CN106441888A - High-speed train rolling bearing fault diagnosis method - Google Patents

Abstract

A high-speed train rolling bearing fault diagnosis method includes the steps: acquiring original vibration signals and decomposing the original vibration signals by an EEMD (ensemble empirical mode decomposition) method, selecting previous a IMF (interactive media forum) components, calculating energy of the components and total energy and normalizing the components to obtain an energy feature vector; determining an RBF (radial basis function) nerve network structure, determining the node number of an input layer, an output layer and a hidden layer, determining training target precision and distribution density, selecting a training sample and a testing sample, taking the training sample as input for training, acquiring a preliminary RBF nerve network diagnosis model after reaching target precision, taking the testing sample as input of a preliminary model to recognize the testing sample, and acquiring a final RBF nerve network diagnosis model for diagnosing a bearing fault type if a fault recognition rate meets an ideal standard. A new idea is provided for improving high-speed train rolling bearing fault diagnosis accuracy and instantaneity, and the performance and running safety of a high-speed train are further ensured.

Description

A kind of bullet train Fault Diagnosis of Roller Bearings

Technical field

The present invention relates to a kind of method for diagnosing faults being modeled using characteristic signal and classifying is and in particular to a kind of base Bullet train Fault Diagnosis of Roller Bearings in EEMD and RBF neural.

Background technology

, as one of the vitals of bullet train, its state quality is most important to safe train operation for rolling bearing. Speedup increment of load is the trend of countries in the world railway development, and having the full train of pull strength is to improve speed, increase freight volume Premise, now as the worth more concerns of rolling bearing of one of bullet train vitals.Rolling as mechanical consumable accessory Dynamic bearing, an outstanding feature is that life-span discreteness is big, and failure cause is complicated.In actual applications, have makes rolling bearing It is far from reaching projected life with the time and various faults but occur, the projected life that exceeds well over having but still can normal work.Cause , in order to prevent bearing fault, the operating condition of monitoring bearing is extremely necessary for this.

At present, it is to analyze its vibration signal mostly to the diagnosis of bearing fault, and vibration signal has non-linear, non-stationary The features such as property.EEMD is a kind of noise assistance data processing method, is suitable for analyzing and processing non-linear, non-stationary signal, utilizes It can obtain the information giving full expression to signal characteristic.Bearing failure diagnosis based on RBF neural diagnose than BP neural network Accuracy is high, speed faster, and is less prone to local minimum, more suitable for carrying out the fault diagnosis of bearing.

Content of the invention

The present invention is unhappy for prior art structure model velocity, and the not high defect of Fault Identification accuracy rate provides one kind Bullet train Fault Diagnosis of Roller Bearings.It can be bullet train rolling bearing fault diagnosis and status monitoring research carries For a kind of new thinking, also the performance for train and traffic safety provide and are further ensured that.

The present invention is directed to the deficiencies in the prior art, provides one kind

To achieve these goals, present invention employs technical scheme below：

A kind of bullet train Fault Diagnosis of Roller Bearings, comprises the following steps：

I, the foundation of fault diagnosis model

Step 1：Bearing is divided into normal bearing, rolling element faulty bearings, outer ring faulty bearings and inner ring faulty bearings four Kind of Status Type, the bearing to above four kinds of Status Types, every kind of gather some groups of original vibration signal respectively；

Step 2：EEMD fault feature vector constructs

1) using EEMD method, every group of original vibration signal is decomposed, chooses and decompose the front a IMF component obtaining, And obtain the ENERGY E of each component respectively_i；

In formula, C_iT is i-th IMF component, i=1,2,3, L, a, C_iIt is the amplitude of discrete point, n is sampled point number,

2) seek energy summation E of each IMF component of every group of original vibration signal；

3) because the Oscillation Amplitude difference suffered by different conditions bearing is larger, each IMF component values is made to differ also larger, So being normalized to energy, energy and the gross energy of each IMF component will seek ratio, obtaining one group of original vibration The feature parameter vectors T of signal；

T=E₁/ E, E₂/ E, L, E_a/E

Step 3：RBF neural models

1) determine RBF neural network structure

Although the quantity increasing hidden layer neuron can improve the non-linear mapping capability of RBF neural, neural First quantity can reduce neural network forecast performance too much, so using three layers of RBF neural of single hidden layer,

2) nodes of input layer are determined

Using the feature parameter vectors T as RBF neural input, the therefore nodes M=a of input layer,

3) nodes of output layer are determined

Preferably output result should be able to directly see out of order classification, so adopting 3 binary codes, the i.e. section of output layer Count as 3, be shown in Table 1,

4) determine the nodes of hidden layer, determine the training objective precision of RBF neural and the distribution of RBF Density SPREAD,

5) from the original vibration signal of every kind of bearing state choose b group as training sample, remaining as test sample, Training sample is trained as the input of RBF neural, training reaches step 4) after the aimed at precision that sets, obtain just Then test sample is diagnosed the input of rudimentary model, to test by step RBF neural diagnostic cast as RBF neural The state of sample bearing is identified,

6) performance evaluation of rudimentary model

Recognition result first according to test sample calculates identification error, when identification error is within the scope of accepting, recognizes Correct for recognition result, on the contrary think recognition result mistake；Then calculate fault recognition rate Q, Q=correctly identifies that number/always is real Test bearing number, RBF neural performance quality is weighed by fault recognition rate Q, if fault recognition rate Q reaches ideal standard, Obtain RBF neural diagnosis final mask, and the bearing failure diagnosis for step II unknown state, otherwise, return step Rapid 5), reselect training sample and test sample, are trained and test,

II, diagnosis bearing fault type

The original vibration signal that the follow-up off-axis of collection unknown state holds, and diagnose final mask as RBF neural Input, by the output of final mask compare with bearing fault output form determine bearing state.

The concrete decomposition step of described EEMD method is as follows：

(1) add frequency spectrum equally distributed white noise a in signal R (t) to be decomposed_mT (), obtains signal S (t)；

(2) EMD is carried out to signal S (t), catabolic process is as follows；

1) all Local Extremum on signal S (t) are determined, upper and lower two envelopes are with cubic spline curve respectively Obtained from all of Local modulus maxima and local minizing point are tied, i.e. s (t)_maxWith S (t)_min；

2) meansigma methodss of the upper lower envelope in each moment are sought, that is,

3) obtain new signal

Y₁(t)=S (t)-u (t)

Judge whether to be symmetrical in local zero-mean, and have identical extreme point and zero crossing, if it is, being designated as C₁(t), It is first IMF component, otherwise repeat step 1) and 2)；

4) by C₁T () separates from S (t), obtain a difference signal V₁(t),

V₁(t)=S (t)-C₁(t)

5) by V₁T () is as initial data, repeat step 1)～3) obtain IMF2, repeat to obtain n IMF component for n time, in It is to have

Work as V_nWhen () meets given end condition t, end condition is V_nT () is monotonic function, loop ends,

By Y₁(t) and V₁T () can get

I.e. primary signal is represented as intrinsic mode functions component and a survival function V_nThe sum of (t), each component C₁(t), C₂ (t) ... C_nT () covers the information of different frequency sections from high to low in primary signal respectively, and the change with signal itself and Change,

(3) repetitive process (1) and (2) after different white noises are added every time；

(4) by integrated average C of each IMF component after multiple decomposition_iAs final result,

Compared with prior art, the beneficial effect that the present invention possesses：

This patent provides a kind of Fault Diagnosis of Roller Bearings based on EEMD and RBF neural, by the party Method can accurately identify 4 kinds of bearing states such as normal bearing, rolling element fault, outer ring fault and inner ring fault, arranges at a high speed for improving The accuracy of car rolling bearing fault diagnosis and real-time provide new approaches, also provide for the performance and traffic safety of train It is further ensured that.

Brief description

Fig. 1 is the idiographic flow of EEMD algorithm.

Fig. 2 is the structure of RBF neural.

Fig. 3 is the primary signal figure of train rolling element faulty bearings.

Fig. 4 is the EEMD exploded view of train rolling element faulty bearings.

Fig. 5 is the feature parameter vectors figure after train rolling element faulty bearings normalization.

Fig. 6 is the training process figure of RBF neural.

Specific embodiment

With reference to specific embodiment, the present invention is described in detail.

Embodiment 1 the present embodiment completes in Matlab R2010a software.

A kind of bullet train Fault Diagnosis of Roller Bearings, comprises the following steps：

I, the foundation of fault diagnosis model

Step 1：Bearing is divided into normal bearing, rolling element faulty bearings, outer ring faulty bearings and inner ring faulty bearings four Kind of Status Type, the bearing to above four kinds of Status Types, every kind of gather some groups of original vibration signal respectively,

The present embodiment concrete grammar is as follows：First by spark erosion technique, rolling element event is arranged on rolling bearing Barrier, outer ring fault and inner ring fault, fault diameter is 0.018 centimetre, is 1.2KHz in sample frequency, and bearing rotating speed is Under the operating mode of 1797r/min, vibration signal is gathered using acceleration transducer, by the DAT recorder collection of 16 passages.Data Adopt 40 groups altogether, i.e. normal bearing, rolling element faulty bearings, outer ring faulty bearings, each 10 groups of inner ring faulty bearings, with rolling element Faulty bearings data instance, sampling number is 12000, and original signal waveform is as shown in Figure 3.

Step 2：EEMD fault feature vector constructs

1) using EEMD method, every group of original vibration signal is decomposed, chooses and decompose the front a IMF component obtaining, The present embodiment a=8, as shown in figure 4, and obtain the ENERGY E of each component respectively_i；

In formula, C_iT () is i-th IMF component, i=1,2,3, L, a, C_iIt is the amplitude of discrete point, n is sampled point number,

2) seek energy summation E of each IMF component of every group of original vibration signal；

3) because the Oscillation Amplitude difference suffered by different conditions bearing is larger, each IMF component values is made to differ also larger, So being normalized to energy, energy and the gross energy of each IMF component will seek ratio, obtaining one group of original vibration The feature parameter vectors T of signal, as shown in Figure 5；

T=[E₁/ E, E₂/ E, L, E_a/E]

The 1 of repeat step 2)～4) respectively EEMD is carried out to each group of data, thus obtaining 40 groups of corresponding different conditions axles The feature parameter vectors holding.

Step 3：RBF neural models

1) determine RBF neural network structure

Although the quantity increasing hidden layer neuron can improve the non-linear mapping capability of RBF neural, neural First quantity can reduce neural network forecast performance too much, so using three layers of RBF neural of single hidden layer,

2) nodes of input layer are determined

Using the feature parameter vectors T as RBF neural input, the therefore nodes M=8 of input layer,

3) nodes of output layer are determined

Preferably output result should be able to directly see out of order classification, so adopting 3 binary codes, the i.e. section of output layer Count as 3, be shown in Table 1,

Table 1 bearing fault output form

4) determine the nodes of hidden layer, determine the training objective precision of RBF neural and the distribution of RBF Density SPREAD,

The present embodiment adopts the method that method experience and test combine, and the different node numbers in 4: 24 are tasted one by one Examination, chooses the hidden node number as model for the interstitial content 22 of best performance, meanwhile, by the training objective of RBF neural It is set as 10^-5, distribution density SPREAD of RBF takes default value 1,

5) from the original vibration signal of every kind of bearing state choose b group as training sample, remaining as test sample, Training sample is trained as the input of RBF neural, training reaches step 4) after the aimed at precision that sets, obtain just Then test sample is diagnosed the input of rudimentary model, to test by step RBF neural diagnostic cast as RBF neural The state of sample bearing is identified,

In the present embodiment, choose 6 groups from every kind of bearing state as training sample, 4 groups as test sample, therefore four Kind of bearing state totally 24 groups of training samples, remaining is test sample, using this 24 groups of training samples as RBF neural input It is trained, the training process of RBF neural, as shown in fig. 6, only need to reach setting accuracy through 18 step training, obtains just Then remaining 16 test samples are diagnosed the defeated of rudimentary model as RBF neural by step RBF neural diagnostic cast Enter, the state of test sample bearing be identified,

6) performance evaluation of rudimentary model

Recognition result first according to test sample calculates identification error, when identification error is within the scope of accepting, recognizes Correct for recognition result, on the contrary think recognition result mistake；Then calculate fault recognition rate Q, Q=correctly identifies that number/always is real Test bearing number, RBF neural performance quality is weighed by fault recognition rate Q, if fault recognition rate Q reaches ideal standard, Obtain RBF neural diagnosis final mask, and the bearing failure diagnosis for step II unknown state, otherwise, return step Rapid 5), reselect training sample and test sample, are trained and test,

The present embodiment adopts 1,2,3 and 4 to represent normal bearing, rolling element faulty bearings, race bearing and inner ring axle respectively Hold, after RBF neural output result, calculate identification error ε, if the present embodiment regulation | ε |≤0.5, illustrate recognition result just Really, on the contrary recognition result mistake, EEMD combines RBF neural diagnostic result as shown in table 2 it is known that RBF neural diagnosis Rudimentary model identification output result is sufficiently close to target detection output result, it can thus be appreciated that this bearing failure diagnosis model can essence Really identification bearing fault type,

Further, every kind of test bearing type is counted and asked for the fault recognition rate of model, be computed normal Bearing, rolling element fault, outer ring fault, the correct recognition rata of inner ring fault are all 100%, so total discrimination is 100%, because This can clearly illustrate that EEMD combines RBF neural network model for the fault diagnosis of rolling bearing, category of model identification effect It is really highly desirable, you can to diagnose final mask as RBF neural,

II, diagnosis bearing fault type

The original vibration signal that the follow-up off-axis of collection unknown state holds, and diagnose final mask as RBF neural Input, by the output of final mask compare with bearing fault output form determine bearing state, can through relatively this enforcement Correct diagnosis bearing fault type.

Table 2 EEMD combines RBF neural fault diagnosis

More excellent than additive method in order to verify the inventive method, it is respectively adopted two methods of EMD and EEMD and extract respectively not With the characteristic vector of state bearing data, then establish EMD and BP, EMD and RBF, EEMD and BP and EEMD and RBF tetra- respectively Plant bearing failure diagnosis model, carry out the Classification and Identification of bearing failure diagnosis, simulation result by using the characteristic vector extracted As shown in table 3.As seen from table, EEMD method is more advantageous than EMD on fault feature vector extracts；BP neural network is not suitable for The rapid modeling of mass data, convergence speed is slow.And RBF neural network model precision is higher, more suitable than BP neural network Pattern recognition together in bearing fault.Therefore EEMD and RBF method has it exclusive excellent on Fault Diagnosis of The Rolling Bearings For Trains Gesture.

The Comparative result of 3 four kinds of bearing failure diagnosis models of table

Conclusion

1) in bearing failure diagnosis, there is modal overlap for EMD method it is proposed that EEMD method, the method The IMF energy of bearing vibration signal can more be accurately acquired, as the characteristic vector of different conditions bearing.

2) simulation result shows, the present invention is a kind of None-linear approximation network with superperformance, can be sufficiently accurately Identification bearing fault type.In network training process, in equating expections error sum of squares and equal input node, output node Under conditions of, the convergence rate of the present invention, apparently higher than BP neural network, not only shortens the learning time of sample and reduces Complexity, and it is less prone to local minimum.

3) relative analyses are passed through, it is feasible for carrying out fault diagnosis using the present invention to train rolling bearing, and The present invention is higher than BP neural network diagnosis efficiency and more accurate, more suitable for carrying out fault diagnosis.

Method therefore presented herein cannot be only used for rolling bearing fault diagnosis, is also fully applicable to gear-box, big The fault diagnosis of type slewing etc., be widely used prospect.

EEMD full name is Ensemble Empirical Mode Decomposition (set empirical mode decomposition), is The innovatory algorithm of EMD (Empirical Mode Decomposition), EMD method be decomposed according to the time scale feature of vibration signal itself and It is not required in advance set any basic function, but the method has modal overlap defect, EEMD method can be good at overcoming above-mentioned lacking Fall into, effectively solve the mixing phenomenon of EMD.The concrete decomposition step of described EEMD method is as follows：

(1) add frequency spectrum equally distributed white noise a in signal R (t) to be decomposed_mT (), obtains signal S (t)；

(2) EMD is carried out to signal S (t), catabolic process is as follows；

1) all Local Extremum on signal S (t) are determined, upper and lower two envelopes are with cubic spline curve respectively Obtained from all of Local modulus maxima and local minizing point are tied, i.e. S (t)_maxWith S (t)_min；

2) meansigma methodss of the upper lower envelope in each moment are sought, that is,

3) obtain new signal

Y₁(t)=S (t)-u (t)

Judge whether to be symmetrical in local zero-mean, and have identical extreme point and zero crossing, if it is, being designated as C₁(t), It is first IMF component, otherwise repeat step 1) and 2)；

4) by C₁T () separates from S (t), obtain a difference signal V₁(t),

V₁(t)=S (t)-C₁(t)

5) by V₁T () is as initial data, repeat step 1)～3) obtain IMF2, repeat to obtain n IMF component for n time, in It is to have

Work as V_nWhen () meets given end condition t, end condition is V_nT () is monotonic function, loop ends,

By Y₁(t) and V₁T () can get

I.e. primary signal is represented as intrinsic mode functions component and a survival function V_nThe sum of (t), each component C₁(t), C₂ (t), C_nT () covers the information of different frequency sections from high to low in primary signal respectively, and changing with signal itself Become and change,

(3) repetitive process (1) and (2) after different white noises are added every time；

(4) by integrated average C of each IMF component after multiple decomposition_iAs final result,

RBF (Radial Basis Function, RBF) neutral net is three layers of feed-forward type neutral net, by Input layer, hidden layer and output layer composition, as shown in Figure 2.Mapping by input layer to output layer is nonlinear, and hidden layer Mapping to output layer is linear.The essence of RBF neural is：Make the linear undistinguishable problem in lower dimensional space in height Linearly can distinguish in dimension space.The weights of the variance of basic function, the center of basic function and hidden layer to output layer are RBF nerve net The parameter of network algorithm requirements solution, in RBF neural, the most frequently used RBF is Gaussian function, therefore RBF neural Activation primitive can be expressed as：

In formula, X_mFor input variable；σ is the variance of Gaussian function；||X_m-c_i| | for European norm；c_iFor Gaussian function Center.

The obtainable network of structure of RBF neural as shown in Figure 2 is output as：

In formula, ω_ijConnection weight for hidden layer to output layer；I=1,2, L, N are hidden layer nodes；y_jBe with defeated Enter the output network of corresponding j-th node of network.

Claims (2)

1. a kind of bullet train Fault Diagnosis of Roller Bearings is it is characterised in that comprise the following steps：

I, the foundation of fault diagnosis model

Step 1：Bearing is divided into normal bearing, rolling element faulty bearings, outer ring faulty bearings and four kinds of shapes of inner ring faulty bearings State type, the bearing to above four kinds of Status Types, every kind of gather some groups of original vibration signal respectively；

Step 2：EEMD fault feature vector constructs

1) using EEMD method, every group of original vibration signal is decomposed, choose and decompose the front a IMF component obtaining, and point Do not obtain the ENERGY E of each component_i；

$E_i={\∫}_{-\mathrm{\∞}}^{+\mathrm{\∞}}|C_i\left(t\right){|}^{2}dt=\underset{1}{\overset{n}{\Σ}}|C_i{|}^2$

In formula, C_iT () is i-th IMF component, i=1,2,3, L, a, C_iIt is the amplitude of discrete point, n is sampled point number,

2) seek energy summation E of each IMF component of every group of original vibration signal；

$E=\underset{i=1}{\overset{a}{\Σ}}{E}_i$

3) because the Oscillation Amplitude difference suffered by different conditions bearing is larger, each IMF component values is made to differ also larger, so Energy is normalized, energy and the gross energy of each IMF component will seek ratio, obtain one group of original vibration signal The feature parameter vectors T；

T=[E₁/ E, E₂/ E, L, E_a/E]

Step 3：RBF neural models

1) determine RBF neural network structure

Although the quantity increasing hidden layer neuron can improve the non-linear mapping capability of RBF neural, neuron number Amount can reduce neural network forecast performance too much, so using three layers of RBF neural of single hidden layer,

2) nodes of input layer are determined

Using the feature parameter vectors T as RBF neural input, the therefore nodes M=a of input layer,

3) nodes of output layer are determined

Preferably output result should be able to directly see out of order classification, so adopting 3 binary codes, the i.e. nodes of output layer For 3, it is shown in Table 1,

Table 1 bearing fault output form

4) determine the nodes of hidden layer, determine the training objective precision of RBF neural and the distribution density of RBF SPREAD,

5) choose b group from the original vibration signal of every kind of bearing state as training sample, remaining will be instructed as test sample Practice sample to be trained as the input of RBF neural, training reaches step 4) after the aimed at precision that sets, obtain preliminary Then test sample is diagnosed the input of rudimentary model, to test specimens by RBF neural diagnostic cast as RBF neural The state of this bearing is identified,

6) performance evaluation of rudimentary model

Recognition result first according to test sample calculates identification error, when identification error is within the scope of accepting it is believed that knowing Other result is correct, otherwise thinks recognition result mistake；Then calculate fault recognition rate Q, Q=correctly identifies that number/always tests axle Hold number, RBF neural performance quality is weighed by fault recognition rate Q, if fault recognition rate Q reaches ideal standard, obtains final product Diagnose final mask, and the bearing failure diagnosis for step II unknown state, otherwise, return to step 5 to RBF neural), Reselect training sample and test sample, be trained and test,

II, diagnosis bearing fault type

The original vibration signal that the follow-up off-axis of collection unknown state holds, and diagnose the defeated of final mask as RBF neural Enter, compared the state determining bearing by the output of final mask with bearing fault output form.

2. bullet train Fault Diagnosis of Roller Bearings as claimed in claim 1 is it is characterised in that described EEMD method has Body decomposition step is as follows：

(1) add frequency spectrum equally distributed white noise a in signal R (t) to be decomposed_mT (), obtains signal S (t)；

(2) EMD is carried out to signal S (t), catabolic process is as follows；

1) all Local Extremum on signal S (t) are determined, upper and lower two envelopes are respectively by institute with cubic spline curve Obtained from some Local modulus maximas and local minizing point are tied, i.e. S (t)_maxWith S (t)_min；

2) meansigma methodss of the upper lower envelope in each moment are sought, that is,

$u\left(t\right)=\frac{S{\left(t\right)}_{max}+S{\left(t\right)}_{min}}{2}$

3) obtain new signal

Y₁(t)=S (t)-u (t)

Judge whether to be symmetrical in local zero-mean, and have identical extreme point and zero crossing, if it is, being designated as C₁(t), as First IMF component, otherwise repeat step 1) and 2)；

4) by C₁T () separates from S (t), obtain a difference signal V₁(t),

V₁(t)=S (t)-C₁(t)

5) by V₁T () is as initial data, repeat step 1)～3) obtain IMF2, repeat to obtain n IMF component for n time, then have

$\left.\begin{array}{c}{V}_1\left(t\right)-C_2\left(t\right)=V_2\left(t\right)\\ .\\ .\\ .\\ V_{n-1}\left(t\right)-C_n\left(t\right)=V_n\left(t\right)\end{array}\right\}$

Work as V_nWhen () meets given end condition t, end condition is V_nT () is monotonic function, loop ends,

By Y₁(t) and V₁T () can get

$S\left(t\right)=\underset{i=1}{\overset{n}{\Σ}}{C}_i\left(t\right)+V_n\left(t\right)$

I.e. primary signal is represented as intrinsic mode functions component and a survival function V_nThe sum of (t), each component C₁(t), C₂ (t) ... C_nT () covers the information of different frequency sections from high to low in primary signal respectively, and the change with signal itself and Change,

(3) repetitive process (1) and (2) after different white noises are added every time；

(4) by integrated average C of each IMF component after multiple decomposition_iAs final result,

$C_i=\frac{1}{M}\underset{m=1}{\overset{M}{\Σ}}{C}_{i,n},(i=1,2,Ln;m=1,2,L,M).$