Fan and blade icing fault diagnosis method and device thereof, storage medium and product

Abstract

The invention discloses a draught fan and a blade icing fault diagnosis method and device thereof, a storage medium and a computer program product, and the method comprises the steps: obtaining the operation data of a draught fan from an SCADA system of a wind power generation system, sequentially carrying out the preprocessing and feature engineering processing, obtaining the feature variables of the draught fan, and dividing a training set and a test set as a sample; optimizing the neural network model by using an HO algorithm to obtain an optimized neural network model; training the optimized neural network model by using the training set to obtain a trained neural network model; evaluating the training neural network model by using the test set to obtain an evaluation result; and correcting the training neural network model according to an evaluation result to obtain a corrected neural network model, and recording the corrected neural network model as a blade icing fault diagnosis model of the fan. According to the scheme, neural network training is carried out based on the data acquired by the SCADA system, the fan blade icing fault diagnosis model is obtained, the cost is low, and the detection precision is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and specifically relates to a method, device, fan, storage medium and computer program product for diagnosing blade icing fault of a wind turbine, and more particularly to a method, device, fan, storage medium and computer program product for intelligently diagnosing blade icing fault of a wind turbine based on a Hippo optimization algorithm. Background Art

[0002] Wind energy is a clean, safe, renewable green energy. It does not pollute the environment or damage the ecology. Its environmental benefits, energy benefits, and cost benefits are becoming more and more significant. It is of great significance to the sustainable development of human society. Wind power generation has thoroughly implemented the national energy security strategy and has become a pioneering force. Wind turbine blades play a vital role in wind power generation. They complete the process of wind power generation by capturing wind energy and converting it into mechanical energy, and then converting the mechanical energy into electrical energy through a generator. In cold areas, the icing of wind turbine blades causes the weight of the blades to increase, the aerodynamic characteristics to change, and the power generation efficiency to be affected, and even major safety hazards such as blade breakage and wind turbine shutdown. Therefore, it is of great significance to study efficient wind turbine blade icing fault diagnosis technology.

[0003] In some schemes, the method for detecting ice on wind turbine blades is to determine whether there is ice formation by directly detecting the ice on the blades, and use cameras, infrared cameras or vibration sensors to monitor the ice adhesion on the blade surface in real time, but the cost is high and the sensor installation and maintenance is inconvenient.

[0004] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The object of the present invention is to provide a method, device, fan, storage medium and computer program product for diagnosing blade icing fault of a fan, so as to solve the problem that the detection method of fan blade icing in related solutions needs to use a camera, an infrared camera or a vibration sensor to monitor the ice adhesion on the blade surface in real time, which has high cost and inconvenient sensor installation. The present invention can achieve the goal of obtaining a fan blade icing fault diagnosis model by training a neural network based on data obtained from a SCADA system, and automatically diagnosing the fan blade icing fault, which is not only low-cost and does not require the installation of sensors, but also can improve the detection accuracy of the model.

[0006] The present invention provides a method for diagnosing blade icing fault of a wind turbine, which is applied to a wind power generation system; the method for diagnosing blade icing fault of a wind turbine comprises: obtaining operation data of the wind turbine from a SCADA system of the wind power generation system; preprocessing the operation data of the wind turbine to obtain preprocessing data of the wind turbine; performing feature engineering processing based on the preprocessing data of the wind turbine to obtain characteristic variables of the wind turbine; and dividing the characteristic variables of the wind turbine into a training set and a test set as samples; optimizing a preselected neural network model by using an HO algorithm to obtain an optimized neural network model; training the optimized neural network model by using the training set to obtain a trained neural network model; evaluating the trained neural network model by using the test set to obtain an evaluation result; and correcting the trained neural network model according to the evaluation result to obtain a corrected neural network model, which is recorded as a blade icing fault diagnosis model of the wind turbine, so as to automatically diagnose the blade icing fault of the wind turbine by using the blade icing fault diagnosis model of the wind turbine.

[0007] In some embodiments, the operating data of the wind turbine include: the operating data of the wind turbine, the environmental data of the wind turbine, and the blade status data of the wind turbine; wherein the operating data of the wind turbine include: the time, wind speed, wind direction of the area where the wind turbine is located, and the speed, power and yaw information of the wind turbine; the environmental data of the wind turbine include: the ambient temperature and cabin temperature of the area where the wind turbine is located; the blade status data of the wind turbine include: the blade speed, blade pitch angle and blade acceleration of the wind turbine.

[0008] In some embodiments, the operation data of the fan is preprocessed to obtain the preprocessed data of the fan, including: classifying the operation data of the fan according to predetermined expert experience to obtain a classification result; and marking data labels according to the classification result; the data labels include: invalid data, normal data, and freezing data; according to the data labels, the invalid data, normal data, and freezing data in the operation data of the fan are cleaned to obtain the normal data and freezing data of the fan; the normal data and freezing data of the fan are oversampled and balanced by using the SMOTE algorithm to obtain the freezing oversampled data of the fan; and the freezing oversampled data of the fan are standardized by using the Z-score to obtain the freezing standardized data of the fan as the preprocessed data of the fan.

[0009] In some embodiments, feature engineering is performed based on the preprocessed data of the fan to obtain the characteristic variables of the fan, including: based on the preprocessed data of the fan, data visualization and physical mechanism analysis of blade icing are performed, global features are extracted using an XGBoost model, and the hyperparameters of the XGBoost model are optimized using a Bayesian algorithm to construct new characteristic variables; the new characteristic variables and the preprocessed data of the fan are used together as inputs of the XGBoost model to obtain the output of the XGBoost model as the characteristic variables of the fan.

[0010] In some embodiments, the new characteristic variable and the preprocessed data of the wind turbine are used as inputs of the XGBoost model to obtain the output of the XGBoost model as the characteristic variable of the wind turbine, including: using the new characteristic variable and the preprocessed data of the wind turbine as inputs of the XGBoost model, and using Bayesian optimization to tune the XGBoost model to obtain important features of the output of the XGBoost model; using the RFECV method to recursively eliminate the important features of the output of the XGBoost model to obtain the feature importance ranking and the optimal number of features of the output of the XGBoost model as the characteristic variables of the wind turbine.

[0011] In some embodiments, the HO algorithm is used to optimize the pre-selected neural network model to obtain an optimized neural network model, including: constructing a stacked CNN-BiLSTM-Transformer neural network model as the pre-selected neural network model; using the FL loss function as the loss function of the CNN-BiLSTM-Transformer neural network model; and using the HO algorithm to optimize the hyperparameters of the CNN-BiLSTM-Transformer neural network model to obtain an optimized neural network model.

[0012] Matching the above method, the present invention provides, on the other hand, a blade icing fault diagnosis device for a wind turbine, which is applied to a wind power generation system; the blade icing fault diagnosis device for a wind turbine comprises: an acquisition unit, configured to acquire the operating data of the wind turbine from the SCADA system of the wind power generation system; a control unit, configured to pre-process the operating data of the wind turbine to obtain the pre-processed data of the wind turbine; the control unit is further configured to perform feature engineering processing based on the pre-processed data of the wind turbine to obtain the characteristic variables of the wind turbine; and the characteristic variables of the wind turbine are used as samples to divide the training set and the test set; the control unit , and is further configured to optimize the pre-selected neural network model by using the HO algorithm to obtain an optimized neural network model; the control unit is further configured to train the optimized neural network model by using the training set to obtain a trained neural network model; the control unit is further configured to evaluate the trained neural network model by using the test set to obtain an evaluation result; and to correct the trained neural network model according to the evaluation result to obtain a corrected neural network model, recorded as the blade icing fault diagnosis model of the fan, so as to automatically diagnose the blade icing fault of the fan by using the blade icing fault diagnosis model of the fan.

[0013] Matching the above-mentioned device, the present invention provides a wind turbine on another aspect, including: the above-mentioned device for diagnosing blade icing fault of the wind turbine.

[0014] In accordance with the above method, the present invention provides a storage medium on the other hand, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the above-mentioned method for diagnosing blade icing faults of the wind turbine.

[0015] In accordance with the above method, the present invention provides a computer program product on another aspect, including a computer program, which implements the steps of the above-mentioned method for diagnosing blade icing faults of a wind turbine when executed by a processor.

[0016] Therefore, the solution of the present invention, for the wind turbine (i.e., wind turbine) used in the wind power generation system, obtains the operating data of the wind turbine from the SCADA system of the wind power generation system, and pre-processes the operating data of the wind turbine to obtain the pre-processed data of the wind turbine; performs feature engineering processing based on the pre-processed data of the wind turbine to obtain the characteristic variables of the wind turbine; optimizes the CNN-BiLSTM-Transformer model using the HO optimization algorithm to obtain the CNN-BiLSTM-Transformer optimization model; trains the CNN-BiLSTM-Transformer optimization model using the characteristic variables of the wind turbine to obtain a fan blade icing fault diagnosis model to automatically diagnose the fan blade icing fault; thus, by training the neural network based on the data obtained from the SCADA system, a fan blade icing fault diagnosis model is obtained, and the fan blade icing fault is automatically diagnosed, which is not only low-cost and does not require the installation of sensors, but also can improve the detection accuracy of the model.

[0017] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0018] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of an embodiment of a method for diagnosing icing fault of a fan blade according to the present invention;

[0020] Figure 2 A schematic diagram of a flow chart of an embodiment of preprocessing the operating data of the wind turbine in the method of the present invention;

[0021] Figure 3 It is a flow chart of an embodiment of performing feature engineering processing based on the pre-processed data of the wind turbine in the method of the present invention;

[0022] Figure 4 A flow chart of an embodiment of the method of the present invention in which the new characteristic variable and the pre-processed data of the wind turbine are used as inputs of the XGBoost model;

[0023] Figure 5 A flow chart of an embodiment of optimizing a pre-selected neural network model by using the HO algorithm in the method of the present invention;

[0024] Figure 6 It is a structural schematic diagram of an embodiment of a device for diagnosing blade icing faults of a fan of the present invention;

[0025] Figure 7This is a flowchart of feature engineering;

[0026] Figure 8 This is a schematic diagram of the structure of the CNN-BiLSTM-Transformer model;

[0027] Fig. 9 It is a flowchart of the HO optimization algorithm;

[0028] Fig.10 It is a flow chart of an intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm of the present invention;

[0029] Fig.11 Schematic diagrams of feature importance rankings of the present invention, wherein (a) is a schematic diagram of feature importance ranking one, (b) is a schematic diagram of feature importance ranking two, (c) is a schematic diagram of feature importance ranking three, and (d) is a schematic diagram of feature importance ranking four;

[0030] Fig.12 Schematic diagram of ice detection effect of the method of the present invention, wherein (a) is a schematic diagram of feature quantity, and (b) is a schematic diagram of importance;

[0031] Fig.13 The data description table is Table 1;

[0032] Fig.14 The comparison result table of the baseline model is Table 2;

[0033] Fig.15 The comparison result table of the optimization algorithm is Table 3;

[0034] Fig.16 The comparison results of model generalization are shown in Table 4.

[0035] In conjunction with the accompanying drawings, the reference numerals in the embodiments of the present invention are as follows:

[0036] 102 - acquisition unit; 104 - control unit. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0038] Considering that there are two types of methods for detecting ice on wind turbine blades: direct method and indirect method. The direct method determines whether there is ice by directly detecting the ice on the blades, and uses cameras, infrared cameras or vibration sensors to monitor the ice adhesion on the blade surface in real time. The indirect method indirectly determines whether ice has occurred by analyzing the operating parameters, environmental conditions or other factors of the wind turbine. However, the above methods are costly, complex to detect, inconvenient to install and maintain sensors, and highly dependent on expert domain knowledge.

[0039] At the same time, with the continuous development of data mining technology and wind turbine data acquisition and supervision (SCADA) systems, data-driven methods for detecting wind turbine blade icing have become a hot topic. For example, by reducing the dimension of the original features of SCADA data, the output features are combined with support vector machines (SVM) to identify the icing state of wind turbine blades. However, since shallow machine learning can only capture simple features and patterns, it is difficult to handle complex nonlinear relationships and learn deep features.

[0040] A fan blade icing fault diagnosis model based on deep learning is proposed. The model extracts features through recursive feature elimination (RFE), synthesizes icing samples using the B-SMOTE method, and uses a bidirectional gated recurrent unit neural network model for blade icing diagnosis. However, this method only focuses on the features at the SCADA data level, ignores the physical meaning of blade icing, does not consider the relationship between the upper and lower time series of the data, and the time information is easily lost.

[0041] Therefore, the scheme of the present invention proposes a method for diagnosing fan blade icing faults, specifically an intelligent method for diagnosing fan blade icing faults based on the Hippopotamus optimization algorithm. The physical mechanism of fan blade icing is studied and physical modeling is performed. A CNN-BiLSTM neural network that can effectively capture spatial features in time series data is introduced, and the output features are globally modeled using the Transformer model. The Hippopotamus (HO) algorithm is used to optimize the hyperparameters of this model. This method is not only low-cost and does not require the installation of sensors, but also can improve the detection accuracy of the model.

[0042] According to an embodiment of the present invention, a method for diagnosing blade icing faults of a fan is provided. Figure 1 The schematic diagram of the flow chart of an embodiment of the method of the present invention is shown in FIG. The method for diagnosing blade icing fault of a wind turbine is applied to a wind power generation system. In the solution of the present invention, as Figure 1 As shown, the method for diagnosing blade icing fault of a wind turbine includes: steps S110 to S160.

[0043] In step S110, the operation data of the wind turbine is obtained from the SCADA system of the wind power generation system.

[0044] In step S120, the operation data of the fan is preprocessed to obtain preprocessed data of the fan.

[0045] In step S130, feature engineering processing is performed based on the preprocessed data of the fan to obtain the characteristic variables of the fan; and the characteristic variables of the fan are used as samples to divide the training set and the test set.

[0046] In step S140, the pre-selected neural network model is optimized using the HO algorithm to obtain an optimized neural network model; wherein the pre-selected neural network model is such as a CNN-BiLSTM-Transformer model, and the optimized neural network model is such as a CNN-BiLSTM-Transformer optimized model.

[0047] In step S150, the optimized neural network model is trained using the training set to obtain a trained neural network model.

[0048] At step S160, the training neural network model is evaluated using the test set to obtain an evaluation result; and the training neural network model is corrected based on the evaluation result to obtain a corrected neural network model, which is recorded as the blade icing fault diagnosis model of the fan, so that the blade icing fault of the fan can be automatically diagnosed using the blade icing fault diagnosis model of the fan.

[0049] The solution of the present invention provides an intelligent diagnosis solution for wind turbine blade icing fault based on a Hippopotamus optimization algorithm, which belongs to the field of wind power operation and maintenance and industrial fault detection. The solution comprises: acquiring multivariate sensor data and wind turbine blade status data, including time series and status feature information, from a SCADA system; performing a series of preprocessing operations such as data labeling, data cleaning, data balancing and data segmentation on the multivariate sensor data; performing a physical mechanism analysis of wind turbine blade icing on the preprocessed data, and generating new status feature variables based on the original status feature information; constructing a stacked CNN-BiLSTM-Transformer neural network model; using a Hippopotamus (HO) optimization algorithm to optimize the hyperparameters of the model; and using a data set for training and testing to complete the diagnosis of wind turbine blade icing fault. In the scheme of the present invention, key features are extracted through feature engineering, which helps to reduce data redundancy and noise, and enables the model to focus on key factors related to icing. The addition of the CNN-BiLSTM-Transformer neural network model can capture the complex feature interaction relationship in the blade status data, establish nonlinear relationships between different features, and enable the model to more comprehensively understand the process of fault occurrence, thereby improving the model's icing detection performance and generalization ability, and can provide information on optimizing maintenance strategies for wind farm operations, ensuring the safe and efficient operation of wind turbines under cold conditions.

[0050] In some embodiments, the operating data of the wind turbine in step S110 includes: the operating data of the wind turbine, the environmental data of the wind turbine, and the blade status data of the wind turbine; wherein the operating data of the wind turbine include: the time, wind speed, wind direction of the area where the wind turbine is located, and the speed, power and yaw information of the wind turbine; the environmental data of the wind turbine include: the ambient temperature and cabin temperature of the area where the wind turbine is located; the blade status data of the wind turbine include: the blade speed, blade pitch angle and blade acceleration of the wind turbine.

[0051] Fig.10 FIG. 1 is a flow chart of an intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm of the present invention. Fig.10 As shown, the intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm, that is, the specific steps of the fan blade icing detection process include:

[0052] Step 1: Obtain the SCADA data of the wind turbine and the status data of the wind turbine blades during operation, including time series and status characteristic information.

[0053] In step 1, multi-sensor data and wind turbine blade status data, wherein the multi-sensor data includes the operating data and environmental data of the wind turbine. The operating data of the wind turbine include: time, wind speed, generator speed (such as the speed of the generator in the cabin), power, wind direction and yaw information (such as yaw position and yaw speed). Environmental data include: ambient temperature and cabin internal temperature (i.e., the temperature inside the wind turbine cabin). The wind turbine blade status data include: blade speed (such as fan blade 1 speed, fan blade 2 speed and fan blade 3 speed), blade pitch angle and blade acceleration.

[0054] The experimental data in the scheme of the present invention is provided by the Ministry of Industry and Information Technology of the People's Republic of China and used in the first industrial big data innovation competition. The data collection frequency is about 7 seconds, and each data set has 29 variables, including fan operation parameters, environmental parameters and state parameters. Fig.13 The data description table is Table 1, and the data description is as follows Fig.13 As shown in Table 1.

[0055] In some implementations, the specific process of preprocessing the operating data of the wind turbine to obtain the preprocessed data of the wind turbine in step S120 is described in the following exemplary embodiments.

[0056] Combine the following Figure 2 The flowchart of an embodiment of preprocessing the operating data of the fan in the method of the present invention further illustrates the specific process of preprocessing the operating data of the fan in step S120, including: steps S210 to S240.

[0057] Step S210, classifying the operation data of the wind turbine according to predetermined expert experience to obtain classification results; and marking data labels according to the classification results; the data labels include: invalid data, normal data, and icing data.

[0058] Step S220, according to the data label, the invalid data, normal data and frozen data in the operation data of the fan are cleaned to obtain the normal data and frozen data of the fan. That is, after the data label is marked, there are three types of data, including invalid data, normal data and frozen data, and only the normal data and frozen data are retained after cleaning.

[0059] Step S230: Perform oversampling balancing processing on the normal data and icing data of the fan by using the SMOTE algorithm to obtain icing oversampling data of the fan.

[0060] Step S240 , using the Z-score, normalizing the icing oversampling data of the fan to obtain icing normalized data of the fan as preprocessing data of the fan.

[0061] like Fig.10 As shown, the intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm, that is, the specific steps of the fan blade icing detection process also include: Step 2, performing a series of preprocessing operations such as data labeling, data cleaning, data balancing and data segmentation on the above-mentioned multi-sensor data.

[0062] In step 2, the data preprocessing process is as follows: first, all multivariate sensor data and fan blade status data are labeled according to expert experience, then invalid data (i.e., data with missing values) are cleaned, and finally, the SMOTE algorithm is used to oversample the icing samples to balance the data samples. Since the icing of fan blades is a slow process, in order to reduce the impact of noise, the data is re-segmented every minute and the mean is taken as the new sample feature.

[0063] Specifically, in step 2, the wind turbine operation data is cleaned and standardized.

[0064] There are about 5% invalid data in the fan data set, so data cleaning is performed on it. The ratio of normal data to icing data is 14.68:1, and the data is extremely unbalanced. SMOTE is used to oversample the minority class samples and adjust the ratio to 1:1. Since the icing of fan blades is a slow process, in order to reduce the impact of noise, the data is re-segmented every minute and the mean is taken as a new feature. In addition, the physical meaning of each feature variable of SCADA data is different, and the size difference is obvious. Therefore, Z-score is used to standardize the data. The formula is as follows:

[0065]

[0066] Among them: Z is the standardized score, i.e. Z-score; X is any sample data; μ is the sample mean; σ is the sample standard deviation.

[0067] In some implementations, the specific process of performing feature engineering processing based on the preprocessed data of the wind turbine in step S130 to obtain the characteristic variables of the wind turbine is described in the following exemplary embodiments.

[0068] Combine the following Figure 3 The flowchart of an embodiment of the method of the present invention for performing feature engineering processing based on the preprocessed data of the fan is shown, which further illustrates the specific process of performing feature engineering processing based on the preprocessed data of the fan in step S130, including: step S310 to step S320.

[0069] Step S310, based on the preprocessed data of the wind turbine, data visualization processing and blade icing physical mechanism analysis processing are performed, the global features are extracted using the XGBoost model, and the hyperparameters of the XGBoost model are optimized using the Bayesian algorithm to construct new feature variables.

[0070] Step S320, taking the new characteristic variable and the preprocessed data of the wind turbine as inputs of the XGBoost model, and obtaining the output of the XGBoost model as the characteristic variable of the wind turbine.

[0071] Figure 7 This is a flow chart of feature engineering. Figure 7 As shown in the figure, the specific implementation steps of feature engineering are as follows: first, the blade pitch angle θ, blade speed ν and pitch motor temperature T mt The distribution of is relatively consistent and has great correlation. The formula for generating new features can be expressed as:

[0072]

[0073] in: is the average blade pitch angle, is the average blade speed, is the average blade pitch motor temperature.

[0074] Secondly, it is found that the ratio of wind speed to power Trend1, the ratio of wind speed to generator speed Trend2, and the ratio of wind speed to power and generator speed Trend3 show a significant upward trend during the blade icing process. The new characteristic formula is:

[0075]

[0076] Where: N g is the engine speed, V is the wind speed, and P is the power.

[0077] Since the fan ambient temperature varies greatly over time, the cabin temperature will also fluctuate. In order to study the impact of the two, the value of the cabin temperature minus the ambient temperature Tdiff is used as a new feature variable, which can be expressed as:

[0078] Tdiff=T in -T out (17).

[0079] Where: T in is the cabin temperature, T out is the ambient temperature.

[0080] Finally, the influence of actual power and predicted power on the icing of wind turbine blades is studied, and the difference between predicted power and actual power is constructed as a new characteristic variable, which is expressed as:

[0081]

[0082] Where: p pre is the power predicted by the baseline model, P true is the actual power of the fan.

[0083] The solution of the present invention extracts key features through feature engineering, which helps to reduce data redundancy and noise and enables the model to focus on key factors related to icing.

[0084] like Fig.10 As shown, the intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm, that is, the specific steps of the fan blade icing detection process also include: Step 3, analyzing the physical mechanism of fan blade icing on the preprocessed data, and generating new state feature variables based on the original state feature information.

[0085] In step 3, the physical mechanism of wind turbine blade icing is analyzed and new feature variables are constructed. Specifically, feature variables are constructed through feature engineering, the XGBoost model is used to extract the global features of the feature variables constructed after feature engineering, and the Bayesian algorithm is used to optimize the XGBoost hyperparameters. For details, see formula (11) to formula (18); wherein the new feature variables include: the mean blade pitch angle, the mean blade speed and the mean pitch motor temperature; the ratio of wind speed to power, the ratio of wind speed to generator speed and the ratio of wind speed to power and generator speed; the difference between cabin temperature and ambient temperature; and the difference between predicted power and actual power.

[0086] In some embodiments, in step S320, the new characteristic variable and the preprocessed data of the wind turbine are used together as the input of the XGBoost model to obtain the output of the XGBoost model as the characteristic variable of the wind turbine. For the specific process, see the following exemplary description.

[0087] Combine the following Figure 4 The flowchart of an embodiment of the method of the present invention in which the new characteristic variables and the pre-processed data of the fan are used as the input of the XGBoost model is shown, and the specific process of using the new characteristic variables and the pre-processed data of the fan as the input of the XGBoost model in step S320 is further explained, including: steps S410 to S420.

[0088] Step S410, taking the new characteristic variables and the pre-processed data of the wind turbine as inputs of the XGBoost model, and using Bayesian optimization to tune the XGBoost model to obtain important features output by the XGBoost model.

[0089] Step S420, using the RFECV method to recursively eliminate the important features output by the XGBoost model, and obtain the feature importance ranking and the optimal number of features output by the XGBoost model as the characteristic variables of the wind turbine.

[0090] like Fig.10 As shown, the intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm, that is, the specific steps of the fan blade icing detection process also include: Step 4, the generated new state features are used together with the state features obtained on the SCADA system as the input of the XGBoost algorithm, and the recursive feature elimination cross validation (RFECV) method is used to recursively eliminate the features to obtain the feature importance ranking and the optimal number of features.

[0091] In step 4, in the fault diagnosis of wind turbine blade icing, there are a lot of redundant features in the SCADA data, which will affect the performance of model construction. The main goal of feature selection is to improve the effectiveness of fault diagnosis by accurately reducing the feature dimension: first, the generated new state features are used as the input of the XGBoost algorithm together with the state features obtained on the SCADA system, and the Bayesian optimization is used to tune the XGBoost model and select important features;

[0092] Then, the features were recursively eliminated using the recursive feature elimination cross-validation (RFECV) method. Finally, the relationship between the number of features and the cross-validation score was obtained by evaluating the model on the validation set. When the number of features was 24, the model cross-validation score was the highest.

[0093] In some implementations, the HO algorithm is used to optimize the pre-selected neural network model in step S140 to obtain a specific process of optimizing the neural network model, see the following exemplary description.

[0094] Combine the following Figure 5 The flowchart of an embodiment of optimizing a pre-selected neural network model by using the HO algorithm in the method of the present invention further illustrates the specific process of optimizing the pre-selected neural network model by using the HO algorithm in step S140, including: steps S510 to S520.

[0095] Step S510, constructing a stacked CNN-BiLSTM-Transformer neural network model as the pre-selected neural network model.

[0096] Step S520, using the FL loss function as the loss function of the CNN-BiLSTM-Transformer neural network model; and using the HO algorithm to optimize the hyperparameters of the CNN-BiLSTM-Transformer neural network model to obtain an optimized neural network model.

[0097] Figure 8 This is a schematic diagram of the structure of the CNN-BiLSTM-Transformer model. Figure 8 As shown in the figure, the structure of the CNN-BiLSTM-Transformer model. The CNN convolution layer extracts features through filters, the pooling layer can reduce the feature dimension and the amount of calculation, and finally the fully connected layer extracts features. The BiLSTM layer captures the contextual relationship in the long-range dependency and time series. The bidirectional mechanism allows the model to consider both past and future information at the same time, and improves the sensitivity to time-dependent features. The calculation formula of CNN-BiLSTM is as follows:

[0098]

[0099]

[0100] Where: x(t) is the input signal at time step t, x(t+i) is the value of the input signal or data at (t+i), ω(i) is the convolution kernel weight at position i, k is the convolution kernel size, b is the bias term, y(t) is the output feature, and “·” is the multiplication operation; represents the input gate of the forward BiLSTM, Represents the weight matrix corresponding to the input gate (forward propagation part), represents the hidden state of the previous time step (forward propagation part), represents the bias term of the input gate (forward propagation part), The forget gate of the forward BiLSTM, Represents the weight matrix corresponding to the forget gate (forward propagation part), Represents the bias term corresponding to the forget gate (forward propagation part); Represents the memory cell state of the forward LSTM at the current time step, represents the memory state of the previous time step, represents the weight matrix of the candidate memory unit (forward propagation), Represents the bias term of the candidate memory unit (forward propagation part); The output gate of the forward BiLSTM, Represents the weight matrix corresponding to the output gate (forward propagation part), Represents the bias term corresponding to the output gate (forward propagation part); represents the input gate of the backward BiLSTM, Represents the weight matrix corresponding to the input gate (back propagation part), represents the hidden state of the next time step (backward propagation part), represents the bias term of the input gate (backward propagation part), The forget gate of the backward BiLSTM, Represents the weight matrix corresponding to the forget gate (backward propagation part), Represents the bias term corresponding to the forget gate (backward propagation part); Represents the memory cell state of the backward LSTM at the current time step, represents the memory state of the next time step, represents the weight matrix of the candidate memory unit (backward propagation), Represents the bias term of the candidate memory unit (backward propagation part); The output gate of the reverse BiLSTM, Represents the weight matrix corresponding to the output gate (backward propagation part), Represents the bias term corresponding to the output gate (backward propagation part); h t represents the final hidden state.

[0101] The structure of the Transformer layer is mainly composed of a self-attention mechanism and a feedforward neural network. The self-attention mechanism allows the model to dynamically focus on information at different positions when processing the input sequence, thereby capturing long-distance dependencies. This feature effectively alleviates the problem that the CNN-BiLSTM network only uses the hidden state at the last moment when constructing the classifier and loses important features and information in the time series. The calculation formula of the Transformer layer is as follows:

[0102] PE (i,2j) = sin(i / 10000 2i / d ) (33);

[0103] PE (i,2j+1) =cos(i / 10000 2i / d ) (34);

[0104]

[0105] FFN(x)=max(0,xW 1 +b 1 )W 2 +b 2 (36);

[0106] Output=x+SubLayer(x) (37);

[0107] Output T =LayerNorm(x+SubLayer(x)) (38);

[0108] F = Flatten (Output T ) (39);

[0109]

[0110] Where: i represents the row of the matrix, j represents the column of the matrix, d represents the feature dimension of the model, PE (i,2j) Indicates the element position of the i-th row and 2j-th column of the matrix, sin() represents the sine function; PE (i,2j+1) represents the element position of the i-th row and the 2j+1-th column of the matrix, cos() represents the cosine function; Multiheadattention() represents multi-head attention; Q, K, V represent query, key, value, QK T is the dot product of the query matrix Q and the transpose of the key matrix K, d k is the dimension of the key, softmax() is the normalization of the attention score; FFN() represents a feedforward neural network, x represents the input vector, and W 1 represents the weight of the input to the hidden layer, W 2 represents the weight of the input to the output layer, b 1 ,b 2 represents the bias vector, max(0,...) represents the ReLU activation function; Output represents the residual connection, SubLayer(x) represents the result after processing by a certain sublayer; Output T represents the normalization layer, LayerNorm() represents the layer normalization operation; F is the flattening layer formula, Flatten() represents converting a multi-dimensional matrix (or tensor) into a one-dimensional vector; K is the total number of categories; Represents the probability of each classification, Dense() represents the fully connected layer; z j is the input score of class j, is to j Perform exponential operations, It is the sum of the scores of all categories after exponential operation, which serves as a normalization factor.

[0111] Fig. 9 Figure 1 is a flow chart of the HO optimization algorithm. Fig. 9 As shown, the specific steps of the HO optimization algorithm process are as follows:

[0112] Step 1: Initialization of the hippo population: Assuming there are n hippos in the d-dimensional space, the position of the i-th hippo in the d-dimensional space is x i , then the population initialization of hippos can be expressed as:

[0113] 2x i =lb j +r·(ub j -lb j )i=1,2,...,N=1,2,...,m (42).

[0114] Where: x i is the position vector of the i-th candidate solution; r is a random number in the range [0,1]; lb j andub j are the lower and upper bounds of the j-th decision variable, respectively; N represents the size of the hippo population in the group, and m represents the number of decision variables.

[0115] Step 2: Hippopotamus position update phase: The male hippopotamus position update formula can be expressed as:

[0116]

[0117] in: represents the position update of the i-th male hippopotamus in the j-th dimension; x i,j Indicates the current position of the hippopotamus, D hippo is the position of the best leader hippo in the current iteration; I 1 is a random integer, usually 1 or 2; y 1 is a random number.

[0118] For a young hippopotamus or a female hippopotamus, the position update formula is:

[0119]

[0120] in: Indicates the best position for a young hippopotamus or a female hippopotamus; x i,j represents the current position of the hippopotamus; MGi represents the average position of the randomly selected hippopotamus; h 1 ,h 2 is a randomly selected value; I 2 is a random integer; T is the probability of an immature hippopotamus and its mother deviating; D hippo is the position of the best performing leader hippo in the current iteration.

[0121] Step 3: The hippopotamus's defense against predators: The hippopotamus's main defense strategy is to quickly turn to the predator and make a loud call to force it to retreat. The predator's position can be represented as:

[0122]

[0123] Where: Predator j is the location of the predator; is a random number between 0 and 1; lb j andub j are the lower and upper bounds of the j-th decision variable respectively; is the distance from the hippopotamus to the predator; x i,j represents the current position of the hippopotamus; m represents the number of decision variables.

[0124] When a predator is very close, the hippopotamus will quickly turn and move towards the predator, forcing it to retreat; if the predator is far away, the hippopotamus will warn the predator of its presence in its territory. This process can be expressed by the formula:

[0125]

[0126] in: It is the position of the hippopotamus facing the predator; is a random vector according to the Levy flight distribution; f, c, d, g are random variables used to model the uncertainty of the predator's position; represents a random vector of 1×m dimensions; is the predator fitness function; F i Represents the fitness function of the current hippopotamus.

[0127] When the number of hippos is less than the original number of hippos, it indicates that the hippo has been hunted, and another hippo will replace it to join the group. Otherwise, the hunter will escape and the hippo will return to the group. This process can be expressed as:

[0128]

[0129] in: Indicates the number of hippos facing predators; F i HippoR represents the fitness function of a hippopotamus after facing a predator.

[0130] Step 4: The hippopotamus escapes from predators: When the hippopotamus faces multiple predators or fails to defend itself, it will choose to escape and try to escape to a nearby water source. This can be expressed as:

[0131]

[0132] in: is the location of the hippopotamus being searched for the nearest safe place; and is the local boundary in the current iteration; x i,jIndicates the current position of the hippopotamus; is a random number or vector that simulates the behavior of a hippopotamus looking for a safe location.

[0133] like Fig.10 As shown, the intelligent diagnosis method for fan blade icing fault based on the Hippo optimization algorithm, that is, the specific steps of the fan blade icing detection process also include: Step 5, constructing a stacked CNN-BiLSTM-Transformer neural network model.

[0134] In step 5, the stacked CNN-BiLSTM neural network is constructed as follows: first, a one-dimensional convolutional layer (Conv1D) is used to extract local features in the input data. Secondly, the features extracted by CNN are passed to the bidirectional LSTM layer to capture the temporal dependency of the features. Finally, a Transformer layer is added after the CNN-BiLSTM structure to extract deeper weighted features, thereby improving the ability to capture complex timing patterns of SCADA data.

[0135] Step 6: Use FL as the loss function of the CNN-BiLSTM-Transformer neural network model to establish an accurate mapping relationship between input and output variables;

[0136] In step 6, using FL as the loss function can better handle the sample imbalance problem and improve the ability to distinguish between categories, which is especially suitable for the binary classification icing fault detection task.

[0137] Step 7. Use the HO optimization algorithm to optimize the hyperparameters of the CNN-BiLSTM-Transformer neural network model, including the convolution kernel of the convolutional neural network, the convolution kernel size, the number of units of BiLSTM, the dropout rate, the learning rate, and the regularization coefficient.

[0138] In step 7, HO optimization is specifically as follows: The hippos represent candidate solutions to the optimization problem. The position update of each hippo in the search space represents the value of the decision variable. The position update strategy simulates the movement behavior of hippos in the natural environment, including exploring new areas and maintaining contact with the group. Hippos will adopt a series of defensive strategies when facing predators, such as turning quickly towards predators and making loud calls to prevent predators from approaching. These behaviors are used in the algorithm to simulate local optimization during the search process and avoid falling into local minima. When hippos cannot repel predators with defensive behaviors, they will try to leave the area. This behavior is used in the algorithm to enhance the global search capability and help the algorithm find better solutions in a wider search space.

[0139] In steps 3 to 7, the physical mechanism of wind turbine blade icing is analyzed and feature engineering is constructed. Specifically, the generated new state features are used together with the state features obtained on the SCADA system as the input of the XGBoost algorithm, and the Bayesian optimization is used to tune the XGBoost model and select important features; then, the recursive feature elimination cross-validation (RFECV) method is used to recursively eliminate the features; finally, the model is evaluated on the validation set, and the relationship between the number of features and the cross-validation score is obtained, as shown in Figure 2. Fig.11 As shown in (a), when the number of features is 24, the model cross-validation score is the highest. Fig.11 (b) It can be seen that the importance of the new features constructed by feature engineering is ranked very high, which shows that studying the physical mechanism of icing plays an important role in the detection of icing on wind turbine blades. Fig.11 Schematic diagrams of feature importance rankings of the present invention, wherein (a) is a schematic diagram of feature importance ranking one, (b) is a schematic diagram of feature importance ranking two, (c) is a schematic diagram of feature importance ranking three, and (d) is a schematic diagram of feature importance ranking four.

[0140] The solution of the present invention combines the architecture of CNN, BiLSTM and Transformer, and the model can capture the complex feature interaction relationship in the blade status data. This multi-module architecture can not only extract local and global information, but also establish nonlinear relationships between different features, so that the model can more comprehensively understand the process of fault occurrence.

[0141] Step 8: Sort the optimal number of features by feature importance and send them to the CNN-BiLSTM-Transformer neural network model optimized by HO.

[0142] In step 8, the optimal number of features is used as the input of the CNN-BiLSTM-Transformer model, and FL is used as the loss function, which can be expressed as:

[0143] FL(p t )=-α t (1-p t ) γ log(p t ) (50).

[0144] Where: FL(p t ) represents the focal loss function value; p t is the predicted probability of the true category; α t is the balance factor; γ is the focus adjustment parameter, (1-p t ) γ Focus factor.

[0145] Step 9: Use the data set for training and testing to complete the diagnosis of wind turbine blade icing faults.

[0146] In step 9, after the model training is completed, it can be evaluated through the test set to determine the performance of the model in the wind blade icing fault diagnosis task.

[0147] The solution of the present invention uses the HO optimization algorithm for parameter tuning of the model, can find the optimal hyperparameter combination, improve the generalization performance of the model, reduce the risk of overfitting, and make the icing detection effect of the model under different working conditions more robust.

[0148] The solution of the present invention proposes an intelligent diagnosis solution for fan blade icing fault based on the Hippopotamus optimization algorithm. By studying the physical mechanism of fan blade icing and physically modeling it, a CNN-BiLSTM neural network that can effectively capture spatial features in time series data is introduced, and the output features are globally modeled using the Transformer model. The Hippopotamus (HO) algorithm is used to optimize the hyperparameters of this model. It is not only low-cost and does not require the installation of sensors, but also can improve the detection accuracy of the model.

[0149] The effect of the solution of the present invention can be verified and explained through experimental results in the following aspects: setting of model parameters, evaluation index of the model, comparison of baseline models, comparison of optimization algorithms, comparison of model generalization, etc.

[0150] In terms of model parameter settings : The solution of the present invention divides the data set into training set, test set and validation set in a ratio of 6:2:2, and then compares them with traditional machine learning RF, SVM and deep learning models CNN, BiLSTM, CNN-BiLSTM, and CNN-BiLSTM-Transformer.

[0151] The solution of the present invention uses HO to optimize the hyperparameters of CNN-BiLSTM-Transformer, and the hyperparameters include the convolution kernel of the convolutional neural network, the convolution kernel size, the number of units of BiLSTM, the drop rate, the learning rate and the regularization coefficient, which are respectively optimized in [16, 128], [1, 10], [50, 200], [0.2, 0.5], [0.0001, 0.01] and [0.00001, 0.01], the number of iterations is 200, and the α and γ of FL are set to 0.25 and 2.0.

[0152] In terms of model evaluation indicatorsIn the diagnosis of wind turbine blade icing fault, the key performance indicators include accuracy, precision, recall, F1 score and area under the curve (AUC), which are specifically expressed as:

[0153]

[0154] Among them: TP is true positive, TN is true negative, FP is false positive, FN is false negative, M is the number of positive samples, N is the number of negative samples, rank insi is the sample index, Indicates the number of samples belonging to the positive set.

[0155] Fig.12 The figure is a schematic diagram of the ice detection effect of the method of the present invention, wherein (a) is a schematic diagram of the number of features and (b) is a schematic diagram of importance. In order to better explain the learning process of the model, the scheme of the present invention uses the t-distributed stochastic neighbor embedding (t-SNE) method to visualize the features of the model at different stages. The results are shown in Figure 2. Fig.12 As shown in the figure, the distribution of the original data is very chaotic, the clustering effect is poor, and the normal data and the ice data are mixed. After learning through the CNN-BiLSTM network, the model can distinguish normal samples from ice samples, the distance between different categories gradually increases, and the clustering effect becomes more obvious. After the Transformer layer, the samples of the same category in the model are closer, and the distance within the class becomes smaller, achieving good results. Although there are still some samples that are difficult to distinguish, overall, after the features of the model are extracted, the feature distribution of the blade ice state and the normal state has been significantly improved.

[0156] In terms of baseline model comparison:The model proposed by the scheme of the present invention outperforms other baseline methods as a whole, and improves the precision, recall, F1 score, accuracy and AUC by 22.9% and 21.4%, 30.2% and 18.2%, 56.3% and 23.8%, 31.3% and 23.7%, 15.9% and 13.1% respectively compared with the traditional machine learning model RF and SVM model. In addition, overall, the deep learning model is better than the traditional machine learning model. Among the deep learning models, the performance of the CNN-BiLSTM model is better than that of CNN and BiLSTM, with an improvement of 9.7% and 4.4%, 9.2% and 0.2%, 9.4% and 2.4%, and 7.2% and 2.2% in precision, recall, F1 score, accuracy and AUC, which shows that the CNN-BiLSTM layer is stronger in learning ability of multi-dimensional time series. By introducing the Transformer layer, the model's precision, recall, F1 score, and accuracy are improved by 4.5%, 6.64%, 4.55%, and 4.43%, respectively, which shows that the multi-head attention mechanism helps to enhance the learning ability of the model. Fig.14 The baseline model comparison results are shown in Table 2. Fig.14 As shown in Table 2.

[0157] Comparison of optimization algorithms :In order to verify the effectiveness of the optimization algorithm proposed in the solution of the present invention, the performance of various optimization algorithms, including AVOA, GWO, SSA and PSO, was analyzed in the experiment. This algorithm is used to optimize the hyperparameters of the CNN-BiLSTM-Transformer model. The optimization target hyperparameters and their search range are consistent with HO, and the fitness function is the F1 score. The number of iterations and population size of all optimization algorithms are set to 10 and 30 times. HO is 3.2%, 2.8%, 3%, 3%, and 0.8% higher than other optimization algorithms in precision, recall, F1 score, accuracy, and AUC on average. Fig.15 The comparison results of the optimization algorithms are shown in Table 3. Fig.15 As shown in Table 3.

[0158] In terms of model generalization comparison : In order to further evaluate the generalization of the model, the solution of the present invention sets the data of fan No. 15 as the training set, 50% of the data of fan No. 21 as the validation set and 50% as the test set, and selects three models with better comparison model effects for comparison. Fig.16 The comparison results of model generalization are shown in Table 4. Fig.16 As shown in Table 4.

[0159] The scheme of the present invention proposes an intelligent diagnosis scheme for fan blade icing fault based on the Hippo optimization algorithm, which effectively solves the problems of insufficient research on the characteristics of fan blade icing mechanism, low sensitivity of time series learning, data imbalance, and the optimization algorithm being prone to fall into local optimal solutions and slow convergence in complex problems. The results show that compared with other methods, this method has better fault diagnosis ability and better generalization. By stacking CNN-BiLSTM layers to capture the time information in SCADA data, CNN-BiLSTM has better detection performance on fan data than RF, SVM, CNN and BiLSTM. The Transformer model is introduced to selectively learn the global information of the SCADA sequence, and the CNN-BiLSTM-Transformer model is superior to the single CNN-BiLSTM network in robustness and detection accuracy. The HO algorithm is superior to other optimization algorithms, thereby further improving the diagnostic performance of the CNN-BiLSTM-Transformer model on fans and the generalization performance of different fans. By analyzing the feature sensitivity of data at different scales, it was found that the data after feature engineering is more sensitive to icing of wind turbine blades, which is beneficial to enhancing the detection accuracy and efficiency of the model.

[0160] By adopting the technical solution of this embodiment, for the wind turbine (i.e., wind turbine) used in the wind power generation system, the operating data of the wind turbine is obtained from the SCADA system of the wind power generation system, and the operating data of the wind turbine is preprocessed to obtain the preprocessed data of the wind turbine; feature engineering processing is performed based on the preprocessed data of the wind turbine to obtain the characteristic variables of the wind turbine; the CNN-BiLSTM-Transformer model is optimized by using the HO optimization algorithm to obtain the CNN-BiLSTM-Transformer optimization model; the CNN-BiLSTM-Transformer optimization model is trained by using the characteristic variables of the wind turbine to obtain a fan blade icing fault diagnosis model to automatically diagnose the fan blade icing fault; thus, by training the neural network based on the data obtained by the SCADA system, the fan blade icing fault diagnosis model is obtained, and the fan blade icing fault is automatically diagnosed, which is not only low-cost and does not require the installation of sensors, but also can improve the detection accuracy of the model.

[0161] According to an embodiment of the present invention, a device for diagnosing blade icing fault of a fan corresponding to the method for diagnosing blade icing fault of a fan is also provided. Figure 6 The schematic diagram of the structure of an embodiment of the device of the present invention is shown. The device for diagnosing blade icing fault of a wind turbine is applied to a wind power generation system; the device for diagnosing blade icing fault of a wind turbine may include: an acquisition unit 102 and a control unit 104.

[0162] The acquisition unit 102 is configured to acquire the operation data of the wind turbine from the SCADA system of the wind power generation system. The specific functions and processing of the acquisition unit 102 refer to step S110.

[0163] The control unit 104 is configured to pre-process the operation data of the fan to obtain pre-processed data of the fan. The specific functions and processing of the control unit 104 refer to step S120.

[0164] The control unit 104 is further configured to perform feature engineering processing based on the preprocessed data of the fan to obtain the characteristic variables of the fan; and use the characteristic variables of the fan as samples to divide the training set and the test set. The specific functions and processing of the control unit 104 are also shown in step S130.

[0165] The control unit 104 is further configured to optimize the pre-selected neural network model using the HO algorithm to obtain an optimized neural network model; wherein the pre-selected neural network model is such as a CNN-BiLSTM-Transformer model, and the optimized neural network model is such as a CNN-BiLSTM-Transformer optimized model. The specific functions and processing of the control unit 104 are also referred to step S140.

[0166] The control unit 104 is further configured to train the optimized neural network model using the training set to obtain a trained neural network model. The specific functions and processing of the control unit 104 are also shown in step S150.

[0167] The control unit 104 is further configured to evaluate the training neural network model using the test set to obtain an evaluation result; and to modify the training neural network model according to the evaluation result to obtain a modified neural network model, which is recorded as the blade icing fault diagnosis model of the fan, so as to automatically diagnose the blade icing fault of the fan using the blade icing fault diagnosis model of the fan. The specific functions and processing of the control unit 104 are also shown in step S160.

[0168] In the scheme of the present invention, multivariate sensor data and wind turbine blade status data, including time series and status feature information, are obtained from the SCADA system; a series of preprocessing operations such as data labeling, data cleaning, data balancing and data segmentation are performed on the above multivariate sensor data; the physical mechanism of wind turbine blade icing is analyzed on the preprocessed data, and new state feature variables are generated on the original state feature information; a stacked CNN-BiLSTM-Transformer neural network model is constructed; the hyperparameters of the model are optimized using the Hippo (HO) optimization algorithm; and the data set is used for training and testing to complete the wind turbine blade icing fault diagnosis. In the scheme of the present invention, key features are extracted through feature engineering, which helps to reduce data redundancy and noise, and allows the model to focus on key factors related to icing. The addition of the CNN-BiLSTM-Transformer neural network model can capture the complex feature interaction relationship in the blade status data, establish nonlinear relationships between different features, enable the model to more comprehensively understand the process of fault occurrence, improve the model's icing detection performance and generalization ability, and provide information for optimizing maintenance strategies for wind farm operations to ensure safe and efficient operation of wind turbines under cold conditions.

[0169] Since the processing and functions implemented by the device of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0170] According to an embodiment of the present invention, a fan corresponding to the blade icing fault diagnosis device of the fan is also provided. The fan may include: the blade icing fault diagnosis device of the fan described above.

[0171] Since the processing and functions implemented by the fan of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned device, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0172] According to an embodiment of the present invention, a storage medium corresponding to a method for diagnosing blade icing fault of a wind turbine is also provided, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the above-mentioned method for diagnosing blade icing fault of a wind turbine.

[0173] Since the processing and functions implemented by the storage medium of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0174] According to an embodiment of the present invention, a computer program product corresponding to the method is also provided, including a computer program, which implements the steps of the above-mentioned method for diagnosing blade icing faults of a wind turbine when executed by a processor.

[0175] Since the processing and functions implemented by the product of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0176] In summary, it is easy for those skilled in the art to understand that, under the premise of no conflict, the above-mentioned advantageous methods can be freely combined and superimposed.

[0177] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A method for diagnosing blade icing faults of a fan, characterized in that: Applied to a wind power generation system; the method for diagnosing blade icing faults of a wind turbine comprises: Obtaining the operation data of the wind turbine from the SCADA system of the wind power generation system; Preprocessing the operating data of the fan to obtain preprocessed data of the fan; Performing feature engineering processing based on the preprocessed data of the fan to obtain the characteristic variables of the fan; and using the characteristic variables of the fan as samples to divide the training set and the test set; The pre-selected neural network model is optimized using the HO algorithm to obtain an optimized neural network model; Using the training set to train the optimized neural network model to obtain a trained neural network model; The training neural network model is evaluated using the test set to obtain an evaluation result; and the training neural network model is corrected according to the evaluation result to obtain a corrected neural network model, which is recorded as the blade icing fault diagnosis model of the fan, so as to automatically diagnose the blade icing fault of the fan using the blade icing fault diagnosis model of the fan.

2. The method for diagnosing blade icing fault of a fan according to claim 1, characterized in that: The operation data of the fan include: operating data of the fan, environmental data of the fan, and blade status data of the fan; The operating data of the wind turbine include: the time, wind speed, wind direction of the area where the wind turbine is located, and the speed, power and yaw information of the wind turbine; The environmental data of the fan include: the ambient temperature of the area where the fan is located and the cabin temperature; The blade status data of the wind turbine include: the blade speed, blade pitch angle and blade acceleration of the wind turbine.

3. The method for diagnosing blade icing fault of a fan according to claim 1 or 2, characterized in that: Preprocessing the operating data of the fan to obtain preprocessed data of the fan includes: According to predetermined expert experience, the operation data of the wind turbine is classified to obtain a classification result; and data labels are marked according to the classification result; the data labels include: invalid data, normal data, and icing data; According to the data tags, invalid data, normal data and freezing data in the operation data of the fan are cleaned to obtain normal data and freezing data of the fan; Using the SMOTE algorithm to perform oversampling balancing processing on the normal data and icing data of the fan to obtain icing oversampling data of the fan; The Z-score is used to perform standardization processing on the icing oversampling data of the fan to obtain the icing standardized data of the fan as the preprocessing data of the fan.

4. The method for diagnosing blade icing fault of a wind turbine according to any one of claims 1 to 3, characterized in that: Feature engineering processing is performed based on the preprocessed data of the fan to obtain characteristic variables of the fan, including: Based on the pre-processed data of the wind turbine, data visualization processing and blade icing physical mechanism analysis processing are performed, the global features are extracted using the XGBoost model, and the hyperparameters of the XGBoost model are optimized using the Bayesian algorithm to construct new feature variables; The new characteristic variable and the preprocessed data of the wind turbine are used together as inputs of the XGBoost model to obtain the output of the XGBoost model as the characteristic variable of the wind turbine.

5. The method for diagnosing blade icing fault of a fan according to claim 4, characterized in that: The new characteristic variable and the pre-processed data of the wind turbine are used as inputs of the XGBoost model to obtain the output of the XGBoost model as the characteristic variable of the wind turbine, including: The new characteristic variables and the preprocessed data of the wind turbine are used as inputs of the XGBoost model, and the XGBoost model is tuned by using Bayesian optimization to obtain important features output by the XGBoost model; The RFECV method is used to recursively eliminate the important features output by the XGBoost model, and the feature importance ranking and the optimal number of features output by the XGBoost model are obtained as the characteristic variables of the wind turbine.

6. The method for diagnosing blade icing fault of a wind turbine according to any one of claims 1 to 5, characterized in that: The HO algorithm is used to optimize the pre-selected neural network model to obtain an optimized neural network model, including: Build a stacked CNN-BiLSTM-Transformer neural network model as the pre-selected neural network model; The FL loss function is used as the loss function of the CNN-BiLSTM-Transformer neural network model; and the HO algorithm is used to optimize the hyperparameters of the CNN-BiLSTM-Transformer neural network model to obtain the optimized neural network model.

7. A device for diagnosing blade icing faults of a fan, characterized in that: Applicable to a wind power generation system; the wind turbine blade icing fault diagnosis device comprises: an acquisition unit configured to acquire the operation data of the wind turbine from the SCADA system of the wind power generation system; A control unit is configured to pre-process the operation data of the fan to obtain pre-processed data of the fan; The control unit is further configured to perform feature engineering processing based on the preprocessed data of the fan to obtain the characteristic variables of the fan; and use the characteristic variables of the fan as samples to divide the training set and the test set; The control unit is further configured to optimize the pre-selected neural network model using the HO algorithm to obtain an optimized neural network model; The control unit is further configured to train the optimized neural network model using the training set to obtain a trained neural network model; The control unit is further configured to evaluate the trained neural network model using the test set to obtain an evaluation result; and to correct the trained neural network model according to the evaluation result to obtain a corrected neural network model, recorded as the blade icing fault diagnosis model of the fan, so as to automatically diagnose the blade icing fault of the fan using the blade icing fault diagnosis model of the fan.

8. A fan, characterized in that: include: A blade icing fault diagnosis device for a wind turbine according to any one of claims 1 to 7.

9. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the method for diagnosing blade icing faults of a wind turbine according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method for diagnosing blade icing faults of a wind turbine according to any one of claims 1 to 7 are implemented.

Images