CN101071076A - Inspection apparatus and method - Google Patents

Abstract

The invention provides an inspection apparatus and an inspection method. The object of the invention is to determine criterion function and in order to make the criterion function used for nonparametric 1-class discrimination form a single on-spec product region. In a discriminant function deciding section 20 in an inspection device, nondefective region discriminating section 26 determines, in an input space in which the discriminant function used for the nonparametric 1-class discrimination plots a sample, whether a region containing the sample discriminated into the class is formed as a single region. Then, when the region containing the sample discriminated into the class is not single, a parameter setting section 24 sets region parameters for defining the size of the region of the basis function of the base of a density function so that the discriminant function constitutes a single region in an input space in which the sample is plotted. Wherein, the region parameters prescribes criterion function and the region size of basis function as basic of criterion function.

Description

Testing fixture and inspection method

Technical field

The present invention relates to from the instrumentation extracting data characteristic quantity of the inspection object of being imported, judge the testing fixture and the inspection method of the state of checking object according to the characteristic quantity that extracts.

Background technology

In automobile and household appliances etc., use the slewing of having assembled drive components such as motor much more very.For example, in automobile, slewing is installed everywhere at engine, power steering gear, automatic seat, wheel box etc.And, in household appliances, in various products such as refrigerator, air-conditioning, washing machine, slewing is installed.So, when the work of these slewings, owing to the rotation of motor etc. produces sound.

The sound that this sound that produces owing to rotation has the operate as normal followed to certainly lead to also has the sound that produces because of defective.For example, this abnormal sound of following defective have bearing unusual, inner unusual contact, imbalance, sneak into foreign matter etc.More specifically, frequency being arranged is that gear whenever revolves generation hypodontia, the foreign matter of nipping, point defect, the rotating part of motor interior and the abnormal sound that the moment friction of fixed part in rotation caused once that turn around.And, for example in 20Hz～20kHz scope that the people can hear, various sound are arranged.As the sound that the people feels under the weather, the sound about 15kHz of for example having an appointment.So, also belong to abnormal sound when producing the sound of this predetermined frequency component.Certainly, abnormal sound is not limited to this frequency.

The sound of following above-mentioned defective is not only uncomfortable, also might produce further fault.Therefore, be purpose with the quality assurance of these various products, have or not the judgement of abnormal sound.In production plant, usually carry out " organoleptic test " according to the sense of hearing and sense of touch etc. based on five senses by the overlooker.Specifically, the overlooker is by listening with ear, touching to confirm to vibrate with hand and check.In addition, organoleptic test's definition to some extent in Japanese organoleptic test's term JIS Z8144.

, the organoleptic test based on inspectoral five senses needs masterful technique.And organoleptic test's result of determination the deviation that causes such as changes greatly because of individual differences and time.In addition, there is the problem of the datumization of result of determination and the difficulty that quantizes, difficult management in the organoleptic test.Therefore, in order to address this is that, comprise that as inspection the unusual testing fixture of product of drive component has the abnormal sound testing fixture.The purpose of abnormal sound testing fixture is quantitatively and based on the stable inspection of clear and definite benchmark.

In abnormal sound testing fixture before this, be purpose to eliminate the generation loss, to reduce the inspection rate, make/improved high performance whether qualified decision algorithm.In addition, said omission is meant unacceptable product (abnormal article) mistake is judged to be certified products (normal product).Omission will cause unacceptable product to dispatch from the factory, so need prevent reliably.Said inspection excessively is meant the certified products mistake is judged to be unacceptable product.Cross inspection and will cause certified products not dispatch from the factory and the processing etc. that goes out of use, need avoid waste and the yield rate reduction.Therefore, increase the quantity of employed characteristic quantity, perhaps increase desired sample size in order to generate better decision rule.

On the other hand, in recent years the consumer for the increasingly stringent that requires of industrial product quality.And, in the manufacturing industry in Multi-varieties and Small-batch Production epoch, not only need to guarantee product quality, and production line started working become important topic.That is, only realize merely that high precision int that abnormal sound checks algorithm not enough.In order to carry the product of better quality, following two kinds of requirements are arranged in the production scene to market.

The firstth, the robotization of checking.That is, the inspection in the production processes such as the size of product and weight comes management quality at the definite respectively administrative standard of each characteristic value of production product usually.For example, in the testing fixture of organoleptic test's robotization that the abnormal sound inspection of the soldering visual examination of printed-wiring board (PWB) and motor car engine is such, from image and waveform, extract a plurality of mass propertys.Then, discrimination model comprehensively judge these characteristics judge whether qualified.

The secondth, vertical debugging (vertical start-up).Generally in the production scene, when production line is debugged, debug production line of batch through producing trial-manufacturing process in batches.The trial-production of said batch is meant in research and development design back and utilizes the means of production identical with batch process to manufacture a product, and determines to have in the operation no problem etc. and could produce in batches.When automatically generating the discrimination model of automatic checking device, if do not gather enough data, then can not modeling, so before producing beginning in batches, can not determine check criteria.Determine the check criteria that uses in the batch process stage in this batch advanced development, begin stable inspection when beginning to produce in batches, this becomes the important topic of the vertical debugging that realizes production line.

Above-mentioned organoleptic test comprehensively judge the size of sound, highly, that various mass propertys such as appearance color and shape are determined is whether qualified.Therefore, organoleptic test's automated system goes out to represent a plurality of characteristic quantities of mass property from the extracting data of utilizing sensors such as microphone and video camera to obtain, utilizes discriminant function to judge that whether qualified pattern-recognition be more effective.Generally, pattern-recognition need prepare to be used for to determine the learning sample of the sufficient amount of discriminant function.

Product examination based on pattern-recognition is described herein.

Figure 24 is the key diagram of the step of expression pattern-recognition.Said pattern-recognition is meant the method for determining the group that (differentiation) these data are affiliated according to the pattern of the characteristic quantity that goes out from extracting data.Therefore, pattern-recognition needs to generate (study) discriminant function automatically according to the data that observed in advance on model space.

And the method for pattern-recognition can be divided into four kinds according to the technique of expression that distributes and the symmetry of distribution.

The technique of expression that distributes can be categorized as " parameter differentiation " and " nonparametric discrimination ".The parameter differentiation utilizes statistical parameter to represent to distribute, and nonparametric discrimination does not utilize statistical parameter to represent to distribute.

And the symmetry of distribution can be categorized as symmetric " two class discrimination models " and symmetric " the class discrimination model " of not supposing distribution that supposition distributes.

Specifically, in learning phase, in parameter is differentiated,, pre-estimate the parameter (for example average, variance) of the shape of the probability density distribution that data of being used to stipulate to belong to each group defer to for a plurality of groups (for example normal and unusual) that the data that observed constitute.And in the differentiation stage, when observing new data, the parameter that use estimates is obtained the degree of membership for each group, determines to belong to which group.It is only can tentation data to conform to effective method can be with the probability density distribution (for example normal distribution) of parameter regulation shape the time that this parameter is differentiated.

And, in learning phase, in nonparametric discrimination, according to each a part of data of organizing all data that direct maintenance observed or helping to differentiate.Perhaps, do not use statistical parameter, and in nonparametric discrimination, directly obtain Density Distribution according to data.Then, in the differentiation stage, when observing new data, according to the data that kept or with the similar degree or the distance that distribute, obtain degree of membership for each group, determine to belong to which group.This nonparametric discrimination is can not tentation data to conform to can be with the probability density distribution of parameter regulation shape the time also effective method.

On the other hand, in learning phase, in two classes are differentiated, use and want two classes (for example certified products and unacceptable product) both sides' of differentiating sample to learn discriminant function.Then, in the differentiation stage, utilize discriminant function to obtain the degree of membership of unknown sample, more belong to which kind of comparative evaluation for two classes.

In addition, in a class is differentiated, in learning phase, only use the learning sample of a class to carry out density Estimation.Then, in the differentiation stage, utilize discriminant function to obtain the degree of membership of unknown sample according to density, and if carry out more than or equal to setting then be judged to be belonging to such, if less than setting then be judged to be the threshold determination that does not belong to such.

The concrete example of four quasi-modes identification is as described below.

(1) parameter two classes are differentiated ... Bayes identification, discriminatory analysis

(2) nonparametric two classes are differentiated ... arest neighbors arbiter (NN arbiter), support vector machine (SVM, support vector machine)

(3) parameter one class is differentiated ... horse field system (Mahalanobis-Taguchi system, MTS)

(4) nonparametric one class is differentiated ... histogram method, Nearest Neighbor Estimates, a class SVM, Parzen window technique, RBF (radial basis function (Radial Basis Function)) network, Density Estimator (kernel density estimation), double sampling bootstrapping method (bootstrap method) etc.

Certified products are identical, but unacceptable product is of all kinds.Therefore, the two common classes of supposing the being distributed symmetrically property of each class on feature space are differentiated the differentiation that is unsuitable for certified products/unacceptable product.And the unacceptable product sample that can collect in product examination is considerably less than the certified products sample.Therefore, a class of only considering the certified products distribution is differentiated more effective to the differentiation of certified products/unacceptable product.

And, check generally need when beginning to produce in batches, begin simultaneously.That is, the limited sample that needs basis to obtain before batch process is identified for differentiating the discriminant function of certified products/unacceptable product.But, before batch process, can not obtain enough certified products samples.Therefore, need the parameter of statistical estimation to differentiate and fully to guarantee performance with a small amount of sample.So determining according to limited sample under the situation of discriminant function, do not needing the nonparametric discrimination of statistical estimation more effective.

In view of above situation, nonparametric one class is differentiated more effective to the pattern-recognition of using in product examination.

In addition, as the prior art document relevant following non-patent literature 1～3 is arranged with the application's invention.

Parameter is differentiated and can't be utilized a small amount of sample to learn or be difficult to guarantee differentiate performance.For example, MTS can not carry out (accurately multicollinearity, accuracy multicollinearity) study during less than feature quantity in learning sample quantity.In addition, during greater than feature quantity, if having only a small amount of sample, then the precision deficiency of statistical estimation can not guarantee to differentiate performance sometimes in learning sample quantity.Therefore, in order to ensure performance, need about 3 times sample of feature quantity from experience.Non-patent literature 1 disclosed method is under the less situation of sample, by using or and guaranteeing performance with nonparametric discrimination.

Non-patent literature 2 disclosed methods are in a class SVM, and the ratio of utilizing support vector is the character by the upper limit of the mistake differentiation rate of cross validation (leave-one-out) method evaluation, regulate parameter and make the support vector number be minimum.But, in the method, because the shape of evaluation region not, so can not solve the problem that the certified products zone may be divided into a plurality of zones.

Non-patent literature 3 disclosed methods judge that for the combination of all learning samples that are judged as certified products whether the line segment that connects them breaks away from the certified products zone, carries out cluster to the learning sample that belongs to same area thus.As shown in figure 25, utilize 0/1 expression to have or not the matrix (Figure 25 (b), Figure 25 (d)) of disengaging, can know the sample that belongs to identical cluster by generation.In addition, the matrix of (Figure 25 (a)) when Figure 25 (b) expression certified products zone is, the matrix of (Figure 25 (c)) when Figure 25 (d) expression certified products zone is two.

Non-patent literature 1: rugged rich, Koji paddy in field and people, middle Island are grand, and " the differentiation モ デ Le of checking め Getting Jin from Move is new skill more " known and can be drilled defeated collected works, pp.243-246 (2005) by シ ス テ system シ Application Port ジ ウ system Talk for the 32nd time

Non-patent literature 2:Nello Cristianini, John Shawe-Taylor, " An Introductionto Support Vector Machines:And Other Kernel-Based Learning Methods ", Cambridge University Press (2000) (day translation: サ Port one ト ベ Network マ シ ン Ru Door, upright altogether the publication)

Non-patent literature 3:Asa Ben-Hur, David Horn, Hava T.Siegelmann, VladimirVapnik, " Support Vector Clustering ", Journal of Machine Learning Research2, pp.125-137 (2001)

To be the center with the desired value have deviation at its periphery (result from the deviation of parts/material and the change of manufacturing installation etc.) to mass property that it is generally acknowledged product, is the single area (Figure 26 (a)) at center so produce that the zone (real certified products zone) of certified products forms with the desired value.

And in fact the discriminant function that obtains from limited sample learning forms the certified products zone different with real certified products zone (study certified products zone).And this difference is more little, differentiates performance good more (Figure 26 (b)).

And, in most nonparametric discrimination, determine the certified products zone, so when having more sparse part in learning sample, the certified products zone might be divided into a plurality of zones (Figure 27 (a)) according to the density of learning sample.And, at learning sample more after a little while, also might be at the original bigger position of density, learning sample is more sparse, differentiates the risk that performance declines to a great extent so exist.

According to above situation, can think and then compare more approaching real certified products zone when being divided as long as the certified products zone is one (Figure 27 (b)), differentiate performance so can expect to improve.Therefore, judge after study whether discriminant function forms single certified products zone, and the adjusting parameter makes the certified products zone get final product for single area.

Summary of the invention

The present invention proposes in view of the above problems, its purpose is, a kind of can support to determine testing fixture, inspection method, the scrutiny program of the discriminant function of use in nonparametric one class is differentiated and the computer readable recording medium storing program for performing that has write down scrutiny program are provided.

In order to address the above problem, testing fixture of the present invention is characterised in that, this testing fixture has the discriminant function identifying unit, and this discriminant function identifying unit is judged whether the discriminant function that uses has been marked and drawed in the input space of sample and formed single area in nonparametric one class is differentiated.

Inspection method of the present invention is an inspection method of determining that the testing fixture of the discriminant function that uses is carried out in checking the condition discrimination of object, it is characterized in that the discriminant function identifying unit that above-mentioned testing fixture has judges whether the discriminant function that uses has marked and drawed the single certified products zone of formation in the input space of sample in nonparametric one class is differentiated.

According to said structure, can judge whether the discriminant function that uses has marked and drawed the single certified products zone of formation in the input space of sample in nonparametric one class is differentiated.In addition, differentiate, specifically, histogram method, Nearest Neighbor Estimates, a class SVM, Parzen window technique, RBF network, Density Estimator, double sampling bootstrapping method etc. are arranged as nonparametric one class.

Therefore, when the result of the above-mentioned judgement certified products single for discriminant function does not form are regional, be that the certified products zone is divided into a plurality of whens zone, for example can change the region parameter of the area size that is used to limit basis function, learn discriminant function once more, so that discriminant function forms single certified products zone.Therefore, the approaching real certified products zone by making certified products zone simplification, thus can improve the differentiation performance.

In addition, testing fixture of the present invention is characterised in that, this testing fixture has the parameter setting unit, this parameter setting unit is being judged to be above-mentioned discriminant function when not forming above-mentioned single certified products zone by above-mentioned discriminant function identifying unit, the setting regions parameter, make this discriminant function form single area in the input space of having marked and drawed sample, wherein this region parameter is stipulated above-mentioned discriminant function, and regulation is as the area size of the basis function on the basis of density function.

In addition, inspection method of the present invention is characterised in that, when being judged to be above-mentioned discriminant function by above-mentioned discriminant function identifying unit when not forming above-mentioned single area, the parameter setting unit setting regions parameter that above-mentioned testing fixture had, make that formation comprises the zone that is judged to be clipped to the sample in the class to this discriminant function as single area in the input space of having marked and drawed sample, wherein this region parameter is defined in the discriminant function that uses in the differentiation of nonparametric one class, and regulation is as the area size of the basis function on the basis of density function.

According to said structure, region parameter can be set at and make this discriminant function in the input space of having marked and drawed sample, form single certified products zone, wherein this region parameter is stipulated above-mentioned discriminant function, and regulation is as the area size of the basis function on the basis of density function.In addition, as region parameter, specifically, at the volume V of histogram method middle finger hypercube, in the radius r of Nearest Neighbor Estimates middle finger hypersphere, in the width cs of a class SVM middle finger gaussian kernel.

Therefore, when discriminant function does not form single certified products zone, when promptly the certified products zone is divided into a plurality of zone, also can changes region parameter, and discriminant function is learnt again, make discriminant function form single certified products zone.Therefore, near real certified products zone, can improve the differentiation performance by making certified products zone simplification.

In addition, the differentiation algorithm of the preferred above-mentioned nonparametric one class differentiation of testing fixture of the present invention is a class support vector machines.

At this, general support vector machine is compared with common nonparametric discrimination algorithm, has (i) versatility (to the differentiation performance of unknown sample) height, (ii) guarantees can not be absorbed in the study way characteristics of local solution.And these characteristics are also identical in a class support vector machines.

Therefore, according to said structure, can realize differentiating performance and the higher testing fixture of learning efficiency.

In addition, the above-mentioned discriminant function identifying unit of testing fixture of the present invention can judge also whether the line segment that connects each sample that is judged as certified products breaks away from certified products zone (Fig. 8 (a) (b)).

In addition, the above-mentioned discriminant function identifying unit of testing fixture of the present invention also can be judged the center that connects sample and whether the line segment that is judged as between the sample of certified products breaks away from certified products zone (Fig. 9 (a) (b)).

According to said structure, to compare when being connected each sample that is judged as certified products, the quantity of line segment reduces, so can suppress calculated amount.

In addition, the above-mentioned discriminant function identifying unit of testing fixture of the present invention can judge also whether the line segment that connects each interval support vector (margin support vector) breaks away from certified products zone (Figure 10 (a) (b)).

In addition, the above-mentioned discriminant function identifying unit of testing fixture of the present invention can judge also whether the center of connection sample and the line segment between the support vector of interval break away from certified products zone (Figure 11 (a) (b)).

According to said structure, to compare during support vector at interval with being connected each, the quantity of line segment reduces, so can suppress calculated amount.

In addition, the above-mentioned discriminant function identifying unit of testing fixture of the present invention also can extract one or more points from above-mentioned line segment, with the above-mentioned discriminant function of each substitution, judges that whether all points are all differentiated is certified products.

According to said structure, only the computing that one or more points are carried out discriminant function gets final product, so can easily carry out the judgement whether line segment breaks away from the certified products zone.And, the quantity of the point that can will extract according to the increase and decrease of desired judgement precision.

In addition, the above-mentioned discriminant function identifying unit of testing fixture of the present invention also can be obtained the minimum value of the above-mentioned discriminant function on the above-mentioned line segment, judges that whether this minimum value is more than or equal to the threshold value of stipulating.

According to said structure, whether the minimum value of also judging the above-mentioned discriminant function on the line segment is more than or equal to the threshold value of stipulating, so can carry out the judgement whether line segment breaks away from the certified products zone accurately.In addition, in order in the scope of line segment, to obtain the minimum value of discriminant function, can utilize nonlinear optimization methods such as Newton method and method of steepest descent.

In addition, above-mentioned testing fixture also can pass through computer realization, when this situation, by the computer readable recording medium storing program for performing that makes computing machine utilize the testing fixture scrutiny program of the above-mentioned testing fixture of computer realization and write down this scrutiny program as above-mentioned each cell operation also within category of the present invention.

As mentioned above, testing fixture of the present invention constitutes has the discriminant function identifying unit, and this discriminant function identifying unit judges whether the discriminant function that uses has marked and drawed the single certified products zone of formation in the input space of sample in nonparametric one class is differentiated.

In addition, inspection method of the present invention is that the discriminant function identifying unit that testing fixture has judges whether the discriminant function that uses has marked and drawed the method that forms single certified products zone in the input space of sample in nonparametric one class is differentiated.

Therefore, can bring into play following effect, when the result of the above-mentioned judgement certified products single for discriminant function does not form are regional, be that the certified products zone is divided into a plurality of whens zone, for example can change the region parameter of the area size that is used for the regulation basis function, discriminant function is learnt again, made discriminant function form single certified products zone.Therefore, near real certified products zone, can improve the differentiation performance by making certified products zone simplification.

In addition, testing fixture of the present invention constitutes has the parameter setting unit, this parameter setting unit is being judged to be above-mentioned discriminant function when not forming above-mentioned single certified products zone by above-mentioned discriminant function identifying unit, the setting regions parameter, so that this discriminant function forms single certified products zone in the input space of having marked and drawed sample, this region parameter stipulate above-mentioned discriminant function, and regulation as the area size of the basis function on the basis of density function.

In addition, inspection method of the present invention is following method, the parameter setting unit setting regions parameter that testing fixture has, so that this discriminant function forms single certified products zone in the input space of having marked and drawed sample, this region parameter is defined in discriminant function that nonparametric one class uses in differentiating and the regulation area size as the basis function on the basis of density function.

Therefore, when discriminant function does not form single certified products zone, when promptly the certified products zone is divided into a plurality of zone, can change region parameter, discriminant function is learnt again, make discriminant function form single certified products zone.Therefore, following effect can be brought into play, near real certified products zone, the differentiation performance can be improved by making certified products zone simplification.

Description of drawings

Fig. 1 is the functional block diagram of structure of the discriminant function determination portion that has of testing fixture of expression one embodiment of the present invention.

Fig. 2 is the key diagram of structure in general of the testing fixture of expression one embodiment of the present invention.

Fig. 3 is that the example that expression nonparametric one class is differentiated is the key diagram of histogram method.

Fig. 4 is that the example that expression nonparametric one class is differentiated is the key diagram of Nearest Neighbor Estimates.

Fig. 5 is that the example that expression nonparametric one class is differentiated is the key diagram of a class SVM.

The key diagram of the input space when Fig. 6 (a) is the no soft interval of expression, Fig. 6 (b) are the key diagrams of the Hilbert space when representing no soft interval.

Fig. 7 (a) is the key diagram of the input space of expression when soft interval is arranged, and Fig. 7 (b) is the key diagram of the Hilbert space of expression when soft interval is arranged.

Fig. 8 be that the certified products region quantity detection unit of discriminant function determination portion shown in Figure 1 generates, connect the key diagram of being differentiated for the line segment of each learning sample of certified products, (a) expression certified products zone is one a situation, and (b) expression certified products zone is two a situation.

Fig. 9 is center (average) and the key diagram of quilt differentiation for the line segment of the learning sample of certified products that the certified products region quantity detection unit of discriminant function determination portion shown in Figure 1 generates, that the connection quilt is differentiated the learning sample that is certified products, (a) expression certified products zone is one a situation, and (b) expression certified products zone is two a situation.

Figure 10 is the certified products region quantity detection unit of a discriminant function determination portion shown in Figure 1 key diagram that generate, that connect the line segment of each interval support vector (MSV), (a) expression certified products zone is one a situation, and (b) expression certified products zone is two a situation.

Figure 11 be that the certified products region quantity detection unit of discriminant function determination portion shown in Figure 1 generates, connect to be differentiated and be the center (average) of the learning sample of certified products and the key diagram of the line segment of support vector at interval, (a) expression certified products zone is one a situation, and (b) expression certified products zone is two a situation.

Even Figure 12 is the certified products zone is one but still the key diagram of inappropriate situation, (a) the porose situation in expression certified products zone, and (b) expression certified products zone is not the situation of convex form.

The key diagram that Figure 13 certified products area size that to be expression caused by the volume V size of the hypercube of histogram method changes, (a) the bigger situation of expression volume V, (b) the medium situation of expression volume V, (c) the less situation of expression volume V.

The key diagram that Figure 14 certified products area size that to be expression caused by the radius r size of the hypersphere of Nearest Neighbor Estimates changes, (a) the bigger situation of expression radius r, (b) the medium situation of expression radius r, (c) the less situation of expression radius r.

The key diagram that Figure 15 certified products area size that to be expression caused by the width cs size of the gaussian kernel of a class SVM changes, (a) the bigger situation of expression width cs, (b) the medium situation of expression width cs, (c) the less situation of expression width cs.

Figure 16 is the key diagram of the data used of discriminant function determination portion shown in Figure 1, (a) expression learning sample, (b) expression parameter candidate, (c) expression discriminant function.

Figure 17 is the process flow diagram that the discriminant function of expression discriminant function determination portion shown in Figure 1 is determined processing.

Figure 18 is the process flow diagram of certified products region quantity determination processing of the certified products region quantity detection unit of expression discriminant function determination portion shown in Figure 1.

Figure 19 is illustrated in the certified products region quantity determination processing of certified products region quantity detection unit of discriminant function determination portion shown in Figure 1, confirms the process flow diagram of discriminant functions more than or equal to the step of threshold value for a plurality of points that extract from line segment.

Figure 20 is illustrated in the certified products region quantity determination processing of certified products region quantity detection unit of discriminant function determination portion shown in Figure 1, the key diagram of the step when confirming discriminant function more than or equal to threshold value for a plurality of points that extract from line segment, (a) expression does not have the situation of disengaging, and (b) there is situation about breaking away from expression.

Figure 21 is illustrated in the certified products region quantity determination processing of certified products region quantity detection unit of discriminant function determination portion shown in Figure 1, for a plurality of points that extract from line segment, be judged to be under the situation of disengaging more than or equal to threshold value by the affirmation discriminant function, the key diagram of the concrete example of data that when line segment extracts, generate, (a) expression extracts 5 points that comprise two ends and registers the example of coordinate time, example when (b) data at registration line segment two ends are omitted in expression, (c) end of expression line segment is differentiated the example for the center of the learning sample of certified products.

Figure 22 is illustrated in the certified products region quantity determination processing of certified products region quantity detection unit of discriminant function determination portion shown in Figure 1, utilize optimization method to obtain the minimum value of discriminant function on the line segment, the process flow diagram of the step when confirming this minimum value more than or equal to threshold value.

Figure 23 is illustrated in the certified products region quantity determination processing of certified products region quantity detection unit of discriminant function determination portion shown in Figure 1, utilize optimization method to obtain the minimum value of the discriminant function on the line segment, the key diagram of the step when confirming this minimum value more than or equal to threshold value, (a) expression does not have the situation of disengaging, and (b) there is situation about breaking away from expression.

Figure 24 is the key diagram of the step of expression pattern-recognition.

Figure 25 utilizes 0/1 to represent whether the line segment between each learning sample of connection breaks away from the key diagram of certified products matrix of regions, (a) matrix when (b) expression certified products zone is, (c) matrix when (d) expression certified products zone is two.

Figure 26 is the key diagram in explanation certified products zone and unacceptable product zone, (a) the generation zone of expression certified products and unacceptable product, (b) real certified products zone and the study certified products zone of expression.

Figure 27 is the key diagram of the quantity in explanation certified products zone, and (a) expression certified products zone is divided into a plurality of situations, and (b) expression certified products zone is single situation.

Embodiment

Below, an embodiment of the invention are described with reference to the accompanying drawings.

Fig. 2 is the key diagram of structure in general of the testing fixture 100 of expression present embodiment.Fig. 1 is the functional block diagram of the structure of the discriminant function determination portion 20 that has of expression testing fixture 100.

The testing fixture 100 of present embodiment is from the instrumentation extracting data characteristic quantity of the inspection object imported.Testing fixture 100 is according to the characteristic quantity that extracts, and whether qualified differentiate sample by nonparametric one class.And particularly, testing fixture 100 has discriminant function determination portion 20.This discriminant function determination portion 20 possesses following function: (1) judges whether the discriminant function that uses forms single certified products zone and (comprise that being differentiated is the zone that belongs to the learning sample of class (certified products): the function zone that comprises many learning samples) in nonparametric one class is differentiated, (2) setting regions parameter, make this discriminant function form the function in single certified products zone, wherein this region parameter is defined in discriminant function that nonparametric one class uses in differentiating and the regulation area size as the basis function on the basis of density function.

Below, at first illustrate the functional profile of discriminant function determination portion 20 its apparatus structure to be described then.

[overview]

(1) key concept of nonparametric one class differentiation

It is not use statistical parameter that nonparametric one class is differentiated, and estimates the density of learning sample, and the area judging of density more than or equal to certain threshold value is certified products, will will be the method for unacceptable product less than the area judging of certain threshold value.As its concrete differentiation algorithm, except that the histogram method of following explanation, Nearest Neighbor Estimates, a class support vector machines (SVM), also have various algorithms such as Parzen window technique, RBF network, Density Estimator, double sampling bootstrapping method.

(a) histogram method

Fig. 3 is the key diagram of expression histogram method.Fig. 3 represents learning sample quantity more than or equal to 1 situation that is made as the certified products zone.And in the drawings, circle is represented sample, and a section is represented the hypercube of volume V.

Histogram method is divided into the hypercube of volume V with the input space, estimates density by the sample size that is comprised is counted.And, be the certified products zone with the learning sample quantity that comprised more than or equal to the area judging of certain threshold value, less than certain threshold value the time, differentiate and be unacceptable product zone (the oblique line part among the figure).

(b) Nearest Neighbor Estimates

Fig. 4 is the key diagram of expression Nearest Neighbor Estimates.In the drawings, circle is represented sample, and arc representation is the hypersphere of radius centered r with the sample.

Nearest Neighbor Estimates will be that the area judging that comprises in the hypersphere of radius centered r is the certified products zone with the learning sample, be unacceptable product zone (the oblique line part among the figure) with the area judging that does not comprise.

(c) a class SVM

Fig. 5 is the key diagram of expression one class SVM.In the drawings, the left side is the input space as former space, and the right side is the multidimensional Hilbert space that has shone upon the input space by Nonlinear Mapping φ.In addition, circle is represented sample.Sample on the identification face of determining in the Hilbert space after study among the figure shown in the downside is support vector (SV), and sample in addition is non-support vector (Non-SV).Zone after changing to the identification face reverse of Hilbert space on the input space is as the border.In a class SVM, be the certified products zone with the area judging in this border, be unacceptable product zone (the oblique line part among the figure) with the area judging outside the border.

More specifically, a class SVM is mapped to the input space on the multidimensional Hilbert space by Nonlinear Mapping φ for the learning sample on the input space, then the linear identification face of study.

At this, used the nuclear mapping of gaussian kernel to have following characteristics, that is, near the learning sample at the more sparse position of input space upper density was mapped in initial point, the learning sample at denser position was mapped in the position away from initial point.

[formula 1]

$K(x,z)=\exp (-\frac{{\left|\right|x-z\left|\right|}^2}{2{\mathrm{\σ}}^2})$

And with in initial point and the lineoid that learning sample separates, a class SVM learns the maximum lineoid (promptly discerning face) of the distance of distance initial point on Hilbert space.Learning sample on the lineoid is called support vector.

At this, n d dimensional vector x={x1 ..., the set of xd} is during as learning sample, is expressed as follows based on the discriminant function of a class SVM, differentiates more than or equal to 0 o'clock at functional value to be certified products, is unacceptable product differentiating less than 0 o'clock.

[formula 2]

$f\left(x\right)=\underset{i}{\mathrm{\Σ}}\left({\mathrm{\α}}_{i}K(x_i,x)\right)-\mathrm{\ρ}$

In following formula, xi represents the mark of learning sample.α i represents to be called as the coefficient of the weight of support vector, determines by study.And p represents constant that support vector xi (learning sample of factor alpha i ≠ 0) substitution following formula is arbitrarily obtained.

[formula 3]

$f\left(x\right)=\underset{i}{\mathrm{\Σ}}\left({\mathrm{\α}}_{i}K(x_i,x)\right)-\mathrm{\ρ}$

In addition, generally on the input space, learn relatively difficulty of non-linear identification face.But, become the linear identification face on the hyperspace if be mapped to, just be easy on hyperspace, use the linear discriminant algorithm to learn.And SVM is the linear discriminant algorithm.In addition, both can as long as can in this study, calculate the inner product of two vectors.Therefore, directly on the input space, do not learn non-linear identification face, and on hyperspace, learn the linear identification face by SVM.

At this, the function of the inner product of two vectors on two vector representation hyperspace utilizing on the input space is called " kernel function ".And,, and needn't carry out the calculating on the hyperspace so long as kernel function just can easily be learnt non-linear identification face.The Calculation Method that this calculating that utilizes kernel function is replaced on the hyperspace calls nuclear skill (kernel trick).

(the soft interval among the class SVM (soft margin))

At this, the key diagram of the input space when Fig. 6 (a) is the no soft interval of expression, Fig. 6 (b) are the key diagrams of the Hilbert space when representing no soft interval.And Fig. 7 (a) is the key diagram of the input space of expression when soft interval is arranged, and Fig. 7 (b) is the key diagram of the Hilbert space of expression when soft interval is arranged.

Usually, a class SVM makes one of learning sample not remain in the study (at interval hard) of the initial point side of identification face.But a class SVM exists under the situation that comprises outlier in the learning sample, differentiates performance owing to crossing the problem that severe judgement does not descend.To this,, can avoid other for the mistake severe judgement of outlier by allowing the remaining certain proportion of initial point side (soft interval) of learning sample at the identification face.

In soft interval, the ratio v that will allow remains in the initial point side imports in the learning algorithm, the discriminant function utilization of learning outcome during with hard interval identical form represent.And, in soft interval, not only be positioned at the learning sample on the identification face, and the learning sample that remains in the initial point side is also referred to as support vector.Therefore, support vector on distinguishing identification face and when remaining in the support vector of initial point side, the former is called support vector at interval, for example in SVM, be called support vector (MSV (margin support vector): be stored in the sample on the identification face in the model) at interval, the latter is called bounded support vector (BSV, bounded support vector).

(2) the certified products zone is single judgement

In the certified products zone for single, be certified products zones under a plurality of situations, certain line segment that breaks away from the certified products zone that exists.Therefore, arbitrary line segment in (i)～(iv) judges whether the point on this line segment breaks away from the certified products zone below for example drawing.In addition, for line segment,, be not limited to (i)～(iv) these four kinds as long as have the part that breaks away from the certified products zone when a plurality of in certified products zones.

(i) connect the line segment of being differentiated for each learning sample of certified products (Fig. 8 (a) (b))

(ii) connect the center (average) and the line segment of being differentiated for the learning sample of certified products (Fig. 9 (a) (b)) differentiated for the learning sample of certified products

(iii) connect each at interval support vector line segment (Figure 10 (a) (b)) of the interval support vector (MSV) among the SVM for example

(iv) connect to be differentiated and be the center (average) of the learning sample of certified products and the support vector at interval line segment (Figure 11 (a) (b)) of the interval support vector among the SVM for example

Herein, Fig. 8～Figure 11 is respectively the key diagram of the line segment of above-mentioned (i)～(iv).Wherein, each figure (a) expression certified products zone is one a situation, and each figure (b) expression certified products zone is two a situation.And, the part that the thick line of the line segment among each figure (b) partly represents to break away from the certified products zone.

In addition, whether the point on the line segment breaks away from the certified products zone, can judge more than or equal to threshold value a plurality of somes affirmation discriminant functions that extract from line segment by (1), also can utilize optimization method to obtain the minimum value of the discriminant function on the line segment, and confirm that this minimum value judges more than or equal to threshold value (size is opposite sometimes according to function) by (2).

In addition, strictly saying, is one even also have the certified products zone, but porose in the zone (Figure 12 (a)) and zone do not produce the possibility that breaks away from for having under the situation of convex form (Figure 12 (b)) yet.Therefore, these situations might be excluded by parameter regulation.In these cases, owing to exist than certified products zone more near the unacceptable product zone of desired value, so the certified products zone that still can not say so suitable.So, think that it is rational by parameter regulation these situations being got rid of.

(3) region parameter of nonparametric discrimination

Figure 13～Figure 15 is the key diagram that the certified products area size that causes of region parameter size of expression nonparametric discrimination changes.Wherein, Figure 13 represents histogram method, and Figure 14 represents Nearest Neighbor Estimates, and Figure 15 represents a class SVM, the bigger situation of each figure (a) expression region parameter, the medium situation of each figure (b) expression region parameter, the less situation of each figure (c) expression region parameter.

The discriminant function of nonparametric discrimination has the region parameter as the function on the basis of density function.Specifically, in histogram method the volume V of hypercube, in Nearest Neighbor Estimates the radius r of hypersphere, in a class SVM, be the width cs of gaussian kernel.

If have that the zone of basis function becomes big then the certified products zone also becomes big character.Therefore, when basis function regional too small, produce the excessive risk raising that then produces second kind of mistake (differentiate be the mistake of certified products with unacceptable product) of error of the first kind (certified products are differentiated mistake for unacceptable product).And in the zone of basis function hour, the certified products zone is divided into a plurality of, then forms single zone easily greatly.

(adjusting of region parameter)

Shown in Figure 26 (b), the certified products zone that the discriminant function that utilization obtains from limited sample learning constitutes (study certified products zone) is in fact different with real certified products zone.When the differing greatly of real certified products zone and study certified products zone, the risk that mistake is differentiated improves.Therefore, in order to make study certified products zone, carry out the adjusting of region parameter near real certified products zone.

When making study certified products zone become big adjusting, the probability that the unacceptable product mistake is differentiated for certified products increases.On the other hand, when the adjusting that study certified products zone is diminished, the probability that the certified products mistake is differentiated for unacceptable product increases.

Under the situation about the priori in real certified products zone not, the control band parameter makes expectation wrong differentiation rate (expected error rate) (probability that occurs unknown sample in the zone that mistake is differentiated) minimum.And more after a little while, the reliability of performance evaluation is low at learning sample.Therefore, in the adjusting of region parameter, mistake differentiation rate is can not abundant minimized possibility bigger.

In addition, for the control band parameter, can use the part of resulting sample be used to differentiate the evaluation of performance, remaining method that is used to learn (cross validation, leave-one-out method).

[apparatus structure]

As shown in Figure 2, testing fixture 100 by amplifier 104 amplify from check object 101 contact/near the microphone 102 that disposes and the signal of acceleration adapter 103.Then, import after being converted to numerical data by AD converter 105.And, though the diagram of omission, but after also can and producing beginning in batches in the batch advanced development, obtain action timing and other data in the production scene from administering the actual PLC (Programmable Logic Controller, programmable logic controller (PLC)) that carries out the control of workpiece (product) manufacturing.Omit diagram about these.And testing fixture 100 is obtained speech data that collects by microphone 102 and the Wave data based on vibration data that collects by acceleration adapter 103, extracts characteristic quantity, and carries out abnormality juding.

In addition, testing fixture 100 is made of the computing machine with input media 100b such as CPU main body 100a, keyboard, mouse and display 100c.And, as required, also can constitute and have external memory, or have communication function, can obtain necessary information with the database communication of outside.

In addition, carry out in the rudimentary algorithm of abnormality detection, be created on the judgement knowledge of using when carrying out abnormality juding according to normal sample, with the qualified certified products that are judged to be, with the ineligible unacceptable product that is judged to be at testing fixture 100.Said abnormality detection refers to the qualified certified products that are judged to be, with the ineligible unacceptable product that is judged to be.

Fig. 1 is the functional block diagram of the structure of the discriminant function determination portion 20 that has of expression testing fixture 100.

Discriminant function determination portion 20 is determined the discriminant function that uses in checking the condition discrimination of object.Specifically, judge whether the discriminant function that uses has marked and drawed the single certified products zone of formation in the input space of sample in nonparametric one class is differentiated.And, if the certified products zone is not single, setting regions parameter in the input space of having marked and drawed sample then, make this discriminant function in the input space of having marked and drawed sample, form single certified products zone, this region parameter regulation discriminant function, and regulation as the area size of the basis function on the basis of density function.Region parameter is the parameter of the area size of regulation basis function.Basis function regulation discriminant function.

Therefore, discriminant function determination portion 20 constitute have learning sample storage part 21, learning sample obtaining section 22, parameter candidate's storage part 23, parameter setting portion (parameter setting unit) 24, discriminant function study portion 25, certified products region quantity detection unit (discriminant function identifying unit) 26, discriminant function efferent 27, discriminant function storage part 28.

When study, learning sample obtaining section 22 is stored in the learning sample in the learning sample storage part 21 when obtaining study, export to discriminant function study portion 25.

Parameter setting portion 24 setting regions parameters, make this discriminant function form single certified products zone in the input space of having marked and drawed sample, this region parameter is defined in discriminant function that nonparametric one class uses in differentiating and the regulation area size as the basis function on the basis of density function.Region parameter regulation becomes the area size of basis function on the basis of density function.Density function is defined in the discriminant function that uses in the differentiation of nonparametric one class.In addition, region parameter also can be selected among the parameter candidate from be stored in parameter candidate storage part 23 in advance.

The region parameter that learning sample that discriminant function study portion 25 use learning sample obtaining sections 22 are obtained and parameter setting portion 24 are selected generates discriminant function.For example, the differentiation algorithm of differentiating in nonparametric one class is the learning sample on the class support vector machines identification face stored in the model when being the nonparametric one class discrimination model of feature, to generate the discriminant function corresponding to the face of identification.

Certified products region quantity detection unit 26 judges whether the discriminant function that uses has marked and drawed the single certified products zone of formation in the input space of sample in nonparametric one class is differentiated.The identical discriminant function of discriminant function that uses when in addition, in this is judged, using with inspection.

Specifically, at first, whether certified products region quantity detection unit 26 is from breaking away from the line segment that certified products when zone use to be chosen in judgement the lower line segment.

(i) connect the line segment of being differentiated for each learning sample of certified products (Fig. 8 (a) (b))

(ii) connect the center (average) and the line segment of being differentiated for the learning sample of certified products (Fig. 9 (a) (b)) differentiated for the learning sample of certified products

(iii) connect each at interval support vector line segment (Figure 10 (a) (b)) of the interval support vector (MSV) among the SVM for example

(iv) connect to be differentiated and be the center (average) of the learning sample of certified products and the support vector at interval line segment (Figure 11 (a) (b)) of the interval support vector among the SVM for example

Then, certified products region quantity detection unit 26 is when judging whether line segment breaks away from the certified products zone, and (1) extracts one or more points from line segment, with the above-mentioned discriminant function of each substitution, judges whether all points are differentiated and is certified products; (2) or, obtain the minimum value of the above-mentioned discriminant function on the line segment, judge that this minimum value is whether more than or equal to the threshold value of regulation.

In addition, the user can utilize input media 100b to select to use the line segment of which kind of type, also can preestablish the line segment that uses which kind of type.And, the user can use input media 100b to select, also can preestablish for a plurality of points that extract from line segment is to confirm that whether discriminant function is more than or equal to threshold value, still utilize optimization method to obtain the minimum value of the discriminant function on the line segment, and judge whether this minimum value is confirmed more than or equal to the threshold value (size is opposite sometimes according to function) of regulation.

At the discriminant function that is judged to be 25 generations of discriminant function study portion by certified products region quantity detection unit 26 is that discriminant function efferent 27 is stored in this discriminant function in the discriminant function storage part 28 when forming the discriminant function in single certified products zone.

With reference to Figure 16, the data that discriminant function determination portion 20 is used are described.

Figure 16 (a) is the key diagram that expression is stored in the data structure of the learning sample in the learning sample storage part 21.Shown in Figure 16 (a), learning sample and the sample ID (ID#), the certified products that are used for recognition sample still be the classification (Class) of unacceptable product and characteristic quantity (x1, x2 ...) be stored in explicitly in the learning sample storage part 21.

Figure 16 (b) is the example that expression is stored in the parameter candidate in the parameter candidate storage part 23.In Figure 16 (b),, the candidate of the width cs of gaussian kernel is shown as the example of a class SVM.

Figure 16 (c) is the example that is stored in the discriminant function in the discriminant function storage part 28.Shown in Figure 16 (c), discriminant function is by following rule predetermining, that is, the value of the formula that will use in differentiation (f (x)) and threshold value (O) compare differentiates certified products/unacceptable product.

Below, the processing of discriminant function determination portion 20 definite discriminant functions that testing fixture 100 has is described.

Figure 17 is that the discriminant function of the discriminant function determination portion 20 that has of expression testing fixture 100 is determined the process flow diagram handled.

At first, learning sample acquisition unit 22 obtains learning sample from learning sample storage part 21, and exports to discriminant function study portion 25 (S1).

Then, parameter setting portion 24 obtains a candidate's region parameter from parameter candidate storage part 23, and exports to discriminant function study portion 25 (S2).

Then, discriminant function study portion 25 uses learning sample of being imported by learning sample obtaining section 22 and the region parameter of being imported by parameter setting portion 24, and a class discriminant function is learnt (S3).

Then, certified products region quantity detection unit 26 judges whether the discriminant function that generates as the learning outcome of discriminant function study portion 25 forms single certified products zone (S4).This discriminant function generates as the learning outcome of discriminant function study portion 25.And, the result of determination of certified products region quantity detection unit 26 be the certified products zone more than or equal to 2 o'clock, return step S2, begin to repeat from the processing of selecting region parameter.On the other hand, result of determination at certified products region quantity detection unit 26 is 1 o'clock for the certified products zone, discriminant function efferent 27 is learnt the discriminant function of portion's 25 generations, the discriminant function that promptly utilizes the region in front parameter learning to obtain with discriminant function, is stored in the discriminant function storage part 28.

At this, be desirably in the region parameter of setting among the step S2 and be divided into a plurality of enough little parameter candidates from the zone, set bigger parameter successively.This is because a plurality of when making certified products zones for the region parameter of single area when existing, will adopt minimum wherein one.Therefore, can prevent to make second kind of mistake (mistake that unacceptable product is judged to be certified products is differentiated) increase (Figure 13 (a), Figure 14 (a), Figure 15 (a)) owing to region parameter is excessive.

On the contrary, the region parameter of setting in step S2 also can be one enough big candidate parameter from the certified products zone really, sets less parameter successively, adopts region quantity to become the previous parameter of the parameter more than 2.

Perhaps, the region parameter of setting in step S2 also can be used all parameter candidates, exports a plurality of certified products zones that make and is the region parameter of single area.When this situation, can decide the parameter of actual inspection by the people, also can be chosen in the parameter of using in the inspection automatically according to certain standard.As the standard of selecting, expectation is selected minimum from the parameter of being exported.

Figure 18 is the process flow diagram of the certified products region quantity determination processing (S4 among Figure 17) of expression certified products region quantity detection unit 26.

Whether at first, certified products region quantity detection unit 26 generates line segment, for unacknowledged line segment, confirm not break away from (S11).In addition, will narrate in the back about the concrete condition of step S11.

Then, when there is disengaging in line segment (S12, "Yes"), is judged to be the certified products zone and has a plurality of zones.On the other hand, when all line segments all do not break away from (S12 "No" and S13 "Yes"), be judged to be the certified products zone for single.

Figure 19 is in the certified products region quantity determination processing (S4 among Figure 17, the S11 among Figure 18) that is illustrated in certified products region quantity detection unit 26, confirms the process flow diagram of discriminant function more than or equal to the step of threshold value for a plurality of points that extract from line segment.Figure 20 is the key diagram of this step of explanation.

At first, certified products region quantity detection unit 26 extracts limited point (Figure 20 (a) (b) * mark) (S21) from line segment.

Then, for extract have a few, utilize the discriminant function identical to judge whether qualified (S22) with the discriminant function that in inspection, uses.And,, also be judged to be and have disengaging even differentiated when only having 1 to be differentiated (S23 "Yes", Figure 20 (b)) for underproof.On the other hand, be underproof point (S23 "No", Figure 20 (a)), then be not judged to be not disengaging if differentiate.

In addition, when point that extracts from line segment and quilt are differentiated for the learning sample of certified products (perhaps interval support vector, for example the interval support vector among the SVM) unanimity, also can omit above-mentioned judgement.In addition, an end of line segment during as the center differentiated for the learning sample of certified products, at the point that extracts from line segment and differentiated and to be the center of the learning sample of certified products when consistent, also can only judge once initial, in the disengaging judgement of each line segment, do not judge separately.

At this, Figure 21 is illustrated in according to step shown in Figure 19 to judge when breaking away from, the key diagram of the concrete example of the data that generate when extracting from line segment.In Figure 21, the example when showing with the line segment quartern.In addition, these data are generated by certified products region quantity detection unit 26, and are stored in the storer (not shown) of discriminant function determination portion 21.

5 points that comprise two ends and the example of registering the situation of coordinate are extracted in Figure 21 (a) expression.Shown in Figure 21 (a), continuous sequence number is put in the extraction of in the data of extracting point, be used to discern the line segment ID (line segment ID#) of line segment, giving continuous sequence number to each line segment and the corresponding characteristic quantity of coordinate that generates with five equilibrium (x1, x2 ...) be associated.Like this, in Figure 21 (a), extract 5 points that comprise two ends and register coordinate.In addition, as line segment ID, the line segment ID that can use the sample ID that connects two end samples to form.

The example of the data conditions at registration line segment two ends is omitted in Figure 21 (b) expression.When promptly being judged as the judgement of learning sample of certified products at the two ends of omitting line segment, certified products region quantity detection unit 26 is these data of registration in storer not also.

One end of Figure 21 (c) expression line segment is differentiated the situation example for the center of the learning sample of certified products.When an end of line segment was differentiated center for the learning sample of certified products, certified products region quantity detection unit 26 can not be registered in it in extraction point of each line segment yet, and center and line segment are registered separately.

Figure 22 is illustrated in the certified products region quantity determination processing of certified products region quantity detection unit 26, utilize optimization method to obtain the minimum value of the discriminant function on the line segment, confirm the process flow diagram of this minimum value more than or equal to the step of threshold value (size is opposite sometimes according to function).And Figure 23 is the key diagram of this step of explanation.

At first, the discriminant score obtained on the line segment of certified products region quantity detection unit 26 is minimum point (S31).

Then, when certified products region quantity detection unit 26 is negative at the discriminant score as smallest point (=be judged as defective) (S32 "Yes"), be judged to be existence and break away from.On the other hand, as the discriminant score of smallest point when negative (S32 "No"), be not judged to be and break away from.

At this, as shown in figure 23, the value of the discriminant function on the line segment becomes the non-linear continuous function of the location parameter t (t is the function of x) on the line segment.Therefore, as long as in the scope of line segment, utilize nonlinear optimization methods such as Newton method and method of steepest descent to obtain the minimum value of discriminant function, certified products region quantity detection unit 26 confirms that its threshold value less than regulation gets final product.For example, it is 0 o'clock in the threshold value of discriminant function, if the minimum value of discriminant function is more than or equal to 0 then be judged to be qualified (Figure 23 (a)), if the minimum value of discriminant function is for negative then be judged to be defective (Figure 23 (b)).

In addition, above-mentioned testing fixture 100 can be applied to the inspection field of extraordinary noise, loading error, output characteristics.And online (in-line) that can be applied to produce in batches check, the off-line (off-line) of inspection etc. that also can be applied to be different from the trial product of batch process is checked.More specifically, above-mentioned testing fixture 100 for example can be as the checkout facility of the driver module of the engine (sound) of automobile, variator automobiles such as (vibrations), and the checkout facility of the motor actuator module of electronic peephole, automatic seat, electric steering wheel automobiles such as (turning device position adjustings), and even the extraordinary noise in the above-mentioned exploitation, loading error, the evaluating apparatus of output characteristics and the evaluating apparatus of the trial-production machine in the development phase.

In addition, can be used as the checkout facility of direct motor drive household electrical appliances such as refrigerator, indoor apparatus of air conditioner and off-premises station, washing machine, suction cleaner, printer, and the evaluating apparatus of the extraordinary noise in the above-mentioned exploitation, loading error, output characteristics.In addition, can also be as the Device Diagnostic machine of the condition discrimination (abnormality/normal condition) that carries out equipment such as NC process equipment, semiconductor equipment, food apparatus.

In embodiments of the present invention, put down in writing the content of the testing fixture that applies the present invention to judge certified products, unacceptable product.But, so long as satisfy the data of following condition (1)～(3), and use nonparametric one class to differentiate specific region and zone in addition, then sample can be any sample.

(1) has the sample of the data of desired value.

(2) a near group's of formation sample desired value.

(3) because external causes such as environment make data the sample of deviation occur.

In addition, in embodiments of the present invention, the differentiation algorithm of having put down in writing the differentiation of nonparametric one class is the situation of a class SVM.But so long as the learning sample on the identification face is stored in nonparametric one class discrimination model in the model, then differentiating algorithm can be any differentiation algorithm.

In addition, as long as " support vector " of record is " being stored in the learning sample in the model " in embodiments of the present invention, then can be any sample.

In addition, as long as " support vector at interval " is " being stored in the learning sample on the identification face in the model ", then can be any sample.

The invention is not restricted to above-mentioned embodiment, can in the scope of claims record, carry out various changes.That is, the embodiment by the technological means that suitably changes formation in the scope that is combined in claims record obtains also is contained in the technical scope of the present invention.

Testing fixture of the present invention is to differentiate by nonparametric one class to judge whether qualified testing fixture, also can constitute to have to confirm that when study the certified products zone is one unit.

In addition, testing fixture of the present invention is to differentiate by nonparametric one class to judge whether qualified testing fixture, also can constitute to have when study and regulate parameter so that the certified products zone is one unit.

In addition, the differentiation algorithm of above-mentioned nonparametric one class of testing fixture of the present invention differentiation also can be a class SVM.

In addition, in testing fixture of the present invention, above-mentioned affirmation certified products zone is that one unit can be the unit that the line segment that confirm to connect each learning sample that is judged as certified products does not break away from the certified products zone.

In addition, in testing fixture of the present invention, above-mentioned affirmation certified products zone is that one unit also can be the unit that the center of affirmation connectionist learning sample and the line segment that is judged as the learning sample of certified products do not break away from the certified products zone.And it must be the sample that is judged as certified products that the learning sample that uses when obtaining the center does not need.

In addition, in testing fixture of the present invention, above-mentioned affirmation certified products zone is that one unit also can be to confirm that the line segment of each interval support vector (MSV) among each interval support vector of connection, for example SVM does not break away from the unit in certified products zone.

In addition, in testing fixture of the present invention, above-mentioned affirmation certified products zone be one unit also can be confirm the center of connectionist learning sample and support vector at interval for example the line segment of the interval support vector (MSV) among the SVM do not break away from the unit in certified products zone.

In addition, in testing fixture of the present invention, the unit that above-mentioned affirmation line segment does not break away from the certified products zone also can be to judge that all points are the unit of certified products when differentiating from limited point that line segment extracts.

In addition, in testing fixture of the present invention, the unit that above-mentioned affirmation line segment does not break away from the certified products zone also can be to utilize optimization method to obtain the minimum value of the discriminant function on the line segment, confirms the unit of this minimum value more than or equal to threshold value (size is opposite sometimes according to function).

At last, the learning sample obtaining section 22 that each functional block of testing fixture 100, particularly discriminant function determination portion 20 have, parameter setting portion 24, discriminant function study portion 25, certified products region quantity detection unit 26 and discriminant function efferent 27 can utilize hardware logic to constitute, and also can utilize software to realize according to following described use CPU.

Promptly, testing fixture 100 has CPU (the central processing unit of the order of carrying out the control program be used to realize various functions, CPU (central processing unit)), stored ROM (the read only memory of said procedure, ROM (read-only memory)), memory storages (recording medium) such as the RAM (randomaccess memory, random access memory) that said procedure is launched, storage said procedure and various memory of data etc.And, purpose of the present invention can realize in the following manner, promptly, the recording medium that software computer-readable with having write down, that be used to realize above-mentioned functions is the program code (execute form program, intermediate code program, source program) of the control program of testing fixture 100 offers above-mentioned testing fixture 100, and the program code of this computing machine (or CPU or MPU) playback record in recording medium also carried out.

As aforementioned recording medium, for example, can use tape series such as tape and cartridge tape, the semiconductor memory series of the card series of dish series of CD such as the disks such as (floppy, registered trademark)/hard disk that comprise floppy disk or CD-ROM/MO/MD/DVD/CD-R, IC-card (comprising storage card)/light-card etc. or MASKROM/EPROM/EEPROM/FLASH ROM etc. etc.

In addition, testing fixture 100 also can constitute and can be connected with communication network, provides the said procedure code by communication network.Be not particularly limited as this communication network, for example, can use the Internet, in-house network, extranets, LAN, ISDN, VAN, CATV communication network, Virtual Private Network (virtual private network), telephone wire road network, mobile radio communication, satellite communication link etc.And, be not particularly limited as the transmission medium that constitutes communication network, for example, can adopt wired modes such as IEEE1394, USB, power line transmission, catv line, telephone wire, adsl line, or wireless mode such as the infrared ray as IrDA and telepilot, Bluetooth (bluetooth, registered trademark), 802.11 wireless, HDR, mobile telephone network, satellite circuit, ground wave digital network.In addition, the present invention also can be according to form that implement, that embed the computer data signal in the carrier wave realizes by the electric transmission of said procedure code.

Testing fixture of the present invention can be determined the discriminant function of use in nonparametric one class is differentiated, make the single certified products zone of formation in the input space of having marked and drawed sample, therefore can be widely used in the testing fixture in the manufacturing line and the evaluating apparatus of machine action.That is, except that product examination, also can be used for the fault detect of manufacturing installation and power equipment etc., people's health condition judging etc.

Claims (13)

1. testing fixture, it is characterized in that, this testing fixture has the discriminant function identifying unit, and this discriminant function identifying unit is judged whether the discriminant function that uses has been marked and drawed in the input space of sample to form as single area and comprised the zone that is judged to be clipped to the sample in the class in nonparametric one class is differentiated.

2. testing fixture according to claim 1, it is characterized in that, this testing fixture has the parameter setting unit, this parameter setting unit is being judged to be above-mentioned discriminant function when not forming above-mentioned single area by above-mentioned discriminant function identifying unit, the setting regions parameter, make this discriminant function form single area in the input space of having marked and drawed sample, wherein this region parameter is stipulated above-mentioned discriminant function, and regulation is as the area size of the basis function on the basis of density function.

3. testing fixture according to claim 1 is characterized in that, the differentiation algorithm that above-mentioned nonparametric one class is differentiated with the sample storage on the identification face in model.

4. according to each described testing fixture in the claim 1～3, it is characterized in that above-mentioned discriminant function identifying unit judges that connection is judged as the line segment that is included in each sample in the above-mentioned single area and whether breaks away from above-mentioned single area.

5. testing fixture according to claim 1 is characterized in that, above-mentioned discriminant function identifying unit is judged the center that connects sample and be judged as the line segment that is included in the sample in the above-mentioned single area whether break away from above-mentioned single area.

6. testing fixture according to claim 3 is characterized in that, above-mentioned discriminant function identifying unit judges whether the line segment of each sample on the identification face of storing in the link model breaks away from above-mentioned single area.

7. testing fixture according to claim 3 is characterized in that, whether the line segment of the sample on the identification face of storing in the center of above-mentioned discriminant function identifying unit judgement connection sample and the model breaks away from above-mentioned single area.

8. testing fixture according to claim 4, it is characterized in that, above-mentioned discriminant function identifying unit extracts one or more points from above-mentioned line segment, with the above-mentioned discriminant function of each substitution, judge whether all points are all differentiated for being included in the above-mentioned single area.

9. testing fixture according to claim 4 is characterized in that above-mentioned discriminant function identifying unit is obtained the minimum value of the above-mentioned discriminant function on the above-mentioned line segment, judges that whether this minimum value is more than or equal to the threshold value of stipulating.

10. inspection method, its testing fixture by the discriminant function of determining to use in the condition discrimination of checking object is carried out, it is characterized in that,

Discriminant function identifying unit that above-mentioned testing fixture had is judged whether the discriminant function that uses has been marked and drawed in the input space of sample to form as single area and is comprised the zone that is judged to be clipped to the sample in the class in nonparametric one class is differentiated.

11. inspection method according to claim 10 is characterized in that,

When being judged to be above-mentioned discriminant function by above-mentioned discriminant function identifying unit when not forming above-mentioned single area, the parameter setting unit setting regions parameter that above-mentioned testing fixture had, make that formation comprises the zone that is judged to be clipped to the sample in the class to this discriminant function as single area in the input space of having marked and drawed sample, wherein this region parameter is defined in the discriminant function that uses in the differentiation of nonparametric one class, and regulation is as the area size of the basis function on the basis of density function.

12. a scrutiny program, it is used for utilizing computer realization to determine the testing fixture of the discriminant function that uses at the condition discrimination of checking object, and this computing machine is played a role, it is characterized in that,

This scrutiny program makes computing machine carry out following the processing:

Discriminant function identifying unit that above-mentioned testing fixture had is judged whether the discriminant function that uses has been marked and drawed in the input space of sample to form as single area and is comprised the zone that is judged to be clipped to the sample in the class in nonparametric one class is differentiated.

13. computer readable recording medium storing program for performing that writes down the described scrutiny program of claim 12.

Images