CN105136454A - Wind turbine gear box fault recognition method - Google Patents

Abstract

The invention discloses a fault identification method for a wind turbine gearbox, which comprises the following steps: acquiring historical data of wind turbine gearbox operation within a certain time range; using autocorrelation analysis to perform wavelet denoising processing on the historical data; using fast Fourier transform , extract the time-domain and frequency-domain characteristic parameters in the denoised historical data; use the kernel principal component analysis method to reduce the dimensionality of the characteristic parameters, and extract several nonlinear principal components with the largest cumulative contribution rate of variance; use the gear The nonlinear principal element extracted from the historical data of the normal operation of the gearbox is used to establish a normal model, and the support vector machine is used for training. For identification, the present invention is beneficial to improve the processing ability of the vibration signal, and is of great significance to the identification of the fault of the gearbox.

Description

A fault identification method for wind turbine gearbox

technical field

The invention relates to the field of fault monitoring and fault diagnosis of a transmission system of a wind turbine, in particular to a fault identification method for a gearbox of a wind turbine based on fast Fourier transform, kernel principal component analysis and support vector machine.

Background technique

With the rapid development of wind energy, a large number of wind turbines have been put into operation, and because most of the wind turbines are installed in remote areas, the load is unstable and other factors, many wind turbines in my country have malfunctioned, which will directly affect the safety of wind power generation sex and economy. In order to make wind power generation more competitive and ensure continuous and efficient operation of wind turbines, the importance of maintenance services such as condition monitoring, fault diagnosis, and repair of wind turbines has attracted widespread attention. Among them, the damage to key mechanical components of wind turbines is particularly serious. According to statistics, the damage rate of gearboxes in wind farms in my country is as high as 40% to 50%, which is the component with the highest failure rate among mechanical components of wind turbines.

When the wind turbine gearbox fails, due to the existence of multi-component coupling vibration and the huge vibration and noise interference during operation, the vibration signal is non-Gaussian, non-stationary, and nonlinear. Commonly used time-domain and frequency-domain feature information extraction methods often contain a lot of redundant information, so that the accuracy of the signal is not high, it is difficult to accurately evaluate and reveal the inherent characteristics of the operating state of the wind turbine, and it cannot effectively reflect the current state of the equipment. The wavelet analysis technology based on the time-frequency domain can meet the above requirements, but in practical applications, the extracted equipment signals often have a strong noise background. How to further process these fault signals is a major obstacle in signal analysis.

In view of this, it is necessary to propose a fault identification method for wind turbine gearboxes, which can analyze the historical data of wind turbine gearboxes after denoising through kernel principal component analysis and variance cumulative contribution rate, and use support vector machines for training and testing, which is beneficial to Improving the processing ability of vibration signals is of great significance to gearbox fault identification.

Contents of the invention

In order to overcome the deficiencies in the above-mentioned prior art, the purpose of the present invention is to provide a wind turbine gearbox fault identification method, which fully considers that there are a large number of nonlinear signals in the wind turbine gearbox fault, and uses support vector machines for fault classification The core principal component analysis method is added on the basis of the fault classification, and the fault recognition rate is improved on the basis of taking into account the fault classification effect, especially for the large number of nonlinear signals in the fault of the wind turbine gearbox, which has stronger adaptability.

In order to achieve the above and other purposes, the present invention proposes a wind turbine gearbox fault identification method, including the following steps:

Step 1, obtain the historical data of wind turbine gearbox operation within a certain time range;

Step 2, using autocorrelation analysis to perform wavelet de-noising processing on historical data;

Step 3, extracting time-domain and frequency-domain characteristic parameters in the denoised historical data through fast Fourier transform;

Step 4, using the nuclear principal component analysis method to reduce the dimensionality of the characteristic parameters, and extract several nonlinear principal components with the largest cumulative contribution rate of variance;

Step five, use the nonlinear principal element extracted from the historical data of the normal operation of the gearbox to establish a normal model, and use the support vector machine for training

Step 6: Import the nonlinear principal element extracted from the historical data of gearbox operation into the trained model, thereby identifying the fault of the gearbox.

Further, step two further includes:

Step 2.1, select an appropriate wavelet and set the level N of wavelet decomposition, and then carry out N-layer wavelet decomposition on the signal containing noise;

Step 2.2, for each layer of high-frequency coefficients from the 1st to the Nth layer, select a threshold to perform soft threshold quantization;

Step 2.3, performing wavelet reconstruction of the one-dimensional signal according to the low-frequency coefficients of the Nth layer decomposed by wavelets and the quantized high-frequency coefficients of the first layer to the Nth layer.

Further, step three further includes:

Step 3.1, performing fast Fourier transform preprocessing, extracting time-domain characteristic parameters and frequency-domain characteristic parameters;

Step 3.2, calculate the 8 most influential time-domain indicators;

Step 3.3, calculating 6 frequency domain indicators;

Step 3.4: The calculated time domain index and frequency domain index constitute the original feature set of the gear state.

Further, the time-domain index includes the mean value, peak value, root mean square value, variance, kurtosis and dimensionless parameters kurtosis factor, pulse factor and margin factor of the vibration signal.

Further, the frequency domain index includes the rotational frequency amplitude, the meshing frequency amplitude, the double frequency amplitude of the meshing frequency, the center of gravity frequency, the mean square frequency, and the frequency variance of the vibration signal.

Further, step four further includes:

Step 4.1, select an appropriate kernel function;

Step 4.2, calculate the kernel matrix K;

Step 4.3, performing eigenvector decomposition on the kernel matrix K, and centering the kernel matrix K to obtain K';

Step 4.4, select normal sample data, calculate its mean and standard deviation, and standardize the sample data to construct a training matrix X;

Step 4.5, calculating eigenvalues and eigenvectors;

Step 4.6, calculate the variance cumulative contribution rate, determine the number of pivots, and extract the eigenvector V corresponding to the largest eigenvalues;

Step 4.7, calculate the projection t _k of the eigenvector V on the eigenspace, that is, the nonlinear pivot of the matrix X.

Further, in step 4.6, the variance contribution rate of the ith component xi in the data set X is:

${\mathrm{<span\; class="notranslate">\; PV</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}$

The variance cumulative contribution rate of the first k components is:

${\mathrm{<span\; class="notranslate">\; CPV</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; PV</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

Where λ _i is the eigenvalue of the covariance matrix Σ, and λ ₁ ≥λ ₂ ...≥λ _m ≥0.

Further, if the cumulative contribution rate of the variance of the current k kernel pivots reaches more than 90%, it is considered that the sought kernel pivots can fully represent the original feature information, and the number of kernel pivots is determined to be k at this time.

Further, in step 1, the historical data of 8 measuring points in total of the input shaft and the output shaft of the wind turbine gearbox within a certain time range are obtained from the data storage module.

Further, in step 1, the historical data obtained are the historical data of the axial and radial measuring points of the input shaft and the output shaft.

Compared with the prior art, a wind turbine gearbox fault identification method according to the present invention, according to the fault characteristics of the wind turbine gearbox, analyzes the historical data of the wind turbine gearbox after denoising through kernel principal component analysis and variance cumulative contribution rate, It is beneficial to deal with nonlinear signals, and then use support vector machine for fault classification, which is conducive to improving the processing ability of vibration signals, and improves the fault recognition rate on the basis of taking into account the effect of fault classification, especially for the large number of faults in the fan gearbox fault. The nonlinear signal has stronger adaptability, which is of great significance to the fault identification of the gearbox, and it is also of great significance to the maintenance and optimization cost of the gearbox of the wind turbine.

Description of drawings

Fig. 1 is a flow chart of the steps of a method for identifying a fault of a wind turbine gearbox according to the present invention.

Detailed ways

The implementation of the present invention is described below through specific examples and in conjunction with the accompanying drawings, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific examples, and various modifications and changes can be made to the details in this specification based on different viewpoints and applications without departing from the spirit of the present invention.

Fig. 1 is a flow chart of the steps of a method for identifying a fault of a wind turbine gearbox according to the present invention. As shown in Fig. 1, a kind of wind turbine gearbox fault identification method of the present invention comprises the following steps:

Step 101 , acquiring historical data of wind turbine gearbox operation within a certain time range.

Step 102, using autocorrelation analysis to perform wavelet de-noising processing on the historical data.

In step 103, feature parameters in the time domain and frequency domain are extracted from the denoised historical data through fast Fourier transform.

In step 104, the kernel principal component analysis method is used to reduce the dimensionality of the characteristic parameters, and several nonlinear principal components with the largest cumulative contribution rate of variance are extracted.

Step 105, using the nonlinear principal element extracted from the historical data of normal operation of the gearbox to establish a normal model, and train it with a support vector machine.

Step 106, importing the nonlinear principal element extracted from the historical data of the later operation of the gearbox into the trained model, so that the fault of the gearbox can be identified.

Further, step 102 also includes the following steps:

Step 2.1 Wavelet decomposition. Select an appropriate wavelet and set the level N of wavelet decomposition, and then carry out N-level wavelet decomposition on the noise-containing signal.

Step 2.2 Threshold quantization of high frequency coefficients of wavelet decomposition. For each layer of high-frequency coefficients from the 1st to the Nth layers, a threshold is selected for soft threshold quantization.

Step 2.3 Reconstruction of one-dimensional wavelet. According to the low-frequency coefficients of the Nth layer decomposed by the wavelet and the high-frequency coefficients of the first layer to the Nth layer after quantization, the wavelet reconstruction of the one-dimensional signal is carried out.

Further, step 103 also includes the following steps:

Step 3.1 Fast Fourier transform preprocessing, extracting time-domain characteristic parameters and frequency-domain characteristic parameters.

Step 3.2 calculates 8 most influential time-domain indicators. The time-domain indicators are the mean value, peak value, root mean square value, variance, kurtosis and dimensionless parameters kurtosis factor, pulse factor and margin factor of the vibration signal.

Step 3.3 Calculate 6 frequency domain indicators, frequency domain indicators are the vibration signal’s rotation frequency amplitude, meshing frequency amplitude, double frequency amplitude of meshing frequency, center of gravity frequency, mean square frequency, and frequency variance

In step 3.4, the calculated time domain index and frequency domain index constitute the original feature set of the gear state.

Further, step 104 also includes the following steps:

Step 4.1: Select an appropriate kernel function;

Step 4.2: Calculate the kernel matrix K;

Step 4.3: Decompose the eigenvector of K, and center the kernel matrix K to obtain K';

Step 4.4: Select normal sample data, calculate its mean and standard deviation, and standardize the sample data to construct a training matrix X;

Step 4.5: Calculate eigenvalues and eigenvectors;

Step 4.6: Calculate the variance cumulative contribution rate, determine the number of pivots, and extract the eigenvectors corresponding to the largest eigenvalues;

Step 4.7: Calculate the projection t _k of the eigenvector V on the eigenspace, that is, the nonlinear pivot of the matrix X.

Each step of the present invention will be described below by specific examples:

Step 1, get historical data. That is, the historical data of 8 measuring points of the wind turbine gearbox input shaft and output shaft (including axial and radial directions) within a certain time range are obtained from the data storage module.

Step 2, using autocorrelation analysis to perform wavelet denoising processing on historical data. Through preliminary comparison, the data in the radial direction of the output shaft can better reflect the change of the vibration signal, so the data of the measuring points in the radial direction of the output shaft are selected for subsequent analysis

Step 3, through fast Fourier transform, extract the time domain and frequency domain characteristic parameters in the denoised historical data. Fast Fourier transform preprocessing to extract time-domain feature parameters and frequency-domain feature parameters. Calculate the 8 most influential time-domain indicators, the time-domain indicators are the mean value, peak value, root mean square value, variance, kurtosis and dimensionless parameters kurtosis factor, pulse factor and margin factor of the vibration signal. Calculate 6 frequency domain indicators, the frequency domain indicators are the vibration signal's rotation frequency amplitude, meshing frequency amplitude, double frequency amplitude of meshing frequency, center of gravity frequency, mean square frequency, and frequency variance. The calculated time domain index and frequency domain index constitute the original feature set of the gear state.

Step 4, using the kernel principal component analysis method to reduce the dimensionality of the characteristic parameters, and extract several nonlinear principal components with the largest cumulative contribution rate of variance. First select the appropriate kernel function: Gaussian radial basis kernel function, calculate the kernel matrix K, decompose the eigenvector of K, center the kernel matrix K to get K', select normal sample data, calculate its mean and standard deviation, and Standardize the sample data, construct the training matrix X, calculate the eigenvalues and eigenvectors, calculate the cumulative contribution rate of the variance, determine the number of pivots, extract the eigenvectors corresponding to the largest eigenvalues, and calculate the projection tk of the eigenvector V on the feature space , which is the nonlinear pivot of matrix X.

The kernel principal component analysis method is used to reduce the number of control indicators, in which the sample variance reflects the size of the carried data information. Therefore, the number of indicators must be determined according to certain criteria, not only to avoid loss of data information, but also to effectively reduce the dimension of parameters. Here, the pivot is determined based on the variance cumulative contribution rate method.

The variance contribution rate of the i-th component xi in the data set X is:

${\mathrm{<span\; class="notranslate">\; PV</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

The variance cumulative contribution rate of the first k components is:

${\mathrm{<span\; class="notranslate">\; CPV</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; PV</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Where λ _i is the eigenvalue of the covariance matrix Σ, and λ ₁ ≥λ ₂ ...≥λ _m ≥0. The greater the contribution, the more important this component is. The cumulative contribution rate of the variance of the current k kernel pivots is over 90%. It is considered that the required kernel pivots can fully represent the original feature information. At this time, the number of kernel pivots is determined to be k.

Step 5 establishes a normal model with the nonlinear principal element extracted from the historical data of the normal operation of the gearbox, and trains it with a support vector machine.

In step 6, the nonlinear principal element extracted from the historical data of the gearbox operation in the later period is imported into the trained model, so that the fault of the gearbox can be identified and the identification conclusion can be drawn.

It can be seen that a wind turbine gearbox fault identification method of the present invention performs wavelet denoising processing on historical data through autocorrelation analysis, and extracts the time domain and frequency domain characteristic parameters in the denoising historical data through fast Fourier transform, and then The kernel principal component analysis method is used to reduce the dimensionality of the characteristic parameters, extract several nonlinear principal elements with the largest cumulative contribution rate of variance, and use the nonlinear principal elements extracted from the historical data of the normal operation of the gearbox to establish a normal model, and use the support The vector machine is trained, and finally the nonlinear principal element extracted from the historical data of the gearbox operation is imported into the trained model, so that the fault of the gearbox can be identified.

In summary, the wind turbine gearbox fault identification method of the present invention is based on the fault characteristics of the wind turbine gearbox, that is, there are a large number of nonlinear signals in the fault of the wind turbine gearbox, and the wind turbine gear is analyzed by the nuclear principal component analysis and the cumulative contribution rate of variance. Analyzing the historical data after the denoising of the vibration box is beneficial to the processing of nonlinear signals, and then using the support vector machine for fault classification is conducive to improving the processing ability of vibration signals, and improving the fault recognition rate on the basis of taking into account the effect of fault classification, especially It has stronger adaptability to a large number of nonlinear signals in wind turbine gearbox faults, which is of great significance for gearbox fault identification, and also has great significance for the maintenance and optimization cost of wind turbine gearboxes.

The above-mentioned embodiments only illustrate the principles and effects of the present invention, but are not intended to limit the present invention. Any person skilled in the art can modify and change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be listed in the claims.

Claims (10)

1. a gearbox of wind turbine fault recognition method, comprises the steps:

Step one, obtains the historical data that the gearbox of wind turbine within the scope of certain hour runs;

Step 2, adopts autocorrelation analysis to carry out wavelet noise process to historical data;

Step 3, by fast fourier transform, extracts the time domain in the historical data after de-noising and frequency domain character parameter;

Step 4, adopts core pivot element analysis method characteristic parameter to be carried out to the dimensionality reduction of dimension, extracts several nonlinear principal components that contribution rate of accumulative total of variance is maximum;

Step 5, the nonlinear principal component that the historical data normally run with gear case is extracted sets up normal model, and utilizes support vector machine to train

Step 6, the nonlinear principal component that the historical data run by later stage gear case is extracted imports the model after training, identifies thus to the fault of gear case.

2. a kind of gearbox of wind turbine fault recognition method as claimed in claim 1, it is characterized in that, step 2 comprises further:

Step 2.1, selects a suitable small echo and sets the level N of wavelet decomposition, then carrying out N layer wavelet decomposition to Noise signal;

Step 2.2, arrives each layer of high frequency coefficient of n-th layer, selects a threshold value to carry out soft-threshold quantification treatment to the 1st;

Step 2.3, according to the low frequency coefficient of the n-th layer of wavelet decomposition and after quantification treatment the 1st layer to the high frequency coefficient of n-th layer, carries out the wavelet reconstruction of one-dimensional signal.

3. a kind of gearbox of wind turbine fault recognition method as claimed in claim 1, it is characterized in that, step 3 comprises further:

Step 3.1, carries out Fast Fourier Transform (FFT) pre-service, extracts time domain charactreristic parameter and frequency domain character parameter;

Step 3.2, calculates 8 time domain indexes had the greatest impact;

Step 3.3, calculates 6 frequency-domain index;

Step 3.4; Time domain index after calculating and frequency-domain index are formed gear condition primitive character collection.

4. a kind of gearbox of wind turbine fault recognition method as claimed in claim 3, is characterized in that: the average of this time domain index involving vibrations signal, peak value, root-mean-square value, variance, kurtosis and the dimensionless group kurtosis factor, the pulse Summing Factor nargin factor.

5. a kind of gearbox of wind turbine fault recognition method as claimed in claim 3, is characterized in that: turn frequency amplitude, meshing frequency amplitude, meshing frequency double frequency amplitude, gravity frequency, all square frequency, the frequency variance of this frequency-domain index involving vibrations signal.

6. a kind of gearbox of wind turbine fault recognition method as claimed in claim 1, it is characterized in that, step 4 comprises further:

Step 4.1, chooses suitable kernel function;

Step 4.2, calculate nuclear matrix K;

Step 4.3, carries out eigendecomposition to this nuclear matrix K, and nuclear matrix K centralization is obtained K ';

Step 4.4, chooses normal sample notebook data, calculates its average and standard deviation, and to sample data standardization, builds training matrix X;

Step 4.5, calculates eigenwert and proper vector;

Step 4.6, calculates contribution rate of accumulative total of variance, determines pivot number, extracts maximum several eigenwert characteristic of correspondence vector V;

Step 4.7, calculates the projection t of proper vector V on feature space _k, i.e. the nonlinear principal component of matrix X.

7. a kind of gearbox of wind turbine fault recognition method as claimed in claim 6, is characterized in that, in step 4.6, in data set X, i-th component xi variance contribution ratio is:

${\mathrm{PV}}_i=\frac{{\mathrm{\&lambda;}}_i}{{\mathrm{\&Sigma;}}_{i=1}^m{\mathrm{\&lambda;}}_i}$

The contribution rate of accumulative total of variance of a front k component is:

${\mathrm{CPV}}_k={\mathrm{\&Sigma;}}_{i=1}^k{\mathrm{PV}}_i$

Wherein λ _ifor the eigenwert of covariance matrix Σ, and λ ₁>=λ ₂>=λ _m>=0.

8. a kind of gearbox of wind turbine fault recognition method as claimed in claim 7, it is characterized in that: if the contribution rate of accumulative total of variance of current k core pivot reaches more than 90%, then think that required core pivot can characterize primitive character information comprehensively, now definite kernel pivot number is k.

9. a kind of gearbox of wind turbine fault recognition method as claimed in claim 1, is characterized in that: in step one, obtains the gearbox of wind turbine input shaft within the scope of certain hour, historical data that output shaft amounts to 8 measuring points from data memory module.

10. a kind of gearbox of wind turbine fault recognition method as claimed in claim 1, is characterized in that: in step one, and the historical data of acquisition is input shaft, the axis of output shaft and the historical data of radial measuring point.