A method and system for fault diagnosis of wind turbine gearboxes based on data fusion

Abstract

This application belongs to the field of wind turbine gearbox fault diagnosis technology, specifically involving a wind turbine gearbox fault diagnosis method and system based on data fusion. It includes a data acquisition unit, a data processing unit, and an output unit. It extracts features of different scales from the vibration data of the wind turbine gearbox and uses a multi-scale convolution module to achieve feature fusion. This method can effectively alleviate the shortcomings of single-scale features in accurately expressing fault information. Multiple scale features complement each other, improve fault detection efficiency, and enhance anti-interference ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fan gearbox fault diagnosis, and particularly relates to a fan gearbox fault diagnosis method and system based on data fusion. BACKGROUND

[0002] In a fan, the gearbox is a key component of the transmission system, and maintaining its health state is of great significance to the fan system. An accurate and reliable fan gearbox fault model is particularly important. The existing data-based fan gearbox fault diagnosis method is mostly a "black box" model, which has low interpretability. Therefore, it is necessary to design a reliable and accurate fault diagnosis method to accurately identify faults in the early stage of failure and ensure the healthy operation of the fan. In addition, the fan system is relatively complex and is in a harsh and variable environment. For gearbox fault diagnosis, time domain features alone are sensitive to impact type faults (such as gear tooth breakage), but it is difficult to handle complex faults and signals with high background noise. Single frequency domain features can better distinguish the corresponding relationship between fault types and frequencies, but the effect on non-stationary signals is poor. SUMMARY

[0003] Based on the above problems, the application aims to provide a data-based fan gearbox feature extraction and fusion and fault diagnosis method to make up for the low interpretability of existing methods and the inability of single scale features to accurately reflect data features. The application adopts the following technical solutions:

[0004] A fan gearbox fault diagnosis method based on data fusion, comprising the following steps:

[0005] S1. Constructing a data set according to the collected fan gearbox vibration data;

[0006] S2. Data processing the constructed data set, and grouping the time series data according to the rotational speed;

[0007] S3. Extracting the time domain, frequency domain, wavelet packet decomposition derived features of each data sample, and the spectral entropy features based on local feature scale decomposition;

[0008] S4. Feature fusion of the extracted features through a multi-scale convolution module to obtain the final data feature expression;

[0009] S5. Inputting the feature data into a data mapping layer first to obtain the expression of the gearbox feature data in the hidden space;

[0010] S6. Constructing a multi-head attention model to explore the global and local dependency relationships in the feature data;

[0011] S7. Constructing a classifier to output the specific fault type.

[0012] Preferably, the missing values in the data set are filled using linear interpolation in step S2, and the formula is as follows: x, y

[0013] (1);

[0014] The mean value is used to fill the error data;

[0015] According to the gearbox speed, the vibration data of each circle is divided into a group, and each group of data is taken as a data sample.

[0016] Preferably, the time domain features in step S3 are: absolute mean, peak value, effective value, square root amplitude, variance, peak-peak value, skewness index, kurtosis index, peak index, waveform index, pulse index, margin index, coefficient of variation, minimum value, average value, rectified average value, 3rd order center distance, 4th order center distance, 3rd order origin distance, and 4th order origin distance, to obtain the time domain feature vector .

[0017] The frequency domain features are: center of gravity frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, spectral peak stability coefficient, 1st frequency band relative energy, 2nd frequency band relative energy, 3rd frequency band relative energy, 4th frequency band relative energy, and 5th frequency band relative energy, to obtain the frequency domain feature vector .

[0018] Preferably, the wavelet packet decomposition in step S3 can simultaneously perform multi-level decomposition on the low-frequency and high-frequency parts of the signal, and the derived features can simultaneously reflect the time domain and frequency domain characteristics of the signal, and the steps are as follows:

[0019] First, the wavelet packet decomposition is performed on each data sample, and the mathematical expression is as follows:

[0020] (2);

[0021] (3);

[0022] wherein, is the coefficient of the jth layer and the nth node, and are low-pass and high-pass filters, respectively;

[0023] Secondly, after wavelet packet decomposition, the relevant derived features are extracted: wavelet packet energy, wavelet packet energy ratio, wavelet packet node 1 energy, wavelet packet node 2 energy, wavelet packet node 3 energy, and wavelet packet node 4 energy, to obtain the time-frequency domain feature vector .

[0024] Preferably, the spectral entropy feature based on local feature scale decomposition is obtained by the following steps: ​

[0025] (1) Extract the extreme points of the signal, find all the local maximum points and local minimum points of the signal;

[0026] (2) Construct the upper and lower envelope lines, interpolate the maximum points to form the upper envelope line , interpolate the minimum points to form the lower envelope line ;

[0027] (3) Calculate the mean envelope line, ;

[0028] (4) Extract the intrinsic scale component, subtract the mean envelope from the original signal to obtain the candidate intrinsic scale component, ;

[0029] Determine whether it is a valid intrinsic scale component, the conditions are: The number of maximum and minimum values of is equal and appears alternately, The number of zero-crossing points of is not more than 1 from the number of extreme points; if the above two requirements are met, It is considered as an intrinsic scale component, denoted as ; otherwise, repeat the above decomposition process for ;

[0030] (5) Recursive decomposition, the residual signal is taken as the new input and continues to be decomposed until the residual signal becomes a monotonic trend or noise;

[0031] (6) After the completion of the local characteristic scale LCD decomposition, the LCD energy spectrum entropy, the LCD singular value spectrum entropy, the LCD envelope spectrum entropy, the LCD component 1 energy, and the LCD component 2 energy are extracted to obtain the LCD spectrum entropy feature vector .

[0032] Preferably, S4 performs feature fusion on the extracted features through a multi-scale convolution module, and the steps are as follows:

[0033] S41. The extracted feature vectors of various types are directly spliced to obtain , that is, , is the input of the multi-scale convolution module;

[0034] S411. Input through a convolution layer with a convolution kernel of to obtain , which can increase the number of channels so that the data can interact with the next convolution layer;

[0035] S412. Through three different operations, obtain , and , three operations are respectively copy, convolution layer with and convolution layer with ;

[0036] S413. , and perform convolution operation with respectively to obtain , and ;

[0037] S42. Concatenate , and to obtain , is the final output of the proposed multi-scale feature extraction and fusion method.

[0038] Preferably, step S5 for a given input sample , X is composed of feature data , the feature expression of the input data in the hidden space is realized by the following linear transformation, and the dimension requirement of the multi-head attention module is met:

[0039] (4);

[0040] wherein, is the mapping weight matrix, is the bias vector, is the mapped feature representation, is the dimension of the hidden space; B is the batch size of the input data, L is the sequence length, D in is the input dimension.

[0041] Preferably, step S6 includes:

[0042] S61. Single-head attention calculation ;

[0043] S62. Multi-head attention is obtained by concatenating the outputs of all heads and performing linear transformation ;

[0044] S63. After the multi-head attention output, the residual connection is added and the layer normalization is performed to obtain ;

[0045] S64. After the FFF output, the residual connection and the layer normalization are performed again to obtain:

[0046] (13);

[0047] complete N The layer self-attention layer stacking process is:

[0048] (14);

[0049] The output of the final self-attention module is:

[0050] (15).

[0051] Preferably, step S7 adopts a three-layer KAN network structure , wherein is the number of fault categories:

[0052] The forward propagation process of the KAN network is:

[0053] (16);

[0054] (17);

[0055] (18);

[0056] wherein is the KAN transformation of the first layer, is the output of the first layer, is the output of the second layer;

[0057] Each layer transformation is composed of a series of learnable B-spline basis functions:

[0058] (19);

[0059] wherein is the B-spline basis function, is the above , , or , is a learnable weight;

[0060] The final output is the predicted probability distribution of each fault category, that is, the final wind turbine gearbox fault diagnosis result.

[0061] A wind turbine gearbox fault diagnosis system based on data fusion includes a data acquisition unit, a data processing unit and an output unit;

[0062] Data acquisition unit: collect wind turbine gearbox vibration data, build data set;

[0063] Data processing unit: Extracts time-domain, frequency-domain, wavelet packet decomposition-derived features and spectral entropy features based on local feature scale decomposition for each data sample; utilizes a multi-head attention model to discover global and local dependencies in the feature data; constructs a classifier to output the specific fault type;

[0064] Output unit: Visualizes the results.

[0065] Compared with the prior art, the beneficial effects of this application are as follows:

[0066] The design incorporates the above four types of features to create a fault extraction and fusion method. The multi-scale features enable the complementarity of signal features, improve the detection rate of early faults, better adapt to complex faults, and enhance their anti-interference capabilities. Attached Figure Description

[0067] Figure 1 This is a flowchart of the fault diagnosis model.

[0068] Figure 2 Here is a flowchart of the feature extraction and fusion process;

[0069] Figure 3 This is a diagram showing the experimental results. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0071] A data-based method for feature extraction, fusion, and fault diagnosis of wind turbine gearboxes includes the following steps:

[0072] S1, Construct a dataset based on the collected vibration data of the wind turbine gearbox.

[0073] S2 processes the constructed dataset, fills in missing values, corrects erroneous data, and groups the time series data according to rotational speed.

[0074] Use linear interpolation to find missing values ​​in the dataset. x, y Fill in the blanks using the following formula:

[0075] (1);

[0076] For obviously erroneous data, fill in the mean. Divide the vibration data for each revolution of the gearbox into groups based on the gearbox speed, and use each group as a data sample.

[0077] S3 extracts the time domain, frequency domain, time-frequency domain, and spectral entropy features of each data sample based on local feature scale decomposition.

[0078] Time-domain features reflect the core characteristics of a signal, such as amplitude, volatility, periodicity, and impulsivity. The following time-domain features are extracted for each data sample: absolute mean, peak value, RMS value, root square amplitude, variance, peak-to-peak value, skewness index, kurtosis index, peak value index, waveform index, impulse index, margin index, coefficient of variation, minimum value, average value, rectified average value, 3rd-order center distance, 4th-order center distance, 3rd-order origin distance, and 4th-order origin distance, resulting in a time-domain feature vector. .

[0079] Frequency domain features can reveal the periodicity and frequency characteristics of a signal. For each data sample, the following frequency domain features are extracted: centroid frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, spectral peak stability coefficient, relative energy of band 1, relative energy of band 2, relative energy of band 3, relative energy of band 4, and relative energy of band 5, resulting in a frequency domain feature vector. .

[0080] Wavelet packet decomposition can simultaneously perform multi-level decomposition of the low-frequency and high-frequency components of a signal, and its derived features can simultaneously reflect the time-domain and frequency-domain characteristics of the signal. First, wavelet packet decomposition is performed on each data sample, mathematically expressed as follows:

[0081] (2);

[0082] (3);

[0083] in, It is the coefficient of the nth node in the j-th layer. and These are low-pass and high-pass filters, respectively. After wavelet packet decomposition, relevant derived features are extracted: wavelet packet energy, wavelet packet energy ratio, wavelet packet node 1 energy, wavelet packet node 2 energy, wavelet packet node 3 energy, and wavelet packet node 4 energy, resulting in a time-frequency domain feature vector. .

[0084] Spectral entropy features based on local feature scale decomposition have the advantage of high sensitivity to minor faults. Therefore, local feature scale decomposition is first performed on each data sample, as follows.

[0085] (1) Extract the extreme points of the signal and find all local maxima and local minima of the signal.

[0086] (2) Construct upper and lower envelopes, and interpolate the maximum points to form the upper envelope. Interpolation is performed on the minimum points to form the lower envelope. .

[0087] (3) Calculate the mean envelope. .

[0088] (4) Extract intrinsic scaling components by subtracting the mean envelope from the original signal to obtain candidate intrinsic scaling components. To determine whether it is an effective intrinsic scale component, the following conditions must be met: The number of maxima and minima is equal and they alternate. The difference between the number of zero-crossing points and the number of extreme points does not exceed 1. If both of the above requirements are met, then... Considered as an intrinsic metric component, denoted as Otherwise, for Repeat the above decomposition process.

[0089] (5) Recursive decomposition, remaining signal As a new input, the decomposition continues until the remaining signal becomes a monotonic trend or noise (intrinsic scale components cannot be extracted).

[0090] After the Local Feature Scale Decomposition (LCD) is completed, the LCD energy spectral entropy, LCD singular value spectral entropy, LCD envelope spectral entropy, LCD component 1 energy, and LCD component 2 energy are extracted to obtain the LCD spectral entropy eigenvector. .

[0091] S4 performs feature fusion on the extracted features using a multi-scale convolution module.

[0092] The extracted feature vectors are directly concatenated to obtain... ,Right now , This is the input for the multi-scale convolutional module. First, the input... After a convolution kernel The convolutional layer is obtained This convolutional layer can increase the number of channels, allowing data to interact with the next convolutional layer.

[0093] Secondly After three different operations, the result was obtained. , and The three operations are copying, convolution kernel... The convolutional layers and convolutional kernels are The convolutional layer.

[0094] Then, , and Perform convolution kernel as The convolution operations respectively yielded , and This convolutional layer not only increases nonlinearity but also reduces the number of parameters and computational cost without compromising the module's expressive power. Furthermore, the two convolutional kernels in this module are... The convolutional layers enable the fusion of data information between different channels.

[0095] Finally, , and spliced ​​together , This is the final output of the proposed multi-scale feature extraction and fusion method.

[0096] S5, first input the data into the data mapping layer to obtain the representation of the gearbox data in the hidden space.

[0097] For a given input sample (from feature data) The following linear transformation is used to represent the features of the input data in the hidden space and to meet the dimensionality requirements of the multi-head attention module.

[0098] (4);

[0099] in, For the mapping weight matrix, For the bias vector, The mapped feature representation, To hide the spatial dimension.

[0100] S6 constructs a multi-head attention module, which includes multiple attention layers. Each attention layer consists of an attention mechanism, a residual normalization layer, and a feedforward network.

[0101] For the Layer input First, a query, key, and value matrix is ​​generated through linear transformation. For the... One point of attention:

[0102] (5);

[0103] (6);

[0104] (7);

[0105] in , , For the number of attention heads.

[0106] Single-head attention is calculated as follows:

[0107] (8);

[0108] scaling factor This is used to prevent the gradient from vanishing due to an excessively large inner product.

[0109] Multi-head attention works by concatenating the outputs of all heads and performing a linear transformation:

[0110] (9);

[0111] in This is for outputting the projection matrix.

[0112] After multi-head attention output, residual connections are added and layer normalization is performed:

[0113] (10);

[0114] The formula for calculating layer normalization is:

[0115] (11);

[0116] in: For layer normalization input, The characteristic mean, For characteristic variance, For learnable scaling and offset parameters, is the numerical stability constant.

[0117] The feedforward network consists of two linear transformations and a ReLU activation function:

[0118] (12);

[0119] in: , , , , This represents the intermediate dimension of the feedforward network.

[0120] After the FFN output, residual connections and layer normalization are performed again:

[0121] (13);

[0122] The complete stacking process of N self-attention layers is as follows:

[0123] (14);

[0124] The final output of the self-attention module is:

[0125] (15);

[0126] for Self-attention output.

[0127] S7. Construct the Kolmogorov-Arnold Network (KAN) classifier to output the specific fault type.

[0128] The Kolmogorov-Arnold Network, based on the Kolmogorov-Arnold representation theorem, achieves nonlinear mapping through learnable spline functions. This paper employs a three-layer KAN network structure. ,in This represents the number of fault categories.

[0129] The forward propagation process of the KAN network is as follows:

[0130] (16);

[0131] (17);

[0132] (18);

[0133] in For the first The KAN transform of each layer consists of a series of learnable B-spline basis functions:

[0134] (19);

[0135] in For the first One B-spline basis function, These are learnable weights.

[0136] Final output For each type of fault.

[0137] The above method yielded the following fault diagnosis results under four operating conditions on the MCC5-THU gearbox dataset: Figure 3 .from Figure 3As can be seen, in operating condition 1, the motor input shaft rotates at a constant speed of 1000 rpm, and the torque of the gearbox output shaft is between 0-10 Nm. In operating condition 2, the motor input shaft rotates at a constant speed of 2000 rpm, and the torque of the gearbox output shaft is between 0-20 Nm. In operating condition 3, the gear output torque is 10 Nm, and the motor input shaft rotates at a speed of 0-2000 rpm. In operating condition 4, the gear output torque is 20 Nm, and the motor input shaft rotates at a speed of 0-2000 rpm. Each operating condition includes data on the gearbox under five fault conditions: missing teeth, wear, pitting, fracture, and cracks, as well as data on its healthy operation.

[0138] A wind turbine gearbox fault diagnosis system based on data fusion includes a data acquisition unit, a data processing unit, and an output unit;

[0139] Data acquisition unit: Collects vibration data from the wind turbine gearbox and constructs a dataset;

[0140] Data processing unit: Extracts time-domain, frequency-domain, wavelet packet decomposition-derived features and spectral entropy features based on local feature scale decomposition for each data sample; utilizes a multi-head attention model to discover global and local dependencies in the feature data; constructs a classifier to output the specific fault type;

[0141] Output unit: Visualizes the results.

[0142] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the application. Various changes and modifications can be made without departing from the spirit and scope of this application, and all such changes and modifications fall within the scope of the claims. The scope of protection of this application is defined by the appended claims and their equivalents.

Claims

1. A data fusion based fault diagnosis method for a fan gearbox, characterized in that, Comprising the following steps: S1. Constructing a data set according to the collected fan gearbox vibration data; S2. Data processing is performed on the constructed data set, and time series data is grouped according to the rotating speed; S3. Extracting the time domain, frequency domain, wavelet packet decomposition derived features of each data sample, and the spectral entropy features based on local characteristic scale decomposition; S4. The extracted features are fused through a multi-scale convolution module to obtain the final data feature expression; S41. The extracted feature vectors of various types are directly spliced to obtain i.e. , is the input of the multi-scale convolution module; S411. Input After a convolution kernel The convolutional layer is obtained This convolutional layer can increase the number of channels, allowing data to interact with the next convolutional layer; S412. After three different operations, the result was obtained. , and The three operations are copying, convolution kernel... The convolutional layers and convolutional kernels are Convolutional layers; S413. 、 and The convolution kernel is The convolution operation is performed respectively to obtain 、 and ; S42. obtaining , and splicing , is the final output of the proposed multi-scale feature extraction and fusion method; S5. The feature data is first input into the data mapping layer to obtain the expression of the gearbox feature data in the hidden space; For a given input sample , X is composed of feature data , the feature expression of the input data in the hidden space is realized by the following linear transformation, and the dimension requirement of the multi-head attention module is met: （4）； wherein, is a mapping weight matrix, is a bias vector, is a mapped feature representation, is a hidden space dimension, B is a batch size of input data, L is a sequence length, D in is an input dimension; S6. A multi-head attention model is constructed to explore the global and local dependency relationships in the feature data; S61. Single-head attention computation ; S62. Multi-head attention is obtained by concatenating the outputs of all heads and applying a linear transformation ; S63. After multi-head attention output, add residual connection and perform layer normalization to obtain ; S64. After the FFN output, residual connection and layer normalization are performed again to obtain: （13）； complete The layer self-attention layer stacking process is: （14）； The output of the final self-attention module is: （15）； S7. A classifier is constructed to output the specific fault type; A three-layer KAN network structure is adopted , wherein is the number of fault categories: The forward propagation process of the KAN network is: （16）； （17）； （18）； wherein is the first layer output, is the KAN transform of the layer, is the first layer output, is the second layer output; Each layer of transformation is composed of a series of learnable B-spline basis functions: （19）； wherein is the th B-spline basis function, is the , or , is a learnable weight; Final output The prediction probability distribution for each fault category, i.e. the final fan gearbox fault diagnosis result.

2. The data fusion based fault diagnosis method of a fan gearbox according to claim 1, characterized in that, In step S2, linear interpolation is used to fill the missing values (x, y) in the data set, and the formula is as follows: （1）； For erroneous data, use mean filling; According to the rotating speed of the gearbox, the vibration data of each cycle is divided into a group, and each group of data is taken as a data sample.

3. The data fusion based fan gear box fault diagnostic method of claim 1, wherein, Step S3 time domain feature: absolute mean value, peak value, effective value, square root amplitude, variance, peak-peak value, skewness index, kurtosis index, peak index, waveform index, pulse index, margin index, coefficient of variation, minimum value, average value, rectified average value, 3-order center distance, 4-order center distance, 3-order origin distance, 4-order origin distance, to obtain a time domain feature vector ; Frequency domain features: center of gravity frequency, mean square frequency, root mean square frequency, frequency variance, frequency standard deviation, spectral peak stability coefficient, relative energy of the first frequency band, relative energy of the second frequency band, relative energy of the third frequency band, relative energy of the fourth frequency band, relative energy of the fifth frequency band, to obtain a frequency domain feature vector .

4. The data fusion based fan gearbox fault diagnosis method of claim 3, wherein, Step S3 wavelet packet decomposition can simultaneously decompose the low-frequency and high-frequency parts of the signal at multiple levels, and its derived features can simultaneously reflect the time domain and frequency domain characteristics of the signal, as follows: First, wavelet packet decomposition is performed on each data sample, and the mathematical expression is as follows: （2）； （3）； wherein, is the coefficient of the jth layer nth node, and are low pass and high pass filters, respectively; Secondly, after wavelet packet decomposition, the relevant derived features are extracted: wavelet packet energy, wavelet packet energy ratio, wavelet packet node 1 energy, wavelet packet node 2 energy, wavelet packet node 3 energy, wavelet packet node 4 energy, to obtain the time-frequency domain feature vector .

5. The data fusion based fan gearbox fault diagnosis method of claim 1, wherein, The spectral entropy feature based on local characteristic scale decomposition is obtained as follows: (1) Extract the extreme points of the signal to find all the local maximum points and local minimum points of the signal; (2) Constructing upper and lower envelope lines, interpolating the maximum points to form the upper envelope line , interpolating the minimum points to form the lower envelope line ; (3) calculating a mean envelope, ; (4) extracting the intrinsic scale component, subtracting the mean envelope from the original signal to obtain a candidate intrinsic scale component, ; whether it is an effective intrinsic scale component, provided that: the maximum and minimum values of the function are equal in number and alternate, the number of zero-crossing points of the function differs from the number of extreme points by no more than 1; if the above two requirements are met, then it is considered an intrinsic scale component, denoted as ; otherwise, repeat the above decomposition process for ; (5) Recursive decomposition, residual signal Continue to decompose as new input until the residual signal becomes monotonic trend or noise; (6) After the local characteristic scale decomposition (LCD) is completed, LCD energy spectrum entropy, LCD singular value spectrum entropy, LCD envelope spectrum entropy, LCD component 1 energy, and LCD component 2 energy are extracted to obtain an LCD spectrum entropy feature vector .

6. A fan gearbox fault diagnosis system based on data fusion, which adopts the fan gearbox fault diagnosis method based on data fusion according to any one of claims 1-5, comprising a data acquisition unit, a data processing unit and an output unit; Data acquisition unit: collect fan gearbox vibration data and construct a data set; Data processing unit: extract the time domain, frequency domain, wavelet packet decomposition derived features of each data sample, and the spectral entropy features based on local characteristic scale decomposition, and use a multi-head attention model to explore the global and local dependency relationships in the feature data; construct a classifier to output the specific fault type; Output unit: visually output the results.

Images