Wind turbine generator blade vibration mode feature identification method

Abstract

The invention discloses a wind turbine generator blade vibration mode feature recognition method, which mainly comprises the following steps of: arranging three-axis piezoelectric acceleration sensors at equal intervals from a blade root to a maximum chord length position, and adaptively selecting an optimal channel as a target input vibration signal according to envelope dispersion and a period proportion; introducing a vibration signal denoising method based on regenerative phase shift sine-assisted empirical mode decomposition, and constructing a Fisher ratio-based multi-dimensional fusion index to remove a noise component; estimating a system order range according to a singular entropy jump value theory, and designing modal similarity and a confidence index to accurately estimate a real order of a blade system; introducing three types of constraints of structure maintenance, modal sparsity and energy smoothness to jointly optimize a low-rank approximation strategy so as to realize optimal reconstruction of the Hankel matrix; constructing a fitness function selected by a clustering center by combining the point set density of the sample and Euclidean distance information, and optimizing a modal extraction result by adopting inter-class dispersivity and an intra-class sample number; according to the method, the environmental noise can be effectively removed, the system order can be accurately determined, and finally the modal parameters of the system can be accurately identified.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine blade fault diagnosis, and in particular to a method for identifying vibration modal features of wind turbine blades. Background Art

[0002] With the rapid growth of demand in my country's wind power market, the operation and maintenance costs of wind turbines are increasing exponentially. Blades, the core components that generate wind energy, are subject to the long-term coupling of complex alternating stresses, making them prone to internal structural damage such as coating whitening and bulging. These damages can cause equipment downtime at best, or even fatalities.

[0003] Currently, the mainstream blade damage identification method still relies primarily on manual inspections supplemented by telescopes. This method suffers from shortcomings such as low efficiency, susceptibility to worker subjective influence, inability to monitor around the clock, and susceptibility to safety accidents. With the rapid development of online detection technology, vibration methods have become the mainstream solution for blade structural condition assessment due to their low sensor cost, real-time online monitoring, and high sensitivity to minor damage. This method primarily relies on accelerometers placed at key blade locations to identify structural damage by analyzing blade modal parameters contained in the vibration signal. However, during operation, wind turbine blades are subject to coupled interference from multiple sources of noise, including aerodynamic, electrical, and mechanical noise, causing damage-sensitive characteristic components to be overwhelmed by strong background noise. Furthermore, existing diagnostic methods often assume that the system input excitation is stationary white noise, but in reality, it exhibits non-stationary, time-varying environmental excitation, resulting in limited effectiveness in detecting blade structural defects. To this end, this paper focuses on the fields of vibration signal denoising, system order estimation, and modal parameter identification, and discloses a method for identifying vibration modal characteristics of wind turbine blades. Summary of the Invention

[0004] In response to the problems and challenges of the above-mentioned blade damage identification method based on vibration data, the present invention provides a method for detecting internal structural damage of wind turbine blades based on improved random subspace identification. The technical solution of the present invention is as follows: Step 1: Arrange a triaxial piezoelectric acceleration sensor array at equal intervals between the target blade root and the maximum chord length position, synchronously collect multi-channel vibration signals, use an envelope smoothing algorithm to enhance feature information, and accurately identify target events based on a double-threshold detection method. Assess data quality by calculating the envelope discreteness and cycle ratio of each channel signal, and then select the optimal channel data as the target input vibration signal. Step 2: RPSEMD (Regenerated Phase-Shifted Sinusoid-Assisted Empirical Mode Decomposition) is combined with the interquartile range method to decompose the target input vibration signal into a series of intrinsic mode functions (IMFs). A multidimensional fusion index based on the Fisher ratio is constructed to remove noise components from the original signal. The details are as follows: (1) Perform empirical mode decomposition (EMD) on the original target input vibration signal and determine the sinusoidal auxiliary signal model based on the IMF obtained by decomposition; (2) A sinusoidal auxiliary signal is superimposed on the original vibration signal to construct a set of enhanced signals with different phases, and the first-order IMF component set of each signal is obtained by EMD decomposition; (3) Based on the correlation coefficients between the decomposed IMF components, the weight index for evaluating the effectiveness of the IMF is obtained: in, Represents the mutual correlation coefficient of two IMFs. The autocorrelation coefficients of the IMF components are screened using the interquartile range method, and the target IMF is reconstructed using the mean strategy; (4) Calculate the energy ratio, related kurtosis and approximate entropy characteristics of IMF, and use Fisher discriminant ratio to determine the The weight of the features: in, represents the number of categories, Indicates the The number of samples of the class, Indicates the All samples of the class The mean of the features, Represents all class samples about the The mean of the features, Indicates the Class Sample No. The value of a feature; (5) Through the weighted fusion strategy, a multidimensional comprehensive index is constructed to quantify the significance of fault characteristics. The interquartile range method is used to screen the target IMF and reconstruct the vibration signal, thereby effectively separating the blade vibration modal components of interest.

[0005] Step 3: Analyze the entropy mutation points in the signal singular value sequence based on the singular entropy jump value theory, preliminarily estimate the order range of the blade system, integrate the modal confidence criterion and similarity to evaluate the order stability and determine the true order of the system, as follows: (1) Obtain the projection matrix of the preprocessed vibration signal through orthogonal projection, and use singular value decomposition to obtain the singular value increment of the projection matrix: in, Indicates the The singular entropy increment of order, For the The singular values ​​of order, is the total number of non-zero singular values; the maximum value of the singular entropy increment is identified according to the jump method, and the entropy mutation point in the singular value sequence is determined according to its index position , and estimate the range of system orders; (3) Based on the discrete random state space theory, the state transfer equation and observation equation of the system are derived, and the modal confidence index for quantifying the extreme stability is constructed based on them: Where N is the upper limit of the preliminary estimated system order range, 、 Representing the The frequency mean of each mode and its Order frequency, 、 Representing the The damped mean of the modes and their Order damping, , indicating the The mode confidence of the mode, Representative The first mode order vibration mode; 、 、 Represent the weight coefficients of frequency, damping and mode shape respectively, and the coefficients of the weights are estimated using the range variance method; (4) Combining the modal consistency criterion and modal vector space theory, a modal similarity index for evaluating extreme point geometric differences is constructed: (5) Based on the modal confidence criterion, whether the pole is a stable point is determined, and potential effective modes are screened out. Then, through the similarity between adjacent orders, false modes and noise interference are eliminated, and finally the true order of the blade system is determined.

[0006] Step 4: Introduce three types of constraints: structure preservation, modal sparsity, and energy smoothness to jointly optimize the low-rank approximation strategy, thereby achieving the optimal reconstruction of the Hankel matrix. Combine the point density and Euclidean distance information of the samples to construct a fitness function for cluster center selection. Finally, use the inter-class dispersion and the number of samples within the class to optimize the modal extraction results, as follows: (1) Based on the true order of the system estimated in step 3, the random subspace method is used to preliminarily construct the original Hankel matrix containing the modal characteristics. The three types of constraints, structure preservation, modal sparsity, and energy smoothness, are introduced to jointly optimize the low-rank approximation strategy to obtain the reconstructed Hankel matrix: in, represents the original Hankel matrix, represents the system order, represents the Frobenius norm, 、 、 Represent the structure preservation, modal sparsity and energy smoothness constraints respectively, 、 、 is the weight coefficient of the corresponding constraint item, express Middle Rank Column elements, represents the left singular vector matrix column vector, express norm; (2) The signal is processed by frame decomposition. The reconstructed Hankel matrix is ​​decomposed and features are extracted frame by frame to construct a global modal parameter matrix. Then, the fitness function for cluster center selection is constructed by combining the point density and Euclidean distance information of the sample: in, Represents sample points The field collection, Indicates the number of signal frames. represents the bandwidth parameter used to control the size of the neighborhood, Indicates the cluster center sample closest to the current sample, Represents sample points and cluster centers The Euclidean distance of Represents the weight factor, and then the dynamic selection of cluster centers is realized based on the k-means algorithm; (3) In order to improve the accuracy and stability of modal recognition results, clustering quality evaluation indicators are designed by quantifying the degree of separation between classes and the degree of clustering within classes: in, Indicates is the number of samples in the cluster center, Indicates is the sample set of the cluster center; based on the optimal cluster center, the nearest neighbor samples are selected as the final modal parameters, thereby realizing the accurate identification of the modal parameters of the blade system.

[0007] The advantages of the present invention are: 1. The present invention designs a RPSEMD filtering method based on multi-dimensional fusion screening criteria, which can effectively suppress vibration signal noise under complex environmental noise interference; 2. The present invention analyzes the entropy mutation points in the signal singular value sequence based on the singular entropy jump value theory, preliminarily estimates the order range of the blade system, integrates the modal confidence and similarity to evaluate the order stability and determine the true order of the system; 3. The present invention introduces three types of constraints, namely structure preservation, modal sparsity and energy smoothness, to jointly optimize the low-rank approximation strategy, thereby achieving the optimal reconstruction of the Hankel matrix. Combining the point density and Euclidean distance information of the samples, the fitness function for cluster center selection is constructed, and finally the inter-class dispersion and the number of samples within the class are used to optimize the modal extraction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the following briefly introduces the practical drawings required in the embodiments or the description of the prior art.

[0009] Figure 1 Schematic diagram of the process of the present invention.

[0010] Figure 2 Schematic diagram of the RPSEMD filtering process based on multi-dimensional fusion screening criteria.

[0011] Figure 3 Schematic diagram of the system order identification process based on singular entropy jump theory and modal stability assessment.

[0012] Figure 4 Schematic diagram of the Hankel matrix reconstruction and adaptive clustering modal extraction process based on multi-constraint optimization. Implementation Method To make the purpose, technical solutions and advantages of the present invention more clear, the following embodiments of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to illustrate the present invention and cannot be used to limit the scope of the present invention.

[0013] A method for identifying vibration modal features of wind turbine blades, such as Figure 1 The specific steps are as follows: Step 1: Filter out the optimal data from the multi-channel data collected by the triaxial piezoelectric acceleration sensor as the target input vibration signal, as follows: (1) Place a triaxial piezoelectric accelerometer at a distance of 1 / 4 of the chord length from the blade root; (2) Perform a fine Hibbert transform on the vibration signal to obtain the envelope signal, and then use the sliding window method for smoothing; (3) The double threshold method is used to locate vibration characteristic events. The data quality of each channel is measured by envelope discreteness and cycle ratio, and the optimal channel data is selected as the target input vibration signal.

[0014] Step 2: Fuse the regenerative phase-shifted sine-assisted empirical mode decomposition and the interquartile range method to decompose the target input vibration signal into a series of IMFs, and construct a multidimensional fusion index based on the Fisher ratio to remove the noise component in the original signal. Figure 2 , as follows: (1) Perform EMD decomposition on the original target input vibration signal and determine the sinusoidal auxiliary signal model based on the IMF obtained by decomposition: in, , Respectively represent the amplitude and frequency determined based on the first-order IMF, For Uniformly sampled phases; (2) A sinusoidal auxiliary signal is superimposed on the original target input vibration signal to construct a set of enhanced signals with different phases, and the first-order IMF component set of each signal is obtained by EMD decomposition; (3) Based on the correlation coefficients between the decomposed IMF components, the weight index for evaluating the effectiveness of the IMF is obtained: in, Represents the mutual correlation coefficient of two IMFs. The autocorrelation coefficients of the IMF components are screened using the interquartile range method, and the target IMF is reconstructed using the mean strategy; (4) Calculate the energy ratio characteristics, approximate entropy and related kurtosis of each IMF component: in, Represents the total energy of the target input vibration signal, Represents the current energy of the IMF, represents the embedding dimension, is the tolerance threshold, represents the entropy measurement function, represents the reconstruction vector, Represents the number of vector pairs whose absolute difference is less than the tolerance threshold, Represents the target input vibration signal The sampling value at the moment, Indicates the target input vibration signal length, is the number of signal shift cycles, is the signal period; Fisher discriminant ratio was used to determine the The weight of the features: in, represents the number of categories, Indicates the The number of samples of the class, Indicates the All samples of the class The mean of the features, Represents all class samples about the The mean of the features, Indicates the Class Sample No. The value of a feature; (5) Through the weighted fusion strategy, a multidimensional comprehensive index is constructed to quantify the significance of fault characteristics. The interquartile range method is used to screen the target IMF and reconstruct the vibration signal, thereby effectively separating the blade vibration modal components of interest.

[0015] Step 3: Analyze the entropy mutation points in the signal singular value sequence based on the singular entropy jump value theory, preliminarily estimate the order range of the blade system, integrate the modal confidence and similarity to evaluate the order stability and determine the true order of the system. Figure 3 , as follows: (1) Obtain the projection matrix of the preprocessed vibration signal through orthogonal projection : in," " represents the pseudo-inverse matrix, and the "past" output matrix and the "future" output matrix The calculation formula is as follows: in, Represents the vibration signal after filtering The sampling value at the moment, the projection matrix is ​​decomposed into singular values: in , is an orthogonal matrix, Sort in descending order The main singular value diagonal matrix, the singular value increment of the projection matrix: in, Indicates the The singular entropy increment of order, For the The singular values ​​of order, Represents the total number of non-zero singular values; the maximum value of the system order is 1.5 times the maximum value of the singular value increment; (2) Based on the Kalman filter sequence theory and the least squares method, the system matrix and output vector are identified, and the frequency, damping, and vibration mode of the system are obtained through eigenvalue decomposition: in, represents the eigenvalue of the system matrix, The magnitude of the undamped natural frequency, represents the output vector, represents the eigenvector matrix; Based on the discrete random state space theory, the state transfer equation and observation equation of the system are derived, and the modal confidence index for quantifying the extreme stability is constructed accordingly: Where N is the upper limit of the preliminary estimated system order range, 、 Representing the The frequency mean of each mode and its Order frequency, 、 Representing the The damped mean of the modes and their Order damping, , indicating the The mode confidence of the mode, Representative The first mode order vibration mode; 、 、 Represent the weight coefficients of frequency, damping and mode shape respectively, and the coefficients of the weights are estimated using the range variance method: in, 、 、 Respectively represent The extreme values ​​of order frequency, damping and vibration mode, 、 、 Respectively represent Order frequency, damping and mode variance, is the comprehensive coefficient of variation, which can be taken as 0.4; (4) Combining the modal consistency criterion and modal vector space theory, a modal similarity index for evaluating extreme point geometric differences is constructed: (5) Based on the modal confidence criterion, whether the pole is a stable point is determined, and potential effective modes are screened out. Then, through the similarity between adjacent orders, false modes and noise interference are eliminated, and finally the true order of the blade system is determined.

[0016] Step 4: Introduce three types of constraints: structure preservation, modal sparsity, and energy smoothness to jointly optimize the low-rank approximation strategy, thereby achieving the optimal reconstruction of the Hankel matrix. Combine the point density and Euclidean distance information of the sample to construct the fitness function for cluster center selection. Finally, use the inter-class scatter and the number of samples within the class to optimize the modal extraction results. Figure 4 , as follows: (1) Based on the true order of the system estimated in step 3, the random subspace method is used to preliminarily construct the original Hankel matrix containing the modal characteristics. The three types of constraints, structure preservation, modal sparsity, and energy smoothness, are introduced to jointly optimize the low-rank approximation strategy to obtain the reconstructed Hankel matrix: in, represents the original Hankel matrix, represents the system order, represents the Frobenius norm, 、 、 Represent the structure preservation, modal sparsity and energy smoothness constraints respectively, 、 、 The weight coefficient corresponding to the constraint item, express Middle Rank Column elements, represents the left singular vector matrix column vector, express norm; (2) The signal is processed by framing, and the reconstructed Hankel matrix is ​​decomposed to extract features frame by frame to construct the global modal parameter matrix: in, Indicates the Frame No. Order frequency, damping, mode shape; Combining the point density and Euclidean distance information of the sample, a fitness function for cluster center selection is constructed: in, Represents sample points The field collection, Indicates the number of signal frames. represents the bandwidth parameter used to control the size of the neighborhood, Indicates the cluster center sample closest to the current sample, Represents sample points and cluster centers The Euclidean distance of represents the weight factor; (3) Design clustering quality evaluation indicators based on the degree of separation between classes and the degree of aggregation within classes: in, Indicates is the number of samples in the cluster center, Indicates is the sample set of cluster centers;

[0017] Subsequently, the optimal cluster center is dynamically selected based on the fitness function. The statistical characteristics of the cluster quality evaluation index are calculated in real time during the iteration process, and the threshold is set to the index mean minus 1.5 times the standard deviation. Then, low-quality clusters are dynamically eliminated based on the threshold until the number of remaining clusters converges to the true order of the system. Finally, the nearest neighbor sample is selected as the final modal parameter based on the optimized cluster center, thereby achieving accurate identification of the modal parameters of the blade system.

Claims

1. A method for identifying vibration modal features of wind turbine blades, characterized in that: The method comprises the following steps: S1. Triaxial piezoelectric accelerometers are evenly spaced from the blade root to the maximum chord length to acquire the target blade vibration signal. The collected multi-channel data is envelope-smoothed and the target event is accurately located using a double-threshold method. The data quality of each channel is measured using envelope dispersion and cycle ratio, and the optimal channel data is selected as the target input vibration signal. S2. Fusion of regenerative phase-shifted sine-assisted empirical mode decomposition and the interquartile range method decomposes the target input vibration signal into a series of intrinsic mode functions (IMFs). A multidimensional fusion index based on the Fisher ratio is constructed to remove noise components from the original signal. S3. Analyze the entropy mutation points in the signal singular value sequence based on the singular entropy jump value theory, preliminarily estimate the order range of the blade system, and integrate the modal confidence criterion and similarity to evaluate the order stability and determine the true order of the system; S4. Three types of constraints, namely structure preservation, modal sparsity, and energy smoothness, are introduced to jointly optimize the low-rank approximation strategy, thereby achieving the optimal reconstruction of the Hankel matrix. The point density and Euclidean distance information of the samples are combined to construct a fitness function for cluster center selection. Finally, the inter-class dispersion and the number of samples within a class are used to optimize the modal extraction results.

2. A method for identifying vibration modal features of a wind turbine blade according to claim 1, characterized in that: The step S2 comprises: S201. Performing empirical mode decomposition (EMD) on the original target input vibration signal, and determining the amplitude and frequency parameters of the sinusoidal auxiliary signal model based on the IMF obtained by the decomposition; S202. A sinusoidal auxiliary signal is superimposed on the original vibration signal to construct a set of enhanced signals of different phases, and the first-order IMF component set of each signal is obtained by EMD decomposition; S203. Calculate the autocorrelation coefficients between the decomposed IMF components, use the interquartile range method to optimize and select effective IMF components, and reconstruct the target IMF using the mean strategy; S204. Calculate the energy ratio, related kurtosis and approximate entropy characteristics of IMF, and use Fisher discriminant ratio to determine the The weight of the features: in, represents the number of categories, Indicates the The number of samples of the class, Indicates the All samples of the class The mean of the features, Represents all class samples about the The mean of the features, Indicates the Class Sample No. The value of a feature; S205. A multidimensional comprehensive index for quantifying the significance of fault features is constructed through a weighted fusion strategy. The target IMFs are screened and the vibration signals are reconstructed using the interquartile range method, thereby effectively separating the blade vibration modal components of interest.

3. The method for identifying vibration modal features of a wind turbine blade according to claim 1, wherein: The step S3 comprises: S301. Obtain the projection matrix of the preprocessed vibration signal by orthogonal projection, and obtain the singular value increment of the projection matrix by singular value decomposition: in, Indicates the The singular entropy increment of order, For the The singular values ​​of order, Represents the total number of non-zero singular values; uses the singular entropy jump value theory to preliminarily estimate the range of system order; S302. Based on the discrete random state space theory, the state transition equation and observation equation of the system are derived, and the modal confidence index for quantifying the extreme stability is constructed based on them: Where N is the upper limit of the preliminary estimated system order range, 、 Representing the The frequency mean of each mode and its Order frequency, 、 Representing the The damped mean of the modes and their Order damping, , indicating the The mode confidence of the mode, Representative The first mode order vibration mode; 、 、 Represent the weight coefficients of frequency, damping and vibration mode respectively, and the coefficients of the weights are estimated using the range variance method, and satisfy ; S303. Combining the modal consistency criterion and modal vector space theory, a modal similarity index for evaluating extreme point geometry differences is constructed: S304. Estimate the true order of the blade system under environmental excitation through confidence threshold screening and adjacent order difference test.

4. The method for identifying vibration modal features of a wind turbine blade according to claim 1, wherein: The step S4 comprises: S401. Based on the true order of the system, a random subspace method is used to preliminarily construct the original Hankel matrix containing the modal characteristics. Subsequently, a low-rank approximation strategy is optimized by introducing three types of constraints: structure preservation, modal sparsity, and energy smoothness. This achieves the optimal reconstruction of the Hankel matrix and addresses the problem of high-dimensional information loss during eigenvalue decomposition. S402. The signal is framed and features are extracted frame by frame by performing matrix decomposition on the reconstructed Hankel matrix to construct a global modal parameter matrix. Then, a fitness function for cluster center selection is constructed by combining the point density and Euclidean distance information of the samples to avoid the problem of the algorithm falling into a local optimum due to random initialization of the cluster centers. S403. Design clustering quality evaluation indicators based on the degree of inter-class separation and intra-class clustering to improve the accuracy and stability of modal recognition results; take the optimal cluster center as the benchmark, select the nearest neighbor samples as the final modal parameters, and thus achieve accurate identification of the blade system modal parameters.

Images