance can be divided into three categories, namely undersampling, oversampling and cost-sensitive learning. In order to retain the original data as much as possible, the present invention selects a cost-sensitive learning method, introduces the Focal Loss loss function, and replaces the original loss function of the XGBoost algorithm. The labeled data and the unlabeled data are input into the tri-XGBoost model for collaborative training. After the base classifier is initialized with the labeled data, the pseudo-labels of the unlabeled data are generated according to the trained base classifier, and the classifier is continuously iterated and updated, and finally the classification model is obtained according to the voting method.

[0077] Parameter optimization and model evaluation. Parameter optimization methods include grid search, random search, and Bayesian optimization. The embodiment of the present invention selects a grid search algorithm to determine the optimal hyperparameters. The test set is subjected to the same data preprocessing and feature engineering to ensure that the feature dimensions are the same, and the processed test set is used for testing to verify the model classification effect.

[0078] The specific process of the tri-training method provided by the embodiment of the present invention is as follows:

[0079] Tri-training first trains three classifiers with labeled data, then generates pseudo labels for unlabeled data based on the trained classifiers. The pseudo-label data and the original labeled data are used to iteratively update the classifiers until they no longer change. It should be noted that in each iteration, the pseudo-label samples from the previous round are reprocessed as unlabeled samples. Finally, the three classifiers are fused into a strong classifier based on a simple voting method.

[0080] If the other two classifiers misclassify, a misclassified sample with noise will be generated, and the misclassified sample will be added to the classifier XGB i After being added to the training set, the classification performance of the classifier will be affected. Therefore, the new i The samples in the training set need to meet the following conditions:

[0081]

[0082] in, represents the upper limit of the misclassification rate of the other two classifiers, L t and Lt-1 Respectively represent the addition to the classifier XGB in the t round and the t-1 round i If the newly added sample to the training set does not meet the above conditions, Perform random downsampling and delete samples. The following formula should be satisfied so that the training set After downsampling, the size is still larger than

[0083]

[0084] 2. Application Examples: In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the technical solution on specific products or related technologies.

[0085] The SCADA data of the wind turbine during operation is obtained. According to the actual data and the manually observed icing conditions, the data is divided into labeled samples and unlabeled samples. The labeled samples include icing samples and normal samples, which are marked as 1 and 0 respectively. A wind speed-power scatter plot of the labeled samples is constructed, and the labeled samples are preliminarily screened to remove outliers and missing values. Normal samples with blade angles greater than 1.5 and limited power are screened and removed to avoid the existence of "dirty data" affecting subsequent wind turbine blade icing detection. The labeled and unlabeled samples after screening of the first two wind turbines are selected as the original training set of tri-XGBoost, and the data of the other wind turbine is not labeled as a test set.

[0086] The short-filtered wind speed, active power, and generator speed in the original features are calculated according to the following formulas: To construct new features, both the training set and the test set must go through the above steps to avoid different data dimensions, which may cause the classifier to be unable to complete the blade icing state recognition. Pearson correlation analysis is performed on the original data features and the newly constructed features, and features with a Pearson correlation coefficient exceeding 0.9 are eliminated to avoid feature redundancy leading to a decrease in model classification accuracy and generalization performance.

[0087] Input the original training set into the tri-XGBoost model, perform Bootstrap sampling on the labeled samples, train them separately, initialize three base classifiers XGB1, XGB2, and XGB3, and customize the loss function of the base classifier to FocalLoss. The definition of Focal Loss is as follows:

[0088]

[0089] in, To predict the probability of the label being 1, γ is a hyperparameter that controls the weight of the loss function for easy-to-classify samples and difficult-to-classify samples. Focal Loss increases the weight of difficult-to-classify samples in the loss function, which helps to improve the accuracy of difficult-to-classify sample classification.

[0090] Upper limit of misclassification rate Set to 0.5, newly added to the classifier XGB i The number of samples in the training set Set to 0, the number of iterations is initialized to 0. Newly added to the classifier XGB i The samples in the training set need to meet the following conditions:

[0091]

[0092] If the sample newly added to the training set does not meet the above conditions, Perform random downsampling and delete samples. The following formula should be satisfied so that the training set After downsampling, the size is still larger than

[0093]

[0094] After iterating the training sets of the three base classifiers, a strong classifier composed of the three base classifiers is finally obtained through the voting method.

[0095] The parameters in the base classifier, such as learning_rate, n_estimators, eta, max_depth, min_child_weight, and subsample, are optimized by grid search, and the candidate list of parameters is set. The optimal parameters of the classifier are calculated by exhaustive method. After obtaining the optimal parameters, the model is finally evaluated, and the accuracy, precision, recall, the harmonic mean of recall and precision (F1), and the Matthews correlation coefficient (MCC) are selected as evaluation criteria.

[0096] 3. Evidence of the effects of the embodiments. The embodiments of the present invention have achieved some positive effects during the development or use process, and indeed have great advantages over the prior art. The following content is described in conjunction with the data, charts, etc. of the test process.

[0097] The experimental data was collected from three wind turbines in a wind farm in a certain province. The SCADA data of the three wind turbines from October 1 to April 29 of a certain year were selected for data analysis and ice detection modeling prediction. The SCADA system collected data every 1 minute. Due to the situation of wind turbine shutdown and SCADA system communication disconnection, there are a large number of invalid sample points in the original data. In order to avoid affecting the classification performance of the model, the invalid sample points should be removed. The wind turbine has limited power operation caused by non-blade icing, so the sample points that meet the above conditions need to be removed. After data preprocessing, the three wind turbines have 177619, 173451, and 170014 sample points respectively. The data of the first two wind turbines are selected as the training set, and the data of the other wind turbine is selected as the test set. The training set is divided into labeled data and unlabeled data. The number of unlabeled sample points is 15840, accounting for about 4.51% of all training sets. There are 67,908 blade icing sample points in the labeled data, accounting for about 19.34% of all training sets.

[0098] The original features of the data mainly include 18 operating parameters such as short-filter wind speed, active power, blade angle, cabin temperature, generator speed, cabin position deviation, cabin position, pitch speed, tower driving direction acceleration, tower non-driving direction acceleration, cabin temperature, wind turbine outlet active power, yaw speed, etc. Based on expert knowledge and data analysis, three new features are integrated Where windSpeed ​​is the short filter wind speed, power is the active power, and generatorSpeed ​​is the generator speed. Figure 4 As shown in the figure, the three new features of the fusion have obvious increases in their values ​​within the time range when the label is 1 (i.e., the wind turbine is frozen).

[0099] Considering the problem of class imbalance in blade icing data, the following five evaluation indicators are selected to evaluate the model performance: accuracy, precision, recall, harmonic mean of recall and precision (F1), and Matthews correlation coefficient (MCC). TP is the number of correct predictions for positive samples, TN is the number of correct predictions for negative samples, FP is the number of incorrect predictions for positive samples, and FN is the number of incorrect predictions for negative samples.

[0100] F1 is a statistical data used to evaluate the generalization performance of the model, which is a weighted average of precision and recall. In the case of an unbalanced data set, the performance of the model can be well evaluated. The higher the F1 value, the better the performance of the classifier. The closer the value of MCC is to 1, the stronger the robustness of the model. Under the same experimental conditions, the algorithm proposed in the present invention is compared with common classification methods such as support vector machine (SVM), logistic regression (LR), nearest neighbor (KNN), random forest (RF) and extreme gradient boosting tree (XGBoost), and the test results are shown in Table 1.

[0101] Table 1 Comparison of results of different detection methods

[0102]

[0103] The experimental results show that the tri-XGBoost algorithm proposed in the present invention outperforms other classification algorithms in all indicators. Although the accuracy of LR, RF, and XGboost is also very high, the reference value of the accuracy is not great for extremely unbalanced data sets. For binary classification problems, MCC is the most intuitive and simple evaluation criterion based on the confusion matrix for the case of unbalanced data sets. Accuracy and F1 are easily misleading in this case. The MCC value of the algorithm proposed in the present invention is significantly higher than other models, and it shows a good classification effect for wind turbine blade icing detection.

[0104] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0105] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for detecting icing of fan blades, characterized in that: The fan blade icing detection method comprises: Based on the XGBoost algorithm, the tri-training semi-supervised learning algorithm is combined to make reasonable use of data with inaccurate labels; new features are constructed based on the original features according to expert knowledge, the Pearson correlation coefficient is used as a model for feature screening, and Focal Loss is selected as the loss function for iterative updates of the algorithm; The specific process of the fan blade icing detection method is as follows: Step 1: Training set construction and data preprocessing, based on the XGBoost algorithm, establish the corresponding mechanism; In the data preprocessing stage, sample points where the fan is operating at limited power due to non-icing are eliminated; Step 2: Feature engineering and feature screening: combine the existing features of the original data and select three new features to input into the model; use the Pearson correlation coefficient to reduce the dimension of the features, screen out the features with too high correlation and remove them; Step 3: Category imbalance processing and model training: select the cost-sensitive learning method, introduce the FocalLoss loss function, and replace the original loss function of the XGBoost algorithm; input the labeled data and unlabeled data into the tri-XGBoost model for collaborative training; Step 4: Parameter optimization and model evaluation: select the grid search algorithm to determine the optimal hyperparameters; use the processed test set to test and verify the model classification effect; In step 1, the specific process of training set construction and data preprocessing is as follows: Based on the tri-XGBoost algorithm, a mechanism is established; in the original SCADA data, the samples in the period before the freezing data are taken as unlabeled samples, and the remaining samples are taken as labeled samples; in the data preprocessing stage, the sample points that cause the fan to operate at limited power due to non-freezing are removed; In step 2, the specific process of selecting three new features and inputting them into the model is as follows: The ratio of short-filtered wind speed to active power, the ratio of short-filtered wind speed to generator speed, and the ratio of short-filtered wind speed to active power and generator speed are input into the model as three new fused features.

2. The method for detecting icing of fan blades according to claim 1, characterized in that: In step 3, the specific process of class imbalance processing and model training is as follows: A cost-sensitive learning method was selected, and the Focal Loss loss function was introduced to replace the original loss function of the XGBoost algorithm. Labeled data and unlabeled data were input into the tri-XGBoost model for collaborative training. After the base classifier was initialized with labeled data, pseudo labels for unlabeled data were generated according to the trained base classifier. The classifier was continuously iterated and updated, and finally the classification model was obtained according to the voting method.

3. The method for detecting icing of fan blades according to claim 1, characterized in that: In step 4, the specific process of parameter optimization and model evaluation is as follows: Select the grid search algorithm to determine the optimal hyperparameters, perform the same data preprocessing and feature engineering on the test set to make the feature dimensions the same, and use the processed test set for testing to verify the model classification effect.

4. A fan blade icing detection system implementing the fan blade icing detection method according to any one of claims 1 to 3, characterized in that: The fan blade icing detection system comprises: The training set construction and data preprocessing module establishes corresponding mechanisms based on the tri-XGBoost algorithm; In the data preprocessing stage, sample points where the fan is operating at limited power due to non-icing are eliminated; The feature engineering and feature screening module combines the existing features of the original data and selects three new features to input into the model; the Pearson correlation coefficient is used to reduce the dimension of the features, and the features with too high correlation are screened out and removed; In the category imbalance processing and model training module, the cost-sensitive learning method is selected, the Focal Loss loss function is introduced, and the original loss function of the XGBoost algorithm is replaced; the labeled data and the unlabeled data are input into the tri-XGBoost model for collaborative training; The parameter optimization and model evaluation module selects the grid search algorithm to determine the optimal hyperparameters; the processed test set is used for testing to verify the model classification effect.

5. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the wind turbine blade icing detection method according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes the steps of the wind turbine blade icing detection method according to any one of claims 1 to 3.

7. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the wind turbine blade icing detection method as described in any one of claims 1 to 3.

Images