CN101871994A - Method for diagnosing faults of analog circuit of multi-fractional order information fusion - Google Patents

Abstract

The invention discloses a method for diagnosing faults of an analog circuit of multi-fractional order information fusion, belonging to the field of network fault test and diagnosis of the analog circuit. The method comprises the following steps of: A inputting N types of effective sinusoidal excitation signals with different frequencies in series at the input end of a circuit to be tested to obtain an initial dataset; B randomly dividing a sample set into k feature space subsets by adopting a random subspace method; C mapping the feature space subsets into k diverse fractional-order time-frequency spaces by using fractional-order Fourier transform; and D obtaining a final diagnosis result by utilizing the weight of a failure class and fusing the diagnosis results of k basic classifying models, and dynamically updating the weight of the failure class after completing the diagnosis on each batch of test data. The invention improves the difference degree of the feature space subsets, increases complementarity among basic classifiers, combines with the gradual change characteristics of device faults and effectively improves the precision of a diagnosing system.

Description

The analog-circuit fault diagnosis method that multi-fractional order information merges

Technical field

The present invention relates to a kind of analog-circuit fault diagnosis method, the analog-circuit fault diagnosis method that especially a kind of multi-fractional order information merges.

Background technology

Mimic channel is the indispensable important component part of electronic circuit; along with extensive Analogous Integrated Electronic Circuits especially developing rapidly of modulus hybrid circuit; the complexity and the closeness of mimic channel constantly increase; in case break down in the somewhere; can not in time be diagnosed, be recovered; gently then shut down, heavy then cause casualties, bring tremendous economic loss.And for the application that high reliability request is arranged, the reliability guarantee of equipment is particularly important.Therefore, greatly developing and technology such as application autonomous formula analog circuit fault diagnosing and failure prediction, become the important content in modern industry production and the national defense construction, also is one of focus of present scientific circles research.

The research of feature extraction and diagnostic method is the important content of analog circuit fault diagnosing research, is the focus of analog-circuit fault diagnosis method research now based on the feature extracting method of signal Processing with based on the intelligent diagnosing method of pattern-recognition.For example Fourier transform, wavelet transformation, Xi Er baud Huang, Fourier Transform of Fractional Order etc., these methods compare with the traditional characteristic extracting method can the acquisition cuicuit state dynamic feature information.Neural network, expert system, fuzzy theory, decision tree, the recognition methods of bayesian theory isotype need not mathematical models, only need the specific operation rule of utilization, measurement space is mapped to decision space, avoided numerous and diverse mathematical operation, only need limited failure message, with regard to the fault element in the energy decision network, it is convenient to compare enforcement in traditional diagnosis method, has higher diagnostic accuracy simultaneously.

Existing analog-circuit fault diagnosis method is mainly studied by identical approach and is carried out feature extraction, obtains the training dataset in the same feature space, and adopts single sorter to carry out fault diagnosis.Yet, show that by practical application this traditional analog-circuit fault diagnosis method has following deficiency: do not have too many otherness between the training set sample that (1) is obtained by the fault signature that extracts under the same approach, cause training the disaggregated model generalization ability that obtains poor; (2) actually show that any single sorter all can not solve the problem that analog circuit fault diagnosing faces completely or reach system requirements.When single disaggregated model runs into the failure classes that certain class is difficult to diagnose, be difficult to by the adjustment model parameter or increase training sample, improve diagnostic accuracy.And multifarious categorizer integration method can effectively address the above problem, and by the complementarity between a plurality of sorters, improves the precision of whole disaggregated model.In recent years, sorter is integrated to be the focus of a research, yet also was in the starting stage in the analog circuit fault diagnosing field.

In addition, the fault that components and parts take place in the mimic channel has the characteristics of gradually changeable.Because the existence of tolerance, the components and parts value is the normal range of tolerable variance inner conversion about nominal value always, think during when the components and parts value and to break down greater than normal range of tolerable variance, rule according to the components and parts damage, always from little soft fault taking place to big soft fault, arrive hard fault again, so the fault of components and parts has gradually changeable in the mimic channel.Components and parts take place after the big soft fault, before the diagnostic model that obtains by the training of the little soft fault sampled data fault signature of coincidence circuit components and parts no longer.Therefore, according to the characteristics of mimic channel components and parts fault gradually changeable, need dynamically update fault diagnosis system, and the research of this respect is still few so far, is not solved.

Summary of the invention

The object of the present invention is to provide a kind of analog-circuit fault diagnosis method that can solve the diagnostic model generalization ability difference of existing analog-circuit fault diagnosis method existence and ignore the problem of fault gradually changeable.

The present invention reaches purpose of the present invention by the following technical programs:

In order to improve the diversity factor between the training sample set, at first adopt the stochastic subspace method that the original sample collection is divided into several features space subclass.Then, select different fractional order p values, adopt Fourier Transform of Fractional Order that the data map of feature space subsets is arrived different fractional order time frequency space for the sample set in each subcharacter space; In order to improve the generalization ability of diagnostic model, adopt the complementarity between the multi-classifier integrating method raising diagnostic model, obtain the last diagnostic result by the diagnostic result that merges a plurality of sorters; In order to solve the characteristics of components and parts fault gradually changeable in the mimic channel, concentrate the ratio that occupies according to up-to-date all kinds of failure classes in all test sample books, dynamically update the weights of every class failure classes, and the weights size of all kinds of failure classes is fused in the diagnostic result.

Specifically, technical scheme of the present invention is carried out analog circuit fault diagnosing according to following each step:

The sinusoidal excitation signal of A, N different frequency in the input end series connection input circuit allowed band of circuit under test, the span of N is 2-5, collection can be surveyed the voltage signal of node, obtains raw data set;

B, the raw data set that steps A is obtained carry out the proper subspace pre-service, be specially: raw data set is divided into k feature space subsets at random with the stochastic subspace method, each feature space subsets is non-intersect each other, and the feature samples number average of each class equates in each feature space subsets; Wherein k for the number of the basic disaggregated model that intend to adopt and k more than or equal to 2;

C, will be mapped in the fractional order time frequency space of k inequality through k the proper subspace that step B pre-service obtains, and obtain k basic disaggregated model by the feature samples training in described k the different fractional order time frequency space with Fourier Transform of Fractional Order;

D, utilize k basic disaggregated model that step C obtains to dynamic acquisition to test data diagnose respectively in batches, utilize the failure classes weights and merge the result of k basic disaggregated model diagnosis according to following formula, obtain final diagnostic result class of z, and after each BT(batch testing) data is finished diagnosis, dynamically update the weights of failure classes:

$\mathrm{class\; of\; z}=\arg \max \underset{j=1}{\overset{f}{\mathrm{\Σ}}}{P}_j\·\left(\underset{i=1}{\overset{k}{\mathrm{\Σ}}}(\mathrm{out}{\mathrm{put}}_i\left(z\right)=j)\right)$

Wherein, z represents sample to be tested; F represents the fault category number; K represents basic disaggregated model number; Output _i(z)=and j represents that test sample book z is diagnosed as the probable value of failure classes j by i basic disaggregated model, j=1, and 2 ... f, i=1,2 ... k; P _jBe the weights that j class failure classes are endowed, calculate by following formula:

$P_j=\frac{s+\mathrm{mg}}{{\mathrm{number}}_j}(j=\mathrm{1,2},...f)$

Wherein, s is the sample number that raw data is concentrated; M is predefined quantitative constant, represents test specimens given figure in every BT(batch testing) sample set; G is that total test specimens given figure deducts the integer-bit numerical value that original sample is counted the merchant who obtains divided by m behind the s; Number _jBe the sample number that is judged as j class failure classes after the diagnosis, and number _jThe sample number that comprises the concentrated j class failure classes of training sample.

Above-mentioned predefined quantitative constant m can choose according to actual conditions.

Compared to existing technology, the present invention program has following technique effect:

(1) the present invention utilizes Fourier Transform of Fractional Order that raw data set is mapped to k different fractional order time frequency space, changed the distribution character of raw data set by inequality fractional order time frequency space, improve the otherness between the different training sample sets of proper subspace, helped improving the integrated complementarity of sorter.

(2) the present invention adopts neural net method as basic disaggregated model, utilize the failure classes weights and merge the result of k basic disaggregated model diagnosis, obtain final diagnostic result, the integrated more traditional single diagnostic method of diagnostic method of multi-categorizer is compared, have more generalization ability, by the integrated complementarity that has increased between the diagnostic model of multifarious multi-categorizer, thus the precision of raising analog circuit fault diagnosing.

(3) to dynamic acquisition to test data diagnose in batches, the ratio that in all data samples, occupies according to up-to-date all kinds of failure classes, dynamically update the weights of failure classes, the characteristics that meet the analog circuit fault gradually changeable, and weights are fused in the diagnostic result, improved the efficient of analog circuit fault diagnosing.

Fourier Transform of Fractional Order is a kind of unified time-frequency conversion, by twiddle factor α, in the time frequency plane rotation, obtains new signal indication, and it is defined as:

$X_{\mathrm{\α}}\left(u\right)=F_{\mathrm{\α}}\left[X\right]\left(u\right)={\∫}_{-\∞}^{+\∞}{\mathrm{\κ}}_a(t,u)f\left(t\right)\mathrm{dt}$

Wherein,

$k_{\mathrm{\α}}(x,u)=\left\{\begin{array}{cc}\sqrt{(1-j\cot \mathrm{\α})}\·e^{\mathrm{j\π}(x^2\cot \mathrm{\α}-2\mathrm{ux}\csc \mathrm{\α}+u^2\cot \mathrm{\α})}& \mathrm{\α}\≠\mathrm{n\π}\\ \mathrm{\δ}(x-u)& \mathrm{\α}=\mathrm{n\π}\\ \mathrm{\δ}(x+u)& \mathrm{\α}=(2n\±1)\mathrm{\π}\end{array}\right.$

α in the formula=p pi/2, p are the fractional order of Fourier Transform of Fractional Order, the fractional number Fourier field on p rank be (t, w) on the plane by the coordinate space that angle produced of counter clockwise direction rotation alpha.The desirable any number of α, generally getting the counter clockwise direction anglec of rotation of 0～pi/2 analyzes, other angles are because the symmetry of fractional number Fourier field and periodicity, the signal analysis result is consistent with 0～pi/2 anglec of rotation, the value of p value generally gets 0～1, when p=0, Fourier Transform of Fractional Order is an original signal; When p=1, equal classical Fourier transform.Along with p changes to 1 from 0, Fourier Transform of Fractional Order changes to common Fourier transform from original function smoothly.

As seen from the above analysis, if use Fourier Transform of Fractional Order, at first must determine fractional order p value.Therefore, utilize the uncertainty of fractional order p value, k random character subspace is mapped to different fractional order time frequency space, thereby increase the diversity factor of subspace feature samples.

In order further to increase the diversity factor of space characteristics sample, can utilize genetic algorithm that fractional order p value is carried out optimizing, and with the diversity factor value between the multi-categorizer as fitness function, the evolution target of genetic algorithm is and makes the fitness function maximization, thereby obtains k optimum fractional order p value; Particularly, the step C in the such scheme comprises following each step:

C1, the value of initial fractional order p is set, the span of p is 0～1; Set up initial population;

C2, utilize the Fourier Transform of Fractional Order method that k proper subspace sample is mapped to k different fractional order time-frequency domain according to fractional order p value;

C3, according to the feature samples of the different fractional order time-frequency subspace of k, training obtains k basic disaggregated model, wherein diagnostic model adopts three layers of BP neural net method, the transport function of neural network and training function adopt " Logsig " and " Traingdx " respectively, and the error training precision is 0.001;

C4, calculate diversity factor value d (all_classfier) between the multi-categorizer according to following formula:

$d(\mathrm{all}\_\mathrm{classfier})=\underset{n,m=1,n\≠m}{\overset{k}{\mathrm{\Σ}}}\frac{d(\mathrm{classfier}\_n,\mathrm{classfier}\_m)}{2}$

Wherein,

$d(\mathrm{classfier}\_n,\mathrm{classfier}\_m)=\frac{1}{\mathrm{prob}\_\mathrm{fail}(\mathrm{classfier}\_n,\mathrm{classfier}\_m)};$

Classfier_n, classfier_m represent k the n in the sorter and m sorter respectively;

(classfier_n classfier_m) is two sorter classfier_n to prob_fail, and classfier_m is to same test sample book while failed probability, promptly simultaneously by the value of the test specimens given figure of two sorter mistaken diagnosis divided by the test sample book sum;

C5, judge whether predefined stop condition is met, in this way, the k that then obtains fractional order p value continues execution in step D for optimum; As not, then population is selected, intersects, make a variation, heavily inserts and operate, obtain new population after, go to step C2; Wherein, described predefined stop condition is meant: the diversity factor value d (all_classfier) that obtains among the step C4 perhaps optimizes iterations greater than predefined second threshold value less than predefined first threshold.

Above-mentioned first threshold, second threshold value can be chosen according to actual conditions.

Description of drawings

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 is the process flow diagram that adopts a plurality of fractional order p values of multi-objective genetic algorithm optimized choice in the specific embodiment of the invention.

Embodiment

Below in conjunction with accompanying drawing technical scheme of the present invention is elaborated:

As shown in Figure 1, carry out analog circuit fault diagnosing according to following steps:

The sinusoidal excitation signal of A, N different frequency in the input end series connection input circuit allowed band of circuit under test, the span of N is 2-5, collection can be surveyed the voltage signal of node, obtains raw data set;

At first, analyze effective passband scope of circuit under test by the frequency sweep analytic approach, wherein the frequency transformation scope of frequency sweep is [0.001Hz 100MHz], in order to increase the multifrequency information in the voltage signal that test node collects, improve the separability of fault data sample, being all 1V, phase place in the series connection of the input end of circuit under test 3 frequency differences, amplitudes is 0 sinusoidal excitation signal, and wherein the frequency of sinusoidal excitation signal selects to represent high, normal, basic different frequency values in the effective passband scope of circuit under test respectively;

Then, select suitable test node, adopt the Node Voltage Sensitivity analytical approach to select optimum test node in this embodiment, Node Voltage Sensitivity is big more, this node changes responsive more to component parameters, help improving the fault identification ability more, specifically according to following steps:

1) sets the various faults classification of circuit under test in advance, the magnitude of voltage of the various faults class output node of collecting test node;

When gathering raw data set, need the artificial fault model of setting, because the analog circuit fault type class is various, in order to make the diagnostic model generalization ability that obtains by the raw data training big as far as possible, the fault type data of gathering need reflect the fault type of mimic channel as far as possible, therefore the fault category of setting in the present embodiment is 6, comprises the components and parts value and becomes big soft fault, components and parts value diminish soft fault, components and parts open circuit hard fault, components and parts short circuit hard fault, a plurality of components and parts soft fault, a plurality of components and parts hard fault;

2) adopt following voltage sensitivity formula to calculate the voltage sensitivity value of each measuring point.Selecting 2～3 test nodes of voltage sensitivity value maximum is the final test node.

$S_{\mathrm{es}}=\frac{{\∂V}_{\mathrm{out}}}{\∂r}$

Wherein r is a component parameter, V _OutBe test node voltage output variable.

B, the raw data set that steps A is obtained carry out the proper subspace pre-service, be specially: raw data set is divided into k feature space subsets at random with the stochastic subspace method, each feature space subsets is non-intersect each other, and the feature samples number average of each class equates in each feature space subsets; Wherein k for the number of the basic disaggregated model that intend to adopt and k more than or equal to 2;

In theory, the number of basic disaggregated model is many more, and promptly k is big more, the precision as a result of fusion diagnosis is high more, but can increase the complexity of diagnostic procedure, and therefore the k value is between the 5-10 usually, the k value is 8 in this concrete enforcement, promptly adopts 8 basic disaggregated models to carry out fusion diagnosis;

C, will be mapped in the fractional order time frequency space of k inequality through k the proper subspace that step B pre-service obtains, and obtain k basic disaggregated model by the feature samples training in described k the different fractional order time frequency space with Fourier Transform of Fractional Order;

This embodiment adopts multi-objective genetic algorithm that fractional order p value is carried out optimizing, and with the diversity factor value between the multi-categorizer as fitness function, the evolution target of genetic algorithm is and makes the fitness function maximization, thereby obtains k optimum fractional order p value; Particularly, as shown in Figure 2, this step comprises following each step:

C1, the value of initial fractional order p is set, the span of p is 0～1, sets up initial population;

C2, utilize the Fourier Transform of Fractional Order method that k proper subspace sample is mapped to k different fractional order time-frequency domain according to fractional order p value;

C3, according to the feature samples of the different fractional order time-frequency subspace of k, training obtains k basic disaggregated model, wherein diagnostic model adopts three layers of BP neural net method, the transport function of neural network and training function adopt " Logsig " and " Traingdx " respectively, and the error training precision is 0.001;

C4, calculate diversity factor value d (all_classfier) between the multi-categorizer according to following formula:

$d(\mathrm{all}\_\mathrm{classfier})=\underset{n,m=1,n\≠m}{\overset{k}{\mathrm{\Σ}}}\frac{d(\mathrm{classfier}\_n,\mathrm{classfier}\_m)}{2}$

Wherein,

$d(\mathrm{classfier}\_n,\mathrm{classfier}\_m)=\frac{1}{\mathrm{prob}\_\mathrm{fail}(\mathrm{classfier}\_n,\mathrm{classfier}\_m)};$

Classfier_n, classfier_m represent k the n in the sorter and m sorter respectively;

(classfier_n classfier_m) is two sorter classfier_n to prob_fail, and classfier_m is to same test sample book while failed probability, promptly simultaneously by the value of the test specimens given figure of two sorter mistaken diagnosis divided by the test sample book sum;

C5, judge whether predefined stop condition is met, in this way, the k that then obtains fractional order p value continues execution in step D for optimum; As not, then population is selected, intersects, make a variation, heavily inserts and operate, obtain new population after, go to step C2; Wherein, described predefined stop condition is meant: the diversity factor value d (all_classfier) that obtains among the step C4 perhaps optimizes iterations greater than predefined second threshold value less than predefined first threshold;

In this embodiment, first threshold, the second threshold value value are respectively 0.01,500, promptly when the diversity factor value less than 0.01 or when optimizing iterations and reaching 500 times, stop iteration, k the fractional order p value of this moment is optimum.

D, utilize k basic disaggregated model that step C obtains to dynamic acquisition to test data diagnose respectively in batches, utilize the failure classes weights and merge the result of k basic disaggregated model diagnosis according to following formula, obtain final diagnostic result class of z, and after each BT(batch testing) data is finished diagnosis, dynamically update the weights of failure classes:

$\mathrm{class\; of\; z}=\arg \max \underset{j=1}{\overset{f}{\mathrm{\Σ}}}{P}_j\·\left(\underset{i=1}{\overset{k}{\mathrm{\Σ}}}(\mathrm{out}{\mathrm{put}}_i\left(z\right)=j)\right)$

Wherein, z represents sample to be tested; F represents the fault category number; K represents basic disaggregated model number; Output _i(z)=and j represents that test sample book z is diagnosed as the probable value of failure classes j by i basic disaggregated model, j=1, and 2 ... f, i=1,2 ... k; P _jBe the weights that j class failure classes are endowed, calculate by following formula:

$P_j=\frac{s+\mathrm{mg}}{{\mathrm{number}}_j}(j=\mathrm{1,2},...f)$

Wherein, s is the sample number that raw data is concentrated; M is predefined quantitative constant, represents test specimens given figure in every BT(batch testing) sample set; G is that total test specimens given figure deducts the integer-bit numerical value that original sample is counted the merchant who obtains divided by m behind the s; Number _jBe the sample number that is judged as j class failure classes after the diagnosis, and number _jThe sample number that comprises the concentrated j class failure classes of training sample.

In this embodiment, in order to simplify the training complexity of neural network base disaggregated model, after the sample set that obtains k fractional order time-frequency characteristics space, extract T dimension principal character in the data sample of employing principal component analysis (PCA) PCA algorithm from each fractional order time-frequency domain it is carried out dimension-reduction treatment, as the input data of neural network base disaggregated model.The PCA method can be reduced to any dimension with former data set by extracting principal character, but the dimension that reduces as PCA is too low, can lose the validity feature information of too much former data set, in this embodiment, the T value is 30, and the feature samples that is about in the fractional order time-frequency domain is reduced to 30 dimensions;

In this embodiment, predefined quantitative constant m value is 500, after intact 500 test sample books of promptly every diagnosis, this BT(batch testing) data set is added original training dataset, recomputates the occupation rate of all kinds of failure classes, upgrades the weights P of all kinds of failure classes _jSize; Like this, certain components and parts gradual change of breaking down in mimic channel, the previous failure classes that obtained by little soft fault sample training of gathering will can not exist, by recomputating and upgrade the weights of failure classes dynamically, the weights of these failure classes can be reduced, the possibility that expression is diagnosed as this type of fault diminishes.

The genetic algorithm of using in this embodiment, fraction Fourier conversion, frequency sweep analytic approach, principal component analysis (PCA) PCA algorithm etc. are prior art, but particular content list of references (Multiobjective Intelligent optimized Algorithm and application thereof, the Randt is bright etc., and 2009, Science Press; Fourier Transform of Fractional Order and application thereof, happy and carefree etc., 2009, publishing house of Tsing-Hua University; The Computer Analysis of mimic channel and design---PSpice program is used, Gao Wenhuan etc., 1998, publishing house of Tsing-Hua University; The intelligent mode recognition methods, Xiao Jianhua, 2006, publishing house of South China Science ﹠ Engineering University).

Claims (4)

1. the analog-circuit fault diagnosis method that multi-fractional order information merges is characterized in that, may further comprise the steps:

The sinusoidal excitation signal of A, N different frequency in the input end series connection input circuit allowed band of circuit under test, the span of N is 2-5, collection can be surveyed the voltage signal of node, obtains raw data set;

B, the raw data set that steps A is obtained carry out the proper subspace pre-service, be specially: raw data set is divided into k feature space subsets at random with the stochastic subspace method, each feature space subsets is non-intersect each other, and the feature samples number average of each class equates in each feature space subsets; Wherein k for the number of the basic disaggregated model that intend to adopt and k more than or equal to 2;

C, will be mapped in the fractional order time frequency space of k inequality through k the proper subspace that step B pre-service obtains, and obtain k basic disaggregated model by the feature samples training in described k the different fractional order time frequency space with Fourier Transform of Fractional Order;

D, utilize k basic disaggregated model that step C obtains to dynamic acquisition to test data diagnose respectively in batches, utilize the failure classes weights and merge the result of k basic disaggregated model diagnosis according to following formula, obtain final diagnostic result class of z, and after each BT(batch testing) data is finished diagnosis, dynamically update the weights of failure classes:

$\mathrm{class\; of\; z}=\arg \max \underset{j=1}{\overset{f}{\mathrm{\Σ}}}{P}_j\·\left(\underset{i=1}{\overset{k}{\mathrm{\Σ}}}(\mathrm{out}{\mathrm{put}}_i\left(z\right)=j)\right)$

Wherein, z represents sample to be tested; F represents the fault category number; K represents basic disaggregated model number; Output _i(z)=and j represents that test sample book z is diagnosed as the probable value of failure classes j by i basic disaggregated model, j=1, and 2 ... f, i=1,2 ... k; P _jBe the weights that j class failure classes are endowed, calculate by following formula:

$P_j=\frac{s+\mathrm{mg}}{{\mathrm{number}}_j}(j=\mathrm{1,2},...f)$

Wherein, s is the sample number that raw data is concentrated; M is predefined quantitative constant, represents test specimens given figure in every BT(batch testing) sample set; G is that total test specimens given figure deducts the integer-bit numerical value that original sample is counted the merchant who obtains divided by m behind the s; Number _jBe the sample number that is judged as j class failure classes after the diagnosis, and number _jThe sample number that comprises the concentrated j class failure classes of training sample.

2. the analog-circuit fault diagnosis method that merges of multi-fractional order information according to claim 1, it is characterized in that: described step C specifically carries out according to following steps:

C1, the value of initial fractional order p is set, the span of p is 0～1, sets up initial population;

C2, utilize the Fourier Transform of Fractional Order method that k proper subspace sample is mapped to k different fractional order time-frequency domain according to fractional order p value;

C3, according to the feature samples of the different fractional order time-frequency subspace of k, training obtains k basic disaggregated model, wherein diagnostic model adopts three layers of BP neural net method, the transport function of neural network and training function adopt " Logsig " and " Traingdx " respectively, and the error training precision is 0.001;

C4, calculate diversity factor value d (all_classfier) between the multi-categorizer according to following formula:

$d(\mathrm{all}\_\mathrm{classfier})=\underset{n,m=1,n\≠m}{\overset{k}{\mathrm{\Σ}}}\frac{d(\mathrm{classfier}\_n,\mathrm{classfier}\_m)}{2}$

Wherein,

$d(\mathrm{classfier}\_n,\mathrm{classfier}\_m)=\frac{1}{\mathrm{prob}\_\mathrm{fail}(\mathrm{classfier}\_n,\mathrm{classfier}\_m)};$

Classfier_n, classfier_m represent k the n in the sorter and m sorter respectively;

(classfier_n classfier_m) is two sorter classfier_n to prob_fail, and classfier_m is to same test sample book while failed probability, promptly simultaneously by the value of the test specimens given figure of two sorter mistaken diagnosis divided by the test sample book sum;

C5, judge whether predefined stop condition is met, in this way, the k that then obtains fractional order p value continues execution in step D for optimum; As not, then population is selected, intersects, make a variation, heavily inserts and operate, obtain new population after, go to step C2; Wherein, described predefined stop condition is meant: the diversity factor value d (all_classfier) that obtains among the step C4 perhaps optimizes iterations greater than predefined second threshold value less than predefined first threshold.

3. the analog-circuit fault diagnosis method that merges as multi-fractional order information as described in the claim 2, it is characterized in that: the described predefined first threshold and the second threshold value value are respectively 0.01,500.

4. the analog-circuit fault diagnosis method that merges of multi-fractional order information according to claim 1, it is characterized in that: the value of described predefined quantitative constant m is 500.

Images