CN104849650B

CN104849650B - One kind is based on improved analog-circuit fault diagnosis method

Info

Publication number: CN104849650B
Application number: CN201510255739.1A
Authority: CN
Inventors: 毛万标; 柴毅; 张可; 熊英志; 张迅捷; 王鸣; 王一鸣
Original assignee: 63790 Troops Of Pla; Chongqing University
Current assignee: Chinese People's Liberation Army 63790 Unit; Chongqing University
Priority date: 2015-05-19
Filing date: 2015-05-19
Publication date: 2018-03-02
Anticipated expiration: 2035-05-19
Also published as: CN104849650A

Abstract

One kind is based on improved analog-circuit fault diagnosis method, and conventional art is improved in terms of two：1st, for the improvement of DAGSVM methods, using the maximum SVM of spacing between class as DAGSVM the superiors node, if root node classification results are i classes, then the selection SVM maximum with i classes between class distance is as this level of child nodes, if root node classification results are j classes, the selection SVM maximum with j classes between class distance is as this level of child nodes；If classification results be both not belonging to i classes or be not belonging to j classes, this two class is excluded, between class distance D is selected in remaining class_ijTwo maximum class SVM continue the above two steps as this node layer, until obtaining diagnostic result.Upper level node diagnostic error can effectively be avoided by so doing, and cause the wrong situation of final result.2nd, to improve the accuracy rate of diagnosis of each child node, parameter optimization is carried out using particle cluster algorithm (PSO) to every sub- SVM, the SVM accuracy rates of diagnosis of each node are improved, to improve whole DAGSVM accuracy rate of diagnosis.

Description

One kind is based on improved analog-circuit fault diagnosis method

Technical field

The present invention relates to analog circuit fault diagnosing field, and in particular to one kind is based on improved analog circuit fault diagnosing Method.

Background technology

With the continuous development of electronic technology and computer technology, the structure composition of equipment becomes increasingly complex.And in equipment Often analog circuit easily goes wrong, and causes its testing time and testing cost very high always.Constantly expand in circuit structure Under the big and development trend that becomes increasingly complex, difficulty of test also becomes higher.The scale of modern circuitry is increasing, complicated Degree also more and more higher, the requirement in terms of circuit system reliability and fault diagnosis is also improved constantly.Theory analysis and reality Using showing, analog circuit is more easy to break down than digital circuit, although the digital circuit in electronic equipment more than 80%, More than 80% failure but comes from analog circuit.

Analog circuit fault diagnosing field has been achieved for certain achievement at present.By wavelet transformation, WAVELET PACKET DECOMPOSITION, Hilbert-Huang transform or principle component analysis processing response signal, extract fault signature, then pass through expert system, nerve net Network, SVMs, fuzzy technology, rough set and parting theory etc. realize fault diagnosis.Analog circuit fault diagnosing master at present To be faced is that fault sample deficiency and diagnostic knowledge are pinpointed the problems for topic, and SVMs (SVM) can solve sample well This problem, realize the classificating knowledge implied in data to greatest extent with limited feature.But SVM is originally for solving two The problem of classification, currently used SVM realize that polytypic method has one-to-many (OVR), one-to-one (OVO) and directed acyclic graph (DAG).OVR training samples are unbalanced so as to cause nicety of grading not high, although and OVO ratio of precision OVR methods are high, due to Voting mechanism, it may appear that no sample belongs to multiclass simultaneously or is not belonging to the situation of any class.DAGSVM is directed to OVO algorithms not Foot is classified using exclusive method, and every sub- SVM classifier excludes a class most unlikely, can finally obtain classification knot Fruit.And DAGSVM algorithms there is also it is certain the problem of：When the classification results mistake of previous node, the classification knot of subsequent node Fruit is also mistake, thus cannot get correct result.

The content of the invention

In view of this, the purpose of the present invention is exactly to propose that one kind is based on improved analog-circuit fault diagnosis method, and it is right If obtaining wrong classification results during previous node-classification mistake in DAGSVM algorithms makes improvement, by classification accuracy highest SVM bring to Front, and each SVM classification accuracy is improved using PSO algorithms, degree reduces upper layer node to greatest extent The situation of classification error, so that fault recognition rate increases.

To reach above-mentioned purpose, the present invention provides following technical scheme：

One kind is based on improved analog-circuit fault diagnosis method, it is characterised in that comprises the steps：Step 1：Signal Collection, signal is gathered from the specific node of analog circuit；Step 2：Fault signature extracts, and the signal collected is carried out small Ripple bag decomposes and normalized, obtains fault signature；Step 3：These fault signatures are respectively used into child node PSOSVM to enter Row training, and calculate between class distance；Step 4：SVM at child node in DAGSVM is replaced with PSOSVM, and by between class Distance determines each child node successively from top to bottom.

Further, the signal acquisition method described in step 1, specifically includes following steps：Imitated using Pspice Very, Monte Carlo analyses are carried out to circuit, output end voltage signal is sampled, 500 data of collection obtain sample；

Further, the feature extracting method described in step 2, specifically includes following steps：31：Using db2 in wavelet systems Small echo carries out WAVELET PACKET DECOMPOSITION to the signal after denoising, then extracts each frequency in all sub-bands of n-th layer from low to high The signal characteristic of rate composition；32：The signal S of each sub-band scope in n-th layer is reconstructed according to WAVELET PACKET DECOMPOSITION coefficient_j；33：According to E_j=∫ | s_j(t)|²Dt=∑s_K=1 ⁿ|x_ik|²The ENERGY E of each sub-band signal_j, x in formula_ikIt is the amplitude of discrete signal reconstruction point, j N is arrived for 1；34：Construction feature vector X=[E1, E2 ..., E_N], and it is normalized to obtain analog circuit fault characteristic vector；

Further, the PSOSVM methods described in step 3, specifically include following steps：41：Fault signature is as training Data input, and primary and initial parameter are generated at random；42：SVM models are established, using SVM accuracy rate as particle Fitness, if this fitness is better than the adaptive optimal control degree of particle, position vector now is stored as the position vector of particle, if grain The fitness of son is better than global optimum's fitness, then position vector is stored as global optimum.43：Repeat above step and know satisfaction Final quasi- side or the greatest iteration step number for having reached setting, the parameter optimized：Punishment parameter C and kernel functional parameter γ；

Further, the improved DAGSVM methods described in step 4, specifically include following steps：51：Pass through D_ij=| | m_i-m_j||²-r_i-r_jCalculate the between class distance of every two classes failure；52：Between class distance is arranged from big to small, takes between class distance D_ijMost Big PSOSVM graders are as the superiors' node classifier；53：If root node classification results are i classes, between selection and i class classes The maximum SVM classifier of distance is as this level of child nodes；If root node classification results are j classes, selection and j classes between class distance are most Big SVM classifier is as this level of child nodes；If classification results be both not belonging to i classes or be not belonging to j classes, this two class is excluded, Between class distance D is selected in remaining class_ijTwo maximum class SVM classifiers continue the above two steps as the superiors' node classifier Suddenly, until obtaining diagnostic result；

The it is proposed of this hair this patent is based on improved analog-circuit fault diagnosis method, can effectively prevent that upper layer node diagnosis is wrong Final result diagnostic error caused by by mistake, improve diagnostic reliability.

Brief description of the drawings

In order that the object, technical solutions and advantages of the present invention are clearer, the present invention is made below in conjunction with accompanying drawing into The detailed description of one step, wherein：

Fig. 1 is DAG-SVM sorting algorithm schematic diagrames；

Fig. 2 is improved analog circuit fault diagnosing schematic flow sheet；

Fig. 3 is PSOSVM schematic flow sheets.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.

Fig. 2 is improved analog circuit fault diagnosing schematic flow sheet.The present invention provides a kind of improved analog circuit event Hinder diagnostic method, mainly include the following steps that：Analog circuit signal acquisition, fault signature are extracted, entered using improved DAGSVM Row fault diagnosis, each several part process comprise the following steps：

S1：Analog circuit signal acquisition, Pspice simulation softwares are utilized in this step, Monte is carried out to circuit Carlo is analyzed, and output end voltage signal is sampled, and 500 data of collection obtain sample；

S2：Fault signature is extracted, and WAVELET PACKET DECOMPOSITION and normalized are utilized in this step, obtains analog circuit fault Feature, it is comprised the following steps that：

S21：WAVELET PACKET DECOMPOSITION is carried out to the signal after denoising using db2 small echos in wavelet systems, then extracts n-th layer from low Signal characteristic of the frequency to each frequency content in all sub-bands of high frequency；

S22：The signal S of each sub-band scope in n-th layer is reconstructed according to WAVELET PACKET DECOMPOSITION coefficient_j；

S23：According to each E_j=∫ | s_j(t)|²Dt=∑s_K=1 ⁿ|x_ik|²The ENERGY E of sub-band signal_j, x in formula_ikIt is discrete letter The amplitude of number reconstruction point, j are 1 to arrive N；

S24：Construction feature vector X=[E1, E2 ..., E_N], and it is normalized obtain analog circuit fault feature to Amount.

S3：Fault diagnosis is carried out using DAGSVM is improved, in this step, improved DAGDVM models is built, realizes mould Intend circuit fault diagnosis, it mainly includes following two parts：Train PSOSVM and build improved DAGSVM models, each several part Comprise the following steps that：

S31：Train PSOSVM

S311：Fault signature inputs as training data, and generates primary and initial parameter at random；

S312：SVM models are established, the fitness using SVM accuracy rate as particle, if this fitness is better than particle most Excellent fitness, position vector now is stored as the position vector of particle, if the fitness of particle is better than global optimum's fitness, Then position vector is stored as global optimum；

S313：Repeat above step and know the greatest iteration step number for meeting final quasi- side or having reached setting, obtain excellent The parameter of change：Punishment parameter C and kernel functional parameter γ.

S32：Build improved DAGSVM models

S321：Pass through D_ij=| | m_i-m_j||²-r_i-r_jCalculate the between class distance of every two classes failure；

S322：Between class distance is arranged from big to small, takes between class distance D_ijMaximum PSOSVM graders save as the superiors Point grader；

S323：) if root node classification results are i classes, the selection SVM classifier maximum with i classes between class distance is used as this Level of child nodes；If root node classification results are j classes, the selection SVM classifier maximum with j classes between class distance is as this straton section Point；If classification results be both not belonging to i classes or be not belonging to j classes, this two class is excluded, between class distance D is selected in remaining class_ijMost Two big class SVM classifiers continue the above two steps as the superiors' node classifier, until obtaining diagnostic result.

By above step, the fault diagnosis of analog circuit can be realized.

Finally illustrate, preferred embodiment above is merely illustrative of the technical solution of the present invention and unrestricted, although logical Cross above preferred embodiment the present invention is described in detail, it is to be understood by those skilled in the art that can be Various changes are made to it in form and in details, without departing from claims of the present invention limited range.

Claims

1. one kind is based on improved analog-circuit fault diagnosis method, it is characterised in that comprises the steps：

Step 1：Signal acquisition, signal is gathered from the specific node of analog circuit；

Step 2：Fault signature is extracted, and the signal collected is carried out into WAVELET PACKET DECOMPOSITION and normalized, obtains failure spy Sign；

Step 3：These fault signatures are respectively used into child node PSOSVM to be trained, and calculate between class distance；

Step 4：SVM at child node in DAGSVM is replaced with PSOSVM, and determined successively from top to bottom by between class distance Fixed each child node；

In PSOSVM methods described in step 3 kind, comprise the following steps that：41：Fault signature inputs as training data, and at random Generate primary and initial parameter；42：SVM models are established, the fitness using SVM accuracy rate as particle, if this is adapted to Degree is better than the adaptive optimal control degree of particle, and position vector now is stored as the position vector of particle, if the fitness of particle is better than Global optimum's fitness, then position vector be stored as global optimum；43：Above step is repeated until meeting final criterion or reaching The greatest iteration step number of setting, the parameter optimized are arrived：Punishment parameter C and kernel functional parameter γ.

2. one kind according to claim 1 is based on improved analog-circuit fault diagnosis method, it is characterised in that：Step 1 Described in signal acquisition method, comprise the following steps that：

Emulated using Pspice, Monte Carlo analyses are carried out to circuit, output end voltage signal is sampled, adopted 500 data of collection obtain sample.

3. one kind according to claim 1 is based on improved analog-circuit fault diagnosis method, it is characterised in that：In step Feature extracting method described in two, is comprised the following steps that：31：The signal after denoising is carried out using db2 small echos in wavelet systems small Ripple bag decomposes, and then extracts the signal characteristic of each frequency content in all sub-bands of n-th layer from low to high；32：According to The signal S of each sub-band scope in WAVELET PACKET DECOMPOSITION coefficient reconstruct n-th layer_j；33：According toRespectively The ENERGY E of sub-band signal_j, x in formula_1kIt is the amplitude of discrete signal reconstruction point, j arrives N for 1；34：Construction feature vector X= [E₁,E₂,…,E_N], and it is normalized to obtain analog circuit fault characteristic vector.

4. one kind according to claim 1 is based on improved analog-circuit fault diagnosis method, it is characterised in that：In step Improved DAGSVM methods described in four, are comprised the following steps that：51：Pass through D_ij=| | m_i-m_j||²-r_i-r_jCalculate every two class The between class distance of failure, wherein m_iRepresent the central point for the feature that the i-th class extracts, m_jRepresent the feature that jth class extracts Central point, r_iRepresent the radius of i classes, r_jRepresent the radius of j classes；52：Between class distance is arranged from big to small, takes between class distance D_ij Maximum PSOSVM graders are as the superiors' node classifier；53：If root node classification results are i classes, selection and i class classes Between the maximum SVM classifier of distance as this level of child nodes；If classification results be both not belonging to i classes or be not belonging to j classes, this is discharged Two classes, between class distance D is selected in remaining class_ijOn two maximum class SVM classifiers continue as the superiors' node classifier The step of face two, until obtaining diagnostic result.