Fan blade fault diagnosis method and device based on differential evolution optimization attention mechanism LSTM, electronic equipment and medium

Abstract

The application discloses a kind of attention mechanism LSTM fan blade fault diagnosis method, device, electronic equipment and medium based on differential evolution optimization.The method can include: collecting SCADA data, determining feature dataset;LSTM neural network is trained by feature dataset, and output state is obtained;Output state is input to attention mechanism model, and output result is obtained;Output result is obtained by output module to obtain prediction result, and prediction result is compared with fan actual state, and output error is calculated;According to output error, differential evolution calculation is carried out, and the parameters of LSTM neural network and attention mechanism network model are updated;Fan data to be tested is input into the model after training, and the diagnosis result of fan blade fault is obtained.The application has higher comprehensive performance and generalization ability, and has the advantages of improving the accuracy and speed of fan blade fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of wind power generation fault detection and diagnosis, and more specifically, to a method, device, electronic device and medium for fault diagnosis of wind turbine blades based on differential evolution optimization attention mechanism LSTM. Background Technology

[0002] With the urgent global demand for clean energy, wind power, as a sustainable and renewable energy source, has received widespread attention. Wind turbines are the core component of wind power systems, and their performance and reliability directly affect the overall system's operational efficiency. As a crucial part of wind turbines, blades are susceptible to icing and other faults due to the complex and variable external environment and mechanical wear, leading to a decline in turbine lifespan, stability, safety, and production efficiency. Therefore, it is essential to quickly detect wind turbine blade failures before they cause further losses. Using traditional mathematical methods to analyze wind turbine operating data makes it difficult to detect imbalance faults caused by blade icing, as the characteristic differences between normal and fault conditions are not significant.

[0003] In recent years, with the development of artificial intelligence technology, more and more methods have been applied to wind power generation systems, such as wind power output prediction. Wind power generation systems generate massive amounts of data, and neural network models have a significant advantage in processing large-scale data.

[0004] By utilizing rich data for classification, prediction, and recognition, the uncertainty caused by insufficient parameters or environmental factors was successfully eliminated. Among these methods, Long Short-Term Memory (LSTM) neural networks, as a deep learning model, improved upon the relatively insensitive nature of Convolutional Neural Networks (CNNs) to temporal information when processing sequential data, and solved the problem of traditional Recurrent Neural Networks (RNNs) losing their learning ability for long-term sequential data.

[0005] To improve the accuracy of model predictions, analyzing and extracting fault features, discovering potential feature information, and integrating it into the model are effective ways to enhance model performance. Wind turbine data is time-series data; therefore, learning the temporal relationships between data is crucial for wind turbine fault prediction. However, current fault detection models perform poorly in handling time-series problems, failing to effectively uncover potential temporal features between data points, resulting in poor generalization ability. Furthermore, the model's running speed needs improvement.

[0006] Therefore, it is necessary to develop a fault diagnosis method, device, electronic equipment, and medium for wind turbine blades based on differential evolution optimization attention mechanism LSTM with strong generalization ability, high prediction accuracy, and fast speed, so as to improve the accuracy and efficiency of fault diagnosis.

[0007] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0008] This invention proposes a method, device, electronic device, and medium for wind turbine blade fault diagnosis based on differential evolution optimization attention mechanism LSTM. It can overcome the shortcomings of existing wind turbine blade fault detection technologies and improve the accuracy and efficiency of wind turbine blade fault diagnosis.

[0009] In a first aspect, embodiments of this disclosure provide a method for fault diagnosis of wind turbine blades based on an attention mechanism LSTM using differential evolution optimization, including:

[0010] Collect SCADA data and determine the feature dataset;

[0011] The LSTM neural network is trained using the aforementioned feature dataset to obtain the output state;

[0012] The output state is input into the attention mechanism model to obtain the output result;

[0013] The output results are passed through the output module to obtain the prediction results. The prediction results are compared with the actual state of the wind turbine to calculate the output error.

[0014] Differential evolution calculation is performed based on the output error to update the parameters of the LSTM neural network and attention mechanism network model;

[0015] The data of the wind turbine to be tested is input into the trained model to obtain the diagnostic results of wind turbine blade failure.

[0016] As a specific implementation of this disclosure, collecting SCADA data and determining the feature dataset includes:

[0017] The SCADA data is the time-series data of the raw wind turbine and blade status collected from the SCADA system, which includes performance data and external environmental data.

[0018] The XGBoost algorithm is used to perform feature correlation analysis on the SCADA data to obtain the feature dataset that has a high correlation with wind turbine blade failure.

[0019] As one specific implementation of this disclosure, the output state is:

[0020] h t =o t ·d t

[0021] Among them, h t For output state, o t It is the output information of the output gate, d t =tanh(C t ), where tanh is the activation function.

[0022] As a specific implementation of this disclosure, the output state is input into the attention mechanism model to obtain the output result, including:

[0023] Calculate the degree of matching between the input data of the LSTM at time k and the output state of the LSTM at time t;

[0024] The element weights are calculated based on the matching degree, and then the corresponding output results are calculated based on the output state.

[0025] As a specific implementation of this disclosure, the output result is as follows:

[0026]

[0027] Among them, Y t The output of the attention mechanism model is given by k,t∈[1,n_step], where a k,t Indicate u k,t The probability distribution at different times, where n_step is the time step of the input data to the LSTM network.

[0028] As a specific implementation of this disclosure, the output error is:

[0029]

[0030] In the formula, L represents the output error, N represents the number of samples, and y (i) y represents the true sample category. ′(i) This indicates the predicted sample category of the model.

[0031] As one specific implementation of this disclosure, the differential evolution calculation includes:

[0032] The output error is used as the fitness function to determine the control parameters of the differential evolution algorithm;

[0033] Randomly generate the initial population;

[0034] Calculate the fitness of each individual in the population;

[0035] Determine whether the termination condition or the number of generations has been reached. If so, terminate the evolution and output the best individual as the optimal solution. If not, perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population, and repeat the fitness calculation.

[0036] Secondly, this disclosure also provides a fault diagnosis device for wind turbine blades based on an attention mechanism LSTM using differential evolution optimization, comprising:

[0037] The acquisition module collects SCADA data and determines the feature dataset;

[0038] The training module trains the LSTM neural network using the feature dataset to obtain the output state;

[0039] The attention mechanism module inputs the output state into the attention mechanism model to obtain the output result;

[0040] The output module obtains the prediction result from the output result, and calculates the output error by comparing the prediction result with the actual state of the wind turbine.

[0041] The differential evolution module performs differential evolution calculations based on the output error to update the parameters of the LSTM neural network and the attention mechanism network model.

[0042] The diagnostic module inputs the data of the wind turbine to be tested into the trained model to obtain the diagnostic results of wind turbine blade failure.

[0043] As a specific implementation of this disclosure, collecting SCADA data and determining the feature dataset includes:

[0044] The SCADA data is the time-series data of the raw wind turbine and blade status collected from the SCADA system, which includes performance data and external environmental data.

[0045] The XGBoost algorithm is used to perform feature correlation analysis on the SCADA data to obtain the feature dataset that has a high correlation with wind turbine blade failure.

[0046] As one specific implementation of this disclosure, the output state is:

[0047] h t =o t ·d t

[0048] Among them, h t For output state, ot It is the output information of the output gate, d t =tanh(C t ), where tanh is the activation function.

[0049] As a specific implementation of this disclosure, the output state is input into the attention mechanism model to obtain the output result, including:

[0050] Calculate the degree of matching between the input data of the LSTM at time k and the output state of the LSTM at time t;

[0051] The element weights are calculated based on the matching degree, and then the corresponding output results are calculated based on the output state.

[0052] As a specific implementation of this disclosure, the output result is as follows:

[0053]

[0054] Among them, Y t The output of the attention mechanism model is given by k,t∈[1,n_step], where a k,t Indicate u k,t The probability distribution at different times, where n_step is the time step of the input data to the LSTM network.

[0055] As a specific implementation of this disclosure, the output error is:

[0056]

[0057] In the formula, L represents the output error, N represents the number of samples, and y (i) y′ represents the true sample class. (i) This indicates the predicted sample category of the model.

[0058] As one specific implementation of this disclosure, the differential evolution calculation includes:

[0059] The output error is used as the fitness function to determine the control parameters of the differential evolution algorithm;

[0060] Randomly generate the initial population;

[0061] Calculate the fitness of each individual in the population;

[0062] Determine whether the termination condition or the number of generations has been reached. If so, terminate the evolution and output the best individual as the optimal solution. If not, perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population, and repeat the fitness calculation.

[0063] Thirdly, embodiments of this disclosure also provide an electronic device, the electronic device comprising:

[0064] Memory, which stores executable instructions;

[0065] A processor that executes the executable instructions in the memory to implement the LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism.

[0066] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism.

[0067] Its beneficial effects are as follows:

[0068] (1) Using features selected based on the XGBoost algorithm as input features for training the diagnostic model is beneficial to improving the accuracy of wind turbine blade fault diagnosis compared to the original features. It reduces the shortcomings of manual feature selection and enhances the diagnostic performance and generalization ability of the model.

[0069] (2) In view of the problem that convolutional neural networks (CNNs) are relatively insensitive to temporal information, and that traditional recurrent neural networks (RNNs) are prone to gradient vanishing and gradient explosion when modeling long-distance dependencies, this invention adopts a long short-term memory (LSTM) neural network model, which can better capture long-range dependencies and improve the model's learning ability and generalization performance.

[0070] (3) To address the problem of numerous parameters in deep neural network models and the low efficiency of manual tuning, this invention proposes using differential evolution to intelligently optimize the hyperparameters of the attention mechanism LSTM neural network. Differential evolution has global search characteristics, which can extensively search the parameter space to find better combinations of hyperparameters, helping to avoid getting trapped in local optima. This optimizes the learning ability of the attention mechanism LSTM network structure and improves the accuracy of fault prediction, which is highly efficient for complex neural network structures with multiple hyperparameters that need to be adjusted.

[0071] (4) Introducing an attention mechanism into the model can enable the model to effectively “pay attention” to target information, quickly find valuable information, improve prediction speed and reduce the cost of wind turbine blade fault diagnosis.

[0072] (5) Given that this method has been successfully applied to the field of wind turbine blade fault prediction, it can be applied to other wind turbines and other fault prediction problems to solve a wider range of fault diagnosis problems.

[0073] The methods and apparatus of the present invention have other features and advantages that will be apparent from or will be set forth in detail in the accompanying drawings and following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0074] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts.

[0075] Figure 1 A flowchart illustrating the steps of a wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism LSTM according to an embodiment of the present invention is shown.

[0076] Figure 2 The diagram shows the overall structure of an LSTM model based on differential evolution optimization attention mechanism according to an embodiment of the present invention.

[0077] Figure 3 A flowchart illustrating the training and testing process of an attention mechanism LSTM model according to an embodiment of the present invention is shown.

[0078] Figure 4 A schematic diagram of the training and testing curves of the model detection results according to an embodiment of the present invention is shown.

[0079] Figure 5 A block diagram of an attention mechanism LSTM wind turbine blade fault diagnosis device based on differential evolution optimization according to an embodiment of the present invention is shown.

[0080] Explanation of reference numerals in the attached figures:

[0081] 201. Data Acquisition Module; 202. Training Module; 203. Attention Mechanism Module; 204. Output Module; 205. Differential Evolution Module; 206. Diagnosis Module. Detailed Implementation

[0082] Preferred embodiments of the invention will now be described in more detail. While preferred embodiments of the invention are described below, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0083] To facilitate understanding of the solutions and effects of the embodiments of the present invention, six specific application examples are given below. Those skilled in the art should understand that these examples are merely for the purpose of understanding the present invention, and any specific details therein are not intended to limit the present invention in any way.

[0084] Example 1

[0085] Figure 1 A flowchart illustrating the steps of a wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism LSTM according to an embodiment of the present invention is shown.

[0086] like Figure 1 As shown, the LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism includes:

[0087] Step 101: Collect SCADA data and determine the feature dataset;

[0088] Step 102: Train the LSTM neural network using the feature dataset to obtain the output state;

[0089] Step 103: Input the output state into the attention mechanism model to obtain the output result;

[0090] Step 104: Obtain the prediction result from the output module, compare the prediction result with the actual state of the wind turbine, and calculate the output error.

[0091] Step 105: Perform differential evolution calculation based on the output error to update the parameters of the LSTM neural network and attention mechanism network model;

[0092] Step 106: Input the wind turbine data to be tested into the trained model to obtain the diagnostic results of wind turbine blade failure.

[0093] In one example, SCADA data is collected, and the characteristic dataset is determined to include:

[0094] The time-series data of the raw wind turbine and blade status collected from the SCADA system is called SCADA data, which includes performance data and external environmental data.

[0095] The XGBoost algorithm was used to perform feature correlation analysis on SCADA data to obtain a feature dataset with high correlation to wind turbine blade failure.

[0096] In one example, the output state is:

[0097] h t =o t ·d t

[0098] Among them, h t For output state, o t It is the output information of the output gate, d t =tanh(C t ), where tanh is the activation function.

[0099] In one example, the output state is input into the attention mechanism model, and the output results include:

[0100] Calculate the degree of matching between the input data of the LSTM at time k and the output state of the LSTM at time t;

[0101] The element weights are calculated based on the degree of matching, and then the corresponding output results are calculated based on the output state.

[0102] In one example, the output is:

[0103]

[0104] Among them, Y t The output of the attention mechanism model is given by k,t∈[1,n_step], where a k,t Indicate u k,t The probability distribution at different times, where n_step is the time step of the input data to the LSTM network.

[0105] In one example, the output error is:

[0106]

[0107] In the formula, L represents the output error, N represents the number of samples, and y (i) y′ represents the true sample class. (i) This indicates the predicted sample category of the model.

[0108] In one example, differential evolution computation includes:

[0109] The output error is used as the fitness function, and the control parameters of the differential evolution algorithm are determined.

[0110] Randomly generate the initial population;

[0111] Calculate the fitness of each individual in the population;

[0112] Determine whether the termination condition or the number of generations has been reached. If so, terminate the evolution and output the best individual as the optimal solution. If not, perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population, and repeat the fitness calculation.

[0113] Specifically, time-series data of the original wind turbine generator and blade status are collected from the SCADA system, including performance data and external environmental data. The XGBoost algorithm is used to perform feature correlation analysis on the original wind turbine time-series data to obtain a feature dataset highly correlated with wind turbine blade failures. Preprocessing operations, including data labeling, data resampling, and data normalization, are performed on the aforementioned SCADA data to divide it into training and test sets. The SCADA data preprocessing includes: labeling the SCADA data according to blade status data; resampling the data by re-dividing the data, resampling the original dataset every minute, and taking the average value of each group of samples as the new sample; and normalization to eliminate the influence of different feature units and improve the model's learning efficiency.

[0114] The training set is input into an LSTM neural network for training, learning the changing trends of multiple features at consecutive time points to obtain the temporal relationships between features. The LSTM neural network module consists of three LSTM layers and one Dropout layer. The basic LSTM unit has three special gating units: the forget gate, the input gate, and the output gate. The most important part of the LSTM network is the basic unit state C. t Furthermore, LSTM can control three gating units to determine whether external data features are stored in this basic unit. The internal operation process of the three gating units of LSTM is as follows:

[0115]

[0116] In the formula, i t f t and o t These are the output information of the input gate, forget gate, and output gate. All three control variables are related to the input time series X. t It is related to the output of the previous time step; i t Its function is to selectively record new feature information into the LSTM unit; f t Its function is to selectively forget some information from LSTM cells; t Its function is to selectively output relevant feature information; the activation function σ is the sigmoid function; W i W f and W o These are the input time series data X corresponding to the input gate, forget gate, and output gate, respectively. t and the output h of the LSTM at the previous time step t-1 The weight matrix; b i b f b o These are the bias vectors inside these three gating units.

[0117] The initial state a of the input data after processing by the activation function t As shown in the following formula:

[0118]

[0119] The core of the basic unit of LSTM is C t It can be calculated using the following formula:

[0120] C t =m t +n t

[0121] Where, the intermediate variable m t =i t ·a t n t =f t ·C t-1 i t with f t The value is between 0 and 1, C t-1 It is the state of the LSTM at the previous time step.

[0122] The basic cell state C of LSTM t After processing by the activation function and the forget gate, the final output state h is obtained. t The processes are as follows:

[0123] d t =tanh(C t )

[0124] h t =o t ·d t

[0125] Where tanh is the activation function, and the output gate state is θ. t The value is between 0 and 1.

[0126] The output state is input into the attention mechanism module to obtain more target information related to blade faults, and the result is output through the fully connected layer of the attention mechanism module; the attention mechanism module includes 1 Flatten layer, 1 Attention layer, 1 Dropout layer and 1 fully connected layer.

[0127] Calculate the degree of matching between the input of the LSTM at time k and the output of the LSTM at time t:

[0128] u k,t =V T ×tanh(W a ×h t +b a )

[0129] Among them, V T W a This is the weight matrix of the attention mechanism; T represents the transpose; represents the bias vector; tanh is the activation function; h t It is the output of the LSTM; u k,t h represents the input of the LSTM at time k and the output of the LSTM at time t. t The score is based on the degree of matching.

[0130] Next, the attention score is numerically transformed using the Softmax function. This normalizes the score, yielding a probability distribution where the sum of all coefficients is 1. Furthermore, the Softmax function's properties highlight the weights of the most important elements.

[0131]

[0132] Among them, a k,t Indicate u k,t Probability distribution at different times, a k,t ∈[1,n_step]. n_step is the time step of the input data to the LSTM network.

[0133] The output feature information of the LSTM is weighted and summed according to the weight coefficients to obtain the output result:

[0134]

[0135] Among them, Y t is the feature vector of the attention mechanism model; k,t∈[1,n_step].

[0136] Understandably, attention mechanisms can quickly select valuable information from a large amount of data and focus on this important information, while ignoring unimportant information, and then quickly classify the target based on this important information. When processing long-term series data, if the data length is too long, the LSTM model may lose important features in the data. Using attention mechanisms can compensate for the shortcomings caused by feature loss due to excessive data length.

[0137] The output results are passed through the output module to obtain the prediction results. The output error is calculated by comparing the model's output results with the actual state of the wind turbine. The output module contains a Sigmoid function, which yields a result between (0,1). A threshold of 0.5 is set. If the result is less than 0.5, it represents a negative sample, and the output is 0; if the result is greater than 0.5, it represents a positive sample, and the output is 1. The output error between the model's output results and the actual values ​​is expressed by the following formula:

[0138]

[0139] In the formula, N represents the number of samples, y (i) y′ represents the true sample class. (i) This indicates the predicted sample category of the model.

[0140] The differential evolution algorithm is used to continuously update the parameters of the LSTM combined with the attention mechanism network model in each iteration, optimize the parameters, and obtain the optimal parameter combination. Differential evolution includes:

[0141] Step 7.1: Determine the control parameters of the differential evolution algorithm and determine the fitness function as the output error;

[0142] Step 7.2: Randomly generate the initial population;

[0143] Step 7.3: Calculate the fitness of each individual in the population;

[0144] Step 7.4: Determine whether the termination condition has been met or the number of generations has reached the preset upper limit. If yes, terminate the evolution and output the best individual as the optimal solution; otherwise, continue to step 7.5.

[0145] Step 7.5: Perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population.

[0146] Step 7.6, increment the generation number by one, and return to step 7.3.

[0147] Load the trained optimal parameter model, input the wind turbine data to be tested into the model, and obtain the diagnostic results of wind turbine blade failure.

[0148] Example 2

[0149] The present invention also provides a fault diagnosis device for wind turbine blades based on differential evolution optimization attention mechanism LSTM, comprising:

[0150] The acquisition module collects SCADA data and determines the feature dataset;

[0151] The training module trains the LSTM neural network using the feature dataset to obtain the output state;

[0152] The attention mechanism module takes the output state as input to the attention mechanism model and obtains the output result.

[0153] The output module outputs the results to obtain the prediction results, compares the prediction results with the actual state of the wind turbine, and calculates the output error.

[0154] The differential evolution module performs differential evolution calculations based on the output error to update the parameters of the LSTM neural network and the attention mechanism network model.

[0155] The diagnostic module inputs the data of the wind turbine to be tested into the trained model to obtain the diagnostic results of wind turbine blade failure.

[0156] In one example, SCADA data is collected, and the characteristic dataset is determined to include:

[0157] The time-series data of the raw wind turbine and blade status collected from the SCADA system is called SCADA data, which includes performance data and external environmental data.

[0158] The XGBoost algorithm was used to perform feature correlation analysis on SCADA data to obtain a feature dataset with high correlation to wind turbine blade failure.

[0159] In one example, the output state is:

[0160] h t =o t ·d t

[0161] Among them, h t For output state, o t It is the output information of the output gate, d t =tanh(C t ), where tanh is the activation function.

[0162] In one example, the output state is input into the attention mechanism model, and the output results include:

[0163] Calculate the degree of matching between the input data of the LSTM at time k and the output state of the LSTM at time t;

[0164] The element weights are calculated based on the degree of matching, and then the corresponding output results are calculated based on the output state.

[0165] In one example, the output is:

[0166]

[0167] Among them, Y t The output of the attention mechanism model is given by k,t∈[1,n_step], where a k,t Indicate u k,t The probability distribution at different times, where n_step is the time step of the input data to the LSTM network.

[0168] In one example, the output error is:

[0169]

[0170] In the formula, L represents the output error, N represents the number of samples, and y (i) y represents the true sample category. ′(i) This indicates the predicted sample category of the model.

[0171] In one example, differential evolution computation includes:

[0172] The output error is used as the fitness function, and the control parameters of the differential evolution algorithm are determined.

[0173] Randomly generate the initial population;

[0174] Calculate the fitness of each individual in the population;

[0175] Determine whether the termination condition or the number of generations has been reached. If so, terminate the evolution and output the best individual as the optimal solution. If not, perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population, and repeat the fitness calculation.

[0176] Specifically, time-series data of the original wind turbine generator and blade status are collected from the SCADA system, including performance data and external environmental data. The XGBoost algorithm is used to perform feature correlation analysis on the original wind turbine time-series data to obtain a feature dataset highly correlated with wind turbine blade failures. Preprocessing operations, including data labeling, data resampling, and data normalization, are performed on the aforementioned SCADA data to divide it into training and test sets. The SCADA data preprocessing includes: labeling the SCADA data according to blade status data; resampling the data by re-dividing the data, resampling the original dataset every minute, and taking the average value of each group of samples as the new sample; and normalization to eliminate the influence of different feature units and improve the model's learning efficiency.

[0177] The training set is input into an LSTM neural network for training, learning the changing trends of multiple features at consecutive time points to obtain the temporal relationships between features. The LSTM neural network module consists of three LSTM layers and one Dropout layer. The basic LSTM unit has three special gating units: the forget gate, the input gate, and the output gate. The most important part of the LSTM network is the basic unit state C. t Furthermore, LSTM can control three gating units to determine whether external data features are stored in this basic unit. The internal operation process of the three gating units of LSTM is as follows:

[0178]

[0179] In the formula, it f t and o t These are the output information of the input gate, forget gate, and output gate. All three control variables are related to the input time series X. t It is related to the output of the previous time step; i t Its function is to selectively record new feature information into the LSTM unit; f t Its function is to selectively forget some information from LSTM cells; t Its function is to selectively output relevant feature information; the activation function σ is the sigmoid function; W i W f and W o These are the input time series data X corresponding to the input gate, forget gate, and output gate, respectively. t and the output h of the LSTM at the previous time step t-1 The weight matrix; b i b f b o These are the bias vectors inside these three gating units.

[0180] The initial state a of the input data after processing by the activation function t As shown in the following formula:

[0181]

[0182] The core of the basic unit of LSTM is C t It can be calculated using the following formula:

[0183] C t =m t +n t

[0184] Where, the intermediate variable m t =i t ·a t n t =f t ·C t-1 i t with f t The value is between 0 and 1, C t-1 It is the state of the LSTM at the previous time step.

[0185] The basic cell state C of LSTM t After processing by the activation function and the forget gate, the final output state h is obtained. t The processes are as follows:

[0186] d t =tanh(C t )

[0187] h t =o t ·d t

[0188] Where tanh is the activation function, and the output gate state is θ. t The value is between 0 and 1.

[0189] The output state is input into the attention mechanism module to obtain more target information related to blade faults, and the result is output through the fully connected layer of the attention mechanism module; the attention mechanism module includes 1 Flatten layer, 1 Attention layer, 1 Dropout layer and 1 fully connected layer.

[0190] Calculate the degree of matching between the input of the LSTM at time k and the output of the LSTM at time t:

[0191] u k,t =V T ×tanh(W a ×h t +b a )

[0192] Among them, V T W a This is the weight matrix of the attention mechanism; T represents the transpose; represents the bias vector; tanh is the activation function; h t It is the output of the LSTM; u k,t h represents the input of the LSTM at time k and the output of the LSTM at time t. t The score is based on the degree of matching.

[0193] Next, the attention score is numerically transformed using the Softmax function. This normalizes the score, yielding a probability distribution where the sum of all coefficients is 1. Furthermore, the Softmax function's properties highlight the weights of the most important elements.

[0194]

[0195] Among them, a k,t Indicate u k,t Probability distribution at different times, a k,t ∈[1,n_step]. n_step is the time step of the input data to the LSTM network.

[0196] The output feature information of the LSTM is weighted and summed according to the weight coefficients to obtain the output result:

[0197]

[0198] Among them, Yt is the feature vector of the attention mechanism model; k,t∈[1,n_step].

[0199] Understandably, attention mechanisms can quickly select valuable information from a large amount of data and focus on this important information, while ignoring unimportant information, and then quickly classify the target based on this important information. When processing long-term series data, if the data length is too long, the LSTM model may lose important features in the data. Using attention mechanisms can compensate for the shortcomings caused by feature loss due to excessive data length.

[0200] The output results are passed through the output module to obtain the prediction results. The output error is calculated by comparing the model's output results with the actual state of the wind turbine. The output module contains a Sigmoid function, which yields a result between (0,1). A threshold of 0.5 is set. If the result is less than 0.5, it represents a negative sample, and the output is 0; if the result is greater than 0.5, it represents a positive sample, and the output is 1. The output error between the model's output results and the actual values ​​is expressed by the following formula:

[0201]

[0202] In the formula, N represents the number of samples, y (i) y represents the true sample category. ′(i) This indicates the predicted sample category of the model.

[0203] The differential evolution algorithm is used to continuously update the parameters of the LSTM combined with the attention mechanism network model in each iteration, optimize the parameters, and obtain the optimal parameter combination. Differential evolution includes:

[0204] Step 7.1: Determine the control parameters of the differential evolution algorithm and determine the fitness function as the output error;

[0205] Step 7.2: Randomly generate the initial population;

[0206] Step 7.3: Calculate the fitness of each individual in the population;

[0207] Step 7.4: Determine whether the termination condition has been met or the number of generations has reached the preset upper limit. If yes, terminate the evolution and output the best individual as the optimal solution; otherwise, continue to step 7.5.

[0208] Step 7.5: Perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population.

[0209] Step 7.6, increment the generation number by one, and return to step 7.3.

[0210] Load the trained optimal parameter model, input the wind turbine data to be tested into the model, and obtain the diagnostic results of wind turbine blade failure.

[0211] Example 3

[0212] The processor used in the experiments of this invention embodiment is i7-9750H, the operating system is Windows 10, the programming language is Python 3.10.14, and the deep learning framework is PyTorch.

[0213] Figure 2 The diagram shows the overall structure of an LSTM model based on differential evolution optimization attention mechanism according to an embodiment of the present invention.

[0214] like Figure 2 As shown, SCADA data and blade status data (normal blade time period and blade failure time period) of wind turbine generators are obtained.

[0215] The SCADA dataset used in this invention is sourced from a wind turbine SCADA dataset provided by a domestic company. Each data entry contains 26 variable features, including wind turbine operating parameters, environmental parameters, and state parameters: wind speed, generator speed, grid-side active power, wind angle, wind direction angle, yaw position, yaw speed, blade 1 angular velocity, blade 2 angular velocity, blade 3 angular velocity, blade 1 velocity, blade 2 velocity, blade 3 velocity, pitch motor 1 temperature, pitch motor 2 temperature, pitch motor 3 temperature, horizontal acceleration, vertical acceleration, ambient temperature, nacelle temperature, pitch control cabinet power supply 1 temperature, pitch control cabinet power supply 2 temperature, pitch control cabinet power supply 3 temperature, pitch control cabinet power supply 1 DC current, pitch control cabinet power supply 2 DC current, and pitch control cabinet power supply 3 DC current. The SCADA system sampling interval is 6 seconds. Furthermore, all monitored variable data have been manually encrypted, differing from the actual values ​​and losing their original physical meaning. The data volume is 384,000 entries, including 360,000 normal data entries and 24,000 fault data entries.

[0216] The XGBoost algorithm was used to perform feature correlation analysis on the original wind turbine time-series data, resulting in a feature dataset highly correlated with wind turbine blade failures. XGBoost can be used for feature importance selection; when constructing the augmentation tree design, feature scores are obtained to indicate the importance of each feature to the training model. The more a feature is used in the key decisions of the augmentation tree, the higher its score. The 26 features of the SCADA dataset were used as input to the XGBoost algorithm to obtain the feature importance of the data, which is the number of times the feature splits into tree nodes. The built-in feature importance calculation method, Gain, was used in XGBoost; at each split, a greedy algorithm was used to select the feature with the largest information gain as the split point.

[0217] After sorting the features by importance, the top 9 features out of the 26 are shown in Table 1 below:

[0218] Table 1 shows the calculation of feature importance using XGBoost.

[0219]

[0220] First, the SCADA data is labeled according to the blade status data: blade failure is a positive sample, labeled 1, and normal blade is a negative sample, labeled 0, thus removing invalid data. Then, resampling is performed, the data is re-divided, and the original dataset is resampled every minute. The average value of each group of samples is used as the new sample, reducing the amount of data and improving computational efficiency. Since the wind turbine has a three-blade structure, the blade angular velocity, turbine blade velocity, pitch motor temperature, pitch control cabinet power supply temperature, and pitch control cabinet power supply DC current data of the three blades in each sample are averaged.

[0221] Data normalization is performed to eliminate the influence of different feature units, improving the model's learning efficiency and resulting in a preprocessed time-series dataset. Following a 70:30 ratio, the first part is used as the training set, and the second part as the test set. Specifically, the first part, consisting of 30,720 original 26-dimensional features, is used for model training; the second part, consisting of 7,680 26-dimensional features, is used for model reliability verification.

[0222] Figure 3 A flowchart illustrating the training and testing process of an attention mechanism LSTM model according to an embodiment of the present invention is shown.

[0223] like Figure 3 As shown, the test set of the selected 9-dimensional feature dataset is input into the LSTM neural network for training to learn the changing trends of multiple features at consecutive time points and obtain the temporal relationship between features.

[0224] The data passes through the forget gate of the LSTM, which determines whether the information in the memory cell at the previous time step needs to be discarded or retained; the data passes through the input gate of the LSTM to obtain the candidate memory cell at the current time step, and the memory cell is updated based on the memory cell at the previous time step and the candidate memory cell at the current time step; the data passes through the output gate of the LSTM to obtain the final output.

[0225] The output of the LSTM neural network is input into the attention mechanism module to obtain more target information related to blade faults, and the output is output through the fully connected layer of the attention mechanism module.

[0226] The output results are used to obtain the prediction results through the output module. The output results of the model are compared with the actual state of the wind turbine to calculate the output error.

[0227] Based on the output error, a differential evolution algorithm is used to continuously update the parameters of the LSTM combined with the attention mechanism network model in each iteration, optimizing the parameters to obtain the optimal parameter combination. After optimization by the differential evolution algorithm, the attention_size parameter is 256, and the n_step size is 96.

[0228] The trained model is loaded, and then the wind turbine data to be tested is input into the model to obtain the diagnostic results of wind turbine blade faults. The model's wind turbine blade fault diagnosis performance is evaluated using a test set.

[0229] Figure 4 A schematic diagram of the training and testing curves of the model detection results according to an embodiment of the present invention is shown.

[0230] The training and testing curves of the wind turbine blade fault diagnosis method model based on a long short-term memory neural network with a differential evolution optimization attention mechanism proposed in this invention are shown below. Figure 4 As shown, the accuracy of the training and test sets tends to a stable value of 0.98 after 150 iterations, indicating that the method is accurate and effective in diagnosing wind turbine blade faults.

[0231] We used an LSTMAM (Attention Mechanism Long Short-Term Memory Neural Network) model based on differential evolution optimization for wind turbine blade fault diagnosis, and compared it with traditional RNN (Recurrent Neural Network), SVM (Support Vector Machine) and LSTM models.

[0232] Table 2 Model Comparison

[0233]

[0234] The comparison results show that recurrent neural networks and support vector machines have lower accuracy than LSTM and LSTMAM, making them unsuitable as models for wind turbine blade fault diagnosis. Meanwhile, the combination of LSTM with an attention mechanism improves the model's feature learning ability and convergence speed, reduces computation time, and ultimately enhances the accuracy of wind turbine blade fault diagnosis.

[0235] Example 4

[0236] Figure 5 A block diagram of an attention mechanism LSTM wind turbine blade fault diagnosis device based on differential evolution optimization according to an embodiment of the present invention is shown.

[0237] like Figure 5 As shown, the LSTM wind turbine blade fault diagnosis device based on differential evolution optimization attention mechanism includes:

[0238] Acquisition module 201 acquires SCADA data and determines the feature dataset;

[0239] Training module 202 trains the LSTM neural network using the feature dataset to obtain the output state;

[0240] Attention mechanism module 203 inputs the output state into the attention mechanism model to obtain the output result;

[0241] Output module 204 outputs the results to obtain the prediction results, compares the prediction results with the actual state of the wind turbine, and calculates the output error.

[0242] Differential evolution module 205 performs differential evolution calculations based on the output error to update the parameters of the LSTM neural network and attention mechanism network model;

[0243] The diagnostic module 206 inputs the wind turbine data to be tested into the trained model to obtain the diagnostic results of wind turbine blade failure.

[0244] In one example, SCADA data is collected, and the characteristic dataset is determined to include:

[0245] The time-series data of the raw wind turbine and blade status collected from the SCADA system is called SCADA data, which includes performance data and external environmental data.

[0246] The XGBoost algorithm was used to perform feature correlation analysis on SCADA data to obtain a feature dataset with high correlation to wind turbine blade failure.

[0247] In one example, the output state is:

[0248] h t =o t ·d t

[0249] Among them, h t For output state, o t It is the output information of the output gate, d t =tanh(C t ), where tanh is the activation function.

[0250] In one example, the output state is input into the attention mechanism model, and the output results include:

[0251] Calculate the degree of matching between the input data of the LSTM at time k and the output state of the LSTM at time t;

[0252] The element weights are calculated based on the degree of matching, and then the corresponding output results are calculated based on the output state.

[0253] In one example, the output is:

[0254]

[0255] Among them, Y t The output of the attention mechanism model is given by k,t∈[1,n_step], where a k,t Indicate u k,t The probability distribution at different times, where n_step is the time step of the input data to the LSTM network.

[0256] In one example, the output error is:

[0257]

[0258] In the formula, L represents the output error, N represents the number of samples, and y (i) y represents the true sample category. ′(i) This indicates the predicted sample category of the model.

[0259] In one example, differential evolution computation includes:

[0260] The output error is used as the fitness function, and the control parameters of the differential evolution algorithm are determined.

[0261] Randomly generate the initial population;

[0262] Calculate the fitness of each individual in the population;

[0263] Determine whether the termination condition or the number of generations has been reached. If so, terminate the evolution and output the best individual as the optimal solution. If not, perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population, and repeat the fitness calculation.

[0264] Example 5

[0265] This disclosure provides an electronic device, comprising: a memory storing executable instructions; and a processor executing the executable instructions in the memory to implement the aforementioned LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism.

[0266] An electronic device according to an embodiment of the present disclosure includes a memory and a processor.

[0267] This memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0268] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory.

[0269] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0270] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0271] Example 6

[0272] This disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism.

[0273] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the methods described in the foregoing embodiments of the present disclosure are performed.

[0274] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0275] Those skilled in the art should understand that the above description of the embodiments of the present invention is only intended to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any of the examples given.

[0276] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism LSTM, characterized in that, include: Collect SCADA data and determine the feature dataset; The LSTM neural network is trained using the aforementioned feature dataset to obtain the output state; The output state is input into the attention mechanism model to obtain the output result; The output results are passed through the output module to obtain the prediction results. The prediction results are compared with the actual state of the wind turbine to calculate the output error. Differential evolution calculation is performed based on the output error to update the parameters of the LSTM neural network and attention mechanism network model; The data of the wind turbine to be tested is input into the trained model to obtain the diagnostic results of wind turbine blade failure.

2. The differential evolution optimization based attention mechanism LSTM wind turbine blade fault diagnosis method of claim 1, wherein, Collect SCADA data and determine the feature dataset, including: The SCADA data is the time-series data of the raw wind turbine and blade status collected from the SCADA system, which includes performance data and external environmental data. The XGBoost algorithm is used to perform feature correlation analysis on the SCADA data to obtain the feature dataset that has a high correlation with wind turbine blade failure.

3. The differential evolution optimization based attention mechanism LSTM wind turbine blade fault diagnosis method of claim 1, wherein, The output state is: h t = o t · d t where h t is the output state, o t is the output information of the output gate, d t = tanh(C t ), tanh is an activation function.

4. The differential evolution optimization based attention mechanism LSTM wind turbine blade fault diagnosis method of claim 1, wherein, The output state is input into the attention mechanism model to obtain the following output results: Calculate the degree of matching between the input data of the LSTM at time k and the output state of the LSTM at time t; The element weights are calculated based on the matching degree, and then the corresponding output results are calculated based on the output state.

5. The LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism according to claim 4, wherein, The output result is as follows: wherein Y t is the output result of the attention mechanism model, k, t ∈ [1, n_step], a k,t represents u k,t the probability distribution at different time, and n_step is the time step of the input data to the LSTM network.

6. The LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism according to claim 1, wherein, The output error is: where L is the output error, N represents the number of samples, y (i) represents the true sample class, y ′(i) represents the predicted sample class of the model.

7. The differential evolution optimization based attention mechanism LSTM wind turbine blade fault diagnosis method of claim 1, wherein, The differential evolution calculation includes: The output error is used as the fitness function to determine the control parameters of the differential evolution algorithm; Randomly generate the initial population; Calculate the fitness of each individual in the population; Determine whether the termination condition or the number of generations has been reached. If so, terminate the evolution and output the best individual as the optimal solution. If not, perform mutation and crossover to generate an intermediate population. Select individuals from the original population and the intermediate population to obtain a new generation population, and repeat the fitness calculation.

8. A differential evolution optimization-based attention mechanism LSTM fan blade fault diagnosis device, characterized in that, include: The acquisition module collects SCADA data and determines the feature dataset; The training module trains the LSTM neural network using the feature dataset to obtain the output state; The attention mechanism module inputs the output state into the attention mechanism model to obtain the output result; The output module obtains the prediction result from the output result, and calculates the output error by comparing the prediction result with the actual state of the wind turbine. The differential evolution module performs differential evolution calculations based on the output error to update the parameters of the LSTM neural network and the attention mechanism network model. The diagnostic module inputs the data of the wind turbine to be tested into the trained model to obtain the diagnostic results of wind turbine blade failure.

9. An electronic device, comprising: The electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the LSTM wind turbine blade fault diagnosis method based on differential evolution optimization as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the LSTM wind turbine blade fault diagnosis method based on differential evolution optimization attention mechanism as described in any one of claims 1-7.

Images