WO2008004559A1 - Clustering system, and defect kind judging device - Google Patents

Abstract

Provided is a clustering system capable of classifying target data more rapidly and precisely than the examples of the prior art. The clustering system classifies input data into each of clusters formed of the populations of learning data, in terms of featuring quantities owned by the input data. The clustering system comprises a featuring set storage unit stored with featuring quantity sets or a combination of featuring quantities to be used for the classifications, in a manner to correspond to the individual clusters, a featuring quantity extracting unit for extracting a preset featuring quantity from the input data, a distance calculating unit for calculating and outputting the distances between the center of the population of the individual clusters and the input data, individually as set distances for each featuring quantity set corresponding to each cluster, on the basis of the featuring quantities contained in the featuring quantity set, and a rank extracting unit for arraying the individual set distances in the order of the smaller length.

Description

 Specification

 Clustering system and defect type determination apparatus

 Technical field

 TECHNICAL FIELD [0001] The present invention relates to a clustering system and a defect type determination device that cut out a partial image of a defect portion in an image of a detection target object, extract a feature signal of the defect from the partial image cover, and classify the defect type. .

 Background art

 A conventional clustering technique based on a distance between unknown data and learning data, for example, a Mahalanobis (Mahalanobis generalized distance) distance, is generally performed. That is, the unknown data is classified by determining whether or not it belongs to a cluster as a population learned in advance, and clustering processing is performed. For example, the determination of which population cluster the unknown data belongs to is made based on the Mahalanobis distance for multiple clusters (see, for example, Patent Document 1).

 In addition, in order to efficiently calculate the above-described distance, a plurality of feature amounts are selected and clustering processing is performed.

 [0003] In addition, a method of determining a cluster to which the unknown data belongs by voting results obtained from a large number of classifiers (classifiers) is also common. The identification result in the identification of unknown data for different areas on the image is used (for example, see Patent Document 2).

 With the clustering method described above, in diagnosing illnesses using parameters obtained as a result of blood tests, that is, in categorizing which disease belongs, a combination of two clusters in a plurality of clusters is set, and for each combination. Whether the test data is judged to be similar to any cluster is performed for all the combinations, and the result of counting the number of judgments determines that the data is classified into a cluster with a large number of judgments. There is a law (for example, Patent Document 3).

[0004] When each defect attached to the LCD glass substrate is classified for each preset defect type, each feature amount used for classification is optimized in accordance with the identification at the time of classification, This Clustering is performed to weight each feature amount so as to correspond to the optimization of each of the clusters, and to determine which cluster the cluster belongs to using the optimized feature amount (for example, Patent Document 4).

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

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

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

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

 Disclosure of the invention

 Problems to be solved by the invention

 [0005] However, in the clustering shown in Patent Document 3, the individual combinations are not optimized, the feature quantities that are the discrimination materials are not fully used, and the discrimination is not performed. When the number of clusters to be increased increases, the number of combinations becomes enormous, and there is a problem that the time required for the determination process increases.

 Further, in the clustering shown in Patent Document 4, the feature amount is weighted based on the determination rate to improve the discrimination accuracy, but there is no concept of feature amount optimization for each cluster as described above. Similar to Patent Document 3, since the feature quantity is not fully utilized, there is a drawback that classification with high accuracy is not performed.

 [0006] The present invention has been made in view of such circumstances, and is utilized when discriminating the extracted feature quantity, which is an object to be classified into the cluster to which it belongs, and compared with the conventional example. Provided are a clustering system and a defect type determination device that can classify classification target data at high speed and with high accuracy, for example, can classify defects on a glass surface into clusters corresponding to defect types.

 Means for solving the problem

[0007] In order to solve the above-described problem, the present invention is different from the conventional example in which the distance between the classification target data and each cluster is calculated with the same type of feature amount to determine the classification destination. Therefore, a set of feature quantities that can obtain differences between each cluster is set for each cluster, and the distance is obtained with different feature quantities for each cluster. Thus, classification is performed with higher accuracy. Since the set of feature values described above is performed based on the characteristics of the learning data belonging to each cluster, the feature value sets are configured with feature values that can be distinguished from other clusters.

 That is, the present invention employs the following configuration.

 [0008] The clustering system of the present invention includes input data (input data) for each cluster formed by the population of learning data (learning data), and features (parameters) of the input data. In a clustering system that classifies according to the above, a feature set (parameter set) that is a combination of features used for classification is stored corresponding to each cluster, and is set in advance from a feature set storage unit and input data. A feature amount extracting unit for extracting the feature amount, and for each feature amount set corresponding to each cluster, based on the feature amount included in the feature amount set, the center of the population of each cluster and the input A distance calculating unit that calculates and outputs the distance to the data as a set distance, and a rank extracting unit that arranges the set distances in ascending order.

 In a preferred clustering system of the present invention, a plurality of the feature quantity sets are set for each cluster.

 [0010] A preferred clustering system of the present invention provides a classification criterion for each cluster of input data set based on the rank of the set distance based on the set distance obtained for each feature amount set. A cluster classification unit for detecting which cluster the input data belongs to by a rule pattern indicating! /

 [0011] In a preferred clustering system of the present invention, the cluster classification unit detects which cluster the input data belongs to based on the rank of the set distance, and there are many set distances where the rank is higher. A cluster is detected as a cluster to which the input data belongs.

 [0012] In a preferred clustering system of the present invention, the cluster classification unit has a threshold for the number of ranks higher than the rank, and the input data belongs if the higher-ranked cluster is equal to or higher than the threshold. Detect as a cluster.

In a preferred clustering system of the present invention, the distance calculation unit multiplies the set distance by a correction coefficient set corresponding to the feature amount set, and sets a set distance between each feature amount set. It is characterized by standardization. [0014] A preferred clustering system of the present invention further includes a feature quantity set creation unit that creates a feature quantity set for each cluster, and the feature quantity set creation unit has a plurality of combinations of each feature quantity. The average value of the learning data of the population of each cluster is used as the origin, the average value of the distance between this origin and each learning data of the population of another cluster is obtained, and the combination of the feature values that has the largest average value is obtained. Then, it is selected as a feature value set used for distinguishing each cluster from other clusters.

 [0015] The defect type determination apparatus according to the present invention is provided with any of the above-described clustering systems, and the input data is image data of a product defect, and the image data includes a feature amount indicating the defect. Defects are classified by defect type.

 In a preferred defect type determination apparatus according to the present invention, the product is a glass article, and the defects of the glass article are classified by defect type.

[0016] A defect detection apparatus according to the present invention detects a defect type of a product provided with the defect type determination apparatus.

 [0017] A manufacturing state determination device according to the present invention is provided with the defect type determination device described above, performs a type of product defect, and in a manufacturing process based on correspondence with an occurrence factor corresponding to the type. Detect the cause of the defect.

 [0018] A preferable manufacturing state determination apparatus of the present invention is provided with any one of the above-described clustering systems, and the input data is a characteristic amount indicating a manufacturing condition in a manufacturing process of the product. Categorize by manufacturing state of each step of the manufacturing process.

 In a preferable manufacturing state determination apparatus of the present invention, the product is a glass article, and the characteristic amount in the manufacturing process of the glass product is classified according to the manufacturing state of each step of the manufacturing process.

 [0019] The manufacturing state detection device of the present invention detects the type of manufacturing state in each step of the product manufacturing process provided with the manufacturing state determination device described above.

 [0020] The product manufacturing management device of the present invention detects the type of manufacturing state in each step of the manufacturing process of the product provided with the above-described manufacturing state determination device, and sets the control item corresponding to the type. Based on this, process control in the manufacturing process is performed.

The invention's effect [0021] As described above, according to the present invention, for each cluster to be classified, an optimal combination of feature amounts that is far from other clusters from a plurality of feature amounts of data to be classified is previously stored. It is set in advance, and the distance between the classification target data and each cluster is calculated, and the classification target data is classified into the cluster with the smallest calculated distance. The classification target data can be accurately classified into the corresponding clusters.

 In addition, according to the present invention, a plurality of the above combinations are set for each cluster, and the calculation result distances of all the clusters and the classification target data are arranged in order, and are included in a preset number of upper groups. Since the data to be classified is classified into the cluster with the largest number, classification can be performed with higher accuracy than in the past.

 Brief Description of Drawings

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

 FIG. 2 is a table explaining a process for selecting a feature set based on a discrimination reference value λ.

FIG. 3 is a table for explaining processing for feature set selection based on a discrimination reference value λ.

FIG. 4 is a diagram showing a histogram for explaining the effect of the discrimination criterion value λ on the feature set selection.

 FIG. 5 is a flowchart showing an operation example in a process of selecting a feature amount set for each cluster according to the first embodiment.

 FIG. 6 is a flowchart showing an operation example in clustering processing for classification target data according to the first embodiment.

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

 FIG. 8 is a flowchart showing an operation example in clustering processing for classification target data according to the second embodiment.

[FIG. 9] Another clustering process for the classification target data according to the second embodiment. It is a flowchart which shows the operation example which takes.

 FIG. 10 is a flowchart showing an operation example in clustering processing for classification target data according to the third embodiment.

[11] This is a flowchart showing an operation example for setting an arithmetic expression as a feature quantity conversion method.

 FIG. 12 is a flowchart showing an operation example of evaluation value calculation in the flowchart of FIG. 11.

[13] This is a flowchart showing an example of the operation of calculating the distance using the feature value converted by using the set conversion method.

 FIG. 14 is a table showing learning data belonging to each cluster.

[15] This is a result table showing the results of classifying the learning data in Fig. 14 by the clustering method of the conventional example.

圆 16] It is a conceptual diagram illustrating a method for calculating the overall correction determination rate.

 FIG. 17 is a result table showing the result of classifying the learning data of FIG. 14 by the clustering system in the first embodiment.

 FIG. 18 is a result table showing the results of classifying the learning data of FIG. 14 by the clustering system in the second embodiment.

 FIG. 19 is a result table showing the results of classifying the learning data of FIG. 14 by the clustering system in the second embodiment.

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

 FIG. 21 is a flowchart showing an operation example of feature quantity set selection in the inspection apparatus of FIG.

 FIG. 22 is a flowchart showing an operation example of clustering processing in the inspection apparatus of FIG.

圆 23] It is a block diagram showing a configuration example of a defect type determination device using the clustering system of the present invention.

FIG. 24 is a block diagram showing a configuration example of a manufacturing management apparatus using the clustering system of the present invention. FIG.

 FIG. 25 is a block diagram showing a configuration example of another manufacturing management apparatus using the clustering system of the present invention.

Explanation of symbols

 1 ··· Feature set creation part

 2. “Feature extraction unit

 3.Distance calculator

 4 · “Feature set storage unit

 5. Cluster database

 100 · Inspection object

 101 ··· Image acquisition unit

 102 ··· Lighting equipment

 103 · Imaging device

 104 · Defect candidate detection unit

 105 · “Clustering Department

 200, 300 ... Control device

 201, 202 ... Image acquisition device

 301, 302 ... Manufacturing equipment

 303 ... Notification Department

 304 ··· Recording section

BEST MODE FOR CARRYING OUT THE INVENTION

The clustering system of the present invention relates to a clustering system that classifies the input data to be classified into the clusters formed by using the learning data as a population according to the feature quantities of the input data, and corresponds to each of the clusters. A feature quantity set storage unit storing a feature quantity set that is a combination of feature quantities used for classification, and a feature quantity extraction unit based on the preset feature quantity set, the input data For each feature quantity set corresponding to each cluster, the distance calculation unit extracts the feature quantity from the feature quantity, and based on the feature quantity included in the feature quantity set, the distance between the population and the input data is obtained. The set distance is calculated, and the rank extraction unit arranges the set distances in ascending order and classifies them into clusters corresponding to the arrangement order.

 <First Embodiment>

 Hereinafter, a clustering 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 a configuration example of the clustering system according to the embodiment.

 As shown in FIG. 1, the clustering system of this embodiment has a feature value set creation unit 1, a feature value extraction unit 2, a distance calculation unit 3, a feature value set storage unit 4, and a cluster database 5.

 The feature value set storage unit 4 stores a feature value set indicating a combination of feature values of classification target data, which is individually set for each cluster, corresponding to the identification information of each cluster. For example, if the classification target data is a set of feature quantities {a, b, c, d}, the feature quantity set for each cluster is [a, b], [a, b, c, d], [c], etc. It is set as a combination of feature types. In the following description, a combination of all feature values or a combination of two or more (in the above example, any two forces or three feature values) from the feature value set, any one of them is referred to as “feature value”. Defined as “combinations of”.

[0026] Here, when clusters A, B, and C are set as classification destination clusters, the feature value set corresponding to each cluster uses learning data that is classified in advance in each cluster, and each cluster Is obtained as a combination of feature quantities having the largest distance from the other clusters, and is stored in the feature quantity set storage unit 4.

 For example, the feature quantity set for cluster A includes the vector that is the average value of each feature quantity of learning data belonging to cluster A, and each feature quantity of learning data belonging to other clusters B and C. The distance from the average vector is set as the largest combination of features.

 Further, the classification target data and the learning data of the population in each cluster are composed of the same set of features.

[0027] When calculating the distance to each cluster from the input classification target data, the feature amount extraction unit 2 reads the feature amount set corresponding to the cluster to be calculated from the feature amount set storage unit 4. The feature quantity corresponding to this feature quantity set is extracted from the plurality of feature quantities of the classification target data, and the extracted feature quantity is output to the distance calculation unit 3.

 The distance calculation unit 3 reads from the cluster database 5 a vector that is the average value of each feature quantity of the learning data of the cluster to be calculated, using the identification information of the cluster to be calculated as a key, and sets the feature quantity set of this cluster The distance between the vector of feature quantity extracted from the classification target data and the vector consisting of the average value of each feature quantity of the training data (the center of gravity center indicating the center of gravity of multiple learning data in the cluster) calculate.

[0028] When calculating the distance, the distance calculation unit 3 eliminates the data unit difference between the feature quantities and standardizes the numerical values between the feature quantities. For each feature v (i), regularity is performed.

 V (i) = (v (i) -avg. (I)) / std. (I)… hi)

 Here, v (i) is the feature value, avg. (I) is the average value of the feature value in the training data in the cluster to be calculated, and std. (I) is the value in the cluster to be calculated. This is the standardized deviation of the feature in the training data, and V (i) is the normalized feature. Therefore, when calculating the distance, the distance calculation unit 3 needs to perform standardization of each feature value for each feature value set.

 In addition, the distance calculation unit 3 performs the normalization process using the average value and the standard deviation of the corresponding feature values of the learning data for each feature value used for calculating the distance in the classification target data.

 [0029] As the distance,! /, Such as standardized Euclidean distance, Mahalanobis distance, Minkowsky distance (Minkowsk ydistance) using the above-mentioned standardized features, may be used! ,.

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

MHD = (lZn) '(V ^T R ^_1 V)---(2)

Each element V (i) in the matrix V in the equation (2) is an average value avg. (I) of the feature value of the learning data in the corresponding cluster with respect to the multidimensional feature value v (i) of the unknown data. And the standard deviation std. (I), the feature amount obtained by the above-described equation (1). n is the degree of freedom. For the state, indicate the number of features that is the number of features in the feature set (described later)!

. As a result, the Mahalanobis square distance is a value obtained by adding the differences of n transformed feature quantities, and the unit distance of the population average becomes 1 due to (Mahalanobis square distance) Zn. V ^T is a transposed matrix of matrix V whose elements are features v (i), and R ^_1 is an inverse matrix of correlation matrix R between each feature in the learning data in the cluster. is there.

 [0030] The feature quantity set creation unit 1 calculates a feature quantity set used when the distance calculation unit 3 calculates the distance between the classification target data and each cluster for each cluster, and calculates the calculation result for each cluster. Corresponding to the cluster identification information, it is written and stored in the feature value set storage unit 4.

 When calculating the feature quantity set, the feature quantity set creation unit 1 for each cluster, the centroid vector (barycentric vector) of the learning data belonging to the target cluster for generating the feature quantity set, and the other clusters excluding the target cluster. Based on the distance from the centroid vector of the learning data belonging to all the clusters, the value λ of the discriminant criterion is calculated by the following equation (3). Hereinafter, a combination of feature amounts will be described as a feature amount set.

[0031] λ = ω ω (μ-μ) ^Ζ / (ω σ ² + ω. Σ. ² ) · '· (3)

 In the above equation (3), μ i is a centroid vector that also serves as an average value of feature quantities in a feature quantity set of “learning data belonging to the target cluster (cluster population)”. σ is a standard deviation of a vector based on a feature amount of learning data belonging to the population in the cluster. ω is the ratio of the number of learning data belonging to the population in the cluster to the learning data belonging to all clusters. In addition, μ is a centroid vector that also represents the average value of feature values in the feature value set of “learning data belonging to clusters other than the target cluster (population outside target cluster)”. σ is a standard deviation of a vector based on a feature amount of learning data belonging to the population outside the target cluster. ω is the learning belonging to the non-cluster population in the learning data belonging to all clusters

 0

 It is the ratio of the number of data. Here, (μ to μ) ί log (logarithm) and a square root value in equation (3) may be used. In addition, when calculating each vector here, the feature value set creation unit 1 calculates and uses a feature value normalized for each feature value according to equation (1). Further, the eigenvalues may be set as numerical values for which the ratios co i and ω are calculated in advance and the separation becomes large.

[0032] Then, the feature value set creation unit 1 uses the above equation (3) for each target cluster and uses another class. The discrimination reference value with the raster is calculated for any force or all combinations of the feature values constituting the learning data, the calculated discrimination reference values are listed in descending order, and the order of the discrimination reference value λ is calculated. Output a list.

 Here, the feature quantity set creation unit 1 uses the combination of feature quantities corresponding to the largest discriminant reference value as the feature quantity set of the target cluster, along with the discriminant reference value and the cluster identification information. And stored in the feature value set storage unit 4.

 [0033] As shown in Fig. 2 (a), the determination criterion value λ is determined by the feature quantity set creation unit 1 when the feature data set of each cluster is set. If there are four feature values, a, b, c, and d, all of the four feature values, multiple, or any one of all combinations of the discrimination reference values λ are calculated.

 Then, the feature quantity set creation unit 1 selects the highest numerical value, for example, the combination of the feature quantities b and c in FIG. 2 (a).

 [0034] As another method for determining the reference value λ, as described in Fig. 2 (b), the BSS method, that is, a reference using all n feature quantities included in the set of classification target data. The value λ is calculated, and then the discriminant reference value is calculated for all combinations in which η−1 is extracted from the set of feature values η. Then, select the combination of the n−l discriminant reference value maximum values, and calculate the discriminant reference value for all the n−1 feature amount forces. To do. In this way, the feature amount is reduced by one by one in the order, and the discrimination reference value is calculated by selecting the combination further reduced by one from the reduced feature amount set, and reducing the feature amount. The feature quantity set creation unit 1 may be configured to select a combination that can be identified by the number of feature quantities.

[0035] Furthermore, as another method for determining the criterion value, as described in Fig. 2 (c), the FSS method, that is, all kinds of feature quantities from n feature quantities included in the set of classification target data are used. Are read one by one, the discrimination reference value λ of each feature quantity is calculated, and the feature quantity having the maximum discrimination reference value is selected from these. Next, a combination of two feature quantities, that is, the feature quantity and other feature quantities is generated, and a discrimination reference value for each combination is calculated. Then, the combination having the maximum discrimination reference value is selected from the combinations. Next, this combination and the combination that is included in this combination and has the three feature quantity powers are generated. And each discrimination reference value is generated. In this way, the feature quantity having the maximum discriminant reference value is selected sequentially from the combination of the immediately preceding feature quantities, and the feature quantity of the combination is increased by one with respect to the combination. The feature value of the increased combination feature value λ is calculated, the combination with the largest combination reference value λ is selected, and the feature value that is not present in this combination is increased by one feature. The feature value set is calculated so that the discrimination criterion value of the combination of amounts is calculated and, finally, the combination with the maximum discrimination criterion value is selected as the feature amount set from all the combinations for which the discrimination criterion value is calculated. Creation unit 1 may be configured.

 Next, FIGS. 3 and 4 show the effectiveness of selecting feature quantity sets used for clustering based on the discrimination reference value.

 FIG. 3 shows combinations of feature amounts a and g, combinations of feature amounts a and h, and feature amounts d and e as combinations for selecting feature amount sets from feature amounts a, b, c, d, and e. From these combinations, cluster 1 and clusters 2 and 3 will be described with reference to selection of feature quantity sets having higher classification characteristics than conventional examples.

 In FIG. 3, 1 corresponds to the above, 2 corresponds to the above, 1 corresponds to the above, σ 2 corresponds to the σ, ω ΐ corresponds to the ω., And ω 2 corresponds to the ω.

 [0037] Among them, the combination has the largest discriminant reference value λ, which is a combination of feature quantities a and h. This combination is defined as a combination of cluster 1 and other clusters. The classification results of cluster 1 and the other clusters (clusters 2 and 3) are confirmed in Fig. 4 for separation. In Fig. 4, the horizontal axis is the log of Mahalanobis distance calculated using the combination of features. The vertical axis represents the number of data to be separated (histogram) having a corresponding numerical value. Here, a numerical value of 1.4 on the horizontal axis means that the numerical value of the log of Mahalanobis distance is less than 1.4 and 1.2 or more (the value on the left side of 1.4). The same applies to the values on the other horizontal axes. In Fig. 4, 1.4≤ indicates that it is 1.4 or more. The Mahalanobis distance in Fig. 4 is calculated for the classification target data belonging to cluster 1 and other clusters using the feature set corresponding to cluster 1.

Fig. 4 (a) shows an example of calculating the Mahalanobis distance using a combination of features a and g. Fig. 4 (b) shows an example of calculating the Mahalanobis distance using a combination of features a and h. Fig. 4 (c) shows an example of calculating a Mahalanobis distance using a combination of features d and e. Looking at the histogram in Fig. 4, it can be seen that if the numerical value of the discrimination criterion is large, cluster 1 is well classified into other clusters.

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

 In the following description, for example, when the classification target data is a set of feature values of scratches attached to a glass article, these feature values are “a: length of scratch”, “b: scratch area”. ”,“ C: width of scratch ”,“ d: transmittance of a predetermined region including a scratch portion ”,“ e: reflectance of a predetermined region including a scratch ”, and the like are obtained from image processing and measurement results. Therefore, the set of feature values (hereinafter referred to as feature value set) is {a, b, c, d, e}. In this embodiment, the distance used for clustering is calculated as the Mahalanobis distance using the standardized feature value. Here, examples of the glass article in the present embodiment include plate glass and a glass substrate for display.

[0039] A. Feature value set creation processing (corresponding to the flowchart in FIG. 5)

 The user detects scratches on the glass, captures this image, obtains image data, and performs image processing to extract feature quantities such as measuring the length of the scratched part. Collecting feature value data consisting of the set of feature values. Then, the user sorts the feature data as learning data based on information such as the cause and shape that is known in advance for each cluster to be classified such as the cause and shape of the scratch, and the learning data of each cluster is sorted. A population is stored in the cluster database 5 in correspondence with the cluster identification information from a processing terminal (not shown) (step Sl).

[0040] Next, when a control instruction for generating a feature quantity set for each cluster is input from the processing terminal, the feature quantity set creation unit 1 corresponds to the identification information of each cluster from the cluster database 5. Read the population of learning data.

Then, the feature quantity set creation unit 1 performs each feature in the cluster population for each cluster. The average value and standard deviation of the quantities are calculated, and the standardized feature quantity in each learning data is calculated from the equation (1) using the average value and standard deviation.

 [0041] Next, the feature quantity set creation unit 1 calculates a discrimination reference value according to equation (3) for each feature quantity set of all combinations of feature quantities included in the feature quantity set.

 At this time, the feature value set creation unit 1 uses, for each cluster, the standardized feature values of the population in the cluster, and the average value of vectors (centroid vectors) μ that also have feature value power corresponding to each feature value set. And the standard deviation σ of the learning data vector corresponding to the feature quantity set in the population within the cluster and the standardized feature quantity of the population outside the cluster, to each feature quantity set. Corresponding feature value power vector mean value (centroid vector) μ, standard deviation σ of learning data vector consisting of features corresponding to feature quantity set in non-cluster population, and intra-cluster population in the total number of learning data The ratio ω of the number of learning data and the ratio ω of the number of learning data of the non-cluster population in the total number of learning data are calculated.

 [0042] Then, the feature quantity set creation unit 1 uses the centroid vectors, μ, standard deviations σ, σ, and ratios ω, ω to calculate the distance from other clusters for each cluster according to equation (3). The discriminant reference value for discriminating is calculated for each cluster feature amount set of all combinations of feature amount sets.

 When the calculation of all the discrimination reference values is completed, the feature quantity set creation unit 1 lists the discrimination reference values in the descending order for each cluster, and selects the feature quantity set corresponding to the largest discrimination reference value. When determining the affiliation to each cluster, it is detected as a feature value set indicating a set of feature value combinations used for calculating the distance (step S2).

 [0043] Next, the feature quantity set creation unit 1 uses the correlation coefficient R between the feature quantities corresponding to each feature quantity set and the population within each cluster for use in the distance calculation by the distance calculation unit 3. The average value avg. (I) and standard deviation std. (I) of the feature values of the learning data at are calculated (step S3).

[0044] Next, the feature value set creation unit 1 calculates a correction coefficient ^{_ (1/2)} from the discrimination reference value. This correction coefficient is standardized between each feature set. Since the distance from other clusters varies depending on the cluster, it is necessary to standardize the feature sets in order to increase the classification accuracy. In addition, the correction coefficient Toshitee ^{_ (1/2)} in Nag log ( λ), or simply (.-can be used. If it is a function that includes λ and can standardize between feature sets, it can be shifted.

 In the above equation (3), when calculating the center of gravity vector in the feature set of the population outside the target cluster, select one of the following three types as the learning data in the population outside the target cluster. To do.

 a. All learning data of the population outside the target cluster in all learning data

 b. Specific learning data corresponding to the purpose of classification in the population outside the target cluster c Learning data of the population outside the target cluster in the learning data used to select the feature amount Here, the purpose of classification in b. The learning data included in other clusters to which this difference is desired is used as the learning data.

Then, the feature quantity set creation unit 1 associates the identification information of each cluster with the feature quantity set, the correction coefficient corresponding to the feature quantity set, and in this embodiment fly_ ^(1/2) , The inverse matrix ^R_1 , the average value avg. (I), and the standard deviation std. (I) are stored as distance calculation data in the feature quantity set storage unit 4 (step S4).

[0045] B. Clustering processing (corresponding to the flowchart in FIG. 6)

 When the classification target data is input, the feature quantity extraction unit 2 reads a feature quantity set corresponding to each cluster from the feature quantity set storage unit 4 based on the identification signal of each cluster.

 Then, the feature quantity extraction unit 2 extracts the feature quantity from the classification target data for each cluster corresponding to the type of feature quantity in the read feature quantity set, and associates it with each cluster identification information. Then, the extracted feature value is stored in the internal storage unit (step Sl l).

[0046] Next, the distance calculation unit 3 obtains each feature amount extracted from the classification target data from the feature amount set storage unit 4 with an average value avg. (I) corresponding to the feature amount and a standard deviation std. (i) is read out, standardized by performing the calculation of equation (2), and the feature quantity stored in the internal storage unit is replaced with the standardized feature quantity.

Then, the distance calculation unit 3 generates the matrix V that also has the elemental force of V (i) obtained as described above, calculates the transposed matrix V ^T of this matrix V, and sequentially classifies it according to equation (3). The Mahalanobis distance between the target data and each cluster is calculated, and the internal data is recorded according to the identification information of each cluster. Store in memory (step S12).

[0047] Next, the distance calculation unit 3 multiplies the Mahalanobis distance of the calculation result by a correction coefficient ^{_ (1/2)} corresponding to the feature quantity set to obtain a correction distance, and each Mahalanobis distance Replace with separation (step S13). Also, when multiplying the correction factor, it may be multiplied after calculating the log or square root of the Mahalanobis distance.

 Then, the distance calculation unit 3 compares the correction distances between the clusters in the internal storage unit (step S14), detects the minimum correction distance, and identifies the cluster of identification information corresponding to the correction distances as the classification target. As the cluster to which the data belongs, the classified data to be classified is stored in the cluster database 5 in correspondence with the identification information of the cluster to be classified (step S15).

 <Second Embodiment>

 In the first embodiment described above, the feature amount set used for clustering is described as one type for each cluster. However, as in the second embodiment described below, the feature amount set is set for each cluster. Multiple, and calculate the Mahalanobis distance corresponding to each feature set, calculate the correction distance, rearrange the correction distances in ascending order, and in advance by the correction distance within the upper predetermined rank As a cluster to which the data to be classified belongs, according to the set rules.

 That is, the distance calculation unit 3 according to the present embodiment is set based on the distance ranking based on the distance between the classification target data obtained for each feature quantity set and each cluster. Based on the rule pattern indicating the classification criteria for each cluster of the classification target data, it is detected whether the classification target data belongs to the misaligned cluster.

 Hereinafter, 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 components, and only the operations different from those of the first embodiment in each configuration are illustrated in FIG. Use 7 to explain. In the second embodiment, there is a process for setting the rule pattern for the learning data power. FIG. 7 is a flowchart showing an example of pattern learning operation for the order of distances for setting rule patterns. 8 and 9 are flowcharts showing an example of clustering operation according to the second embodiment.

Further, in the first embodiment, when creating a feature quantity set, a feature quantity set creation unit 1 For each cluster, a discrimination reference value is calculated for a plurality of feature quantity sets as a combination of feature quantities, and a feature quantity set corresponding to the maximum value of the plurality of obtained discrimination reference values is calculated for each cluster. It was set as a feature amount set.

 On the other hand, in the second embodiment, the feature value set creation unit 1 has one or more combinations of other clusters for each cluster. By setting the maximum value of the feature quantity set corresponding to the number of combinations, a plurality of judgment reference values are obtained, and a plurality of feature quantity sets for separating each cluster from each other is set.

 Then, the feature quantity set creation unit 1 obtains distance calculation data for each feature quantity set, and associates a plurality of feature quantity sets with the distance calculation data of each feature quantity set in accordance with the cluster identification information. Stored in the feature set storage unit 4.

Then, in FIG. 7, when learning data is input, the feature quantity extraction unit 2 receives a plurality of feature quantity sets corresponding to each cluster from the feature quantity set storage unit 4 according to the identification signal of each cluster. read out.

 Then, the feature quantity extraction unit 2 extracts the feature quantity from the learning data for each cluster corresponding to the type of feature quantity in each read feature quantity set, and associates it with the identification information of each cluster. Then, the extracted feature quantity is stored in the internal storage unit for each feature quantity set (step S21).

 [0052] Next, the distance calculation unit 3 extracts each feature quantity extracted from the learning data from the feature quantity set storage unit 4 and the average value avg. (I) corresponding to the feature quantity for each feature quantity set and the standard The deviation std. (I) is read out and standardized by performing the calculation of equation (2) above, and the feature quantity stored in the internal storage unit is replaced with the standardized feature quantity.

Then, the distance calculation unit 3 generates a matrix V that also has the elemental force of V (i) obtained as described above, calculates a transposed matrix V ^T of this matrix V, and sequentially learns according to Equation (3). The Mahalanobis distance between the data and each cluster is calculated, and stored in the internal storage unit for each feature quantity set in correspondence with the identification information of each cluster (step S22).

[0053] Next, the distance calculation unit 3 multiplies the Mahalanobis distance of the calculation result by a correction coefficient ^{_ (1/2)} corresponding to the feature amount set to obtain a correction distance, and each Mahalanobis distance Replace with release (step S23).

 Then, the distance calculation unit 3 sorts the correction distances between the clusters in the internal storage unit in ascending order (sorts the smaller correction distances in a higher rank), that is, the correction distance from the classification target data becomes smaller. Then, arrange them in the order of higher cluster identification information (step S24).

 [0054] Next, the distance calculation unit 3 detects the identification information of the cluster corresponding to each of the correction distances from the smaller (higher) to the n-th, and the identification information for each cluster included in the n is detected. Count the number, that is, vote for each cluster.

 Then, the distance calculation unit 3 detects a rule pattern in which the number of identification information counts of each cluster of each learning data is common to the learning data included in the same cluster. For example, if n is set to 10 and the learning data for cluster B is detected as a count pattern with 5 clusters A, 3 clusters B, and 2 clusters C, this is detected. Is rule R1.

 Also, in the case of learning data for cluster C, if 3 clusters C are detected, 7 clusters A and 0 clusters B will always be cluster C. If the count of C is 3 or more, rule R2 is set to cluster C regardless of the count of other clusters.

 In addition, in the case of learning data for cluster A, when cluster A has the first and second patterns from the top, even if the number of counts for cluster B is 8, the number of counts for other clusters Regardless of rule R3, cluster A is used.

[0055] As described above, the regularity of the number of counts of each cluster included in each learning data classified into the same cluster is detected and stored internally as a pattern table for each piece of identification information of each cluster. Here, one rule may be set for each cluster, or a plurality of rules may be set. In the above description, the distance calculation unit 3 extracts the rule pattern. However, in order for the user to change the accuracy of classification into each cluster, the count number or the order rule pattern can be set arbitrarily. Also good.

Some clusters have similar characteristics to the characteristics of other clusters, and the relevance of multiple clusters, that is, the count number of each cluster or the pattern from the top In some cases, it is more accurate to classify the classification target data from the target pattern, and this embodiment complements this point.

 Next, the clustering process of the second embodiment using the rules described in the above table will be described with reference to the flowchart of FIG.

 When the classification target data is input, the feature quantity extraction unit 2 reads a plurality of feature quantity sets corresponding to each cluster from the feature quantity set storage unit 4 in accordance with the identification signal of each cluster. Then, the feature quantity extraction unit 2 extracts the feature quantity from the classification target data for each cluster corresponding to the feature quantity type in each read feature quantity set, and each of the cluster identification information. Correspondingly, the extracted feature quantity is stored in the internal storage unit for each feature quantity set (step S31).

 [0057] Next, the distance calculation unit 3 extracts each feature quantity extracted from the classification target data from the feature quantity set storage unit 4 for each feature quantity set by calculating an average value avg. (I) And the standard deviation std. (I) are read out and standardized by performing the calculation of equation (2), and the feature quantity stored in the internal storage unit is replaced with the standardized feature quantity.

 Then, the distance calculation unit 3 generates a matrix V that also has the elemental force of V (i) obtained as described above.

Then, the transpose matrix V ^T of this matrix V is calculated, and the Mahalanobis distance between the data to be classified and each cluster is calculated sequentially by equation (3), and each feature is correlated with the identification information of each cluster. Each quantity set is stored in the internal storage unit (step S32).

[0058] Next, the distance calculation unit 3 multiplies the Mahalanobis distance of the calculation result by a correction coefficient ^{_ (1/2)} corresponding to the feature amount set to obtain a correction distance, and each Mahalanobis distance Replace with release (step S33).

 Then, the distance calculation unit 3 rearranges the correction distances between the clusters in the internal storage unit in ascending order, that is, arranges the identification information of the clusters with the small correction distances to the classification target data in the order of higher rank (step S34).

 After rearrangement, the distance calculation unit 3 detects the identification information of the clusters corresponding to the correction distances from the smaller (higher) to the nth, and calculates the number of identification information for each cluster included in the n pieces. Counting, that is, voting is performed for each cluster.

[0059] Next, the distance calculation unit 3 applies to each cluster in the top n pieces of each classification target data. Count number pattern (or arrangement pattern) 1S Check whether it exists in the table stored in the inside (step S35).

 When the distance calculation unit 3 detects that the rule pattern matching the target pattern of the classification target data is described in the table as a result of the above-described collation, the distance calculation unit 3 identifies the classification target data corresponding to the matched rule. It is determined that it belongs to the cluster of information, and the classification target data is classified into this cluster (step S36).

Further, another clustering process of the second embodiment using the rules described in the above-described table will be described with reference to the flowchart of FIG.

 In the other clustering processing shown in FIG. 9, the processing from step S31 to step S35 is the same as the processing shown in FIG. 8, and the distance calculation unit 3 stores the table in the table as already described in step S35. Based on the stored rule pattern, collation processing with the target pattern of the classification target data is performed.

 Then, the distance calculation unit 3 detects whether or not a rule pattern that matches the target pattern is found in the collation result. If the rule calculation unit 3 detects that a rule pattern that matches the target pattern is found, the process proceeds to step S47. On the other hand, if it is detected that no matching rule pattern is found, the process proceeds to step S48 (step S46).

[0061] When it is detected that a matching rule pattern has been searched, the distance calculation unit 3 determines that this classification target data belongs to the cluster of identification information corresponding to the matching rule, and determines the classification target data. The classification target data is classified and stored in the cluster database 5 in correspondence with the identification information of the classification destination cluster (step S47). On the other hand, when it is detected that the matching rule pattern is not searched, the distance calculation unit 3 detects the identification information having the largest number of counts, that is, the number of votes, and classifies the classification target data into the cluster corresponding to the identification information. To do.

 Then, the distance calculation unit 3 stores the classified target data in the cluster database 5 in association with the identification information of the cluster to which it belongs (step S48).

<Third Embodiment>

The second embodiment described above prepares a table of rule patterns in the top n from the V V (similarity is large) distance from each cluster of the calculated classification target data. Although it has been described that clustering processing is performed on each classification target data depending on whether or not the rule pattern corresponding to the rule pattern in the table is selected, the feature amount is set for each cluster as in the third embodiment described below. Set multiple sets, calculate the Mahalanobis distance corresponding to each feature set, calculate the correction distance, and select the cluster with the most correction distances within the upper predetermined rank as the cluster to which the classification target data belongs. Also good.

 Hereinafter, 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 components, and only the operations different from those of the second embodiment are performed in the respective configurations. Is explained using FIG. In the third embodiment, step S48 in FIG. 9 is performed directly after the process of setting the rules from the learning data. FIG. 10 is a flowchart showing an example of clustering operation in the third embodiment.

[0063] 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 calculation unit 3 performs the step as described above. In S34, the correction distances between the clusters in the internal storage unit are rearranged in ascending order, that is, the correction distances from the classification target data are arranged in ascending order of the cluster identification information (step S34).

 Next, the distance calculation unit 3 detects cluster identification information corresponding to each of the correction distances from the smaller (higher) to the nth, and counts the number of identification information for each cluster included in the n. In other words, a voting process is performed for each cluster (step S55).

 Then, the distance calculation unit 3 detects the identification information of the largest count value (the number of votes) in the voting result, and designates the cluster corresponding to this identification information as the cluster to which the classification target data belongs, with respect to the cluster database 5. Then, the classified data to be classified is stored in correspondence with the identification information of the cluster to which it belongs (step S 56).

[0064] In addition, when the user sets a threshold value for the number of votes in advance in the distance calculation unit 3 for each identification information, and the number of votes for the identification information with the largest number of votes is less than this threshold value, Do not belong to any cluster.

For example, when classifying data to be classified for three clusters A, B, and C, the number of votes for the identification information of cluster A is five, and the number of votes for the identification information of cluster B is five. If there are three and the number of votes for cluster C is two, the most votes Cluster A and the distance calculation unit 3 detect the identification information.

 However, if the above threshold value for cluster A is set to 6, the distance calculation unit 3 must not belong to any cluster because the number of votes for the identification information of cluster A is less than the threshold value. Judgment is made.

 As a result, it is possible to improve the reliability of the classification process for the cluster of the classification target data in the clustering for the cluster whose feature quantity is not slightly different from other clusters.

[0065] <Feature Conversion Method>

 Clustering is performed with the expectation that the population of each feature is a normal distribution, but depending on the type of feature (area, length, etc.), the distribution may not be a normal distribution and the population may be biased. In some cases, the calculation of the distance between the classification target data and each cluster, that is, the accuracy in determining the similarity between the classification target data and each cluster may be reduced. Therefore, depending on the feature value, it is necessary to convert the feature value of the population by a predetermined method and to improve the accuracy of similarity determination by bringing it close to the normal distribution.

This normal distribution can be converted to either a logarithm, n-root such as square root (), cubic root ( ³ ), factorial, or an arithmetic expression including a function obtained by numerical calculation. Convert.

 [0066] Hereinafter, setting processing of a conversion method for each feature amount will be described with reference to FIG. FIG. 11 is a flowchart showing an operation example of the setting process of the conversion method of each feature quantity. This conversion method is set for each cluster in units of feature values included in the cluster. This conversion method is set using learning data belonging to each cluster. The following processing is described as being performed by the feature value set creation unit 1, but it is not necessary to provide another processing unit corresponding to this processing.

 The feature value set creation unit 1 uses the identification information of the cluster to be classified as a key, reads the learning data included in this cluster from the cluster database 5, and calculates (normalizes) the feature value of each learning data (step S61). ).

[0067] Next, the feature value set creation unit 1 performs feature value conversion by calculating each of the read learning data using any of the arithmetic expressions for performing feature value conversion stored therein. (Step S62).

 When the conversion of the feature values of all learning data is completed, the feature value set creation unit 1 calculates an evaluation value indicating whether the distribution obtained by the conversion process is close to the normal distribution (step S63) .

 [0068] Next, the feature value set creation unit 1 detects whether or not the force is calculated by calculating the evaluation value for all the arithmetic expressions stored therein, that is, preset as a conversion method. , And if it is detected that the evaluation value of the distribution obtained by converting the feature values in all the arithmetic expressions is calculated, the process proceeds to step S65, while If the calculation of the feature quantity is completed and it is detected, the process returns to step S62 (step S64) in order to process the arithmetic expression that is set next.

 When the feature value conversion by all the calculation formulas is completed, the feature value set creation unit 1 detects the distribution with the smallest evaluation value in the distribution obtained in the set calculation formula, that is, the distribution closest to the normal distribution. Then, the arithmetic expression used to create the detected distribution is determined as a conversion method and set internally as a conversion method for the feature quantity of the cluster (step S65).

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

[0069] Next, the calculation of the evaluation value in step S63 will be described with reference to FIG. Fig. 12 is a flow chart for explaining an example of the processing for obtaining the evaluation value of the distribution obtained by the arithmetic expression.

 The feature quantity set creation unit 1 converts the feature quantity of each learning data belonging to the target cluster by the set arithmetic expression (step S71).

 After converting the feature values of all the learning data, the feature value set creation unit 1 calculates the average value μ and standard deviation σ of the distribution (population) obtained from the converted feature values (step) S72).

 Then, the feature value set creation unit 1 calculates the ζ value (1) from (X− / ζ) Ζσ using the average value of the population and the standard deviation σ (step S73).

[0070] Next, the feature value set creation unit 1 calculates the cumulative probability in the population (steps). P74).

 After the calculation, the feature value set creation unit 1 calculates the z value (2) as the value of the inverse function of the cumulative distribution function of the standard normal distribution based on the calculated cumulative probability in the population (step S75).

 Then, the feature quantity set creation unit 1 calculates the difference between the two z values of the feature quantity distribution, that is, the difference between the z value (1) and the z value (2), that is, the error between the two z values in the distribution (step S76). . When the error of the z value is obtained, the feature value set creation unit 1 calculates the sum of the errors of the two z values, that is, the sum of the errors (square sum) as an evaluation value (step S77).

 The smaller the error between the two z-values mentioned above, the closer the distribution is to the normal distribution if there is no z-value error. On the other hand, the larger the error is, the larger the error is.

Next, calculation of the feature quantity of the classification target data will be described with reference to FIG. 13 before performing the clustering process in the first to third embodiments. FIG. 13 is a flowchart showing an operation example of calculating feature amount data of classification target data.

 The distance calculation unit 3 extracts the feature quantity of the identification target from the input classification target data corresponding to the feature quantity set set for each cluster, and performs the normalization process already described (step S81). .

 Next, the distance calculation unit 3 converts the feature quantity used for classification into the cluster to be classified in the classification target data V according to the conversion method (arithmetic formula) set for the feature quantity of this cluster. (Step S82).

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

[0072] Next, the distance calculation unit 3 converts the feature amounts for all the clusters to be classified by a conversion method set in correspondence with the feature amounts of each cluster, and the cluster and the converted feature amounts. If it is detected that the distance is calculated for all the clusters to be classified, the process proceeds to step S85, while the cluster to be classified is detected. If it is detected that remains, the process returns to step S82 (step S84). Then, in each of the first to third embodiments, processing of the point-in-time force when the distance calculation is completed is started (step S85).

With the processing described above, the Mahalanobis distance used in this embodiment is When calculating the distance between the target data and each cluster, we expect that the feature quantity is a normal distribution. Therefore, the closer the distribution of each feature quantity of the population is to the normal distribution, the closer to each cluster. In addition, it can be expected that an accurate distance (similarity) can be obtained and the accuracy of classification for each cluster is improved.

 Example

 [0073] <Calculation example>

 Next, using the clustering systems of the first, second, and third embodiments described above, the accuracy of classification with the conventional example based on the sample data shown in FIG. 14 was confirmed. Although the number of samples is small, it can be seen that the correct answer rate is higher than that of the conventional example, even though the number of features used is small. In Fig. 14, 10 learning data are defined for each of category 1, force category 2 and category 3 as clusters, and each learning data has feature quantities a, b, c, d, e, f, g, h There are eight of them. In this example, a feature amount set used for clustering is determined from the learning data belonging to each cluster shown in FIG. 14, and then clustering is similarly performed using the learning set as classification target data.

[0074] As a calculation result, Fig. 15 shows a conventional calculation method, using feature amounts a and g as a combination of feature amounts, and for each learning data shown in Fig. 14 of cluster 1 to cluster 3, Mahalanobis. The distance is calculated and the determination result is shown. In Fig. 15 (a), the cluster column is the Mahalanobis distance to cluster 1, the cluster 2 column is the Mahalanobis distance to cluster 2, and the cluster 3 column is the Mahalanobis distance to cluster 3. The category column indicates the cluster to which each learning data actually belongs, and the judgment result indicates the learning data and the mano, and the cluster with the smallest Ranobis distance. The numbers in the category and the judgment result match! The feature data that is correctly classified is displayed.

In FIG. 15B, the column number indicates the cluster to which the learning data actually belongs, and the row number indicates the determined cluster. For example, “8” in mark R1 is determined as cluster 1 in 8 of 10 clusters in cluster 1, and “2” in mark R2 is determined as cluster 3 in 2 out of 10 clusters in cluster 1. It has been shown. ρθ indicates the match rate between correct answer and answer, pi indicates the probability of coincidence of both, and κ is the overall correction determination rate, which can be obtained by the following formula. The higher κ, the higher the classification accuracy. κ = (p0 -pl) / (l -pl)

 pO = (a + d) / (a + b + c + d)

pi = [(a + b) · (a + c) · (b + d) · (c + d)] · (a + b + c + d) ²

[0076] The relationship between a, b, c, and d in the above equation will be described with reference to FIG.

 The data belonging to cluster 1 is the number power classified as cluster 1, the data belonging to cluster 1 is the number power ¾ classified as cluster 2, and a + b indicates the number of data belonging to cluster 1 . Similarly, the number of data belonging to cluster 2 classified as cluster 2 is d, the number of data belonging to cluster 2 classified as cluster 1 is c, and c + d is data belonging to cluster 2 Show the number! / a + c is the number classified as cluster 1 in all data a + b + c + d, b + d is the number classified as cluster b in all data a + b + c + d It is.

Next, FIG. 17 shows the determination result by calculating the Mahalanobis distance for each learning data shown in FIG. 14 of cluster 1 to cluster 3 using the calculation method of the first embodiment. . The way of viewing FIGS. 17 (a) and (b) is the same as in FIG. 15, and therefore the description thereof is omitted. It can be seen that the correct answer rate ρθ, the coincidence probability pi, and the overall correction judgment rate _K are equivalent to the conventional calculation method in FIG. Here, a feature value set corresponding to each cluster was calculated using a method of selecting a combination having the maximum discriminant reference value for each cluster from the entire combinations described above. The feature set corresponding to cluster 1 is a combination of feature quantities a and h, the feature set corresponding to cluster 2 is a combination of feature quantities a and d, and the feature set corresponding to cluster 3 is A combination of features a and g was used.

Next, FIG. 18 shows the determination result by calculating the Mahalanobis distance for each learning data shown in FIG. 14 of cluster 1 to cluster 3 using the calculation method of the second embodiment. . 18 (a) and 18 (b) are the same as FIG. 15 and will not be described. The correct answer rate pO is 0.8333, the probability of coincidence pi is 0.3333, the overall correction judgment rate κ is 0.75, and the classification accuracy is improved compared to the conventional calculation method of Fig. 15. You can see that Here, each class is selected using a method of selecting the combination having the third highest discriminant reference value λ for each cluster from the above-described total combinations. The feature amount set corresponding to the star was calculated. As a feature set corresponding to cluster 1, three combinations of feature quantities a 'h, ag, d' e are used. As a feature set corresponding to cluster 2, feature quantities a'f, a-d , a 'b, and three combinations of feature quantities e' g, a-c, and a 'g were used as the feature quantity set corresponding to cluster 3.

 In addition, for voting, we calculated the number of clusters that entered the third from the smallest Mahalanobis distance, and calculated the number of the third cluster from the smallest, and set the largest number of clusters to which the classification target data belongs.

Next, FIG. 19 uses the calculation method of the second embodiment, calculates the Mahalanobis distance for each learning data shown in FIG. 14 for cluster 1 to cluster 3, and further calculates the Mahalanobis distance of the calculation result. ^Is multiplied by the correction coefficient (λ) ^{_1 / 2} , and then the ranking of the distance is performed, and the determination result is shown. The way of viewing FIGS. 19 (a) and 19 (b) is the same as FIG. The correct answer rate ρθ is 0.8333, the probability of coincidence pi is 0.3333, the overall correction judgment rate κ is 0.75, and the classification accuracy is improved compared to the conventional calculation method of Fig. 15. You can see that Here, a feature value set corresponding to each cluster was calculated using a method of selecting combinations having the third highest discriminant reference value λ for each cluster from the above-described total combination power. The feature set corresponding to cluster 1 uses three combinations of feature quantities a'h, a -g, d 'e, and the feature set corresponding to cluster 2 is feature quantity a'f, a-d, Three combinations of a 'b were used, and three combinations of features e' g, a- c, and a 'g were used as the feature set corresponding to cluster 3, and the Mahalanobis distance was used for voting The number of clusters that fall in order from the smallest number to the third is calculated, and the largest number of clusters is the cluster to which the classification target data belongs.

[0080] From the classification results shown in FIGS. 15, 17, 18, and 19 described above, it can be seen that the clustering process of this embodiment is faster and more accurate than the conventional example. The superiority of the form over the conventional example was confirmed.

<Application example of the present invention>

A. Inspection equipment As shown in FIG. 20, an inspection apparatus (defect detection apparatus) for classifying the types of scratches on the surface of an inspection object, for example, a glass substrate will be described. FIG. 21 is a flowchart for explaining an operation example for selecting a feature quantity set, and FIG. 22 is a flowchart for explaining an operation example in the clustering process.

 First, the feature quantity set selection operation will be described. The learning data collection capability in step S1 in the flowchart of FIG. 5 corresponds to steps S101 to S105 in the flowchart of FIG.

 Steps S2 to S4 in FIG. 21 are the same as those in the flowchart in FIG.

 [0082] By the operation of the operator, a sample for learning data corresponding to each cluster to be classified into scratch types is collected (step S101).

 The flaw shape collected as learning data by the image acquisition unit 101 is irradiated by the illumination device 102, and the image data of the flaw portion is acquired by the imaging device 103 (step S102). Then, the flaw feature amount of each learning data is calculated from the image data acquired by the image acquisition unit 101 (step S 103).

 The feature quantities of the obtained learning data are assigned to the classification destinations obtained visually, and the learning data in each cluster is specified (step S104).

 Then, the processing from step S101 to step S102 is repeated until the learning data of each cluster reaches a predetermined number (preset number of samples), for example, about 300 pieces each. The clustering unit 105 performs the processing after step S2. Here, the clustering unit 105 is the clustering system in the first or second embodiment.

 Next, with reference to FIG. 22, the clustering process in the inspection apparatus of FIG. 4 will be described.

 Here, steps S31 to S34, S55, and S56 in FIG. 22 are the same as the flowchart in FIG.

In the inspection apparatus of FIG. 20, when inspection is started, the illumination device 102 illuminates the glass substrate that is the object to be inspected 100, and the imaging device 103 captures the surface of the glass substrate and images the captured image. Output to the acquisition unit 101. As a result, the defect candidate detection unit 104 When a portion different from the planar shape is detected in the captured image input from the image acquisition unit 101, it is set as a defect candidate to be classified (step S201).

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

 Then, the defect candidate detection unit 104 calculates the feature amount from the image data of the classification target data, and outputs the classification target data including the extracted feature amount set to the clustering unit 105 (step S 202).

 The subsequent clustering process has already been described in step 10 in FIG. As described above, the inspection apparatus of the present invention can classify scratches on a glass substrate with high accuracy for each type of scratch.

[0085] B. Defect type determination device

 The defect type determination apparatus shown in FIG. 23 corresponds to the clustering system of the present invention already described by the clustering unit 105.

 The image acquisition device 201 includes the image acquisition unit 101, the illumination device 102, and the imaging device 103 in FIG.

 The learning data of each cluster to which the classification target data is classified has already been acquired and prepared in the cluster database 5 of the clustering apparatus 105. Therefore, the feature set selection in FIG. 5 is also completed.

[0086] A defect candidate is detected from the captured image input from the image acquisition device 202 attached to each manufacturing device !, and the image data is cut out, and the feature amount is extracted and output to the data collection device 203. To do. The control device 200 transfers the classification target data input to the data collection device 203 to the clustering unit 105. As described above, the clustering unit 105 classifies the input classification target data for each cluster corresponding to the type of scratch.

[0087] C. Manufacturing management device

As shown in FIG. 24, the manufacturing management apparatus of the present invention is composed of a control device 300, manufacturing devices 301, 302, a notification unit 303, a recording unit 304, a defective device determination unit 305, and a defect type determination device 306. ing. Here, the defect type determination device 306 is the defect described in the section B above. This is the same as the type determination device.

 The defect type determination device 306 performs image processing on the captured images from the image acquisition devices 201 and 202 provided in the manufacturing device 301 and the manufacturing device 302, respectively, and performs feature processing on the corresponding defect candidate detection unit 104. And classify the data to be classified.

Next, the defective device determination unit 305 has a table indicating the relationship between the identification information of the classified cluster and the generation factor corresponding to the cluster, and is input from the defect type determination device 306. The generation factor corresponding to the identification information of the cluster to be classified is read from the table, and the manufacturing apparatus that is the generation factor is determined. That is, the defective device determination unit 305 detects the cause of the defect in the product manufacturing process in accordance with the cluster identification information.

 Then, the defective device determination unit 305 notifies the operator from the notification unit 303, and also notifies the recording unit 304 of the identification number of the cluster into which the defect is classified, the generation factor, and the manufacturing process corresponding to the determined date and time. The device identification information is stored as a history. Further, the control device 300 stops the manufacturing device determined by the defective device determination unit 305 or controls the control parameters.

[0089] D. Manufacturing management device

 As shown in FIG. 25, another production management apparatus of the present invention includes a control device 300, production devices 301 and 302, a notification unit 303, a recording unit 304, and a clustering unit 105. Here, the clustering unit 105 has the same configuration as that described in the above sections A and B.

 In the clustering unit 105, unlike the cases A to C described above, the feature data of the classification target data is the manufacturing condition (material quantity, processing temperature, pressure, processing speed, etc.) force in the manufacturing process of industrial products, for example, glass substrates. Are classified according to the manufacturing status of each stage of the manufacturing process. The feature amount is input as a feature amount to the clustering unit 105 as process information detected by sensors provided in the respective manufacturing apparatuses 301 and 302.

[0090] That is, the clustering unit 105 needs to adjust the manufacturing state of the glass manufacturing process in each process of each manufacturing apparatus according to the feature quantity of the classification target data as “normal state” and “defects are likely to occur. Status ”,“ dangerous and needs adjustment ”, etc. Similar. Then, the clustering unit 105 notifies the operator of the classification result by the notification unit 303, outputs the cluster identification information of the classification result to the control device 300, and also notifies the recording unit 304 at the determined date and time. Correspondingly, the identification number of the classified cluster of the manufacturing state of each process, the manufacturing condition which is the most problematic feature amount, and the identification information of the manufacturing apparatus are stored as history.

 The control device 300 has a table indicating the correspondence between the cluster identification information and the adjustment items for returning the manufacturing conditions to normal and the data, and the manufacturing apparatus corresponding to the cluster identification information input from the clustering unit 105. The adjustment items that return the conditions to normal and their data are read, and the corresponding manufacturing equipment is controlled by the read data.

[0091] It should be noted that a program for realizing the functions of the clustering system in FIG. 1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into the computer system and executed. You may perform clustering processing of the classification target data. The “computer system” here includes OS and hardware such as peripheral devices. “Computer system” also includes a WWW system equipped with a home page provision environment (or display environment). The “computer-readable recording medium” means a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built in the computer system. Furthermore, the “computer-readable recording medium” means a volatile memory (RA) inside a computer system that becomes a server or client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. As shown in M), the program is held for a certain period of time.

[0092] The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting a program refers to 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. The program may be for realizing a part of the functions described above. Furthermore, what can realize the above-mentioned functions in combination with programs already recorded in the computer system, so-called differences. Even a minute file (difference program).

Industrial applicability

 The present invention can be applied to the field of classifying and discriminating information having various types of features with high accuracy, such as detection of defects in glass articles, etc., and can also be used for manufacturing state detection devices and product manufacturing management devices. . It should be noted that the entire contents of the specification, patent claims, drawings and abstract of Japanese Patent Application 2006-186628 filed on July 6, 2006 are incorporated herein by reference. It is something that is incorporated.

Claims

The scope of the claims  [1] In a clustering system that classifies input data according to the features of the input data into each cluster formed by the population of learning data,  A feature value set that is a combination of feature values used for classification is stored corresponding to each of the clusters, and a feature value set storage unit,  A feature quantity extraction unit that extracts preset feature quantities from the input data, and for each feature quantity set corresponding to each cluster, based on the feature quantities included in the feature quantity set, the population of each cluster A distance calculation unit that calculates and outputs the distance between the center and the input data as a set distance,  A rank extracting unit for arranging the set distances in ascending order;  A clustering system characterized by comprising:  [2] The clustering system according to claim 1, wherein a plurality of the feature quantity sets are set for each cluster.  [3] According to the rule pattern indicating the classification criteria for each cluster of the input data set based on the set distance obtained for each feature quantity set and based on the rank of the set distance, the input data The clustering system according to claim 2, further comprising a cluster classification unit that detects which cluster belongs to.  [4] The cluster classification unit detects which cluster the input data belongs to based on the rank of the set distance, and a cluster having a large set distance with the rank higher belongs to the input data. The clustering system according to claim 3, wherein the clustering system is detected as a cluster.  5. The cluster classification unit according to claim 4, wherein the cluster classification unit has a threshold for the number of ranks higher than the rank, and detects a cluster to which input data belongs if the higher rank cluster is equal to or higher than the threshold. Clustering system.  [6] The distance calculation unit is characterized by multiplying the set distance corresponding to the feature amount set by a correction coefficient and standardizing a set distance between each feature amount set. Item 6. The clustering system according to any one of Items 1 to 5. [7] It further includes a feature quantity set creation unit that creates a feature quantity set for each cluster, For each of a plurality of combinations of feature amounts, the feature value set creation unit uses the average value of the learning data of the population of each cluster as the origin, and calculates the distance between this origin and each learning data of the population of other clusters. The average value is obtained, and the combination of the feature values having the largest average value is selected as a feature value set used for distinguishing each cluster from other clusters. V, the clustering system described in somewhere. [8] The clustering system according to any one of claims 1 to 7 is provided, wherein the input data is image data of a product defect, and the defect in the image data is detected by a feature amount indicating the defect. Defect type determination device for classifying by defect type. [9] The defect type determination device according to claim 8, wherein the product is a glass article, and the defects of the glass article are classified by defect type. [10] A defect detection apparatus for detecting a defect type of a product, provided with the defect type determination apparatus according to claim 8 or 9. [11] The defect type judging device according to claim 8 or claim 9 is provided, the type of defect of the product is determined, and the cause of the defect in the manufacturing process based on the correspondence with the cause corresponding to the type Manufacturing state determination device that detects the above. [12] The clustering system according to any one of claims 1 to 7 is provided, and the input data is a feature value indicating a manufacturing condition in a product manufacturing process, and the feature value is converted into a manufacturing process. The manufacturing state determination apparatus which classifies according to the manufacturing state of each process. 13. The manufacturing state determination device according to claim 12, wherein the product is a glass article, and the feature amount in the manufacturing process of the glass article is classified according to the manufacturing state of each step of the manufacturing process.  [14] A manufacturing state detecting device provided with the manufacturing state determining device according to claim 12 or 13, for detecting a type of manufacturing state in each step of a product manufacturing process.  [15] The manufacturing state determination device according to claim 12 or claim 13 is provided to detect the type of manufacturing state in each step of the product manufacturing process, and based on the control item corresponding to the type, Product manufacturing management device for process control in the manufacturing process