Wind turbine generator fault prediction method based on multi-strategy improved hybrid optimization algorithm

Abstract

The invention discloses a wind turbine generator fault prediction method based on a multi-strategy improved hybrid optimization algorithm, and belongs to the technical field of wind turbine generator fault diagnosis. The method comprises the following steps: collecting vibration signals of a wind turbine generator in different states; optimizing the initial population quality of a sparrow algorithm by using dynamic reverse learning, fusing an exploration strategy of a northern eagle algorithm in a search stage of the sparrow algorithm, introducing a Gaussian random walk strategy to update the position of the sparrow, obtaining an improved hybrid optimization algorithm, and determining the number k of modal decomposition and a penalty factor alpha; decomposing the vibration signal by using a modal decomposition method to obtain an IMF component; calculating to obtain a multi-dimensional feature vector; and introducing the mixed kernel function into an extreme learning machine to obtain a mixed kernel learning machine, and performing state recognition on the multi-dimensional feature vector. Under a test data set, the accuracy rate of the fault diagnosis method is as high as 100%, and in actual fault early warning classification, the accuracy rate is as high as 97.5%, and an effective solution is provided for fault early warning of the wind turbine generator.

Description

Technical Field

[0001] The invention belongs to the technical field of wind turbine fault diagnosis, and discloses a wind turbine fault prediction method based on a multi-strategy improved hybrid optimization algorithm. Background Art

[0002] With the widespread adoption of renewable energy, wind power has become one of the most important green energy sources worldwide. As crucial equipment for wind energy conversion, the operational stability and reliability of wind turbines are directly linked to the power generation efficiency and economic benefits of wind farms. Therefore, ensuring the proper operation of wind turbines and promptly identifying and addressing potential faults have become key challenges facing the wind power industry.

[0003] Currently, the development of wind turbine fault warning technology focuses primarily on key areas such as data monitoring, analysis, and prediction. With the rapid development of the Internet of Things and big data technologies, data collection on the operating status and environmental conditions of wind turbines has become increasingly convenient. By deploying a variety of sensors in wind turbines to monitor operating parameters (such as speed, vibration, and temperature) in real time, a wealth of raw data can be provided to the fault warning system. After processing and analysis, this data can help operations and maintenance personnel understand the health status of the turbine and identify potential fault hazards. However, relying solely on data monitoring is not sufficient for effective fault warning.

[0004] Currently, fault diagnosis research generally involves three steps: vibration signal processing, fault feature extraction, and feature state classification. Common signal processing methods include Fourier transform, empirical mode decomposition, ensemble empirical mode decomposition, and complementary empirical mode decomposition. However, due to the fixed nature of the window function, Fourier transform methods struggle to achieve optimal results simultaneously in time and frequency. EMD methods, on the other hand, face challenges with endpoint effects and modal mixing. Furthermore, under strong noise interference, EMD methods often perform poorly on fault signals. While variational mode decomposition (VMD) signal decomposition methods avoid the drawbacks of traditional EMD methods, in practice, it is difficult to directly determine the number of modal decompositions, k, and the penalty factor, α. These two parameters directly impact the decomposition performance of VMD. Attempts have been made to optimize parameters using intelligent algorithms such as the sparrow search algorithm (SSA) and the northern goshawk optimization algorithm (NGO). However, these algorithms suffer from problems such as susceptibility to local optima, poor generalization, and slow convergence. In terms of feature extraction, traditional methods tend to use various time-domain indicators or entropy values ​​to characterize fault characteristics, but a single indicator cannot fully characterize the characteristics of the fault signal. The essence of feature state classification is to use a large amount of data to train the network model to achieve the effect of identifying different states. In recent years, various neural networks have been widely used in feature classification tasks. Among them, the Extreme Learning Machine (ELM) is a single hidden layer feedforward neural network with fast training speed and good generalization ability. However, due to its single activation function and hidden layer node configuration, it lacks the ability to adapt to changes in data characteristics, making it difficult to apply to the complex and large-scale data analysis of wind turbines. Summary of the Invention

[0005] The purpose of the present invention is to provide a wind turbine fault prediction method based on a multi-strategy improved hybrid optimization algorithm to solve the problems of poor fault prediction accuracy caused by unsatisfactory signal processing effect, incomplete feature extraction information, and poor applicability of feature classification network models in existing methods.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A wind turbine fault prediction method based on a multi-strategy improved hybrid optimization algorithm is characterized by comprising the following steps:

[0008] (1) Collect vibration signals of wind turbines in different states;

[0009] (2) Dynamic reverse learning is used to optimize the initial population quality of the sparrow algorithm. The exploration strategy of the northern goshawk algorithm is integrated into the search phase of the sparrow algorithm. The Gaussian random walk strategy is introduced to update the sparrow position. An improved hybrid optimization algorithm is obtained and the number of modal decompositions k and the penalty factor α are determined using this hybrid optimization algorithm.

[0010] (3) Decomposing the vibration signal using the modal decomposition method in step (2) to obtain the corresponding IMF components, calculating the envelope entropy of the IMF components, and selecting the IMF component with the smallest envelope entropy value as the optimal component;

[0011] (4) Calculate the sample entropy, energy entropy, and permutation entropy of the optimal IMF component to obtain a multi-dimensional feature vector;

[0012] (5) The hybrid kernel function is introduced into the extreme learning machine to obtain a hybrid kernel learning machine, the feature set is divided into a training set and a test set, the feature set is input into the hybrid kernel learning machine, the training model is used to obtain a wind turbine fault prediction model, and the test set is input into the wind turbine fault prediction model for state recognition, thereby completing the fault diagnosis of the wind turbine vibration signal.

[0013] A further technical solution is that the specific steps of step (2) are:

[0014] (1) Assume that the population size is n:

[0015]

[0016] Where: m is the dimension of the optimization variable;

[0017] (2) Initialize the population based on the dynamic reverse learning strategy

[0018] X d =X+r1×(r2×(LB+UB-X)-X)

[0019] Where: X is the randomly generated initial population, r1 and r2 are random numbers between 0 and 1, and X d is a new population generated based on the dynamic reverse learning strategy; LB represents the lower limit and UB represents the upper limit;

[0020] (3) Introducing the exploration strategy of the Northern Goshawk optimization algorithm in the initial search phase of the Sparrow algorithm:

[0021]

[0022] P i represents the position of the i-th prey, Indicates the fitness value of the position; is the updated position, r∈[0,1], I is 1 or 2;

[0023] (4) Further position updates are as follows:

[0024]

[0025] Where: is the optimal position of the finder in the t+1th iteration, is the global worst position in the tth iteration; is the t-th position, Q is a random number that obeys the normal distribution, A + , L represents a matrix in which each element is randomly assigned a value of 1 or -1;

[0026] (5) Introducing the Gaussian random walk strategy, the new sparrow position is updated as follows:

[0027]

[0028] Where: is a random individual, t represents the number of iterations, Indicates this iteration;

[0029] (6) Assuming that the number of individuals responsible for vigilance in a population accounts for 20% of the total population, the mathematical expression for the number of individuals responsible for vigilance in a population can be expressed as follows:

[0030]

[0031] Where: is the current global optimal position, β and k are the control step coefficients, and f i is the current position fitness value, f b 、f w are the best fitness value and the worst fitness value in this iteration respectively, ε is a constant, f i ≠f w , represents the global worst position, Updated location, Indicates the position before updating;

[0032] (7) Use the above method to determine the number k of modal decomposition VMD and the penalty factor α.

[0033] A further technical solution is that the specific steps of step (3) are:

[0034] 3-1) VMD decomposition variational model is:

[0035]

[0036] By introducing the Lagrange multiplication operator and the quadratic penalty term, it becomes an unconstrained problem extension:

[0037]

[0038] In the above formula: α is the penalty factor, λ is the Lagrangian operator, u k and ω k is the center frequency, represents the frequency domain result, k is the number of decompositions, and k and α are determined using a hybrid optimization algorithm;

[0039] 3-2) Use the modal decomposition method VMD to decompose the vibration signal to obtain k IMF components, and calculate the envelope entropy of each IMF component. The IMF component corresponding to the minimum envelope entropy value is taken as the optimal component.

[0040] A further technical solution is that the specific steps of step (4) are:

[0041] 4-1) Sample entropy:

[0042]

[0043] Where: SampEn(m,r,N) represents the sample entropy; m is the dimension, the value is 2; r is the similarity tolerance parameter;

[0044] N is the number of sample points, A m (r) represents the probability of matching m+1 points under the similarity tolerance r; B m (r) represents the probability of matching m points under the similarity tolerance r;

[0045] Energy entropy:

[0046] H(j)=-p j log2 p j

[0047] Where: H(j) represents energy entropy; P j It represents the ratio of characteristic energy to total energy;

[0048] Permutation Entropy:

[0049]

[0050] Where: N is the number of sample points; P m The relative frequency of a sequence pattern in a subsequence sorted in ascending order by element value;

[0051] 4-2) The sample entropy, energy entropy, and permutation entropy corresponding to the optimal IMF component are used as the multidimensional feature vector after vibration signal extraction.

[0052] A further technical solution is that the specific steps of step (5) are:

[0053] 5-1) The activation function of the extreme learning machine ELM is g(w i x i +b i ), then the output matrix H(x) of the hidden layer nodes is:

[0054]

[0055] Where: w is the input layer weight; b represents the hidden layer bias;

[0056] 5-2) After entering the kernel function, the ELM model is expressed as

[0057]

[0058] 5-3) The kernel function is expressed as follows:

[0059] K hybrid =wK Poly (x i ,x j )+(1-w)K RBF (x i ,x j )

[0060] Where: w is the weight coefficient, and w∈[0,1];

[0061] K Poly (x i ,x j )=(x i ,x j +c) d

[0062] Where: d and c are the kernel parameters of the polynomial kernel respectively;

[0063]

[0064] Where: δ is the kernel parameter of RBF kernel;

[0065] 5-4) Divide the feature set into a training set and a test set, input the feature set into a hybrid kernel learning machine, train the model to obtain a wind turbine fault prediction model, input the test set into the wind turbine fault prediction model for state recognition, and complete the fault diagnosis of the wind turbine vibration signal.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. The present invention adopts a multi-strategy improved hybrid optimization algorithm to optimize the two key parameters of variational mode decomposition (VMD), namely the decomposition number k and the penalty factor α, to achieve the optimal decomposition of the vibration signal. The envelope entropy is used as the fitness function to screen the optimal components from the decomposed components. Then, the multidimensional characteristic indicators of these optimal components are calculated to construct the characteristic vector of the wind turbine vibration signal. Finally, the constructed characteristic vector is input into the HKELM for learning and training to form a wind turbine fault early warning model. The model is then compared with other models to verify the superiority of the proposed method.

[0068] 2. The multi-strategy improved hybrid optimization algorithm incorporates the exploration strategy of the Northern Goshawk algorithm into the search stage of the Sparrow algorithm, so that the algorithm can more effectively avoid falling into the local optimal solution, enhance the global search capability, and use dynamic reverse learning to improve the quality of the initial population and reduce the impact of the random position of the population. Finally, a random walk strategy is introduced at the end of the algorithm, so that during local search, the variation amplitude can be adaptively adjusted according to the current state of the population to seek the optimal parameter group in VMD, thereby effectively improving the efficiency and accuracy of signal processing, thereby improving the overall algorithm performance and providing more reliable support for the advancement of fault warning technology of wind turbines.

[0069] 3. A more comprehensive approach is adopted when extracting feature vectors. Time-domain and frequency-domain feature indicators and entropy values ​​are extracted from the signal to construct a multi-dimensional feature vector. This allows for a comprehensive characterization of wind turbine fault information, thereby improving the accuracy of fault diagnosis.

[0070] 4. The hybrid kernel function is introduced into the extreme learning machine (ELM) to obtain the HKELM. By introducing the hybrid kernel method, multiple kernel functions can be combined to better capture the complex characteristics of the data, improving the model's flexibility and performance. At the same time, by adjusting the weights of the hybrid kernels, the HKELM can effectively improve the model's accuracy, demonstrating stronger modeling capabilities and adaptability, further enhancing the accuracy of wind turbine fault prediction.

[0071] 5. The present invention proposes a multi-strategy improved hybrid optimization algorithm for extracting the multidimensional features of the optimal IMF components in combination with VMD, and using the multidimensional features as input to HKELM to achieve feature classification of signals in different states. First, the hybrid optimization algorithm is used to determine the optimal decomposition parameters of VMD and extract high-quality modal components. Then, the multidimensional features of the modal components are calculated, and a comprehensive feature vector is constructed based on this. Finally, the multidimensional feature vector is input into the HKELM model for feature classification. Experimental results show that under the test data set, the fault diagnosis method has an accuracy rate of up to 100%, and in the actual fault warning classification, the accuracy rate is as high as 97.5%, providing an effective solution for wind turbine fault warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a flow chart of the present invention.

[0073] Figure 2 It is the vibration signal of the wind turbine generator system in different states simulated in Example 1.

[0074] Figure 3 1 is a graph of convergence curves under different optimization algorithms in Example 1.

[0075] Figure 4 1 is a diagram of fault diagnosis results under different models in Example 1.

[0076] Figure 5 It is the vibration signal of the wind turbine generator set in Example 2 during actual operation.

[0077] Figure 6 1 is a diagram of fault diagnosis results under different feature extraction methods in Example 2. DETAILED DESCRIPTION

[0078] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0079] Example 1

[0080] like Figure 1 As shown, a wind turbine fault prediction method based on a multi-strategy improved hybrid optimization algorithm according to an embodiment of the present invention is characterized by comprising the following steps:

[0081] (1) Collect vibration signals of wind turbines in different states.

[0082] The simulation experiment was conducted using the fault test bench data. The data set contains 400 samples, 100 groups of each sample, and the sampling frequency is 25.6kHz. There are four states: normal, collision and friction, imbalance, and misalignment to simulate the vibration signals of wind turbines in different states, such as Figure 2 shown.

[0083] (2) Dynamic reverse learning is used to optimize the initial population quality of the sparrow algorithm. The exploration strategy of the northern goshawk algorithm is integrated into the search phase of the sparrow algorithm. The Gaussian random walk strategy is introduced to update the sparrow position. An improved hybrid optimization algorithm is obtained, and this hybrid optimization algorithm is used to determine the number of modal decompositions k and the penalty factor α.

[0084] The Sparrow Algorithm (SSA) has attracted attention for its simplicity and ease of implementation, as well as its excellent global optimization capabilities. However, when solving complex optimization problems, the algorithm suffers from limited convergence and global search capabilities, and is significantly affected by the initial position of the population. To overcome these shortcomings, the exploration strategy of the Northern Goshawk algorithm is incorporated into the Sparrow Algorithm's search phase, enabling the algorithm to more effectively avoid being trapped in local optimal solutions, enhancing its global search capabilities. Dynamic reverse learning is also used to improve the quality of the initial population and mitigate the impact of random population positions. Finally, a random walk strategy is introduced at the end of the algorithm, enabling it to adaptively adjust the variation amplitude during local search based on the current population state, further improving its performance on both multimodal and unimodal optimization problems.

[0085] The specific steps are:

[0086] (1) Assume that the population size is m:

[0087]

[0088] Where: m is the dimension of the optimization variable;

[0089] (2) Initialize the population based on the dynamic reverse learning strategy

[0090] X d =X+r1×(r2×(LB+UB-X)-X)

[0091] Where: X is the randomly generated initial population, r1 and r2 are random numbers between 0 and 1, and X d is a new population generated based on the dynamic reverse learning strategy; LB represents the lower limit and UB represents the upper limit;

[0092] (3) Introducing the exploration strategy of the Northern Goshawk optimization algorithm in the initial search phase of the Sparrow algorithm:

[0093]

[0094] P i represents the position of the i-th prey, Indicates the fitness value of the position; is the updated position, r∈[0,1], I is 1 or 2;

[0095] (4) Further position updates are as follows:

[0096]

[0097] Where: is the optimal position of the finder in the t+1th iteration, is the global worst position in the tth iteration; is the t-th position, Q is a random number that obeys the normal distribution, A+ , L represents a matrix in which each element is randomly assigned a value of 1 or -1;

[0098] (5) Introducing the Gaussian random walk strategy, the new sparrow position is updated as follows:

[0099]

[0100] Where: is a random individual, t represents the number of iterations, Indicates that the cosine function of this iteration is used to balance the global search and local development of the algorithm;

[0101] (6) Assuming that the number of individuals responsible for vigilance in a population accounts for 20% of the total population, the mathematical expression for the number of individuals responsible for vigilance in a population can be expressed as follows:

[0102]

[0103] Where: is the current global optimal position, β and k are the control step coefficients, and f i is the current position fitness value, f b 、f w are the best fitness value and the worst fitness value in this iteration respectively, ε is a constant, f i ≠f w , represents the global worst position, Updated location, Indicates the position before the update; when f i >f b This means that the sparrows are at the edge of their population and are extremely vulnerable to predators. i =f g This suggests that sparrows in the middle of the flock are aware of the danger and need to stay close to other sparrows to minimize their risk of being preyed upon.

[0104] In order to verify the effect of the hybrid optimization algorithm improved by the above multi-strategy, six common test functions are used to compare the hybrid optimization algorithm with three common optimization algorithms and analyze their convergence curves. Figure 3 As shown. Figure 3 As can be seen from the figure, the hybrid optimization algorithm always reaches the global minimum fitness, indicating that it has a stronger optimization capability. (GWO represents the Gray Wolf Optimization Algorithm, NGO represents the Northern Goshawk Optimization Algorithm, SSA represents the Sparrow Search Algorithm, and RNGSSA represents the hybrid optimization algorithm proposed in this paper).

[0105] (7) Use the above method to determine the number k of modal decomposition VMD and the penalty factor α.

[0106] First, the VMD parameters k and α were optimized. The parameters for the hybrid optimization algorithm (RNGSSA) were set as follows: population size 25, maximum number of iterations 50, and upper and lower bounds [10, 2] and [2500, 100], respectively. Using the normal operation data sample as an example, the optimal VMD decomposition parameters and optimal VMD components obtained using different optimization algorithms are shown in Table 1.

[0107] Table 1

[0108] Optimization Algorithm k α Optimal IMF component NGO 7 2500 5 SSA 8 1661 6 RNGSSA 8 1534 1

[0109] (3) Decomposing the vibration signal using the modal decomposition method in step (2) to obtain IMF components, calculating the envelope entropy of the IMF components, and selecting the IMF component with the smallest envelope entropy value as the optimal component;

[0110] 3-1) VMD decomposition variational model is:

[0111]

[0112] By introducing the Lagrange multiplication operator and the quadratic penalty term, it becomes an unconstrained problem extension:

[0113]

[0114] In the above formula: α is the penalty factor, λ is the Lagrangian operator, u k and ω k is the center frequency, represents the frequency domain result, k is the number of decompositions, and k and α are determined using a hybrid optimization algorithm;

[0115] 3-2) Use the modal decomposition method VMD to decompose the vibration signal to obtain k IMF components, and calculate the envelope entropy of each IMF component. The IMF component corresponding to the minimum envelope entropy value is taken as the optimal component.

[0116] (4) For the optimal IMF component, sample entropy, energy entropy, and permutation entropy are used to calculate the multi-dimensional feature vector.

[0117] 4-1) Sample entropy:

[0118]

[0119] Where: SampEn(m,r,N) represents the sample entropy; m is the dimension, the value is 2; r is the similarity tolerance parameter; N is the number of sample points, A m (r) represents the probability of matching m+1 points under the similarity tolerance r; B m (r) represents the probability of matching m points under the similarity tolerance r;

[0120] Energy entropy:

[0121] H(j)=-p j log2 p j

[0122] Where: H(j) represents energy entropy; p j It represents the ratio of characteristic energy to total energy;

[0123] Permutation Entropy:

[0124]

[0125] Where: N is the number of sample points; P m The relative frequency of occurrence of a sequence pattern in a subsequence sorted in ascending order by element value;

[0126] 4-2) The sample entropy, energy entropy, and permutation entropy corresponding to the optimal IMF component are used as the multidimensional feature vector after vibration signal extraction.

[0127] (5) A hybrid kernel function is introduced into an extreme learning machine to obtain a hybrid kernel learning machine. The feature set is divided into a training set and a test set. The feature set is input into the hybrid kernel learning machine, and the training model is used to obtain a wind turbine fault prediction model. The test set is input into the wind turbine fault prediction model for state recognition, thereby completing the fault diagnosis of the wind turbine vibration signal. Specifically:

[0128] Compared with traditional neural networks, ELM only needs to adjust the number of hidden layer nodes during operation. Therefore, ELM has faster computing efficiency and is effective in the field of fault diagnosis.

[0129] 5-1) The activation function of the extreme learning machine ELM is g(w i x i +b i ), then the output matrix H(x) of the hidden layer nodes is:

[0130]

[0131] Where: w is the input layer weight; b represents the hidden layer bias;

[0132] 5-2) The ELM model is expressed as follows:

[0133] F(x)=βH(x)

[0134] β is the weight between the hidden layer and the output layer, and its calculation formula is:

[0135]

[0136] Where: T is the label of the training data, is the Moore-Penrose generalized inverse matrix of H.

[0137]

[0138] Where: C is the regularization coefficient;

[0139] Furthermore, the expression of ELM is:

[0140]

[0141] However, incorrect selection of ELM hidden layer nodes will lead to defects such as unstable operation results in the algorithm. To solve these problems, the kernel function is introduced into ELM to obtain stronger stability and generalization ability. Its calculation formula is as follows:

[0142]

[0143] Where k(·) represents the kernel function.

[0144] 5-3) The combination of multiple kernel functions in KELM constitutes HKELM. Common kernel functions are divided into global kernel functions and local kernel functions. Polymorphic kernel functions, as classic global kernel functions, can better obtain the global information of distant data points, while local kernel functions represented by RBF kernel functions can effectively extract local information of data points.

[0145] Polynomial kernel function:

[0146] K Poly (x i ,x j )=(x i ,x j +c) d

[0147] Where d and c are the kernel parameters of the polynomial kernel respectively.

[0148] RBF kernel function:

[0149]

[0150] Where: δ is the kernel parameter of the RBF kernel.

[0151] The expression of the hybrid function formula obtained by combining the above two kernel functions is as follows:

[0152] K hybrid =wK Poly (x i ,x j )+(1-w)K RBF (x i ,x j )

[0153] Where: w is the weight coefficient, and w∈[0,1].

[0154] 5-4) Divide the feature set into a training set and a test set, input the feature set into a hybrid kernel learning machine, train the model to obtain a wind turbine fault prediction model, input the test set into the wind turbine fault prediction model for state recognition, and complete the fault diagnosis of the wind turbine vibration signal.

[0155] The multi-dimensional feature vector index of the IMF component is used as the input of HKELM for feature classification. (80% of the data is used as a training set and 20% as a test set) In order to demonstrate the effectiveness and efficiency of the constructed model, different combination models are introduced for comparison. The comparison results of RNGSSA-VMD-HKELM and SSA-VMD-HKELM are listed. The classification results are as follows: Figure 4 The recognition results of other different models are shown in Table 2.

[0156] Table 2

[0157]

[0158] according to Figure 4 As can be seen from Table 2, the fault diagnosis performance of the model optimized by RNTSSA is significantly better than that of other optimization models, indicating the good performance of this algorithm in optimizing fault diagnosis models. At the same time, when a single indicator is used as a feature set input model to diagnose system faults, the accuracy is 81.25%. When the comprehensive indicator proposed in this application is used as a feature vector for fault diagnosis, the accuracy jumps to more than 90%, achieving an improvement of about 8%, of which the accuracy of the model proposed in this application reaches 100%. This significant progress strongly demonstrates that in the absence of noise interference in laboratory data, the comprehensive indicator can more effectively capture signal feature information, thereby enhancing the generalization ability of the model and data processing efficiency.

[0159] Example 2

[0160] Using the method described in Example 1, the WD200-5000 wind turbine of a wind farm of Guangxi Longyuan Electric Power Company was used as the actual analysis object. The actual operating vibration data of the wind turbine in the wind farm was used for verification and fault diagnosis of the wind turbine in the wind farm. The data set contains 200 samples, of which 114 are the normal operating status data of the main shaft front bearing and 86 are the fault operating status data of the main shaft front bearing. Each sample contains 4096 data, and the vibration curve is as follows Figure 5 shown.

[0161] First, the hybrid optimization algorithm is used to search for the optimal decomposition parameters of VMD. The optimal decomposition parameters and optimal IMF components of the actual vibration signal are shown in Table 3.

[0162] Table 3

[0163] Vibration state k α Optimal IMF component Normal operating status 10 2498 1 Fault operating status 10 1063 1

[0164] The multidimensional characteristic index of the best IMF component in the above optimization results is obtained and used as the input of HKELM for feature classification. In order to reflect the effectiveness of the constructed model in actual engineering applications, the model in the above simulation analysis is introduced for comparison. A different model from the above simulation signal analysis is selected, and the comparison results of sample entropy-HKELM and multidimensional characteristic index-HKELM are listed. The classification results are as follows Figure 6 The recognition results of other different models are shown in Table 4.

[0165] Table 4

[0166]

[0167] As can be seen from the table above, compared to the laboratory vibration data, the overall diagnostic accuracy of the model decreased when processing actual data, indicating that there are certain differences between the laboratory simulated fault data and the actual vibration data. At the same time, when using a single indicator as the feature set input model to diagnose system faults, the accuracy was 72.5%. After optimization using the hybrid optimization algorithm, the accuracy increased by 10%. Moreover, on this basis, after using the comprehensive indicator as the feature vector, the accuracy increased again by 15%, reaching 97.5%. This shows that the model proposed in this application is a practical wind turbine fault warning model that can be used.

[0168] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A wind turbine fault prediction method based on a multi-strategy improved hybrid optimization algorithm, characterized in that The steps include: (1) Collect vibration signals of wind turbines in different states; (2) Dynamic reverse learning is used to optimize the initial population quality of the sparrow algorithm. The exploration strategy of the northern goshawk algorithm is integrated into the search phase of the sparrow algorithm. The Gaussian random walk strategy is introduced to update the position of the sparrows. The improved hybrid optimization algorithm is obtained and the number of modal decompositions k and the penalty factor α are determined using this hybrid optimization algorithm. (3) Decomposing the vibration signal using the modal decomposition method in step (2) to obtain corresponding IMF components, calculating the envelope entropy of the IMF components, and selecting the IMF component with the smallest envelope entropy value as the optimal component; (4) For the best IMF component, calculate its sample entropy, energy entropy, and permutation entropy to obtain a multi-dimensional feature vector; (5) The hybrid kernel function is introduced into the extreme learning machine to obtain a hybrid kernel learning machine. The feature set is divided into a training set and a test set. The feature set is input into the hybrid kernel learning machine. The training model is used to obtain a wind turbine fault prediction model. The test set is input into the wind turbine fault prediction model for state recognition, thereby completing the fault diagnosis of the wind turbine vibration signal.

2. The method according to claim 1, characterized in that: The specific steps of step (2) are as follows: (1) Assuming the population size is n: Where: m is the dimension of the optimization variable; (2) Initialize the population based on the dynamic reverse learning strategy X d =X+r1×(r2×(LB+UB-X)-X) Where: X is the randomly generated initial population, r1 and r2 are random numbers between 0 and 1, and X d is a new population generated based on the dynamic reverse learning strategy; LB represents the lower limit value, and UB represents the upper limit value; (3) Introducing the exploration strategy of the Northern Goshawk optimization algorithm in the initial search phase of the Sparrow algorithm: P i represents the position of the i-th prey, Indicates the fitness value of the position; is the updated position, r∈0,1], I is 1 or 2; (4) Further position updates are as follows: Where: is the optimal position of the finder in the t+1th iteration, is the global worst position in the tth iteration; is the t-th position, Q is a random number that follows a normal distribution, A + , L represents a matrix in which each element is randomly assigned a value of 1 or -1; (5) Introducing the Gaussian random walk strategy, the new sparrow position is updated as follows: Where: is a random individual, t represents the number of iterations, Indicates this iteration; (6) Assuming that the number of individuals responsible for vigilance in the population accounts for 20% of the total population, the mathematical expression for the number of individuals responsible for vigilance in the population can be expressed as follows: Where: is the current global optimal position, β and k are the control step length coefficients, and f i is the fitness value of the current position, f b 、f w are the best fitness value and the worst fitness value in this iteration, ε is a constant, f i ≠f w , represents the global worst position, Updated location, Indicates the position before updating; (7) Use the above method to determine the number k of modal decomposition VMD and the penalty factor α.

3. The method according to claim 1, characterized in that: The specific steps of step (3) are: 3-1) The VMD decomposition variational model is: x=∑ k u k (t) By introducing the Lagrange multiplication operator and the quadratic penalty term, it becomes an unconstrained problem expansion: In the above formula: α is the penalty factor, λ is the Lagrangian operator, u k and ω k is the center frequency, represents the frequency domain result, k is the number of decompositions, and k and α are determined using a hybrid optimization algorithm; 3-2) Use the modal decomposition method VMD to decompose the vibration signal to obtain IMF components, calculate the envelope entropy of each IMF component, and take the IMF component corresponding to the minimum envelope entropy value as the optimal component.

4. The method according to claim 1, characterized in that: In the step (4): 4-1) Sample entropy: Where: SampEn(m,r,N) represents sample entropy; m is the dimension, the value is 2; r is the similarity tolerance parameter; N is the number of sample points, A m (r) represents the probability of matching m+1 points under similarity tolerance r; B m (r) represents the probability of matching m points under similarity tolerance r; Energy entropy: H(j)=-p j log2pj Where: H(j) represents energy entropy, p j It represents the ratio of characteristic energy to total energy; Permutation Entropy: Where: N is the number of sample points; P m is the relative frequency of occurrence of a sequence pattern in a subsequence sorted in ascending order by element value; 4-2) The sample entropy, energy entropy and permutation entropy corresponding to the optimal IMF component are used as the multidimensional feature vector after the vibration signal is extracted.

5. The method according to claim 1, characterized in that: The specific steps of step (5) are: 5-1) The activation function of the extreme learning machine ELM is g(w i x i +b i ), then the output matrix H(x) of the hidden layer nodes is: Where: w is the input layer weight; b represents the hidden layer bias; 5-2) After entering the kernel function, the ELM model is expressed as 5-3) The kernel function is expressed as follows: K hybrid =wK Poly (x i ,x j )+(1-w)K RBF (x i ,x j ) Where: w is the weight coefficient, and w∈[0,1]; in: K Poly (x i ,x j )=(x i ,x j +c) d Where: d and c are the kernel parameters of the polynomial kernel respectively; Where: δ is the kernel parameter of the RBF kernel; 5-4) The feature set is divided into a training set and a test set, the feature set is input into a hybrid kernel learning machine, the training model is used to obtain a wind turbine fault prediction model, the test set is input into the wind turbine fault prediction model for state recognition, and the fault diagnosis of the wind turbine vibration signal is completed.

Images