CN106323635A - Rolling bearing fault on-line detection and state assessment method - Google Patents

Abstract

A rolling bearing fault on-line detection and state assessment method is disclosed. The method comprises the following steps: twelve dimensional dimensionless parameters are extracted; the twelve dimensional dimensionless parameters comprise six dimensional time domain statistical parameters, three dimensional frequency domain statistical parameters and three dimensional dimensionless parameters in a small wave envelope spectrum; standardized reconstruction characteristic vectors can be obtained; whether a rolling bearing malfunctions is determined, and a state of the rolling bearing is assessed. Via the rolling bearing fault on-line detection and state assessment method, the twelve dimensional dimensionless parameters which can be used for effectively representing the state of the rolling bearing can be automatically extracted, the twelve dimensional dimensionless parameters are subjected to decorrelation and standardization operation, standardized reconstruction characteristic vectors that are distributed to form a hypersphere with an original point being a sphere center, and fault detection and state assessment of the rolling bearing can be realized via 2-norms of the standardized reconstruction characteristic vectors; difficult problems of long on line training time, low efficiency, and hard-to-obtain fault samples and the like of a rolling bearing state assessing model can be solved.

Description

A kind of rolling bearing fault on-line checking and state evaluating method

Technical field

The invention belongs to rolling bearing fault Intelligent Measurement and state evaluating method field, be specifically related to a kind of rolling bearing On-line Fault Detection and state evaluating method.

Background technology

Owing to rolling bearing fault sample is often difficult to obtain, and bearing fault type complexity is various, and a few class faults may Exist simultaneously, so, what the state estimation of rolling bearing often faced is test in data domain problem, the feature evaluation i.e. taked Method shall apply to the situation of only normal sample.

Since the on-line monitoring of rolling bearing substantially can be regarded as the description problem to border, bearing normal data territory, So it is necessary to study multidimensional characteristic vectors distribution spatially, sets up a conjunction to better profiting from priori Suitable model.Generally, the dimension of feature is the highest (more than 3-dimensional), and its distribution cannot be carried out visual displaying, but according to low-dimensional feelings The popularization of condition, it is conceivable that the distribution of characteristic vector is super ellipsoids shape, and according to the difference of the feature selected, super ellipsoids main shaft Direction and length all it may happen that change.Such border is complicated, if wanting to describe complicated border, algorithm needs tool Having strong nonlinearity ability, the method such as Support Vector data description, self organizing neural network and gauss hybrid models has such Ability, and be used as solving such problem and all achieving good effect in test.

But just due to said method, there is strong nonlinearity ability, the most inevitably face a problem during its training: fortune Calculate complexity relatively big, and therefore, it is difficult to realize dynamic training.But, as a example by Aeroengine Ball Bearings, its state estimation Model is to be expected to be embedded in aeroengine control system, and engine control system needs optimized distribution modestly to calculate money Source ensures that the safety of vital task (such as electromotor control etc.), limited calculation resources constitute one with the complexity of model To contradiction.Therefore, the assessment models that structure is effective and computational complexity is low, it is possible to achieve the dynamic training of rolling bearing is with in real time Assessment, has important engineering significance.

No. 201310015619 Chinese patents disclose a kind of rolling bearing fault testing method based on vibration detection, first The rolling bearing data collected by acceleration transducer carry out 3 layers of WAVELET PACKET DECOMPOSITION, then solve third layer wavelet packet coefficient The energy of reconstruction signal, changes then according to third layer each band energy value, chooses the frequency range of energy concentration to reconstruct original letter Number approximate evaluation；Utilize cepstrum that reconstruction signal is further analyzed, the last and fault signature of Theoretical Calculation Frequency and side frequency Property comparison are with tracing trouble.This kind of method needs the fault data of rolling bearing, but either reality is examined Surveying or Theoretical Calculation, fault data is often difficult to obtain, and diagnosis process needs artificial participation, be unfavorable for automatic on-line training with Detection.

Summary of the invention

It is an object of the invention to provide rolling bearing fault detection and state evaluating method, the party of a kind of on-line training Method can solve the problem that Rolling Bearing Status assessment models on-line training is the longest, inefficient difficulties.

For achieving the above object, the present invention is achieved by the following technical solutions:

A kind of rolling bearing fault on-line checking and state evaluating method, comprise the following steps:

S1: extract 12 dimension dimensionless groups, described 12 dimension dimensionless groups include: 6 dimension Time-domain Statistics parameter, 3-dimensional frequency domains 3-dimensional dimensionless group in statistical parameter and Based on Wavelet Envelope spectrum；

S2: obtain standardization reconstruct characteristic vector；And

S3: judge whether described rolling bearing breaks down；Wherein,

Described step S1 specifically includes:

S1-1: gather vibration acceleration signal, the acceleration signal segment store that shakes that will gather, obtain some samples；

S1-2: extract 6 dimension Time-domain Statistics parameters, described 6 dimension Time-domain Statistics from described sample by Time-domain Statistics parameter Parameter includes: form factor T_SI, peak index T_CI, pulse index T_MI, margin index T_CLI, kurtosis T_KUWith flexure T_SK；

S1-3: extract 3-dimensional frequency domain statistical parameter from described sample by frequency domain statistical parameter, described 3-dimensional frequency domain is added up Parameter includes gravity frequency F_FC, mean square frequency F_MSFWith frequency variance F_VF；And

S1-4: by wavelet transformation obtain described sample Based on Wavelet Envelope compose, extract described Based on Wavelet Envelope spectrum in 3-dimensional without Dimensional parameters, the 3-dimensional dimensionless group in described Based on Wavelet Envelope spectrum includes: inner ring failure-frequency f_ICharacteristic of correspondence W_BPFI, outer ring Failure-frequency f_OCharacteristic of correspondence W_BPFOWith ball failure-frequency f_BCharacteristic of correspondence W_BSF；

Described step S2 specifically includes:

S2-1: based on minimal reconstruction error criterion, above-mentioned 12 dimension dimensionless groups are applied constraints, obtains 12 dimensions and go Relevant reconstruct characteristic vector；

S2-2: be standardized processing to the reconstruct characteristic vector of described 12 dimension decorrelations, it is thus achieved that 12 dimension sample averages are 0 And the standardization reconstruct characteristic vector that sample standard deviation is 1；And

Described step S3 specifically includes:

S3-1: training, obtains the sample of 2 norms of described standardization reconstruct characteristic vector under rolling bearing normal operating condition This average and sample standard deviation, and set the threshold value that Rolling Bearing Status is abnormal；

S3-2: test, by the characteristic vector of sample under rolling bearing unknown state to the properly functioning shape of described rolling bearing The Euclidean distance of the sample average characteristic vector under state, as deterioration index, compares deterioration index and described threshold value, when deterioration refers to When mark is more than described threshold value, it is determined that rolling bearing is in malfunction；When deteriorating index less than or equal to described threshold value, sentence Determine rolling bearing and be in normal operating condition.

Further, described deterioration index is the biggest, then judge that the degradation of rolling bearing is the biggest.

Further, the concrete operations extracting described 6 dimension Time-domain Statistics parameters in step S1-2 are:

Form factor T_SIBe given by formula (1), peak index T_CIBe given by formula (2), pulse index T_MIBy formula (3) Be given, margin index T_CLIBe given by formula (4), kurtosis T_KUBe given by formula (5), flexure T_SKBe given by formula (6),

$T_{SI}=\frac{\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left(y_i^2\right)}}{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left|y_i\right|}---\left(1\right)$

$T_{CI}=\frac{\underset{i=1}{\overset{10}{\mathrm{\Σ}}}{y}_{pi}}{\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(y_i\right)}^2}}---\left(2\right)$

$T_{MI}=\frac{\underset{i=1}{\overset{10}{\mathrm{\Σ}}}{y}_{pi}}{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left|y_i\right|}---\left(3\right)$

$T_{CLI}=\frac{\underset{i=1}{\overset{10}{\mathrm{\Σ}}}{y}_{pi}}{{\[\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{\left|y_i\right|}\]}^2}---\left(4\right)$

$T_{KU}=\frac{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^4}{{\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^2\right)}^2}---\left(5\right)$

$T_{SK}=\frac{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^3}{{\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^2\right)}^{\frac{3}{2}}}---\left(6\right)$

Wherein, y_iShake described in being the data in acceleration signal；y_piIt is that after described data are divided into 10 sections, every segment data is exhausted Maximum to value.

Further, the concrete operations extracting 3-dimensional Time-domain Statistics parameter in step S1-3 are:

Gravity frequency F_FCBe given by formula (7), mean square frequency F_MSFBe given by formula (8), frequency variance F_VFBe given by formula (9),

$F_{FC}=\frac{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{f}_{i}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S\left(f_i\right)}---\left(7\right)$

$F_{MSF}=\frac{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{f}_i^{2}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S\left(f_i\right)}---\left(8\right)$

$F_{VF}=\frac{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{(f_i-F_{FC})}^{2}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S\left(f_i\right)}---\left(9\right)$

Wherein, S (f_i) it is vibration acceleration signal frequency spectrum function.

Further, the concrete operations of described step S1-4 include following three steps:

S1-4-1, calculates inner ring failure-frequency f according to formula (10)-(12)_I, outer ring failure-frequency f_OWith ball failure-frequency f_B,

$f_I=\frac{1}{2}Z(1+\frac{d}{D}cos\mathrm{\α})f_r---\left(10\right)$

$f_O=\frac{1}{2}Z(1-\frac{d}{D}cos\mathrm{\α})f_r---\left(11\right)$

$f_B=\frac{D}{2d}\[1-{\left(\frac{d}{D}\right)}^2{\cos }^2\mathrm{\α}\]f_r---\left(12\right)$

Wherein d represents rolling element diameter, and D represents bearing pitch diameter, and Z represents rolling element number, and α represents contact angle, f_rRepresent Rolling bearing inner ring and outer ring relatively rotate frequency；

S1-4-2, uses db8 small echo that described vibration acceleration signal carries out wavelet decomposition and obtains 6 signals, including 5 Detail signal d1, d2, d3, d4, d5 and 1 approximate signal a5；

S1-4-3, extracts the 3-dimensional dimensionless group in Based on Wavelet Envelope spectrum: be located in wavelet packet envelope spectrum, fault signature frequency With the presence of feature spectral peak near rate and each rank frequency multiplication thereof, if a width of f of envelope spectrum analysis band_e, envelope spectrum is W (f), if W (f) spectral line Number be N_e, then S_eaFor

$S_{ea}=\frac{1}{N_e}\underset{i=0}{\overset{N_e}{\mathrm{\Σ}}}W\left(f_i\right)---\left(13\right)$

Make S again_edFor the spectral line meansigma methods at the frequency multiplication of each rank of fault characteristic frequency in envelope spectrum, if failure-frequency in envelope spectrum Spectral line number be n_e, then

$S_{ed}=\frac{1}{n_e}\underset{i=0}{\overset{n_e}{\mathrm{\Σ}}}W\left({\mathrm{if}}_d\right)---\left(14\right)$

Construct a dimensionless group:

${\mathrm{\ΔS}}_e=\frac{S_{ed}}{S_{ea}}---\left(15\right)$

Described in step S1-4-2 6 signal is carried out envelope spectrum analysis respectively, obtains inner ring failure-frequency f_ICorresponding Feature W_BPFI, outer ring failure-frequency f_OCharacteristic of correspondence W_BPFO, ball failure-frequency f_BCharacteristic of correspondence W_BSF。

Further, the concrete operations of described step S2-1 include following three steps:

S2-1-1: be standardized processing to 12 dimension dimensionless groups, obtain x_i, then have Σ_i x_i=0；

S2-1-2: assuming that the new coordinate system obtained after projective transformation is { w₁,w₂,…,w_d, wherein, d is intrinsic dimensionality；Execute The wi that is constrained to added is normal orthogonal base vector, meets | | w_i||₂=1, w_i ^Tw_j=0 (i ≠ j), optimization object function is

$\begin{array}{cc}\underset{W}{\min }& -tr\left(W^T{\mathrm{XX}}^{T}W\right)\\ s.t.& W^{T}W=I\end{array}---\left(16\right)$

S2-1-3: solve formula (16) and obtain projection matrix W, be calculated the reconstruct characteristic vector of decorrelation according to formula (17) z_i

z_i=(z_i1,z_i2,…,z_id), z_id=w_j ^Tx_i (17)。

Further, described step S2-2 Playsization reconstruct characteristic vector z_i ^*Be given by formula (18)

$z_i^{*}=(z_{i1}^{*},z_{i2}^{*},\mathrm{...},z_{id}^{*}),z_{ij}^{*}=\frac{z_{ij}-{\widehat{\mathrm{\μ}}}_j}{{\widehat{\mathrm{\σ}}}_j}---\left(18\right)$

Wherein

${\widehat{\mathrm{\μ}}}_j=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{z}_{ij},{\widehat{\mathrm{\σ}}}_j^2=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(z_{ij}-{\widehat{\mathrm{\μ}}}_j)}^2---\left(19\right).$

Further, the described threshold value in step S3-1 is given by formula (20)

$D_{max}={\widehat{\mathrm{\μ}}}_D+3{\widehat{\mathrm{\σ}}}_D---\left(20\right)$

Wherein,WithIt is D respectively_iSample average and sample standard deviation.

Compared with prior art, the method have the advantages that

The online test method of a kind of rolling bearing fault that the present invention proposes, the method can automatically extract out Efficient Characterization The multidimensional dimensionless group of bearing state, and by primitive character is carried out decorrelation and standardization, obtain with initial point be The standardization reconstruct characteristic vector of the super spherical distribution of the centre of sphere, it is only necessary to 2 norms of normalized reconstruct characteristic vector just can be real The fault detect of existing rolling bearing and state estimation, have following differences in the significant advantage of traditional method:

1) in terms of Feature Fusion, it is many that the present invention fully combines temporal signatures, frequency domain character and Based on Wavelet Envelope spectrum signature Dimension information, and expressed by one-dimensional deterioration index, efficiently solve the different characteristic sensitivity to different faults traditionally The inconsistent troubleshooting issue brought；

2) in terms of model complexity, the present invention improves the sky of characteristic vector with looking for another way by linear projection conversion Between be distributed, and therefore greatly simplifie the complexity of sorter model, simultaneously as without adjusting parameter and iterative computation, The present invention can realize rolling bearing fault detection and the on-line training of state estimation model and dynamically renewal well；

3) in terms of the Data Source of model training, the present invention only needs the properly functioning data of rolling bearing to instruct for model Practice, it is to avoid traditional method training needs rolling bearing fault data, and fault data is often difficult to the defect that obtains；

4) seeing on the whole, the present invention fully combines multidimensional characteristic information, improves characteristic vector by eigentransformation Spatial distribution, can improve fault recognition rate, closer to engineering demand while effective simplified model complexity further.

Accompanying drawing explanation

Fig. 1 is the method flow diagram of the present invention；

Fig. 2 (a) is the distribution of incoherent two-dimensional feature space；

Fig. 2 (b) is relevant two-dimensional feature space distribution；

The distribution of Fig. 3 data and the inconsistent error schematic diagram brought of description；

Fig. 4 is Distance Discrimination Analysis schematic diagram；

Fig. 5 is that Rolling Bearing Status assesses schematic diagram；

Fig. 6 (a) is the inventive method scatterplot to bearing different faults testing result；

Fig. 6 (b) is the Support Vector data description method scatterplot to bearing different faults testing result；And

Fig. 6 (c) is the self organizing neural network method scatterplot to bearing different faults testing result.

Detailed description of the invention

Below in conjunction with the accompanying drawings the content of rolling bearing fault on-line checking a kind of to the present invention and state evaluating method make into One step explanation.

As it is shown in figure 1, the invention discloses a kind of rolling bearing fault on-line checking and state evaluating method, concrete by with Lower step is implemented:

S1:12 dimension dimensionless group extracts, and specifically includes following four step:

S1-1: gather vibration acceleration signal, the vibration acceleration signal segment store that will gather, obtain some samples；

S1-2: calculated by Time-domain Statistics parameter and extract 6 dimension Time-domain Statistics parameters, including form factor T_SI, peak index T_CI, pulse index T_MI, margin index T_CLI, kurtosis T_KU, flexure T_SK；

S1-3: calculated by frequency domain statistical parameter and extract 3-dimensional Time-domain Statistics parameter, including gravity frequency F_FC, mean square frequency F_MSF, frequency variance F_VF；

S1-4: obtain Based on Wavelet Envelope by wavelet transformation and compose, extracts the corresponding inner ring failure-frequency f of Based on Wavelet Envelope spectrum_ICorresponding Feature W_BPFI, outer ring failure-frequency f_OCharacteristic of correspondence W_BPFO, ball failure-frequency f_BCharacteristic of correspondence W_BSF；

S2: obtain standardization reconstruct characteristic vector, specifically include following two step:

S2-1: based on minimal reconstruction error criterion, applies corresponding constraints, obtain decorrelation 12 dimension reconstruct features to Amount；

S2-2: by 12 dimension reconstruct characteristic vectors be standardized process, obtain each feature samples average be 0, sample Standard deviation is the standardization reconstruct characteristic vector of 1；

The detection of S3: rolling bearing fault based on Distance Discrimination Analysis and state estimation, specifically include following two step Rapid:

S3-1: training, is calculated the sample standard deviation of the 2-norm of sampling feature vectors under rolling bearing normal operating condition Value and sample standard deviation, the threshold value that definition Rolling Bearing Status is abnormal；

S3-2: test, is calculated under rolling bearing unknown state sampling feature vectors to normal sample mean vector Euclidean distance, as deterioration index, compares the threshold value that deterioration index sets with S3-1, it is judged that whether bearing breaks down, and assesses The degradation of bearing.

The concrete operations of described step S1-2 are:

Form factor T_SIBe given by formula (1), peak index T_CIBe given by formula (2), pulse index T_MIBe given by formula (3), abundant Degree index T_CLIBe given by formula (4), kurtosis T_KUBe given by formula (5), flexure T_SKBe given by formula (6), wherein, y_iIt it is the acceleration letter that shakes Vibration data in number i.e. initial data；y_piIt is that initial data is divided into the maximum of every segment data absolute value after 10 sections.

$T_{SI}=\frac{\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left(y_i^2\right)}}{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left|y_i\right|}---\left(1\right)$

$T_{CI}=\frac{\underset{i=1}{\overset{10}{\mathrm{\Σ}}}{y}_{pi}}{\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left(y_i\right)}^2}}---\left(2\right)$

$T_{MI}=\frac{\underset{i=1}{\overset{10}{\mathrm{\Σ}}}{y}_{pi}}{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left|y_i\right|}---\left(3\right)$

$T_{CLI}=\frac{\underset{i=1}{\overset{10}{\mathrm{\Σ}}}{y}_{pi}}{{\[\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\sqrt{\left|y_i\right|}\]}^2}---\left(4\right)$

$T_{KU}=\frac{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^4}{{\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^2\right)}^2}---\left(5\right)$

$T_{SK}=\frac{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^3}{{\left(\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{y_i}^2\right)}^{\frac{3}{2}}}---\left(6\right)$

The concrete operations of described step S1-3 are:

Gravity frequency F_FCBe given by formula (7), mean square frequency F_MSFBe given by formula (8), frequency variance F_VFBe given by formula (9), Wherein, S (f_i) it is vibration acceleration signal frequency spectrum function

$F_{FC}=\frac{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{f}_{i}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S\left(f_i\right)}---\left(7\right)$

$F_{MSF}=\frac{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{f}_i^{2}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S\left(f_i\right)}---\left(8\right)$

$F_{VF}=\frac{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{(f_i-F_{FC})}^{2}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\mathrm{\Σ}}}S\left(f_i\right)}---\left(9\right)$

The concrete operations of described step S1-3 include following three steps:

S1-3-1) according to formula (10)-(12) calculating fault features frequency, wherein d represents rolling element diameter, and D represents bearing Pitch diameter, Z represents rolling element number, and α represents contact angle, f_rRepresent rolling bearing inner ring and outer ring relatively rotates frequency

$f_I=\frac{1}{2}Z(1+\frac{d}{D}cos\mathrm{\α})f_r---\left(10\right)$

$f_O=\frac{1}{2}Z(1-\frac{d}{D}cos\mathrm{\α})f_r---\left(11\right)$

$f_B=\frac{D}{2d}\[1-{\left(\frac{d}{D}\right)}^2{\cos }^2\mathrm{\α}\]f_r---\left(12\right)$

S1-3-2) use db8 small echo vibration acceleration signal is carried out wavelet decomposition obtain 5 detail signal d1, d2, D3, d4, d5 and 1 approximate signal a5；

S1-3-3) the automatically extracting of Based on Wavelet Envelope spectrum signature: be located in wavelet packet envelope spectrum, fault characteristic frequency and each With the presence of feature spectral peak near the frequency multiplication of rank, if a width of f of envelope frequency spectrum analytic band_e, envelope spectrum is W (f), if the number of W (f) spectral line For N_e, then S_eaFor

$S_{ea}=\frac{1}{N_e}\underset{i=0}{\overset{N_e}{\mathrm{\Σ}}}W\left(f_i\right)---\left(13\right)$

Make S again_edFor the spectral line meansigma methods at the frequency multiplication of each rank of fault characteristic frequency in envelope spectrum, if failure-frequency in envelope spectrum Spectral line number be n_e, then

$S_{ed}=\frac{1}{n_e}\underset{i=0}{\overset{n_e}{\mathrm{\Σ}}}W\left({\mathrm{if}}_d\right)---\left(14\right)$

Construct a dimensionless group:

${\mathrm{\ΔS}}_e=\frac{S_{ed}}{S_{ea}}---\left(15\right)$

To step S1-3-2) 6 signals carry out envelope spectrum analysis respectively, corresponding inner ring can be obtained by automatically calculating Failure-frequency f_ICharacteristic of correspondence W_BPFI, outer ring failure-frequency f_OCharacteristic of correspondence W_BPFO, ball failure-frequency f_BCharacteristic of correspondence W_BSF。

After being standardized the feature extracted processing, sample average μ=0 of each feature, sample standard deviation σ=1, because of Its distribution can preferably be compared in same yardstick by this.It can be seen that the border of feature distribution can be near in Fig. 2 (a) Like describing with a circle (in figure, the radius of circle is 2.5), but in Fig. 2 (b), feature distribution presents obvious ellipticity, and oval Main shaft and coordinate main shaft be approximated to 45° angle.Take two kinds from 12 dimensional features (i.e. 12 dimension dimensionless group) times to be combined having 66 kinds of situations, but all without departing from two kinds of distributions in shown in Fig. 2.The distribution of Fig. 2 (a) can preferably be described by a circle, but When describing the distribution of similar Fig. 2 (b), can inevitably bring two class errors, as shown in Figure 3.Wherein, error of first kind is Mistakenly fault sample is identified as caused by normal sample, i.e. false positive example；Error of the second kind is mistakenly by normal sample Originally it is judged as caused by fault sample, i.e. false counter-example.

Further Fig. 2 (b) is carried out it has been observed that be distributed and why present ellipticity and be because between feature also existing bigger Dependency, i.e. when gravity frequency is bigger, frequency variance is the biggest, comprises gravity frequency in frequency variance formula, and this is The basic reason that its dependency exists.More generally, between feature, dependency can be weighed with correlation coefficient, correlation coefficient ρ_xyDetermine Justice is as shown in formula (16):

${\mathrm{\ρ}}_{xy}=\frac{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i{y}_i-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\·\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i}{\sqrt{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i^2-{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i\right)}^2}\·\sqrt{n\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i^2-{\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i\right)}^2}}---\left(16\right)$

In formula, x_i,y_iThe eigenvalue of corresponding i-th sample different characteristic, n is sample size.In Fig. 2 between character pair Correlation coefficient has marked in figure.

The conclusion of above-mentioned analysis can extend to the space of more higher-dimension: if each feature the most not phase being described can be promoted Close, then characteristic vector distribution in hyperspace can be hypersphere shape, it is possible to be described with a hypersphere simply； Otherwise, the distribution of characteristic vector then presents super ellipsoids shape, is not easy to describe.

The concrete operations of described step S2-1 include following three steps:

S2-1-1) assume to be standardized processing to the 12 dimension the most former characteristic vectors of dimensionless group, obtain x_i, then have Σ_ix_i=0；

S2-1-2) the new coordinate system obtained after supposing projective transformation is { w₁,w₂,…,w_d}.Wherein, d is intrinsic dimensionality；Execute The wi that is constrained to added is normal orthogonal base vector, meets | | w_i||₂=1, w_i ^Tw_j=0 (i ≠ j), if sample point x_iIn new coordinate system In be projected as z_i=(z_i1,z_i2,…,z_id), wherein z_id=w_j ^Tx_iIt is x_iThe coordinate of the jth dimension under new coordinate system.Based on z_i Reconstruct x_i, then have:

${\hat{x}}_i=\underset{j=1}{\overset{d}{\mathrm{\Σ}}}{z}_{ij}{w}_j---\left(17\right)$

Consider whole training set, former characteristic vector x_iWith reconstruct characteristic vectorBetween distance be:

$\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\right|\underset{j=1}{\overset{d}{\mathrm{\Σ}}}{z}_{ij}{w}_j-x_i|{|}_2^2=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{z}_i^T{z}_i-2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{z}_i^T{W}^T{x}_i+const\∝-tr\left(W^T\right(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i{x}_i^T\left)W\right)---\left(18\right)$

According to minimal reconstruction error criterion, formula (18) should be minimized.In view of w_jIt is orthonormal basis, Σ_ix_ix_i ^TIt it is association Variance matrix, therefore, optimization object function is

$\begin{array}{cc}\underset{W}{\min }& -tr\left(W^T{\mathrm{XX}}^{T}W\right)\\ s.t.& W^{T}W=I\end{array}---\left(19\right)$

S2-1-3) solve formula (19) and obtain projection matrix W, be calculated reconstruct characteristic vector z of decorrelation according to formula (20)_i

z_i=(z_i1,z_i2,…,z_id), z_id=w_j ^Tx_i (20)

Reconstruct characteristic vector z obtained by decorrelation_iDistribution be still super ellipsoids shape, but the major axes orientation of super ellipsoids Approximate parallel with coordinate axes.Therefore, by being standardized processing to reconstruct characteristic vector further, can be by characteristic vector The suprasphere distribution that it is the hypersphere heart with zero that distribution is converted to.

Described step S2-2 Playsization reconstruct characteristic vector z_i* be given by formula (21)

$z_i^{*}=(z_{i1}^{*},z_{i2}^{*},\mathrm{...},z_{id}^{*}),z_{ij}^{*}=\frac{z_{ij}-{\widehat{\mathrm{\μ}}}_j}{{\widehat{\mathrm{\σ}}}_j}---\left(21\right)$

Wherein,

${\widehat{\mathrm{\μ}}}_j=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{z}_{ij},{\widehat{\mathrm{\σ}}}_j^2=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(z_{ij}-{\widehat{\mathrm{\μ}}}_j)}^2---\left(22\right)$

Distance Discrimination Analysis is on the premise of classification determines, uses statistical analysis means, according to each feature of new samples Its ownership of the distance discrimination of value and known class.Its basic thought be exactly calculate certain individual and each overall between distance, and recognize For this individuality belong to closest with it totally.

In d dimensional feature space, it is considered to use a d dimensional vector x₀Represent n sample point of certain sample set, Wo Menxi Hope this vector x₀With each sample i (i=1 ..., square distance sum n) is the smaller the better, define square error criterion function J₀(x₀) as follows:

$J_0\left(x_0\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\right|x_0-x_i|{|}^2---\left(23\right)$

Then

$x_0=\widehat{\mathrm{\μ}}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i---\left(24\right)$

This conclusion may certify that as follows:

$\begin{array}{c}{J}_0\left(x_0\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\right|(x_0-\widehat{\mathrm{\μ}})-(x_i-\widehat{\mathrm{\μ}})|{|}^2\\ =\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\right|x_0-\widehat{\mathrm{\μ}}|{|}^2-2{(x_0-\widehat{\mathrm{\μ}})}^T\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(x_i-\widehat{\mathrm{\μ}})+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}||x_i-\widehat{\mathrm{\μ}}|{|}^2\\ =\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left|\right|x_0-\widehat{\mathrm{\μ}}|{|}^2+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}||x_i-\widehat{\mathrm{\μ}}|{|}^2\end{array}---\left(25\right)$

Wherein, the right Section 2 of formula (25) does not relies on x₀, so expression formula minimalization under the conditions of formula (24).Cause This, the sample average of normal sample can be as the zero dimension description to normal sample.

In the case of the most normal sample, problem is converted to: calculate characteristic vector to normal sample according to certain measurement criterion Distance D of this mean vector_i, and with arrange distance threshold D_maxIt is compared to judge bearing state (sample ownership), the most right In arbitrary sample, if meeting D_i≤D_max, then this sample is classified as normally, otherwise classifies as exception.As shown in Figure 4, distance is sentenced The border not described be radius be D_maxHypersphere, according to Euclidean distance as measurement criterion, then have

$D_i=\left|\right|z_i^{*}|{|}_2---\left(26\right)$

Meanwhile, D_iThere is clear and definite meaning, the degradation of rolling bearing can be reflected, can be as rolling bearing health shape The evaluation index of state, as shown in Figure 5.Damage of the bearing is the most serious, the D of corresponding sample_iAlso can be the biggest.Further, by accordingly The multiple threshold value of Standard-making can provide the warning limit of bearing, abnormal limit etc..

In described step S3-1, the formulation of threshold value is given by formula (27)

$D_{max}={\widehat{\mathrm{\μ}}}_D+3{\widehat{\mathrm{\σ}}}_D---\left(27\right)$

Wherein,WithIt is D respectively_iSample average and sample standard deviation.

The aeroengine rotor exerciser using the band casing of Shenyang electromotor design and research institute development carries out fault mould Intend test to verify effectiveness of the invention.

This exerciser, on structure designs, first considers consistent with the casing of engine core machine in shape, and size contracts Little is 1/3, and internal structure has made necessary simplification: core engine is reduced to 020 supporting structure forms, and multistage compressor simplifies For the disc structure of single-stage, blade is reduced to tilting plane.In view of in true aero-engine, sensor is difficult to be placed in On bearing block, therefore in test vibration acceleration sensor is arranged in turbine casing both vertically and horizontally.Test In, vibration signal is acquired by NI USB9234 data acquisition unit, and acceleration transducer model is B&K 4805, sampling frequency Rate is 10.24kHz.

Subjects is 6206 type ball bearings, and basic geometric parameters is as shown in table 1.Use Wire EDM the most right The impact that housing washer, inner ring raceway and rolling element processing groove are given birth to simulation damage of the bearing.Rated speed in test For 1500rpm.

The physical dimension (unit: mm) of table 1 bearing

As a example by the vibration acceleration signal that measuring point records above casing, according to step 1) it is extracted 12 (dimension) dimensionless Parameter.

The effectiveness of extracting method in order to verify, by neural with Support Vector data description and self-organizing for the method for the present invention Network compares.Assessment models all uses the half of normal sample to be trained, and remaining sample is used for as unknown sample Test (containing 110 samples under every kind of state).

D in modus ponens of the present invention (26)_iIndex is deteriorated, shown in the scatterplot obtained such as Fig. 6 (a) as Rolling Bearing Status； The MATLAB workbox LibSVM-3.18-SVDD that Support Vector data description method is expanded by LibSVM software kit realizes, core letter Number takes radially base core, and relevant parameter is selected by cross validation, uses decision value to deteriorate index as bearing, and the scatterplot obtained is such as Shown in Fig. 6 (b)；SOM method is realized by MATLAB-SOM workbox, and neuron number and iterations are selected by cross validation, Use characteristic vector to the minimal matching span of neuron as bearing state evaluation index, the scatterplot obtained such as Fig. 6 (c) institute Show.Each method is shown in Table 2 to the classification accuracy of bearing different conditions sample.

The accuracy of table 2 distinct methods classification

It can be seen that the inventive method ensure that higher level on classification accuracy rate, and it is substantially better than support vector number According to describing and the result of self organizing neural network.

The method of the present invention has the advantage without adjusting ginseng, the miscellaneous degree of model is little, can preferably solve rolling bearing and supervise online The problem dynamically updated of assessment models in survey；Meanwhile, the method combines time domain, frequency domain and the 12 of Based on Wavelet Envelope spectrum fully Effective information in dimensional feature so that the accuracy of fault detect has obtained further raising.

Last it is noted that above-described each embodiment is merely to illustrate technical scheme, rather than to it Limit；Although the present invention being described in detail with reference to previous embodiment, it will be understood by those within the art that: Technical scheme described in previous embodiment still can be modified by it, or enters wherein part or all of technical characteristic Row equivalent；And these amendments or replacement, do not make the essence of appropriate technical solution depart from various embodiments of the present invention technical side The scope of case.

Claims (8)

1. a rolling bearing fault on-line checking and state evaluating method, it is characterised in that: comprise the following steps:

S1: extract 12 dimension dimensionless groups, described 12 dimension dimensionless groups include: 6 dimension Time-domain Statistics parameters, 3-dimensional frequency domain statistics 3-dimensional dimensionless group in parameter and Based on Wavelet Envelope spectrum；

S2: obtain standardization reconstruct characteristic vector；And

S3: judge whether described rolling bearing breaks down；Wherein,

Described step S1 specifically includes:

S1-1: gather vibration acceleration signal, the acceleration signal segment store that shakes that will gather, obtain some samples；

S1-2: extract 6 dimension Time-domain Statistics parameters, described 6 dimension Time-domain Statistics parameters from described sample by Time-domain Statistics parameter Including: form factor T_SI, peak index T_CI, pulse index T_MI, margin index T_CLI, kurtosis T_KUWith flexure T_SK；

S1-3: extract 3-dimensional frequency domain statistical parameter, described 3-dimensional frequency domain statistical parameter from described sample by frequency domain statistical parameter Including gravity frequency F_FC, mean square frequency F_MSFWith frequency variance F_VF；And

S1-4: the Based on Wavelet Envelope being obtained described sample by wavelet transformation is composed, extracts the 3-dimensional dimensionless in described Based on Wavelet Envelope spectrum Parameter, the 3-dimensional dimensionless group in described Based on Wavelet Envelope spectrum includes: inner ring failure-frequency f_ICharacteristic of correspondence W_BPFI, outer ring fault Frequency f_OCharacteristic of correspondence W_BPFOWith ball failure-frequency f_BCharacteristic of correspondence W_BSF；

Described step S2 specifically includes:

S2-1: based on minimal reconstruction error criterion, above-mentioned 12 dimension dimensionless groups are applied constraints, obtains 12 dimension decorrelations Reconstruct characteristic vector；

S2-2: be standardized processing to the reconstruct characteristic vector of described 12 dimension decorrelations, it is thus achieved that 12 dimension sample averages are 0 and sample This standard difference is the standardization reconstruct characteristic vector of 1；And

Described step S3 specifically includes:

S3-1: training, obtains the sample standard deviation of 2 norms of described standardization reconstruct characteristic vector under rolling bearing normal operating condition Value and sample standard deviation, and set the threshold value that Rolling Bearing Status is abnormal；

S3-2: test, by under the characteristic vector of sample under rolling bearing unknown state to described rolling bearing normal operating condition The Euclidean distance of sample average characteristic vector as deterioration index, compare deterioration index and described threshold value, when deterioration index is big When described threshold value, it is determined that rolling bearing is in malfunction；When deteriorating index less than or equal to described threshold value, it is determined that rolling Dynamic bearing is in normal operating condition.

Rolling bearing fault on-line checking the most according to claim 1 and state evaluating method, it is characterised in that: described bad Change index the biggest, then judge that the degradation of rolling bearing is the biggest.

Rolling bearing fault on-line checking the most according to claim 1 and state evaluating method, it is characterised in that: step The concrete operations extracting described 6 dimension Time-domain Statistics parameters in S1-2 are:

Form factor T_SIBe given by formula (1), peak index T_CIBe given by formula (2), pulse index T_MIBe given by formula (3), Margin index T_CLIBe given by formula (4), kurtosis T_KUBe given by formula (5), flexure T_SKBe given by formula (6),

$T_{SI}=\frac{\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\left(y_i^2\right)}}{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\left|y_i\right|}---\left(1\right)$

$T_{CI}=\frac{\underset{i=1}{\overset{10}{\Σ}}{y}_{pi}}{\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{\left(y_i\right)}^2}}---\left(2\right)$

$T_{MI}=\frac{\underset{i=1}{\overset{10}{\Σ}}{y}_{pi}}{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\left|y_i\right|}---\left(3\right)$

$T_{CLI}=\frac{\underset{i=1}{\overset{10}{\Σ}}{y}_{pi}}{{\[\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}\sqrt{\left|y_i\right|}\]}^2}---\left(4\right)$

$T_{KU}=\frac{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{y_i}^4}{{\left(\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{y_i}^2\right)}^2}---\left(5\right)$

$T_{SK}=\frac{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{y_i}^3}{{\left(\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{y_i}^2\right)}^{\frac{3}{2}}}---\left(6\right)$

Wherein, y_iShake described in being the data in acceleration signal；y_piBe described data are divided into 10 sections after every segment data absolute value Maximum.

Rolling bearing fault on-line checking the most according to claim 3 and state evaluating method, it is characterised in that: step The concrete operations extracting 3-dimensional Time-domain Statistics parameter in S1-3 are:

Gravity frequency F_FCBe given by formula (7), mean square frequency F_MSFBe given by formula (8), frequency variance F_VFBe given by formula (9),

$F_{FC}=\frac{\underset{i=0}{\overset{n}{\Σ}}{f}_{i}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\Σ}}S\left(f_i\right)}---\left(7\right)$

$F_{MSF}=\frac{\underset{i=0}{\overset{n}{\Σ}}{f}_i^{2}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\Σ}}S\left(f_i\right)}---\left(8\right)$

$F_{VF}=\frac{\underset{i=0}{\overset{n}{\Σ}}{(f_i-F_{FC})}^{2}S\left(f_i\right)}{\underset{i=0}{\overset{n}{\Σ}}S\left(f_i\right)}---\left(9\right)$

Wherein, S (f_i) it is vibration acceleration signal frequency spectrum function.

Rolling bearing fault on-line checking the most according to claim 3 and state evaluating method, it is characterised in that: described The concrete operations of step S1-4 include following three steps:

S1-4-1, calculates inner ring failure-frequency f according to formula (10)-(12)_I, outer ring failure-frequency f_OWith ball failure-frequency f_B,

$f_I=\frac{1}{2}Z(1+\frac{d}{D}cos\mathrm{\α})f_r---\left(10\right)$

$f_O=\frac{1}{2}Z(1-\frac{d}{D}cos\mathrm{\α})f_r---\left(11\right)$

$f_B=\frac{D}{2d}\[1-{\left(\frac{d}{D}\right)}^2{\cos }^2\mathrm{\α}\]f_r---\left(12\right)$

Wherein d represents rolling element diameter, and D represents bearing pitch diameter, and Z represents rolling element number, and α represents contact angle, f_rRepresent the axis of rolling That holds inner ring and outer ring relatively rotates frequency；

S1-4-2, uses db8 small echo that described vibration acceleration signal carries out wavelet decomposition and obtains 6 signals, including 5 details Signal d1, d2, d3, d4, d5 and 1 approximate signal a5；

S1-4-3, extract Based on Wavelet Envelope spectrum in 3-dimensional dimensionless group: be located in wavelet packet envelope spectrum, fault characteristic frequency and With the presence of feature spectral peak near its each rank frequency multiplication, if a width of f of envelope spectrum analysis band_e, envelope spectrum is W (f), if the number of W (f) spectral line Mesh is N_e, then S_eaFor

$S_{ea}=\frac{1}{N_e}\underset{i=0}{\overset{N_e}{\Σ}}W\left(f_i\right)---\left(13\right)$

Make S again_edFor the spectral line meansigma methods at the frequency multiplication of each rank of fault characteristic frequency in envelope spectrum, if the spectrum of failure-frequency in envelope spectrum Line number is n_e, then

$S_{ed}=\frac{1}{n_e}\underset{i=0}{\overset{n_e}{\Σ}}W\left({\mathrm{if}}_d\right)---\left(14\right)$

Construct a dimensionless group:

${\mathrm{\ΔS}}_e=\frac{S_{ed}}{S_{ea}}---\left(15\right)$

Described in step S1-4-2 6 signal is carried out envelope spectrum analysis respectively, obtains inner ring failure-frequency f_ICharacteristic of correspondence W_BPFI, outer ring failure-frequency f_OCharacteristic of correspondence W_BPFO, ball failure-frequency f_BCharacteristic of correspondence W_BSF。

Rolling bearing fault on-line checking the most according to claim 1 and state evaluating method, it is characterised in that: described The concrete operations of step S2-1 include following three steps:

S2-1-1: be standardized processing to 12 dimension dimensionless groups, obtain x_i, then have Σ_ix_i=0；

S2-1-2: assuming that the new coordinate system obtained after projective transformation is { w₁,w₂,…,w_d, wherein, d is intrinsic dimensionality；Apply Being constrained to wi is normal orthogonal base vector, meets | | w_i||₂=1, w_i ^Tw_j=0 (i ≠ j), optimization object function is

$\begin{array}{cc}\underset{W}{min}& -tr\left(W^T{\mathrm{XX}}^{T}W\right)\\ s.t.& W^{T}W=I\end{array}---\left(16\right)$

S2-1-3: solve formula (16) and obtain projection matrix W, be calculated reconstruct characteristic vector z of decorrelation according to formula (17)_i

z_i=(z_i1,z_i2,…,z_id), z_id=w_j ^Tx_i (17)。

Rolling bearing fault on-line checking the most according to claim 6 and state evaluating method, it is characterised in that: described Step S2-2 Playsization reconstruct characteristic vector z_i ^*Be given by formula (18)

$z_i^{*}=(z_{i1}^{*},z_{i2}^{*},...,z_{id}^{*}),z_{ij}^{*}=\frac{z_{ij}-{\widehat{\mathrm{\μ}}}_j}{{\widehat{\mathrm{\σ}}}_j}---\left(18\right)$

Wherein,

${\widehat{\mathrm{\μ}}}_j=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}{z}_{ij},{\widehat{\mathrm{\σ}}}_j^2=\frac{1}{n-1}\underset{i=1}{\overset{n}{\Σ}}{(z_{ij}-{\widehat{\mathrm{\μ}}}_j)}^2---\left(19\right).$

Rolling bearing fault on-line checking the most according to claim 1 and state evaluating method, it is characterised in that: step Described threshold value in S3-1 is given by formula (20)

$D_{max}={\widehat{\mathrm{\μ}}}_D+3{\widehat{\mathrm{\σ}}}_D---\left(20\right)$

Wherein,WithIt is D respectively_iSample average and sample standard deviation.