TWI434229B - Cluster system and defect type determination device - Google Patents

Description

Cluster system and defect type determining device

The present invention relates to a cluster system for removing a defective portion of an image of a detection object, a feature signal for extracting a defect from the partial image, and a classification system and a defect type determination device classified according to the type of the defect.

The distance between unknown data and learning materials, such as the clustering technique produced by the Mahalanobis generalized distance, has been conventionally performed. That is, the clustering process is performed by determining whether or not the unknown material belongs to a cluster as a parent group that has been learned beforehand. For example, by judging the size of the majority cluster by the Mahalanobis distance, it is determined which cluster of the parent group the unknown data belongs to (for example, refer to Patent Document 1).

Furthermore, in order to efficiently calculate the above distance, a plurality of feature quantities are selected to perform the process of clustering.

Furthermore, based on the votes obtained by the majority of the classifiers, the method of determining the cluster decision to which the unknown data belongs is also generally visible, using the recognition result of the output of the different sensors, or on a portrait. The different fields identify the recognition result of the unknown data and the like (for example, refer to Patent Document 2).

By the above clustering method, the disease diagnosis is performed on the parameters obtained according to the results of the blood test, that is, which kind of condition belongs to the clustering, which is set in each of the plurality of clusters to set each of the two clusters as a combination, for all group Each of the combinations is determined to determine whether or not the detected data is similar to the cluster, and the statistical result of the number of determinations determines the method of classifying the cluster into a determined number (for example, refer to Patent Document 3).

When each defect attached to the LCD glass substrate is classified according to each defect type set in advance, the optimization of each feature amount used for classification is performed corresponding to the identification at the time of classification to correspond to In the optimization method, weighting is performed for each feature amount, and clustering of which cluster belongs is determined using the optimized feature amount (for example, Patent Document 4).

Patent Document 1: Japanese Laid-Open Patent Publication No. 2005-214682

Patent Document 2: Japanese Laid-Open Patent Publication No. 2001-56861

Patent Document 3: Japanese Laid-Open Patent Publication No. 07-105166

Patent Document 4: Japanese Laid-Open Patent Publication No. 2002-99916

However, in the clustering shown in Patent Document 3, the optimization of each combination is not performed, and the feature quantity that becomes the discrimination material is not effectively utilized, and when the cluster to be discriminated is increased, the number of combinations is large, and the determination processing is required. The problem of increasing time.

Further, in the clustering shown in Patent Document 4, the weight of the feature amount is increased based on the determination rate, and the discrimination accuracy is improved. However, the concept of optimizing the feature amount of each cluster is not the same as in Patent Document 3 described above. Since the feature amount cannot be effectively utilized, there is a disadvantage that the classification of high precision cannot be performed.

The present invention has been made in view of such a problem, and provides a feature amount that can be effectively used for discriminating the classification target data classified into objects belonging to the cluster, and provides higher speed and higher than the conventional example. Accuracy classifies the classification target data, for example, a defect system on the glass surface can be classified into a cluster system corresponding to the cluster of defect types, and a defect type determination device.

In order to solve the above problems, the present invention differs from the conventional example in which the classification target data and the clusters are calculated by the same type of feature quantity, and the classification target is determined, and each cluster setting can be obtained between clusters. The set of the feature quantities of the difference, and the distance is obtained by different feature quantities between the clusters, so the classification with higher precision is performed than in the past.

Since the set of the above-described feature quantities is executed based on the characteristics of the learning materials belonging to each cluster, it is composed of feature quantities that can be distinguished from other clusters.

That is, the present invention adopts the following constitution.

The cluster system of the present invention belongs to a cluster in which input data is classified into a cluster formed by a population of learning data by a parameter of the input data. The system is characterized in that: the feature quantity set storage unit is a parameter set corresponding to each cluster and memorizing a combination of feature amounts used for classification; a feature quantity extracting unit, The feature amount extracting unit extracts a feature amount set in advance from the input data, and a distance calculating unit that corresponds to each Each feature quantity set of the cluster calculates the distance between the center of the parent group of each cluster and the input data according to the feature quantity contained in the feature quantity set, and outputs the distance as the set distance; and the order extraction unit, The order extracting unit arranges the respective set distances from the smallest to the largest.

In the preferred clustering system of the present invention, a plurality of the above-described feature amount sets are set in each cluster.

The optimal clustering system of the present invention further has a cluster classification unit that classifies the clusters into the clusters by input data indicating the order of the set distances in the set distances obtained in each of the feature amount sets. The rule mode of the classification criterion detects which cluster the input data belongs to.

In the optimal clustering system of the present invention, the cluster classification unit detects, by the rank of the set distance, which cluster the input data belongs to, and sets the rank as a cluster of the upper set distance as the input data. It is detected by the cluster it belongs to.

In the optimum clustering system of the present invention, the cluster classification unit has a threshold value that is higher than the rank, and if the upper cluster is equal to or greater than the threshold value, it is detected as a cluster to which the input data belongs.

In the optimum clustering system of the present invention, the distance calculating unit multiplies the set distance by the correction coefficient set corresponding to the feature amount set, and normalizes the collective distance between the feature amount sets.

The optimal clustering system of the present invention further has a feature quantity set cooperation unit for creating a feature quantity set of each cluster, and the feature quantity set cooperation part is a needle For each of the majority of the combinations of the feature quantities, the average of the learning materials of the parent group of each cluster is taken as the origin, and the average of the distances of the learning materials of the origin and the parent group of the other clusters is obtained. And will be the combination of the feature quantities of the largest average, selected as the set of feature quantities used to distinguish each cluster from other clusters.

The defect type determining apparatus according to the present invention is provided with any one of the cluster systems described above, wherein the input data is an image data of a product defect, and the defect in the image data is classified according to the type of the defect by indicating the feature amount of the defect. sort.

In the optimum defect type determining device of the present invention, the product is a glass article, and the defects of the glass article are classified according to the type of the defect.

The defect detecting device of the present invention is provided with the defect type determining device described above for detecting the type of defect of the product.

The manufacturing state determination device according to the present invention is provided with the defect type determination device described above, and performs classification of defects of the product, and detects the cause of the defect in the process based on the correspondence with the cause of occurrence of the category.

The optimum manufacturing state judging device of the present invention is provided with any one of the cluster systems described above, wherein the input data is a feature amount indicating a manufacturing condition in a process of the product, and the manufacturing state is in accordance with the manufacturing state of each process of the process. The feature quantities are classified.

In the optimum manufacturing state judging device of the present invention, the product is a glass article, and the feature amount in the process of the glass article is classified according to the manufacturing state of each process of the process.

The manufacturing state detecting device of the present invention is provided with the manufacturing state determining device as described above, and detects the type of manufacturing state in each of the processes of the product process.

The product manufacturing management device of the present invention is provided with the manufacturing state determining device as described above, and performs a process for detecting a manufacturing state in each process of the product manufacturing process, and performs a process in the process engineering according to a control item corresponding to the type. control.

As described above, according to the present invention, since each cluster of the classification target has a plurality of feature quantities having the classification target data, the combination of the optimal feature amounts that are distant from the other clusters is set in advance, and the classification is respectively calculated. The distance between the object data and each cluster is classified into the cluster with the smallest distance calculated, so that the classification object data can be classified into the corresponding cluster more correctly than in the past.

Furthermore, according to the present invention, since a plurality of the above combinations are set for each cluster, the distances of the calculation results of the full cluster and the classified object data are arranged in ascending order, and the classified object data is classified into the previously set number. The upper cluster has the largest number of clusters, so it can perform high-precision classifications compared to the past.

The cluster system of the present invention relates to classifying input data of a classification object into learning materials as a parent set by using the feature quantity of the input data. The cluster system of each cluster formed by the group has a feature quantity set storage unit for storing a feature quantity set corresponding to the combination of the feature quantities used for classification in accordance with each of the above clusters, and the feature quantity extracting unit is set in advance according to The feature quantity set extracts the feature quantity from the input data, and the distance calculation unit pairs the parent group and the input data according to the feature quantity included in the feature quantity set for each feature quantity corresponding to each of the cluster sets. The set distance is calculated as a set distance, and the set extracting units are arranged in the order of small to large, and the clustering is performed corresponding to the arrangement order.

[First Embodiment]

Hereinafter, a cluster system according to a first embodiment of the present invention will be described with reference to the drawings. Fig. 1 is a block diagram showing an example of the configuration of a cluster system according to the same embodiment.

The cluster system of the present embodiment has a feature amount set cooperation unit 1, a feature amount extracting unit 2, a distance calculating unit 3, a feature amount set storage unit 4, and a cluster database 5 as shown in Fig. 1 .

The feature amount set storage unit 4 stores, in association with the identification information of each cluster, a feature amount set that is set in a combination of the feature amounts of the classification target data for each cluster. For example, when the classification object data is a set of feature quantities {a, b, c, d}, the feature quantity set of each cluster is treated as [a, b], [a, b, c, d], [c] The combination of the feature quantities of the types. In the following description, any combination of the feature amounts and a plurality of combinations (in the above example, any two of the sets, three feature amounts) are defined as "feature quantities". combination".

Here, when the clusters A, B, and C are set as the clusters of the classification targets, the learning materials classified in the clusters are used in advance to obtain the distance between the clusters and the other clusters. The combination of the feature amounts is stored in the feature amount set storage unit 4.

For example, the feature quantity set set for the cluster A is a vector composed of the average values of the feature amounts of the learning materials belonging to the cluster A, and an average value of the respective feature quantities of the learning materials belonging to the other clusters B and C. The distance of the constituent vectors is set to the combination of the largest feature quantities.

Furthermore, the classification object data and the learning materials of the parent group in each cluster are composed of the same set of feature quantities.

When the distance between the clusters is calculated from the input classification target data, the feature amount extracting unit 2 reads out the feature amount set corresponding to the cluster to be calculated from the feature amount set storage unit 4, and extracts the feature amount of the classification target data. The extracted feature amount is output to the distance calculating unit 3 in accordance with the feature amount of the feature amount combination.

The distance calculation unit 3 is a self-cluster database 5, and uses the identification information of the cluster of calculation objects as a key, and reads a vector composed of the average values of the feature amounts of the learning data of the cluster to be calculated, according to the cluster. The feature quantity set calculates a vector formed by the feature quantity extracted by the classification target data, and a distance formed by a vector composed of an average value of each feature quantity of the learning material (a center of gravity vector indicating a position of a center of gravity of a plurality of learning materials in the cluster) .

In the calculation of the above-described distance, the distance calculating unit 3 normalizes the numerical value of the feature quantity so that the data unit of the feature quantity does not differ, and the feature quantity v(i) of the classification target data by the following formula (1) Execution positive Regulation.

V(i)=(v(i)-avg.(i))/std.(i)...(1)

Here, v(i) is the feature quantity, avg.(i) is the average value of the feature quantity in the learning material in the cluster of the calculation object, and std.(i) is the feature in the learning material in the cluster of the calculation object. The standard deviation of the quantity, V(i) is the characteristic quantity normalized. Therefore, when calculating the distance, the distance calculation unit 3 must perform normalization of each feature amount for each feature amount set.

Further, the distance calculating unit 3 executes the above-described normalization processing using the average value and the standard deviation of the feature amounts corresponding to the respective learning materials for each feature amount for calculating the distance in the classification target data.

Further, as the distance, a standardized Euclidean distance, a Mahalanobis distance, a Minkowskiy distance, or the like using the above-described standardized feature quantity is used. Either one can be.

Here, when the Mahalanobis distance is used, the Mahalanobis squared distance MHD is obtained by the following formula (2).

MHD=(1/n)‧(V ^T R ^-1 V)...(2)

Each element V(i) in the matrix V in the above formula (2) is a feature quantity v(i) of a plurality of elements of unknown data, and an average value avg. of the feature quantity of the learning data in the cluster. i) and the standard deviation std. (i), the feature quantity obtained by the above formula (1). n is a degree of freedom. In the present embodiment, the number of feature quantities indicating the number of feature quantities in the feature amount set (described later) is shown. Thereby, the Mahalanobis squared distance is a value obtained by adding the difference of the n transformed feature quantities, and the unit distance of the parent group average becomes 1 by (Marananobis squared distance) /n. Furthermore, V ^T is a transposed matrix of the matrix V with the feature quantity v(i) as the element, and R ^-1 V is the inverse of the correlation matrix R between the feature quantities in the learning data in the cluster. matrix.

The feature amount set cooperation unit 1 calculates a feature amount set used when the distance calculation unit 3 calculates a distance between the classification target data and each cluster for each cluster, and writes the calculation result to each cluster identification information, and writes the feature to the feature. The quantity is stored in the memory unit 4.

When calculating the feature quantity set, the feature quantity set cooperation part 1 is a barycentric vetor of the learning material belonging to the object cluster of the generated feature quantity set for each cluster, and belongs to all other than the object cluster. The distance of the center of gravity vector of the learning data of the cluster is used as a reference, and the value λ of the discriminant criterion is calculated by the following equation (3). Hereinafter, the combination of the feature amounts will be described as a feature amount set.

λ=ω _o ω _i (μ _o -μ _i ) ² /(ω _o σ _o ² +ω _i σ _i ² )...(3)

In the above (3), μ _i is a centroid vector composed of an average value of the feature quantities in the feature quantity set of the learning material belonging to the target cluster (the parent group in the cluster). σ _i is the standard deviation of the vector generated by the feature quantities of the learning materials belonging to the parent group within the cluster. ω _i is the ratio of the number of learning materials belonging to the parent group within the cluster to the number of learning materials belonging to the full cluster. Further, μ _o is a centroid vector composed of an average value of the feature quantities in the feature quantity set of the learning material (object cluster outer group) belonging to the cluster outside the cluster. σ _o is the standard deviation of the vector generated by the feature quantities of the learning materials belonging to the parent group of the object cluster. Ω _o is the ratio of the number of learning materials belonging to the cluster parent group to the number of learning materials belonging to the full cluster. Here, the formula (3) in the (μ _o -μ _i) even if log (logarithmic), and also the value of the square root. Further, when the vectors are calculated here, the feature amount set cooperation unit 1 calculates the feature amount for which each feature amount is normalized by the equation (1). Further, even if the eigenvalue is set as the value obtained by dividing the ratio ω _i and ω _{o in} advance, the separation becomes large.

Then, the feature amount set cooperation unit 1 calculates the discrimination reference value λ which is discriminated from the other clusters by using the above equation (3) for each of the target clusters and for any or all of the feature amounts constituting the learning material. The calculated discrimination reference value λ is arranged in descending order, and a ranking list of the discrimination reference value λ is output.

Here, the feature quantity set cooperation unit 1 is a combination of the feature quantities corresponding to the maximum discrimination reference value λ as the feature quantity set of the target cluster, and the value of the discrimination reference value λ, and corresponds to the identification information of the cluster, and the memory The feature quantity set storage unit 4.

The determination of the discrimination reference value λ is as shown in Fig. 2(a), and when the feature quantity set cooperation unit 1 sets the feature quantity set of each cluster, the feature quantity of the learning data and the classification target data is a, In the case of four of b, c, and d, the discriminant reference value λ of all of the four feature quantities, all of them, and one of them is calculated.

Then, the feature quantity set cooperation unit 1 selects the highest value, for example, In Fig. 2(a), a combination of feature quantities b and c is selected.

Further, in the method of discriminating the reference value λ , there is a BSS method as described in FIG. 2(b), that is, a discriminant reference value λ for calculating all the feature amounts included in the set of classification target data. Then, a combination of n-1 is extracted from the set of n feature quantities, and the discrimination reference value λ is calculated. Then, the combination of the maximum values is selected from the n-1 discrimination reference values λ, and the reference value λ is calculated from the n-1 feature quantities for all combinations of n-2. In this way, even if the feature quantity is reduced by the self-collection by the order, the set of the feature quantity is reduced, and the combination is reduced by a combination to calculate the reference value λ, and the combination can be determined with less feature quantity. In this manner, the feature quantity set cooperation unit 1 may be configured.

Further, in another method of discriminating the reference value λ, the FSS method as described in FIG. 2(c) is a feature quantity included in the set of self-classification target data. For all types, the discrimination reference value λ of each feature amount is calculated, and the feature amount having the largest discrimination reference value is selected. Next, a combination of the feature amounts and the other feature amounts of the feature amounts other than the above is generated, and the discrimination reference value λ with respect to each combination is calculated. Then, the combination having the largest discrimination reference value is selected from its combination. Next, a combination of the three feature quantities of the combination and the feature amount not included in the combination is generated, and each discrimination reference value λ is generated. In this way, even if the feature quantity having the largest discriminant reference value λ is selected in the order from the combination of the feature quantities in the order, the feature quantity of the combination is increased by one feature amount that does not exist in the combination with respect to the combination, and the calculation is increased. The discriminating reference value λ of the combined feature quantity, and the combination having the largest discriminant reference value λ is selected from the combination, and the discriminant reference value λ of the combination of the feature amounts that are not present in the combined feature quantity is calculated. Finally, it is also possible to form the feature amount set cooperation unit 1 by calculating all combinations in which the reference value λ is determined and the combination in which the discrimination reference value λ is the largest is selected as the feature amount set.

Next, by determining the reference value λ, the validity of the selection of the feature amount set used in the cluster is represented by the third and fourth figures.

Figure 3 is a combination of the feature quantities a and g, the combination of the feature quantities a and h, and the combination of the feature quantities d and e for the feature quantities a, b, c, d, e. The combination of sets, selected from clusters 1, cluster 2, and cluster 3 from these combinations, is described with a set of feature quantities having higher classification characteristics than the conventional examples.

In Fig. 3, μ1 corresponds to μ _i , μ 2 corresponds to μ _o , σ ₁ corresponds to σ _i , σ 2 corresponds to σ _o , ω 1 corresponds to ω _i , and ω 2 corresponds to ω _o .

Wherein, in the above combination, the value of the discrimination reference value λ is at most a combination of the feature quantities a and h, and the combination is used for the separation of the cluster 1 and the other clusters, and the cluster 1 is confirmed by the fourth graph and Classification results for clusters other than clusters (clusters 2 and 3).

In Fig. 4, the horizontal axis represents the value of the log of the Mahalanobis distance calculated using the combination of the feature amounts, and the vertical axis represents the number of the separated object data (histogram) having the corresponding numerical value. Here, the value of 1.4 on the horizontal axis means that the value of the log of the Mahalanobis distance is less than 1.4 and 1.2 or more (the value on the left side of 1.4). The values on the other horizontal axes are also the same. again In Fig. 4, 1.4≦ means 1.4 or more. The Mahalanobis distance of Fig. 4 is calculated using the feature quantity set corresponding to cluster 1 and the classification object data belonging to cluster 1 and other clusters.

Fig. 4(a) shows an example of calculating a Mahalanobis distance using a combination of feature quantities a and g, and Fig. 4(b) is a calculation of Mahalanobis using a combination of feature quantities a and h. In the case of the distance, Fig. 4(c) is an example of calculating the Mahalanobis distance using a combination of the feature quantities d and e. When viewing the histogram in FIG. 4, when the value of the discrimination reference value λ is large, it is understood that the classification of cluster 1 and other clusters is well performed.

Next, the operation of the cluster system according to the embodiment of Fig. 1 will be described with reference to Figs. 5 and 6 . Fig. 5 is a flowchart showing an operation example of the feature quantity set cooperation unit 1 of the cluster according to the first embodiment, and Fig. 6 is a flowchart showing an operation example of the cluster of classification target data.

In the following description, for example, when the classification target data is a collection of the feature amount of the scratch on the glass article, "a: the length of the scratch" which is the feature amount is obtained from the self-image processing or the measurement result, "b: the area of the scratch", "c: the width of the scratch", "d: the transmittance of the specific region containing the scratched portion", and "e: the reflectance of the specific region containing the scratch". Therefore, the set of feature quantities (hereinafter, the feature quantity set) becomes {a, b, c, d, e}. Furthermore, in the present embodiment, the distance used for the clustering is calculated as the Mahalanobis distance using the normalized feature amount. Here, the glass article in the present embodiment may be, for example, a plate glass or a glass substrate for a display.

A. Feature set integration process (corresponding to the flowchart of Figure 5)

The user detects that there is a scratch on the glass, and captures the image to obtain the image data, and performs image processing to extract the feature amount such as the length of the scratched portion from the image data, and collects the feature amount. The characteristic quantity data of the composition. Then, for each cluster to which the user wants to classify the cause or shape of the scratches, the feature amount data is divided into learning materials according to the information of the cause or shape known in advance, and is regarded as the parent group of the learning materials of the clusters. The processing terminal from the non-graphic processing corresponds to the identification information of the cluster, and is memorized to the cluster database 5 (step S1).

Next, when the feature amount set cooperation unit 1 inputs a control command for generating a feature amount set from each cluster from the cluster, the cluster data library 5 reads the identification information of each cluster and reads the parent group of the learning materials.

Then, the feature quantity set cooperation unit 1 calculates the average value and the standard deviation of each feature quantity in the parent group in the cluster for each cluster, and calculates the specifications in each learning material from the formula (1) using the average value and the standard deviation. The amount of features.

Next, the feature amount set cooperation unit 1 is for each feature amount of all combinations of the feature amounts included in each feature amount set, and the discrimination reference value λ is calculated by the equation (3).

In this case, the feature quantity set cooperation unit 1 calculates the normalized feature quantity of the parent group in the cluster for each cluster, and calculates the average value (center of gravity vector) μ _i of the vectors composed of the feature amounts corresponding to the respective feature amount sets. The normalized deviation σ _i of the vector of the learning material formed by the feature quantity corresponding to the feature quantity set in the parent group in the cluster, and the normalized feature quantity of the cluster outer group is used to calculate the corresponding feature quantity set The average value (center of gravity vector) μ _{o of the} vector formed by the feature quantity, and the standard deviation σ _o of the vector of the learning material composed of the feature quantity corresponding to the feature quantity set in the cluster outer group, and the full learning data The ratio ω _{i of the} number of learning materials of the parent group within the cluster, and the ratio of the number of learning materials of the cluster of the parent group in the total number of learning materials ω _o .

Then, the feature quantity set cooperation unit 1 uses the above-described barycenter vectors μ _i , μ _o , standard deviation σ _i , σ _o , ratio ω _i , ω _o , by equation (3), for each cluster, for the feature quantity set A set of feature quantities of all combinations, and a discriminant reference value λ for discriminating the distance between each cluster and other clusters is calculated.

When the calculation of all the discrimination reference values λ is completed, the feature amount set cooperation unit 1 arranges the discrimination reference value λ in descending order for each cluster, and detects the feature amount set corresponding to the largest discrimination reference value λ as a determination. When belonging to each cluster, a feature amount set indicating a set of combinations of feature amounts used in the distance calculation (step S2).

Next, the feature amount set cooperation unit 1 calculates the correlation coefficient R between the feature amounts corresponding to each feature amount set and the feature amount of the learning material in the parent group in each cluster for use in the distance calculation by the distance calculation unit 3. The average value avg. (i) and the standard deviation std. (i) (step S3).

Next, the feature quantity set cooperation unit 1 calculates the correction coefficient λ ^{- (1/2)} from the above-described discrimination reference value λ . The correction coefficient λ ^{- (1/2)} is obtained by standardization between sets of feature quantities. Since the distance from other clusters varies depending on the cluster, in order to improve the classification accuracy, it is necessary to perform standardization between the feature quantity sets. Furthermore, not based on λ ^{- (1/2)} as the correction coefficients, even if the log (λ), or simply using (μ _o -μ _i) may, if it is a function of [lambda] containing executable feature amount Even if you are a standardizer between sets, you can do it.

When Further, in (3) above, in the feature quantity calculation target cluster outside the parent group of the set of centroid vector μ _o, select any of the following three categories of objects is calculated in one cluster in the parent group of the outer Learning data.

a. All the learning materials of the external parent group in the whole learning materials

b. The specific learning materials of the above-mentioned object cluster external parent group corresponding to the purpose of classification

c. Feature data selection of the learning materials used in the learning materials of the external parent group learning materials

Here, the classification purpose of b. is to distinguish it from the cluster that is clearly noticed, and the learning material is the learning material included in the other clusters to which the difference is to be given.

Then, the feature quantity set cooperation unit 1 corresponds to the identification information of each cluster, and sets the feature quantity; the correction coefficient corresponding to the feature quantity set is the value of λ ^-(1/2) in the present embodiment; the inverse matrix R ^-1 ; the average value avg. (i); and the standard deviation std. (i) are stored as the distance calculation data in the feature quantity set storage unit 4 (step 4).

B. Cluster processing (corresponding to the flowchart of Figure 6)

When the classification target data is input, the feature amount extracting section 2 by each cluster The identification signal is read from the feature amount set storage unit 4 for the feature amount set corresponding to each cluster.

Then, the feature quantity extracting unit 2 extracts the feature quantity from each of the classified object data to each cluster based on the type of the feature quantity in the read feature quantity set, and extracts the extracted feature quantity memory corresponding to each piece of identification information of the cluster. The internal memory unit (step S11).

Next, the distance calculating unit 3 reads out the average value avg.(i) and the standard deviation std(i) corresponding to the feature amount from the feature amount set storage unit 4, and performs the operation of the above formula (2). The classification target data feature amount is normalized, and the feature quantity memorized in the internal memory unit is replaced with the normalized feature quantity.

Then, the distance calculating unit 3 generates a matrix V composed of the elements of V(i) obtained as described above, calculates a transposed matrix V ^T of the matrix V, and sequentially calculates the classification target data and the equation (3). The Mahalanobis distance between the clusters is stored in the internal memory corresponding to the identification information of each cluster (step S12).

Next, the distance calculating unit 3 calculates the Mahalanobis distance of the calculation result, multiplies the correction coefficient λ ^-(1/2) corresponding to the feature quantity set, and obtains the correction coefficient, respectively, and the Mahalano ratio. The distance is replaced (step S13). Furthermore, when multiplying the correction coefficient, even after calculating the log or square root of the Mahalanobis distance, it is also possible.

Then, the distance calculating unit 3 compares the correction distances between the clusters in the internal memory unit (step S14), detects the smallest correction distance, and uses the cluster of the identification information corresponding to the correction distance as the classification target data. The cluster of genus, for the cluster database 5, corresponding to the identification information of the cluster of classification targets, and memorizes the classified object data classified (step S15).

[Second Embodiment]

In the first embodiment, the feature amount set used when the clustering is performed is described as one type for each cluster. However, as in the second embodiment described below, a plurality of features are set for each cluster. The amount is calculated by the Mahalanobis distance corresponding to each feature quantity set, and the correction distance is calculated, and the correction distance is rearranged in the order of small to large, and the correction distance within a specific order of the upper position is set in advance. The rules can also be used as a cluster to which the classification object belongs.

In other words, the distance calculating unit 3 according to the present embodiment is configured to represent each cluster by the classification target data set based on the rank of the distance between the classification target data acquired in each feature amount set and the distance of each cluster. A rule pattern of the classification criteria to detect which cluster the classification object data belongs to.

The configuration of the second embodiment is the same as that of the first embodiment shown in Fig. 1, and the same reference numerals are given to the respective configurations. Only the operations different from the first embodiment in the respective configurations will be described with reference to Fig. 7. In the second embodiment, there is a process of setting the rule mode by the self-learning material. Fig. 7 is a flow chart showing an example of the operation of mode learning with respect to the distance of the set rule mode. Fig. 8 and Fig. 9 are flowcharts showing an example of the operation of the clustering in the second embodiment.

Furthermore, in the first embodiment, when the feature quantity set is created, The levy set cooperation unit 1 calculates a discrimination reference value λ for each of a plurality of feature quantity sets that are a combination of feature quantities, and sets a feature quantity set corresponding to the maximum value of the plurality of obtained discrimination reference values λ. A collection of feature quantities for each cluster.

Further, in the second embodiment, the feature amount set cooperation unit 1 sets the feature amount set corresponding to the combination number of the feature amounts for each cluster, one or a plurality of combinations of other clusters, or all other clusters. The maximum value is obtained by obtaining a plurality of discrimination reference values λ, and a plurality of feature amount sets for separating from other clusters are set for each cluster.

Then, the feature quantity set cooperation unit 1 obtains distance calculation data for each feature quantity set, and correspondingly sets the plurality of feature quantity sets and the distance calculation data of each feature quantity set in the feature quantity set storage unit 4 corresponding to the cluster identification information. .

Then, in the seventh figure, when the learning material is input, the feature amount extracting unit 2 reads out a plurality of feature amount sets corresponding to each cluster from the feature amount set storage unit 4 by the identification signals of the clusters.

Then, the feature amount extracting unit 2 extracts the feature amount from the learning material to each cluster in accordance with the category of the feature amount in each of the read feature quantity sets, so that each of the identification information corresponding to the cluster respectively extracts the extracted feature amount. Each set of feature quantities is stored in the internal memory (step S21).

Next, the distance calculating unit 3 is the self-featured set storage unit 4, and reads the average value avg.(i) and the standard deviation std.(i) corresponding to the feature amount for each feature amount set by executing the above formula. (2) The calculation makes the feature quantities extracted from the self-learning materials normalized and will be memorized internally. The feature quantity of the memory unit is replaced with a normalized feature quantity.

Then, the distance calculating unit 3 generates a matrix V composed of the elements of V(i) obtained as described above, calculates a transposed matrix V ^T of the matrix V, and sequentially calculates learning materials and each by equation (3). The Mahalanobis distance between the clusters corresponds to the identification information of each cluster, and each feature amount set is memorized in the internal memory (step S22).

Next, the distance calculating unit 3 calculates the Mahalanobis distance of the calculation result, multiplies the correction coefficient λ ^-(1/2) corresponding to the feature quantity set, and obtains the correction coefficient, and respectively, and the Mahalano Biss distance replacement (step S23).

Then, the distance calculating unit 3 rearranges the correction distances between the clusters in the internal memory unit in the order of small to large (rearranged to the smaller the correction distance, the higher the ranking), that is, the classification data. The correction distance is such that the identification information of the small clusters is ranked in the upper order (step S24).

Next, the distance calculating unit 3 detects the identification information of the cluster corresponding to each of the corrected distances from one of the small (upper) to the nth, and counts the number of pieces of identification information of each of the clusters included in the n, that is, The cluster performs voting processing.

Then, the distance calculation unit 3 is a pattern in which the pattern of the identification amount of the identification information of each cluster of each learning material is detected, and the learning data included in the same cluster is shared.

For example, when n is set to 10, in the case of the learning material of the cluster B, it is detected that the cluster A is five, the cluster B is three, and the cluster C is two. When the mode of counting is set, this is set to rule R1.

Furthermore, when three clusters C are detected in the case of cluster C learning data, even if cluster A is 7 and cluster B is 0, it is not necessarily cluster C, and in this case, If the count amount of the cluster C is three or more, the cluster C may be regarded as the rule R2 irrespective of the count amount of the other clusters.

Furthermore, for the learning data of the cluster A, when the cluster A is in the arrangement pattern of occupying the first and second bits from the upper position, even if the count amount of the cluster B is eight, it is regarded as irrelevant to the counts of other clusters. Cluster A, this is set to rule R3.

As described above, the regularity of the count amount of each cluster included in each learning material classified into the same cluster is detected, and the identification information for each cluster is stored as a pattern list. Here, even if one rule is set in each cluster, a plurality of rules may be set. Further, in the above description, although the distance calculation unit 3 extracts the rule mode, the user may arbitrarily set the count amount or the arrangement rule pattern in order to change the classification accuracy for each cluster.

Depending on the cluster, it is sometimes similar to other clusters in the characteristics of feature information, and there are also correlations between multiple clusters, that is, the count of each cluster or the object pattern from the pattern of the upper order to perform classification. In the case where the accuracy of one of the classifications of the object data is relatively high, this embodiment supplements this point.

Next, the process of clustering using the second embodiment of the rule described in the above list will be described using the flowchart of Fig. 8.

When the classification target data is input, the feature amount extracting unit 2 reads out a plurality of feature amount sets corresponding to each cluster from the feature amount set storage unit 4 by the identification signals of the clusters.

Then, the feature quantity extracting unit 2 is a category corresponding to the feature quantity in each of the read feature quantity sets, and extracts the feature quantity from the classification target data to each cluster, respectively corresponding to the identification information of the cluster, and extracts the feature quantity extracted. Each set of feature quantities is memorized in the internal memory (step S31).

Next, the distance calculating unit 3 is the self-featured set storage unit 4, and reads the average value avg.(i) and the standard deviation std.(i) corresponding to the feature amount for each feature amount set by executing the above formula. (2) The calculation is performed to normalize each feature amount extracted from the classification target data, and replace the feature quantity stored in the internal memory unit with the normalized feature quantity.

Then, the distance calculating unit 3 generates a matrix V composed of the elements of V(i) obtained as described above, calculates a transposed matrix V ^T of the matrix V, and sequentially calculates learning materials and each by equation (3). The Mahalanobis distance between the clusters corresponds to the identification information of each cluster, and each feature amount set is memorized in the internal memory (step S32).

Next, the distance calculating unit 3 calculates the Mahalanobis distance of the calculation result, multiplies the correction coefficient λ ^-(1/2) corresponding to the feature quantity set, and obtains the correction coefficient, respectively, and the Mahalano ratio. The distance is replaced (step S33).

Then, the distance calculating unit 3 rearranges the correction distances between the clusters in the internal memory unit in the order of small to large, that is, the identification information in which the clustering information of the classification target data is smaller is arranged in the upper order. (Step S34).

After the rearrangement, the distance calculating unit 3 detects the identification information of the cluster corresponding to each of the corrected distances from one of the small (upper) to the nth, and counts the number of pieces of identification information of each of the clusters included in the n, that is, The voting process is performed on each cluster.

Next, the distance calculation unit 3 executes a mode (or a mode of arrangement) in which the count amounts of the clusters among the upper n pieces of the classification target data are present, and whether or not there is a comparison process stored in the internal list (step S35).

When the distance calculation unit 3 detects the result of the comparison and the rule pattern matching the target pattern of the classification target data is described in the list, the distance calculation unit 3 determines that the classification target data belongs to the cluster of the identification information corresponding to the matching rule. The classification object data is classified in the cluster (step S36).

Further, the processing of the other clustering using the second embodiment of the rules described in the above list will be described using the flowchart of FIG.

In the other clustering processing shown in FIG. 9, the processing of steps S31 to S35 is the same as the processing shown in FIG. 8, and the distance calculating unit 3 performs self-memory as described in step S35. The rule mode of the list is compared with the object mode of the classified object data.

Then, the distance calculating unit 3 detects whether or not the rule pattern matching the target pattern is searched for, and when detecting the rule pattern in which the matching is found, the process proceeds to step S47, and when the detection is performed, When the matching rule mode is not retrieved, the process proceeds to step S48 (step S46).

When detecting the rule pattern in which the agreement is found, the distance calculation unit 3 determines that the classification target data belongs to the identification information corresponding to the rule that is consistent with it. The clusters are classified into the clusters, and the cluster database 5, corresponding to the identification information of the clusters of the classification targets, memorizes the classified object data (step S47).

Further, when it is detected that there is no rule pattern in which the search is unmatched, the distance calculation unit 3 detects the count amount, that is, the identification information having the largest number of votes, and classifies the classification target data corresponding to the cluster of the identification information.

Then, the distance calculation unit 3 stores the classified classification target data for the cluster database 5 corresponding to the identification information of the cluster of the attribution target (step S48).

[Third embodiment]

In the second embodiment, a list of rule patterns from the one of the clusters of the calculated classification target data from the small (similarity is large) to the upper n is prepared by whether or not it corresponds to the existence of the list. The rule mode is described by performing the process of clustering the classification target data. However, as in the third embodiment below, a plurality of feature amount sets are set for each cluster, and the Mahalano corresponding to each feature amount set is calculated. The distance of the distance is calculated, and the correction distance within the specific order of the upper rank is calculated as the cluster to which the classification target data belongs.

In the following, the configuration of the third embodiment is the same as that of the first and second embodiments shown in Fig. 1, and the same reference numerals are given to the respective configurations. Only the operations different from the second embodiment in the respective configurations will be described using FIG. . In the third embodiment, the processing of setting the above rule is not performed on the self-learning material, and step S48 in Fig. 9 is directly executed. Figure 10 is a diagram showing the third implementation A flow chart of an example of the operation of a cluster in a pattern.

In the other clustering processing shown in FIG. 10, the processing from step S31 to step S34 is the same as the processing shown in Fig. 8, and the distance calculating unit 3, as described earlier, in step S34, In the order from small to large, the correction distance between the clusters in the internal memory portion is rearranged, that is, the cluster information in which the correction distance of the classification target data is small is the upper order (step S34).

Next, the distance calculating unit 3 detects the identification information of the cluster corresponding to each of the corrected distances from one of the small (upper) to the nth, and counts the number of pieces of identification information of each of the clusters included in the n, that is, Each cluster performs voting processing (step S55).

Then, the distance calculating unit 3 detects the identification information of the most count value (the number of votes) in the voting result, and regards the cluster corresponding to the identification information as the cluster to which the classification target data belongs, and corresponds to the cluster database 5 The identification information of the cluster belonging to the target, and the classified object data of the memory classification (step S56).

Furthermore, the distance calculation unit 3 sets a threshold value for the number of votes that the user has pre-screened for each piece of identification information, and when the number of votes of the identification information having the highest number of votes does not reach the threshold value, even if the execution does not belong to Any cluster can be processed as well.

For example, when the classification object data is classified into three clusters of clusters A, B, and C, the number of votes relative to the identification information of cluster A is five, and the number of votes relative to the identification information of cluster B is three. When the number of votes relative to cluster C is 2, the identification information of the most votes is detected. Set A and distance calculation unit 3.

However, when the above-described threshold value of the cluster A is set to six, since the number of votes for the identification information with respect to the cluster A does not reach the critical value, the distance calculating unit 3 determines that it does not belong to any cluster.

Accordingly, in the clustering of clusters in which the feature quantity is only slightly different from other clusters, the reliability of the classification processing of the classification object data to the clustering can be improved.

[Method of transforming feature quantity]

Although the parent group of each feature quantity is expected to perform clustering for the normal distribution, it is considered that there is a case where the parent group has a skew distribution depending on the type (area, length, etc.) of the feature quantity, and the classification is performed at this time. The calculation of the distance between the object data and each cluster is a decrease in the accuracy of the classification object data and the similarity of each cluster.

Therefore, depending on the feature quantity, it is necessary to perform a specific method to change the feature quantity of the parent group, so that the determination accuracy of the similarity is improved close to the normal distribution.

As a conversion method for transforming to the normal distribution, any one of the expressions obtained by the n-square root or factorial or numerical calculation by log or square root (√), cube root ( ³ √), or the like is used. To transform the feature quantity.

Hereinafter, the setting process of the conversion method of each feature amount will be described using FIG. Fig. 11 is a flowchart showing an operation example of setting processing of each feature amount conversion method. Moreover, the transformation method is based on the special features contained in the cluster. The levy unit is set for each cluster. Furthermore, the setting of the conversion method is performed using learning materials belonging to each cluster. The following processing will be described with the feature amount set cooperation unit 1 being executed, but the processing unit corresponding to the processing may be provided elsewhere.

The feature amount set cooperation unit 1 uses the identification information of the cluster of classification objects as a key, reads the learning data included in the cluster from the cluster database, and calculates (normalizes) the feature amount of each learning material (step S61).

Then, the feature amount set cooperation unit 1 performs the conversion of the feature amount by transporting the read learning materials by using any one of the arithmetic expressions for performing the feature amount conversion stored therein (step S62).

When the conversion of the feature amounts of all the learning materials is completed, the feature amount set cooperation unit 1 calculates an evaluation value indicating whether or not the distribution obtained by the conversion processing is close to the normal distribution (step S63).

Next, the feature amount set cooperation unit 1 detects whether or not it is stored internally, that is, all the arithmetic expressions set in advance as the conversion method, and the estimated value is calculated, and the feature value is calculated by calculating the feature amount in all the arithmetic expressions. When the evaluation value of the distribution is performed, the process proceeds to step S65, and when it is detected that the calculation of the feature amount using all the arithmetic expressions has not yet been completed, the processing of the next set arithmetic expression is executed, so the process returns to step S62 ( Step S64).

When the transformation using the feature quantities of all the expressions ends, the feature quantity set cooperation unit 1 detects the distribution in which the evaluation value is the smallest among the distributions obtained in the set expression, that is, the distribution closest to the normal distribution. Determining the expression used to create the detected distribution as a transformation method The method of transforming the feature amount of the cluster is set internally (step S65).

The feature amount set cooperation unit 1 performs the above-described processing for each feature amount of each cluster, and sets a conversion method corresponding to each feature amount of each cluster.

Next, the calculation of the evaluation value in the above step S63 will be described using FIG. Fig. 12 is a flow chart showing an example of the operation of the process of evaluating the evaluation value of the distribution obtained by the arithmetic expression.

The feature amount set cooperation unit 1 converts the feature amounts of the learning materials belonging to the target cluster by the set arithmetic expression (step S71).

After the feature quantity of all the learning materials is converted, the feature quantity set cooperation unit 1 calculates the average value μ and the standard deviation σ of the distribution (parent group) obtained by the converted feature quantity (step S72).

Then, the feature quantity set cooperation unit 1 calculates the z value (1) from (x - μ) / σ using the average value μ and the standard deviation σ of the parent group (step S73).

Next, the feature amount set cooperation unit 1 calculates the cumulative accuracy rate in the parent group (step S74).

After the calculation, the feature amount set cooperation unit 1 calculates the z value (2) as the inverse function of the cumulative distribution function of the standard normal distribution by the cumulative reliability ratio in the obtained parent group (step S75).

Then, the feature quantity set cooperation part 1 is to obtain two z values of the distribution of the feature quantity, that is, the difference between the z value (1) and the z value (2), that is, the error of the two z values in the distribution (step S76).

When the error of the z value is obtained, the feature quantity set cooperation unit 1 calculates the sum of the errors of the above two z values, that is, the sum of the errors (self-multiplication) to be regarded as The value is evaluated (step S77).

The smaller the error of the above two z values, the closer the distribution is to the normal distribution. If there is no error of z value, it is a normal distribution. In addition, the farther the distribution is from the normal distribution, the larger the error.

Next, before performing the processing of the clusters in the first to third embodiments, the feature amount of the classification target data will be described using FIG. Fig. 13 is a flow chart showing an example of the operation of calculating the feature amount data of the classification target data.

The distance calculation unit 3 extracts the feature amount of the recognition target from the input classification target data in accordance with the feature amount set set for each cluster, and performs all the normalization processing described (step S81).

Next, the distance calculation unit 3 converts the feature amount for classifying the classification target in the classification target data by the conversion method (calculation formula) set for the feature amount of the cluster (step S82).

Then, the distance calculating unit 3 calculates the distance from the cluster to be classified as described in the first to third embodiments (step S83).

Next, the distance calculation unit 3 converts the feature quantity to all the clusters of the classification target by the conversion method set corresponding to the feature quantity of each cluster, and performs detection to detect whether or not the cluster is calculated by the feature quantity of the transformation. When it is detected that the distance is obtained for all the clusters of the classification target, the process proceeds to step S85, and when the cluster in which the classification target remains is detected, the process returns to step S82 (step S84).

Then, in each of the first to third embodiments, the processing from the point at which the calculation of the distance is completed is started (step S85).

For the above reasons, the Mahalanobi used in the present embodiment When the distance between the classification target data and each cluster is obtained, the expected feature quantity is a normal distribution. Therefore, the distribution of each feature quantity of the parent group is closer to the normal distribution, and the cluster can be obtained. The correct distance (similarity) can be expected to classify the accuracy of each cluster.

Example [Calculation example]

Next, using the cluster system of the first, second, and third embodiments, the classification accuracy of the conventional example is confirmed by the sample data shown in Fig. 14. It can be seen that although the number of samples is small, the accuracy of the conventional example or the above is obtained regardless of the amount of features used. In the fourteenth figure, as the cluster, 10 learning materials are defined for each of the types 1, 2, and 3, and each learning material has the characteristic quantities a, b, c, d, e, f, g, and h. One. In this example, the learning material belonging to each cluster shown in Fig. 14 determines the feature amount set used by the cluster, and then the learning set is also used as the classification target data to perform clustering.

As a result of the calculation, Fig. 15 is a conventional calculation method, and using the feature quantities a and g as a combination of feature amounts, the Mahalanobis distance is calculated for each learning material shown in Fig. 14 of cluster 1 to cluster 3. , indicating the result of the judgment. In Figure 15 (a), Cluster1 is listed as the Mahalanobis distance from Cluster 1, Cluster 2 is listed as the Mahalanobis distance from Cluster 2, and Cluster3 is listed as the Horse with Cluster 3. Harlanobis distance. Furthermore, the category column actually represents the cluster to which each learning material belongs, and the judgment result indicates that the learning material and the Mahalanobis distance are the smallest clusters. table The number indicating the correct classification type and the judgment result is the same characteristic quantity data.

In Fig. 15(b), the number of the column indicates the cluster to which the learning material actually belongs, and the number of the row indicates the cluster to be determined. For example, "8" of the flag R1 is determined to be the cluster 1 of the 10 clusters of the cluster 1 and the "2" of the marker R2 is the cluster 2 of the 10 clusters of the cluster 1. P0 is a coincidence rate indicating a positive solution and an answer, p1 is a rate indicating the coincidence of the two, and k is a total correction rate, which is obtained by the following equation. Indicates that the higher the k, the higher the classification accuracy.

k =(p0-p1)/(1-p1)

P0=(a+d)/(a+b+c+d)

P1=[(a+b)‧(a+c)‧(b+d)‧(c+d)]‧(a+b+c+d) ²

The relationship of a, b, c, and d in the above formula will be described using Fig. 16.

The data belonging to cluster 1 is classified as a cluster 1 and the data belonging to cluster 1 is treated as cluster 2 as the number b, and a+b is the number of data belonging to cluster 1. Furthermore, the data belonging to cluster 2 is treated as cluster 2, the number is d, and the data belonging to cluster 2 is treated as cluster 1 and the number is c, indicating that c+d belongs to cluster 2. a+c is the number of clusters 1 classified into the full data a+b+c+d, and b+d is classified into the number of clusters b in the full data a+b+c+d.

Next, in the seventh embodiment, the Mahalanobis distance is calculated for each learning material shown in Fig. 14 of the cluster 1 to the cluster 3, using the calculation method of the first embodiment, and the determination result is shown. The views of FIGS. 17(a) and (b) are the same as those of Fig. 15, and therefore their description will be omitted. It can be seen that the accuracy rate p0, the accidental coincidence rate p1, and the total correction rate k are equal to the conventional calculation method of Fig. 15. Here, a method of selecting a combination of the maximum discrimination reference values λ for each cluster from the above-described combination is used, and a feature amount set corresponding to each cluster is calculated. The feature quantity set corresponding to the cluster 1 is a combination of the feature quantities a and h, and the feature quantity set corresponding to the cluster 2 is the feature quantity a and d, and the feature quantity set corresponding to the cluster 3 is A combination of feature quantities a, g is used.

Next, in the 18th figure, using the calculation method of the second embodiment, the Mahalanobis distance is calculated for each learning material shown in Fig. 14 of the cluster 1 to the cluster 3, and the determination result is shown. The views of FIGS. 18(a) and (b) are the same as those of Fig. 15, and the description thereof will be omitted. The correct rate p0 is 0.8333, the coincidence accuracy rate p1 is 0.3333, and the total correction rate k is 0.75. It can be seen that the classification accuracy is improved compared with the conventional calculation method of Fig. 15. Here, a method of selecting a combination of the discrimination reference values λ up to the third digit in each cluster is selected from the combination of the above, and the feature amount set corresponding to each cluster is calculated. Use feature quantity a. h, a. g, d. 3 combinations of e to be used as a set of feature quantities corresponding to cluster 1, using feature quantity a. f, a. d, a. The three combinations of b are taken as the feature quantity set corresponding to cluster 2, and the feature quantity e is used. g, a. c, a. The three combinations of g are taken as the feature quantities corresponding to the cluster 3.

Furthermore, as the judgment of the vote, the number of clusters from the few to the third digits is calculated in the order of the distance from the Mahalanobis, and the cluster of the largest number is regarded as the cluster to which the classification object data belongs. .

Next, in Fig. 19, using the calculation method of the second embodiment, the Mahalanobis distance is calculated for each learning material shown in Fig. 14 of the cluster 1 to the cluster 3, and the Mahalano ratio is calculated. After multiplying the correction coefficient (λ) ^-1/2 , the distance of the execution distance is arranged to indicate the determination result. The method for viewing the above-described 19th (a) and (b) is the same as that of Fig. 15, and therefore the description thereof will be omitted. The correct rate p0 is 0.8333, the coincidence accuracy rate p1 is 0.3333, and the total correction rate k is 0.75. It can be seen that the classification accuracy is improved compared with the conventional calculation method of Fig. 15. Here, a combination of the above-described total combinations is used to select a combination of the discrimination reference values λ having the number of the upper third digits in each cluster, and the feature amount sets corresponding to the clusters are calculated. Three combinations of feature quantities a‧h, a‧g, and d‧e are used as the feature quantity set corresponding to cluster 1, and three combinations of feature quantities af, a‧d, and a‧b are used as Corresponding to the feature quantity set of the cluster 2, three combinations of the feature quantities e‧g, a‧c, and a‧g are used as the feature quantities corresponding to the cluster 3.

Furthermore, as the judgment of the vote, the number of clusters from the few to the third digits is calculated in the order of the distance from the Mahalanobis, and the cluster of the largest number is regarded as the cluster to which the classification object data belongs. .

From the respective classification results shown in the above-mentioned 15th, 17th, 18th, and 19th, it is judged that the present embodiment performs high-speed and high-precision clustering processing as compared with the conventional example, and the superiority of the present embodiment with respect to the conventional example is confirmed.

[Application example of the present invention] A. Inspection device

An inspection apparatus (defect inspection apparatus) for classifying an inspection object such as a scratch type on the surface of a glass substrate as shown in Fig. 20 will be described. Figure 21 is A flowchart illustrating an operation example of selection of a feature amount set, and FIG. 22 is a flowchart illustrating an operation example in the clustering process.

First, the selection action of the feature amount set will be described. The collection of the learning materials in step S1 in the flowchart of Fig. 5 corresponds to steps S101 to S105 of the flowchart of Fig. 21.

Steps S2 to S4 of Fig. 21 are the same as those of the flowchart of Fig. 5, and therefore description thereof will be omitted.

By the operation of the operator, samples for learning materials respectively corresponding to the clusters of the types of scratches to be classified are collected (step S101).

The illumination device 102 irradiates the image acquisition unit 101 with the shape of the scratches collected by the learning material, and the image capturing device 103 acquires the image data of the scratched portion (step S102).

Then, the image data acquired by the self-image acquisition unit 101 calculates the feature amount of the scratch of each learning material (step S103).

Each of the acquired feature amounts of the learning materials is assigned to the classified target obtained by visual observation, and the specificity of the learning materials in each cluster is executed (step S104).

Then, the learning data of each cluster is a specific number (the number of samples set in advance), for example, about 30 each, and the processing from step S101 to step S102 is repeated. When the number is a specific number, the clustering unit 105 executes all the fifth graphs. The processing after step S2 described in the above. Here, the clustering unit 105 is the clustering system in the first or second embodiment.

Next, the processing of the clustering in the inspection apparatus of Fig. 4 will be described with reference to Fig. 22. Here, step S31 to step S34 of FIG. 22, S55 and S56 are the same as the flowchart of Fig. 10, and therefore the description thereof will be omitted.

In the inspection apparatus of Fig. 20, when the inspection is started, the illumination device 102 is illuminated for the glass substrate of the inspection object 100, and the imaging device 103 photographs the surface of the glass substrate and outputs the captured image to the image acquisition unit 101. In this way, the defect candidate detecting unit 104 detects a portion different from the plane shape in the photographed image input by the self-image obtaining unit, and sets the defect candidate to be classified (step S201).

Next, the defect candidate detecting unit 104 cuts out the image data of the defect candidate portion from the photographed image as the classification target data.

Then, the defect candidate detecting unit 104 calculates the feature amount from the image data of the classification target data, and outputs the classification target data composed of the extracted feature amount to the cluster unit 105 (step S202).

The processing for the subsequent clustering is omitted because it is explained in the step of Fig. 10. As described above, the inspection apparatus of the present invention can classify the scratches on the glass substrate into each of the types of scratches with high precision.

B. Defect type determination device

The defect type determining device shown in Fig. 23, the clustering unit 105 corresponds to the cluster system of the present invention which has been described.

The image acquisition device 201 is composed of the image acquisition unit 101, the illumination device 102, and the imaging device 103 in Fig. 20 .

The learning materials of the clusters in which the classification object data is classified are obtained, and are prepared in the cluster database 5 of the cluster device 105. Therefore, the feature amount set in Fig. 5 is also ended.

The imaging device input from the image acquisition device 202 mounted on each manufacturing device detects the defect candidate, extracts the image data, extracts the feature amount, and outputs it to the data collection device 203. The control device 200 transfers the classification target data input to the material collection device 203 to the clustering unit 105. Then, as described above, the clustering unit 105 is classification object data input for each cluster classification corresponding to the type of the scratch.

C. Manufacturing management device

The manufacturing management device of the present invention is constituted by the control device 300, the manufacturing devices 301 and 302, the notifying unit 303, the recording unit 304, the inferior device determining unit 305, and the defect type determining device 306, as shown in Fig. 24. Here, the defect type determining means 306 is the same as the defect type determining means described in the above item B.

The defect type determining unit 306 performs image processing on the image capturing images of the image capturing devices 201 and 202 provided in each of the manufacturing device 301 and the manufacturing device 302 in the corresponding defect candidate detecting unit 104, and extracts the feature amount. Perform classification of classified object data.

Next, the poor device determining unit 305 has a table indicating the identification information of the cluster to be classified and a relationship corresponding to the cause of the cluster, and reads the cluster of the classification targets input from the defect type determining device 306 from the table. Identify the cause of the occurrence of the information and determine the manufacturing device that is the cause of the occurrence. In other words, the poor condition device determining unit 305 detects the occurrence of a defect in the manufacturing process of the product in accordance with the identification information of the cluster.

Then, the poor condition device determining unit 305 is notified by the notifying unit 303 to The operator, when the recording unit 304 corresponds to the date determined, memorizes the identification number of the cluster of the defect classification, the cause of the occurrence, and the identification information of the manufacturing apparatus as the history. Furthermore, the control device 300 is a manufacturing device determined by the stop condition poor device determining unit 305, and controls the control parameters.

D. Manufacturing management device

The other manufacturing management apparatus of the present invention is constituted by the control device 300, the manufacturing apparatuses 301 and 302, the notifying unit 303, the recording unit 304, and the clustering unit 105 as shown in Fig. 25. Here, the clustering unit 105 is the same as the configuration described in the above items A and B.

The clustering unit 105 is different from the above-described cases of A to C, and the feature data of the classified object data is constituted by, for example, manufacturing conditions (material component, processing temperature, pressure, processing speed, etc.) in the manufacturing process of the glass substrate. The feature quantity is classified according to the manufacturing status of each process of the process. The above-described feature amount is input to the cluster portion 115 as the feature amount by the engineering information detected by the sensors provided in the respective manufacturing apparatuses 301 or 302.

In other words, the clustering unit 105 classifies the manufacturing state of the glass manufacturing process in each of the manufacturing apparatuses by the feature amount of the classification target data into a "normal state" and a "state in which defects are likely to be adjusted". A cluster of "dangerous and need to adjust the state". Then, the clustering unit 105 notifies the operator of the classification result by the notifying unit 303, and outputs the identification information of the clustering result of the classification result to the control device 300, and further, when the recording unit 304 corresponds to the date determined, The classification identification number of the manufacturing status of each of the above-mentioned projects as a history; The manufacturing conditions of the levy; and the identification information of the manufacturing device.

The control device 300 has a table corresponding to the adjustment item for which the identification information of the cluster and the manufacturing condition are returned to normal, and the data thereof, and reads the adjustment information corresponding to the cluster identification information input from the cluster unit 105, and returns the manufacturing condition to the normal adjustment item. The data is controlled by the read data to control the corresponding manufacturing device.

Further, even if the program for realizing the function of the cluster system in FIG. 1 is recorded on a computer-readable recording medium, the computer system reads the program recorded on the recording medium, and performs clustering of the classified object data. Processing can also be. Further, the term "computer system" as used herein includes hardware such as an OS or a peripheral device. Furthermore, the "computer system" also includes a WWW system with a webpage providing environment (or a display environment). Furthermore, "computer-readable recording medium" refers to a portable device such as a floppy disk, a magneto-optical disk, a ROM, a CD-ROM, or the like, and a hard disk embedded in a computer system. Further, the "computer-readable recording medium" is a volatile memory (RAM) inside the computer system of the server or the client when the program is transmitted via a communication loop such as a network such as the Internet or a telephone line. , to keep the programmer for a certain period of time.

Furthermore, the program may be transmitted to another computer system via a transmission medium or a transmission wave in the transmission medium, even if it is stored in a computer system such as a memory device. Here, the "transmission medium" of the transmission program is a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. Furthermore, the above program may be used to implement one of the above functions. And that is It is also possible to implement the above functions by a combination with a program already recorded in a computer system, or a differential file (differential program).

[Industry use feasibility]

The present invention is applicable to a field in which information having a plurality of types of feature amounts is classified and classified with high precision, such as detection of defects such as glass articles, and the like, and can also be utilized in a manufacturing state detecting device or a product manufacturing management device.

The entire contents of the specification, the claims, the drawings, and the abstract of the application of the Japanese Patent Application No. 2006-186628, filed on Jul. 6, 2006, are hereby incorporated by reference.

1‧‧‧Characteristics

2‧‧‧Characteristic Extraction Department

3‧‧‧ Distance Calculation Department

4‧‧‧Characteristics collection memory

5‧‧‧ Cluster Database

100‧‧‧Inspected objects

101‧‧‧Portrait Acquisition Department

102‧‧‧Lighting device

103‧‧‧ camera

104‧‧‧Defective candidate detection department

105‧‧ ‧ Cluster Department

200, 300‧‧‧ control devices

201, 202‧‧‧Portrait acquisition device

301, 302‧‧‧ manufacturing equipment

303‧‧‧Notice

304‧‧‧Record Department

Fig. 1 is a block diagram showing a configuration example of a cluster system according to the first and second embodiments of the present invention.

Fig. 2 is a diagram for explaining a process of selecting a feature set by discriminating a reference value γ.

FIG. 3 is a list for explaining a process of selecting a feature set by discriminating a reference value γ.

Fig. 4 is a histogram illustrating the effect of selecting a feature set by discriminating the reference value γ.

Fig. 5 is a flowchart showing an operation example of a process of selecting a feature amount set for each cluster by the first embodiment.

Fig. 6 is a view showing execution of classification data by the first embodiment; A flowchart of an example of the operation of the clustering process.

Fig. 7 is a flowchart showing an operation example of a table for generating a rule pattern used in the process of clustering in the second embodiment.

Fig. 8 is a flowchart showing an operation example of a process of performing clustering on classification target data by the second embodiment.

Fig. 9 is a flowchart showing an operation example of a process of performing other clustering on the classification target data by the second embodiment.

Fig. 10 is a flowchart showing an operation example of a process of performing clustering on classification target data by the third embodiment.

Fig. 11 is a flowchart showing an example of an operation of setting an arithmetic expression as a method of converting a feature amount.

Fig. 12 is a flow chart showing an operation example of calculating an evaluation value in the flowchart of Fig. 11.

Fig. 13 is a flowchart showing an example of an operation for calculating a distance of a feature amount converted using the set conversion method.

Figure 14 is a list showing the learning materials belonging to each cluster.

Fig. 15 is a list of results showing the results of classifying the learning materials of Fig. 14 by the clustering method of the conventional example.

Fig. 16 is a conceptual diagram for explaining a method of calculating the total correction determination rate.

Fig. 17 is a view showing a result of the result of classifying the learning materials of Fig. 14 by the cluster system in the first embodiment.

Fig. 18 is a view showing a result list of the learning materials classified in Fig. 14 by the cluster system in the second embodiment.

Fig. 19 is a view showing a result list of the learning materials of Fig. 14 classified by the cluster system in the second embodiment.

Fig. 20 is a block diagram showing an example of the configuration of an inspection apparatus using the cluster system of the present invention.

Fig. 21 is a flow chart showing an example of the selection operation of the feature amount set in the inspection apparatus of Fig. 20.

Fig. 22 is a flow chart showing an operation example of the process of clustering in the inspection apparatus of Fig. 20.

Fig. 23 is a block diagram showing the constitution of a defect type judging device using the cluster system of the present invention.

Fig. 24 is a block diagram showing an example of the configuration of a manufacturing management apparatus using the cluster system of the present invention.

Fig. 25 is a block diagram showing an example of the configuration of another manufacturing management apparatus using the cluster system of the present invention.

1‧‧‧Characteristics

2‧‧‧Characteristic Extraction Department

3‧‧‧ Distance Calculation Department

4‧‧‧Characteristics collection memory

5‧‧‧ Cluster Database

Claims (13)

A cluster system, which is a cluster system that classifies input data into respective clusters formed by a parent group of learning materials by using the feature quantity of the input data, and has a feature quantity set memory unit. Corresponding to each of the clusters, a feature quantity set in which a combination of feature quantities used for classification is stored; a feature quantity extracting unit extracting a feature quantity set in advance from the input data; and a distance calculating unit corresponding to each Each feature quantity set of the cluster calculates the distance between the center of the parent group of each cluster and the input data according to the feature quantity contained in the feature quantity set, and outputs the distance as the set distance; and the order extraction unit, The above-mentioned respective set distances are arranged from small to large, and a plurality of the above-mentioned feature quantity sets are set for each cluster, and a cluster classification part is also represented by the above-mentioned set distances obtained according to each feature quantity set. The input data of the set distance of the set distance is classified into a rule pattern of the classification reference of each cluster, and the input data is detected. Which one cluster. The clustering system according to the first aspect of the invention, wherein the cluster classification unit detects, by the rank of the set distance, which cluster the input data belongs to, and sets the rank as a cluster of a higher set distance. , detected as a cluster to which the above input data belongs Out. The clustering system according to the second aspect of the invention, wherein the cluster classification unit has a threshold value that is higher than a rank, and if the upper cluster is greater than the threshold, the data is regarded as an input data. It is detected by clustering. The clustering system according to any one of claims 1 to 3, wherein the distance calculating unit multiplies the set distance by the correction coefficient set corresponding to the feature amount set, and sets each feature amount set. The set distance between them is standardized. The clustering system according to any one of claims 1 to 3, further comprising a feature quantity set cooperation unit for creating a feature quantity set of each cluster, wherein the feature quantity set cooperation part is for each feature quantity For each of the majority of the combinations, the average of the learning materials of the parent group of each cluster is taken as the origin, and the average of the distances between the origins and the learning materials of the parent group of the other clusters is obtained, and will become The combination of the feature amounts of the maximum average is selected as a set of feature quantities for distinguishing each cluster from other clusters. A defect type determining apparatus, characterized in that: the clustering system according to any one of the above-mentioned claims, wherein the input data is an image data of a product defect, by indicating a feature quantity of the defect, The defects in the image data are classified according to the type of the defect. Defect type determination device as described in item 6 of the patent application scope The above product is a glass article, and the defects of the glass article are classified according to the type of the defect. A defect detecting device characterized by providing a defect type determining device as described in claim 6 or 7 for detecting a defect type of a product. A manufacturing state judging device characterized in that the defect type judging device described in the sixth or seventh aspect of the patent application is provided, and the classification of the defect of the product is performed, and the correspondence is detected based on the correspondence with the cause corresponding to the category. The cause of defects in the process. A manufacturing state judging device characterized by being provided with a cluster system as described in any one of claims 1 to 5, wherein the input data is a feature quantity indicating a manufacturing condition in a process of the product, according to the process The feature quantity of each project is classified. The manufacturing state judging device according to claim 10, wherein the product is a glass article, and the feature amount in the process of the glass article is classified according to the manufacturing state of each process of the process. A manufacturing state detecting device characterized in that the manufacturing state judging device as set forth in claim 10 or 11 is provided, and the type of manufacturing state in each of the processes of the product process is detected. A product manufacturing management apparatus, comprising: a manufacturing state judging device as described in claim 10 or 11 of the patent application, executing a category of a manufacturing state in each of the processes for detecting a product process, and according to the category corresponding to the category Control the project and carry out process control in the process engineering.