Wind turbine generator bearing fault feature extraction method based on improved AMVMD

Abstract

The invention discloses a wind turbine generator bearing fault feature extraction method based on improved adaptive multivariate variational mode decomposition, and belongs to the technical field of wind turbine generator bearing fault diagnos.The method comprises the steps that S1, logistic-tent chaotic mapping is used for replacing a traditional random number generation method, particle positions in a QPSO algorithm are initialized, and initialization uniformity and randomness are ensured; s2, adaptively selecting a modal number K and a penalty factor through an improved QPSO algorithm, performing parameter optimization by taking an energy error as a fitness function, and decomposing the processed signal by using MVMD to obtain a plurality of modal components representing different frequency components; and S3, calculating the kurtosis value and the Shannon entropy of each component, selecting a reconstruction component for subsequent feature extraction, analyzing the change of the signal through a Teager energy operator, and carrying out envelope analysis to extract the fault frequency so as to identify the specific fault features of the bearing. The method has practicability and high efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine bearing fault diagnosis. Specifically, a wind turbine bearing fault feature extraction method based on improved adaptive multivariate variational mode decomposition is proposed to solve the problem that key parameters need to be manually set during signal decomposition, which affects subsequent fault feature extraction. Background Art

[0002] As a clean and renewable energy source, wind energy has been increasingly valued in recent years, and the construction and operation of wind turbines has gradually increased. In the context of global energy transformation, the proportion of wind power has continued to increase, which has promoted the research and development of related technologies. During the operation of wind turbines, they also face the challenge of frequent failures of key components, resulting in equipment downtime and increased maintenance costs. Bearing failure is one of the main causes of wind turbine failure, so it is crucial to detect and diagnose faults in a timely and effective manner.

[0003] Rolling bearings are the basic units in wind turbines. Condition monitoring and fault diagnosis of rolling bearings can reduce downtime caused by faults, reduce economic losses, and avoid major safety accidents. This is of great significance. The existing methods for extracting rolling bearing fault features face the problem of manually setting parameters during signal decomposition, which affects the effect of subsequent fault feature extraction. Therefore, it is necessary to use some new, practical and efficient methods for extracting bearing fault features. Traditional bearing fault diagnosis usually performs time domain and frequency domain analysis on bearing vibration signals. Time domain analysis calculates the time domain feature quantities of the signal, such as kurtosis, kurtosis, root mean square, margin factor, etc. The main methods for fault extraction in the past include: calculating the kurtosis average of the sub-signal bandpass filter and using the method of converting the fast average kurtosis into a two-dimensional kurtosis diagram, which is conducive to subsequent fault identification; envelope analysis and energy spectrum decomposition; introducing multiple compression transformations to construct a time-frequency diagram of energy aggregation, which realizes the comprehensive extraction of multiple fault features; performing wavelet packet decomposition on the vibration signal, extracting the energy spectrum of the key frequency band information as the feature vector, combining principal component analysis and hierarchical clustering methods, and then effectively extracting the fault features of non-stationary vibration signals.

[0004] In order to effectively reduce the interference of extreme noise points, a method is proposed to construct the random feature energy spectrum of the signal using a sparse random feature model, adaptively decompose the spectrum using the average harmonic intensity ratio as an indicator, and reconstruct the random features. Wavelet transform and empirical mode decomposition (EMD) are commonly used methods for time-frequency domain analysis, among which EMD has the problem of decomposition endpoint effect and pseudo component.

[0005] Multivariate variational mode decomposition (MVMD) is a generalization of multi-channel data. It can obtain a series of modal aligned subsequences and is widely used in mechanical fault diagnosis, water conservancy prediction and other fields. In the MVMD algorithm, the fusion parameter combining power spectrum cross entropy and reconstruction error is introduced as the fitness function to achieve the optimal selection of parameters and effectively identify tunnel structures with different degrees of damage. The MVMD method is used to extract the intrinsic principle between a single parameter and multiple related variables, and the Transformer model is combined to achieve accurate prediction of the set parameter. The application of the off-target data recovery method combining MVMD and a fully convolutional network can mine the spatiotemporal correlation between multi-channel data and achieve data recovery. The above method determines the key parameters of MVMD by constructing indicators, but can only determine one parameter and has a large amount of calculation. In addition, as the amount of data increases, training these models may take a long time or even fail. Summary of the invention

[0006] In view of the above problems, the present invention proposes a new adaptive variational mode decomposition method to solve the problem that the key parameters need to be manually set during signal decomposition, which affects the subsequent fault feature extraction: logistic-tent chaotic mapping is used to replace random number initialization, energy error is used as a new fitness function, and the key parameters of MVMD, the intrinsic mode number and the penalty factor, are optimized to achieve adaptive decomposition of the signal. The optimal reconstruction component is selected by the kurtosis value and Shannon entropy, and the fault feature extraction is achieved by combining teager demodulation and envelope analysis.

[0007] The purpose of the present invention is to solve the defects of the existing model. On the basis of extracting the rolling bearing fault features, a model based on improved adaptive multivariate variational modal decomposition is adopted to extract the features of the rolling bearings of wind turbines, which improves the accuracy of fault extraction and avoids the adverse effects of artificially set parameters.

[0008] To achieve the above purpose, the technical solution adopted by the present invention is:

[0009] A new method for extracting wind turbine bearing fault features: based on improved adaptive multivariate variational mode decomposition (AMVMD), including the following steps:

[0010] S1. Use logistic-tent chaotic mapping instead of traditional random number generation method to initialize the particle positions in the QPSO algorithm to ensure the uniformity and randomness of initialization, which specifically includes the following sub-steps:

[0011] S11, initialize the QPSO algorithm, including particle population size, particle dimension, maximum number of iterations, convergence threshold, upper and lower bounds of K and α;

[0012] S12, control parameter, set the control parameter r in the logistic-tent chaotic mapping. Appropriate parameters can generate a good chaotic sequence. Usually, the chaotic mapping value r is selected in the range of (0,1), at which time the function is in a chaotic state. When r is 0.5, the function mapping is relatively uniform.

[0013] S13, initialize the seed value, select an initial seed value x0, usually in the range of (0,1). This value is the starting point for generating the chaotic sequence;

[0014] S14, generate a chaotic sequence, where the logistic mapping formula is as follows:

[0015] x n+1 = r·x n ·(1-x n )

[0016] where x n Represents the state after the nth iteration. In each iteration, the next value x is calculated according to the above formula n+1 , until the desired number of values ​​are generated;

[0017] S15. Generate a logistic-tent chaotic mapping sequence. The calculation formula is as follows:

[0018]

[0019] Using chaotic mapping instead of random numbers to initialize the population makes the initialization more uniform and improves the global optimization ability of the algorithm;

[0020] S16, the generated chaotic sequence value x n Map to the position of the particle in the search space and initialize the particle swarm to ensure that each particle has an initial position according to the set number of particles.

[0021] S2. Adaptively select the modal number (K) and penalty factor through the improved QPSO algorithm, use the energy error as the fitness function, perform parameter optimization, decompose the processed signal using MVMD, and obtain multiple modal components representing different frequency components, which specifically includes the following sub-steps:

[0022] S21, initialize a set of particle positions using logistic-tent chaotic mapping, where the horizontal and vertical coordinates of the particles represent K and α respectively;

[0023] S22. Perform MVMD decomposition according to the initialized K and α. MVMD is a non-recursive signal processing method that decomposes the original signal into a series of modal components with limited bandwidth, iteratively searches for each center frequency so that these modes can jointly reproduce the input signal.

[0024] First, the input signal is transformed into an analytical signal through Hilbert transform to obtain its one-sided spectrum and exponential term Multiply and modulate to the corresponding baseband. Calculate the L of the demodulated signal 2 The norm square estimates the bandwidth of each mode, and the established MVMD constrained optimization mathematical model is as follows:

[0025]

[0026] where w k is the center frequency of the kth modal component.

[0027] Secondly, two penalty terms are introduced to transform the constrained variational problem into an unconstrained variational problem, which improves the reconstruction accuracy of the signal under noise interference and ensures that the constraints are strictly met. The formula obtained is as follows:

[0028]

[0029] Finally, the alternating multiplier method is used to update the parameters and divide the complete optimization problem into multiple simple sub-optimization problems, which are transformed into the following sub-optimization problems:

[0030]

[0031] Transforming the above iterative problem into the frequency domain, the formula is as follows:

[0032]

[0033] The center frequency and Lagrangian operator are shown in the formula:

[0034]

[0035] in and x c (t) and λ c Fourier transform of (t); for u k (t) Fourier transform of the n+1th iteration.

[0036] Use the MVMD method to decompose into n modal components, see the formula:

[0037]

[0038] S23. Using energy error as a new fitness function, quantitatively evaluate the decomposition effect of MVMD under different parameters to find the optimal solution for parameter setting. The energy E0 of the original signal x(t) is expressed as:

[0039]

[0040] The sum of the energy of each modal component E total It can be expressed as:

[0041]

[0042] When the decomposed modal components are completely orthogonal to the original signal, the sum of the energies of the modal components E total Equal to the energy E0 of the original signal x(t):

[0043] E total =E1+E2+…+E n =E0

[0044] When the modal components are not completely orthogonal to the original signal, the sum of the energies of the components E total The energy E0 of the target signal is not equal, and the absolute value of the calculated difference is recorded as the energy error |E er |:

[0045] |E er |=|E0-E total |=|E0-(E1+E2+…+E n )|

[0046] Energy error |E er The smaller the | value, the better the orthogonality of the modal components obtained by signal decomposition, and the better the effect of the modal components jointly reproducing the original signal, which can reduce the loss of information carried in the original signal and is more conducive to the subsequent fault feature extraction.

[0047] Calculate the fitness value of each particle |E er |, and save the current minimum value, recorded as min(|E er |);

[0048] S24, update the particle position, calculate the fitness value again and compare it with min(|E er |) comparison, if the current particle fitness value is less than min(|E er |) then update the minimum value, where the formula for updating the particle position is as follows:

[0049] v ij =ωv ij +c1r1(p ij -x ij )+c2r2(g j -x ij );

[0050] S25. Repeat the above steps until the maximum number of iterations is reached or the convergence threshold is reached, stop the optimization and output the best particle position, which is the optimal parameters K and α of MVMD.

[0051] S3, calculate the kurtosis value of each component and the Shannon entropy to select the reconstructed component for subsequent feature extraction, analyze the change of the signal through the Teager energy operator, and perform envelope analysis to extract the fault frequency to identify the specific fault characteristics of the bearing, which specifically includes the following sub-steps:

[0052] S31. Calculate the kurtosis value. The kurtosis is a statistic used to measure the sharpness of a signal, and is usually used to identify impact signals. The higher the value, the more transient information the signal contains. Calculate the kurtosis value of each modal component decomposed by AVMMD, and select the modal component with a higher kurtosis value for subsequent processing;

[0053] S32. Calculate Shannon entropy, which is a measure of the uncertainty of random variables and reflects the complexity and information content of the signal. Calculate the Shannon entropy value for each modal component and select the mode with the lower entropy value;

[0054] S33, reconstructing the modal components, combining the calculation results of the kurtosis value and Shannon entropy, and selecting suitable modal components for reconstruction. These components are usually those that are most representative in fault feature extraction;

[0055] S34. The input signal is converted into an analytical signal through Hilbert transform, and the selected modal components are analyzed using the Teager energy operator to capture the transient characteristics of the signal. The specific fault characteristics of the bearing are identified by analyzing the envelope signal.

[0056] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0057] In the present invention, the improved AVMMD method can more accurately decompose the signal, extract the characteristic frequency related to the fault, and improve the accuracy of fault diagnosis.

[0058] In the present invention, by introducing quantum uncertainty and chaotic mapping, the improved QPSO algorithm enhances the diversity and search ability of particles and effectively avoids local optimal solutions. The improved QPSO algorithm can automatically optimize the modal number (K) and the penalty factor (α), reduce manual intervention, and improve optimization efficiency and applicability.

[0059] The present invention proposes kurtosis value and Shannon entropy as indicators for component selection, uses Teager energy operator and envelope analysis to extract fault features, and compares the simulation signal and actual bearing data with the multivariate empirical mode decomposition method, which is practical and efficient.

[0060] Comparing the method of the present application with the slime mold optimization algorithm, the method of the present application is not easy to fall into the local optimum and has higher optimization efficiency; the experimental results are compared with the MEMD method, and the method is verified by simulation signals and actual data. It can clearly extract the frequencies of different types of faults and has certain practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of a new method for extracting fault characteristics of wind turbine bearings of the present invention;

[0062] Figure 2 It is a schematic diagram of the optimization experiment results in the present invention;

[0063] Figure 3 It is a comparison diagram of the convergence process of the optimization algorithm of the present invention through simulation verification (A), experimental verification (B) and experimental data verification (C);

[0064] Figure 4 The three groups (a), (b), (c), and (d) are respectively comparison diagrams of the decomposition results of AVMMD and MEMD in the present invention in simulation verification, experimental verification, and experimental data verification;

[0065] Figure 5 (a), (b), (c), and (d) are the index comparison diagrams of each component of AMVMD and MEMD verified by simulation in the present invention respectively;

[0066] Figure 6 It is a comparison diagram of the reconstruction results of AVMMD and MEMD verified by simulation and experimental data in the present invention;

[0067] Figure 7 (a) and (b) are respectively the reconstructed component envelope diagrams of AMVMD and MEMD experimentally verified in the present invention;

[0068] Figure 8 (a)(b)(c)(d)(e)(f) are respectively comparison diagrams of the reconstruction results of the two methods AVMMD and MEMD experimentally verified in the present invention;

[0069] Fig. 9 (a)(b)(c)(d)(e)(f) are respectively comparison diagrams of the reconstruction results of the AVMMD and MEMD methods verified by the experimental data of the present invention; DETAILED DESCRIPTION

[0070] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0071] A new method for extracting wind turbine bearing fault features: based on improved adaptive multivariate variational mode decomposition, including the following steps:

[0072] S1. Use logistic-tent chaotic mapping instead of traditional random number generation method to initialize the particle positions in the QPSO algorithm to ensure the uniformity and randomness of initialization, which specifically includes the following sub-steps:

[0073] S11, initialize the QPSO algorithm, including particle population size, particle dimension, maximum number of iterations, convergence threshold, upper and lower bounds of K and α;

[0074] S12, control parameter, set the control parameter r in the logistic-tent chaotic mapping. Appropriate parameters can generate a good chaotic sequence. Usually, the chaotic mapping value r is selected in the range of (0,1), at which time the function is in a chaotic state. When r is 0.5, the function mapping is relatively uniform.

[0075] S13, initialize the seed value, select an initial seed value x0, usually in the range of (0,1). This value is the starting point for generating the chaotic sequence;

[0076] S14, generate a chaotic sequence, where the logistic mapping formula is as follows:

[0077] x n+1 = r·x n ·(1-x n )

[0078] where x n Represents the state after the nth iteration. In each iteration, the next value x is calculated according to the above formula n+1 , until the desired number of values ​​are generated;

[0079] S15. Generate a logistic-tent chaotic mapping sequence. The calculation formula is as follows:

[0080]

[0081] Using chaotic mapping instead of random numbers to initialize the population makes the initialization more uniform and improves the global optimization ability of the algorithm;

[0082] S16, the generated chaotic sequence value x n Map to the position of the particle in the search space and initialize the particle swarm to ensure that each particle has an initial position according to the set number of particles.

[0083] S2. Adaptively select the modal number (K) and penalty factor through the improved QPSO algorithm, use the energy error as the fitness function, perform parameter optimization, decompose the processed signal using MVMD, and obtain multiple modal components representing different frequency components, which specifically includes the following sub-steps:

[0084] S21, initialize a set of particle positions using logistic-tent chaotic mapping, where the horizontal and vertical coordinates of the particles represent K and α respectively;

[0085] S22. Perform MVMD decomposition according to the initialized K and α. MVMD is a non-recursive signal processing method that decomposes the original signal into a series of modal components with limited bandwidth, iteratively searches for each center frequency so that these modes can jointly reproduce the input signal.

[0086] First, the input signal is transformed into an analytical signal through Hilbert transform to obtain its one-sided spectrum and exponential term Multiply and modulate to the corresponding baseband. Calculate the L of the demodulated signal 2 The norm square estimates the bandwidth of each mode, and the established MVMD constrained optimization mathematical model is as follows:

[0087]

[0088] where w k is the center frequency of the kth modal component.

[0089] Secondly, two penalty terms are introduced to transform the constrained variational problem into an unconstrained variational problem, which improves the reconstruction accuracy of the signal under noise interference and ensures that the constraints are strictly met. The formula obtained is as follows:

[0090]

[0091] Finally, the alternating multiplier method is used to update the parameters and divide the complete optimization problem into multiple simple sub-optimization problems, which are transformed into the following sub-optimization problems:

[0092]

[0093] Transforming the above iterative problem into the frequency domain, the formula is as follows:

[0094]

[0095] The center frequency and Lagrangian operator are shown in the formula:

[0096]

[0097] in and x c (t) and λ c Fourier transform of (t); for u k (t) Fourier transform of the n+1th iteration.

[0098] Use the MVMD method to decompose into n modal components, see the formula:

[0099]

[0100] S23. Using energy error as a new fitness function, quantitatively evaluate the decomposition effect of MVMD under different parameters to find the optimal solution for parameter setting. The energy E0 of the original signal x(t) is expressed as:

[0101]

[0102] The sum of the energy of each modal component E total It can be expressed as:

[0103]

[0104] When the decomposed modal components are completely orthogonal to the original signal, the sum of the energies of the modal components E total Equal to the energy E0 of the original signal x(t):

[0105] E total =E1+E2+…+E n =E0

[0106] When the modal components are not completely orthogonal to the original signal, the sum of the energies of the components E total The energy E0 of the target signal is not equal, and the absolute value of the calculated difference is recorded as the energy error |E er |:

[0107] |E er |=|E0-E total |=|E0-(E1+E2+…+E n )|

[0108] Energy error |E er The smaller the | value, the better the orthogonality of the modal components obtained by signal decomposition, and the better the effect of the modal components jointly reproducing the original signal, which can reduce the loss of information carried in the original signal and is more conducive to the subsequent fault feature extraction.

[0109] Calculate the fitness value of each particle |E er |, and save the current minimum value, recorded as min(|E er |);

[0110] S24, update the particle position, calculate the fitness value again and compare it with min(|E er |) comparison, if the current particle fitness value is less than min(|E er |) then update the minimum value, where the formula for updating the particle position is as follows:

[0111] v ij =ωv ij+c1r1(p ij -x ij )+c2r2(g j -x ij );

[0112] S25. Repeat the above steps until the maximum number of iterations is reached or the convergence threshold is reached, stop the optimization and output the best particle position, which is the optimal parameters K and α of MVMD.

[0113] S3, calculate the kurtosis value of each component and the Shannon entropy to select the reconstructed component for subsequent feature extraction, analyze the change of the signal through the Teager energy operator, and perform envelope analysis to extract the fault frequency to identify the specific fault characteristics of the bearing, which specifically includes the following sub-steps:

[0114] S31. Calculate the kurtosis value. The kurtosis is a statistic used to measure the sharpness of a signal, and is usually used to identify impact signals. The higher the value, the more transient information the signal contains. Calculate the kurtosis value of each modal component decomposed by AVMMD, and select the modal component with a higher kurtosis value for subsequent processing;

[0115] S32. Calculate Shannon entropy, which is a measure of the uncertainty of random variables and reflects the complexity and information content of the signal. Calculate the Shannon entropy value for each modal component and select the mode with the lower entropy value;

[0116] S33, reconstructing the modal components, combining the calculation results of the kurtosis value and Shannon entropy, and selecting suitable modal components for reconstruction. These components are usually those that are most representative in fault feature extraction;

[0117] S34. The input signal is converted into an analytical signal through Hilbert transform, and the selected modal components are analyzed using the Teager energy operator to capture the transient characteristics of the signal. The specific fault characteristics of the bearing are identified by analyzing the envelope signal.

[0118] The Teager energy operator is an effective signal processing tool for capturing transient features in signals. First, a representative modal component showing fault-related features is selected from the multivariate variational mode decomposition (MVMD). Then, the Teager energy operator is used to calculate the instantaneous energy of the selected modal component. The calculation formula of the Teager energy operator is as follows:

[0119] Ψ[x(n)]=x(n)2-x(n-1)·x(n+1)

[0120] The feature extraction is performed by observing the Teager energy signal, and then the envelope analysis is performed on the Teager signal. The spectrum features are extracted by Fourier transforming the envelope signal. The extracted envelope signal is analyzed to identify the fault frequency and features.

[0121] according to Figure 3 It can be seen that the improved QPSO algorithm converges faster, is not prone to fall into local optimality, has higher optimization efficiency, and has certain advantages. The decomposition results of the AVMMD method are compared with the MEMD method. Figure 4 As shown in , the AVMMD method can decompose the frequencies of a certain bandwidth into the same modal components, and the aliasing phenomenon of adjacent modal components is effectively improved, which is more suitable for subsequent feature extraction. Figure 6 From the reconstruction result comparison chart, we can see that AVMMD can clearly extract the fault frequencies of the outer and inner rings, while MEMD can only extract the single frequency of the outer ring fault feature, and the inner ring fault feature is completely submerged in the noise.

[0122] In summary, the present invention proposes a new adaptive variational modal decomposition method: using logistic-tent chaotic mapping instead of random number initialization, using energy error as a new fitness function, optimizing the key parameters of MVMD, the intrinsic mode number and the penalty factor, and realizing the adaptive decomposition of the signal. The optimal reconstruction component is selected by the kurtosis value and Shannon entropy, and the fault feature extraction is realized by combining teager demodulation and envelope analysis. The above method is compared with the slime mold optimization algorithm. This method is not easy to fall into the local optimum and has a higher optimization efficiency. Compared with the MEMD method, the experimental results are verified by simulation signals and actual data. This method can clearly extract the frequencies of different types of faults and has certain practical value.

Claims

1. A wind turbine bearing fault feature extraction method based on improved AVMMD, characterized by The following steps are involved: S1. Use logistic-tent chaotic mapping instead of traditional random number generation method to initialize the particle positions in the QPSO algorithm to ensure the uniformity and randomness of initialization; S2, the modal number K and penalty factor are adaptively selected through the improved QPSO algorithm, the energy error is used as the fitness function, the parameters are optimized, and the processed signal is decomposed by MVMD to obtain multiple modal components representing different frequency components; S3. Calculate the kurtosis value and Shannon entropy of each component to select the reconstructed component for subsequent feature extraction. Analyze the signal changes through the Teager energy operator and perform envelope analysis to extract the fault frequency to identify the specific fault characteristics of the bearing.

2. A wind turbine bearing fault feature extraction method based on improved AVMMD according to claim 1, characterized in that Step S1 specifically includes the following steps: S11, initialize the QPSO algorithm, including particle population size, particle dimension, maximum number of iterations, convergence threshold, upper and lower bounds of K and α; S12, control parameter, set the control parameter r in the logistic-tent chaotic mapping, select the chaotic mapping value r in the range of (0,1), then the function is in a chaotic state, when r is 0.5, the function mapping is more uniform; S13, initialize the seed value, select an initial seed value x0, usually in the range of (0,1), this value is the starting point for generating the chaotic sequence; S14, generate a chaotic sequence, where the logistic mapping formula is as follows: x n+1 =r·x n ·(1-x n ) where x n Represents the state after the nth iteration. In each iteration, the next value x is calculated according to the above formula n+1 , until the desired number of values ​​are generated; S15. Generate a logistic-tent chaotic mapping sequence. The calculation formula is as follows: Using chaotic mapping instead of random numbers to initialize the population makes the initialization more uniform and improves the global optimization ability of the algorithm; S16, the generated chaotic sequence value x n Map to the position of the particle in the search space and initialize the particle swarm to ensure that each particle has an initial position according to the set number of particles.

3. The wind turbine bearing fault feature extraction method based on improved AVMMD according to claim 1 is characterized in that Step S2 specifically includes the following steps: S21, initialize a set of particle positions using logistic-tent chaotic mapping, where the horizontal and vertical coordinates of the particles represent K and α respectively; S22, performing MVMD decomposition according to the initialized K and α. MVMD is a non-recursive signal processing method that decomposes the original signal into a series of modal components with limited bandwidth, iteratively searches for each center frequency, and makes these modes jointly reproduce the input signal; S23, using energy error as a new fitness function, quantitatively evaluates the decomposition effect of MVMD under different parameters to find the optimal solution for parameter settings. The energy E0 of the original signal x(t) is expressed as follows: The sum of the energy of each modal component E total It can be expressed as: When the decomposed modal components are completely orthogonal to the original signal, the sum of the energies of the modal components E total Equal to the energy E0 of the original signal x(t): AND total =E1+E2+…+E n =E0 When the modal components are not completely orthogonal to the original signal, the sum of the energies of the components E total The energy E0 of the target signal is not equal, and the absolute value of the calculated difference is recorded as the energy error |E er |: |E er |=|E0-E total |=|E0-(E1+E2+…+E n )| Energy error |E er The smaller the value, the better the orthogonality of the modal components obtained by signal decomposition, and the better the effect of the modal components jointly reproducing the original signal, which can reduce the loss of information carried in the original signal and is more conducive to the subsequent fault feature extraction; Calculate the fitness value of each particle |E er |, and save the current minimum value, recorded as min(E er |); S24, update the particle position, calculate the fitness value again and compare it with min(|E er |) comparison, if the current particle fitness value is less than min(|E er |) then update the minimum value, where the formula for updating the particle position is as follows: v ij =ω vij +c1r1(p ij -x ij )+c2r2(g j -x ij ); S25. Repeat the above steps until the maximum number of iterations is reached or the convergence threshold is reached, stop the optimization and output the best particle position, which is the optimal parameters K and α of MVMD.

4. A wind turbine bearing fault feature extraction method based on improved AVMMD according to claim 3, characterized in that Step S22 specifically includes the following steps: First, the input signal is transformed into an analytical signal through Hilbert transform to obtain its one-sided spectrum and exponential term Multiply, modulate to the corresponding baseband, and calculate the L of the demodulated signal 2 The norm square estimates the bandwidth of each mode, and the established MVMD constrained optimization mathematical model is as follows: where w k is the center frequency of the kth modal component; Secondly, two penalty terms are introduced to transform the constrained variational problem into an unconstrained variational problem, which improves the reconstruction accuracy of the signal under noise interference and strictly satisfies the constraints. The formula is as follows: Finally, the alternating multiplier method is used to update the parameters and divide the complete optimization problem into multiple simple sub-optimization problems, which are transformed into the following sub-optimization problems: Transforming the above iterative problem into the frequency domain, the formula is as follows: The center frequency and Lagrangian operator are shown in the formula: in and x c (t) and λ c Fourier transform of (t); for u k (t) Fourier transform of the n+1th iteration; Use the MVMD method to decompose into n modal components, see the formula:

5. The wind turbine bearing fault feature extraction method based on improved AVMMD according to claim 1 is characterized in that Step S3 specifically includes the following steps: S31, calculating the kurtosis value, calculating the kurtosis value of each modal component decomposed by AVMMD, and selecting the modal component with a higher kurtosis value for subsequent processing; S32, calculating the Shannon entropy, calculating the Shannon entropy value for each modal component, and selecting the modal component with a lower entropy value; S33, reconstructing modal components, combining the calculation results of the kurtosis value and the Shannon entropy, selecting suitable modal components for reconstruction, wherein the modal components are those most representative parts in fault feature extraction; S34. The input signal is converted into an analytical signal through Hilbert transform, and the selected modal components are analyzed using the Teager energy operator to capture the transient characteristics of the signal. The specific fault characteristics of the bearing are identified by analyzing the envelope signal.

6. The method for extracting wind turbine bearing fault features based on improved AVMMD according to claim 5 is characterized in that: Kurtosis is a statistic used to measure the sharpness of a signal and is used to identify impact signals. The higher its value, the more transient information the signal contains. Shannon entropy is a measure of the uncertainty of a random variable and reflects the complexity and amount of information in the signal.

7. The method for extracting wind turbine bearing fault features based on improved AVMMD according to claim 5 is characterized in that: The Teager energy operator is an effective signal processing tool for capturing transient features in signals. First, a representative modal component showing fault-related features is selected from the multivariate variational mode decomposition (MVMD). Then, the Teager energy operator is used to calculate the instantaneous energy of the selected modal component. The calculation formula of the Teager energy operator is as follows: Ψ[x(n)]=x(n)2-x(n-1)·x(n+1) The feature extraction is performed by observing the Teager energy signal, and then the envelope analysis is performed on the Teager signal. The spectrum features are extracted by Fourier transforming the envelope signal. The extracted envelope signal is analyzed to identify the fault frequency and features.

Images