CN104833534A - Train running fault diagnosis device based on multi-source information fusion, and method - Google Patents

Abstract

The invention discloses a train running fault diagnosis device based on multi-source information fusion, and a method. The train running fault diagnosis device based on multi-source information fusion comprises a trackside fault early warning apparatus and a vehicle on-board fault diagnosis apparatus. The trackside fault early warning apparatus comprises a trackside vibration sensor disposed at one side of the track, and a temperature sensor. The vehicle on-board fault diagnosis apparatus comprises an axle box vibration sensor and a framework vibration sensor. The diagnosis comprises the steps of firstly adopting the singular value decomposition and de-noising technology to conduct noise-reducing process on various signals; secondly adopting HHT to extract fault characteristic constant from various vibration signals; thirdly adopting SVM and a threshold determining method to conduct train running local fault diagnosis; finally adopting improved D-S evidence theory to conduct decision fusion on the local fault diagnosis result, and realizing accurate diagnosis of running fault. The train running fault diagnosis device based on multi-source information fusion, and a method are advantageous in that the cost is low, precision is high, and device is simple; the problem of a single sensor that diagnosis may be not accurate can be resolved; the accurate diagnosis of the train running fault can be completed.

Description

A kind of train traveling system's trouble-shooter based on Multi-source Information Fusion and method

Technical field

The present invention relates to city railway train monitoring in transit and safe early warning key technology area, particularly a kind of train traveling system's method for diagnosing faults and device.

Background technology

Along with Chinese society's rapid development of economy, by 2016, the process of Chinese Urbanization gradually deeply, is estimated that city gauge lines total length will reach 3500 kilometers, is had and exceed putting into effect in the year two thousand twenty more than 30,000 city rail vehicles of 60 cities.The fast development of city rail traffic and coming into operation of a large amount of vehicle, make the operation security of vehicle become the focus paid close attention in recent years.

City rail vehicle traveling system is due to working environment complexity, when load impacting is excessive or be subject to the factors such as mounting design is improper, lubricating status is bad affect time, the various fault of easy appearance, it is one of the most flimsy system in city rail vehicle, carrying out Real-Time Monitoring to it is the important means ensureing city rail vehicle traffic safety, also be to provide the key of higher quality of passenger service simultaneously, therefore carry out Real-Time Monitoring and fault pre-alarming and diagnosis for the state of city rail vehicle traveling system and be of great immediate significance.

At present, it is main that China still rests on hand dipping mode to train traveling system fault diagnosis, and hand dipping not only labour intensity is large, and process is loaded down with trivial details, and due to the backwardness of survey instrument, is difficult to avoid artificial measuring error.If apply online monitoring system in maintenance link, form an automatic diagnostic device, testing staff not only can be made to free from the work of high strength and complicated data analysis, and reliable technology can be provided to ensure for maintenance.For the research of city rail vehicle traveling system automatic diagnostic device, mainly concentrate on the time frequency analysis to vibration signal at present, as wavelet analysis method, Short Time Fourier Transform, empirical mode decomposition, amplitude spectrum feature etc., these methods are all for single-sensor, single failure pattern is analyzed, and city rail vehicle traveling system fault effects many factors, and it is interrelated, the diversity of fault, complicacy between uncertain and various fault constitutes the difficult point of fault diagnosis, therefore only city rail vehicle traveling system fault Precise Diagnosis has been difficult to by the method for single-sensor and single failure characteristic quantity.

Summary of the invention

The object of the present invention is to provide that a kind of cost is low, the train traveling system trouble-shooter of Multi-source Information Fusion, trackside vibration signal, axle temperature signal, traveling system axle box vibration signal and framework vibration signal are carried out multi-source fusion, real time on-line monitoring traveling system quality condition.

Realizing technical solution of the present invention is: a kind of train traveling system trouble-shooter based on Multi-source Information Fusion, comprises trackside fault pre-alarming device and vehicle-mounted fault diagnosis device;

Described trackside fault pre-alarming device comprises the trackside vibration transducer and temperature sensor that are arranged in track side, for gathering trackside vibration signal and the axle temperature signal of traveling system;

Described vehicle-mounted fault diagnosis device comprises axle box vibration transducer and framework vibration transducer, for gathering axle box vibration signal and the framework vibration signal of traveling system.

As a kind of specific embodiment, described trackside vibration transducer, axle box vibration transducer and framework vibration transducer are piezoelectric acceleration transducer; Described temperature sensor is online infrared temperature-test sensor.

Another object of the present invention is to provide a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion.

Realizing technical solution of the present invention is: a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion, comprises the following steps:

Step 1: svd de-noising is carried out to vibration signal x (t) of traveling system, obtains signal y (t) after de-noising;

Step 2: carry out population mean empirical mode decomposition to signal y (t) after de-noising, obtains intrinsic mode function IMF _n;

Step 3: adopt Type B degree of association method to described intrinsic mode function IMF _ncarry out chaff component identification, extract true component, utilize the energy of true component as fault feature vector T, and described fault feature vector T is normalized obtains T ';

Step 4: the fault feature vector T ' after described normalization is carried out local fault diagnosis as the training set of support vector machine, exports the local fault diagnosis result of vibration signal;

Step 5: carry out traveling system local fault diagnosis based on axle temperature signal, exports the local fault diagnosis result of axle temperature signal;

Step 6: adopt the D-S evidence theory improved to carry out Decision fusion to the local fault diagnosis result of described vibration signal and the local fault diagnosis result of axle temperature signal, obtain the final fault diagnosis result of traveling system.

Further, carry out svd de-noising to vibration signal x (t) of traveling system in step 1, concrete steps are as follows:

(1.1) vibration signal x (t) is expressed as x=[x ₁, x ₂, x ₃..., x _n], structure attractor track matrix D _m:

$D_m={\left(\begin{array}{cccc}{x}_1& x_2& ...& x_n\\ x_{l\×\mathrm{\τ}+1}& x_{l\×\mathrm{\τ}+2}& ...& x_{l\×\mathrm{\τ}+n}\\ ...& ...& ...& ...\\ x_{(m-1)\×\mathrm{\τ}+1}& x_{(m-1)\×\mathrm{\τ}+2}& ...& x_{(m-1)\×\mathrm{\τ}+n}\end{array}\right)}_{m\×n}$

In formula: τ is time delay, m is Embedded dimensions, and N is the component number of vibration signal, n=N-(m-1) × τ;

(1.2) complex autocorrelation method select time is adopted to postpone τ, time series { x _ncomplex autocorrelation function be:

$R_{\mathrm{xx}}^m\left(\mathrm{\τ}\right)=\underset{j=1}{\overset{m-1}{\mathrm{\Σ}}}{R}_{\mathrm{xx}}\left(\mathrm{j\τ}\right)$

In formula, m is Embedded dimensions, and j is the multiple of delay time, gets first zero crossing be time delay τ;

(1.3) to matrix D _mcarry out svd:

D _m＝USV'

Wherein, U ∈ R ^{m × n}, V' ∈ R ^{n × n}, and UU'=I, VV'=I, S=diag (σ ₁, σ ₂..., σ _r), m is Embedded dimensions, and n is the n-th vibration signal component, (σ ₁, σ ₂..., σ _r) be matrix D _mdiagonal matrix after svd, wherein σ _rfor singular value component, R ^{m × n}for m × n ties up real number matrix;

(1.4) singular value corresponding for noise signal is set to 0, constructs new eigenmatrix, utilize this matrix carry out anti-singular value calculate de-noising after signal y (t).

Further, in step 2, population mean set of modes Empirical Mode Decomposition is carried out to signal y (t) after de-noising, obtain intrinsic mode function IMF _n, concrete steps are as follows:

(2.1) initialization population mean number of times M and the noise amplitude added, first time decomposes number of times p=1;

(2.2) to add the signal after making an uproar for the m time be y _p(t), y _pt () equals y (t) and the noise n added for the p time _p(t) sum:

y _p(t)＝y(t)+n _p(t)

(2.3) utilize EMD to decompose and add the signal y after making an uproar _pt (), obtains one group of IMF _q,p, IMF _q,pbe decompose q the IMF obtained the p time;

(2.4) IMF decomposed for P time is calculated _q,ppopulation mean obtain final IMF _q:

${\mathrm{IMF}}_q=\frac{1}{P}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}{\mathrm{IMF}}_{q,p},p=\mathrm{1,2},...,P$

(2.5) to y _p(t) and IMF _qdifference repeat step (2.3) ~ (2.4).

Further, adopt Type B degree of association method to described intrinsic mode function IMF in step 3 _ncarry out chaff component identification, extract true component, utilize the energy of true component as fault feature vector T, and be normalized described fault feature vector T and obtain T ', concrete steps are as follows:

(3.1) to described intrinsic mode function IMF _ncarry out sliding-model control with signal y (t) after de-noising, obtain IMF _n(k), y (k), k=1,2 ..., l.Order:

$d_0=\underset{k=1}{\overset{l}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)|$

$d_1=\underset{k=1}{\overset{l-1}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n(k+1)-y(k+1)-{\mathrm{IMF}}_n\left(k\right)+y\left(k\right)|$

$d_2=\frac{1}{2}\underset{k=2}{\overset{l-1}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n(k+1)-y(k+1)-2[{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)]+[{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)\left]\right|$

In formula, l is sliding-model control number, d ₀for physical features displacement difference, d ₁for physical features velocity contrast, d ₂for physical features acceleration is poor;

Then Type B calculation of relationship degree formula is:

$\mathrm{\ψ}({\mathrm{IMF}}_n,y\left(t\right))=\frac{1}{1+\frac{1}{l}{d}_0+\frac{1}{l-1}{d}_1+\frac{1}{l-2}{d}_2}$

The normalization Type B degree of association, according to the order of magnitude of numerical value, rejects chaff component, extracts true component;

(3.2) to each true component IMF extracted _αask its ENERGY E _α:

$E_{\mathrm{\α}}={\∫}_{-\∞}^{+\∞}{\left|c_{\mathrm{\α}}\left(t\right)\right|}^2\mathrm{dt},\mathrm{\α}=\mathrm{1,2},...,H$

In formula, Η is the true component IMF extracted _αnumber, c _αt () is true component IMF _αamplitude;

(3.3) with ENERGY E _αfor element, to construct a fault feature vector T as follows:

T＝[E ₁,E ₂,…E _Η]

(3.4) proper vector T is normalized:

T′＝[E ₁/E,E ₂/E,…E _Η/E]

In formula, t ' is the fault feature vector after normalization.

Further, in step 4, the fault feature vector T ' after normalization is carried out fault diagnosis as the training set of support vector machine, export the local fault diagnosis result of vibration signal, concrete steps are as follows:

(4.1) training dataset is set up;

According to the fault feature vector T ' after described normalization, set up training dataset (ξ _s, ζ _s), ξ _sfor the value of input variable, ζ _sfor corresponding output variable value, s is training set number.

(4.2) definite kernel function

Adopt gaussian radial basis function kernel function: K (ξ _s, ζ _s)=exp (-| ξ _s-ζ _s| ²/ σ ²), in formula, K (ξ _s, ζ _s) be kernel function, σ is the variance of Gaussian function.

(4.3) a certain amount of training dataset (ξ is selected _s, ζ _s) support vector machine is trained, set up the fault diagnosis model of support vector machine.

(4.4) treat diagnostic sample by the support vector machine trained and carry out diagnosis output, export the local fault diagnosis result of vibration signal.

As a kind of specific embodiment, in step 5, carry out traveling system local fault diagnosis based on axle temperature signal, specific as follows:

Whether normal according to the threshold decision axle temperature of the temperature difference of axle box and environment temperature;

If the temperature difference of axle box and environment temperature (20 DEG C, 40 DEG C] in, then for running thermal level, judge that axle temperature is normal further;

If the temperature difference of axle box and environment temperature (40 DEG C, 70 DEG C] in, be then low-grade fever level, judge axle temperature fault further;

If the temperature difference of axle box and environment temperature (70 DEG C, 100 DEG C] in, be then heat-flash level, judge axle temperature fault further;

If the temperature difference of axle box and environment temperature is more than 100 DEG C, then for swashing thermal level, judge axle temperature fault further.

Further, adopt the local fault diagnosis result of the D-S evidence theory of improvement to described vibration signal and axle temperature signal to carry out Decision fusion in step 6, obtain the final fault diagnosis result of traveling system, concrete steps are as follows:

(6.1) the identification framework Θ={ A of decision system is constructed ₁, A ₂..., A _χ, χ is fault mode number, A _χfor fault mode;

(6.2) structure is according to the evidence source V of identification framework _g, g=1,2 ..., G, G are evidence source number;

(6.3) basic reliability distribution function is constructed;

Basic the overall of trust degree partition function is made up of two parts: Part I is the Output rusults sum of support vector machine and threshold decision, and Part II is uncertainty estimation value;

$U_r=\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{\mathrm{\ζ}}_w+E_c$

In formula, U _rfor totally, ζ _wfor the diagnostic result of support vector machine and threshold decision, χ is fault mode number, E _cfor uncertain estimated value;

Described uncertain estimated value E _csolve according to following formula:

E _c＝E ₁+E ₂

In formula, E ₁for the root-mean-square error of diagnostic result, E ₂for other uncertainty estimations.

E ₁solve according to following formula:

$E_1=\sqrt{\frac{1}{\mathrm{\χ}}\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{({\mathrm{\ζ}}_w-C_w)}^2}$

In formula, C _wfor the desired output vector that support vector machine and threshold decision export.

E ₂solve according to following formula:

$E_2=\mathrm{\κ}\sqrt{\frac{1}{\mathrm{\χ}}\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{({\mathrm{\ζ}}_w-C_w)}^2}$

In formula, κ is uncertainty coefficient, often gets 0.1;

Then basic reliability distribution function is:

$m\left({\mathrm{\ζ}}_w\right)=\frac{{\mathrm{\ζ}}_w}{U_r}$

$m\left(E_c\right)=\frac{E_c}{U_r}$

(6.4) the basic reliability distribution m (ζ under utilizing D-S composition rule to calculate the synergy of each evidence body between two _w), its concrete steps are:

For V ₁and V ₂two evidence sources, m ₁and m ₂the basic reliability distribution function corresponding with two evidence sources, A _wand B _pbe respectively corresponding fault mode, then V ₁and V ₂synthesize according to following formula:

$m^{\′}\left(A\right)=\left\{\begin{array}{c}\frac{\underset{A_o\∩B_p=A}{\mathrm{\Σ}}{m}_1\left(A_w\right)m_2{B}_p}{1-K},A\≠\mathrm{\φ}\\ 0,A=\mathrm{\φ}\end{array}\right.$

$k=\underset{A_o\∩B_p=\mathrm{\φ}}{\mathrm{\Σ}}{m}_1\left(A_w\right)m_2\left(B_p\right)$

In formula, w=1,2 ..., χ; P=1,2 ..., χ.

Repeat step 6.4, by V ₁and V ₂fusion results and other evidence sources are merged successively, obtain final basic reliability distribution function M _r, thus obtain last diagnostic result.

As a kind of specific embodiment, described vibration signal x (t) comprises trackside vibration signal, axle box vibration signal and framework vibration signal.

The present invention compared with prior art, its remarkable advantage is: (1) is based on the train traveling system method for diagnosing faults of Multi-source Information Fusion, avoid the diagnosis uncertain problem that single-sensor may exist, complete train traveling system fault Precise Diagnosis.(2) cost is low, avoids the high cost using personal monitoring to bring; (3) engineering construction is good, and vibration transducer and shaft temperature sensor can bear on-the-spot rugged surroundings, and easy for installation; (4) on-line real time monitoring, can Timeliness coverage burst with the traveling system state change of long-term accumulated, thus provide and safeguard early warning timely.

Accompanying drawing explanation

Fig. 1 is the train traveling system method for diagnosing faults process flow diagram of Multi-source Information Fusion of the present invention.

Fig. 2 is the D-S evidence theory fusion diagnosis process flow diagram that the present invention improves.

Fig. 3 is trackside vibration transducer of the present invention and temperature sensor scheme of installation.

Fig. 4 is the vehicle-mounted axle box vibration transducer of the present invention and framework vibration transducer scheme of installation.

Fig. 5 a is axle box vibration signal de-noising front signal time-domain diagram in the embodiment of the present invention 1.

Fig. 5 b is signal time-domain diagram after the de-noising of axle box vibration signal in the embodiment of the present invention 1.

Fig. 6 a ~ Fig. 6 g is respectively seven intrinsic mode function time-domain diagrams that the signal after removing eliminated noise in the embodiment of the present invention 1 obtains after EEMD decomposes.

Fig. 6 h is for removing the residual components of the signal after eliminated noise after EEMD decomposes in the embodiment of the present invention 1.

Embodiment

Below in conjunction with drawings and the specific embodiments, the present invention is described in further detail.

Composition graphs 1 ~ 2, the train traveling system method for diagnosing faults of Multi-source Information Fusion of the present invention, first the collection of traveling system trackside vibration signal and axle temperature signal and vehicle-mounted axle box vibration signal and framework vibration signal is completed, secondly denoising is carried out to the signal collected, Hilbert-Huang transform is again adopted (to be called for short: HHT) process signal, and then try to achieve local fault proper vector, then local fault diagnosis is completed by support vector machine, the D-S evidence theory of improvement is finally adopted to merge local fault diagnosis, realize train traveling system Precise Diagnosis.

Composition graphs 2, the present invention adopts the D-S evidence theory of improvement to merge local fault diagnosis, comprises to set up identification framework, conclusion evidence source, construct basic trust degree partition function, combine the steps such as different evidence source and fusion diagnosis.D-S evidence theory does not provide the common form of basic reliability distribution function, and it is formed often according to subjective experience, needs make concrete analyses of concrete problems; The present invention builds basic reliability distribution function according to support vector machine local fault diagnosis result, local fault diagnosis result and D-S evidence theory is organically combined.

Composition graphs 3 ~ 4, the train traveling system trouble-shooter of Multi-source Information Fusion of the present invention comprises trackside fault pre-alarming device and vehicle-mounted fault diagnosis device, and trackside fault pre-alarming device comprises 8 trackside vibration transducer L ₁-L ₄and R ₁-R ₄with 2 temperature sensor T ₁, be all arranged in track side, for gathering vibration signal and the temperature signal of traveling system; Vehicle-mounted fault diagnosis device comprises axle box vibration transducer 1 and framework vibration transducer 2, and 4 axle box vibration transducers 1 and 4 framework vibration transducers 2 are installed for gathering the vibration information of traveling system by each traveling system.Described trackside vibration transducer L ₁-L ₄and R ₁-R, axle box vibration transducer 1 and framework vibration transducer 2 all select piezoelectric acceleration transducer, the output voltage range of this vibration transducer is 1 ~ 5V, acceleration range is 0 ~ 100g, and the advantage of this sensor is good airproof performance, has electrostatic protection function, sturdy and durable; Described temperature sensor adopts online infrared temperature-test sensor, and this infrared temperature-test sensor range is-40 DEG C ~ 900 DEG C, has perfect modulating output, and several data interface module is optional.

The present invention is based on the train traveling system method for diagnosing faults of Multi-source Information Fusion, comprise the following steps:

Step 1: svd de-noising is carried out to vibration signal x (t) of traveling system, obtains signal y (t) after de-noising; ; Concrete steps are as follows:

(1.1) vibration signal x (t) is expressed as x=[x ₁, x ₂, x ₃..., x _n], structure attractor track matrix D _m:

$D_m={\left(\begin{array}{cccc}{x}_1& x_2& ...& x_n\\ x_{l\×\mathrm{\τ}+1}& x_{l\×\mathrm{\τ}+2}& ...& x_{l\×\mathrm{\τ}+n}\\ ...& ...& ...& ...\\ x_{(m-1)\×\mathrm{\τ}+1}& x_{(m-1)\×\mathrm{\τ}+2}& ...& x_{(m-1)\×\mathrm{\τ}+n}\end{array}\right)}_{m\×n}$

In formula: τ is time delay, m is Embedded dimensions, and N is the component number of vibration signal, n=N-(m-1) × τ;

(1.2) complex autocorrelation method select time is adopted to postpone τ, time series { x _ncomplex autocorrelation function be:

$R_{\mathrm{xx}}^m\left(\mathrm{\τ}\right)=\underset{j=1}{\overset{m-1}{\mathrm{\Σ}}}{R}_{\mathrm{xx}}\left(\mathrm{j\τ}\right)$

In formula, m is Embedded dimensions, and j is the multiple of delay time, gets first zero crossing be time delay τ;

(1.3) to matrix D _mcarry out svd:

D _m＝USV'

Wherein, U ∈ R ^{m × n}, V' ∈ R ^{n × n}, and UU'=I, VV'=I, S=diag (σ ₁, σ ₂..., σ _r), m is Embedded dimensions, and n is the n-th vibration signal component, (σ ₁, σ ₂..., σ _r) be matrix D _mdiagonal matrix after svd, wherein σ _rfor singular value component, R ^{m × n}for m × n ties up real number matrix;

(1.4) singular value corresponding for noise signal is set to 0, construct new eigenmatrix, utilize this matrix to carry out anti-singular value and calculate fault-signal y (t) after can obtaining noise reduction.

Step 2, carries out population mean empirical mode decomposition to signal y (t) after de-noising, obtains intrinsic mode function IMF _n; Concrete steps are as follows:

(2.1) initialization population mean number of times M and the noise amplitude added, first time decomposes number of times p=1;

(2.2) to add the signal after making an uproar for the m time be y _p(t), y _pt () equals y (t) and the white noise n added for the p time _p(t) sum:

y _p(t)＝y(t)+n _p(t)

(2.3) population mean empirical mode decomposition is utilized (to be called for short: EEMD) decompose and add the signal y after making an uproar _pt (), obtains one group of IMF _q,p, IMF _q,pbe decompose q the IMF obtained the p time;

(2.4) IMF decomposed for P time is calculated _q,ppopulation mean obtain final IMF _q:

${\mathrm{IMF}}_q=\frac{1}{P}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}{\mathrm{IMF}}_{q,p},p=\mathrm{1,2},...,P$

(2.5) to y _p(t) and IMF _qdifference repeat step (2.3) ~ (2.4).

Step 3, adopts the intrinsic mode function IMF that Type B degree of association method obtains the 2nd step _ncarry out chaff component identification, extract true component, utilize the energy of true component as fault feature vector T, and described fault feature vector T is normalized obtains T '; Concrete steps are as follows:

(3.1) to intrinsic mode function IMF _ncarry out sliding-model control with signal y (t) after de-noising, obtain IMF _n(k), y (k), k=1,2 ..., l.Order:

$d_0=\underset{k=1}{\overset{l}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)|$

$d_1=\underset{k=1}{\overset{l-1}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n(k+1)-y(k+1)-{\mathrm{IMF}}_n\left(k\right)+y\left(k\right)|$

$d_2=\frac{1}{2}\underset{k=2}{\overset{l-1}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n(k+1)-y(k+1)-2[{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)]+[{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)\left]\right|$

In formula, l is sliding-model control number, d ₀for physical features displacement difference, d ₁for physical features velocity contrast, d ₂for physical features acceleration is poor.

The Type B degree of association is as one of method most widely used in grey relational grade, for describing correlativity between each things, between factor " measuring ", its basic thought be according to the curve describing factor characteristic between similarity degree judge between each factor degree of correlation, its computing formula is:

$\mathrm{\ψ}({\mathrm{IMF}}_n,y\left(t\right))=\frac{1}{1+\frac{1}{l}{d}_0+\frac{1}{l-1}{d}_1+\frac{1}{l-2}{d}_2}$

The normalization Type B degree of association, rejects chaff component from numbered magnitude, extracts true component.Particularly, as being less than 10 when numerical value magnitude ^-3, be then chaff component, when magnitude is more than or equal to 10 ^-2time, be then true component.

(3.2) its ENERGY E is asked to each IMF component extracted _α:

$E_{\mathrm{\α}}={\∫}_{-\∞}^{+\∞}{\left|c_{\mathrm{\α}}\left(t\right)\right|}^2\mathrm{dt},\mathrm{\α}=\mathrm{1,2},...,H$

In formula, Η is real IMF _αcomponent number, c _αt () is component IMF _αamplitude.

(3.3) with energy be that element constructs a proper vector T as follows:

T＝[E ₁,E ₂,…E _Η]

(3.4) proper vector T is normalized:

T′＝[E ₁/E,E ₂/E,…E _Η/E]

In formula, t ' is the fault feature vector after normalization.

Step 4, (is called for short the proper vector T ' after normalization: training set SVM) carries out fault diagnosis as support vector machine; Concrete steps are as follows:

(4.1) training dataset is set up;

According to each fault feature vector T ' that step 2 obtains, set up training dataset (ξ _s, ζ _s), ξ _sfor the value of input variable, ζ _sfor corresponding output variable value, s is training set number.

(4.2) definite kernel function

Because radial basis function only needs to determine a variable, compare other kernel functions and easily determine, therefore, adopt gaussian radial basis function kernel function, make SVM give play to optimal performance:

K(ξ _s,ζ _s)＝exp(-|ξ _s-ζ _s| ²/σ ²)

In formula, K (ξ _s, ζ _s) be kernel function, σ is the variance of Gaussian function.

(4.3) a certain amount of training dataset (ξ is selected _s, ζ _s) SVM is trained, make it finally tend towards stability, reach classificating requirement, set up the fault diagnosis model of SVM.

(4.4) treat diagnostic sample with the SVM trained and carry out diagnosis output, what obtain each fault mode is subordinate to angle value.

Step 5, carries out traveling system local fault diagnosis based on axle temperature signal; Concrete steps are as follows:

To the extent of injury of operation security, axle temperature is divided into four grades by hot box, the heat that namely operates, low-grade fever level, heat-flash level and sharp thermal level, the threshold value of each grade is as shown in table 1.

The each grade threshold table of table 1

According to the infrared shaft temperature sensor T of trackside ₁detection data, get the temperature of maximal value as axle box of detection data.

If the temperature difference of axle box and environment temperature (20 DEG C, 40 DEG C] in for running thermal level,

If the temperature difference (40 DEG C, 70 DEG C] interior time be low-grade fever level, if the temperature difference (70 DEG C, 100 DEG C] interior time be heat-flash level, if when the temperature difference is more than 100 DEG C for swash thermal level, finally judge the fault mode of bearing according to the thermal level of axle temperature.

Wherein, normal rolling bearing is generally in running thermal level, when being in low-grade fever level, heat-flash level, sharp thermal level, judges bearing inner race fault, bearing outer ring fault, rolling body fault.

Step 6, adopts the D-S evidence theory improved to carry out Decision fusion to local fault diagnosis result; Concrete steps are as follows:

(6.1) on the basis that system is analysed in depth, the identification framework Θ={ A of structure decision system ₁, A ₂..., A _χ, χ is fault mode (burnt unit) number.

The fault intersection occurred for municipal rail train traveling system represents with identification framework Θ, then traveling is that contingent fault mode is all corresponding with the subset of in identification framework Θ.Traveling is that common fault mode comprises bearing inner race fault, bearing outer ring fault, rolling body fault, wheel tread fault, because of various uncertain factor causing trouble pattern None-identified, uses uncertain replacement, thus to set up identification framework be Θ={ A ₁, A ₂, A ₃, A ₄, A ₅, wherein A ₁represent bearing inner race fault, A ₂represent bearing outer ring fault, A ₃represent rolling body fault, A ₄represent wheel tread fault, A ₅represent that fault mode is uncertain.It should be noted that, A ₁, A ₂, A ₃, A ₄, A ₅the not necessarily above-mentioned A of order ₁represent bearing inner race fault, A ₂represent bearing outer ring fault, A ₃represent rolling body fault, A ₄represent wheel tread fault, A ₅represent that fault mode is uncertain, also can A ₁represent bearing outer ring fault, A ₂represent rolling body fault etc.

(6.2) structure is according to the evidence source V of identification framework _g, g=1,2 ..., G, G are evidence source number;

Train fault information is from trackside vibration transducer L ₁-L ₄and R ₁-R ₄, trackside shaft temperature sensor T ₁, traveling system axle box vibration transducer 1 and framework vibration transducer 2, wherein trackside vibration transducer L ₁-L ₄and R ₁-R ₄, traveling system axle box vibration transducer 1 and framework vibration transducer 1 data obtain local diagnosis result respectively by SVM, axle temperature data obtain local diagnosis result by threshold classification process, and evidence source is above local diagnosis result everywhere.

(6.3) basic reliability distribution function is constructed;

Based on the local fault diagnosis Output rusults of SVM and threshold decision, using the uncertain factor of all the other uncertain factors in the diagnosis root-mean-square error of each sample and diagnostic procedure as basic trust degree partition function.Be shown below, be totally made up of two parts: Part I is the Output rusults sum of SVM and threshold decision, and Part II is uncertainty estimation value.

$U_r=\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{\mathrm{\ζ}}_w+E_c$

In formula, U _rfor totally, ζ _wfor SVM and threshold decision diagnostic result, χ is fault mode (burnt unit) number, E _cfor uncertain estimated value.

E _csolve according to following formula:

E _c＝E ₁+E ₂

In formula, E ₁for diagnostic result root-mean-square error, E ₂for other uncertainty estimations.

E ₁solve according to following formula:

$E_1=\sqrt{\frac{1}{\mathrm{\χ}}\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{({\mathrm{\ζ}}_w-C_w)}^2}$

In formula, C _wfor the desired output vector that SVM and threshold decision export.

E ₂solve according to following formula:

$E_2=\mathrm{\κ}\sqrt{\frac{1}{\mathrm{\χ}}\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{({\mathrm{\ζ}}_w-C_w)}^2}$

In formula, κ is uncertainty coefficient, often gets 0.1.

Then basic reliability distribution function is:

$m\left({\mathrm{\ζ}}_w\right)=\frac{{\mathrm{\ζ}}_w}{U_r}$

$m\left(E_c\right)=\frac{E_c}{U_r}$

(6.4) the basic reliability distribution m (ζ under utilizing D-S composition rule to calculate the synergy of each evidence body between two _w), its concrete steps are:

For V ₁and V ₂two evidence sources, m ₁and m ₂the basic reliability distribution function corresponding with two evidence sources, A _wand B _pbe respectively corresponding burnt unit, then V ₁and V ₂synthesize according to following formula:

$m^{\′}\left(A\right)=\left\{\begin{array}{c}\frac{\underset{A_o\∩B_p=A}{\mathrm{\Σ}}{m}_1\left(A_w\right)m_2{B}_p}{1-K},A\≠\mathrm{\φ}\\ 0,A=\mathrm{\φ}\end{array}\right.$

$k=\underset{A_o\∩B_p=\mathrm{\φ}}{\mathrm{\Σ}}{m}_1\left(A_w\right)m_2\left(B_p\right)$

In formula, w=1,2 ..., χ; P=1,2 ..., χ.

Repeat step 6.4, by V ₁and V ₂fusion results and other evidence sources are merged successively, obtain final basic reliability distribution function M _r, thus obtain last diagnostic result.

Embodiment:

The present embodiment data come from traveling system fault diagnosis analog platform, and the wheel diameter of traveling system model is 0.35m, and test bearing model is 6205-2RS JEM SKF, and deep-groove ball roller contact angle is 0 °, and rolling body number is 9.Utilize Discharge Processing Technology artificially to manufacture diameter is 0.5mm, the degree of depth is 0.2mm bearing inner race fault, bearing outer ring fault, rolling body fault and tread fault.Utilize 4 different sensors record respectively occur this 4 kinds of faults time signal.

Composition graphs 5a and Fig. 5 b, be vehicle-mounted axle box vibration signal for traveling during bearing inner race fault, the original signal of collection contains a large amount of burr details, is the noise of axle box vibration, by svd by these noise filterings, as can be seen from the figure after denoising, much noise is disallowable.

Composition graphs 6a-Fig. 6 h, carries out EEMD decomposition to the signal after de-noising, and obtain 7 IMF components and 1 residual volume, as can be seen from the figure, 7 components all meet the feature of intrinsic mode function.

The Type B degree of association calculated respectively between each a 7 IMF components and original signal step of going forward side by side is normalized, and result of calculation is as shown in table 2, divides front 5 IMF components of extracting to be true component from result of calculation.

Table 2 axle box vibration signal each IMF component chaff component differentiates

According to (3.2) ~ (3.4) joint, extract the energy of front 5 IMF components, the fault feature vector obtaining axle box vibration transducer bearing inner race fault is (166.0542,4.6238,1.8632,0.5638,0.1142), after normalization, obtaining proper vector is (0.9995,0.0278,0.0112,0.0034,0.0007).According to the method described above, can in the hope of one group of fault feature vector of axle box vibration transducer bearing outer ring fault, rolling body fault and tread fault.

Table 3 axle box vibration transducer fault-signal fault feature vector

According to the method described above, obtain the proper vector of 30 groups of axle box vibration transducers, and get wherein 25 groups of training samples as support vector machine, remain 5 groups as test sample book.There is the failure mode that 4 kinds different in the present invention, can represent that each failure mode exports in the following ways: bearing inner race fault: (1,0,0); Bearing outer ring fault: (0,1,0); Rolling body fault: (0,0,1); Tread fault: (1,1,0).

Sensor one group of test sample book is sent into support vector machine, and adopt the D-S evidence theory improved to merge local diagnosis result, last diagnostic result is as shown in table 4.

Basic reliability distribution value after table 4 merges and diagnostic result diagnostic result

From table 4, after Multi-source Information Fusion diagnosis, fiduciary level and the accuracy of fault are obtained for raising, there is not the situation that local diagnosis existence is judged by accident and diagnosis effect is undesirable.

In sum, the present invention compared with prior art, avoids the diagnosis uncertain problem that single-sensor may exist, has the advantages such as cost is low, precision is high, device is simple, achieve the Precise Diagnosis of traveling system fault.

It should be noted that; by reference to the accompanying drawings embodiments of the present invention are explained in detail above; but invention is not limited to above-mentioned embodiment; in the ken that one skilled in the relevant art possesses; can also make a variety of changes under the prerequisite not departing from present inventive concept, and these distortion all belong to the protection domain of the claims in the present invention.

Claims (10)

1., based on a train traveling system trouble-shooter for Multi-source Information Fusion, it is characterized in that:

Comprise trackside fault pre-alarming device and vehicle-mounted fault diagnosis device;

Described trackside fault pre-alarming device comprises the trackside vibration transducer and temperature sensor that are arranged in track side, for gathering trackside vibration signal and the axle temperature signal of traveling system;

Described vehicle-mounted fault diagnosis device comprises axle box vibration transducer and framework vibration transducer, for gathering axle box vibration signal and the framework vibration signal of traveling system.

2. a kind of train traveling system trouble-shooter based on Multi-source Information Fusion according to claim 1, is characterized in that:

Described trackside vibration transducer, axle box vibration transducer and framework vibration transducer are piezoelectric acceleration transducer;

Described temperature sensor is online infrared temperature-test sensor.

3., based on a train traveling system method for diagnosing faults for Multi-source Information Fusion, it is characterized in that, comprise the following steps:

Step 1: svd de-noising is carried out to vibration signal x (t) of traveling system, obtains signal y (t) after de-noising;

Step 2: carry out population mean empirical mode decomposition to signal y (t) after de-noising, obtains intrinsic mode function IMF _n;

Step 3: adopt Type B degree of association method to described intrinsic mode function IMF _ncarry out chaff component identification, extract true component, utilize the energy of true component as fault feature vector T, and described fault feature vector T is normalized obtains T ';

Step 4: the fault feature vector T ' after described normalization is carried out local fault diagnosis as the training set of support vector machine, exports the local fault diagnosis result of vibration signal;

Step 5: carry out traveling system local fault diagnosis based on axle temperature signal, exports the local fault diagnosis result of axle temperature signal;

Step 6: adopt the D-S evidence theory improved to carry out Decision fusion to the local fault diagnosis result of described vibration signal and the local fault diagnosis result of axle temperature signal, obtain the final fault diagnosis result of traveling system.

4. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to claim 3, is characterized in that:

Carry out svd de-noising to vibration signal x (t) of traveling system in step 1, concrete steps are as follows:

(1.1) vibration signal x (t) is expressed as x=[x ₁, x ₂, x ₃..., x _n], structure attractor track matrix D _m:

$D_m={\left(\begin{array}{cccc}{x}_1& x_2& ...& x_n\\ x_{l\×\mathrm{\τ}+1}& x_{l\×\mathrm{\τ}+2}& ...& x_{l\×\mathrm{\τ}+n}\\ ...& ...& ...& ...\\ x_{(m-1)\×\mathrm{\τ}+1}& x_{(m-1)\×\mathrm{\τ}+2}& ...& x_{(m-1)\×\mathrm{\τ}+n}\end{array}\right)}_{m\×n}$

In formula: τ is time delay, m is Embedded dimensions, and N is the component number of vibration signal, n=N-(m-1) × τ;

(1.2) complex autocorrelation method select time is adopted to postpone τ, time series { x _ncomplex autocorrelation function be:

$R_{\mathrm{xx}}^m\left(\mathrm{\τ}\right)=\underset{j=1}{\overset{m-1}{\mathrm{\Σ}}}{R}_{\mathrm{xx}}\left(\mathrm{j\τ}\right)$

In formula, m is Embedded dimensions, and j is the multiple of delay time, gets first zero crossing be time delay τ;

(1.3) to matrix D _mcarry out svd:

D _m＝USV'

Wherein, U ∈ R ^{m × n}, V' ∈ R ^{n × n}, and UU'=I, VV'=I, S=diag (σ ₁, σ ₂..., σ _r), m is Embedded dimensions, and n is the n-th vibration signal component, (σ ₁, σ ₂..., σ _r) be matrix D _mdiagonal matrix after svd, wherein σ _rfor singular value component, R ^{m × n}for m × n ties up real number matrix;

(1.4) singular value corresponding for noise signal is set to 0, constructs new eigenmatrix, utilize this matrix carry out anti-singular value calculate de-noising after signal y (t).

5. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to claim 3, is characterized in that:

In step 2, population mean empirical mode decomposition is carried out to signal y (t) after de-noising, obtain intrinsic mode function IMF _n, concrete steps are as follows:

(2.1) initialization population mean number of times M and the noise amplitude added, first time decomposes number of times p=1;

(2.2) to add the signal after making an uproar for the m time be y _p(t), y _pt () equals y (t) and the noise n added for the p time _p(t) sum:

y _p(t)＝y(t)+n _p(t)

(2.3) the signal y after utilizing population mean empirical mode decomposition to add to make an uproar _pt (), obtains one group of IMF _q,p, IMF _q,pbe decompose q the IMF obtained the p time;

(2.4) IMF decomposed for P time is calculated _q,ppopulation mean obtain final IMF _q:

${\mathrm{IMF}}_q=\frac{1}{P}\underset{p=1}{\overset{P}{\mathrm{\Σ}}}{\mathrm{IMF}}_{q,p},p=\mathrm{1,2},...,P$

(2.5) to y _p(t) and IMF _qdifference repeat step (2.3) ~ (2.4).

6. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to claim 3, is characterized in that:

Adopt Type B degree of association method to described intrinsic mode function IMF in step 3 _ncarry out chaff component identification, extract true component, utilize the energy of true component as fault feature vector T, and be normalized described fault feature vector T and obtain T ', concrete steps are as follows:

(3.1) to described intrinsic mode function IMF _ncarry out sliding-model control with signal y (t) after de-noising, obtain IMF _n(k), y (k), k=1,2 ..., l.Order:

$d_0=\underset{k=1}{\overset{l}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)|$

$d_1=\underset{k=1}{\overset{l-1}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n(k+1)-y(k+1)-{\mathrm{IMF}}_n\left(k\right)+y\left(k\right)|$

$d_2=\frac{1}{2}\underset{k=2}{\overset{l-1}{\mathrm{\Σ}}}|{\mathrm{IMF}}_n(k+1)-y(k+1)-2[{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)]+[{\mathrm{IMF}}_n\left(k\right)-y\left(k\right)\left]\right|$

In formula, l is sliding-model control number, d ₀for physical features displacement difference, d ₁for physical features velocity contrast, d ₂for physical features acceleration is poor;

Then Type B calculation of relationship degree formula is:

$\mathrm{\ψ}({\mathrm{IMF}}_n,y\left(y\right))=\frac{1}{1+\frac{1}{l}{d}_0+\frac{1}{l-1}{d}_1+\frac{1}{l-2}{d}_2}$

The normalization Type B degree of association, according to the order of magnitude of numerical value, rejects chaff component, extracts true component;

(3.2) to each true component IMF extracted _αask its ENERGY E _α:

$E_{\mathrm{\α}}={\∫}_{-\∞}^{+\∞}{\left|c_{\mathrm{\α}}\left(t\right)\right|}^2\mathrm{dt},\mathrm{\α}=\mathrm{1,2},...,H$

In formula, Η is the true component IMF extracted _αnumber, c _αt () is true component IMF _αamplitude;

(3.3) with ENERGY E _αfor element, to construct a fault feature vector T as follows:

T＝[E ₁,E ₂,…E _Η]

(3.4) proper vector T is normalized:

T′＝[E ₁/E,E ₂/E,…E _Η/E]

In formula, t ' is the fault feature vector after normalization.

7. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to claim 3, is characterized in that:

In step 4, the fault feature vector T ' after normalization is carried out fault diagnosis as the training set of support vector machine, export the local fault diagnosis result of vibration signal, concrete steps are as follows:

(4.1) training dataset is set up;

According to the fault feature vector T ' after described normalization, set up training dataset (ξ _s, ζ _s), ξ _sfor the value of input variable, ζ _sfor corresponding output variable value, s is training set number.

(4.2) definite kernel function

Adopt gaussian radial basis function kernel function: K (ξ _s, ζ _s)=exp (-| ξ _s-ζ _s| ²/ σ ²), in formula, K (ξ _s, ζ _s) be kernel function, σ is the variance of Gaussian function.

(4.3) a certain amount of training dataset (ξ is selected _s, ζ _s) support vector machine is trained, set up the fault diagnosis model of support vector machine.

(4.4) treat diagnostic sample by the support vector machine trained and carry out diagnosis output, export the local fault diagnosis result of vibration signal.

8. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to claim 3, is characterized in that, carry out traveling system local fault diagnosis in step 5 based on axle temperature signal, specific as follows:

Whether normal according to the threshold decision axle temperature of the temperature difference of axle box and environment temperature;

If the temperature difference of axle box and environment temperature (20 DEG C, 40 DEG C] in, then for running thermal level, judge that axle temperature is normal further;

If the temperature difference of axle box and environment temperature (40 DEG C, 70 DEG C] in, be then low-grade fever level, judge axle temperature fault further;

If the temperature difference of axle box and environment temperature (70 DEG C, 100 DEG C] in, be then heat-flash level, judge axle temperature fault further;

If the temperature difference of axle box and environment temperature is more than 100 DEG C, then for swashing thermal level, judge axle temperature fault further.

9. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to claim 3, is characterized in that:

Adopt the local fault diagnosis result of the D-S evidence theory of improvement to described vibration signal and axle temperature signal to carry out Decision fusion in step 6, obtain the final fault diagnosis result of traveling system, concrete steps are as follows:

(6.1) the identification framework Θ={ A of decision system is constructed ₁, A ₂..., A _χ, χ is fault mode number, A _χfor fault mode;

(6.2) structure is according to the evidence source V of identification framework _g, g=1,2 ..., G, G are evidence source number;

(6.3) basic reliability distribution function is constructed;

Basic the overall of trust degree partition function is made up of two parts: Part I is the Output rusults sum of support vector machine and threshold decision, and Part II is uncertainty estimation value;

$U_r=\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{\mathrm{\ζ}}_w+E_c$

In formula, U _rfor totally, ζ _wfor the diagnostic result of support vector machine and threshold decision, χ is fault mode number, E _cfor uncertain estimated value;

Described uncertain estimated value E _csolve according to following formula:

E _c＝E ₁+E ₂

In formula, E ₁for the root-mean-square error of diagnostic result, E ₂for other uncertainty estimations.

E ₁solve according to following formula:

$E_1=\sqrt{\frac{1}{\mathrm{\χ}}\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{({\mathrm{\ζ}}_w-C_w)}^2}$

In formula, C _wfor the desired output vector that support vector machine and threshold decision export.

E ₂solve according to following formula:

$E_2=\mathrm{\κ}\sqrt{\frac{1}{\mathrm{\χ}}\underset{w=1}{\overset{\mathrm{\χ}}{\mathrm{\Σ}}}{({\mathrm{\ζ}}_w-C_w)}^2}$

In formula, κ is uncertainty coefficient, often gets 0.1;

Then basic reliability distribution function is:

$m\left({\mathrm{\ζ}}_w\right)=\frac{{\mathrm{\ζ}}_w}{U_r}$

$m\left(E_c\right)=\frac{E_c}{U_r}$

(6.4) the basic reliability distribution m (ζ under utilizing D-S composition rule to calculate the synergy of each evidence body between two _w), its concrete steps are:

For V ₁and V ₂two evidence sources, m ₁and m ₂the basic reliability distribution function corresponding with two evidence sources, A _wand B _pbe respectively corresponding fault mode, then V ₁and V ₂synthesize according to following formula:

$m^{\′}\left(A\right)=\left\{\begin{array}{c}\frac{\underset{A_o\∩B_p=A}{\mathrm{\Σ}}{m}_1\left(A_w\right)M_2\left(B_p\right)}{1-K},A\≠\mathrm{\φ}\\ 0,A=\mathrm{\φ}\end{array}\right.$

$k=\underset{A_o\∩B_p=\mathrm{\φ}}{\mathrm{\Σ}}{m}_1\left(A_w\right)m_2\left(B_p\right)$

In formula, w=1,2 ..., χ; P=1,2 ..., χ.

Repeat step 6.4, by V ₁and V ₂fusion results and other evidence sources are merged successively, obtain final basic reliability distribution function M _r, thus obtain last diagnostic result.

10. a kind of train traveling system method for diagnosing faults based on Multi-source Information Fusion according to any one of claim 3-9, is characterized in that:

Described vibration signal x (t) comprises trackside vibration signal, axle box vibration signal and framework vibration signal.