Wind turbine generator anti-impact noise fault identification method based on feature embedding deep learning

Abstract

The invention discloses a wind turbine generator anti-impact noise fault identification method based on feature embedding deep learning, and the method comprises the steps: carrying out the feature mode decomposition of an original vibration signal of a wind turbine generator, screening out an optimal mode component, and converting a time domain signal of the optimal mode component into an envelope spectrum; using the minimum envelope entropy as a fitness function, and using a sparrow search algorithm to globally optimize the filter length and the decomposition modal number of characteristic modal decomposition; and extracting time-frequency domain features, constructing a multi-dimensional time-frequency domain feature vector, inputting the multi-dimensional time-frequency domain feature vector into the combined fault recognition model, and outputting a fault classification result. According to the method, on the basis of an FMD and SSA-MEE joint optimization framework, the sensitivity limitation of a traditional envelope demodulation method on impact noise is broken through, modal aliasing is restrained, and noise robustness is enhanced. According to the method, the vibration signals are subjected to characteristic mode decomposition, the influence of early impact noise is inhibited, a CNN-GRU-Attention fault recognition model is provided, and the accuracy of fault recognition is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault identification methods, and particularly relates to a fault identification method for wind turbine anti-impulse noise based on feature embedding deep learning. Background Technique

[0002] With the rapid growth of global wind power installed capacity, wind power, as a clean and renewable energy form, is of great significance for achieving the carbon neutrality goal. However, with the increase in the number of wind turbines, its operation and maintenance costs also show an upward trend. Early fault identification of wind turbines can effectively reduce the impact of faults and ensure the safe and stable operation of equipment.

[0003] However, in actual operation, the coupling mechanism of fault characteristics under complex working conditions is unclear. Especially for progressive faults such as early wear of generator bearings and pitting of gearboxes, the signal-to-noise ratio of the characterization signals is very low; traditional signal processing methods (such as wavelet transform, empirical mode decomposition, etc.) have the inherent defect of feature extraction distortion under the interference of strong background noise (especially non-stationary impulse noise); conventional deep learning models (such as standard convolutional neural networks) have insufficient feature fusion capabilities for multi-source heterogeneous sensing data, resulting in very low fault classification accuracy in a noisy environment.

[0004] Therefore, how to effectively extract the features of weak fault signals in a noisy environment has become a key issue in improving the performance of wind turbine fault identification systems. Therefore, researching and developing a fault identification method for wind turbine anti-impulse noise based on feature embedding deep learning is of great significance for improving the operation safety of wind turbines and reducing maintenance costs. Summary of the Invention

[0005] The purpose of the present invention is to provide a fault identification method for wind turbine anti-impulse noise based on feature embedding deep learning, which solves the problem of low accuracy of early fault identification of existing wind turbines.

[0006] The technical solution adopted by the present invention is: a fault identification method for wind turbine anti-impulse noise based on feature embedding deep learning, including the following steps:

[0007] Step 1: Perform feature modal decomposition on the original vibration signal of the wind turbine, separate the global and local features in the signal through a multi-scale decomposition strategy, screen out the optimal modal components highly correlated with the early fault impact components, and transform the optimal modal time-domain signal into an envelope spectrum;

[0008] Step 2: Take the minimum envelope entropy of the envelope spectrum obtained in Step 1 as the fitness function, and use the sparrow search algorithm to globally optimize the filter length and decomposition modal number of the feature modal decomposition to obtain the optimized feature modal decomposition parameters;

[0009] Step 3: Using the optimized parameters obtained in Step 2, re - execute the feature mode decomposition to decompose the original vibration signal and extract the time - frequency domain features, and construct a multi - dimensional time - frequency domain feature vector;

[0010] Step 4: Construct a CNN - GRU - Attention joint fault identification model, input the feature vector obtained in Step 3, and output the early fault classification result of the wind turbine.

[0011] The beneficial effects of the present invention are as follows: The wind turbine anti - impact noise fault identification method based on feature - embedded deep learning, based on the FMD and SSA - MEE joint optimization framework, breaks through the sensitivity limitation of the traditional envelope demodulation method to impact noise, suppresses modal aliasing, and enhances noise robustness. By performing feature mode decomposition on the vibration signal, the influence of early impact noise is suppressed, and the traditional one - dimensional convolutional network model is improved by introducing the gated recurrent unit (GRU) and the attention mechanism (Attention), and a CNN - GRU - Attention fault identification model is proposed, which greatly improves the accuracy of fault identification. Description of the Drawings

[0012] Figure 1 is the schematic diagram of the feature mode decomposition algorithm in the present invention;

[0013] Figure 2 is the flow chart of the sparrow search algorithm in the present invention;

[0014] Figure 3 is the overall framework diagram of the wind turbine anti - impact noise fault identification method based on feature - embedded deep learning;

[0015] Figure 4(a) is the time - domain waveform of the original vibration signal of the wind turbine gear with tooth missing;

[0016] Figure 4(b) is the frequency - domain waveform of the original vibration signal of the wind turbine gear with tooth missing;

[0017] Figure 5 is the comparison chart of the fitness optimization curves of the sparrow search algorithm, whale optimization algorithm, and particle swarm algorithm;

[0018] Figure 6(a) is the improved feature mode decomposition result of the tooth - missing signal;

[0019] Figure 6(b) is the time - domain waveform of the optimal mode of the improved feature mode decomposition;

[0020] Figure 7 is the envelope spectrum of the tooth - missing signal after being processed by the improved feature mode decomposition;

[0021] Figure 8(a) is the relationship diagram between the fault identification accuracy of the wind turbine gear and the number of iterations;

[0022] Figure 8(b) is a graph showing the relationship between the loss value of wind turbine gear fault identification and the number of iterations;

[0023] Figure 9 is the confusion matrix of wind turbine gear fault identification. Specific implementation manners

[0024] The present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0025] Embodiment 1

[0026] The present invention provides a method for identifying impact noise faults of wind turbines based on feature embedding deep learning, including the following steps:

[0027] Step 1: Aiming at the non-stationarity, non-linearity of the vibration signals of wind turbines and the low-amplitude and short-time periodic pulse characteristics of early fault features, the Feature Mode Decomposition (FMD) is used to directly perform adaptive decomposition on the original vibration signals without traditional signal envelope preprocessing. The FMD method is based on the impact, periodicity and autocorrelation characteristics of fault signals. By designing an adaptive filter bank with the Correlated Kurtosis (CK) as the optimization objective function, the decomposition parameters are dynamically adjusted to suppress the interference of impact noise on the local extreme value distribution and avoid the modal aliasing and envelope distortion problems caused by abnormal extreme values in traditional envelope demodulation methods. Further, through a multi-scale decomposition strategy, the optimal modal components highly correlated with the early fault impact components are separated from the global and local features of the signal, overcoming the instantaneous pulse interference and modal component mixing defects introduced by signal envelope calculation in traditional methods, and significantly improving the identification accuracy and anti-interference ability of early fault features;

[0028] Step 2: On the basis of the Feature Mode Decomposition (FMD) in Step 1, the Sparrow Search Algorithm (SSA) is introduced to globally optimize the filter length L and the decomposition modal number m of FMD to solve the core influence of parameter selection on the decomposition accuracy, modal aliasing suppression and anti-interference ability. L determines the impact component capture ability of the FMD filter bank, and m needs to be dynamically adapted to the actual frequency components of the signal. Aiming at the low-amplitude and short-time periodic impact characteristics of the early fault signals of wind turbines, the Minimum Envelope Entropy (MEE) is used as the fitness function of SSA. The envelope energy of the effective fault modal components is concentrated near the characteristic frequency, and the entropy value is significantly lower than the random high-entropy distribution of noise and interference components. MEE can quantitatively characterize the significance of the periodic impact characteristics in the decomposed modal components, suppress the instantaneous pulse interference caused by impact noise, and directly reflect the comprehensive influence of the parameter combination on modal orthogonality, aliasing suppression and fault feature identifiability through the joint evaluation of L and m;

[0029] Step 3: Using the optimized parameters obtained in Step 2, re - perform the eigen - mode decomposition on the original vibration signal, extract the time - frequency domain features from the decomposed signal, and construct a multi - dimensional time - frequency domain feature vector.

[0030] Step 4: Aiming at the time - series evolution characteristics of early faults in wind turbines and the challenge of the mixture of effective information and noise in the time - frequency domain features of vibration signals, based on the traditional one - dimensional convolutional neural network (CNN), fuse the gated recurrent unit (GRU) and the attention mechanism (Attention) to construct a CNN - GRU - Attention joint fault recognition model. Input the time - frequency domain features of the vibration signal extracted in Step 3 for early fault recognition of wind turbines. By introducing GRU and the attention mechanism, the inherent defects of traditional CNN in time - series modeling and feature importance quantification are made up for, forming an all - chain optimization of feature extraction - classification decision - making with the improved FMD decomposition method in Steps 1, 2, and 3, and jointly solving the recognition problem caused by the weakness, time - series property, and coupling of impulsive noise interference of early fault signals in wind turbines.

[0031] Example 2

[0032] The present invention provides a method for anti - impact noise fault recognition of wind turbines based on feature - embedding deep learning. On the basis of Example 1, the specific process of Step 1 is preferably as follows:

[0033] Step 1.1: Aiming at the typical frequency - domain characteristics of early fault signals in wind turbines (such as the fault characteristic frequencies of gearbox bearings and their harmonics distributed in 0.1 - 2 kHz, and the high - frequency resonance components caused by blade cracks concentrated in 3 - 5 kHz), combined with the sidelobe suppression advantage of the Hanning window function, conduct directional division of the fault - sensitive frequency band of the original vibration signal and initialize the filter bank. To realize the initialization of the Hanning window, first evenly divide the frequency band of the original signal into K sub - frequency bands, and determine the cut - off frequency parameters of each sub - frequency band based on the signal power spectral density characteristics. The mathematical expression is:

[0034]

[0035] where f s is the sampling frequency of the original signal; f L is the lower cut - off frequency; f H is the upper cut - off frequency.

[0036] Step 1.2: The low-amplitude, short-time periodic pulse characteristics of the early fault signal of the wind turbine are easily submerged by the impact noise, and the impact component of the early fault is similar to the random noise in a single time domain or frequency domain indicator (such as kurtosis and envelope entropy). It is proposed to use the correlated kurtosis (CK) as the optimization target for filter update and period estimation. Due to the lack of periodic regularity, the CK value of the impact noise is significantly lower than the real fault component. By maximizing CK, the noise-dominated modal component can be suppressed. CK can simultaneously evaluate the pulse intensity and periodic regularity by quantifying the high-order cumulative characteristics of the signal at a specific period T. With the correlated kurtosis as the optimization target, in the [f L ,f H Design the kth FIR filter f within the frequency band k , filter update and period estimation are performed, and its mathematical expression is:

[0037]

[0038] where v k (n) is the kth decomposition mode; f k represents the kth FIR filter of length L; t x is the input period measured in number of samples; M is the shift order.

[0039] From the above formula, we can see that the input period t x It has a direct impact on the filter coefficients that need to be updated. In the eigenmode decomposition, t x The autocorrelation spectrum is used for estimation in each iteration. The autocorrelation number R(τ) of the input signal sequence x(n) can be expressed as:

[0040]

[0041] Where N is the length of x(n); τ is the time delay. If R x (τ1) is the local maximum of the autocorrelation spectrum after zero crossing, so the estimated period is T x =τ1.

[0042] Step 1.3: In order to eliminate redundant modes and reduce the amount of calculation, a mode selection strategy is adopted. For the K modal components v output in step 1.2 k (n)(k=1,…i,j…K) performs the following processing and calculates each mode v by the following formula i (n) and v j The correlation coefficient C between (n)(i, j∈[1,K]) ij , determine the maximum C ij The mode, C ijA large value indicates that the two modes contain more identical components:

[0043]

[0044] Repeat the above process until the number of finally retained modes is equal to the preset number of modes m (m < K). Although the above m modes are structurally independent, the significance of fault characteristics may vary greatly. At the same time, in order to select the modes with more fault information, remove the modes with smaller CK values, and finally obtain the optimal mode v with the largest CK value opt (n):

[0045]

[0046] Step 1.4. Faults usually manifest as amplitude modulation and phase modulation. In order to accurately extract the fault characteristic frequency to judge the type of fault, use the Hilbert demodulation analysis method to transform the optimal mode time-domain signal v opt (n) into an envelope spectrum, and analyze and extract the fault characteristic frequency from the envelope spectrum.

[0047] Perform the Hilbert transform on the optimal mode component v opt (t) to obtain its imaginary part signal

[0048]

[0049] Take v opt (t) as the real part, as the imaginary part, and construct the optimal mode analytic signal Z(t):

[0050]

[0051] The modulus (envelope) A(t) and phase φ(t) of the analytic signal Z(t) are respectively:

[0052]

[0053] Perform the Fourier transform on the envelope signal A(t) to obtain the optimal mode envelope spectrum S(f):

[0054]

[0055] By analyzing the positions of the significant spectral peaks in S(f), extract the characteristic frequencies corresponding to the faults of the wind turbine (such as bearing damage, gear crack).

[0056] Example 3

[0057] The present invention provides a method for identifying impact noise faults of wind turbines based on feature embedding deep learning. On the basis of Example 1, the specific process of Step 2 is preferably:

[0058] Aiming at the problem that the early fault signals of wind turbines are vulnerable to parameter sensitivity in empirical mode decomposition (such as mode aliasing or feature loss caused by the deviation in the selection of filter length and the number of modes), a framework for adaptive optimization of FMD parameters based on the sparrow search algorithm (SSA) is proposed, and the minimum envelope entropy (MEE) is used as the fitness function.

[0059] The quasi-periodic impact of early faults is manifested as the sparsity of the envelope sequence, and its envelope entropy (EE) value is significantly lower than that of noise (with uniform energy distribution); the impact noise has a high envelope entropy value due to the lack of a stable periodic pattern, and MEE optimization can suppress the generation of noise-dominated modes; MEE only requires single-channel signal analysis, avoiding the computational complexity of multi-dimensional feature extraction and adapting to the iterative optimization requirements of SSA. Taking the minimum envelope entropy of the optimal mode envelope spectrum obtained in step 1 as the fitness function, the envelope entropy calculation formula is as follows:

[0060]

[0061] where A opt (n) is the envelope amplitude of the signal Z(n), p k is the normalized envelope probability density, and N is the signal length.

[0062] SSA imitates the foraging behavior of sparrows and searches for the global optimal solution through cooperation and competition. The discoverer searches for the optimal solution, and the followers perform local searches around individuals with high fitness. The sparrow search algorithm is used to globally optimize the filter length L and the number of decomposition modes m of empirical mode decomposition. The parameters L and m to be optimized are encoded as the position vector X of the sparrow individuals i = [L i , m i (i = 1, 2,..., N pop , N pop is the population size), and its physical meaning is: L i is the filter length (a positive integer) represented by the i-th sparrow individual; m i is the number of decomposition modes represented by the i-th sparrow individual (m i ≤K max , K max is the maximum number of sub-bands preset in step 1.1).

[0063] The update expression for the position of the discoverer is as follows:

[0064]

[0065] where b is the current iteration number; α ∈ (0, 1); is the position of the i-th sparrow in the j-th dimension; bmax is the maximum number of iterations; M is a 1×d order search step adjustment matrix for parameters; Q is a random number obeying the standard normal distribution; R2 is the warning value, with a value range of [0,1]; S r is the safety value, with a value range of [0,1]. When R2≥S r , the followers reach the safe area along the path of the discoverers. The expression for updating the followers' positions is as follows:

[0066]

[0067] where, is the optimal position of the sparrow population at the (b + 1)-th iteration; is the worst position of the sparrow population at the b-th iteration; A represents a 1×d order matrix with element values of 1 or -1, and A + = A T (AA T ) -1 .

[0068] In addition, the algorithm also introduces a "guard" mechanism, which simulates that some sparrows adjust their positions by avoiding predators to avoid falling into local optima. The formula for updating their positions is:

[0069]

[0070] where, β is a random number obeying the standard normal distribution, used to control the step size of individual movement; R is a random number, representing the direction of movement; ε is a very small constant for controlling the step size parameter to avoid the denominator being zero; f i is the fitness value of the i-th individual; f g is the optimal fitness value; f w is the worst fitness value.

[0071] For each sparrow individual X i = [L i , m i , the following operations are performed: with L i as the filter length and m i as the number of modes, perform the eigenmode decomposition in steps 1.1 - 1.3; calculate the minimum envelope entropy EE of the decomposition result as the fitness value f i = -EE (the smaller the better). When the maximum number of iterations b max is reached or the fitness value converges, output the global optimal solution X best = [L, m].

[0072] Example 4

[0073] The present invention provides a method for identifying impact-noise faults of wind turbine generators based on feature-embedded deep learning. On the basis of Embodiment 1, the specific process of Step 3 is preferably as follows:

[0074] Step 3.1: Substitute the optimized filter length L and the number of decomposition modes m into the feature mode decomposition algorithm, and re-perform feature mode decomposition to decompose the original vibration signal, generating m modal components. Specifically: evenly divide the signal frequency band into K sub-frequency bands according to the filter length L, and calculate the cut-off frequency f of each sub-frequency band L and f H ; within each sub-frequency band, with the maximization of the correlation kurtosis as the goal, extract the modal component v x (n) through filter update and period estimation T k ; screen and retain m modal components, and then select the modal with the largest CK value among the m modes as the optimal mode.

[0075] Step 3.2: Construct a feature sequence input layer. For different working conditions, based on the optimal modal components obtained in 3.1, respectively extract the time-domain features and frequency-domain features of the optimal mode of the vibration signal, and use a multi-dimensional feature fusion strategy to perform characterization analysis on the vibration signal, where the time-frequency domain features are shown in Table 1.

[0076] Table 1 Time-frequency domain features of vibration signals

[0077]

[0078]

[0079] Step 3.3: Concatenate the time-domain feature vector and the frequency-domain feature vector obtained in Step 3.2 to form a multi-dimensional time-frequency domain feature vector.

[0080]

[0081] Embodiment 5

[0082] The present invention provides a method for identifying impact-noise faults of wind turbine generators based on feature-embedded deep learning. On the basis of Embodiment 1, the specific process of Step 4 is preferably as follows:

[0083] In view of the time series correlation, weak feature and noise interference coupling characteristics of the early faults of wind turbines, traditional single deep learning models have limitations, including the separation of local features and global time series modeling, which makes it difficult to capture the long-term time series dependence of fault evolution; noise sensitivity and fuzzy feature weights, which cannot focus on key fault features in the noise background. The GRU can analyze the time series evolution law of fault features and solve the problem of time series information loss caused by the "local receptive field" of the traditional CNN. The attention mechanism can quantify the feature importance. Therefore, a CNN-GRU-Attention fault recognition model is built. Through the feature interaction and dynamic weight allocation mechanism, accurate identification of fault features is realized. Specifically:

[0084] Step 4.1: Adopt a three-layer one-dimensional convolutional structure, and the input is the multi-dimensional time-frequency domain feature vector F generated in Step 3.3 joint , and its main parameter structure is shown in Table 2. The convolutional layer is responsible for extracting input features, performing multi-depth convolution operations on the data with convolutional kernels of odd sizes, and then using activation functions such as relu and leaky-relu to perform non-linear mapping on neurons; the pooling layer summarizes the features obtained by the convolution operation and performs data dimensionality reduction through operations such as maximum pooling and average pooling; the fully connected layer is embedded in the bottom layer of the network in the form of a BP neural network, combines the pooled features, and calculates the results of classification or regression.

[0085] Table 2 Main parameters of CNN

[0086] Input N Size Stride Conv1D_1 128 7×1 1 MaxPooling1D_1 2×128 2 Conv1D_2 64 5×1 1 MaxPooling1D_2 2×64 2 Conv1D_3 32 3×1 1 MaxPooling1D_3 2×32 2

[0087] Step 4.2: Build a gated recurrent unit (GRU) network, input the time series feature sequence output by the CNN into the GRU network of the gated recurrent unit, and use it to capture the time series information of the signal. The GRU consists of an update gate and a reset gate. The update gate z t and the reset gate r t realize their respective functions by applying the σ function to the hidden state information h t-1 at the previous moment and the input x t at this moment respectively. The update gate determines the allocation ratio of the hidden state information h t at the (t - 1)th moment to the hidden state information h t-1 at the tth moment and the candidate state information at the tth moment. The larger the value, the more information from the (t - 1)th moment is retained; the reset gate determines the correlation degree between the candidate state information at the tth moment and the hidden state information h t-1 at the (t - 1)th moment. The smaller the value, the greater the degree of information forgetting at the (t - 1)th moment. The update expressions of each gating mechanism and the hidden state at the mathematical level are as follows:

[0088] zt = σ g (W z x t + U z h t-1 + b z )

[0089] r t = σ g (W r x t + U r h t-1 + b r )

[0090]

[0091] where the learnable parameters include all the weight matrices and the bias vector b ∈ d . These parameters are shared across all time steps and learned through model training. The symbol denotes the dot product of matrices, and d and k represent the dimensional hyperparameters of the input and hidden vectors respectively.

[0092] Step 4.3: To further enhance the model's ability to focus on key information, an attention mechanism is introduced. The attention mechanism automatically focuses on the most important information by assigning different weights to the input features and reduces the interference of irrelevant features.

[0093] To calculate the attention weights, the model uses a fully connected network to take the hidden layer vector obtained from the GRU module as input and calculates the scores of each vector to measure its contribution to the final output. After being activated by the tanh function, these scores result in s i , and then s i is normalized by the softmax function to obtain the weight coefficients a i . Since the output feature vector of the GRU is h i , the context vector c i can be calculated through this vector, thus completing the weighted sum of the GRU output feature vector.

[0094]

[0095] s i = tanh(w T h i )

[0096]

[0097] where w T is a model parameter; s i is for hi and c i degree of correlation with; a i is the attention weight.

[0098] Embodiment 6

[0099] The present invention provides a method for identifying impact noise faults of wind turbine generators based on feature embedding deep learning, which applies an improved feature mode decomposition algorithm and a convolutional network that combines a gated recurrent unit and an attention mechanism. It has strong anti-noise and fault feature extraction capabilities, can effectively identify weak fault signals in complex environments, and effectively improves the accuracy of early fault identification of wind turbine generators. The specific implementation steps are as follows:

[0100] Step 1, as Figure 1 shown, is a schematic diagram of the feature mode decomposition algorithm. The feature mode decomposition algorithm is based on the impact, periodicity, and autocorrelation characteristics of the fault signal. By designing an adaptive filter bank with correlated kurtosis (CK) as the optimization objective function, the decomposition parameters are dynamically adjusted to suppress the interference of impact noise on the local extreme value distribution, and avoid the problems of mode mixing and envelope distortion caused by abnormal extreme values in the traditional envelope demodulation method. Further, through a multi-scale decomposition strategy, the optimal mode components highly correlated with the early fault impact components are separated from the global and local features of the signal, overcoming the instantaneous pulse interference and mode component mixing defects introduced by the signal envelope calculation in the traditional method, and significantly improving the identification accuracy and anti-interference ability of early fault features. The specific process is as follows:

[0101] Step 1.1, Initialize the FIR filter bank. Combining the sidelobe suppression advantage of the Hanning window function, the original vibration signal is directionally divided into fault-sensitive frequency bands and the filter bank is initialized. To implement the Hanning window initialization, first, the frequency band of the original signal is evenly divided into K sub-frequency bands, and based on the signal power spectral density characteristics, the cut-off frequency parameters of each sub-frequency band are determined, and its mathematical expression is:

[0102]

[0103] where f s is the sampling frequency of the original signal; f L is the lower cut-off frequency; f H is the upper cut-off frequency.

[0104] Step 1.2, update the filter and estimate the period. Due to the low amplitude, short-time periodic pulse characteristics of the early fault signal of the wind turbine, the fault component is neither the most periodic nor the most impulsive, but it has the maximum correlation kurtosis (CK) value that takes into account both the periodicity and impulse of the signal. In view of this characteristic, establishing CK as the objective function of filter optimization can more effectively capture the fault characteristics hidden in the signal than traditional methods. This optimization criterion based on multi-dimensional feature fusion significantly improves the identification accuracy of weak characteristic components in mechanical fault diagnosis by balancing the coupling relationship between periodic components and transient shocks. Taking the correlation kurtosis as the optimization target, in step 1.1, [f L ,f H Design the kth FIR filter f within the frequency band k , filter update and period estimation are performed, and its mathematical expression is:

[0105]

[0106] where v k (n) is the kth decomposition mode; f k represents the kth FIR filter of length L; t x is the input period measured in number of samples; M is the shift order.

[0107] From the above formula, we can see that the input period t x It has a direct impact on the filter coefficients that need to be updated. In the eigenmode decomposition, t x The autocorrelation spectrum is used for estimation in each iteration. The autocorrelation number R(τ) of the input signal sequence x(n) can be expressed as:

[0108]

[0109] Where N is the length of x(n); τ is the time delay. If R x (τ1) is the local maximum of the autocorrelation spectrum after zero crossing, so the estimated period is T x =τ1.

[0110] Step 1.3: In order to eliminate redundant modes and reduce the amount of calculation, a mode selection strategy is adopted. For the K modal components v output in step 1.2 k (n)(k=1,…i,j…K) performs the following processing and calculates each mode v by the following formula i (n) and v j The correlation coefficient C between (n)(i, j∈[1,K]) ij , determine the maximum C ij The mode, C ij Large means that the two modes contain more identical components:

[0111]

[0112] Repeat the above process until the number of finally retained modes is equal to the preset number of modes m (m < K). Although the above m modes are structurally independent, the significance of their fault characteristics may vary greatly. At the same time, in order to select the modes with more fault information, remove the modes with smaller CK values, and finally obtain the optimal mode v with the largest CK value. opt (n).

[0113]

[0114] Step 1.4. Faults usually manifest as amplitude modulation and phase modulation. In order to accurately extract the fault characteristic frequency to judge the type of fault, use the Hilbert demodulation analysis method to transform the optimal mode time-domain signal v opt (n) into an envelope spectrum, and analyze and extract the fault characteristic frequency from the envelope spectrum.

[0115] Perform the Hilbert transform on the optimal mode component v opt (t) to obtain its imaginary part signal.

[0116]

[0117] The Cauchy principal value P = 1. Let h(t) = 1 / (πt), perform the Fourier transform to obtain the sign function sgn(ω), and obtain the Hilbert transform frequency:

[0118]

[0119] Take v opt (t) as the real part, as the imaginary part, and construct the optimal mode analytic signal Z(t):

[0120]

[0121] The modulus (envelope) A(t) and phase φ(t) of the analytic signal Z(t) are respectively:

[0122]

[0123] Perform the Fourier transform on the envelope signal A(t) to obtain the optimal mode envelope spectrum S(f):

[0124]

[0125] By analyzing the positions of the significant spectral peaks in S(f), extract the characteristic frequencies corresponding to the faults of the wind turbine (such as bearing damage, gear crack).

[0126] Step 2: As Figure 2 , the sparrow search algorithm (SSA) is used to optimize the FMD parameters. The quasi-periodic shock of early faults is manifested as the sparsity of the envelope sequence, and its envelope entropy (EE) value is significantly lower than that of noise (uniform energy distribution); the impulsive noise has a high envelope entropy value due to the lack of a stable periodic pattern, and the optimization of the minimum envelope entropy (MEE) can suppress the generation of noise-dominated modes; MEE only requires single-channel signal analysis, avoiding the computational complexity of multi-dimensional feature extraction and adapting to the iterative optimization requirements of SSA. Therefore, MEE is used as a signal complexity quantization index, and a fitness evaluation system is constructed through iterative calculation to effectively guide the algorithm to locate the parameter configuration with the optimal decoupling performance in the solution space. Taking the minimum envelope entropy of the optimal modal envelope spectrum obtained in Step 1 as the fitness function, its definition is as follows:

[0127]

[0128] where A opt (n) is the envelope amplitude of the signal Z(n), p opt (n) is the normalized envelope probability density, and N is the signal length.

[0129] SSA simulates the division of labor and cooperation mechanism in the foraging strategy of the sparrow population, and constructs a dynamic balance global optimization mode through the combination of the independent exploration of the discoverer and the follow-up learning of the followers. Among them, the discoverer searches for the optimal solution, and the followers conduct local searches around high-fitness individuals. The sparrow search algorithm is used to globally optimize the filter length L and the number of decomposition modes m of the eigenmode decomposition. The parameters L and m to be optimized are encoded as the position vector X i = [L i , m i (i = 1, 2,..., N pop , N pop is the population size).

[0130] The update expression of the discoverer's position is as follows:

[0131]

[0132] where b is the current iteration number; α ∈ (0, 1); is the position of the i-th sparrow in the j-th dimension; b max is the maximum iteration number; M is a 1×d matrix; Q is a random number obeying the standard normal distribution; R2 is the warning value, with a value range of [0, 1]; S r is the safety value, with a value range of [0, 1]. When R2 ≥ S rWhen the time comes, the followers reach the safe area along the path of the discoverer. The expression for updating the position of the followers is as follows:

[0133]

[0134] Among them, is the optimal position of the sparrow population at the (b + 1)-th iteration; is the worst position of the sparrow population at the b-th iteration; A represents a 1×d matrix with element values of 1 or -1, and A + = A T (AA T ) -1 .

[0135] In addition, the algorithm also introduces a "guard" mechanism, which simulates that some sparrows adjust their positions by avoiding predators, so as to avoid falling into local optima. The formula for updating its position is:

[0136]

[0137] Among them, β is a random number obeying the standard normal distribution, which is used to control the step size of individual movement; R is a random number, which is the direction of movement; ε is a very small constant that controls the step size parameter to avoid the denominator being zero; f i is the fitness value of the i-th individual; f g is the optimal fitness value; f w is the worst fitness value.

[0138] For each sparrow individual X i = [L i , m i , the following operations are performed: with L i as the filter length and m i as the number of modes, run the eigenmode decomposition in steps 1.1 - 1.3; calculate the minimum envelope entropy EE of the decomposition result as the fitness value f i = -EE (the smaller the better). When the maximum number of iterations b max or the fitness value converges, output the global optimal solution X best = [L, m].

[0139] The specific process of step 3 is as follows:

[0140] Step 3.1: Substitute the optimized filter length L and the number of decomposition modes m into the eigenmode decomposition algorithm, and re - execute the eigenmode decomposition to decompose the original vibration signal, generating m modal components. Specifically: According to the filter length L, the signal frequency band is evenly divided into K sub - frequency bands, and calculate the cut - off frequencies f L and f H; within each sub - band, aiming at maximizing the relevant kurtosis, update the filter and estimate the period T x Extract the modal component v k (n); screen and retain m modal components, and then select the modal with the largest CK value among the m modes as the optimal mode.

[0141] Step 3.2: Construct the input layer of the feature sequence. For different working conditions, based on the optimal modal components obtained in 3.1, extract the time - domain features and frequency - domain features of the optimal mode of the vibration signal respectively, and use the multi - dimensional feature fusion strategy to perform characterization analysis on the vibration signal. The time - frequency domain features are shown in Table 1.

[0142] Step 3.3: Concatenate the time - domain feature vector and the frequency - domain feature vector obtained in Step 3.2 to form a multi - dimensional time - frequency domain feature vector.

[0143] Step 4: As Figure 3 , on the basis of the traditional one - dimensional convolutional neural network (CNN), fuse the gated recurrent unit (GRU) and the attention mechanism (Attention) to construct a CNN - GRU - Attention joint fault recognition model. By introducing GRU and the attention mechanism, the inherent defects of the traditional CNN in time - series modeling and feature importance quantification are made up for. Input the time - frequency domain features of the vibration signal to realize the full - chain optimization of feature extraction - classification decision - making, and jointly complete the early fault recognition of wind turbines. Build a CNN - GRU - Attention fault recognition model to solve the problems of time - series correlation, weak feature, and noise interference coupling in the early faults of wind turbines. Traditional single deep - learning models are difficult to capture the long - term time - series dependencies of fault evolution and cannot focus on key fault features in the presence of noise. While GRU can analyze the time - series evolution law of fault features, and the attention mechanism can quantify the feature importance. Therefore, construct a CNN - GRU - Attention hybrid model, and through the feature interaction and dynamic weight allocation mechanism, achieve the accurate recognition of fault features. Specifically:

[0144] Step 4.1: A hierarchical (3-layer) one-dimensional convolutional processing architecture is constructed. The detailed core parameter configuration is shown in Table 2. The feature extraction layer uses a multi-scale convolutional kernel group to perform deep feature capture, conducts sliding window analysis on the signal based on the parameter sharing mechanism, constructs a multi-level abstract feature space through an odd-sized filter bank, and introduces non-linear transformation units such as ReLU and its improved version Leaky-ReLU to enhance features. The feature compression layer applies an adaptive downsampling strategy, and performs spatial compression processing on the features using modes such as extreme value retention (max pooling) and regional averaging (mean pooling), optimizing the computational efficiency while maintaining the significance of key features. The feature fusion layer adopts a fully connected topological structure, performs weighted aggregation on high-level semantic features through a backpropagation optimization mechanism, and finally constructs a non-linear mapping channel from the original signal to the target output to achieve end-to-end recognition results.

[0145] Step 4.2: A gated recurrent unit (GRU) network is constructed. The sequential feature sequence output by the CNN is input into the GRU network to capture the sequential information of the signal. The GRU consists of an update gate and a reset gate. The update gate z t and the reset gate r t achieve their respective functions by applying the σ function to the hidden state information h t-1 at the previous moment and the input x t at the current moment respectively. The update gate determines the allocation ratio of the hidden state information h t at the (t - 1)th moment to the hidden state information h t-1 at the tth moment and the candidate state information at the tth moment. The larger the value, the more information from the (t - 1)th moment is retained; the reset gate determines the correlation degree between the candidate state information at the tth moment and the hidden state information h t-1 at the (t - 1)th moment. The smaller the value, the greater the degree of information forgetting at the (t - 1)th moment. The update expressions of each gating mechanism and the hidden state at the mathematical level are as follows:

[0146] z t = σ g (W z x t + U z h t-1 + b z )

[0147] r t = σ g (W r x t + U r h t-1 + b r )

[0148]

[0149] The learnable parameters include all the weight matrices and the bias vectors These parameters are shared across all time steps and are learned through model training. The symbol denotes the dot product of matrices, where d and k represent the dimensional hyperparameters of the input and hidden vectors respectively.

[0150] Step 4.3: To further enhance the model's ability to focus on key information, an attention mechanism is introduced. The attention mechanism automatically focuses on the most important information by assigning different weights to the input features and reduces the interference of irrelevant features.

[0151] To calculate the attention weights, the model uses a fully connected network to take the hidden layer vectors obtained from the GRU module as input, calculates the scores of each vector to measure its contribution to the final output. After being activated by the tanh function, these scores result in s i , and then s i is normalized by the softmax function to obtain the weight coefficients a i . Since the output feature vector of the GRU is h i , the context vector c i can be calculated through this vector, thus completing the weighted sum of the GRU output feature vectors.

[0152]

[0153] s i = tanh(w T h i )

[0154]

[0155] where w T is a model parameter; s i is the correlation degree between h i and c i ; a i is the attention weight.

[0156] Through the above method, the wind turbine anti-impulse noise fault identification method based on feature embedding deep learning, based on the FMD and SSA-MEE joint optimization framework, breaks through the sensitivity limitation of traditional envelope demodulation methods to impulse noise, suppresses modal aliasing, and enhances noise robustness. By performing feature modal decomposition on vibration signals, the influence of early impulse noise is suppressed, and the traditional one-dimensional convolutional network model is improved by introducing a gated recurrent unit (GRU) and an attention mechanism (Attention), and a CNN-GRU-Attention fault identification model is proposed, which greatly improves the accuracy of fault identification. In addition, it also has the following characteristics:

[0157] 1) The present invention breaks through the efficiency of parameter adaptive optimization. Through the global optimization mechanism of the SSA algorithm, the sensitive dependence of the FMD parameters (L, m) on the decomposition performance is solved. The MEE fitness function quantifies the entropy difference between the periodicity of fault impulses and the randomness of noise, effectively avoiding local optimal solutions globally, thereby finding the most suitable parameter combination, ensuring the efficiency and accuracy of signal decomposition.

[0158] 2) The input of the identification of the present invention is the time-frequency domain features of vibration signals, rather than time series data, which improves the input quality and thus ensures the generality and accuracy of the identification.

[0159] 3) Compared with the traditional neural network fault identification model, the CNN-GRU-Attention hybrid model overcomes the defects of traditional models in focusing on time series evolution and key features through multi-modal architecture design.

[0160] Result Analysis

[0161] The present invention is verified by the planetary gearbox fault experimental data collected from the wind turbine drive system test bench. The experiment collects vibration acceleration signals through vibration sensors, and sets 5 health conditions, namely normal condition, gear tooth breakage, gear wear, tooth root crack and missing tooth. The sampling frequency is 48 kHz, the rotation frequency is 20 Hz, and the meshing frequency is 436.56 Hz.

[0162] (1) Taking the missing tooth of the gear as a typical example, analyze the fault characteristic frequency

[0163] Taking the missing tooth of the gear as the analysis object, Figure 4 shows the time-domain waveform and frequency-domain waveform of the sampled signal. It can be found from Figure 4(a) that the acceleration amplitude at the gearbox measurement point has been at a relatively high level, and its impact amplitude is very large. And it can be found that although obvious impact signals appear in both the time-domain and frequency-domain diagrams, there are a large number of resonance spectral peak components in the spectrum of Figure 4(b), so the fault frequency cannot be judged. The periodic characteristics are submerged in strong background noise, making it difficult to identify.

[0164] Therefore, three optimization algorithms (SSA, Whale Optimization Algorithm WOA, Particle Swarm Optimization Algorithm PSO) are used to optimize the parameters L and m in FMD. Figure 5 The curves of the fitness function of FMD under three optimization algorithms changing with the population evolution iteration number are given. It can be found that the SSA, WOA, and PSO algorithms converge after the 4th, 5th, and 6th iterations respectively, and their fitness values are 0.34074, 0.36726, and 0.35238 respectively. The results show that the SSA algorithm has a stronger search ability, and its corresponding optimal parameter combination is (100, 5). Therefore, the filter length parameter is set to L = 100, and the decomposition mode number m = 5. Subsequently, the optimized parameters are used to guide the FMD algorithm to decompose the vibration signal, and the decomposed modes are obtained. The optimal mode is selected, as shown in Figure 6.

[0165] Figure 6(a) shows the five modes after decomposition. The high-frequency impact details are successively stripped off, the amplitude decays gradually, and the frequency complexity increases. The decomposed modes effectively separate the fault impact characteristics and background noise mixed in the signal. Figure 6(b) is the time-domain waveform of the optimal mode. It can be seen from Figure 6(b) that the original signal has high-amplitude fluctuations and many spikes, contains a large amount of high-frequency noise, and it is difficult to observe obvious periodicity from the figure. After empirical mode decomposition, in the optimal mode selected by the maximum correlation in the article, the amplitude of the noise is significantly lower, and the fluctuation is smoother. Most of the noise is effectively filtered out, the signal is clearer, and the main components in the signal are highlighted. Although there is still a certain amount of fluctuation, its periodicity is easier to observe. Then perform Hilbert transform on it. The envelope spectrum is obtained as Figure 7 shown. Multiple obvious frequency peaks can be observed, and these frequencies are f n1 , 2f n1 , 3f n1 , 4f n1 , 5f n1 and 6f n1 . This indicates that during the operation of the gear, due to gear tooth missing, the energy of the vibration signal is significantly enhanced at the fundamental frequency and its multiples. It is speculated that the tooth missing fault is due to the periodic impact caused by the missing tooth part during the gear meshing process, resulting in a significant increase in the frequency doubling component of the envelope demodulation in the vibration signal. From the characteristic frequency distribution, it shows that the dynamic performance of the gear system is greatly affected.

[0166] (2) Gear fault identification

[0167] Process the signals under the remaining working conditions according to the same steps as the missing tooth fault. After processing, for each type of working condition, 7,170 groups of vibration signal samples are respectively sampled, and each group of samples contains 2,048 data points. The data composition is shown in Table 3, where the ratio of the training set to the test set is set at 0.8 and 0.2. The present invention constructs a CNN-GRU-Attention fault recognition model under the Kears framework of the Python compilation platform.

[0168] Table 3 Composition of gear fault data

[0169] Fault type Sample length Training set / Test set Fault label Broken gear tooth 2048 5736 / 1434 0 Tooth root crack 2048 5736 / 1434 1 Normal 2048 5736 / 1434 2 Gear wear 2048 5736 / 1434 3 Missing gear tooth 2048 5736 / 1434 4

[0170] Train the fault features of the vibration signals through the CGA model and conduct fault recognition. The neural network model has the ability to automatically extract features, but has high requirements for the quality of the input data. Therefore, the input is the feature sequences under 5 working conditions. In order to evaluate the performance of the fault recognition model, two quantitative indicators, accuracy and loss value, are introduced.

[0171]

[0172] where T p is the number of positive samples correctly predicted; T n is the number of negative samples correctly predicted; F p is the number of positive samples wrongly predicted; F n is the number of negative samples wrongly predicted.

[0173]

[0174] where n is the total number of samples; k is the total number of categories; y ij is the true label of the i-th sample in the j-th category; is the predicted probability of the i-th sample in the j-th category.

[0175] The accuracy rates during the training and testing processes of the model are shown in Fig. 8(a), and its loss values are shown in Fig. 8(b). As can be seen from Fig. 8(a), the accuracy rates of the training samples and the testing samples increase with the increase of the number of iterations, and finally the accuracy rate curve of the test set gradually stabilizes at about 97.1%. As can be seen from Fig. 8(b), the method used in the present invention has an extremely fast convergence speed and strong stability, and to a certain extent, can suppress the overfitting of the model. When the number of iterations increases, the loss values of the training set and the test set continuously decrease and approach 0.

[0176] Finally, the confusion matrix of the recognition results of the CNN-GRU-Attention model for the fan gear faults is as Figure 9As shown, it intuitively demonstrates the classification results of the model under five different categories. The numbers on the diagonal of the matrix represent the number of correctly classified samples. Generally speaking, the classification accuracy of the model is relatively high, and in most cases, the samples can be accurately classified.

[0177] Table 4 Comparison of the performance of different models

[0178]

[0179]

[0180] Table 4 compares the recognition results of the method of the present invention with those of CNN-LSTM, traditional 1DCNN, SVM, and deep neural network (DNN). The average accuracy of fault recognition of the present invention is the highest, reaching 97.1%. It is 7.4% higher than the average accuracy of CNN-LSTM, 10.9% higher than that of 1DCNN and DNN, and 13.3% and 20.45% higher than the accuracy of the SVM and DNN fault recognition models respectively.

Claims

1. A wind turbine anti-impact noise fault identification method based on feature embedding deep learning, characterized in that: The following steps are involved: Step 1: Perform characteristic mode decomposition on the original vibration signal of the wind turbine, separate the global and local features in the signal through a multi-scale decomposition strategy, screen out the optimal mode component that is highly correlated with the early fault impact component, and transform the optimal mode time domain signal into an envelope spectrum; Step 2: Taking the minimum envelope entropy of the envelope spectrum obtained in step 1 as the fitness function, the sparrow search algorithm is used to globally optimize the filter length and the number of decomposed modes of the eigenmode decomposition to obtain the optimized eigenmode decomposition parameters; Step 3: Using the optimization parameters obtained in step 2, re-execute the eigenmode decomposition to decompose the original vibration signal and extract the time-frequency domain features, and construct a multi-dimensional time-frequency domain feature vector; Step 4: Build a CNN-GRU-Attention joint fault recognition model, input the feature vector obtained in step 3, and output the early fault classification result of the wind turbine.

2. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning according to claim 1, characterized in that: The step 1 specifically comprises the following steps: Step 1.1: According to the frequency domain characteristics of typical faults of wind turbines and the sidelobe suppression advantage of the Hanning window function, the frequency band of the original vibration signal is evenly divided into K sub-bands, and the cutoff frequency of each sub-band is calculated: In the formula, f L is the lower cutoff frequency; f H is the upper cutoff frequency; f s is the sampling frequency of the original vibration signal; Step 1.2: Taking the relevant kurtosis as the optimization target, in step 1.1, the [f L ,f H Design the kth FIR filter f within the frequency band k , perform filter update and period estimation, and maximize the relevant kurtosis CK value to capture the periodicity and impulse characteristics of the fault impact: In the formula, v k (n) is the kth decomposition mode; f k represents the kth FIR filter of length L; T x is the input period measured using the number of samples; M is the shift order; the autocorrelation number R(τ) of the input signal sequence x(n) is expressed as: Where N is the length of x(n); τ is the time delay. If R x (τ1) is the local maximum of the autocorrelation spectrum after zero crossing, so the estimated period is T x =τ1; Step 1.3: For the K modal components v output from step 1.2 k (n) Perform the following processing, k = 1, 2, 3 ... i, j ... K, and calculate each mode v by the following formula i (n) and v j (n) Correlation coefficient C ij ,i,j∈[1,K],select the one with the largest C ij The modality: Repeat the above process until the number of remaining modes is equal to the preset value m, where m < K. Then, perform a secondary screening based on the CK value to select the single mode with the largest CK value, which can best represent the fault impact, that is, the optimal mode v opt (n): Step 1.4: Use the Hilbert demodulation analysis method to convert the optimal modal time domain signal v output in step 1.3 into opt (n) is transformed into an envelope spectrum and the fault characteristic frequency is extracted from it.

3. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning according to claim 2 is characterized in that: The step 1.4 is specifically as follows: For the optimal modal component v opt (t) Perform Hilbert transform to obtain the imaginary signal v opt (t) as the real part, As the imaginary part, construct the optimal modal analysis signal Z(t): The modulus value of the analytical signal Z(t), i.e., the envelope signal A(t) and the phase φ(t), are respectively: Perform Fourier transform on the envelope signal A(t) to obtain the optimal modal envelope spectrum S(f): By analyzing the significant spectral peak positions in S(f), the characteristic frequencies corresponding to wind turbine faults are extracted.

4. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning as claimed in claim 3 is characterized in that: The step 2 specifically includes the following steps: Step 2.1: The minimum envelope entropy of the envelope spectrum S(f) obtained in step 1 is used as the fitness function, where the calculation formula of the minimum envelope entropy is expressed as: In the formula, A opt (n) is the envelope amplitude of the signal Z(n), p opt (n) is the normalized envelope probability density, N is the signal length; Step 2.2: Use the sparrow search algorithm to globally optimize the filter length L and the number of decomposed modes m of the characteristic mode decomposition.

5. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning as claimed in claim 4 is characterized in that: The step 2.2 is specifically as follows: The sparrow search algorithm imitates the foraging behavior of sparrows and seeks the global optimal solution through cooperation and competition. The discoverer seeks the optimal solution, and the follower performs local search around the high fitness individuals. The parameters to be optimized L and m are encoded as the position vector X of the sparrow individual. i =[L i ,m i ],i=1,2,...,N pop , N pop is the population size; L i is the filter length representing the i-th sparrow individual; m i is the number of decomposition modes represented by the i-th sparrow individual, m i ≤K max , K max The maximum number of sub-bands preset for step 1.1; The expression for updating the discoverer's position is as follows: Where b is the number of iterations at the current stage; α∈(0,1); is the position of the i-th sparrow in the j-th dimension; b max is the maximum number of iterations; M is a 1×d-order parameter search step adjustment matrix; Q is a random number that obeys the standard normal distribution; R2 is the warning value, which is [0,1]; S r is a safe value, ranging from [0,1]; when R2≥S r When , the follower follows the path of the discoverer to reach the safe area, and the follower position update expression is as follows: In the formula, is the optimal position of the sparrow population in the b+1th iteration; is the worst position of the sparrow population in the bth iteration; A is represented by a 1×d-order matrix with element values ​​1 or -1, and A + =A T (AA T ) -1 ; It also simulates some sparrows adjusting their positions by escaping predators. The position update formula is: In the formula, β is a random number that follows a standard normal distribution and is used to control the step length of individual movement; R is a random number, which is the direction of movement; ε is a very small constant that controls the step size parameter to avoid the denominator being zero; f i is the fitness value of the ith individual; f g is the optimal fitness value; f w is the worst fitness value; For each sparrow individual X i =[L i ,m i ] Perform the following operations: i is the filter length, m i is the number of modes, run the characteristic mode decomposition of steps 1.1 to 1.3; calculate the minimum envelope entropy EE of the decomposition result as the fitness value f i =-EE, reaching the maximum number of iterations b max Or when the fitness value converges, output the global optimal solution X best =[L, m].

6. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning according to claim 5, characterized in that: The step 3 is specifically as follows: Step 3.1, substitute the optimized filter length L and the number of decomposed modes m into the characteristic mode decomposition algorithm, re-execute the characteristic mode decomposition to decompose the original vibration signal, and generate m modal components; specifically: divide the signal frequency band into K sub-bands evenly according to the filter length L, calculate the cutoff frequency f of each sub-band L With f H ; In each sub-band, with the goal of maximizing the correlation kurtosis, the filter is updated and the period is estimated by T x Extract the modal component v k (n); screen and retain m modal components, and then select the mode with the largest CK value among the m modes as the optimal mode; Step 3.2, extracting time domain features from the time domain signal of the modal component obtained in step 3.1 to obtain a time domain feature vector; extracting frequency domain features from the frequency domain signal of the modal component obtained in step 3.1 to obtain a frequency domain feature vector; Step 3.3: Concatenate the time domain feature vector obtained in step 3.2 with the frequency domain feature vector to form a multidimensional time-frequency domain feature vector F joint : In the formula, μ opt is the mean of the time domain features, PP opt is the peak-to-peak value of the time domain feature, Kurtosis opt is the kurtosis value of the time domain feature, RMS opt is the root mean square of the time domain feature, Shape-factor opt is the shape factor of the time domain feature, Impluse-factor opt is the impact factor of the time domain characteristics, Crest-factor opt is the peak factor of the time domain feature, is the dominant frequency of the frequency domain feature, is the spectral energy of the frequency domain feature, is the spectral entropy of the frequency domain feature.

7. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning according to claim 6, characterized in that: The CNN-GRU-Attention joint fault recognition model constructed in step 4 specifically includes the following steps: Step 4.1: The convolutional neural network CNN adopts a three-layer one-dimensional convolution structure, and the input is the multi-dimensional time-frequency domain feature vector F generated in step 3.3 joint The convolution layer is responsible for extracting input features, performing multi-depth convolution operations on the data with odd-sized convolution kernels, and then using activation functions to perform nonlinear mapping on neurons; the pooling layer summarizes the features obtained by the convolution operation, and then performs data dimension reduction through pooling operations; the fully connected layer is embedded in the bottom layer of the network in the form of a BP neural network, merges the pooled features, and calculates the classification or regression results; Step 4.2: Input the temporal feature sequence output by CNN into the gated recurrent unit GRU network. The GRU network consists of an update gate and a reset gate. The update gate z t and reset gate r t By applying the σ function to the hidden state information h of the previous moment t-1 And the input x at this moment t To realize their respective functions, the update gate determines the hidden state information h at the current time t t The hidden state information h at time t-1 t-1 and the candidate state information at time t The larger the value, the more information is retained at time t-1; the reset gate determines the candidate state information at time t and the hidden state information h at time t-1 t-1 The smaller the value, the greater the degree of information forgotten at time t-1; Step 4.3: The hidden state sequence h of GRU output t We use the attention mechanism to perform weighted fusion and automatically focus on the most important information by assigning different weights to input features, and reduce the interference of irrelevant features. When calculating the attention weight, the hidden layer vector obtained from GRU is used as input through the fully connected network to calculate the scores of each vector. After the scores are activated by the tanh function, the output feature vector h of GRU is obtained. i and the environment vector c i The degree of relevance i , and then use the softmax function to i Perform normalization to obtain the attention weight coefficient a i , complete the weighted summation of the GRU output feature vector.

8. The wind turbine anti-impact noise fault identification method based on feature embedding deep learning according to claim 7, characterized in that: The parameter configuration of the convolutional neural network CNN in step 4.1 is: Layer: Conv1D_1; Dimension: 128; Size: 7×1; Stride: 1; Tier: MaxPooling1D_1; Size: 2×128; Step: 2; Layer: Conv1D_2; Dimension: 64; Size: 5×1; Step length: 1; Tier: MaxPooling1D_2; Size: 2×64; Step: 2; Layer: Conv1D_3; Dimension: 32; Size: 3×1; Step length: 1; Tier: MaxPooling1D_3; Size: 2×32; Step: 2.

Images