A wind turbine gearbox fault diagnosis method and system

Abstract

This invention discloses a method and system for fault diagnosis of wind turbine gearboxes, comprising: performing VMD variational mode decomposition on the original vibration signal to obtain intrinsic mode components (IMFs), and taking the correlation coefficient ρ. i The largest intrinsic mode component (IMF) is subjected to Hilbert transform to obtain the Hilbert-Huang spectrum. The Hilbert-Huang spectra corresponding to each original vibration signal are divided into source domain data samples and target domain data samples. A deep residual network model is constructed, and the deep residual network model is trained using the source domain data samples and the target domain data samples to obtain a transfer diagnostic model for the target domain. The target domain test data samples are input into the transfer diagnostic model to obtain the diagnostic results. The MK-MMD loss function is used to adjust the deep residual network model so that the model can achieve high-precision fault diagnosis with only a small number of samples.

Description

Technical Field

[0001] This invention belongs to the field of gearbox fault diagnosis, specifically relating to a method for fault diagnosis of wind turbine gearboxes based on improved deep residual networks and transfer learning. Background Technology

[0002] Gearboxes belong to the category of rotating machinery. With the rapid development of industrial technology, people have put forward high reliability requirements for the operation of general rotating machinery systems. However, in actual working conditions, gearboxes are widely used in large and complex mechanical equipment such as engineering machinery. Low speed, heavy load and harsh working environment often lead to serious failures in key parts of the gearbox. Therefore, it is crucial to study effective fault diagnosis models for gearboxes to ensure their safe and reliable operation.

[0003] Gearboxes are widely used in large and complex mechanical equipment such as engineering machinery. Low-speed, heavy-load, and harsh working environments often lead to serious failures in critical components of the gearbox. Furthermore, strong background noise drowns out many effective vibration signals, making fault detection difficult. In addition, in actual gearbox operation, normal data samples are often the majority, while typical fault-labeled data samples are few. Moreover, most collected vibration signals are not labeled, and manual labeling is time-consuming and labor-intensive. Summary of the Invention

[0004] This invention provides a method and system for diagnosing gearbox faults in wind turbines, which solves the problems of existing technologies where models are difficult to train, cannot converge, and have low fault diagnosis accuracy under small sample conditions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The first aspect of this invention provides a method for diagnosing gearbox faults in wind turbine generators, comprising:

[0007] The original vibration signals of various types of gears are acquired; the original vibration signals are subjected to VMD variational mode decomposition to obtain intrinsic mode components (IMFs);

[0008] Calculate the correlation coefficient ρ between the intrinsic modal components (IMF) and the original vibration signal. i Take the correlation coefficient ρ i The largest intrinsic mode component (IMF) is subjected to Hilbert transform to obtain the Hilbert-Huang spectrum; the Hilbert-Huang spectrum corresponding to each original vibration signal is divided into source domain data samples and target domain data samples;

[0009] A deep residual network model is constructed, and the deep residual network model is trained using source domain data samples and target domain data samples to obtain a transfer diagnostic model for the target domain; the target domain test data samples are input into the transfer diagnostic model to obtain the diagnostic results.

[0010] Furthermore, methods for obtaining intrinsic mode components (IMFs) by performing VMD variational mode decomposition on the original vibration signal include:

[0011] VMD variational mode decomposition is performed on the original signal to obtain K modes. In each mode The sum equals the original signal as a constraint; each mode center frequency and mode A constrained variational model is constructed with the minimum as the optimization objective, expressed by the following formula:

[0012]

[0013]

[0014] In the formula, δ(t) is the impulse function. t represents time; j represents the imaginary unit; "*" indicates convolution; Represented as the mode in the nth iteration The center frequency; This represents the partial derivative with respect to time; x(t) represents the original vibration signal.

[0015] The constrained variational model is solved by introducing the Lagrange multiplier λ and the quadratic penalty factor α, and by iteratively solving the model using the alternating direction multiplier algorithm. The Lagrange multiplier λ is updated after each iteration, as expressed by the following formula:

[0016]

[0017] Where Γ represents the noise margin, λ n λ is represented as the Lagrange multiplier in the nth iteration; n-1 Represented as the Lagrange multiplier in the (n-1)th iteration;

[0018] The optimal mode is output when the convergence condition is met. As intrinsic mode components (IMFs), the convergence conditions include:

[0019]

[0020] Where ε represents the precision setting value; Represents the square of the 2-norm of a matrix or determinant.

[0021] Furthermore, the correlation coefficient ρ between the intrinsic modal components (IMF) and the original vibration signal is calculated.i The calculation formula is:

[0022]

[0023] In the formula, ρ i The correlation coefficient between the intrinsic modal components (IMF) and the original vibration signal is expressed as x, where L is the length of the original vibration signal; h′ (t) represents the original vibration signal of segment h′; This is represented as the mean of the original vibration signal; R is expressed as the mean of the intrinsic mode components (IMFs); i,h′ It is represented as the intrinsic mode component (IMF) of the k-th segment h′.

[0024] Furthermore, methods for obtaining Hilbert-Huang spectra by performing Hilbert transform on intrinsic mode components (IMFs) include:

[0025] Take the correlation coefficient ρ i The largest intrinsic mode component (IMF), denoted as component I(t), is given by the Hilbert transform formula for its complex conjugate function y(t):

[0026]

[0027] In the formula, P is the value of Cauchy's principle;

[0028] Based on the component I(t) and the complex conjugate function y(t), the corresponding analytic signal z(t) is constructed, expressed by the following formula:

[0029] z(t)=I(t)+jy(t)=a(t)e jθ(t)

[0030]

[0031]

[0032] The reconstructed signal Z(t) is obtained by reconstructing the analytic signal z(t), and the formula is as follows:

[0033]

[0034]

[0035] Reconstruct the signal Let Z(t) be the real part of the reconstructed signal, expressed as: Based on reconstructed signal Construct the Hilbert-Huang spectrum.

[0036] Furthermore, the deep residual network model sequentially includes an input layer, a depthwise separable convolutional layer, a first ReLU+BN layer, a pointwise convolutional layer, a second ReLU+BN layer, a nonlinear transformation layer, and an output layer; the input features of the input layer are identically mapped to the output layer, expressed by the following formula:

[0037] H(x′)=F(x′)+x′

[0038] In the formula, x′ represents the input feature identity mapping value of the input layer; F(x′) represents the residual mapping value; and H(x′) represents the feature of the deep residual network model.

[0039] Furthermore, the soft and hard thresholds are combined and inserted as a nonlinear transform layer between the second ReLU+BN layer and the output layer. The formula for calculating the soft and hard thresholds is as follows:

[0040]

[0041] In the formula, x * y is the input feature of the nonlinear transform layer. * τ represents the output characteristics of the nonlinear transform layer, τ is the threshold, and δ is the adjustment parameter.

[0042] Furthermore, methods for obtaining a transfer diagnostic model for the target domain by training a deep residual network model using source domain data samples and target domain data samples include:

[0043] The original vibration signals corresponding to the source domain data samples are obtained from public datasets under various gear working conditions. The source domain data samples are used to train and debug a deep residual network model to obtain the optimal source domain fault diagnosis model. The weight parameters of the source domain fault diagnosis model are transferred to a new deep residual network model to obtain the target domain fault diagnosis model.

[0044] The original vibration signals corresponding to the target domain data samples are data generated in the actual operation of the wind turbine gearbox. The target domain data samples are used to train the target domain fault diagnosis model. The parameters of the target domain fault diagnosis model are adjusted by the MK-MMD loss function, and finally the transfer diagnosis model is obtained.

[0045] Furthermore, the calculation formula for the MK-MMD loss function is as follows:

[0046]

[0047] In the formula, θ represents the parameters in the target domain fault diagnosis model; n t D represents the number of data samples in the target domain. MK-MMD The difference between the target domain data samples and the source domain data samples; J(·) is the cross-entropy function; λ MK-MMDWeights for the results calculated by the MK-MMD function; Represented as target domain data sample; This represents the diagnostic results for the target domain data sample.

[0048] In a second aspect, the present invention provides a wind turbine gearbox fault diagnosis system, comprising:

[0049] The acquisition module is used to acquire the original vibration signals of various types of gears; the original vibration signals are subjected to VMD variational mode decomposition to obtain intrinsic mode components (IMFs);

[0050] The feature extraction module is used to calculate the correlation coefficient between the intrinsic modal components (IMFs) and the original vibration signals. The IMF with the largest correlation coefficient is selected and subjected to Hilbert transform to obtain the Hilbert-Huang spectrum. The Hilbert-Huang spectrum corresponding to each original vibration signal is divided into source domain data samples and target domain data samples.

[0051] The training module is used to build a deep residual network model. It uses source domain data samples and target domain data samples to train the deep residual network model to obtain a transfer diagnostic model for the target domain.

[0052] The diagnostic module is used to input the target domain data sample into the transfer diagnostic model to obtain the diagnostic results.

[0053] In a third aspect, the present invention provides an electronic device including a storage medium and a processor; the storage medium is used to store instructions; characterized in that the processor is used to operate according to the instructions to execute the method described in the first aspect.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] This invention obtains intrinsic mode components (IMFs) by performing VMD variational mode decomposition on the original vibration signal, selects the IMF with the largest Pearson correlation coefficient and performs HHT transformation to obtain its Hilbert-Huang spectrum, which can extract information features more comprehensively; trains a deep residual network model using source domain data samples and target domain data samples to obtain a transfer diagnostic model for the target domain; and adjusts the deep residual network model using the MK-MMD loss function so that the model can achieve high-precision fault diagnosis with only a small number of samples. Attached Figure Description

[0056] Figure 1 This is a flowchart of the wind turbine gearbox fault diagnosis method provided in Example 1;

[0057] Figure 2 This is a diagram of the residual block structure in the residual network of Example 1;

[0058] Figure 3 This is a structural diagram of the depth-separable convolutional layer in Example 1;

[0059] Figure 4 This is a structural diagram of the deep residual network model provided in Example 1. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0061] Example 1

[0062] like Figures 1 to 4 As shown, a method for diagnosing gearbox faults in wind turbines includes:

[0063] Methods for obtaining the original vibration signals of various types of gears and performing VMD variational mode decomposition on the original vibration signals to obtain intrinsic mode components (IMFs) include:

[0064] VMD variational mode decomposition is performed on the original signal to obtain K modes. The formula is as follows:

[0065]

[0066]

[0067]

[0068] Among them, A k (t) represents the k-th mode u k The amplitude, For the k-th mode u k Phase; center frequency of each component For u k The instantaneous frequency of (t); mode u k Represented as modality The initial value; x(t) represents the original vibration signal; ω k (t) represents the mode The center frequency; For u k The instantaneous frequency of (t).

[0069] In each mode The sum equals the original signal as a constraint; each mode center frequency and mode A constrained variational model is constructed with the minimum as the optimization objective, expressed by the following formula:

[0070]

[0071]

[0072] In the formula, δ(t) is the impulse function. t represents time; j represents the imaginary unit; "*" indicates convolution; Represented as the mode in the nth iteration The center frequency; This represents the partial derivative with respect to time; x(t) represents the original vibration signal.

[0073] By introducing the Lagrange multiplier λ and the quadratic penalty factor α, λ ensures the strictness of the constraints, and α ensures the accuracy of signal reconstruction under Gaussian noise conditions; the extended Lagrange expression is:

[0074]

[0075] The constrained variational model is solved iteratively using the alternating direction multiplier algorithm.

[0076]

[0077]

[0078] In the formula, They correspond to Fourier transforms of x(t) and λ(t); An iterative adaptive solution method can also be used; ω represents the center frequency.

[0079] After each iteration, the Lagrange multiplier λ is updated, expressed as follows:

[0080]

[0081] Where Γ represents the noise margin, λ n λ is represented as the Lagrange multiplier in the nth iteration; n-1 It is represented as the Lagrange multiplier in the (n-1)th iteration.

[0082] The optimal mode is output when the convergence condition is met. As intrinsic mode components (IMFs), the convergence conditions include:

[0083]

[0084] Where ε represents the precision setting value; Represents the square of the 2-norm of a matrix or determinant.

[0085] The decomposed intrinsic mode components (IMFs) contain three components: information-dominated component, noise-dominated component, and pure noise component. The correlation coefficient between the IMFs and the original vibration signal is calculated using the following formula:

[0086]

[0087] In the formula, ρ i The correlation coefficient between the intrinsic modal components (IMF) and the original vibration signal is expressed as x, where L is the length of the original vibration signal; h′ (t) represents the original vibration signal of segment h′; This is represented as the mean of the original vibration signal; R is expressed as the mean of the intrinsic mode components (IMFs); i,h′ It is represented as the intrinsic mode component (IMF) of the k-th segment h′.

[0088] Take the correlation coefficient ρ i Methods for obtaining the Hilbert-Huang spectrum by performing Hilbert transform on the largest intrinsic mode component (IMF) include:

[0089] Taking the intrinsic mode component (IMF) with the largest correlation coefficient, denoted as component I(t), the formula for finding its complex conjugate function y(t) using the Hilbert transform is:

[0090]

[0091] In the formula, P is the value of Cauchy's principle;

[0092] Based on the component I(t) and the complex conjugate function y(t), the corresponding analytic signal z(t) is constructed, expressed by the following formula:

[0093] z(t)=I(t)+jy(t)=a(t)e jθ(t)

[0094]

[0095]

[0096] The reconstructed signal Z(t) is obtained by reconstructing the analytic signal z(t), and the formula is as follows:

[0097]

[0098]

[0099] Reconstruct the signal Let Z(t) be the real part of the reconstructed signal, expressed as: Based on reconstructed signal Constructing Hilbert-Huang spectra; selecting appropriate IMF components ensures that the generated Hilbert-Huang spectra contain fault information, while eliminating interference components and highlighting local signal features.

[0100] A deep residual network model is constructed, which sequentially includes an input layer, a depthwise separable convolutional layer, a first ReLU+BN layer, a pointwise convolutional layer, a second ReLU+BN layer, a nonlinear transformation layer, and an output layer. The input features of the input layer are identically mapped to the output layer, as expressed by the following formula:

[0101] H(x′)=F(x′)+x′

[0102] In the formula, x′ represents the input feature identity mapping value of the input layer; F(x′) represents the residual mapping value; and H(x′) represents the feature of the deep residual network model.

[0103] Considering that the feature learning ability of deep residual network models often declines when processing high-noise vibration signals, the model may fail to detect fault-related features due to noise interference. In this case, the high-level features learned in the output layer often lack sufficient discriminative power to correctly classify the fault. To improve the feature learning ability of deep residual networks for high-noise vibration signals, the soft and hard thresholding combination is inserted as a nonlinear transform layer between the second ReLU+BN layer and the output layer. Depthwise separable convolution and channel reconstruction attention mechanisms address the issues of decreased diagnostic performance and excessive computation time caused by the increased depth of deep residual network models. While maintaining excellent network performance, this approach strengthens the connections between feature channels and reduces the number of trainable weight parameters.

[0104] The formula for calculating the soft and hard thresholds is as follows:

[0105]

[0106] In the formula, x * y is the input feature of the nonlinear transform layer. * τ represents the output characteristics of the nonlinear transform layer, τ is the threshold, and δ is the adjustment parameter.

[0107] Instead of setting negative features in the ReLU activation function to zero, soft and hard thresholding methods set features close to zero to zero, thus preserving useful negative features. During soft and hard thresholding, the derivative of the output with respect to the input is either 1 or 0, effectively preventing gradient vanishing and exploding problems. Its derivative can be expressed as:

[0108]

[0109] Pointwise convolution is essentially the same as standard convolution, its main function being to weight and combine the output features along the channel directions. First, depthwise convolution is used to extract features from each channel separately, and then pointwise convolution is used to integrate the channel output features. If the input feature size is W... n If the number of channels is M, the kernel width is W, and the number of kernels is K, then the computational complexity of depthwise separable convolution is:

[0110] T1 = W n ×M×W+K×M×W n

[0111] Standard convolution can be divided into two operations: feature extraction and feature merging. The computational cost of standard convolution is:

[0112] T2 = W n ×M×K×W

[0113] Comparing the computational cost of depthwise separable convolution and standard convolution reveals that:

[0114]

[0115] In the formula, the kernel width W n Typically, the values ​​are 3, 5, or 7. Since the number of kernels K is greater than 1, the comparison value above is less than 1, meaning that the computational cost of depthwise separable convolution is less than that of standard conventional convolution.

[0116] The method of dividing the Hilbert-Huang spectra corresponding to each original vibration signal into source domain data samples and target domain data samples; and using the source domain data samples and target domain data samples to train a deep residual network model to obtain a transfer diagnostic model for the target domain includes:

[0117] The original vibration signals corresponding to the source domain data samples are obtained from public datasets under various gear working conditions. The source domain data samples are used to train and debug a deep residual network model to obtain the optimal source domain fault diagnosis model. The weight parameters of the source domain fault diagnosis model are transferred to a new deep residual network model to obtain the target domain fault diagnosis model.

[0118] The original vibration signal corresponding to the target domain data sample is the data generated by the actual operation in the wind turbine gearbox. The target domain data sample is used to train the target domain fault diagnosis model. The parameters of the target domain fault diagnosis model are adjusted by the MK-MMD loss function, and finally the transfer diagnosis model is obtained.

[0119] The basic definition of maximum mean difference (MMD) is as follows:

[0120]

[0121]

[0122] In the formula, σ is the bandwidth; and These are source domain data samples and target domain data samples, respectively. For RKHS and A nonlinear mapping function; k(·) is the Gaussian kernel function; n s n represents the number of data samples in the source domain. t This represents the number of data samples in the target domain.

[0123] Low-level features have strong transferability, while the features of high-level convolutional layers are abstract features related to specific tasks. For datasets with large differences, it is necessary to train and update a large number of high-level convolutional layer parameters. At this time, the transfer learning method is adopted to extract shallow fault features by using the source domain fault diagnosis model as a pre-trained model, which can further learn deep fault features. The low-level network structure and parameters are frozen, and the high-level network structure uses target domain data to adjust the model parameters.

[0124] The formula for calculating the MK-MMD loss function is as follows:

[0125]

[0126] In the formula, θ represents the parameters in the target domain fault diagnosis model; n t D represents the number of data samples in the target domain. MK-MMD The difference between the target domain data samples and the source domain data samples; J(·) is the cross-entropy function; λ MK-MMD Weights for the results calculated by the MK-MMD function; Represented as target domain data sample; This represents the diagnostic results for the target domain data sample.

[0127] The MK-MMD loss function is used to better transfer the distribution of learned features on source domain data samples to the distribution of target domain data samples. The accuracy of the algorithm is verified using target domain data samples, and the model structure is adjusted so that the model can achieve high-precision fault diagnosis with only a small number of samples.

[0128] The target domain data sample is input into the transfer diagnostic model to obtain the diagnostic results. The fault types include, but are not limited to, missing teeth, broken teeth, wear, tooth root cracks, bearing inner ring faults, outer ring faults, ball faults, and mixed faults in rotating machinery gearboxes.

[0129] Example 2

[0130] This embodiment provides a wind turbine gearbox fault diagnosis system. The diagnosis system described in this embodiment can be applied to the diagnosis method described in Embodiment 1, including:

[0131] The acquisition module is used to acquire the original vibration signals of various types of gears; the original vibration signals are subjected to VMD variational mode decomposition to obtain intrinsic mode components (IMFs);

[0132] The feature extraction module is used to calculate the correlation coefficient between the intrinsic modal components (IMFs) and the original vibration signals. The IMF with the largest correlation coefficient is selected and subjected to Hilbert transform to obtain the Hilbert-Huang spectrum. The Hilbert-Huang spectrum corresponding to each original vibration signal is divided into source domain data samples and target domain data samples.

[0133] The training module is used to build a deep residual network model. It uses source domain data samples and target domain data samples to train the deep residual network model to obtain a transfer diagnostic model for the target domain.

[0134] The diagnostic module is used to input the target domain data sample into the transfer diagnostic model to obtain the diagnostic results.

[0135] Example 3

[0136] This embodiment provides an electronic device, including a storage medium and a processor; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the method described in Embodiment 1.

[0137] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0138] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0139] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0140] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0141] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A wind turbine generator set gearbox fault diagnostic method characterized by, include: The original vibration signals of various types of gears are acquired; the original vibration signals are subjected to VMD variational mode decomposition to obtain intrinsic mode components (IMFs). Calculate the correlation coefficient between the intrinsic mode components (IMFs) and the original vibration signals. Take the IMF with the largest correlation coefficient and perform a Hilbert transform to obtain the Hilbert-Huang spectrum. Divide the Hilbert-Huang spectrum corresponding to each original vibration signal into source domain data samples and target domain data samples. A deep residual network model is constructed, and the deep residual network model is trained using source domain data samples and target domain data samples to obtain a transfer diagnostic model for the target domain; the target domain test data samples are input into the transfer diagnostic model to obtain the diagnostic results; The deep residual network model sequentially comprises an input layer, a depthwise separable convolutional layer, a first ReLU+BN layer, a pointwise convolutional layer, a second ReLU+BN layer, a nonlinear transformation layer, and an output layer; the input features of the input layer are identically mapped to the output layer, expressed by the following formula: ； In the formula, represents the input feature identity mapping value as the input layer; represents the residual mapping value; represents the feature as the deep residual network model; The soft and hard thresholds are combined and inserted as a nonlinear transform layer between the second ReLU+BN layer and the output layer. The formula for calculating the soft and hard thresholds is as follows: ； wherein is an input feature of the nonlinear transformation layer, is an output feature of the nonlinear transformation layer, is a threshold value, is a tuning parameter.

2. The wind turbine gearbox fault diagnosis method according to claim 1, characterized in that, Methods for obtaining intrinsic mode components (IMFs) from the original vibration signal by performing VMD variational mode decomposition include: VMD variational mode decomposition is performed on the original signal to obtain K modes. ; in each mode The sum equals the original signal as a constraint; each mode center frequency and mode A constrained variational model is constructed with the minimum as the optimization objective, expressed by the following formula: ； st. ; In the formula, For impulse functions, , ; Time; j is the imaginary unit; "*" indicates convolution; Represented as the mode in the nth iteration The center frequency; This represents taking the partial derivative with respect to time; This is represented as the original vibration signal; By introducing Lagrange multipliers and secondary penalty factor The constrained variational model is solved iteratively using the alternating direction multiplier algorithm; the Lagrange multipliers are updated after each iteration. The formula is as follows ； in, Indicates noise tolerance. Represented as the Lagrange multiplier in the nth iteration; Represented as the Lagrange multiplier in the (n-1)th iteration; The optimal mode is output when the convergence condition is met. As intrinsic mode components (IMFs), the convergence conditions include: ； in, Represents the precision setting value; Represents the square of the 2-norm of a matrix or determinant.

3. The wind turbine gearbox fault diagnosis method according to claim 1, characterized in that, Calculate the correlation coefficient between the intrinsic modal components (IMF) and the original vibration signal. The calculation formula is: ； In the formula, It is expressed as the correlation coefficient between the intrinsic modal components (IMF) and the original vibration signal. The length of the original vibration signal; Represented as the first Original vibration signal of segment; This is represented as the mean of the original vibration signal; It is expressed as the mean of the intrinsic modal components (IMFs). Represented as the k-th The intrinsic mode components (IMF) of the segment.

4. The wind turbine gearbox fault diagnosis method according to claim 1, characterized in that, Methods for obtaining Hilbert-Huang spectra by performing Hilbert transform on intrinsic mode components (IMFs) include: Take the correlation coefficient The largest intrinsic mode component (IMF) is denoted as component. Then, the Hilbert transform is used to find its complex conjugate function. The formula is: ； In the formula, This is the value of Cauchy's principle; Based on components and complex conjugate function Construct the corresponding analytic signal The formula is as follows: ； ； ； Analyze the signal Reconstruction is performed to obtain reconstruction signals. The formula is as follows: ； ； Reconstruct the signal Set as reconstructed signal The real part is expressed by the formula: Based on reconstructed signals Construct the Hilbert-Huang spectrum.

5. The wind turbine gearbox fault diagnosis method according to claim 1, characterized in that, Methods for obtaining a transfer diagnostic model for the target domain by training a deep residual network model using source domain data samples and target domain data samples include: The original vibration signals corresponding to the source domain data samples are obtained from public datasets under various gear working conditions. The source domain data samples are used to train and debug a deep residual network model to obtain the optimal source domain fault diagnosis model. The weight parameters of the source domain fault diagnosis model are transferred to a new deep residual network model to obtain the target domain fault diagnosis model. The original vibration signals corresponding to the target domain data samples are data generated in the actual operation of the wind turbine gearbox. The target domain data samples are used to train the target domain fault diagnosis model. The parameters of the target domain fault diagnosis model are adjusted by the MK-MMD loss function, and finally the transfer diagnosis model is obtained.

6. The wind turbine gearbox fault diagnosis method according to claim 5, characterized in that, The formula for calculating the MK-MMD loss function is as follows: ； In the formula, These are the parameters in the target domain fault diagnosis model; The number of data samples in the target domain; The difference between the target domain data sample and the source domain data sample; It is the cross-entropy function; Weights for the results calculated by the MK-MMD function; Represented as target domain data sample; This represents the diagnostic results for the target domain data sample.

7. The application system of the wind turbine gearbox fault diagnosis method according to any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire the original vibration signals of various types of gears; the original vibration signals are subjected to VMD variational mode decomposition to obtain intrinsic mode components (IMFs); The feature extraction module is used to calculate the correlation coefficient between the intrinsic modal components (IMFs) and the original vibration signals. The IMF with the largest correlation coefficient is selected and subjected to Hilbert transform to obtain the Hilbert-Huang spectrum. The Hilbert-Huang spectrum corresponding to each original vibration signal is divided into source domain data samples and target domain data samples. The training module is used to build a deep residual network model. It uses source domain data samples and target domain data samples to train the deep residual network model to obtain a transfer diagnostic model for the target domain. The diagnostic module is used to input the target domain data sample into the transfer diagnostic model to obtain the diagnostic results.

8. An electronic device, comprising a storage medium and a processor; said storage medium for storing instructions; said processor for operating according to said instructions to perform the method of any one of claims 1 to 6.

Images