Fault diagnosis method and system for blade icing and blade mass imbalance of wind turbines

Abstract

Disclosed in the present invention are a fault diagnosis method and system for blade icing and blade mass imbalance of wind turbines. The method comprises the steps of: acquiring operation data and blade icing information of a plurality of wind turbines; labeling the operation data with an icing state label on the basis of the blade icing information, so as to obtain fault data; extracting fault features in the fault data, and ranking the fault features according to the degree of importance, so as to generate an optimal feature set; on the basis of a criterion of minimizing a squared error, selecting optimal features in the optimal feature set to generate an optimal decision tree; and performing classification on the basis of the optimal decision tree, so as to obtain a diagnosis test result including fault information. The present invention has the advantages of a high level of diagnostic accuracy, etc.

Description

Fan blade icing and blade mass imbalance fault diagnosis method and system

[0001] Cross-reference to related applications

[0002] [According to Rule 91, correct 21.05.2025] This application is based on the Chinese patent application with the application date of "2024.09.10", the application number of "2024112647834", and the invention name of "a fan blade icing and blade mass imbalance fault diagnosis method and system", and claims priority thereto, the whole text of the Chinese patent application is hereby incorporated into the present application as part of the present application.

TECHNICAL FIELD

[0003] The present application mainly relates to the field of wind power technology, in particular to a fan blade icing and blade mass imbalance fault diagnosis method and system.

BACKGROUND

[0004] With the rapid development of wind power, some problems have inevitably arisen, such as wind turbine icing problems, wind turbine blade mass imbalance problems, etc. When the fan runs in a fault state, the power quality is low, and it cannot be normally connected to the grid, thus causing serious resource waste, and when the fan runs in a fault state for a period of time, the entire system will start to malfunction to varying degrees, causing the fan to be unable to operate under normal working conditions.

[0005] The units of land-based wind power are generally installed in high-altitude cold regions and places with rich wind resources such as ridges. These places are usually high in altitude and low in temperature, which can easily cause blade icing, thus leading to a decrease in fan efficiency, a change in blade shape, and even the induction of blade fracture. In addition, most wind turbines mainly operate outdoors in relatively harsh environments, and wind and sand abrasion or icing may cause the impeller mass to be unbalanced. Continued operation under the condition of impeller mass imbalance will exacerbate the fatigue vibration of the blade and other components, which can cause serious damage to the wind turbine. If the state change of the fan is monitored and discovered in time, the fault state and fault position are identified, the fan can be maintained and repaired in time, thereby greatly reducing the resource loss and waste of wind power generation. Therefore, it is of great significance to study the fault diagnosis of wind turbine icing problems and wind turbine blade mass imbalance problems.

[0006] Solving the problem of wind turbine blade imbalance and icing through modeling is actually to track and analyze the internal physical change process during the blade failure process, so as to determine the fault state through the corresponding data. This method is based on the research of mechanism, and usually needs to place additional sensors for the wind turbine blade. The actual application is not convenient, the operation and maintenance cost of the wind farm is increased, etc. which makes it unable to be widely applied in practice.

[0007] The data-driven modeling method does not need to install additional sensors, data acquisition and monitoring are performed through a supervisory control and data acquisition (SCADA) system, and a feature quantity is mined from the SCADA system, so that the fault state is diagnosed. This method often needs a large amount of data as support, and has the disadvantages of lack of explainability and operation guidance. SUMMARY

[0008] In view of the technical problems of the prior art, the present application provides a fan blade icing and blade mass imbalance fault diagnosis method and system with high diagnosis accuracy.

[0009] To solve the above technical problems, the technical scheme provided by the present application is as follows:

[0010] A fan blade icing and blade mass imbalance fault diagnosis method, comprising the steps of:

[0011] obtaining running data and blade icing information of multiple fans;

[0012] annotating the running data with icing state labels according to the blade icing information to obtain fault data;

[0013] extracting fault features in the fault data, and sorting the fault features according to importance to generate an optimal feature set;

[0014] selecting optimal features in the optimal feature set according to the criterion of minimizing square error to generate an optimal decision tree;

[0015] classifying based on the optimal decision tree to obtain a diagnosis detection result containing fault information.

[0016] Preferably, the running data of the fan includes wind speed, power, cabin temperature, and ambient temperature and cabin front, back, left and right accelerations.

[0017] Preferably, the fault features include power residual features, wind speed to power ratio, rotational speed residual features, temperature difference features, and vibration acceleration vector synthesis features.

[0018] Preferably, the optimal feature set is generated by sorting the fault features according to importance through logistic regression and support vector machine, and specifically includes:

[0019] dividing the fault data into a training set D 1j and a validation set D 2j through resampling;

[0020] training a random forest fault diagnosis model M ij (θ) using all features j 1j of the training set D 1j , and verifying M 2j on the validation set D .ij (θ) evaluation;

[0021] By training the model M ij (θ), calculate its performance score, and select the feature set S 1j from high to low in importance, and then select the feature set S ij , where i = 1, 2,..., m;

[0022] For the feature set, use the feature set S 1j of the training set D ij to train the model M ij (θ), and perform model evaluation on the validation set D 2j to obtain the score of the model M ij (θ) on the feature set S ij ;

[0023] Average the model scores of the same feature set S ij with h times resampling, and obtain the feature set with the highest score;

[0024] Use the feature set S max on the original training set D to execute the algorithm and obtain the optimal feature set S m .

[0025] Preferably, the random forest classification fault diagnosis model integrates multiple decision tree basic units through the idea of ensemble learning, including:

[0026] The optimal feature is selected by the criterion of minimizing the square error to generate a decision tree; wherein the decision tree is defined as:

[0027] In the formula, x is the input, f(x) is the output, M represents the division of the input space into M units: R1, R2,..., R m , and each unit R m has a fixed output value mc;

[0028] The left and right branch regions of the jth feature and the feature split point s of the decision tree model are defined as: R1(j,s) = {x|x (j) ≤s}, R2(j,s) = {x|x (j) >s}

[0029] The optimal split feature j and split point s are obtained by solving the following formula:

[0030] In the formula, y i is the corresponding label value of x i , c1 and c2 are: C1 = ave(y i |x i∈R1(j,s)),C2=ave(y i |x i ∈R2(j,s))

[0031] Starting from the root node, find the optimal segmentation features and segmentation points to divide the input space into two regions. Then, repeat this segmentation process for each region until m partitioning units R1, R2, ... R are obtained. m The optimal output value of each unit is obtained based on the criterion of minimizing the squared error; each unit R m The optimal value is the sum of all samples x in this unit. i Corresponding y i The mean is defined as: C m =ave(y i |x i ∈R m ).

[0032] Preferably, the rules for generating the decision tree include:

[0033] Select k subsets of training samples from the total training sample set and build k decision trees;

[0034] At each node of the decision tree, m indicators are randomly selected from n indicators. The optimal feature is selected from the m candidate indicators according to the principle of minimizing node impurity. The node is then classified and grown. The decision tree is allowed to grow fully until the impurity of each leaf node is minimized. At the same time, the decision tree is not pruned.

[0035] Repeat the above steps to traverse the pre-built k decision trees, and form a random forest from the k decision trees;

[0036] The classification result of a new unknown sample is predicted based on the k well-grown decision trees, and the decision is made by the majority vote of the k decision trees.

[0037] Preferably, the decision tree is pruned by starting from the bottom of the complete decision tree T0 and continuously pruning to the root node, forming a subtree sequence {T0, T1, ..., T}. n Then, cross-validation is used to test the subtree on the validation set, and the optimal subtree is selected.

[0038] The loss function of the subtree in the above pruning operation is defined as: C α (T)=C(T)+α|T|

[0039] In the formula, T represents any subtree, C(T) is the prediction error of the training data, |T| is the number of leaf nodes in the subtree, and is a parameter greater than 0 used to balance the fit of the training data with the complexity of the model. α (T) represents the overall loss of subtree T when the parameter is α;

[0040] From T0, calculate the alpha of each node from bottom to top, cut the subtree with the smallest alpha in T0 to get T1, and then take T1 as the whole tree, calculate the alpha of each node from bottom to top, cut the subtree with the smallest alpha to get T2, until T n Finally, test the sub-trees {T0, T1,..., T n} with the verification set, and the decision tree with the smallest square error is the optimal decision tree.

[0041] The application further discloses a fan blade icing and blade mass imbalance fault diagnosis system.

[0042] The data acquisition module is used for acquiring operation data and blade icing information of multiple fans.

[0043] The data preprocessing module is used for labeling the operation data with icing state labels according to the blade icing information, and obtaining fault data.

[0044] The feature selection module is used for extracting fault features in the fault data, sorting the fault features according to importance, and generating an optimal feature set.

[0045] The decision tree generation module is used for selecting optimal features in the optimal feature set to generate an optimal decision tree according to the square error minimization criterion.

[0046] The random forest module is used for classification based on the optimal decision tree to obtain a diagnosis detection result containing fault information.

[0047] The application further discloses a computer readable storage medium, which stores a computer program.

[0048] The application further discloses a computer device, which comprises a memory and a processor connected to each other.

[0049] Compared with the prior art, the application has the following advantages:

[0050] In the feature selection, the application adopts a wrapped feature selection method and combines cross-validation automatic optimization, has stronger feature judgment ability compared with a single model, and effectively avoids overfitting by extracting part of samples and features from all decision trees, is stable in performance, is suitable for parallel computing, and has higher diagnosis performance compared with other diagnosis methods while simplifying model training complexity. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flow chart of the fault diagnosis method of the present application in an embodiment.

[0052] Figure 2 is a schematic diagram of data label classification in data preprocessing in the present application.

[0053] Figure 3 is a schematic diagram of data label statistical results in the present application.

[0054] Figure 4 is a graph of the change of classification accuracy of the feature selection method proposed in the present application with the number of selected feature variables.

[0055] Figure 5 is a structure diagram of the random forest algorithm in the present application.

[0056] Figure 6 is a confusion matrix result diagram obtained by the algorithm model of the present application in fault diagnosis on the blade icing test set.

[0057] Figure 7 is a confusion matrix result diagram obtained by the algorithm model of the present application in fault diagnosis on the blade quality imbalance test set.

DETAILED DESCRIPTION

[0058] The present application is further described below in conjunction with the accompanying drawings and specific embodiments.

[0059] As shown in Figure 1, the embodiment of the present application provides a fan blade icing and blade quality imbalance fault diagnosis method based on a random forest model, which specifically includes the following steps:

[0060] Collecting data: collecting real-time operation data of multiple wind turbine generators and blade icing information through a SCADA system;

[0061] Icing state labeling: labeling the icing state label according to the blade icing information of each wind turbine generator in each operation period; preprocessing the operation data (including wind turbine operation parameters: unit operation state, wind speed, power, cabin temperature, environment temperature, cabin front and rear and left and right acceleration information, etc., see Table 1) to obtain preprocessed data, and dividing the data into a test set and a training set;

[0062] Table 1 SCADA system state data

[0063] Feature selection is performed on the data after preprocessing, power residual features, wind speed to power ratio, rotational speed residual features, temperature difference features, and vibration acceleration vector synthesis features are obtained according to the wind turbine operation time series data, and feature dimension reduction is performed to obtain an optimal feature subset;

[0064] The optimal feature subset is used to establish a random forest fault diagnosis model, the features in the optimal feature subset are taken as inputs, the current icing state label and future icing state label and icing state confidence are taken as outputs of the model, and fault diagnosis accuracy verification is performed.

[0065] As shown in FIGS. 2-3, the data preprocessing mainly includes marking the fault data rows as 1 (89% in FIG. 3), marking the normal data rows as 0 (6% in FIG. 3), and not participating in the evaluation of invalid data (5% in FIG. 3). The variables useless for modeling are deleted, the influence of invalid information on the model is reduced, the problems of missing and abnormal data are solved, and the data is better adapted to the model. The data set is divided into a training set and a test set according to a 2:1 ratio.

[0066] As shown in FIG. 4, the above feature selection method adopts the RFECV method, and outputs the feature importance ranking through logistic regression and support vector machine, including:

[0067] The data set is divided into a training set D 1j and a validation set D 2j by resampling.

[0068] The training set D 1j is used to train the model M ij (θ) with all features j 1j , and M ij (θ) is evaluated on the validation set D 2j .

[0069] The performance score of the trained model M ij (θ) is calculated, and j 1j is ranked from high to low according to importance. The features are selected in order from important to unimportant to form a feature set S ij , where i = 1, 2,..., m.

[0070] For the feature set, the training set D 1j is used to train the model M ij (θ) with the feature set S ij , and the model M ij (θ) is evaluated on the validation set D 2j to obtain the score of the model M ij (θ) on the feature set S ij .

[0071] The model scores of the same feature number set S ij from h times of resampling are averaged, and the feature number with the highest score is obtained.

[0072] The feature number is used to execute the algorithm on the original training set D to obtain the optimal feature set S max .

[0073] As shown in FIG. 5, the random forest classification fault diagnosis model integrates multiple decision tree basic units through the idea of ensemble learning, including:

[0074] The optimal feature is selected through the criterion of minimum square error to generate the above decision tree; wherein the decision tree is defined as:

[0075] In the formula, x is the input, f(x) is the output, M represents the division of the input space into M units: R1, R2,..., R m , and each unit R m has a fixed output value mc;

[0076] The left and right branch regions of the jth feature and the split point s of the above decision tree model are defined as: R1(j, s) = {x|x (j) ≤s}, R2(j, s) = {x|x (j) >s}

[0077] The optimal split feature j and split point s are obtained by solving the following formula:

[0078] In the formula, y i is the corresponding label value of x i , c1 and c2 are: C1 = ave(y i |x i ∈R1(j, s)), C2 = ave(y i |x i ∈R2(j, s))

[0079] From the above root node, the optimal split feature and split point are found to divide the input space into two regions, and then the division process is repeated for each region, and finally m division units R1, R2,..., R m are obtained, and the optimal output value of each unit is obtained according to the criterion of minimum square error; the optimal value of each unit R m is the mean value of the corresponding y i of all samples x i in the unit, which is defined as: C m = ave(y i |x i ∈R m ).

[0080] The above generation rule of decision tree includes:

[0081] k decision trees are established by sampling and selecting k sub-training sample sets from the total training sample set;

[0082] Randomly select m features from n features at each node of the decision tree, select the optimal feature from the m candidate features according to the principle of minimum node impurity, and grow the node by classification, so that the decision tree is fully grown until the impurity of each leaf node reaches the minimum, and the decision tree is not pruned;

[0083] Repeat the above steps to traverse the pre-built k decision trees, and form a random forest from the k decision trees;

[0084] According to the grown k decision trees, the classification results of new unknown samples are predicted, and the majority voting result of the k decision trees is determined.

[0085] The pruning of the decision tree is performed as follows: firstly, the pruning is performed from the bottom of the complete decision tree T0 to the root node to form a sub-tree sequence {T0, T1,..., T n}, and then the cross-validation method is used to test the sub-trees on the validation set to select the optimal sub-tree;

[0086] The loss function of the sub-tree in the pruning operation is defined as: C α (T)=C(T)+α|T|

[0087] In the formula, T is any sub-tree, C(T) is the prediction error of the training data, |T| is the number of leaf nodes of the sub-tree, and is a parameter greater than 0, which is used to balance the fitting degree of the training data and the complexity of the model, and C α (T) represents the overall loss of the sub-tree T when the parameter is a.

[0088] Starting from T0, calculate the a of each node from bottom to top, cut the sub-tree with the smallest a in T0 to get T1, then take T1 as the whole tree, calculate the a of each node from bottom to top, cut the sub-tree with the smallest a in T1 to get T2, and so on, until T n Finally, the sub-trees {T0, T1,..., T n} are tested on the validation set, and the decision tree with the smallest squared error is the optimal decision tree.

[0089] Specifically, the parameter setting of the random forest includes:

[0090] max_features: the maximum number of feature values considered when dividing the data set, which needs to be appropriately reduced for a large number of feature values.

[0091] max_depth: the maximum depth of the decision tree, which can not be set when the data is small or the features are few, and the default value can be used to complete the model training. If the sample size and the number of features of the model are large, the value can be limited, and the commonly used value range is 10-100, which can improve the training accuracy and efficiency while avoiding overfitting of the data.

[0092] min_samples_split: The minimum number of samples required to split an internal node (i.e. decision condition). If the node sample size is less than min_samples_split, it will not continue to try to split the decision tree. When the sample size is small, the value should be small to improve the training accuracy. If the sample size is very large, the value should be increased to speed up the training of the decision tree and avoid overfitting of the data.

[0093] min_samples_leaf: The minimum number of samples to reach a leaf node (i.e. classification). This value limits the minimum number of samples in a leaf node. If the number of leaf nodes is less than the sample size, it will be pruned together with its sibling nodes. When the sample size is small, the value should be small. If the sample size is very large, the value should be increased to speed up the training of the decision tree and improve efficiency.

[0094] min_weight_fraction_leaf: The minimum sum of weights of all samples in a leaf node (i.e. classification). This value limits the minimum sum of weights of all samples in a leaf node. If it is less than this value, it will be pruned together with its sibling nodes. The default is 0, which means that the weight issue is not considered and all samples have the same weight. For special data, the weight needs to be considered.

[0095] max_leaf_nodes: The maximum number of leaf nodes, with a default value of 0. This value limits the maximum number of leaf nodes, which is intended to prevent overfitting. If a limit is added, the algorithm will build the optimal decision tree within the maximum number of leaf nodes. If there are not many features, this value can be ignored. However, if there are many features, a limit can be added to achieve the best training model results.

[0096] min_impurity_decrease: The minimum impurity of the node, with a default value of 0. This value limits the growth of the decision tree, and the impurity of the node (Gini coefficient, information gain, mean square error, absolute difference) must be greater than this threshold, otherwise the node will not generate child nodes.

[0097] presort: Pre-sort the data, with a default value of False, indicating whether to pre-sort the data before fitting to speed up the construction of the tree. For a very large data set, presort=true will cause the entire classification to become slow; when the data set is small and the tree depth is limited, presort=true will only speed up the classification. It needs to be set according to the actual situation whether to turn it on.

[0098] The application adopts a wrapped feature selection method in feature selection, and combines cross-validation automatic optimization, has stronger feature judgment ability compared with a single model, and adopts a random forest classification method, extracts part of samples and features for all decision trees, effectively avoids overfitting, has stable performance and is suitable for parallel computing, compared with other diagnostic methods, has high diagnostic performance while simplifying model training complexity.

[0099] The model needs to be evaluated and optimized after being established. For binary classification problems, the confusion matrix related indicators are commonly used for evaluation:

[0100] The model needs to be evaluated and optimized after being established. For binary classification problems, the confusion matrix related indicators are commonly used for evaluation:

[0101] Wherein, TP is the number of samples that are true positive and predicted positive, FN is the number of samples that are true positive and predicted negative, FP is the number of samples that are true negative and predicted positive, and TN is the number of samples that are true negative and predicted negative.

[0102] Accuracy, precision, recall and F1-score:

[0103] Accuracy refers to the ratio of the total number of positive examples and negative examples predicted by the trained model to the total number of samples, defined as:

[0104] Precision refers to the proportion of true positive samples in the predicted positive samples, defined as:

[0105] Recall refers to the proportion of correctly predicted positive samples in the total number of positive samples, defined as:

[0106] The concept of F1-score is to average the precision and recall, defined as:

[0107] Figure 6 is a confusion matrix result obtained by random forest classification for fault diagnosis on the leaf icing test set, according to the data shown in Figure 6, the random forest classification model can be seen that the corresponding accuracy is 99.98%, the precision is 100%, the recall is 99.49%, and the F1-score is 1; Figure 7 is a confusion matrix result obtained by random forest classification for fault diagnosis on the leaf imbalance test set, according to the data shown in Figure 7, the random forest classification model can be seen that the corresponding accuracy is 99.71%, the recall is 100%, the recall is 99.67%, and the F1-score is 1. It can be seen that the random forest classification diagnosis model has high accuracy in diagnosing leaf icing and leaf imbalance, and realizes the goal of ensuring the performance of the model on the basis of simplifying the complexity of model training.

[0108] The embodiment of the present application also provides a fan blade icing and blade quality imbalance fault diagnosis system based on a random forest model, comprising:

[0109] A preprocessing module is configured to preprocess fault data to obtain preprocessed data and divide the preprocessed data into a test set and a training set.

[0110] A data input module is configured to input the preprocessed data output by the preprocessing module into a pre-trained random forest model.

[0111] A feature selection module is configured to use the RFECV method to output a feature importance ranking and generate an optimal feature subset by using a logistic regression and a support vector machine based on the preprocessed data input by the data input module.

[0112] A decision tree generation and pruning module is configured to select an optimal feature to generate a decision tree according to a least square error minimization criterion, continuously prune from a bottom end to a root node to form a sub-tree sequence, and obtain an optimal sub-tree through cross-validation.

[0113] A random forest module is configured to set random forest parameters, perform classification based on the optimal decision tree, and obtain a diagnosis detection result containing fault information.

[0114] The present application further discloses a computer readable storage medium having a computer program stored thereon, wherein the computer program performs the steps of the above method when executed by a processor. The present application also discloses a computer device comprising a memory and a processor connected to each other, wherein the memory has a computer program stored thereon, and the computer program performs the steps of the above method when executed by the processor. The system, medium and device of the present application correspond to the above method and also have the advantages of the above method.

[0115] The present application can realize all or part of the processes in the above-mentioned embodiment methods, and can also be completed by computer program instruction related hardware. The computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned method embodiment can be realized. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable storage medium includes any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. The memory is used to store computer programs and / or modules. The processor realizes various functions by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device, etc.

[0116] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiment. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principle of the present application shall be considered as the protection scope of the present application.

Claims

1. A wind turbine blade icing and blade mass imbalance fault diagnosis method, characterized in that, The method comprises the steps of: obtaining running data of multiple wind turbines and blade icing information; annotating the running data with icing state labels according to the blade icing information to obtain fault data; extracting fault features from the fault data and sorting the fault features according to importance to generate an optimal feature set; selecting optimal features in the optimal feature set according to the criterion of minimizing square error to generate an optimal decision tree; performing classification based on the optimal decision tree to obtain a diagnosis and detection result containing fault information.

2. The fan blade icing and blade mass imbalance fault diagnostic method of claim 1, wherein, The running data of the wind turbine includes wind speed, power, cabin temperature, and environmental temperature and cabin front, back, left and right acceleration.

3. The fan blade icing and blade mass imbalance fault diagnostic method of claim 2, wherein, The fault features include power residual feature, wind speed to power ratio, rotational speed residual feature, temperature difference feature and vibration acceleration vector synthesis feature.

4. The fan blade icing and blade mass imbalance fault diagnostic method according to claim 1 or 2 or 3, characterized by, The optimal feature set is generated by sorting the fault features according to importance through logistic regression and support vector machine, specifically including: Divide the faulty data into a training set D by resampling 1j and a validation set D 2j ; using the training set D 1j all features j 1j training a random forest failure diagnosis model M ij (0) on the validation set D 2j evaluating M ij (0) on the validation set D Through training the model M ij Calculate the performance score of j 1j Sort the features from high to low in importance, and select the features in turn from important to unimportant to form a feature set S ij Where i = 1, 2,..., m; For the feature set, use the training set D 1j Feature set S ij Training model M ij (θ), in the validation set D 2j Model evaluation was performed, and model M was obtained. ij (θ) in the feature set S ij The score on; The set S of the number of identical features after h resampling ij The average of the model scores is used to obtain the number of features with the highest scores; The algorithm is executed on the original training set D using the number of features to obtain the optimal feature set S max .

5. The fan blade icing and blade mass imbalance fault diagnostic method of claim 4, wherein, The random forest classification fault diagnosis model integrates multiple decision tree basic units through the idea of ensemble learning, including: The optimal feature is selected by a criterion of least square error minimization; wherein the decision tree is defined as: where x is the input, f(x) is the output, M represents the division of the input space into M cells: R1, R2,..., R m , and there is a fixed output value mc on each cell R m . The left and right branch regions of the jth feature and the feature split point s of the decision tree model are defined as: R1 (j,s) = {x I x (j) ≤ s}, R2 (j,s) = {x I x (j) > s} The optimal segmentation feature j and segmentation point s are obtained by solving the following equation: where y i is x i The corresponding label values, c1 and c2 are: C1= ave(y i |x i ∈R1(j,s)), C2= ave(y i |x i ∈R2(j,s)) The optimal split feature and split point are found from the root node to divide the input space into two regions, and then the process is repeated for each region to finally obtain m partition units R1, R2,... Rm m , and the optimal output value of each unit is obtained according to the minimum square error criterion; the optimal value of each unit R m is the mean value of y i corresponding to all samples x i in the unit, which is defined as: C m = ave(y i |x i ∈R m ).

6. The fan blade icing and blade mass imbalance fault diagnostic method of claim 5, wherein, The rules for generating the decision tree include: k sub-training sample sets are selected from the total training sample set by sampling, and k decision trees are established; At each node of the decision tree, m features are randomly selected from n indexes, and the optimal feature is selected from the m candidate indexes according to the principle of minimum node impurity to classify and grow the node, so that the decision tree is fully grown until the impurity of each leaf node reaches the minimum, and the decision tree is not pruned; The above steps are repeated to traverse the pre-built k decision trees to form a random forest from the k decision trees; The classification results of new unknown samples are predicted according to the grown k decision trees, and the majority vote result of the k decision trees is determined.

7. The fan blade icing and blade mass imbalance fault diagnostic method of claim 6, wherein, The pruning of the decision tree is performed by first pruning from the bottom of the complete decision tree T0 to the root node to form a sub-tree sequence {T0, T1,..., Tn}, and then testing the sub-trees on the validation set using the cross-validation method to select the optimal sub-tree. n}, and then testing the sub-trees on the validation set using the cross-validation method to select the optimal sub-tree. The loss function of the sub-tree in the above pruning operation is defined as: C α (T) = C(T) + a |T| where T is an arbitrary sub-tree, C(T) is the prediction error of the training data, |T| is the number of leaf nodes of the sub-tree, is a parameter greater than 0 to balance the fitting degree of the training data and the complexity of the model, and C α (T) represents the overall loss of the sub-tree T when the parameter is a. From T0, calculate the α of each node from bottom to top, cut the subtree with the minimum α in T0 to get T1, take T1 as the whole tree, calculate the α of each node from bottom to top, cut the subtree with the minimum α to get T2, and so on until T n Finally, test the subtrees {T0, T1,..., T n} with the verification set, and the decision tree with the minimum square error is the optimal decision tree. 8.A wind turbine blade icing and blade mass imbalance fault diagnosis system, characterized in that, It comprises: a data acquisition module for obtaining running data of multiple wind turbines and blade icing information; a data preprocessing module for annotating the running data with icing state labels according to the blade icing information to obtain fault data; a feature selection module for extracting fault features from the fault data and sorting the fault features according to importance to generate an optimal feature set; a decision tree generation module for selecting optimal features in the optimal feature set according to the criterion of minimizing square error to generate an optimal decision tree; a random forest module for performing classification based on the optimal decision tree to obtain a diagnosis and detection result containing fault information.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, performs the steps of the method of any one of claims 1-7.

10. A computer device comprising a memory and a processor connected to each other, wherein the memory has stored a computer program, characterized in that, The computer program, when executed by the processor, performs the steps of the method of any one of claims 1-7.

Images