CN103728551A - Analog circuit fault diagnosis method based on cascade connection integrated classifier - Google Patents

Abstract

The invention discloses an analog circuit fault diagnosis method and an implementation method of the analog circuit fault diagnosis method. The content includes the first part of analog circuit fault feature information extraction, the second part of fault classifier construction, and the third part of implementation of algorithm software. The analog circuit fault diagnosis method includes the following steps of constructing a fault feature information base, selecting an optimal mother wavelet through an information entropy maximizing principle, conducting wavelet decomposition on response nodes of a measured circuit, extracting the optimal feature of the measured circuit, conducting dimensionality reduction on the fault features through principal component analysis, conducting fault classification and intelligent diagnosis, constructing a fault diagnosis device according to the obtained fault feature information and through a multi-classifier cascade connection model and the classifier integration technology so as to recognize existing faults and causes of the faults, and conducting specific implementation on the algorithm through a C#.NET platform and through combination with the Weka software. The diagnosis method and the implementation method have the advantages of being high in fault diagnosis performance, wider in diagnosis range, higher in algorithm robustness and higher in interpretability.

Description

A kind of analog-circuit fault diagnosis method based on cascade integrated classifier

Technical field

The present invention relates to a kind of method for diagnosing faults and implementation thereof of mimic channel.

Background technology

The fault diagnosis of mimic channel starts from the sixties in 20th century, to its theoretical research from network element parameter solvability, but because the difficulty of its uniqueness is as complicacy of the tolerance of malfunction diversity, component parameters, information deficiency and structural model etc., make for the research and development of the fault diagnosis of mimic channel relatively slowly, its test and fault diagnosis all become a difficult problem for puzzlement circuit test industry all the time.After the nineties in 20th century, development along with artificial intelligence technology, fuzzy theory, wavelet technique and some machine learning methods all sequential use in this field and obtained good effect, but it all exists one-sidedness, to solving actual analog circuit fault diagnosing and problem analysis, all also more or less there is a certain distance.Meanwhile, the actual demand of analog circuit fault diagnosing but constantly increases.Therefore, study a kind ofly to analog circuit board fault detect accurately and rapidly and Fault Locating Method, shorten to detect maintenance time and reduce maintenance cost, for the guarantee maintenance that completes analog circuit board in electronic equipment, be significant.

Summary of the invention

The present invention discloses a kind of method for diagnosing faults and implementation thereof of mimic channel, comprising: fault signature extraction, the Classification and Identification of fault and the software of algorithm of signal are realized three aspects:.The method is comprised of following steps: (1) structure fault characteristic information storehouse, according to circuit-under-test signal characteristic, employing information maximum entropy principle (MEP), choose optimum female small echo, the responsive node of circuit-under-test is carried out to wavelet decomposition, the optimal characteristics of extraction circuit-under-test, then utilizes principal component analysis (PCA) (PCA) thereby to every layer, carrying out dimensionality reduction obtains fault characteristic information.(2) fault analysis and intelligent diagnostics, according to the fault characteristic information parameter obtaining, utilize multi-categorizer cascade model and integrated (Ensemble) [homomorphism and differential mode] technical construction intelligent trouble diagnosis device to pick out fault and the reason thereof that may exist.A) fault diagnosis device adopts multi-categorizer cascade model, first solve normal sample and initial failure sample and be difficult to the problem of distinguishing, first the sample of the feature samples of normal circuit and all faulty circuits is formed respectively to two disjoint subsets, the integrated technology structure support vector machine classifier that adopts homomorphism, forms level G ₀, for distinguishing normal and malfunction; Secondly to fault sample, the integrated technology of employing differential mode trains the base sorter of algorithms of different, then utilizes Nearest Neighbor with Weighted Voting algorithm to merge sorter, forms level G ₁, for distinguishing different faults state, this two hierarchical structure has just formed polytypic cascade reasoning thought.B) sorter of cascade model structure adopts integrated technology, first for level G ₀, adopt homomorphism integrated technology, utilize the Bagging algorithm of monolateral sampling, the imbalance problem that solves data trains integrated support vector machine classifier.Then for level G ₁, the synthetic sorter that adopts differential mode integrated technology to train based on Bayes, decision tree and algorithm of support vector machine is weighted ballot output to sample, increases the extensive precision of fault diagnosis system.(3) software of algorithm is realized and is adopted the C#.NET of Microsoft platform, the weka.jar file of Weka software project is converted to the weka.dll procedure set that can be called by .NET by IKVM.NET instrument, some class in weka.dll is rewritten, complete and adopt three-tier architecture model to write software to after the specific implementation of algorithm, realize the concrete analysis of fault and diagnosis.

Accompanying drawing explanation

Fig. 1 fault signature extracts process flow diagram

Fig. 2 fault decision flow diagram

Fig. 3 software architecture figure

Embodiment

Analog Circuit Fault Diagnosis Technology based on knowledge is a pattern recognition and classification problem in essence.Therefore, the validity feature that how to extract fault is gordian technique and an important ring of analog circuit fault diagnosing, and the final purpose of simultaneously extracting feature is to test sample book structural classification device, realizes the correct Classification and Identification to different faults kind.Finally to reach such object, complete the true realization to fault diagnosis, must carry out to algorithm the realization of software.

In order to achieve the above object, method of the present invention is achieved in that

1, the optimal wavelet of analog circuit fault characteristic information extracts

The small echo fault characteristic information extracting method of processing as signal is current study hotspot, wavelet analysis belongs to multiresolution analysis, it is a kind of meticulous Time-Frequency Analysis Method, signal is carried out to multilayer decomposition, be conducive to obtain more sampled signal local detail characteristic, yet because dissimilar small echo has different time-frequency characteristics, fault characteristic information for more effective extraction circuit, should make the time-frequency characteristics of small echo and the time-frequency characteristics of Circuit responce node match, therefore, the present invention uses the female small echo system of selection of a kind of optimum based on information maximum entropy principle to solve this problem.Concrete step is as follows, its flow process as shown in Figure 1:

(1) establishing any given node response signal is f (t), according to the definition of wavelet transformation,

$W_f(a,b)=\frac{1}{\sqrt{a}}\&Integral;f\left(t\right){\mathrm{\&psi;}}^{*}\left(\frac{t-b}{a}\right)\mathrm{dt}=<f\left(t\right),{\mathrm{\&psi;}}_{a,b}\left(t\right)>,$ In formula: ψ is wavelet mother function, a is scale parameter, and b is time centre parameter, W _f(a, b) be the coefficient of wavelet decomposition of signal, signal is carried out to wavelet transformation, but transform generally can only carry out approximate treatment by computing machine, in order to reach higher precision and the simplicity of calculating, generally adopt trapezoidal method to do numerical approximation integral operation, get t=n Δ T, b=k Δ T, computing formula is

$W_f(a,k)=\frac{\mathrm{\&Delta;T}}{2\sqrt{a}}\underset{n}{\mathrm{\&Sigma;}}[f\left(n\right)\mathrm{\&psi;}\left(\frac{n-k}{a}\right)+f(n+1)\left(\frac{n+1-k}{a}\right)],$ In formula: Δ T is sampling interval, supposes a=2 ⁱ(i=1,2..., n), for each given a=2 ⁱ, calculate successively the asynchronous value of k, get different wavelet mother functions simultaneously and calculate, one group of wavelet conversion coefficient in the time of can obtaining getting different wavelet mother function.

(2) adopt different wavelet transformations, the information entropy of counting circuit, selects optimum female small echo to respond and carry out wavelet transformation circuit node.If responsive node normal signal is r (t), failure response signal f _i(t) (i=1 ..., c), wherein i is failure mode, and c is fault sum, and the information entropy computing method of circuit are as follows: respectively normal signal r (t) and fault are rung to signal f _i(t) carry out corresponding wavelet transformation, decomposition level is n, gets respectively n layer low frequency Coefficients of Approximation and the 1st ..., n floor height frequently Coefficients of Approximation forms a vector, and establishing normal signal is R (k), and fault-signal is F _i(k), wherein k is signal sampling number of samples, and normal signal R (k) is designated as to F ₀(k) vector, forming after all signal wavelet transformations can be expressed as F _i(k) (i=0,1 ..., c), then calculate respectively the cosine similarity between every two vectors

according to information entropy basic theories, definable circuit information entropy

can select optimum wavelet mother function according to the maximum principle of circuit information entropy.For example through can be calculated impulse response signals, be suitable for converting with Haar small echo.

(3) before carrying out PCA, for fear of the impact of data dimension, every layer coefficients of response signal wavelet transformation is carried out respectively to data normalization, adopt following formula to carry out:

wherein l is n layer wavelet decomposition,

n is the number of every layer of wavelet coefficient.After normalization, every layer of wavelet coefficient merged and form a new vector

all normalized sample vectors are combined into matrix X, set up correlation matrix,

m is number of samples, by R, can obtain eigenvalue λ _iwith proper vector a _i(i=1,2 ..., n), calculate the contribution rate of i pivot to population variance, by contribution rate is descending, arrange, choose successively k pivot and make to accumulate contribution rate sum and be greater than 90%.Foundation afterwards

calculate each required pivot value, form final proper vector sample.

2, the structure of fault grader

The object of fault diagnosis is that test sample book is carried out to Classification and Identification, common way is the different sorting algorithm of design, such as present conventional sorter comprises neural network classifier, support vector machine classifier, Bayes classifier etc., in order to realize recognition performance as well as possible, usually can design different classification schemes, yet be no matter that any sorter effect that different problems is obtained is always not best, thereby conventional way has improving one's methods of various sorters and the technology based on sorter integrated (Ensemble) etc. now, and integrated study utilizes the output of a plurality of base sorters can improve the precision of traditional classifier, obtained good effect.The present invention is the integrated technology based on sorter, adopt a kind of cascade model to carry out constructive inference fault diagnosis sorter, its basic ideas are: first solve normal sample and initial failure sample and be difficult to the problem of distinguishing, first the sample of the feature samples of normal circuit and all faulty circuits is formed respectively to two disjoint subsets, the integrated technology structure support vector machine classifier that adopts homomorphism, forms level G ₀, for distinguishing normal and malfunction; Secondly to fault sample, the integrated technology of employing differential mode trains the base sorter of algorithms of different, then utilizes Nearest Neighbor with Weighted Voting algorithm to merge sorter, forms level G ₁, for distinguishing different faults state, by two such hierarchy Model, formed a kind of diagnostic reasoning thought.Thinking and implementation step that it is concrete are as follows, flow process as shown in Figure 2:

(1) sample of the feature samples of normal circuit and all faulty circuits is divided into two disjoint subset X _nand X _f, establish X _nfor positive class sample instance, X _ffor anti-class sample instance, yet we know according to practical experience, the probability that normal class occurs in data centralization is very large, and the probability that failure classes occur is very little, will cause like this sample size of normal class to want the sample obviously forming more than other failure classes, this data set is called unbalanced data collection.The sorting technique main thought of processing this unbalanced data collection has: (a) use imbalanced class distribution is had to fine adaptive base sorter, and Ensemble Learning Algorithms is constant, (b) using traditional classifier, is that the sorter finally obtaining can adapt to imbalanced class distribution problem by revising Ensemble Learning Algorithms.The second thinking is occupied an leading position at present.Bagging is simple, respond well as a kind of important Ensemble Learning Algorithms implementation method.The present invention adopts a kind of method that is called set of homomorphisms constituent class technology, uses monolateral sampling Bagging Ensemble Learning Algorithms to sample set X _nand X _ftrain, and base sorter uses same classification learning algorithm-support vector machine, it has overcome the deficiency of neural network, shows the features such as simple in structure, global optimum, generalization ability are strong in solving the problems such as small sample, non-linear and higher-dimension pattern-recognition.Monolateral sampling Bagging Ensemble Learning Algorithms based on support vector machine, be performed such training, so take turns and first extract positive class sample instance out at each, from anti-class sample instance, have at random and put back to the example sample extracting with positive class as much again, composing training collection T together with all positive class examples _i, then use base classification learning algorithm-support vector machine from T _iin train base sorter, finally each is taken turns to the base sorter of learning out and merges, form the first level G of our fault diagnosis sorter ₀, can, for the classification of normal and fault, be normally output as 0, and fault be output as 1.

(2) work as G ₀sorter is output as at 1 o'clock, and sample is fault sample, and needing further judgement is any fault, so also just need to construct one or several fault grader again.For the more problem of this kind number, except some traditional sorting algorithms, as neural network etc.A lot of improved algorithms have also been there are at present, such as first based on clustering algorithm, carry out rough sort again to each rough sort structural classification device and a kind of multi-categorizer merge algorithm that is called differential mode integrated technology, use identical data sample, and base sorter adopts different algorithms to train, finally the base sorter obtaining is merged etc.The present invention adopts differential mode Ensemble classifier technology, constructs the second level G ₁sorter, base sorter is selected bayesian algorithm, decision Tree algorithms and algorithm of support vector machine, again various base sorters are assessed, adopt the ballot mode Output rusults of weighting, the classification curve of the integrated classifier of structure will be obviously level and smooth like this, also has the feature of strong robustness simultaneously.

3, a kind of software realization mode of algorithm

Algorithm is the soul of dealing with problems, and the realization of algorithm just makes soul have the human body depending on, and has just had real realistic meaning.Realize algorithm as above, it is a uninteresting and difficult thing, yet fortunately, Weka, as a disclosed data mining workbench, has gathered a large amount of machine learning algorithms that can bear data mining task, comprises data are carried out to pre-service, classification, recurrence, cluster etc., and more valuable, developer can improve the code of increasing income, even utilize the framework of Weka to develop more data mining algorithm.Therefore the present invention is based on C# platform and utilize Weka project software to complete the realization to above-mentioned algorithm, concrete steps are as follows:

(1) software of algorithm is realized and is adopted the C#.NET of Microsoft platform, to under .NET, can call the weka.jar file of Weka software project, need to utilize IKVM.NET instrument weka.jar file to be converted to the weka.dll procedure set that can be called by .NET, only need to carry out ikvmc-target:library weka.jar, and during the project that weka.dll imports .NET is quoted.

(2) the one-dimensional wavelet transform function of realizing in C++ is exported as to the function that can call in C#, the following method of employing:

using?System.Runtime.InteropServices；

[DllImport(“Wavelet1D.dll”，CharSet＝CharSet.Auto)]

public?static?extern?int[]Wavelet1D(string?filename，int?level，string?wname，refint[]length)；

In C#, call afterwards optimum female small echo signal is carried out to wavelet transformation function, obtain after wavelet conversion coefficient it to carry out dimension normalization, then call the principal component analysis (PCA) that PrincipalComponents class in weka realizes sample.Should be noted before using weka PrincipalComponents class needs to use using statement to lead weka.filters.unsupervised.attribute, org.antlr.stringtemplate, org.antlr.stringtemplate.language NameSpace.

(3) realize the rewriting to Bagging class, what realize due to weka.classifiers.meta.Bagging class in weka is that the ground sample mode of putting back to of standard operates training set, and the present invention's employing is monolateral sampling Bagging algorithm, so need the method in its class be rewritten called after SSBagging class.Directly call afterwards LibSVM base sorter and train the first level G ₀sorter, called after SSBaggingClassify class, its basic code is as follows:

(4) constructed the first level G ₀sorter, when it is output as 1, which kind of fault it is to need further judgement, according to above-mentioned thinking, need construct the sorter based on Ensemble technology, need be by the crucial class for classifying in weka

weka.classifiers.functions.LibSVM，weka.classifiers.tress.J48，

Weka.classifiers.bayes.NaiveBayes and for the weka.classifiers.meta.Vote of integrated technology.Its basic code is as follows:

Finally, software adopts three-tier architecture model to realize, and its structural drawing, as Fig. 3, repeats no more.

Claims (1)

1. the present invention relates to a kind of method for diagnosing faults and implementation thereof of mimic channel.Its content mainly comprises that the fault signature of signal extracts, the Classification and Identification of fault and the software of algorithm are realized three aspects:.It is characterized in that the method carries out according to the following steps: (1) structure fault characteristic information storehouse, according to circuit-under-test signal characteristic, employing information maximum entropy principle (MEP), choose optimum female small echo, the responsive node of circuit-under-test is carried out to wavelet decomposition, the optimal characteristics of extraction circuit-under-test, then utilizes principal component analysis (PCA) (PCA) thereby to every layer, carrying out dimensionality reduction obtains fault characteristic information.(2) fault analysis and intelligent diagnostics, according to the fault characteristic information parameter obtaining, utilize multi-categorizer cascade model and integrated (Ensemble) [homomorphism and differential mode] technical construction intelligent trouble diagnosis device to pick out fault and the reason thereof that may exist.A) fault diagnosis device adopts multi-categorizer cascade model, first solve normal sample and initial failure sample and be difficult to the problem of distinguishing, first the sample of the feature samples of normal circuit and all faulty circuits is formed respectively to two disjoint subsets, the integrated technology structure support vector machine classifier that adopts homomorphism, forms level G ₀, for distinguishing normal and malfunction; Secondly to fault sample, the integrated technology of employing differential mode trains the base sorter of algorithms of different, then utilizes Nearest Neighbor with Weighted Voting algorithm to merge sorter, forms level G ₁, for distinguishing different faults state, this two hierarchical structure has just formed polytypic cascade reasoning thought.B) sorter of cascade model structure adopts integrated technology, first for level G ₀, adopt homomorphism integrated technology, utilize the Bagging algorithm of monolateral sampling, the imbalance problem that solves data trains integrated support vector machine classifier.Then for level G ₁, the synthetic sorter that adopts differential mode integrated technology to train based on Bayes, decision tree and algorithm of support vector machine is weighted ballot output to sample, increases the extensive precision of fault diagnosis system.(3) software of algorithm is realized and is adopted the C#.NET of Microsoft platform, the weka.jar file of Weka software project is converted to the weka.dll procedure set that can be called by .NET by IKVM.NET instrument, some class in weka.dll is rewritten, complete and adopt three-tier architecture model to write software to after the specific implementation of algorithm, realize the concrete analysis of fault and diagnosis.Method for diagnosing faults of the present invention has advantages of that performance of fault diagnosis is higher, diagnostic area is wider and algorithm robustness, interpretation are stronger.

Images