CN105204493A - State monitoring and fault diagnosing method for rotating mechanical equipment - Google Patents

Abstract

The invention discloses a state monitoring and fault diagnosing method for rotating mechanical equipment based on systematic cluster analysis and Fisher discriminant analysis. The method includes the steps that empirical mode decomposition is carried out on obtained monitoring data in a normal state and various fault states to extract energy feature vectors; based on systematic cluster analysis, the separability among samples is verified, and the influence on the dispersibility of sample data is eliminated; on the basis, discriminant analysis functions are extracted, a discrimination totality is built, and faults are diagnosed according to the real-time state monitoring data of the rotating mechanical equipment. Hot problems of online fault diagnosis for rotating mechanical equipment are solved, and intelligent fault diagnosis for the rotating mechanical equipment is achieved. A state monitoring and fault diagnosing model can be set up without the fault state data of a plenty of samples or full-life state monitoring data, dependency on the data is reduced, operability is improved, and the state monitoring and fault diagnosing method has high engineering applicability.

Description

A kind of rotating machinery condition monitoring and fault diagnosis method

Technical field

The invention belongs to the Intelligent Fault Diagnosis Technique field of rotating machinery, be specifically related to a kind of rotating machinery method for diagnosing faults combined with Fei Xier discriminatory analysis (FDA) based on hierarchial-cluster analysis (HCA).

Background technology

Current less desirable equipment failure shut down remain missile weapon system especially terrestrial weapon equip existing significant problem, serious task, safety effects will be caused thus.To be correlated with land equipment for missile weapon system, as a kind of effective mode, equipment maintenance can ensure that weaponry has lasting high reliability to a certain extent.But; passive correction maintenance is difficult to the generation avoiding fault; stop time long, MISSILE LAUNCHING task will be caused to postpone and even high consequence was lost in failure and maintenance cost and continuity; and preventative maintenance blindly easily causes maintenance superfluous, unnecessary maintenance down also directly has influence on the availability of equipment.Therefore; for reducing equipment failure stop time, reducing life cycle cost, improve equipment availability, reducing security risk, how to equipment carry out real-time online status monitoring and fast and accurately fault diagnosis become one of study hotspot of device intelligence maintenance.

HCA is a kind of non-supervisory formula method for classifying modes, and namely its analytic process is the multiple class or the overall set of physics or abstract object be grouped into as being made up of similar object.Each cluster generated is the set of some data object, and the object in these objects and same cluster has very high similarity, and has obvious difference with the object in other clusters.Based in the fault diagnosis of Condition Monitoring Data, in order to avoid the impact of obtained normal condition and all kinds of malfunction sample data dispersiveness, improve the accuracy of condition monitoring and fault diagnosis, the sample data separability analysis based on HCA will be absolutely necessary.

FDA, as a kind of spatial transform technique, can determine a series of linear transformation vector and according to maximization inter _ class relationship, the criterion simultaneously minimizing within-cluster variance in spatial mappings process all kinds of totally between be farthest separated.In spatial alternation process, choose optimum discriminant vector and make fisher criterion function maximum, thus the data space of higher-dimension along acquired Fisher characteristic direction projection, thus can achieve Data Dimensionality Reduction and different classes of data be farthest separated.

As the crucial pith of missile weapon system land equipment, the fault of rotating machinery will have a strong impact on the availability of weaponry with inefficacy and impact security, task and economy.At present, although the technology such as real-time state monitoring, intelligent trouble diagnosis obtains extensive concern, but seldom relate to rotating machinery, and existing condition monitoring and fault diagnosis method generally all needs using the complete life-cycle degenerate state Monitoring Data of equipment to be monitored and the fault state data of great amount of samples as input, seldom consider the abundant digging utilization to fragmentary data (namely only having the Monitoring Data under normal condition and malfunction).But, the life-cycle degraded data often more difficult acquisition of equipment, especially for the equipment that some are are newly installed and used, there is no the fault state data of life-cycle degraded data and great amount of samples at all, fault diagnosis accuracy is low, poor robustness, is therefore limited by very large in practical engineering application.

Summary of the invention

The technical problem to be solved in the present invention is to provide a kind of rotating machinery condition monitoring and fault diagnosis method combined based on hierarchial-cluster analysis and Fei Xier discriminatory analysis, to solve in missile weapon system land equipment is correlated with rotating machinery condition monitoring and fault diagnosis, existing method depends on life-cycle degraded data and the problem of accurate low, the poor robustness of fault diagnosis.

For solving the problems of the technologies described above, the present invention adopts following technical proposals:

Step one, extract the feature parameter vectors based on empirical mode decomposition (EMD): under normal and all kinds of malfunction, the vibration monitoring signal of rotating machinery has distribution in each frequency range usually.EMD method is based on the local feature time scale of vibration signal, and the signal function of complexity is decomposed into limited intrinsic mode function (IMF) sum, and each IMF contains the signal frequency composition from high frequency to low frequency successively.Carrying out statistics and analysis to decomposing the IMF obtained, forming the energy indexes of reflected signal feature.Under the normal operating conditions and all kinds of fail operation state of rotating machinery, gather vibration signal and carry out EMD decomposition thus extract the feature parameter vectors sample.

Step 2, sample separability are analyzed: the sample extracted under often kind of duty (comprising normal operating conditions and all kinds of fail operation state) is carried out hierarchial-cluster analysis (HCA), based on the calculating of the distance between sample or similarity coefficient and between class distance, the sample of various state is assembled for different class, thus judges the similarity of sample in the otherness of sample between all kinds of state and all kinds of state.

Step 3, extraction discriminatory analysis function: the feature parameter vectors sample extracted under often kind of duty (comprising normal operating conditions and all kinds of fail operation state) composition training set is carried out Fei Xier discriminatory analysis (FDA), realize the spatial alternation of higher dimensional space to lower dimensional space, and build differentiation overall (comprising normally overall and all kinds of fault population) and extract discriminatory analysis function.

Step 4, obtain Condition Monitoring Data: by Real-time Collection to the vibration signal of rotating machinery under t duty carry out EMD decomposition, choose front p the IMF component comprising major degenerative feature, extraction the feature parameter vectors x.

Step 5, fault diagnosis: in conjunction with discriminatory analysis function, calculate the feature parameter vectors x and the mahalanobis distance normally between overall and all kinds of fault population, can judge current operating state whether fault localizing faults pattern based on set decision rule.

Beneficial effect of the present invention is as follows:

Technical scheme of the present invention utilizes empirical mode decomposition, obtain intrinsic mode function component, not only there is significant gradual ripple bag feature on frequency domain, also in time domain, there is local characteristic, and between each intrinsic mode function component, there is good orthogonality, for condition monitoring and fault diagnosis provides high-quality input data; The present invention is using the non-supervisory formula method for classifying modes of hierarchial-cluster analysis as invention, separability between the test sample book obtained under fully can verifying normal condition and all kinds of malfunction, overcome the impact of test sample book data scatter, the accuracy of fault diagnosis can be significantly improved; The present invention make full use of the powerful dimension-reduction treatment ability of Fei Xier discriminatory analysis and mahalanobis distance intuitively meter to seek peace discriminant classification ability, can realize fault mode location while fault detect, method is simple, and workable, diagnosis effect is remarkable; The inventive method can set up fault diagnosis model based on normal condition and various fault state data, reduces the dependence to historical data, can realize the intelligent maintenance of rotating machinery, has very strong versatility and very high engineer applied value.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described in further detail;

Fig. 1 illustrates the schematic diagram of a kind of rotating machinery condition monitoring and fault diagnosis method of the present invention;

Fig. 2 illustrates the schematic diagram of empirical mode decomposition method of the present invention;

Fig. 3 illustrates the schematic diagram of Hierarchical Clustering process of the present invention;

Fig. 4 illustrates Xi Er discriminatory analysis logical process figure of the present invention;

Fig. 5 illustrates the fault diagnosis mapping graph based on hierarchial-cluster analysis and Fei Xier discriminatory analysis of the present invention;

Fig. 6 illustrates the process flow diagram of the fault diagnosis based on hierarchial-cluster analysis and Fei Xier discriminatory analysis of the present invention;

Fig. 7 illustrates the time domain beamformer of a certain group of inner ring fault vibration signal in the embodiment of the present invention;

Fig. 8 illustrates the EMD decomposition result figure of the present invention's a certain group of inner ring fault vibration signal;

Fig. 9 illustrates in the embodiment of the present invention and utilizes HCA to carry out the schematic diagram of separability checking to training set.

Embodiment

In order to be illustrated more clearly in the present invention, below in conjunction with preferred embodiments and drawings, the present invention is described further.Parts similar in accompanying drawing represent with identical Reference numeral.It will be appreciated by those skilled in the art that specifically described content is illustrative and nonrestrictive, should not limit the scope of the invention with this below.

A kind of important branch and the study hotspot that real-time state monitoring and fault diagnosis have become intelligent maintenance is carried out to rotating machinery.Existing certain methods is all carry out condition monitoring and fault diagnosis according to the fault state data of life-cycle degenerate state Monitoring Data and great amount of samples substantially, but these data are difficult to obtain, and diagnoses poor in timeliness, and diagnostic result is inaccurate.The present invention is directed to structure and the signal characteristic thereof of rotating machinery, propose a kind ofly to differentiate based on Hierarchical Clustering and Fei Xier the method for diagnosing faults combined.The core concept of the inventive method be overall by differentiation that running status that computing equipment is current is corresponding with normal condition and all kinds of malfunction between distance, can judge that whether normal current operating state is based on set decision rule, if having fault and energy localizing faults pattern simultaneously.The inventive method utilizes the pattern-recognition of malfunction, achieves the condition monitoring and fault diagnosis of rotating machinery.

As shown in Figure 1, what the present invention adopted is a kind of design feature for rotating machinery and failure mechanism, and adopt the method for diagnosing faults that hierarchial-cluster analysis and Fei Xier discriminatory analysis combine, concrete steps are as follows:

Step one, extract the feature parameter vectors based on empirical mode decomposition: to the vibration signal gathered under the normal operating conditions of rotating machinery and all kinds of fail operation state, carry out empirical mode decomposition, detailed process as shown in Figure 2, obtains the IMF component c of different frequency composition _i(t) i=1,2 ..., n.Choose front p the component comprising major degenerative feature, ask the gross energy E of each component _ck, k=1,2 ..., p:

$E_{ck}={\∫\left|c_k\left(t\right)\right|}^{2}dt=\underset{k=1}{\overset{n}{\Σ}}|c_{kj}{|}^2---\left(1\right)$

In formula, c _kj(k=1,2 ..., p; J=1,2 ..., l) represent the amplitude of a kth IMF component jth discrete point, l is signal length.

Due on rotating machinery during certain component failure, larger impact can be had on the energy of signal in each frequency.Therefore, be that element builds a normalized proper vector T with energy:

T＝[E _c1/E,E _c2/E,…,E _cp/E](2)

Wherein, E _c1, E _c2..., E _cpbe respectively the energy of front p IMF component, and

$E={\left(\underset{k=1}{\overset{p}{\Σ}}\right|E_{ck}{|}^2)}^{1/2}---\left(3\right)$

Proper vector T is and decomposes based on EMD the feature parameter vectors extracted.

Step 2, the feature parameter vectors sample separability are analyzed: the feature parameter vectors sample composition training set that will extract under often kind of duty (comprising normal operating conditions and all kinds of fail operation state), carry out hierarchial-cluster analysis, to verify the separability between Different categories of samples, eliminate the impact of data scatter between sample.

The basic thought of hierarchical clustering method is: first, assuming that each sample becomes a class separately, at this moment, the distance between all kinds of is exactly the distance between each sample; Then, according to the similarity between two between sample, the most close two samples are merged into a new class; Calculate the distance between new class and other class again, two nearest classes are merged, reduces a class, so at every turn till all samples all become a class.Finally, classification results is drawn according to determined class number.General cluster process as shown in Figure 3.

The cluster principle of hierarchical clustering method is decided by the definition of distance between sample or similarity coefficient and between class distance.With d _ijrepresent sample X _(i)and X _(j)between distance, with D _pqrepresentation class G _pand G _qbetween distance.The class method of average be a kind of utiliztion ratio comparatively extensively, the good method of Clustering Effect, the class method of average by the mean value square being defined as the square distance between any two of each sample in this two class of the distance between two classes, namely

${D_{pq}}^2=\frac{1}{n_p{n}_q}\underset{i\∈G_P,j\∈G_q}{\Σ}{d_{ij}}^2---\left(4\right)$

Recursion formula is:

${D_{rk}}^2=\frac{n_p}{n_r}{D_{pk}}^2+\frac{n_q}{n_r}{D_{pk}}^2\left(k\≠p,q\right)---\left(5\right)$

In formula, n _pclass G _pthe number of middle sample, n _qclass G _qthe number of middle sample, class G _pwith class G _qmerge into class G _r, G _rthe number of middle sample is n _r=n _p+ n _q.

Step 3, extract discriminatory analysis function: between checking Different categories of samples separability basis on, Different categories of samples is formed training set, carries out the study of FDA, extract discriminatory analysis function and to build each differentiation overall.Concrete grammar principle is as follows:

If known k differentiates overall (comprising normally overall and all kinds of fault population):

G ₁～(μ ₁,∑ ₁),G ₂～(μ ₂,∑ ₂),…,G _k～(μ _k,∑ _k)

Wherein, μ _iand ∑ _ig respectively _imean vector and covariance matrix, i=1,2 ..., k.If x is sample to be sentenced.Through to the linear comprehensive differentiating overall variable index, obtain corresponding one dimension sample y=a ' x to be sentenced.It may come from overall G ₁ ^*~ (a ' μ ₁, a ' ∑ ₁a), G ₂ ^*~ (a ' μ ₂, a ' ∑ ₂a) ..., G _k ^*~ (a ' μ _k, a ' ∑ _ka).Copy the thought of variance analysis, differentiate overall to distinguish each better, the selection of vectorial a (transposition of a ' expression a) should make " poor between group " to expand as far as possible, makes " group interpolation " little as far as possible, order:

$B_0=\underset{i=1}{\overset{k}{\Σ}}{\left(a^{\′}{\mathrm{\μ}}_i-a^{\′}\stackrel{\‾}{\mathrm{\μ}}\right)}^2=a^{\′}\[\underset{i=1}{\overset{k}{\Σ}}({\mathrm{\μ}}_i-\stackrel{\‾}{\mathrm{\μ}}){({\mathrm{\μ}}_i-\stackrel{\‾}{\mathrm{\μ}})}^{\′}\]a=a^{\′}Ba---\left(6\right)$

$E_0=\underset{i=1}{\overset{k}{\Σ}}{a}^{\′}{\mathrm{\Σ}}_{i}a=a^{\′}\left({\mathrm{\Σ}}_1+{\mathrm{\Σ}}_2+...+{\mathrm{\Σ}}_k\right)a=a^{\′}\mathrm{\Σ}a---\left(7\right)$

Wherein, $\stackrel{\‾}{\mathrm{\μ}}=\frac{1}{k}({\mathrm{\μ}}_1+{\mathrm{\μ}}_2+...+{\mathrm{\μ}}_k),B=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}({\mathrm{\μ}}_i-\stackrel{\‾}{\mathrm{\μ}}){({\mathrm{\μ}}_i-\stackrel{\‾}{\mathrm{\μ}})}^{\′},\mathrm{\Σ}=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\Σ}}_i.$ B ₀be equivalent to sum of squares between groups, E ₀be equivalent to organize interior quadratic sum.Therefore, the selection of vectorial a should make:

$\underset{a}{max}(\frac{B_0}{E_0}=\frac{a^{\′}Ba}{a^{\′}\mathrm{\Σ}a})---\left(8\right)$

Reach maximum.This is equivalent to preferentially problem

$\left\{\begin{array}{c}{a}^{\′}Ba\→max\\ a^{\′}\mathrm{\Σ}a=1\end{array}\right.$

Thus, when a is chosen as and Σ ⁰¹the eigenvalue of maximum λ of B ₁characteristic of correspondence vector a ₁time, formula (8) reaches maximal value λ ₁, therefore y ₁=a ₁' x.

If x is comprehensively become a canonical variable y ₁=a ₁' x can't distinguish each totally well, can also by Σ ⁰¹second eigenvalue λ of B ₂corresponding proper vector a ₂set up second canonical variable y ₂=a ₂' x; If not enough, can λ be used ₃set up y ₃=a ₃' x, the rest may be inferred.Generally establish Σ ⁰¹front m the eigenvalue of maximum of B is followed successively by λ ₁>=λ ₂>=...>=λ _m, characteristic of correspondence vector is respectively a ₁, a ₂..., a _m, and contribution rate of accumulative total (p is the dimension of sample to be sentenced, pem) reaches certain threshold value (being set as 85%), then can obtain m mutual incoherent canonical variable y ₁, y ₂... y _mfor discriminatory analysis.This is equivalent to that the p dimension indicator of sample x to be sentenced is compressed into m dimension indicator and differentiates.

Thus, y _i=a _i' x, i=1,2 ..., m is exactly the discriminatory analysis function that will extract, and detailed process as shown in Figure 4.

Step 4, status monitoring: by Real-time Collection to the vibration signal of rotating machinery under t duty carry out EMD decomposition, extract the feature parameter vectors, and be set as x.

Step 5, fault diagnosis is carried out to rotating machinery: using x as sample to be sentenced, in conjunction with discriminatory analysis function y _i=a _i' x, (i=1,2 ..., m), after dimension-reduction treatment, in new lower dimensional space, sample x to be sentenced is to the overall G of differentiation _j(j=1,2 ..., mahalanobis distance MD=d (x, G k) _j) can by y=(y ₁, y ₂..., y _m) ' to G _j ^*(j=1,2 ..., distance k) calculates:

$d^2(x,G_j)={(x-{\mathrm{\μ}}_j)}^{\′}{\mathrm{\Σ}}^{-1}(x-{\mathrm{\μ}}_j)=\underset{i=1}{\overset{m}{\Σ}}{(y_i-a_i{\mathrm{\μ}}_j)}^2---\left(9\right)$

According to sample x to be sentenced and the mahalanobis distance normally between overall and all kinds of fault population, in conjunction with decision rule, determine that whether current operating state is normal, if there is fault to occur, then localizing faults pattern while.Decision rule is as follows:

If $d^2(x,G_l)=\underset{1\≤j\≤k}{\min }{d}^2(x,G_j),$ So x ∈ G _l

In formula, G _jfor normal overall and all kinds of fault population.That is, the overall corresponding state that mahalanobis distance is minimum and between sample x to be sentenced is exactly the duty of t.

As illustrated in Figures 5 and 6, for based on the fault diagnosis mapping relations of HCA and FDA and flow process.By Fei Xier discriminatory analysis, normal condition data sample, all kinds of fault state data sample and sample to be sentenced are mapped to new lower dimensional space by original higher dimensional space.New lower dimensional space calculate sample mapping point to be sentenced and respectively judges overall between mahalanobis distance MD, by comparing MD, if sample mapping point to be sentenced and normally totally between mahalanobis distance minimum, then current device belongs to normal operating conditions; If the mahalanobis distance between sample mapping point to be sentenced and a certain fault population is minimum, then current device belongs to malfunction, and locates the fault mode that its fault mode characterizes for this fault population.

Below by one group of embodiment, the present invention will be further described:

This example adopts the experimental data of bearing data center of Washington Catholic University of America to verify.This testing table by the motor of a 2HP (horsepower), a torque converter/scrambler, a dynamometer machine and control circuit.Be contained in the deep-groove ball test bearing that model on motor is 6205-2RSJEMSKF, inject inner ring, outer shroud and rolling body Single Point of Faliure respectively by electrosparking.Under normal, inner ring fault, outer shroud fault and rolling body fault four kinds of states, gather vibration data by the acceleration transducer be connected on magnetic base shell, setting shaft rotating speed is 1750RPM, and sample frequency is 12000HZ.

Step one, based on EMD decompose extract energy feature.

Be in the duty of normal and inner ring, outer shroud, rolling body fault at bearing under, gather 8 groups of vibration signals respectively, carry out EMD decomposition to each group of data, wherein the original signal of one group of inner ring fault as shown in Figure 7, and its EMD decomposition result as shown in Figure 8.Decompose obtain 8 IMF components (imf1, imf2 ..., imf8) and discrepance R.As seen from the figure, front 6 IMF components contain main fault signature.Therefore, choose front 6 IMF components and calculate its energy according to formula (1)-(3), building the feature parameter vectors.Wherein, the study that front 4 composition of sample training sets under often kind of state are used for FDA (for the ease of identification, corresponds respectively to normal condition, inner ring, outer shroud, rolling body malfunction, they is numbered N_1 to N_4, I_1 to I_4, O_1 to O_4, B_1 to B_4), remaining 4 form test set (equally, they are numbered N_T1 to N_T4, I_T1 to I_T4, O_T1 to O_T4, B_T1 to B_T4).Training set and test set the feature parameter vectors are as shown in table 1,2.

Table 1 bearing training sample proper vector

Num	1	2	3	4	5	6
							N_1	0.8111	0.5012	0.0524	0.2964	0.0002	0.0018
N_2	0.8259	0.4665	0.0544	0.3112	0.0002	0.0019
							N_3	0.7891	0.5274	0.0524	0.3097	0.0002	0.0018

N_4	0.8137	0.493	0.0553	0.3023	0.0002	0.0018
							I_1	0.0795	0.2297	0.5852	0.1497	0.0014	0.0081
I_2	0.0823	0.2226	0.5831	0.1464	0.0015	0.0082
							I_3	0.081	0.2341	0.6031	0.1474	0.0014	0.0084
I_4	0.0723	0.2279	0.5922	0.1505	0.0014	0.0092
							O_1	0.0069	0.0096	0.4907	0.0156	0.0082	0.0137
O_2	0.0065	0.0096	0.4945	0.0161	0.0081	0.0137
							O_3	0.0069	0.0098	0.4831	0.0145	0.0076	0.0121
O_4	0.0065	0.01	0.5217	0.016	0.007	0.0124
							B_1	0.045	0.0449	0.4876	0.0184	0.0005	0.0027
B_2	0.0443	0.0424	0.457	0.0189	0.0005	0.0025
							B_3	0.0426	0.0436	0.464	0.0187	0.0005	0.0025
B_4	0.0412	0.042	0.4584	0.0179	0.0004	0.0022

Table 2 bearing test sample book proper vector

Num	1	2	3	4	5	6
							N_T1	0.804	0.5112	0.0511	0.2987	0.0002	0.0016
N_T2	0.8119	0.5007	0.0528	0.2948	0.0002	0.0018
							N_T3	0.8208	0.4824	0.0538	0.3006	0.0002	0.0017
N_T4	0.7854	0.5344	0.0532	0.307	0.0002	0.0018
							I_T1	0.0734	0.2138	0.58	0.1423	0.0015	0.008
I_T2	0.087	0.2276	0.5929	0.15	0.0014	0.0089
							I_T3	0.0769	0.2242	0.6017	0.1463	0.0016	0.0084
I_T4	0.071	0.2271	0.5888	0.1535	0.0014	0.0082
							O_T1	0.0066	0.0091	0.4939	0.0159	0.0073	0.0132
O_T2	0.0073	0.0098	0.4945	0.0166	0.0075	0.0128
							O_T3	0.0068	0.0108	0.5057	0.0159	0.0073	0.0137
O_T4	0.0069	0.0098	0.4976	0.0154	0.0077	0.0133
							B_T1	0.0438	0.0431	0.483	0.0187	0.0005	0.0024
B_T2	0.0438	0.0439	0.4753	0.018	0.0006	0.0026
							B_T3	0.046	0.0411	0.471	0.0197	0.0005	0.0023
B_T4	0.043	0.0436	0.4627	0.0187	0.0005	0.0023

Step 2, training sample separability are analyzed.

Application HCA carries out separability checking to training set, and as shown in Figure 9, the dendrogram of gained clearly show that whole cluster process.First, assuming that each sample becomes a class separately; After first time, class merged, it is a class that sample B _ 1 to B_4 and O_1 to O_4 gathers, and sample I_1 to I_4 gathers for another kind of, and it is the 3rd class that sample N_1 to N_4 gathers.As shown in table 3, can find out that from clustering relationships result all normal samples gather for overall 1, all inner ring fault samples gather for overall 2, and all outer shroud samples gather for overall 3, and all rolling body fault samples gather for overall 4.The separability of the proper vector sample extracted under the SPSS specialty Output rusults dendrogram of statistical software and clustering relationships result sufficient proof four kinds of states.

Table 3 training sample clustering relationships

Sample	Cluster index
		1:N_1	1
2:N_2	1
		3:N_3	1
4:N_4	1
		5:I_1	2
6:I_2	2
		7:I_3	2
8:I_4	2
		9:O_1	3
10:O_2	3
		11:O_3	3
12:O_4	3
		13:B_1	4
14:B_2	4
		15:B_3	4
16:B_4	4

Step 3, extraction discriminatory analysis function, structure differentiate overall.

On the basis demonstrating sample separability, utilize training set (N_1 to N_4, I_1 to I_4, O_1 to O_4, B_1 to B_4) extract discriminant function, build each differentiation overall, i.e. normal overall, inner ring fault population, outer shroud fault population and rolling body fault population, and be labeled as G_N respectively, G_I, G_O and G_B.

Step 4, bearing failure diagnosis.

Calculate the mahalanobis distance between 16 energy feature samples (N_T1 to N_T4, I_T1 to I_T4, O_T1 to O_T4, B_T1 to B_T4) in test set and normal overall and inner ring, outer shroud, rolling body fault population, result of calculation is as shown in table 4.Known by comparative analysis, test sample book N_T1 to N_T4 is minimum with the mahalanobis distance of overall G_N, and equally, test sample book I_T1 to I_T4, O_T1 to O_T4, B_T1 to B_T4 are minimum with the mahalanobis distance of overall G_I, G_O and G_B respectively.Thus can judge, sample N_T1 to N_T4 belongs to normal condition, sample I_T1 to I_T4, O_T1 to O_T4, B_T1 to B_T4 belongs to inner ring fault, outer shroud fault and rolling body fault respectively, namely determine their fault mode, while status monitoring, fault detect, reach the object of localization of fault.

Table 4 bearing failure diagnosis result table

By the detailed description of above method for diagnosing faults and result, visible Method for Bearing Fault Diagnosis of the present invention, workable, diagnosis effect is good, has significant engineer applied and is worth.

In sum:

(1) the present invention have studied the integrated approach of a kind of rotating machinery status monitoring, fault detect and localization of fault, can realize the intelligent maintenance of rotating machinery;

(2) the IMF component of EMD decomposition gained not only has significant gradual ripple bag feature on frequency domain, also in time domain, there is local characteristic, and have good orthogonality between each IMF component, these characteristics are that condition monitoring and fault diagnosis provides high-quality input data;

(3) HCA is as the non-supervisory formula method for classifying modes of one, separability between the test sample book obtained under fully can verifying normal condition and all kinds of malfunction, overcome the impact of test sample book data scatter, the accuracy of fault diagnosis can be significantly improved;

(4) make full use of the powerful dimension-reduction treatment ability of Fei Xier discriminatory analysis and mahalanobis distance intuitively meter to seek peace discriminant classification ability, can realize fault mode location while fault detect, method is simple, and workable, diagnosis effect is remarkable;

(5) the inventive method can set up fault diagnosis model based on normal condition and various fault state data, reduces the dependence to historical data, has very strong versatility and very high engineer applied value.

Obviously; the above embodiment of the present invention is only for example of the present invention is clearly described; and be not the restriction to embodiments of the present invention; for those of ordinary skill in the field; can also make other changes in different forms on the basis of the above description; here cannot give exhaustive to all embodiments, every belong to technical scheme of the present invention the apparent change of extending out or variation be still in the row of protection scope of the present invention.

Claims (5)

1. a rotating machinery condition monitoring and fault diagnosis method, is characterized in that, the step of the method comprises

S1, under the normal operating conditions and various faults duty of rotating machinery, gather vibration signal, and carry out empirical mode decomposition, extract the feature parameter vectors sample;

S2, separability analysis is carried out to the feature parameter vectors sample, judge the similarity of sample in the otherness of sample between every kind state and every kind state;

S3, the sample composition training set that will extract under normal operating conditions and various faults duty, carry out Fei Xier discriminatory analysis, and the normal differentiation of structure is overall and various faults judges totally, to extract discriminatory analysis function respectively;

S4, the vibration signal of Real-time Collection rotating machinery under t duty, and empirical mode decomposition is carried out to it, choose front p the intrinsic mode function component comprising major degenerative feature, extract the feature parameter vectors x;

S5, utilize described discriminatory analysis function, calculate the feature parameter vectors x and normal mahalanobis distance totally and between often kind of fault population, based on set decision rule, to current operating state, whether fault judges, and localizing faults pattern.

2. according to method according to claim 1, it is characterized in that, described various faults duty comprises inner ring fault, outer shroud fault or rolling body fault.

3. according to method according to claim 1, it is characterized in that, described step S1 comprises

S11, utilize empirical mode decomposition method, based on the local feature time scale of vibration signal, the signal function of complexity is decomposed into limited intrinsic mode function sum, and wherein, each intrinsic mode function contains the signal frequency composition from high frequency to low frequency successively;

S12, statistics and analysis is carried out to the intrinsic mode function obtained, form the energy indexes of reflected signal feature.

4. method according to claim 1, is characterized in that, described step S2 comprises

S21, the sample extracted under normal operating conditions and various faults duty is carried out hierarchial-cluster analysis;

S22, calculating based on the distance between sample or similarity coefficient and between class distance, assemble the sample of various state for different classifications.

5. method according to claim 1, is characterized in that, described step S5 comprises

S51, by Fei Xier discriminatory analysis, normal condition data sample, all kinds of fault state data sample and sample to be sentenced are mapped to new lower dimensional space by original higher dimensional space;

S52, calculate the mapping point of sample to be sentenced and the mahalanobis distance between respectively judging totally at new lower dimensional space;

If S53 sample to be sentenced mapping point and normal overall between mahalanobis distance minimum, then current device belongs to normal operating conditions; If the mahalanobis distance between sample mapping point to be sentenced and a certain fault population is minimum, then current device belongs to malfunction, and locates the fault mode that its fault mode characterizes for this fault population.