Fan blade icing fault diagnosis method based on GWO-REKELM
By improving the gray wolf optimized kernel extreme learning machine model, combining random forest and oversampling techniques, and combining the error count and root mean square error objective function, the accuracy of wind turbine blade icing fault diagnosis is improved, and the problems of local minima convergence and low diagnosis error rate in traditional methods are solved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI DIANJI UNIV
- Filing Date
- 2023-04-23
- Publication Date
- 2026-05-19
AI Technical Summary
In existing technologies, in wind turbine blade icing fault diagnosis methods based on SCADA data, the objective functions of traditional optimization algorithms, such as the number of diagnostic errors and root mean square error, suffer from local minima convergence and low diagnostic accuracy when optimizing model parameters.
An improved gray wolf optimized kernel extreme learning machine model is adopted. By combining the error count and root mean square error of the test results as the objective function, and combining random forest feature selection and oversampling techniques, the wind turbine operation data is optimized to improve the diagnostic accuracy.
It improves the accuracy of diagnosing wind turbine blade icing faults and solves the problems of diagnostic errors caused by the number of diagnostic errors easily converging to local minima and the decrease in RMSE value in traditional methods.
Smart Images
Figure CN116467664B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for diagnosing icing faults in wind turbine blades, and more particularly to a method for diagnosing icing faults in wind turbine blades based on GWO-REKELM. Background Technology
[0002] Fault diagnosis methods for wind turbine blade icing typically employ fault prediction methods based on SCADA (Supervisory Control and Data Acquisition) data. Generally, the dataset undergoes preprocessing; to address imbalance issues, oversampling techniques are used to balance the dataset. Then, machine learning or deep learning models are employed for fault diagnosis, with model parameters often optimized using optimization algorithms. The objective function used in these algorithms to optimize model parameters is typically the root mean square error of the test results or the number of diagnostic errors. However, using traditional objective functions in the optimization process results in low diagnostic accuracy.
[0003] When optimization algorithms optimize model parameters, the objective function for particles is either the diagnostic error count (ERROR) or the root mean square error (RMSE) of the test results. Both objective functions have problems:
[0004] 1) When optimizing model parameters by using the number of diagnostic errors as the objective function of the particle, the particle can be regarded as a continuously decreasing extreme value as the ERROR value decreases. However, there may be a phenomenon that the objective function value of the particle cannot find a better direction after decreasing to a certain minimum value, which will cause the particle to converge to a local minimum value.
[0005] 2) When using the root mean square error (RMSE) of the test results as the objective function to optimize model parameters, there is a problem where the RMSE value decreases while the ERROR value increases. Assume the true label distribution of two samples is 1, 0, meaning the first is a positive example and the second is a negative example. The rule is that a result greater than 0.5 is a positive example, and less than 0.5 is a negative example. In the current round of prediction, the results for the two samples are 0.55 and 0.45, indicating both samples are correctly judged. In the next round of prediction, the prediction results for the two samples are 0.99 and 0.55, indicating positive and negative results respectively. Although the overall RMSE value of the samples decreases, the second sample may be predicted incorrectly, leading to a decrease in diagnostic accuracy. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a wind turbine blade icing fault diagnosis method based on GWO-REKELM. The objective function of particle optimization in the gray wolf optimized kernel extreme learning machine model has been improved to improve the accuracy of wind turbine blade icing fault diagnosis.
[0007] The objective of this invention can be achieved through the following technical solutions:
[0008] A method for diagnosing wind turbine blade icing faults based on GWO-REKELM, the method comprising the following steps:
[0009] 1) Preprocess and normalize all wind turbine operating data;
[0010] 2) Use random forest to perform feature filtering on all wind turbine operation data;
[0011] 3) Oversampling techniques are used to balance the dataset after feature filtering of the wind turbine operation data;
[0012] 4) The dataset was trained and tested using a gray wolf optimized kernel extreme learning machine model with an improved objective function to complete the diagnosis of wind turbine blade icing faults.
[0013] Furthermore, the preprocessing process in 1) is as follows: First, all wind turbine operating data are adjusted and groups are deleted; then, samples within the same timestamp are averaged and samples with power greater than 2KW are deleted; finally, the wind turbine blade pitch angle, wind turbine blade pitch speed, and wind turbine blade pitch motor temperature data are averaged.
[0014] Furthermore, 2) specifically refers to: using random forest to filter features of all wind turbine operation data to obtain the features of the top eight with the highest scores.
[0015] Furthermore, in step 3), half of the dataset is divided into a training set and the other half into a test set.
[0016] Furthermore, the gray wolf optimization specifically refers to the following: In the gray wolf optimization algorithm, each gray wolf represents a candidate solution in the population. The optimal solution in the population is called α, the second best solution is called β, the third best solution is called δ, and the remaining candidate solutions are collectively called ω. In each iteration of the gray wolf optimization algorithm, α, β, and δ are used as the top three best solutions to determine the prey's position and direct ω to be randomly updated around the prey, continuously approaching the prey.
[0017] Furthermore, the nuclear extreme learning machine specifically refers to:
[0018]
[0019] Where f(x) is the output of the test set, β is the connection weight between the hidden layer and the output layer, h(x) is the output of the hidden layer neurons, T is the expected output of the training set, I is the identity matrix, C is the regularization coefficient, N is the number of samples in the training set, Ω is the kernel matrix, and K(x) is the output of the hidden layer neurons. i ,x j ) represents the kernel function.
[0020] Furthermore, the kernel matrix is a Gaussian kernel, and the kernel function formula is as follows:
[0021] Ω ij =K(x) i ,x j )=K(x j ,x i )=exp(-||x i -x j || 2 / g)
[0022] Where i and j represent the elements in the i-th row and j-th column of the kernel matrix Ω, respectively, and g is the kernel parameter, g > 0.
[0023] Furthermore, the improved objective function is specifically defined as follows: during the optimization process of the particle pair model parameters, the error count and root mean square error of the test results are combined, and the objective function is set as follows:
[0024] 0.001*RMSE+ERROR
[0025] Where ERROR is the number of errors in the test results, and RMSE is the root mean square error of the test results.
[0026] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.
[0027] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] I. This invention improves the objective function of the gray wolf optimized kernel extreme learning machine model by combining the error count and root mean square error of the test results, thereby solving the problem of limited diagnostic accuracy of the traditional gray wolf optimized kernel extreme learning machine model and improving the diagnostic accuracy of wind turbine blade icing faults.
[0030] Second, this invention uses SCADA data, preprocesses and normalizes the data, and then performs oversampling to achieve category balance of wind turbine operation data, thereby further improving the accuracy of diagnosis. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0033] Example
[0034] like Figure 1 As shown, a method for diagnosing icing faults in wind turbine blades based on GWO-REKELM (an improved gray wolf optimized kernel extreme learning machine model) is described, comprising the following steps:
[0035] 1) Preprocess and normalize all wind turbine operating data.
[0036] The data in this example comes from an industrial big data innovation competition. The preprocessing process is as follows: First, all wind turbine operating data are adjusted and groups are deleted; then, samples within the same time stamp are averaged and samples with power greater than 2KW are deleted; finally, the wind turbine blade pitch angle, wind turbine blade pitch speed, and wind turbine blade pitch motor temperature data are averaged.
[0037] 2) Use random forest to filter features of all wind turbine operation data to obtain the features of the top eight with the highest scores.
[0038] 3) Oversampling techniques are used to balance the wind turbine operation data after feature filtering; half of the dataset is divided into training set and half into test set.
[0039] 4) The dataset was trained and tested using a gray wolf optimized kernel extreme learning machine model with an improved objective function to complete the diagnosis of wind turbine blade icing faults.
[0040] The gray wolf optimization algorithm is as follows: In the gray wolf optimization algorithm, each gray wolf represents a candidate solution in the population. The best solution in the population is called α, the second best solution is called β, the third best solution is called δ, and the remaining candidate solutions are collectively called ω. In each iteration of the gray wolf optimization algorithm, α, β, and δ are used as the top three best solutions to determine the prey's position and direct ω to be randomly updated around the prey, continuously approaching the prey.
[0041] The formula for surrounding prey is as follows:
[0042] D = |C × X P (t)-X(t)|
[0043] X(t+1)=X P (t)-A×D
[0044] Where D represents the distance between the individual gray wolf and its prey, X(t+1) represents the updated position of the gray wolf, t represents the current iteration number, and X PX(t) represents the prey's position, X(t) represents the gray wolf's position, and A and C are parameters. The calculation formula is as follows:
[0045] A = a × (2 × r1 - 1)
[0046] C = 2 * r²
[0047] Where a is the convergence factor, which decreases linearly from 2 to 0 as the number of iterations decreases, and r1 and r2 are random numbers between [0,1].
[0048] Assuming α, β, and δ are the top three optimal solutions in history, other gray wolf individuals update their positions based on the positions of these three optimal gray wolf individuals, gradually approaching the prey.
[0049] The mathematical model formula for an individual gray wolf tracking its prey is as follows:
[0050]
[0051]
[0052] The final position of each gray wolf is determined by the direction and step length it takes towards α, β, and δ:
[0053] X(t+1)=X1(t)+X2(t)+X3(t).
[0054] The nuclear extreme learning machine is specifically:
[0055] In extreme learning machines:
[0056] f(x)=h(x)β
[0057]
[0058]
[0059] Where f(x) is the output of the test set, β is the connection weight between the hidden layer and the output layer, h(x) and H are the outputs of the hidden layer neurons, T is the expected output of the training set, I is the identity matrix, C is the regularization coefficient, N is the number of samples in the training set, L is the number of hidden layer neurons, w represents the input weight connecting the hidden layer neurons and the input layer neurons, and b represents the bias of the hidden layer neurons.
[0060] However, when h(x) is generalized to implicit mappings, its expression is unknown. Therefore, a kernel limit learning machine is introduced, whose expression is:
[0061]
[0062] Where Ω is the kernel matrix, K(x) i ,x j) represents the kernel function.
[0063] The kernel matrix is a Gaussian kernel, and the kernel function formula is as follows:
[0064] Ω ij =K(x) i ,x j )=K(x j ,x i )=exp(-||x i -x j || 2 / g)
[0065] Where i and j represent the elements in the i-th row and j-th column of the kernel matrix Ω, respectively, and g is the kernel parameter, g > 0.
[0066] The improved objective function is as follows: In the process of optimizing the particle pair model parameters, to address the problems of traditional objective functions such as ERROR easily converging to local values and RMSE decreasing while diagnostic error increases, the objective function is set by combining the error count and root mean square error of the test results:
[0067] 0.001*RMSE+ERROR
[0068] Where ERROR is the number of errors in the test results, and RMSE is the root mean square error of the test results.
[0069] Since the ERROR term is much larger than the 0.001*RMSE term, the particle prioritizes the number of errors in the test results as its optimization direction. When the number of errors cannot be reduced further, the particle will move in the direction with a smaller root mean square error to increase the probability of finding a place with a smaller number of errors, thereby improving the classification accuracy.
[0070] The GWO-REKELM model was used to diagnose wind turbine blade icing faults, and compared with traditional BP (neural network model), ELM (extreme learning machine model), KELM (kernel extreme learning machine model), GWO-RKELM (gray wolf optimized kernel extreme learning machine model with RMSE objective function) and GWO-EKELM (gray wolf optimized kernel extreme learning machine model with ERROR objective function).
[0071] To facilitate comparison between models, 400 samples were randomly selected from both the icing and normal sample groups for each experiment. Half of these samples were used as the training set, and the other half as the test set. All models used the randomly selected data simultaneously, and a total of 50 experiments were conducted, with the average value taken. Fault diagnosis experiments were performed using wind turbines No. 15 and No. 21. The specific diagnostic results for wind turbines No. 15 and No. 21 on each model are shown in Table 1.
[0072] Table 1. Fault diagnosis accuracy of wind turbines 15 and 21 on all models.
[0073]
[0074] The experimental results show that the accuracy of the Extreme Learning Machine (ELM) and neural network models is lower than that of the Kernel Extreme Learning Machine (KELM), making them unsuitable as models for diagnosing wind turbine blade icing faults. Furthermore, a comparison of models 3, 4, 5, and 6 reveals that the KELM model suffers from limited diagnostic accuracy due to the lack of particle optimization of parameters. The improved Gray Wolf Optimized Kernel Extreme Learning Machine model, after adopting an improved objective function, not only solves the problem of decreased RMSE value but also increased diagnostic errors when using RMSE as the objective function; it also addresses the issue of particles easily converging to limited values and failing to find a better direction when using ERROR as the objective function. Ultimately, the accuracy of wind turbine blade icing fault diagnosis is improved.
[0075] The present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.
[0076] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0077] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for diagnosing wind turbine blade icing faults based on GWO-REKELM, characterized in that, The method includes the following steps: 1) Preprocess and normalize all wind turbine operating data; 2) Use random forest to perform feature filtering on all wind turbine operation data; 3) Oversampling techniques are used to balance the dataset after feature filtering of the wind turbine operation data; 4) The kernel extreme learning machine model is trained and tested on the dataset using the gray wolf optimization algorithm with an improved objective function. The trained and optimized kernel extreme learning machine model is then used for wind turbine blade icing fault diagnosis. The improved objective function is specifically defined as follows: during the optimization of particle pair model parameters, the error count and root mean square error of the test results are combined, and the objective function is set as follows: Where ERROR is the number of errors in the test results, and RMSE is the root mean square error of the test results.
2. The method for diagnosing wind turbine blade icing faults based on GWO-REKELM according to claim 1, characterized in that, The preprocessing process in 1) is as follows: First, all wind turbine operating data are adjusted and groups are deleted; then, samples within the same timestamp are averaged and samples with power greater than 2KW are deleted; finally, the wind turbine blade pitch angle, wind turbine blade pitch speed, and wind turbine blade pitch motor temperature data are averaged.
3. The method for diagnosing wind turbine blade icing faults based on GWO-REKELM according to claim 1, characterized in that, Specifically, 2) involves using random forests to filter features from all wind turbine operation data, and then identifying the features of the top eight turbines with the highest scores.
4. The method for diagnosing wind turbine blade icing faults based on GWO-REKELM according to claim 1, characterized in that, In step 3), the dataset is divided into half training set and half test set.
5. The method for diagnosing wind turbine blade icing faults based on GWO-REKELM according to claim 1, characterized in that, The gray wolf optimization specifically refers to the following: In the gray wolf optimization algorithm, each gray wolf represents a candidate solution in the population, and the optimal solution in the population is called... The suboptimal solution is called The third optimal solution is called The remaining candidate solutions are collectively referred to as Each iteration of the Grey Wolf optimization algorithm will , and As one of the top three optimal solutions, determine the prey's location and command. It updates randomly around its prey, constantly closing in on it.
6. The method for diagnosing wind turbine blade icing faults based on GWO-REKELM according to claim 1, characterized in that, The nuclear extreme learning machine is specifically: in, For the output of the test set, For the connection weights between the hidden layer and the output layer, The output of the hidden layer neurons, The expected output of the training set, It is the identity matrix; The regularization coefficient is . The number of samples in the training set. For the kernel matrix, Represents the kernel function.
7. The method for diagnosing wind turbine blade icing faults based on GWO-REKELM according to claim 6, characterized in that, The kernel matrix is a Gaussian kernel, and the kernel function formula is as follows: in, i , j Representing the kernel matrix respectively The i line, number j Column elements, For kernel parameters, .
8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.