CN103761372A - Multilevel inverter fault diagnosis strategy based on principal component analysis and multi-classification related vector machine(PCA-mRVM) - Google Patents

Abstract

The invention discloses a multilevel inverter fault diagnosis strategy based on a principal component analysis and multi-classification related vector machines (PCA-mRVM). The multilevel inverter fault diagnosis strategy includes: subjecting primary samples to dimensionality reduction through principal component analysis, and extracting multiple principal components with fault features so as to form training samples; subjecting the training samples to fault diagnosis through the multi-classification related vector machine, outputting probabilities of fault classifications, and taking the fault classifications with the maximum probabilities as diagnosis results. The multilevel inverter fault diagnosis strategy has the advantages in the fault diagnosis with larger sample space and more classifications, is high in model sparseness, low in computation complexity and the like; most importantly, the probabilities of classification members can be outputted through the mRVM, probability and statistic significance is achieved, and uncertain problems can be conveniently analyzed.

Description

A kind of multi-electrical level inverter Fault Diagnosis Strategy based on pivot analysis and many classification Method Using Relevance Vector Machines

Technical field:

The present invention relates to the fault diagnosis of the multi-electrical level inverter in field of power electronics, relate in particular to a kind of method for diagnosing faults based on PCA-mRVM.

Background technology:

Along with the development of Power Electronic Technique and the reduction of power electronic devices production cost, high-voltage high-power converter is widely used in various electrical equipments, such as high-power AC motor transmission, active power filtering and new-energy grid-connected etc.And in order to meet the power system development demand day by day improving, multi-electrical level inverter is as a kind of novel transducer, because of advantages such as its main circuit structure are simple and control circuit is very flexible, as the high-power conversion of field of power electronics mesohigh aspect most active Yi Ge branch, in commercial production, obtained promoting rapidly.Wherein Cascade H bridge type inverter does not need a large amount of catching diodes or electric capacity because of it, without balanced capacitance voltage, is easy to modularization and expansion in structure, and work efficiency is high, has therefore obtained application comparatively widely.Although Cascade H bridge type inverter is to export same quantity level in multi-electrical level inverter and minimum a kind of of required device, its needed switching device quantity or be multiplied along with the increase of system level number.A large amount of power semiconductor devices that use have increased the possibilities that fault occurs (switch open circuit or short circuit), and along with the rising of voltage, the probability that fault occurs also becomes large.Although the generation of Cascade H bridge type multi-electrical level inverter provides a lot of facilities in the application of high pressure, large-power occasions for Power Electronic Technique, once break down, gently cause enterprise to stop production, heavy can cause catastrophic failure, to society, bring huge loss.Research shows, in the frequency conversion speed-adjusting system of inverter power supply, the fault of switching device accounts for 82.5% of whole drive system fault, is the link the most easily breaking down in drive system.

Multi-electrical level inverter open fault diagnostic method mainly contains at present: Rothenhagen has proposed the method for diagnosing faults of IGBT open circuit, the method needs a large amount of check point and sensor, can make like this circuit become more complicated, cost is higher, and is easily subject to external disturbance and causes wrong diagnosis; The people such as soup clear spring have proposed according to the Fault Identification scheme that detects inverter outgoing side PWM voltage and output current polarity and diagnose open-circuit fault of power tubes, the method is carried out subregion to the mode of operation of inverter, not only can be used for single power tube breaks down, the situation that also diagnosable a plurality of power tube breaks down, but because output current is subject to load effect, cause the method generalization ability poor, and along with the increasing of inverter level number, mode of operation is carried out to subregion and also can become very complicated; The people such as Yang Zhong woods has proposed just can realize the fault diagnosis of inverter switching device pipe by detecting the frequency content of inverter direct-flow side electric current in addition, when each switching tube is working properly, in DC side electric current, there is no harmonic components, and when having switching tube to break down, in DC side, there will be harmonic components.Although the method can detect switching tube fault, cannot specifically navigate to the position of breakdown switch pipe, and when circuit is more complicated, is easily subject to external disturbance and causes wrong diagnosis.From analyzing above, at present for the troubleshooting issue of multi-electrical level inverter, difficulty below main existence: the more difficult extraction of (1) fault signature sample; (2) sample data dimension is larger, processes compared with difficult and calculated amount is large; (3) current diagnostic method only can be exported diagnostic accuracy, can not export the concrete outcome of each classification, and uncertain problem etc. is analyzed in inconvenience.

Summary of the invention:

Technical matters to be solved by this invention is a kind of Cascade H bridge multi-electrical level inverter Fault Diagnosis Strategy that is applicable to proposing for the technical matters of above-mentioned prior art, its object is exactly poor in order to overcome the generalization ability existing in prior art, cost is high, the shortcomings such as diagnosis speed is slow, by utilizing FFT that sample data is transformed into frequency domain from time domain, thereby be convenient to the extraction of fault signature sample, secondly by PCA, carry out dimensionality reduction, reduce calculated amount etc., finally adopt mRVM, can export membership probability of all categories, output has probability statistics meaning, be convenient to analyze uncertain problem.

To achieve these goals, the present invention has adopted following technical scheme:

1, a multi-electrical level inverter realistic model of structure

Adopt Matlab/Simulink to build single-phase Cascade H bridge five level inverter circuit system models, then add failsafe link, carry out successively the single switching transistor open fault test of diverse location, extract the fault characteristic signals of one-period, set up fault signature collection;

2, data pre-service

Utilize FFT that signal is transformed into the comparatively obvious frequency domain of fault signature and extracts fault characteristic signals;

3, extract fault pivot

Because fault signature sample dimension is larger, the computation complexity of directly processing is higher, and processing speed is slower, therefore adopts PCA to carry out dimensionality reduction to sample herein, extract the influential major component of classifying, thus the processing speed of significantly improving;

4, the model training stage

Initialization mRVM parameter, the present invention adopts RBF kernel function as the kernel function of mRVM, and carrys out definite kernel ginseng with cross-validation method.Then by the fault signature sample substitution mRVM model training after simplifying; After training finishes, by test sample book, this Diagnostic Strategy is tested, if the nicety of grading that training obtains reaches the accuracy requirement of setting, trained; If training result does not reach both provisioning requests, jump to characteristic extraction step and change the core ginseng of mRVM.The accuracy requirement of training can be set voluntarily according to condition of work.

5, the real-time diagnosis stage

In this stage, directly sample data to be sorted is imported in the model training, can obtain the probability output result of each sample, and using the sample of maximum probability wherein as classification diagnosis result.

Technique effect of the present invention is: PCA-mRVM strategy of the present invention, first, on the basis of traditional RVM classification diagnosis algorithm, added dimensionality reduction link, to meet the situation with larger samples, and PCA algorithm of the present invention, the feature samples extracting is more accurately and have more representativeness; Secondly, this Diagnostic Strategy can directly be exported membership probability of all categories, and output has probability statistics meaning, is convenient to analyze uncertain problem.Because this Diagnostic Strategy only adopts output end voltage signal as its input signal, therefore avoided a large amount of uses of sensor, and reduced complexity and the construction cost of system.

Accompanying drawing explanation

Fig. 1 is multi-electrical level inverter real-time fault diagnosis policy construction figure of the present invention, wherein:

1: threephase asynchronous machine

2: Cascade H bridge multi-electrical level inverter

3: voltage sampling circuit

Fig. 2 is the model training process flow diagram flow chart of Fault Diagnosis Strategy of the present invention;

Fig. 3 is the output voltage waveform of inverter single switching transistor fault in the present invention;

Fig. 4 is for carrying out FFT spectrogram afterwards to output voltage waveforms;

Fig. 5 is layering Bayesian model structural drawing;

Fig. 6 is the program flow diagram of Fast Type-II ML.

Embodiment:

For technological means, creation characteristic that the present invention is realized, reach object and effect is easy to understand, below in conjunction with concrete diagram, further explain the present invention.

A kind of multi-electrical level inverter real-time fault diagnosis policy construction figure as shown in Figure 1.Diagnostic Strategy is mainly comprised of four parts: data pre-service and feature extraction, classification diagnosis model, output diagnostic result, switching mode calculate.The present invention mainly studies first three part, and switching mode calculates and puts aside here.This strategy principle of work is as follows: first the output voltage signal of inverter is carried out to pre-service and to its feature extraction, thereby reduce the dimension of sample; Secondly, the data after processing are brought into the mRVM model (only need train once when this model is applied to strange sample) having trained when initial and carry out diagnostic classification; Then export membership probability of all categories, and using the sample of maximum probability wherein as classification diagnosis result.

The waveform that the people such as Surin Khomifoi have analyzed its output voltage of cascade connection multi-level is to be closely related with the break-make of switch, and the single-phase Cascade H bridge of take is example, if there is a short trouble, some switched voltage outputs will be zero always.And if there is an open fault, output voltage be (the interior average of one-period).When breaking down (open circuit or short circuit), switch can not produce the output voltage waveforms of expectation.Analyze knownly, when different H bridges unit breaks down, its fault signature is different, and output phase voltage signal is closely related with fault type and position, and exports phase voltage not with load variations.Therefore the present invention adopts output voltage as fault diagnosis signal.

The PCA-mRVM multi-electrical level inverter Diagnostic Strategy that article the present invention below proposes, mainly comprises model construction, data pre-service, extracts fault pivot, model training stage and test phase etc.Fig. 2 by the training process process flow diagram of proposition Fault Diagnosis Strategy, concrete step is as follows.

Step1: model construction

Adopt Matlab/Simulink to build single-phase Cascade H bridge five level inverter circuit system models, utilize SHPWM(Subharmonic PWM) to control, frequency of modulated wave is 50Hz, and carrier frequency is 4kHz, and sample frequency is 40kHz.Then add failsafe link, carry out successively the single switching transistor open fault test of diverse location, extract the output voltage signal of every kind of fault one-period of inverter, as shown in Figure 3, set up fault signature collection;

Step2: data pre-service

The discrete output voltage signal of the one-period that sampling obtains

after FFT conversion, obtain data set X _{a * b}, wherein a is observed reading number, b is harmonic wave number (corresponding 0 to b-1 subharmonic).FFT transformation for mula:

here W _b=e ^{-j2 π/b}, k=0,1 ..., b/2-1.By pretreated fault sample data X _{a * b}l is corresponding one by one with its tag along sort vector, and as input sample data [X, l], Fig. 4 has provided the spectrogram after one group of sample preprocessing, above-mentioned X _{a * b}is X here _{9 * 400}; Step3: extract fault pivot

(1) by step 2, obtain pretreated data sample X ∈ R ^{a * b}, wherein a is corresponding data sample points, b is corresponding sample dimension.

(2) calculate covariance matrix, solve covariance matrix R _xeigenvalue λ and corresponding feature value vector P:

Covariance matrix: R _x=E{[X-E (X)] [X-E (X)] ^t;

Pass through solving equation | λ I-R _x|=0 He | λ _ii-R _x| p _i=0, i=1,2 ..., b tries to achieve eigenvalue λ and feature value vector P.Wherein, λ _ifor R _xi eigenwert, meet λ ₁>=λ ₂>=...>=λ _b, p _icorresponding to eigenvalue λ _iproper vector, P=[p ₁, p ₂..., p _b] ^t.

(3) according to contribution rate of accumulative total, select relative principal component number m:

$\mathrm{CPV}\left(m\right)=\left(\frac{{\mathrm{\Σ}}_{i=1}^m{\mathrm{\λ}}_i}{{\mathrm{\Σ}}_{i=1}^b{\mathrm{\λ}}_i}\right)\×100\%\≥\mathrm{CL}$

Wherein, CL is a reference value, and its value can User Defined, generally chooses CL=85%.

(4) matrix that counts the score:

T=X * P, wherein, T=[t ₁, t ₂..., t _b];

After pivot analysis, X principal component model is:

$X=\mathrm{TP}=\underset{i=1}{\overset{b}{\mathrm{\Σ}}}{t}_i{p}_i^T=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{t}_i{p}_i^T+\underset{i=m+1}{\overset{b}{\mathrm{\Σ}}}{t}_i{p}_i^T=X_p+E$

Wherein, m is the number of got pivot, t _ipivot score vector, p _ibeing load vector, is again the proper vector of X covariance matrix.X _pfor a front m pivot, carry out the Main change of representative data, E is error matrix.

Step4: model training stage

(1) to pivot sample X _pset gradually class label l ∈ 1,2 ... C}, C is fault sum, data sample is now X _p∈ R ^{n * M}, and training sample (X is set respectively _train, l _train) and test sample book (X _test, l _test).

(2) selected kernel function: the present invention chooses radial basis kernel function (RBF), the wide parameter of height that wherein σ is kernel function.

(3) m-RVM model construction

Introduce auxiliary regression target Y ∈ R ^{c * N}with weight parameter W ∈ R ^{n * C}, obtain standard noise regression model:

Introduce multinomial probability Copula, regressive object be converted into class label:

$l_n=i,y_{\mathrm{ni}}>y_{\mathrm{nj}},\∀j\≠i---\left(2\right)$

Consequent multinomial probability likelihood function is as follows, wherein u N (0,1):

$P(t_n=i|W,k_n)={\mathrm{\ϵ}}_{p\left(u\right)}\left\{\underset{j\≠i}{\mathrm{\Π}}\mathrm{\φ}(u+{(w_i-w_j)}^T{k}_n)\right\}---\left(3\right)$

In order to ensure the sparse property of model, for weight vectors, introduce zero-mean, variance is

standard normal prior distribution.By priori parameter alpha _ncthe matrix forming is designated as A ∈ R ^{n * C}, α _ncobeying super parameter is τ, and the Gamma of υ distributes, and Fig. 5 has provided layering Bayesian model structure.

According to the Bayesian model shown in Fig. 5, derive posterior probability, wherein A _cdiagonal matrix for the c row derivation by A:

$P\left(W\right|Y)\∝P\left(Y\right|W)P\left(W\right|A)\∝\underset{c=1}{\overset{C}{\mathrm{\Π}}}N({({\mathrm{KK}}^T+A_c)}^{-1}{\mathrm{Ky}}_c^T,{({\mathrm{KK}}^T+A_c)}^{-1})---\left(4\right)$

By maximum a posteriori probability, estimate to such an extent that method can obtain:

$\hat{W}=\arg {\max }_{w}P\left(W\right|Y,A,K)---\left(5\right)$

Therefore during given classification the update method based on maximum a posteriori (MAP) of weight as shown in the formula:

${\hat{w}}_c={({\mathrm{KK}}^T+A_c)}^{-1}{\mathrm{Ky}}_c^T---\left(6\right)$

According to above formula, the E-step form of the auxiliary variable of can deriving, for i class as shown in the formula:

${\tilde{y}}_{\mathrm{in}}\←{\hat{w}}_i^T{k}_n-(\underset{j\≠i}{\mathrm{\Σ}}{\tilde{y}}_{\mathrm{jn}}-{\hat{w}}_j^T{k}_n)---\left(7\right)$

For

as shown in the formula:

${\hat{y}}_{\mathrm{cn}}\←{\hat{w}}_c^T{k}_n-\frac{{\mathrm{\ϵ}}_{P\left(u\right)}\left\{N_u({\hat{w}}_c^T{k}_n-{\hat{w}}_i^T{k}_n,1){\mathrm{\Φ}}_u^{n,i,c}\right\}}{{\mathrm{\ϵ}}_{P\left(u\right)}\left\{\mathrm{\Φ}({u+\hat{w}}_i^T{k}_n-{\hat{w}}_c^T{k}_n){\mathrm{\Φ}}_u^{n,i,c}\right\}}---\left(8\right)$

The posterior probability of weight vectors priori parameter distribute as shown in the formula:

$P\left(A\right|W)\&Proportional;P\left(W\right|A)P\left(A\right|a,b)\&Proportional;\underset{c=1}{\overset{C}{\mathrm{\&Pi;}}}\underset{n=1}{\overset{N}{\mathrm{\&Pi;}}}G(a+\frac{1}{2},\frac{{\mathrm{<figure-callout\; id="2"\; label="w\; nc"\; filenames="HDA0000453393840000011.png"\; state="\{\{state\}\}">w</figure-callout>}}_{\mathrm{nc}}^{}+2b}{2})---\left(9\right)$

The average that Gamma distributes is:

${\widetilde{\mathrm{\&alpha;}}}_{\mathrm{nc}}=\frac{2a+1}{{\mathrm{<figure-callout\; id="2"\; label="w\; nc"\; filenames="HDA0000453393840000011.png"\; state="\{\{state\}\}">w</figure-callout>}}_{\mathrm{nc}}^{}+2b}---\left(10\right)$

(4) m-RVM model learning

The present invention is by constructing an empty model and obtaining degree of rarefication to adding or remove sample in it, adding or removing sample is the contribution margin to model according to them, here adopt quick type-II maximum likelihood (Fast Type-II ML) parameter updating method, Fig. 6 has provided the program flow diagram of Fast Type-II ML.

Marginal likelihood function is P (Y|K, A)=log ∫ P (Y|K, W) P (W|A) dW, asks for, wherein by logarithm mode

can be decomposed into

wherein

when representative removes i sample

value, can derive:

Marginal likelihood function further can be decomposed now as shown in the formula:

$\mathrm{\&zeta;}\left(A\right)=\mathrm{\&zeta;}\left(A_{-i}\right)+\underset{c=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA0000453393840000011.png"\; state="\{\{state\}\}">C</figure-callout>}}{\mathrm{\&Sigma;}}}\frac{1}{2}[\log {\mathrm{\&alpha;}}_i-(\log {\mathrm{\&alpha;}}_i+s_i)+\frac{{\mathrm{<figure-callout\; id="2"\; label="q\; ci"\; filenames="HDA0000453393840000011.png"\; state="\{\{state\}\}">q</figure-callout>}}_{\mathrm{ci}}^{}}{{\mathrm{\&alpha;}}_i+s_i}]---\left(13\right)$

Wherein

for the sparse factor,

for quantizing factor.By solving $\&PartialD;\mathrm{\&zeta;}\left(A\right)/\&PartialD;{\mathrm{\&alpha;}}_i=0$ Can obtain stationary point:

${\mathrm{\&alpha;}}_i=\frac{{\mathrm{Cs}}_i^2}{\underset{c-1}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2-{\mathrm{Cs}}_i}i,\mathrm{if}\underset{c=1}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2>{\mathrm{Cs}}_i---\left(14\right)$

α _i=∞， $\mathrm{if}\underset{c=1}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2\&le;{\mathrm{Cs}}_i---\left(15\right)$

I sample defines contribution factor θ to the contribute information amount of model for convenience of description _i:

${\mathrm{\&theta;}}_i=\underset{c=a}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2-{\mathrm{Cs}}_i---\left(16\right)$

If every step iterative process all exists active set O, by D dimension active samples, formed.If certain sample θ value, for just, is not still but included in O, we just add this sample in O to; Otherwise, if certain sample in O, but θ value is for negative, this sample is removed.Therefore in the learning process of model, the correction of MAP update method formula is as follows, wherein K _*∈ R ^{d * N}, A _*∈ R ^{d * D}, D N:

${\hat{W}}_{*}={(K_{*}{K}_{*}^T+A_{*})}^{-1}{K}_{*}{\tilde{Y}}^T---\left(17\right)$

By formula (10) and given

now provide α _iinitialization formula as follows:

(5) m-RVM end condition

When m-RVM model three conditions below the training stage meets, training stops:

(a)

comprise all contribution samples;

(b)

reject all non-contribution samples;

(c) | logA ^(K)-logA ^(K-1)| < ε, priori parameter matrix A no longer upgrades in iterative process.

Step5: test phase

First by test sample book (T _test, l _test) after the processing of step 3 and step 4, bring in the mRVM model training, obtain the probability output result of each sample, and using the sample of maximum probability wherein as classification diagnosis result, specifically can be referring to formula (1) (2).

Application example of the present invention:

In order to verify validity and the diagnosis performance of carried Diagnostic Strategy, the present invention has built single-phase five level Cascade H bridge inverters, and the output voltage signal when obtaining the normal work of 8 switching tubes and opening a way is respectively as experimental data (only considering to only have a switching tube to break down) simultaneously, experimental data collection is at m _aunder=0.86 condition, obtain (m _afor the which amplitude modulation ratio of modulating wave with carrier wave), utilize SHPWM(Subharmonic PWM) control, frequency of modulated wave is 50Hz, and carrier frequency is 4kHz, and sample frequency is that 40kHz number of samples N is 800, to test performance of fault diagnosis of the present invention.The output voltage signal obtaining after sampling is transformed into frequency domain form through FFT.

(1) model training

First data are carried out to FFT conversion, trained and test sample book, because the front semiperiod after FFT conversion is identical with the frequency spectrum in later half cycle, therefore only choose the data of front semiperiod, i.e. training sample T _train∈ (270*400), l _train∈ (9*270), test sample book T _test∈ (180*400), l _test∈ (9*180).Adopt PCA to carry out feature extraction, CL gets 85%, and the sample obtaining after dimensionality reduction is

${\hat{l}}_{\mathrm{train}}\&Element;(9*270),{\hat{T}}_{\mathrm{test}}\&Element;(180*2),{\hat{l}}_{\mathrm{test}}\&Element;(9*180).$ Here mRVM model adopts RBF kernel function, and by repeatedly training, getting wherein optimum core ginseng is σ=0.23.Table 1 is fault type division and tag along sort setting.

Table 1 fault type is divided and tag along sort setting

Sequence number	Fault category	Tag along sort
			1	Normally	[1,0,0,0,0,0,0,0,0]T
2	H1S1 open circuit	[0,1,0,0,0,0,0,0,0]T
			3	H1S2 open circuit	[0,0,1,0,0,0,0,0,0]T
4	H1S3 open circuit	[0,0,0,1,0,0,0,0,0]T
			5	H1S4 open circuit	[0,0,0,0,1,0,0,0,0]T
6	H2S1 open circuit	[0,0,0,0,0,1,0,0,0]T
			7	H2S2 open circuit	[0,0,0,0,0,0,1,0,0]T
8	H2S3 open circuit	[0,0,0,0,0,0,0,1,0]T
			9	H2S4 open circuit	[0,0,0,0,0,0,0,0,1]T

(2) fault diagnosis platform test

Above-mentioned test data is brought in the model training, obtain model probability Output rusults, get that wherein one group of Q(is as follows).When the row vector in Q all represents corresponding with it test data as input with column vector, the state of inversion system is the target output value of normal, H1S1 open circuit, H1S2 open circuit, H1S3 open circuit, H1S4 open circuit, H2S1 open circuit, H2S2 open circuit, H2S3 open circuit and H2S4 open circuit, specifically can be with reference to the division in table 1.With test sample book class label l _testrelatively, not only can obtain this Model Diagnosis accuracy rate is 100%, and exports the nonzero value in Q by comparison, can show that this situation belongs to the possibility of other classifications, for further analyzing uncertain problem, lays a good foundation.

$Q=\left\{\begin{array}{ccccccccc}0.9749& 0.0140& 0.0107& 0.0001& 0.0003& 0.0000& 0.0000& 0.0000& 0.0000\\ 0.0020& 0.9851& 0.0065& 0.0000& 0.0055& 0.0009& 0.0000& 0.0000& 0.0000\\ 0.0001& 0.0179& 0.9570& 0.0012& 0.0000& 0.0006& 0.0231& 0.0000& 0.0000\\ 0.0000& 0.0000& 0.0024& 0.9897& 0.0000& 0.0000& 0.0000& 0.0079& 0.0000\\ 0.0000& 0.0018& 0.0000& 0.0000& 0.9958& 0.0000& 0.0000& 0.0000& 0.0024\\ 0.0000& 0.0002& 0.0000& 0.0000& 0.0000& 0.8515& 0.0154& 0.0000& 0.1329\\ 0.0000& 0.0000& 0.0000& 0.0000& 0.0000& 0.0269& 0.8654& 0.1077& 0.0000\\ 0.0000& 0.0000& 0.0003& 0.0005& 0.0000& 0.0000& 0.1821& 0.8172& 0.0000\\ 0.0000& 0.0003& 0.0000& 0.0000& 0.0006& 0.1826& 0.0000& 0.0000& 0.8165\end{array}\right\}$

More than show and described ultimate principle of the present invention, principal character and advantage of the present invention.The technician of the industry should understand; the present invention is not restricted to the described embodiments; that in above-described embodiment and instructions, describes just illustrates principle of the present invention; the present invention also has various changes and modifications without departing from the spirit and scope of the present invention, and these changes and improvements all fall in the claimed scope of the invention.The claimed scope of the present invention is defined by appending claims and equivalent thereof.

Claims (1)

1. the multi-electrical level inverter Fault Diagnosis Strategy with many classification Method Using Relevance Vector Machines based on pivot analysis, is characterized in that, contains successively following steps:

Step 1: construct a multi-electrical level inverter realistic model

Adopt Matlab/Simulink to build single-phase Cascade H bridge five level inverter circuit system models, utilize SHPWM to control, frequency of modulated wave is 50Hz, carrier frequency is 4kHz, sample frequency is 40kHz, then adds failsafe link, carries out successively the single switching transistor open fault test of diverse location, extract the output voltage signal of every kind of fault one-period of inverter, set up fault signature collection;

Step 2: data pre-service

The discrete described output voltage signal of the one-period that step 1 sampling obtains

after FFT conversion, obtain data set X _{a * b}, wherein a is observed reading number, b is harmonic wave number (corresponding 0 to b-1 subharmonic), FFT transformation for mula:

here W _b=e ^{-j2 π/b}, k=0,1 ..., b/2-1, by pretreated fault sample data X _{a * b}l is corresponding one by one with its tag along sort vector, as input sample data [X, l];

Step 3: extract fault pivot

(1) by step 2, obtain pretreated data sample X ∈ R ^{a * b}, wherein a is corresponding data sample points, b is corresponding sample dimension;

(2) calculate covariance matrix, solve covariance matrix R _xeigenvalue λ and corresponding feature value vector P:

Covariance matrix: R _x=E{[X-E (X)] [X-E (X)] ^t;

Pass through solving equation | λ I-R _x|=0 He | λ _ii-R _x| p _i=0, i=1,2 ..., b tries to achieve eigenvalue λ and feature value vector P, wherein, and λ _ifor R _xi eigenwert, meet λ ₁>=λ ₂>=...>=λ _b, p _icorresponding to eigenvalue λ _iproper vector, P=[ _p1, _p2..., _pb] ^t;

(3) according to contribution rate of accumulative total, select relative principal component number m:

$\mathrm{CPV}\left(m\right)=\left(\frac{{\mathrm{\&Sigma;}}_{i=1}^m{\mathrm{\&lambda;}}_i}{{\mathrm{\&Sigma;}}_{i=1}^b{\mathrm{\&lambda;}}_i}\right)\&times;100\%\&GreaterEqual;\mathrm{CL}$

Wherein, CL is a reference value, and its value can User Defined, generally chooses CL=85%;

(4) matrix that counts the score:

^t= ^x* ^p, wherein, T=[t ₁, t ₂..., t _b];

After pivot analysis, X principal component model is:

$X=\mathrm{TP}=\underset{i=1}{\overset{b}{\mathrm{\&Sigma;}}}{t}_i{p}_i^T=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{t}_i{p}_i^T+\underset{i=m+1}{\overset{b}{\mathrm{\&Sigma;}}}{t}_i{p}_i^T=X_p+E$

Wherein, m is the number of got pivot, t _ipivot score vector, p _ibeing load vector, is again the proper vector of X covariance matrix, X _pfor a front m pivot is carried out the Main change of representative data, ^efor error matrix; Step 4: model training stage

(1) by pivot sample X _pbe divided into two parts, be respectively training sample X _trainwith test sample book X _test, and set gradually class label l ∈ 1,2 ... C}, C is fault sum,

(2) selected kernel function: choose radial basis kernel function, the wide parameter of height that wherein σ is kernel function,

(3) m-RVM model construction

Introduce auxiliary regression target Y ∈ R ^{c * N}with weight parameter W ∈ R ^{n * C}, obtain standard noise regression model:

Introduce multinomial probability Copula, regressive object be converted into class label:

$l_n=i,y_{\mathrm{ni}}>y_{\mathrm{nj}},\&ForAll;j\&NotEqual;i$

Consequent multinomial probability likelihood function is as follows, wherein u N (0,1):

$P(t_n=i|W,k_n)={\mathrm{\&epsiv;}}_{p\left(u\right)}\left\{\underset{j\&NotEqual;i}{\mathrm{\&Pi;}}\mathrm{\&phi;}(u+{(w_i-w_j)}^T{k}_n)\right\}$

In order to ensure the sparse property of model, for weight vectors, introduce zero-mean, variance is

standard normal prior distribution, by priori parameter alpha _ncthe matrix forming is designated as A ∈ R ^{n * C}, α _ncobeying super parameter is τ, and the Gamma of υ distributes;

According to Bayesian model, derive posterior probability, wherein A _cfor the diagonal matrix of the c row derivation by A,

$P\left(W\right|Y)\&Proportional;P\left(Y\right|W)P\left(W\right|A)\&Proportional;\underset{c=1}{\overset{C}{\mathrm{\&Pi;}}}N({({\mathrm{KK}}^T+A_c)}^{-1}{\mathrm{Ky}}_c^T,{({\mathrm{KK}}^T+A_c)}^{-1})$

By maximum a posteriori probability, estimate to such an extent that method can obtain:

$\hat{W}=\arg {\max }_{w}P\left(W\right|Y,A,K)$

Therefore during given classification the update method based on maximum a posteriori (MAP) of weight as shown in the formula:

${\hat{w}}_c={({\mathrm{KK}}^T+A_c)}^{-1}{\mathrm{Ky}}_c^T$

According to above formula, the E-step form of the auxiliary variable of can deriving, for i class as shown in the formula:

${\tilde{y}}_{\mathrm{in}}\&LeftArrow;{\hat{w}}_i^T{k}_n-(\underset{j\&NotEqual;i}{\mathrm{\&Sigma;}}{\tilde{y}}_{\mathrm{jn}}-{\hat{w}}_j^T{k}_n)$

For

as shown in the formula:

${\hat{y}}_{\mathrm{cn}}\&LeftArrow;{\hat{w}}_c^T{k}_n-\frac{{\mathrm{\&epsiv;}}_{P\left(u\right)}\left\{N_u({\hat{w}}_c^T{k}_n-{\hat{w}}_i^T{k}_n,1){\mathrm{\&Phi;}}_u^{n,i,c}\right\}}{{\mathrm{\&epsiv;}}_{P\left(u\right)}\left\{\mathrm{\&Phi;}({u+\hat{w}}_i^T{k}_n-{\hat{w}}_c^T{k}_n){\mathrm{\&Phi;}}_u^{n,i,c}\right\}}$

The posterior probability of weight vectors priori parameter distribute as shown in the formula:

$P\left(A\right|W)\&Proportional;P\left(W\right|A)P\left(A\right|a,b)\&Proportional;\underset{c=1}{\overset{C}{\mathrm{\&Pi;}}}\underset{n=1}{\overset{N}{\mathrm{\&Pi;}}}G(a+\frac{1}{2},\frac{w_{\mathrm{nc}}^2+2b}{2})$

The average that Gamma distributes is:

${\widetilde{\mathrm{\&alpha;}}}_{\mathrm{nc}}=\frac{2a+1}{w_{\mathrm{nc}}^2+2b}$

(4) m-RVM model learning

By constructing an empty model and obtaining degree of rarefication to adding or remove sample in it, adding or removing sample is the contribution margin to model according to them, adopts quick t _ype-II maximum likelihood (Fast T _ype-II ML) parameter updating method;

Marginal likelihood function is P (Y|K, A)=log ∫ P (Y|K, W) P (W|A) dW, asks for, wherein by logarithm mode

can be decomposed into

wherein

when representative removes i sample

value, can derive:

By marginal likelihood function further decompose as shown in the formula:

$\mathrm{\&zeta;}\left(A\right)=\mathrm{\&zeta;}\left(A_{-i}\right)+\underset{c=1}{\overset{C}{\mathrm{\&Sigma;}}}\frac{1}{2}[\log {\mathrm{\&alpha;}}_i-(\log {\mathrm{\&alpha;}}_i+s_i)+\frac{q_{\mathrm{ci}}^2}{{\mathrm{\&alpha;}}_i+s_i}]$

Wherein

for the sparse factor,

for quantizing factor.By solving $\&PartialD;\mathrm{\&zeta;}\left(A\right)/\&PartialD;{\mathrm{\&alpha;}}_i=0$ Can obtain stationary point:

${\mathrm{\&alpha;}}_i=\frac{{\mathrm{Cs}}_i^2}{\underset{c-1}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2-{\mathrm{Cs}}_i}i,\mathrm{if}\underset{c=1}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2>{\mathrm{Cs}}_i$

α _i=∞， $\mathrm{if}\underset{c=1}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2\&le;{\mathrm{Cs}}_i$

I sample defines contribution factor θ to the contribute information amount of model for convenience of description _i:

${\mathrm{\&theta;}}_i=\underset{c=a}{\overset{C}{\mathrm{\&Sigma;}}}{q}_{\mathrm{ci}}^2-{\mathrm{Cs}}_i$

If all there is active set O in every step iterative process, by M dimension active samples, formed, if certain sample θ value is for just, be not still but included in O, we just add this sample in O to; Otherwise, if certain sample in O, but θ value is for negative, this sample is removed; Therefore in the learning process of model, the correction of MAP update method formula is as follows, wherein K _*∈ R ^{m * N}, A _*∈ R ^{m * M}, M N:

${\hat{W}}_{*}={(K_{*}{K}_{*}^T+A_{*})}^{-1}{K}_{*}{\tilde{Y}}^T$

By above-mentioned stationary point and given

now provide α _iinitialization formula as follows:

(5) m-RVM end condition

When m-RVM model three conditions below the training stage meets, training stops:

(a)

comprise all contribution samples;

(b)

reject all non-contribution samples;

(c) | logA ^(K)-logA ^(K-1)| < ε, priori parameter matrix A no longer upgrades in iterative process;

Step 5: test phase

First by test sample book (T _test, l _test) after the processing of step 3 and step 4, bring in the mRVM model training, obtain the probability output result of each sample, and using the sample of maximum probability wherein as classification diagnosis result.

Images