CN104714171A - Switching circuit fault classifying method based on wavelet transform and ICA feature extraction - Google Patents

Abstract

The invention discloses a switching circuit fault classifying method based on wavelet transform and ICA feature extraction. The method comprises the following steps: (1) generating a pseudo random signal as a test stimulation signal; (2) defining a fault mode; (3) acquiring the original response data of the circuit; (4) pre-treating the original response data by a Haar wavelet orthogonal filter; (5) extracting the fault feature parameters, and calculating the entropy and kurtosis as well as fuzzy sets thereof of low-frequency approximate information and high-frequency detail information for the pre-treated signal respectively; and (6) constructing a fault dictionary based on the extracted fault feature parameters so as to realize fault classification of the switching circuit. The method disclosed by the invention has the advantages of skillful concept, easiness in implementation and simulation proof and can distinguish the fault types more accurately than the existing method.

Description

Based on the on-off circuit Fault Classification of wavelet transformation and ICA feature extraction

Technical field

The present invention relates to a kind of on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction.

Background technology

Switching current (Switched Current, SI) technology is the analog sampled data signal processing technology proposed late 1980s.As the substitute technology of switched capacitor technique, it is by discrete time sampled-data processing simulating signal continuous time, has the advantages such as the little and high-frequency characteristic of low-voltage, high speed, low-power consumption, chip area is good, obtains develop fast in the nearest more than ten years.SI technology does not comprise linear capacitance and high performance operational amplifier, and complete and digital CMOS process technical compatibility, is easy to the single-chip integration realizing extensive Digital Analog Hybrid Circuits.But the test and diagnostic one of simulating part in digital-to-analogue mixed signal circuit is slowly straight.Although the test of mimic channel in recent years and fault diagnosis achieve many achievements in research, but the test of the analogue technique-Switched-Current Circuit of digital technology and fault diagnosis lag behind the design and manufaction of Switched-Current Circuit, and this greatly hinders the development of SI technology always.And due to the difficulty that the imperfect performance of metal-oxide-semiconductor field effect transistor in SI circuit, switch-charge flow into, the factor such as limited frequency span and non-zero input and output conductance ratio further add the feature extraction of SI fault, make SI circuit fault diagnosis also be faced with very large challenge.

In recent years, in Switched-Current Circuit fault diagnosis fault diagnosis field, some the effectively test of practicality and method for diagnosing faults are emerged.In these diagnostic methods, feature extraction plays relative crucial effect to test and test macro.At document [Huang Jun, He Yigang. the Primary Study of Switched-Current Circuit fault diagnosis technology. modern electronic technology, 2007,30 (9): 76-78] in, author simply discusses the test of SI basic unit of storage with reference to analog circuit test and diagnostic method, has carried out hard fault test to the basic unit of storage circuit without MOS switch.Due to measurement is current parameters, causes the dependent failure quantity of information for testing and diagnose imperfect, so that can not carry out localization of fault exactly.

And at document [Guo, J., R., He, Y.G., Liu M.R..Wavelet neural network approach for testingof switched-current circuits.J Electron Test, 27:611-625,2011.], in [being designated as document [3]], author utilizes wavelet neural network to carry out diagnosis to SI circuit can correctly diagnose out all hard faults, but very low to diagnosis during muting sensitivity failed transistor especially to soft fault, be only about 80%.In addition, at document [Long, Y., He, Y.G., & Yuan, L.F.Faultdictionary based switched current circuit fault diagnosis using entropy as a preprocessor.Analog Integrated Circuits and Signal Processing, 66 (1), 2011:93-102.] in [being designated as document [1]], author's first time introduces fault signature pre-service concept in SI circuit test and diagnosis, by carrying out the extraction of information entropy preprocessed features to the failure response signal collected, computing information entropy fuzzy set builds fault dictionary, carry out failure modes, decrease calculating and failure diagnosis time, rate of correct diagnosis reaches about 95%, but the method is only adapted to the fault diagnosis of middle and small scale Switched-Current Circuit.At document [Zhang, Z., Duan, Z., Long, Y., & Yuan, L.F.A new swarm-SVM-based fault diagnosis approach for switched currentcircuit by using kurtosis and entropy as a preprocessor.Analog Integrated Circuits and SignalProcessing, vol.81, no.1.2014.] in [being designated as document [2]], author adds a characteristic parameter-kurtosis, propose based on the information entropy of particle swarm support vector machine and the pretreated SI circuit test of kurtosis and diagnostic method, soft fault diagnosis accuracy has had further raising, reach about 99%, but still can not reach the failure diagnosis of 100%.

Therefore, be necessary to design a kind of novel on-off circuit Fault Classification.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction, should be easy to implement based on the on-off circuit Fault Classification of wavelet transformation and ICA feature extraction, compare existing method, various fault type can be distinguished more accurately.

The technical solution of invention is as follows:

Based on an on-off circuit Fault Classification for wavelet transformation and ICA feature extraction, the method is used for the failure modes of Switched-Current Circuit, comprises the following steps:

Step 1: produce pseudo random signal as test and excitation signal;

Pseudo random signal is pseudo-random pulse sequence;

Step 2: failure definition pattern:

Based on circuit simulation, carry out sensitivity analysis to Switched-Current Circuit to be measured, the change obtaining component parameters changes the single order of electric network system features, with the fault element most possibly broken down in positioning circuit; And divide fault mode based on fault element location; The quantity of fault element is N, then the kind of fault mode is 2*N; N is natural number;

Step 3: the original response data of Acquisition Circuit:

Encourage tested Switched-Current Circuit by pseudo random signal, with ASIZ software, the various malfunction of tested Switched-Current Circuit and normal condition are emulated, collect original response data from the output terminal of Switched-Current Circuit; These original response data are curtage data; Can set the parameter variation range of each fault element during emulation, thus obtain the response data under various fault mode, emulation can obtain the response data under normal mode equally.

Step 4: adopt Haar small echo orthogonal filter to carry out pre-service to original response data;

Utilize Haar small echo orthogonal filter as the pretreatment system of acquisition sequence, obtain low-frequency approximation information and the detail of the high frequency of observation signal;

Step 5: Fault characteristic parameters extracts;

Entropy and the kurtosis of low-frequency approximation information and detail of the high frequency is calculated respectively for pretreated signal; Obtain following Fault characteristic parameters: low-frequency approximation entropy, low-frequency approximation kurtosis, low-frequency approximation entropy fuzzy set, low-frequency approximation kurtosis fuzzy set, high frequency detail entropy, high frequency detail kurtosis, high frequency detail entropy fuzzy set and high frequency detail kurtosis fuzzy set;

Step 6: based on the Fault characteristic parameters structure fault dictionary extracted, thus realize on-off circuit failure modes.

The computing method of described information entropy are:

Information entropy $J\left(x\right)=k_1{\left(E\right\{\mathrm{x\; exp}(-x^2/2)\left\}\right)}^2+k_2{\left(E\right\{\left|x\right|\}-\sqrt{2/\mathrm{\π}})}^2;$ In formula, $k_1=36/(8\sqrt{3}-9)$ And k ₂=1/ (2-6/ π), x are the data that the primary current response data of the circuit-under-test output terminal extracted obtains through wavelet transformation; E represents expectation value (namely E is operation of averaging); The computing method of described kurtosis are:

Kurtosis kurt (x)=E{x ⁴-3 [E{x ²] ², x is the data that the primary current response data of the circuit-under-test output terminal extracted obtains through wavelet transformation, and E represents expectation value (namely E is operation of averaging).

Fuzzy set is transistor transconductance value g _mthe information entropy obtained when range of tolerable variance changes for 5% or 10% or the constant interval of kurtosis;

Low-frequency approximation information entropy fuzzy set and detail of the high frequency entropy fuzzy set are a numerical intervals; Normal mode refers to the pattern that circuit does not break down; And set a failure code to each fault mode and normal mode;

Fault mode, normal mode, failure code and fault eigenvalue and fault signature fuzzy set are become a table as one group of data rows, if fault signature fuzzy set is enough to isolate all faults, namely set up the fault dictionary being used for Switched-Current Circuit failure modes by existing information.

In step 1, pseudo random signal is 255 pseudo-random sequences that employing 8 rank linear feedback shift register produces.

In step 2, Switched-Current Circuit specialty simulation software ASIZ emulation is adopted to carry out sensitivity analysis with localizing faults element to circuit.

In step 3, carry out time-domain analysis and 30 Monte Carlo Analysis to various fault mode and normal condition, sample to failure response signal with the sample frequency of 250KHZ at the output terminal of circuit, the sampled signal obtained is original response data simultaneously.

Described Switched-Current Circuit is the oval band pass filter circuit in six rank chebyshev low-pass filter circuit or six rank.

Pseudo-random pulse sequence is pulse voltage signal, is produced by shift register, is exported during test by the IO port of microprocessor (as DSP).Pseudo-random sequence series has good randomness and the related function close to white noise, and has confirmability in advance and repeatability.Because Switched-Current Circuit belongs to analog sampled data treatment technology, it is the analogue technique of digital technology.So adopt pseudo-random sequence test better than sinusoidal signal.

For each element, there are the fault of transconductance value higher than nominal value and the fault lower than nominal value; ↓ and ↑ represent the fault that the mutual conductance of some elements is lower than nominal value and the fault higher than nominal value respectively;

Fuzzy interval is transistor transconductance value g _mthe constant interval of the information entropy obtained when range of tolerable variance changes for 5% or 10%.Not a concept with the fuzzy set in fuzzy control. with there being Monte Carlo (Monte-Carla) analytic function during ASIZ software emulation, carry out 30 Monte Carlos (Monte-Carla) during emulation to analyze, analyze each time and can obtain one group of time domain response data, corresponding low frequency and high-frequency information entropy can be calculated, 30 analyses can obtain 30 low frequencies and high-frequency information entropy, these 30 information entropys, in an interval, namely constitute fuzzy interval, i.e. fuzzy set;

Fuzzy set is transistor transconductance value g _mthe constant interval of the information entropy obtained during (below get 5% illustrate) change for 5% or 10% in range of tolerable variance.Such as: Mg1 ↓ fault mutual conductance nominal value is 1.9134, and fault value is 0.9567, and change by tolerance 5%, fault value constant interval is 0.9089-1.0045, and the information entropy constant interval calculated is 4.7353-5.5344, i.e. information entropy fuzzy set.

All faults can be distinguished by fuzzy interval and figure, fuzzy interval does not overlap and just can distinguish (as can be seen from Table 2, first see low-frequency information entropy, can find out that in 13 fault modes, some fault divides be not very clear, the information entropy fuzzy set as Mi1 ↓ fault and Mi1 ↓ fault, Mb ↓ fault and Mk ↑ fault and Mf1 ↑ fault and normal condition is very close.These six kinds of malfunctions need to distinguish further by low-frequency approximation information entropy.

Also by scheming to distinguish, ordinate is high and low frequency information entropy, and they just can not distinguish corresponding failure at same level line.As shown in Figure 6, can find out have 6 faults to divide be not very clear, these six kinds of malfunctions need to distinguish further by high-frequency characteristic, and as seen in Figure 7, it is more clear that each fault signature divides.

Following explanation is done to theoretical foundation of the present invention and technical conceive:

Wavelet transformation basic theories

When wavelet transformation (WT) (Wavelet Transform) is one-and frequency method of localization, have the feature of multiresolution analysis, the frequency resolution of its low frequency part is higher, and temporal resolution is lower.And the temporal resolution of HFS is higher, frequency resolution is lower.Wavelet transformation decomposed signal is general picture part and thin looks part, and morther wavelet ψ (x) is obtained little wave system through flexible and translation process by it, then carries out corresponding conversion process to it.Little wave system can be expressed as:

${\mathrm{\ψ}}_{a,b}\left(x\right)=\frac{1}{\sqrt{a}}\mathrm{\ψ}\left(\frac{x-b}{a}\right)---\left(1\right)$

For signal f (x) and morther wavelet ψ (x), its wavelet transformation can be expressed as:

$c(a,b)=<f\left(x\right),{\mathrm{\&psi;}}_{a,b}\left(x\right)>=\frac{1}{\sqrt{a}}{\&Integral;}_{-\&infin;}^{+\&infin;}f\left(x\right)\mathrm{\&psi;}\left(\frac{x-b}{a}\right)\mathrm{dx}---\left(2\right)$

In formula: a, b are respectively the flexible of morther wavelet and translation parameters, the wavelet coefficient that c (a, b) is signal.

In fact, wavelet multi_resolution analysis is carried out to signal f (x), if { V _j} _{j ∈ Z}for orthogonal multiresolution analysis, { W _j} _{j ∈ Z}for the wavelet space that it decomposes, then signal f (x) is at V _jon rectangular projection be Pv _jf=Pv _j+1f+Pw _j+1f, that is:

$P_{V_j}f=\underset{k\&Element;Z}{\mathrm{\&Sigma;}}{c}_{j+1}^k{\mathrm{\&phi;}}_{j+1,k}+\underset{k\&Element;Z}{\mathrm{\&Sigma;}}{d}_{j+1}^k{\mathrm{\&psi;}}_{j+1,k}---\left(3\right)$

In formula, for f (x) is 2 ^j+1scale coefficient under resolution, for f (x) is 2 ^j+1wavelet coefficient under resolution, that is, c _j+1and d _j+1that f (x) is 2 ^j+1general picture part under resolution and thin looks part.So, { V _j} _{j ∈ Z}to spatial decomposition below be carried out:

$\begin{array}{c}{V}_j=W_{j+1}\&CirclePlus;V_{j+1}\\ =W_{j+1}\&CirclePlus;(W_{j+2}\&CirclePlus;V_{j+2})\\ =W_{j+1}\&CirclePlus;W_{j+2}\&CirclePlus;(W_{j+3}\&CirclePlus;V_{j+3})\\ =W_{j+1}\&CirclePlus;W_{j+2}\&CirclePlus;W_{j+3}\&CirclePlus;\&CenterDot;\&CenterDot;\&CenterDot;\end{array}---\left(4\right)$

In order to improve time frequency resolution, wavelet transformation can decompose in wavelet analysis further not through the high fdrequency component of segmentation.The block scheme of wavelet decomposition as shown in Figure 1.

The pre-service of Haar small echo orthogonal filter

Wavelet transformation grows up the nearly more than ten years and is applied to rapidly a kind of mathematical tool of the various fields such as data compression, signal transacting and feature selecting, and it is another important breakthrough before more than 100 years after invention Fourier analysis.In numerous orthogonal function, Haar wavelet function is the simplest orthogonal function, and compared with other orthogonal function, it has the feature of simple structure, convenience of calculation.The orthogonal set of Haar wavelet function to be some amplitudes be+1 and-1 square wave, and have value in one section of interval, other interval is zero, and this makes Haar wavelet transformation (HWT) faster than other wavelet transformation.Haar wavelet function represents with ψ (t) usually, and it is defined as follows

$\mathrm{\&psi;}\left(t\right)=\left\{\begin{array}{cc}1& \mathrm{for}0<t<1/2\\ -1& \mathrm{for}1/2<t<1\\ 0& \mathrm{otherwise}\end{array}\right.$

Haar wavelet basis function φ (t) is the collection of functions of a set of segmentation constant function composition, and it is defined as:

Haar orthogonal wavelet transformation can be equivalent to the process of an arrangement of mirrors as filtering, namely signal decomposes Hi-pass filter Sum decomposition low-pass filter by one, the high frequency components of Hi-pass filter output corresponding signal, i.e. detailed information, low-pass filter exports the relatively low frequency component part of original signal, i.e. approximate information.This filtering decomposition algorithm utilizes down-sampled method namely in 2 that export, only to get a data point, produces the sequence that two are original signal data length half, is designated as the generic term that CA and CD. high and low frequency is wavelet decomposition field.

Orthogonal filter adopts the little filter bank of filters of Haar.Relation between two multichannel analysis and the input and output of synthesis filter banks can be described as:

$\left[\begin{array}{c}{Y}_0\left(z\right)\\ Y{\left(z\right)}_1\end{array}\right]=\frac{1}{2}\left[\begin{array}{cc}{H}_0\left(z^{\frac{1}{2}}\right)& H_0\left({-z}^{\frac{1}{2}}\right)\\ H_1\left(z^{\frac{1}{2}}\right)& H_1\left(z^{\frac{1}{2}}\right)\end{array}\right]\left[\begin{array}{c}X\left(z^{\frac{1}{2}}\right)\\ X\left({-z}^{\frac{1}{2}}\right)\end{array}\right]$

$\begin{array}{c}\hat{X}\left(z\right)=\left[\begin{array}{cc}{G}_0\left(z\right)& G_1\left(z\right)\end{array}\right]\left[\begin{array}{c}{Y}_0\left(z^2\right)\\ Y_1\left(z^2\right)\end{array}\right]\\ =\left[\begin{array}{cc}{G}_0\left(z\right)& G_1\left(z\right)\end{array}\right]\frac{1}{2}\left[\begin{array}{cc}{H}_0\left(z^{\frac{1}{2}}\right)& H_0\left({-z}^{\frac{1}{2}}\right)\\ H_1\left(z^{\frac{1}{2}}\right)& H_1\left(z^{\frac{1}{2}}\right)\end{array}\right]\left[\begin{array}{c}X\left(z\right)\\ X(-z)\end{array}\right]\\ =\frac{1}{2}[G_0\left(z\right)H_0\left(z\right)+G_1\left(z\right)H_1\left(z\right)]X\left(z\right)\\ +\frac{1}{2}[G_0\left(z\right)H_0(-z)+G_1\left(z\right)H_1(-z)]X(-z)\\ =T\left(z\right)X\left(z\right)+\hat{T}\left(z\right)X(-z)\end{array}$

Wherein:

$\left\{\begin{array}{c}T\left(z\right)=\frac{1}{2}[G_0\left(z\right)H_0\left(z\right)+G_1\left(z\right)H_1\left(z\right)]\\ \hat{T}\left(z\right)=\frac{1}{2}[G_0\left(z\right)H_0(-z)+G_1\left(z\right)H_1(-z)]\end{array}\right.$

Independent component analysis

Independent component analysis (ICA) is a kind of method finding internal factor or composition from polynary (multidimensional) statistics, and what it was found is not only statistical iteration but also the composition of non-Gaussian system.The very directly perceived and important principle that ICA estimates is very big non-Gaussian system ^[23].Its thinking is, according to central limit theorem, non-gaussian random variables sum than former variable closer to gaussian variable.For ICA process, if hybrid variable is the linear hybrid of multiple independent source variable, thus, compared with each independent source variable, hybrid variable and Gaussian distribution more close, i.e. Gaussian strong (non-Gaussian system is weak in other words) of hybrid variable more each independent source variable.Therefore, by measuring the non-Gaussian system (or Gaussian) of separating resulting, the independence between separating resulting can be monitored.If the non-Gaussian system of separating resulting maximum (or Gaussian minimum), illustrates the separation completed isolated component.In estimating at ICA, adopt non-Gaussian system, need a quantitative decision rule to measure the non-Gaussian system of stochastic variable, this decision rule has kurtosis (kurtosis), negentropy (negentropy) etc.

Non-Gaussian system is measured by kurtosis

In estimating at ICA, use non-Gaussian system, be a kind ofly called that the Fourth amount of kurtosis has because of it some useful qualitys that fourth central square do not have, make it in ICA algorithm, obtain application.Kurtosis is the another kind of call of the fourth order cumulant of stochastic variable, and it statistically can indicate the most simply measuring of a stochastic variable non-Gaussian system.Can illustrate, if x has Gaussian distribution, its kurtosis is zero.The kurtosis kurt (x) of x may be defined as:

kurt(x)＝E{x ⁴}-3[E{x ²}] ²(5)

Non-Gaussian system can be measured by the absolute value of kurtosis usually, and kurtosis or its absolute value are used as the tolerance of non-Gaussian system widely at ICA and association area, this mainly because of its from calculating or all very simple in theory.From the angle calculated, kurtosis can use the Fourth-order moment of sample data to estimate simply (if the variance of sample data remains unchanged).Theoretical analysis also becomes simple because of linear feature below: if x ₁and x ₂be two independently stochastic variables, then two formula perseverances are set up below:

kurt(x ₁+x ₂)＝kurt(x ₁)+kurt(x ₂) (7)

kurt(αx ₁)＝α ⁴kurt(x ₁) (8)

In formula, α is constant.

Non-Gaussian system is measured by negentropy

In actual applications, because the value of kurtosis can only be estimated from measurement sample, kurtosis may be extremely responsive to outlier (outliers), makes kurtosis method also there are some shortcomings.Therefore using negentropy as second of non-Gaussian system important tolerance, negentropy is all contrary with kurtosis feature in a lot, its robustness well but calculation of complex.

The concept of negentropy is from this information theory parameter of differential entropy, and differential entropy is called entropy by simply here.Entropy is information-theoretical key concept, and a density is p _x(η) stochastic variable, its differential entropy is defined as:

H(x)＝-∫p _x(η)log p _x(η)dη (9)

In order to derive rational non-Gaussian system tolerance, be the value of non-negative, and be zero to its value of gaussian variable, can utilize a kind of amount being called negentropy, it is actually a kind of normalized version of differential entropy.Negentropy J is defined as follows:

J(x)＝H(x _gauss)-H(x) (10)

Wherein x _gaussit is the Gaussian random vector with x with identical relevant (and covariance) matrix.

A useful thinking is promoted Higher Order Cumulants approximation method, uses the expectation of the non-quadratic function of general type, can by polynomial function x ³and x ⁴replace the general function G becoming other ⁱ(i is label but not power here), this method gives based on expectation E{G ⁱ(x) } carry out the reduced form of approximate negentropy.As a simple special case, any two non-quadratic function G can be used ¹and G ²as long as make G ¹be odd function and G ²be even function, estimated as follows:

J(x)≈k ₁(E{G ¹(x)}) ²+k ₂(E{G ²(x)}-E{G ²(v)}) ²(11)

In formula, k ₁and k ₂be normal number, v is the gaussian variable of zero mean unit variance (i.e. standardization).Variable x is also standardized (having zero mean unit variance).Note, even if approximate at this is not that very accurately in situation, formula (11) still may be used for the tolerance building non-Gaussian system.

After collecting original time domain response data at Switched-Current Circuit output terminal, through the pre-service of Haar small echo orthogonal filter, select two function G ¹and G ², according to formula (11), the negentropy of signal after pre-service can being calculated, in order to measure bimodal/openness, selecting Laplce's logarithmic function density ^[1]

G ²(x)＝|x| (12)

In order to measure skew-symmetry, use function G below ¹:

G ¹(x)＝xexp(-x ²/2) (13)

According to formula (8), obtain information entropy:

$J\left(x\right)=k_1{\left(E\right\{\mathrm{x\; exp}(-x^2/2)\left\}\right)}^2+k_2{\left(E\right\{\left|x\right|\}-\sqrt{2/\mathrm{\&pi;}})}^2---\left(14\right)$

In formula, and k ₂=1/ (2-6/ π)

We obtain the negentropy approximate evaluation that can obtain very well compromise between kurtosis and the classical non-Gaussian system tolerance of negentropy two like this, and its concept is simple, and calculated amount is little, and has good statistical property, particularly robustness.

Method explanation

First adopt linear feedback shift register (LFSR) generating period pseudo-random sequence, produce 255 pseudo random sequence length by 8 rank LFSR, obtain band limited white noise test and excitation.Then failure definition pattern, Acquisition Circuit original response data, utilize Haar small echo orthogonal filter as the pretreatment system of acquisition sequence, realize a road input two-way and export, obtain low-frequency approximation information and the detail of the high frequency of observation signal.Finally calculate corresponding information entropy and fuzzy set thereof, extract optimum fault signature, build fault dictionary, this dictionary is for completing the Accurate classification of the fault of each fault mode.Its Troubleshooting Flowchart as shown in Figure 3.

Step 1 is for producing pseudo random testing pumping signal.

Compare the advantage of sinusoidal signal excitation to embody pseudo random testing excitation, the present invention gives 255 pseudo random signal excitations and the soft fault class fault dictionary under sinusoidal signal excitation, as shown in table 4.Still adopt failure classes identical in table 3, can find out, compared with sinusoidal signal excitation, pseudo random signal test can reach a high failure modes rate.

Beneficial effect:

On-off circuit Fault Classification based on wavelet transformation and ICA feature extraction of the present invention, adopt pseudo random signal excitation through Monte Carlo Analysis, Haar small echo orthogonal filter decomposes and the calculating of information entropy, kurtosis and fuzzy set realizes the acquisition of Switched-Current Circuit fault dictionary, thus the classification of complete paired fault, core of the present invention is to adopt linear feedback shift register (LFSR) generating period pseudo-random sequence, obtains band limited white noise test and excitation.Utilize Haar small echo orthogonal filter to decompose, obtain low-frequency approximation information and the detail of the high frequency of original response data.Calculate corresponding information entropy, kurtosis and fuzzy set thereof, extract optimum fault signature, build fault dictionary.

Embodiment part has carried out Simulation experiments validate to six rank chebyshev low-pass filters and to the oval bandpass filter in 6 rank, describe and adopt pseudo random signal excitation more can obtain result accurately compared to sinusoidal signal excitation, and therefore obtain the high failure modes accuracy rate based on information entropy, compare with other method, the accuracy that experimental result shows the inventive method localizing faults is higher, possesses significant superiority.

It is key of the present invention that Haar small echo orthogonal filter decomposes, because define high-frequency information and low-frequency information, thus provide great convenience for follow-up precise classification, otherwise depend merely on a kind of data and classify, its accuracy will be had a greatly reduced quality.

Generally speaking, the present invention obtains the negentropy approximate evaluation of very well compromise between kurtosis and the classical non-Gaussian system tolerance of negentropy two, its concept is simple, calculated amount is little, and has good statistical property, particularly robustness, in addition, form final fault dictionary in conjunction with wavelet decomposition, kurtosis, fuzzy set, various fault can accurately be distinguished, relative to existing Fault Classification, there is obvious advantage.

The meaning of fault dictionary is, based on fault dictionary, fault diagnosis can be carried out to circuit under test further by the sorter such as neural network or support vector machine, thus distinguish fault fault more accurately, if not with reference to fault dictionary, there is comparatively big error in the possibility of result that sorter exports, because, if certain 2 fault cannot well be made a distinction by fault dictionary, so the structure of sorter output is if a fault, b fault may be actually, and had fault dictionary, if a and b fault is not well distinguished, then need to use another sorter to do further differentiation to a and b, thus realize diagnosing accurately, and fault dictionary just serves the effect splitting respective fault in diagnostic procedure.

Embodiment

Below with reference to the drawings and specific embodiments, the present invention is described in further details:

As shown in Figure 2, first adopt linear feedback shift register (LFSR) to produce periodically pseudo-random sequence, choose reasonable pseudo random sequence length, obtain band limited white noise test and excitation.Then failure definition pattern, carry out fault simulation, Acquisition Circuit original response data, utilize Haar small echo orthogonal filter as the pretreatment system of acquisition sequence, obtain low-frequency approximation information and the detail of the high frequency of original response data, realize a road input two-way and export.Next carry out the extraction of ICA fault signature, respectively its differential (bearing) entropy and kurtosis and fuzzy set thereof are calculated to high and low frequency two-way output signal, obtain optimum fault signature.These optimum fault signatures are applied to the fault diagnosis that neural network failure sorter carries out Switched-Current Circuit the most at last, correctly and effectively identify to realize fault element.

The present invention considers two typical switching current diagnostic circuits, i.e. the oval band pass filter circuit of six rank chebyshev low-pass filter circuit and six rank.Wherein, low-pass filter considers soft fault and hard fault situation, and soft fault also contemplates muting sensitivity failed transistor situation.And soft fault situation is only considered for band pass filter circuit.Apply test and excitation to respectively these two filter circuits, test and excitation signal adopts 255 pseudo-random sequence signal produced by 8 rank linear feedback shift registers, the circuit that pseudo-random sequence can strengthen normal circuit mode and fault mode exports the difference responded, and is easy to the location realizing fault.And be easy to produce high-quality test id signal, decrease testing cost.Under pseudo-random sequence excitation, carry out sampling at two circuit output ends and can obtain respective fault time domain response signal.

For the classification of diagnostic circuit 1-six rank chebyshev low-pass filter circuit

Be six rank chebyshev low-pass filter circuit shown in Fig. 3, in figure, give also the normalization transconductance value of MOS transistor.Circuit cutoff frequency is 5MHz, and cutoff frequency is 1: 4 with the ratio of clock frequency, and clock frequency is 20MHz, ripple 0.5dB in band.In filter circuit, MOS transistor mutual conductance g _mrange of tolerable variance be 5% or 10% respectively.When simulating circuit applying test and excitation signal, suppose that in circuit, transistor transconductance value changes in respective range of tolerable variance, then indication circuit is normal condition (NF).

Utilize ASIZ switching current specialty simulation software to carry out sensitivity analysis to this circuit, analysis result shows Mg1, and the change of Mf1, Mi1, Mb, Mh and Mk value exports response impact comparatively greatly to circuit, therefore selects these 6 transistors to carry out Analysis on Fault Diagnosis.When its nominal value 50% of any one downward shift of these 6 transistors, and when other five transistors change in its range of tolerable variance, carry out ASIZ emulation to circuit and can obtain failure response, now circuit there occurs soft fault.Can obtain like this 13 fault categories be respectively Mg1 ↑, Mg1 ↓, Mf1 ↑, Mf1 ↓, Mi1 ↑, Mi1 ↓, Mb ↑, Mb ↓, Mh ↑, Mh ↓, Mk ↑, Mk ↓ and NF.Here ↑ and ↓ mean that fault value is higher or lower than nominal g _mthe malfunction corresponding to 50% of value, table 1 gives the nominal value of six rank chebyshev low-pass filter muting sensitivity transistors and fault value with the fault category of correspondence and failure code.

The soft fault classification of table 1 low-pass filter muting sensitivity transistor and respective value thereof

In order to compared with document [1], invention also contemplates that fault category identical in six rank chebyshev low-pass filter diagnosis examples in document [1], when all hypothesis breaks down, transistor transconductance value offset by 50%, has 11 kinds of malfunctions as shown in table 2.

The soft fault classification of table 2 six rank low-pass filter transistor and respective value

Hard fault situation: hard fault and bust, six kinds of busts are grid source short trouble (GSS), drain-source short trouble (DSS), source electrode open fault (SOP) open-drain fault (DOP), grid leak short trouble (GDS) and open-grid fault (GOP) respectively.Above-mentioned six kinds of busts can be simulated in the following ways: during simulation grid source short trouble, a small resistor can be added between grid and source electrode; Then large resistance etc. is added in source terminal during analog source electrode open fault.When in Fig. 3, Mb and Mk creates bust, these bust time domain response signals extract its fault signature after pretreater, form 13 kinds of fault modes as shown in table 3.

Failure modes is carried out for the oval band pass filter circuit in diagnostic circuit 2-six rank

Be the six oval band pass filter circuits in rank shown in Fig. 4, its transistor transconductance value as shown in FIG..The tolerance of mutual conductance is 5% or 10%, and centre frequency is 1KHz.Table 4 gives the soft fault classification of circuit and the nominal value of transistor and fault value prescribed situation.

The hard fault classification of table 3 six rank low-pass filter transistor and respective value

The oval bandpass filter transistor rating in table 4 six rank and fault value and soft fault classification

The pre-service of Haar small echo orthogonal filter

In order to more effectively extract fault signature, the pre-service of Haar small echo orthogonal filter is carried out to failure response signal.This process can carry out meticulousr decomposition to high-frequency information, and this decomposition is irredundant, nothing careless omission, can bring line frequency and temporal resolution analysis into simultaneously, improve time frequency resolution at the high and low frequency of signal.

Haar orthogonal wavelet transformation can be compared to an arrangement of mirrors as filtering, filtering is carried out through two groups of wave filters by signal to be analyzed, its course of work is: signal is input to one respectively and decomposes in Hi-pass filter Sum decomposition low-pass filter, the high frequency components (detailed information) of signal exports from Hi-pass filter, and the low frequency component part (approximate information) of signal exports from low-pass filter.Approximate signal and the detail signal of signal can be obtained.The process flow diagram of what Fig. 5 represented is two multichannel analysis and synthesis filter banks.As can be seen from Figure, H ₀z () is low-pass filter, H ₁z () is Hi-pass filter.Orthogonal filter in Fig. 5 adopts the little filter bank of filters of Haar.For oval bandpass filter soft fault diagnosis (table 4) in six rank, Haar small echo orthogonal filter preprocessing process is described, detailed decomposable process is as follows: implement time-domain analysis and 30 Monte Carlos (Monte-Carla) analyses to oval band pass filter circuit 15 kinds of malfunctions, getting sample frequency during analysis is 100KHZ, obtains the failure response signal with 158 sampled points.That is, often kind of fault category can obtain 30 time domain failure response samples, and each sample packages is containing 158 sampled points.Then the pre-service of Haar small echo orthogonal filter is implemented to these 30 sample signals, obtain low-frequency approximation information and the detail of the high frequency of original response data, realize a road input two-way and export.So for often kind of fault category, its time domain failure response feature has 30 samples, and each sample packages is containing 2 attributes (low-frequency approximation information and detail of the high frequency).15 kinds of fault categories constitute 900 time domain response samples altogether altogether.

ICA fault signature extracts

The Independent Component Analysis proposed according to Section 3 and very big non-Gaussian system decision rule, carry out the extraction of ICA fault signature to above two Switched-Current Filter circuit.

ICA refers to the english abbreviation of independent component analysis (Independent Component Analysis).

Fault signature represents kurtosis and (bearing) entropy two non-Gaussian system characteristic parameters.

Under MATLAB environment, calculate the low-frequency approximation kurtosis of various fault category and (bear) entropy and high frequency detail kurtosis and (bearing) entropy, obtain the Fault characteristic parameters of various fault category like this.Next, the extraction of ICA fault signature is carried out to the time domain failure response sample after 30 Monte Carlo Analysis, obtain often kind of corresponding fault signature fuzzy set of fault category.

The following data of concrete extraction: low-frequency approximation entropy, low-frequency approximation kurtosis, low-frequency approximation entropy fuzzy set, low-frequency approximation kurtosis fuzzy set, high frequency detail entropy, high frequency detail kurtosis, high frequency detail entropy fuzzy set, high frequency detail kurtosis fuzzy set (i.e. each column data of the first row in table 5)

Non-Gaussian system is measured with kurtosis and (bearing) entropy.Under MATLAB environment, calculate the low-frequency approximation kurtosis of various fault category and (bear) entropy and high frequency detail kurtosis and (bearing) entropy, obtain the Fault characteristic parameters of various fault category like this.Next, the extraction of ICA fault signature is carried out to the time domain failure response sample after 30 Monte Carlo Analysis, obtain often kind of corresponding fault signature fuzzy set of fault category.

First the soft fault feature of six rank chebyshev low-pass filters is extracted.Table 5 gives six rank chebyshev low-pass filter muting sensitivity transistor soft fault class fault signatures.First the characteristics of low-frequency distribution plan of 13 kinds of fault categories can be obtained, as shown in Figure 6 according to table 5.As can be seen from Figure 6, there occurs more serious classification between F3, F6, F10 fault category and between F4, F9, F13 fault category overlapping.Except these two groups of fault categories, other each fault category all obtains reasonable separation.Fig. 7 is the high-frequency characteristic distribution plan of above two groups of fault categories (6 failure classes), and in the figure 7,6 fault categories are obtained for effective differentiation.

Table 5 six rank chebyshev low-pass filter muting sensitivity transistor soft fault class fault signature

In order to the convenience compared with document [1], the present invention selects fault category (as shown in table 2) identical with it.Table 6 gives 6 rank chebyshev low-pass filter soft fault class fault signatures, can obtain six rank chebyshev low-pass filters, 11 kinds of soft fault classification characteristics of low-frequency distribution plan (Fig. 8) and 4 kinds of soft fault classification high-frequency characteristics distribution plan (Fig. 9) respectively by table 6.In fig. 8, there are two groups of fault categories (F2 and F11, F1 and F5) to occur larger overlap in 11 kinds of fault categories, need to distinguish further by high-frequency characteristic distribution plan.And in fig .9, these two groups of fault categories obtain good differentiation.

Then, the hard fault classification of six rank chebyshev low-pass filters shown in his-and-hers watches 3 carries out the extraction of ICA fault signature.

Finally, the extraction of ICA fault signature is carried out to the oval band pass filter circuit in second diagnostic circuit i.e. six rank.Table 8 gives the soft fault class fault signature of the oval bandpass filter in six rank.Figure 10-Figure 11 sets forth this circuit 15 kinds of soft fault classification characteristics of low-frequency distribution plans and 4 kinds of soft fault classification high-frequency characteristic distribution plans.In Fig. 10, between fault category F6 and F8 and between F1 and F4, more serious overlapping phenomenon all occurs, and in fig. 11, these two groups of fault categories have had further separately, obtain and are separated completely.

Table 6 six rank chebyshev low-pass filter transistor soft fault class fault signature

Table 7 six rank chebyshev low-pass filter transistor hard fault class fault signature

Table 8 six rank oval bandpass filter transistor soft fault class fault signature

Diagnostic result is analyzed

The BP neural network determined according to the fault category low frequency shown in Fig. 6-Figure 11 and high-frequency characteristic distribution plan and formula (15) carries out fault diagnosis to switch current filter circuit.In order to the diagnostic result that the present invention is directed to six rank chebyshev low-pass filters is carried out comparing of diagnosis efficiency with the diagnostic result for this circuit in document [3], document [1] and document [2] by the dominance embodying the inventive method other literature method relative, as shown in table 9.

In document [1], hard fault class number is 9, and define GSS, GDS, SOP and DOP tetra-kinds of hard fault types, hard fault diagnosis efficiency is 100%, shows to distinguish all hard fault classifications.Soft fault class number is 11, due to the sinusoidal signal that test and excitation adopts, and further optimization is not extracted to fault signature, cause diagnosis effect not to be desirable especially, soft fault diagnosis efficiency only has about 95%, can not distinguish Mg1 ↓, Mi ↑, Mg1 ↑ and Mi ↓ tetra-kind of soft fault classification.And in the methods of the invention, hard fault class number is 13, add GSS and GOP two kinds of hard fault types, can distinguish all hard faults equally, diagnosis efficiency is 100%.Soft fault classification is identical with document [1], and soft fault diagnosis efficiency reaches 100%, and successfully all soft fault states have been distinguished in diagnosis.

The various method for diagnosing faults of table 9 six rank chebyshev low-pass filter compare

In addition, document [3] adopts Wavelet Neural Network Method to carry out fault test to six rank chebyshev low-pass filters.The method, for GSS, GDS, SOP, DOP, GSS and GOP six kinds of hard fault type tests, reaches the rate of correct diagnosis of 100%.But when in circuit during muting sensitivity transistor generation soft fault, because sensitivity is low, its failure response is most and normal condition is close, causing trouble conductively-closed and the diagnosis effect that can not reach soft fault, therefore the muting sensitivity transistor soft fault diagnosis efficiency of document [3] method is only 80%.But in the methods of the invention, soft fault class number is 13, soft fault is test for the muting sensitivity transistor can not correctly distinguished in document [3] method equally.The inventive method also reaches the correct classification rate of 100% to muting sensitivity transistor soft fault classification.

Finally, in document [2], the method is not diagnosed for hard fault type.And in soft fault diagnosis, although document [1] relatively, diagnosis efficiency increases, by four kinds of soft fault classifications undistinguishable in document [1] (Mg1 ↓, Mi ↑, Mg1 ↑ and Mi ↓) successfully distinguished three kinds, but also have Mg1 ↑ fault can not correctly distinguish.And the inventive method successfully can distinguish all soft fault classifications.

The present invention has also carried out soft fault diagnosis for the oval bandpass filter in six rank, and reach good diagnosis effect equally, rate of correct diagnosis reaches 10) 0%.

5 conclusions

Diagnostic result of the present invention shows, the fault diagnosis system based on wavelet transformation and ICA feature extraction that the present invention proposes can implement the fault diagnosis of Switched-Current Circuit effectively, two kinds of filter circuits are achieved to the fault diagnosis accuracy of almost 100%, and the method that the inventive method and other literature method adopt has been done comparative analysis, demonstrate the present invention adopt the superiority of method, diagnosis effect is satisfactory.Wavelet transformation can extract fault feature effectively, utilizes Haar small echo orthogonal filter as the pretreatment system of acquisition sequence, realizes a road input two-way and exports, obtain low-frequency approximation information and the detail of the high frequency of observation signal.Next carry out the extraction of ICA fault signature, respectively its (bearing) entropy and kurtosis and fuzzy set thereof are calculated to high and low frequency two-way output signal, obtain optimum fault signature.By to the soft fault of the oval bandpass filter in six rank chebyshev low-pass filters and six rank and the hard fault simulation results show high efficiency of the method, it is a kind of diagnostic method for failure of switch current circuit of high diagnosis efficiency.

Accompanying drawing explanation

Fig. 1 is small echo (wavelet packet) decomposing schematic representation;

Fig. 2 is process flow diagram of the present invention.

Fig. 3 is six rank chebyshev low-pass filter circuit theory diagrams;

Fig. 4 is the structural drawing of the oval bandpass filter in 6 rank;

Fig. 5 is two multichannel analysis and synthesis filter banks block diagram;

Fig. 6 is low-pass filter muting sensitivity transistor 13 kinds of soft fault classification characteristics of low-frequency distribution plans;

Fig. 7 is low-pass filter muting sensitivity transistor 6 kinds of soft fault classification high-frequency characteristic distribution plans;

Fig. 8 is six rank chebyshev low-pass filters, 11 kinds of soft fault classification characteristics of low-frequency distribution plans;

Fig. 9 is six rank chebyshev low-pass filters, 4 kinds of soft fault classification high-frequency characteristic distribution plans;

Figure 10 is the oval bandpass filter in six rank 15 kinds of soft fault classification characteristics of low-frequency distribution plans;

Figure 11 is the oval bandpass filter in six rank 4 kinds of soft fault classification high-frequency characteristic distribution plans.

Claims (7)

1., based on an on-off circuit Fault Classification for wavelet transformation and ICA feature extraction, the method is used for the failure modes of Switched-Current Circuit, it is characterized in that, comprises the following steps:

Step 1: produce pseudo random signal as test and excitation signal;

Pseudo random signal is pseudo-random pulse sequence;

Step 2: failure definition pattern:

Based on circuit simulation, carry out sensitivity analysis to Switched-Current Circuit to be measured, the change obtaining component parameters changes the single order of electric network system features, with the fault element most possibly broken down in positioning circuit; And divide fault mode based on fault element location; The quantity of fault element is N, then the kind of fault mode is 2*N; N is natural number;

Step 3: the original response data of Acquisition Circuit:

Encourage tested Switched-Current Circuit by pseudo random signal, with ASIZ software, the various malfunction of tested Switched-Current Circuit and normal condition are emulated, collect original response data from the output terminal of Switched-Current Circuit; These original response data are curtage data;

Step 4: adopt Haar small echo orthogonal filter to carry out pre-service to original response data;

Utilize Haar small echo orthogonal filter as the pretreatment system of acquisition sequence, obtain low-frequency approximation information and the detail of the high frequency of observation signal;

Step 5: Fault characteristic parameters extracts;

Entropy and the kurtosis of low-frequency approximation information and detail of the high frequency is calculated respectively for pretreated signal; Obtain following Fault characteristic parameters: low-frequency approximation entropy, low-frequency approximation kurtosis, low-frequency approximation entropy fuzzy set, low-frequency approximation kurtosis fuzzy set, high frequency detail entropy, high frequency detail kurtosis, high frequency detail entropy fuzzy set and high frequency detail kurtosis fuzzy set;

Step 6: based on the Fault characteristic parameters structure fault dictionary extracted, thus realize on-off circuit failure modes.

2. the on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction according to claim 1, it is characterized in that, the computing method of described information entropy are:

Information entropy $J\left(x\right)=k_1{\left(E\right\{\mathrm{xexp}(-x^2/2)\left\}\right)}^2+k_2{\left(E\right\{\left|x\right|\}-\sqrt{2/\mathrm{\&pi;}})}^2;$ In formula, $k_1=36/(8\sqrt{3}-9)$ And k ₂=1/ (2-6/ π), x are the data that the primary current response data of the circuit-under-test output terminal extracted obtains through wavelet transformation; E represents expectation value;

The computing method of described kurtosis are:

Kurtosis kurt (x)=E{x ⁴-3 [E{x ²] ², x is the data that the primary current response data of the circuit-under-test output terminal extracted obtains through wavelet transformation, and E represents expectation value.

3. the on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction according to claim 2, it is characterized in that, fuzzy set is transistor transconductance value g _mthe information entropy obtained when range of tolerable variance changes for 5% or 10% or the constant interval of kurtosis;

Low-frequency approximation information entropy fuzzy set and detail of the high frequency entropy fuzzy set are a numerical intervals; Normal mode refers to the pattern that circuit does not break down; And set a failure code to each fault mode and normal mode;

Fault mode, normal mode, failure code and fault eigenvalue and fault signature fuzzy set are become a table as one group of data rows, if fault signature fuzzy set is enough to isolate all faults, namely set up the fault dictionary being used for Switched-Current Circuit failure modes by existing information.

4. the on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction according to claim 2, is characterized in that, in step 1, pseudo random signal is 255 pseudo-random sequences that employing 8 rank linear feedback shift register produces.

5. the on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction according to claim 4, it is characterized in that, in step 2, Switched-Current Circuit specialty simulation software A S I Z emulation is adopted to carry out sensitivity analysis with localizing faults element to circuit.

6. the on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction according to claim 5, it is characterized in that, in step 3, time-domain analysis and 30 Monte Carlo Analysis are carried out to various fault mode and normal condition, sample to failure response signal with the sample frequency of 250KHZ at the output terminal of circuit, the sampled signal obtained is original response data simultaneously.

7. the on-off circuit Fault Classification based on wavelet transformation and ICA feature extraction according to any one of claim 1-6, it is characterized in that, described Switched-Current Circuit is the oval band pass filter circuit in six rank chebyshev low-pass filter circuit or six rank.