Fan bearing fault feature extraction method based on vibration signal density distribution

Abstract

This invention relates to a method for extracting fault features of wind turbine bearings based on vibration signal density distribution. It uses phase to quantify vibration signals such as vibration frequency and amplitude. Different faults yield different gradient feature maps. Based on PCA dimensionality reduction technology, these features are reduced to three dimensions, and a sample feature library is constructed. Compared with traditional methods, this method has stronger discriminative power. This invention focuses on the three key elements of vibration signals: vibration magnitude, vibration frequency, and vibration phase, without performing excessive feature extraction, thus reducing the influence of low-correlation factors on the results. Furthermore, it employs statistical theory to process the original vibration signal, highlighting the signal characteristics of vibration signals that lack strong features, thereby improving the accuracy of feature extraction for high-speed wind turbine bearings.

Description

Technical Field

[0001] This invention relates to wind turbine bearings, and in particular to a method for extracting fault features of high-speed wind turbine bearings based on vibration signal density distribution. Background Technology

[0002] Vibration signal degradation in high-speed wind turbine bearings is a significant indicator of bearing failure. Capturing locally degraded vibration signals can promptly identify potential faults in high-speed wind turbine bearings. Feature extraction is a prerequisite for fault type identification. Current methods, such as pulse waveform features and wavelet features, have high reliability in identifying signals with obvious fault characteristics. However, their reliability decreases when vibration signal characteristics are not obvious (e.g., pulse peaks are not prominent). Therefore, improvement and innovation in these methods are imperative.

[0003] This invention proposes a new signal quantization rule based on statistical graph method, which uses phase to quantify vibration signals such as vibration frequency and vibration amplitude. Different faults can yield different gradient feature maps. Based on PCA dimensionality reduction technology, these features are reduced to three dimensions to create a sample feature library. Compared with traditional methods, this method has stronger recognition accuracy. Summary of the Invention

[0004] In view of the above situation and to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method for extracting fault features of wind turbine bearings based on vibration signal density distribution, which can effectively solve the problem of improving the identification accuracy of bearing fault feature extraction.

[0005] The technical solution solved by this invention is:

[0006] A method for extracting fault features of wind turbine bearings based on vibration signal density distribution includes the following steps:

[0007] Step S1: Collect vibration signals

[0008] Vibration signals of all rolling bearings of the same model as the one from which the fault characteristics of the wind turbine are to be extracted are collected, and the vibration signals are processed into time-domain curves;

[0009] Step S2: Process bearing failure sample data

[0010] (1) Divide the degree of rotation of the rolling bearing into W equal intervals, where W≥8;

[0011] The vibration of a rolling bearing in one revolution is represented by a vibration histogram. Let T be the period of one revolution of the rolling bearing, i represent the i-th period, and j represent the j-th interval. The horizontal axis of the vibration histogram is the interval number, and the vertical axis is the vibration extracted from the time-domain curve. Each bar in each vibration histogram corresponds to one interval, and the height of the bar in the vibration histogram is the average value of the vibration amplitude within the corresponding interval.

[0012] If a bearing rotates n times, n vibration histograms can be obtained for each rolling bearing. The average height values ​​of the same intervals in each of the n vibration histograms are then calculated to obtain the average vibration amplitude of the bearing across different intervals after n rotations. This yields the average vibration amplitude histogram for the bearing after n rotations. Used to represent the distribution of the average vibration of a rolling bearing after quantization according to an angular range;

[0013] (2) The original vibration signal in each interval of the rolling bearing is subjected to fast Fourier transform to obtain the vibration spectrum of each interval; the frequency of the interval is counted in each vibration spectrum, and the frequency at the maximum of the vertical axis is taken as the vibration frequency of the interval, thus obtaining the vibration frequency of each interval.

[0014] The vibration frequency of a rolling bearing rotating one revolution is represented by a vibration frequency bar chart. The horizontal axis of the vibration frequency bar chart is the interval number, and the vertical axis is the vibration frequency. Each bar in each vibration frequency bar chart corresponds to one interval, and the height of the vibration frequency bar chart is the vibration frequency within the corresponding interval.

[0015] If a bearing rotates n times, n vibration frequency histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in each of the n vibration frequency histograms, the average vibration frequency of different intervals after n rotations of the bearing is obtained, resulting in a vibration frequency distribution histogram for the bearing after n rotations. Used to represent the distribution of vibration frequencies of rolling bearings after quantization according to angular intervals;

[0016] (3) Sum the vibration sample values ​​of each interval extracted from the time-domain curve to obtain the total vibration of each interval;

[0017] The total vibration of the rolling bearing in one revolution is represented by a total vibration bar chart. The horizontal axis of the total vibration bar chart is the interval number, and the vertical axis is the total vibration. Each bar in each total vibration bar chart corresponds to one interval, and the height of the total vibration bar chart is the total vibration within the corresponding interval.

[0018] If a bearing rotates n times, n total vibration histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in the n total vibration histograms, the average total vibration of the bearing in different intervals after n rotations is obtained, resulting in a total vibration distribution histogram for the bearing after n rotations. Used to represent the distribution of the total vibration of a rolling bearing after quantification according to an angular range;

[0019] (4) The peak-to-peak value of the vibration in each interval of the original vibration signal is the maximum vibration in each interval.

[0020] The maximum vibration of a rolling bearing in one revolution is represented by a bar chart of maximum vibration. The horizontal axis of the maximum vibration bar chart is the interval number, and the vertical axis is the maximum vibration. Each bar in each maximum vibration bar chart corresponds to one interval, and the height of the maximum vibration bar chart is the maximum vibration within the corresponding interval.

[0021] If a bearing rotates n times, n maximum vibration histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in the n maximum vibration histograms, the average maximum vibration of different intervals in the bearing after n rotations is obtained, resulting in a maximum vibration distribution histogram for the bearing after n rotations. Used to represent the distribution of the maximum vibration of a rolling bearing after quantization according to an angular range;

[0022] Step S3: Vibration signal feature extraction

[0023] (1) Transformation

[0024] According to the following formula, the average vibration amplitude histograms are plotted separately. Vibration frequency distribution histogram Total vibration distribution bar chart Maximum vibration distribution histogram The distribution is transformed into a density function, and the calculation formula is as follows:

[0025]

[0026] In the formula: y j p represents the ordinate value of each histogram in the j-th interval. j This represents the percentage of the total value in the j-th interval;

[0027] (2) Calculate the expected value and variance

[0028] The expected value μ of the density distribution is calculated using the following formula:

[0029]

[0030] In the formula: x j p represents the interval containing the x-axis and its corresponding angular value. j This represents the percentage of the total value in the j-th interval;

[0031] The variance of the density distribution is calculated using the following formula:

[0032]

[0033] (3) Calculate the skewness S k Steepness Ku, Number of peak points Pe:

[0034] Skewness and kurtosis are used to determine the symmetry and protrusion of the probability density distribution of a vibration signal relative to a Gaussian distribution. Their calculation methods are as follows:

[0035]

[0036]

[0037] Definition of peak point: In a sequence of vibration signals, if the value of a certain point is greater than the values ​​of the two adjacent points, then this point can be called the peak point of that interval. The number of peak points in a set of signals is denoted as Pe. The method for finding Pe is as follows:

[0038] Let the coordinates of a point in a set of signals be (x i y i The coordinates of the two points before and after it are: (x i-1 y i-1 ), (x i+1 y i+1 When the data at these three points satisfy the following formula, that point can be determined as a peak point:

[0039] y i -y i-1 >0 and y i+1 -y i <0 (7)

[0040] (4) Calculate similarity C:

[0041] The correlation coefficient method is used to calculate the similarity between two distributions, expressed by the following formula:

[0042]

[0043] In the formula: p j Let W represent the percentage of the total value in the j-th interval; let W represent the number of intervals.

[0044] At this point, the slope S has been extracted. k The four feature values ​​are: kurtosis Ku, number of peak points Pe, and similarity C.

[0045] (5) The composition method of the characteristic sequence:

[0046] From the average vibration amplitude histogram Extract a set of feature data [S] k1 K e1 P e1 [C1], and then from the vibration frequency distribution histogram Total vibration distribution bar chart Maximum vibration distribution histogram We can extract three more sets of feature data to form a 16-dimensional vector:

[0047] A = [S] k1 K e1 P e1 C1, S k2 K e2 P e2 C2, S k3 K e3 P e3 C3, S k4 K e4 P e4 C4]

[0048] (6) Feature sequence preprocessing and dimensionality reduction:

[0049] The PCA dimensionality reduction algorithm is used to reduce the dimensionality of the above 16-dimensional vector. Before dimensionality reduction, the sample data is preprocessed as follows:

[0050] There are four types of bearing samples: inner ring fault, outer ring fault, rolling element fault, and healthy. Let their corresponding numbers be A, B, C, and D. Rotating the bearing n times yields a set of 16-dimensional vectors. For an inner ring fault, denoted as A1, then k sets of 16-dimensional vectors are obtained. Therefore, n×k rotations are needed to obtain A1, A2, ..., A... k Similarly, for outer ring faults, rolling element faults, and healthy conditions, we can obtain: B1, B2...B k C1, C2......C k D1, D2......D k A total of 4×k groups of 16-dimensional vectors were generated to complete the data preprocessing.

[0051] After data preprocessing, the PCA dimensionality reduction algorithm is used to reduce the 4×k groups of 16-dimensional vectors to 4×k groups of 3-dimensional vectors. This completes the extraction of wind turbine bearing fault features. Based on the above method, four different types of 3-dimensional feature vectors can be obtained. Each type of feature vector represents a bearing state (fault or healthy). The 3-dimensional feature vectors obtained in the above steps can be used to establish a sample data feature library. For subsequent use, simply label each type of feature vector in the feature library with the fault type to serve as the basis for identifying unknown samples.

[0052] This invention proposes a novel signal quantization rule based on statistical graphing, using phase to quantify vibration signals such as vibration frequency and amplitude. Different faults yield different gradient feature maps. PCA dimensionality reduction technology is used to reduce these features to three dimensions, creating a sample feature library. Compared with traditional methods, this approach offers stronger discriminative power. This invention focuses on the three key elements of vibration signals: vibration magnitude, vibration frequency, and vibration phase, minimizing feature extraction and reducing the influence of less correlated factors. Furthermore, statistical theory is employed to process the original vibration signals, highlighting their characteristics even in signals lacking strong features, thus improving the accuracy of feature extraction for high-speed wind turbine bearings. Attached Figure Description

[0053] Figure 1 This is a time-domain curve of bearing vibration according to the present invention.

[0054] Figure 2 This is a schematic diagram illustrating the degree division of the rolling bearing of the present invention.

[0055] Figure 3 This is a bar chart showing the change in amplitude of a certain ring of the present invention with rotation angle (inner ring fault data).

[0056] Figure 4 This is a bar chart showing the average vibration amplitude of the present invention.

[0057] Figure 5 This is a schematic diagram of vibration frequency acquisition according to the present invention, wherein (a) is the original vibration signal within a certain interval; and (b) is the vibration spectrum within a certain interval.

[0058] Figure 6 This is a schematic diagram illustrating the extraction of the maximum vibration amount in this invention.

[0059] Figure 7 This is a diagram of the original sample data for an application example of the present invention.

[0060] Figure 8 This is a bar chart showing the average vibration amplitude in an application example of the present invention.

[0061] Figure 9 This is a distribution diagram of the maximum vibration in the inner fault zone of an application example of the present invention, where (a) is an amplified view of the sampled signal and (b) is a bar chart of the maximum vibration distribution.

[0062] Figure 10 This is a comparison chart showing the effect of fault feature extraction in an application example of the present invention. Detailed Implementation

[0063] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0064] Depend on Figure 1-6The present invention provides a method for extracting fault features of wind turbine bearings based on vibration signal density distribution, comprising the following steps:

[0065] Step S1: Collect vibration signals

[0066] Vibration signals of all rolling bearings of the same model as the wind turbine with the fault characteristics to be extracted (including normal bearings and bearings with inner and outer ring faults) are collected, and the vibration signals are processed into time domain curves.

[0067] like Figure 1 As shown, the bearing vibration time-domain signal shows that the inner ring fault signal has a significant pulse characteristic, which enables envelope spectrum analysis to effectively capture its fault characteristics and make classifications. However, envelope spectrum analysis may not be able to distinguish between the outer ring fault signal and the normal signal, which do not have obvious peak values.

[0068] Step S2: Process bearing failure sample data

[0069] (1) Divide the degree of rotation of the rolling bearing into W equal intervals, where W≥8;

[0070] A bearing rotates 360° in one revolution. Now, let's divide 360° into N intervals, such as... Figure 2 As shown, the degree of one circle is divided into 22.5° / parts, N=16;

[0071] The vibration of a rolling bearing in one revolution is represented by a vibration histogram. Let T be the period of one revolution of the rolling bearing, i represent the i-th period, and j represent the j-th interval. The horizontal axis of the vibration histogram is the interval number, and the vertical axis is the vibration extracted from the time-domain curve. Each bar in each vibration histogram corresponds to one interval, and the height of the bar in the vibration histogram is the average value of the vibration amplitude within the corresponding interval.

[0072] like Figure 3 The figure shown is a bar chart showing the change of amplitude with rotation angle for a certain circle (inner circle fault data). The method to obtain the average value of the vibration amplitude within the corresponding interval is as follows: amplitude sampling is performed in the time domain curve. For example, if there are three amplitude sampling points in the same interval (the number of sampling points depends on the sampling frequency), then the height of the bar chart is the average value of the three vibration amplitudes.

[0073] If a bearing rotates n times, n vibration histograms can be obtained for each rolling bearing. The average height values ​​of the same intervals in each of the n vibration histograms are then calculated to obtain the average vibration amplitude of the bearing across different intervals after n rotations. This yields the average vibration amplitude histogram for the bearing after n rotations. Used to represent the distribution of the average vibration of a rolling bearing after quantization according to an angular range;

[0074] (2) The original vibration signal in each interval of the rolling bearing is subjected to fast Fourier transform to obtain the vibration spectrum of each interval; the frequency of the interval is counted in each vibration spectrum, and the frequency with the maximum vertical axis is taken as the vibration frequency of the interval (because the larger the vertical axis value, the greater the proportion of this frequency in this vibration, which is the main frequency), thus obtaining the vibration frequency of each interval.

[0075] The vibration frequency of a rolling bearing rotating one revolution is represented by a vibration frequency bar chart. The horizontal axis of the vibration frequency bar chart is the interval number, and the vertical axis is the vibration frequency. Each bar in each vibration frequency bar chart corresponds to one interval, and the height of the vibration frequency bar chart is the vibration frequency within the corresponding interval.

[0076] If a bearing rotates n times, n vibration frequency histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in each of the n vibration frequency histograms, the average vibration frequency of different intervals after n rotations of the bearing is obtained, resulting in a vibration frequency distribution histogram for the bearing after n rotations. Used to represent the distribution of vibration frequencies of rolling bearings after quantization according to angular intervals;

[0077] like Figure 5 As shown in (b), the frequency at the maximum on the vertical axis is the vibration frequency f = 0.625 Hz. Repeating the above steps will yield a statistical chart of all frequencies for one revolution of the bearing (N = 16, resulting in 16 intervals). Representing the frequency distribution as a bar chart for n revolutions will yield the vibration frequency distribution. You can get and Figure 4 A similar histogram showing the distribution of vibration frequencies.

[0078] (3) Sum the vibration sample values ​​of each interval extracted from the time-domain curve to obtain the total vibration of each interval;

[0079] The total vibration of the rolling bearing in one revolution is represented by a total vibration bar chart. The horizontal axis of the total vibration bar chart is the interval number, and the vertical axis is the total vibration. Each bar in each total vibration bar chart corresponds to one interval, and the height of the total vibration bar chart is the total vibration within the corresponding interval.

[0080] If a bearing rotates n times, n total vibration histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in the n total vibration histograms, the average total vibration of the bearing in different intervals after n rotations is obtained, resulting in a total vibration distribution histogram for the bearing after n rotations. Used to represent the distribution of the total vibration of a rolling bearing after quantification according to an angular range;

[0081] The total vibration amount represents the sum of all vibration amounts within a certain interval. The average vibration amount distribution has already been calculated above, so the total vibration amount can also be calculated by multiplying the average vibration amount of each interval by the corresponding number of samples (the number of samples in each interval is different because the sampling frequency is the same, but the bearing speed may change during actual sampling, resulting in fewer samples in each interval at high speeds).

[0082] (4) The peak-to-peak value of the vibration in each interval of the original vibration signal is the maximum vibration in each interval.

[0083] The maximum vibration of a rolling bearing in one revolution is represented by a bar chart of maximum vibration. The horizontal axis of the maximum vibration bar chart is the interval number, and the vertical axis is the maximum vibration. Each bar in each maximum vibration bar chart corresponds to one interval, and the height of the maximum vibration bar chart is the maximum vibration within the corresponding interval.

[0084] If a bearing rotates n times, n maximum vibration histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in the n maximum vibration histograms, the average maximum vibration of different intervals in the bearing after n rotations is obtained, resulting in a maximum vibration distribution histogram for the bearing after n rotations. Used to represent the distribution of the maximum vibration of a rolling bearing after quantization according to an angular range;

[0085] The maximum vibration amplitude in a certain interval is represented by the peak-to-peak value of the vibration signal within that interval, such as... Figure 6 As shown, simply divide the original vibration signal into predetermined intervals and then count the peak-to-peak value within each interval.

[0086] Step S3: Vibration signal feature extraction

[0087] (1) Transformation

[0088] According to the following formula, the average vibration amplitude histograms are plotted separately. Vibration frequency distribution histogram Total vibration distribution bar chart Maximum vibration distribution histogram The distribution is transformed into a density function, and the calculation formula is as follows:

[0089]

[0090] In the formula: y j p represents the ordinate value of each histogram in the j-th interval. j This represents the percentage of the total value in the j-th interval;

[0091] (2) Calculate the expected value and variance

[0092] The expected value μ of the density distribution is calculated using the following formula:

[0093]

[0094] In the formula: x j p represents the interval containing the x-axis and its corresponding angular value. j This represents the proportion of the value in the j-th interval to the total value. To link the expected value and variance with the actual physical quantity, X is introduced. J The x-axis value represents the probability distribution, which is the interval index value in the spectrum (without specific physical meaning). x is defined as... j =x J ×2π / W;

[0095] The variance of the density distribution is calculated using the following formula:

[0096]

[0097] (3) Calculate the skewness S k Steepness Ku, Number of peak points Pe:

[0098] Skewness and kurtosis are commonly used to determine whether a signal distribution is symmetrical and whether the signal is Gaussian. In this invention, they are used to determine the symmetry and bulge of the probability density distribution of a vibration signal relative to a Gaussian distribution. The calculation method is as follows:

[0099]

[0100]

[0101] Definition of peak point: In a sequence of vibration signals, if the value of a certain point is greater than the values ​​of the two adjacent points, then this point can be called the peak point of that interval. The number of peak points in a set of signals is denoted as Pe. The method for finding Pe is as follows:

[0102] Let the coordinates of a point in a set of signals be (x i y i The coordinates of the two points before and after it are: (x i-1 y i-1 ), (x i+1 y i+1 When the data at these three points satisfy the following formula, that point can be determined to be a peak point:

[0103] and

[0104] Considering that the x-coordinates of the above points are actually arranged from 0 to 360°, forming an increasing sequence, i.e., x... i+1 >x i >xi-1 Therefore, the above formula can be simplified to:

[0105] y i -y i-1 >0 and y i+1 -y i <0 (7)

[0106] (4) Calculate similarity C:

[0107] The correlation coefficient method is used to calculate the similarity between two distributions, expressed by the following formula:

[0108]

[0109] In the formula: p i Let W represent the percentage of the total value in the j-th interval; let W represent the number of intervals.

[0110] The principle can be explained as follows: The similarity between two sets of distributed data is calculated by sliding multiplication, and the result is denoted as C. The value of C is defined as 0-1. When C=0, it indicates no correlation; C=1 indicates a linear correlation. C between 0 and 1 indicates a certain linear relationship.

[0111] At this point, the slope S has been extracted. k The four feature values ​​are: kurtosis Ku, number of peak points Pe, and similarity C.

[0112] (5) The composition method of the characteristic sequence:

[0113] From the average vibration amplitude histogram Extract a set of feature data [S] k1 K e1 P e1 [C1], and then from the vibration frequency distribution histogram Total vibration distribution bar chart Maximum vibration distribution histogram We can extract three more sets of feature data to form a 16-dimensional vector:

[0114] A = [S] k1 K e1 P e1 C1, S k2 K e2 P e2 C2, S k3 K e3 P e3 C3, S k4 K e4 P e4 C4]

[0115] (6) Feature sequence preprocessing and dimensionality reduction:

[0116] The PCA dimensionality reduction algorithm is used to reduce the dimensionality of the above 16-dimensional vector (a 16-dimensional vector has too high a dimension, and the mapping relationship for feature extraction is complex, resulting in extremely low accuracy; therefore, the PCA algorithm is used to reduce its dimensionality). Before dimensionality reduction, the sample data is preprocessed as follows:

[0117] There are four types of bearing samples: inner ring fault, outer ring fault, rolling element fault, and healthy. Let their corresponding numbers be A, B, C, and D. Rotating the bearing n times yields a set of 16-dimensional vectors. For an inner ring fault, denoted as A1, then k sets of 16-dimensional vectors are obtained. Therefore, n×k rotations are needed to obtain A1, A2, ..., A... k Similarly, for outer ring faults, rolling element faults, and healthy conditions, we can obtain: B1, B2...B k C1, C2......C k D1, D2......D k A total of 4×k groups of 16-dimensional vectors were generated to complete the data preprocessing.

[0118] As shown in the table below:

[0119]

[0120] After data preprocessing, the PCA dimensionality reduction algorithm is used to reduce the 4×k groups of 16-dimensional vectors to 4×k groups of 3-dimensional vectors. At this point, the extraction of wind turbine bearing fault features is completed.

[0121] For inner ring faults, the dimensionality-reduced feature vectors are represented as a1, a2, ..., a k ;

[0122] For outer ring faults, the dimensionality-reduced feature vectors are represented as b1, b2, ..., b k ;

[0123] For rolling element faults, the dimensionality-reduced feature vector is represented as c1, c2, ..., c. k ;

[0124] For health, the dimensionality-reduced feature vectors are represented as d1, d2, ..., d... k ;

[0125] The feature vectors of each sample after dimensionality reduction can be represented by the following table:

[0126]

[0127] Based on the above method, four different types of 3D feature vectors can be obtained. The set of feature vectors of each type represents a state of the bearing (fault or health). The 3D feature vectors finally obtained in the above steps can be used to build a sample data feature library. For subsequent use, it is only necessary to label each type of feature vector with a fault type in the feature library, which can then be used as a basis for identifying unknown samples.

[0128] In practical applications, this invention has achieved good technical results, as exemplified by data collected from the high-speed bearing of a 2MW wind turbine. Figure 7 The image shows the original sample data for an application example, with a sampling frequency f. s =9000Hz.

[0129] Through steps S1-S2 of the method of this invention, 16 bar charts can be obtained from the four types of samples. Considering that the data collected for inner ring faults is relatively small, a spectrum of the sample is generated every 100 revolutions of the bearing (n=100). Figure 7-9 As shown (only a portion of the sample spectrum is displayed).

[0130] After step S3, four sets of 16-dimensional vectors are obtained, denoted as A1, B2, C3, and D4; this is repeated 50 times (k = 50, i.e., 50 × 100 rotations, resulting in 50 sets of 16-dimensional feature vectors for each type). Then, the PCA dimensionality reduction algorithm is used to reduce the dimensionality of these 4 × 50 sets of 16-dimensional feature vectors, resulting in 4 × 50 sets of 3-dimensional feature vectors.

[0131] To demonstrate the high consistency of each feature extracted by this invention, the aforementioned three-dimensional vectors are represented as a three-dimensional graph, and compared with the results of traditional feature extraction methods. The results are as follows: Figure 10 As shown, the evaluation of a feature vector extraction method mainly depends on whether the extracted feature vectors have high consistency. It is clearly visible from the figure that the 3D scatter plot of the feature vectors obtained by this invention is significantly more concentrated and the features are clearly defined compared to traditional methods. This further verifies that this invention is superior to traditional fault feature extraction methods.

Claims

1. A method for extracting fault features of wind turbine bearings based on vibration signal density distribution, characterized in that, Includes the following steps: Step S1: Collect vibration signals Vibration signals of all rolling bearings of the same model as the one from which the fault characteristics of the wind turbine are to be extracted are collected, and the vibration signals are processed into time-domain curves; Step S2: Process bearing failure sample data (1) Divide the degree of rotation of the rolling bearing into W equal intervals, where W≥8; The vibration of a rolling bearing in one revolution is represented by a vibration histogram. Let T be the period of one revolution of the rolling bearing, i represent the i-th period, and j represent the j-th interval. The horizontal axis of the vibration histogram is the interval number, and the vertical axis is the vibration extracted from the time-domain curve. Each bar in each vibration histogram corresponds to one interval, and the height of the bar in the vibration histogram is the average value of the vibration amplitude within the corresponding interval. If a bearing rotates n times, n vibration histograms can be obtained for each rolling bearing. The average height values ​​of the same intervals in each of the n vibration histograms are then calculated to obtain the average vibration amplitude of the bearing across different intervals after n rotations. This yields the average vibration amplitude histogram for the bearing after n rotations. Used to represent the distribution of the average vibration of a rolling bearing after quantization according to an angular range; (2) The original vibration signal in each interval of the rolling bearing is subjected to fast Fourier transform to obtain the vibration spectrum of each interval; the frequency of the interval is counted in each vibration spectrum, and the frequency at the maximum of the vertical axis is taken as the vibration frequency of the interval, thus obtaining the vibration frequency of each interval. The vibration frequency of a rolling bearing rotating one revolution is represented by a vibration frequency bar chart. The horizontal axis of the vibration frequency bar chart is the interval number, and the vertical axis is the vibration frequency. Each bar in each vibration frequency bar chart corresponds to one interval, and the height of the vibration frequency bar chart is the vibration frequency within the corresponding interval. If a bearing rotates n times, n vibration frequency histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in each of the n vibration frequency histograms, the average vibration frequency of different intervals after n rotations of the bearing is obtained, resulting in a vibration frequency distribution histogram for the bearing after n rotations. Used to represent the distribution of vibration frequencies of rolling bearings after quantization according to angular intervals; (3) Sum the vibration sample values ​​of each interval extracted from the time-domain curve to obtain the total vibration of each interval; The total vibration of the rolling bearing in one revolution is represented by a total vibration bar chart. The horizontal axis of the total vibration bar chart is the interval number, and the vertical axis is the total vibration. Each bar in each total vibration bar chart corresponds to one interval, and the height of the total vibration bar chart is the total vibration within the corresponding interval. If a bearing rotates n times, n total vibration histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in the n total vibration histograms, the average total vibration of the bearing in different intervals after n rotations is obtained, resulting in a total vibration distribution histogram for the bearing after n rotations. Used to represent the distribution of the total vibration of a rolling bearing after quantification according to an angular range; (4) The peak-to-peak value of the vibration in each interval of the original vibration signal is the maximum vibration in each interval. The maximum vibration of a rolling bearing in one revolution is represented by a bar chart of maximum vibration. The horizontal axis of the maximum vibration bar chart is the interval number, and the vertical axis is the maximum vibration. Each bar in each maximum vibration bar chart corresponds to one interval, and the height of the maximum vibration bar chart is the maximum vibration within the corresponding interval. If a bearing rotates n times, n maximum vibration histograms can be obtained for each rolling bearing. By averaging the height values ​​of the same intervals in the n maximum vibration histograms, the average value of the maximum vibration in different intervals after n rotations of the bearing is obtained, resulting in a maximum vibration distribution histogram for the bearing after n rotations. Used to represent the distribution of the maximum vibration of a rolling bearing after quantization according to an angular range; Step S3: Vibration signal feature extraction (1) Transformation According to the following formula, the average vibration amplitude histograms are plotted separately. Vibration frequency distribution histogram Total vibration distribution bar chart Maximum vibration distribution histogram The distribution is transformed into a density function, and the calculation formula is as follows: In the formula: y j p represents the ordinate value of each histogram in the j-th interval. j This represents the percentage of the total value in the j-th interval; (2) Calculate the expected value and variance The expected value μ of the density distribution is calculated using the following formula: In the formula: x j p represents the interval of the x-axis. j This represents the percentage of the total value in the j-th interval. The variance of the density distribution is calculated using the following formula: (3) Calculate skewness S k , steepness Ku, peak point number Pe: Skewness and kurtosis are used to determine the symmetry and protrusion of the probability density distribution of a vibration signal relative to a Gaussian distribution. Their calculation methods are as follows: Definition of peak point: In a sequence of vibration signals, if the value of a certain point is greater than the values ​​of the two adjacent points, then this point can be called the peak point of that interval. The number of peak points in a set of signals is denoted as Pe. The method for finding Pe is as follows: Let the coordinates of a point in a set of signals be (x i ,y i The coordinates of the two points before and after it are: (x i-1 ,y i-1 ), (x i+1 ,y i+1 When the data at these three points satisfy the following formula, that point can be determined as a peak point: y i -y i-1 > 0 and y i+1 -y i < 0 (7) (4) Calculate similarity C: The correlation coefficient method is used to calculate the similarity between two distributions, expressed by the following formula: In the formula: p j Let W represent the percentage of the total value in the j-th interval; let W represent the number of intervals. So far, the slope S k , kurtosis Ku, peak point number Pe, similarity C, a total of four groups of characteristic values; (5) The composition method of the characteristic sequence: From the average vibration amplitude histogram Extract a set of feature data [S] k1 ,K e1 ,P e1 [C1], and then from the vibration frequency distribution histogram Total vibration distribution bar chart Maximum vibration distribution histogram We can extract three more sets of feature data to form a 16-dimensional vector: A=[S k1 ,K e1 ,P e1 ,C1,S k2 ,K e2 ,P e2 ,C2,S k3 ,K e3 ,P e3 ,C3,S k4 ,K e4 ,P e4 ,C4] (6) Feature sequence preprocessing and dimensionality reduction: The PCA dimensionality reduction algorithm is used to reduce the dimensionality of the above 16-dimensional vector. Before dimensionality reduction, the sample data is preprocessed as follows: There are four types of bearing samples: inner ring fault, outer ring fault, rolling element fault, and healthy. Let their corresponding numbers be A, B, C, and D. Rotating the bearing n times yields a set of 16-dimensional vectors. For an inner ring fault, denoted as A1, then k sets of 16-dimensional vectors are obtained. Therefore, n×k rotations are needed to obtain A1, A2, ..., A... k Similarly, for outer ring faults, rolling element faults, and healthy conditions, we can obtain: B1, B2...B k C1, C2......C k D1, D2......D k A total of 4×k groups of 16-dimensional vectors were generated to complete the data preprocessing. After data preprocessing, the PCA dimensionality reduction algorithm is used to reduce the 4×k groups of 16-dimensional vectors to 4×k groups of 3-dimensional vectors. At this point, the extraction of wind turbine bearing fault features is completed.

Images