JP4431988B2 - Knowledge creating apparatus and knowledge creating method - Google Patents

Description

  The present invention relates to a knowledge creation device and a knowledge creation method. More specifically, the present invention relates to a test for extracting feature quantities from input measurement data to be examined and determining a state based on the extracted feature quantities. The present invention relates to a technique for generating inspection logic (effective feature amount, parameter adjustment of feature amount, etc.) in an apparatus.

  In automobiles and home appliances, a rotating device in which a drive system component such as a motor is incorporated is used very often. For example, taking an automobile as an example, rotating equipment is mounted everywhere in the engine, power steering, power seat, mission and others. Household appliances include refrigerators, air conditioners, washing machines and various other products. When the rotating device is actually operated, a sound is generated with the rotation of the motor or the like.

  Some of these sounds are inevitably generated along with normal operation, and other sounds are generated due to defects (failures). Examples of abnormal sounds associated with the failure include bearing abnormalities, internal abnormal contact, imbalance, and foreign matter contamination. More specifically, there are gear chipping, foreign object biting, spot flaws, and abnormal noise such that the rotating part and the fixed part inside the motor rub for a moment during rotation. Further, as sounds that people feel uncomfortable, for example, there are various sounds from 20 Hz to 20 kHz that humans can hear, for example, about 15 kHz. And when the sound of the predetermined frequency component is generated, it becomes an abnormal sound. Of course, abnormal sounds are not limited to this frequency.

  The sound associated with such a defect is not only unpleasant, but may cause further failure. Therefore, for the purpose of quality assurance for each of these products, production factories usually perform “sensory inspection” that relies on the five senses such as hearing and tactile sensation, and determine the presence or absence of abnormal sounds. Specifically, it is done by listening with the ear or touching it with the hand to check the vibration. The sensory test is defined by the sensory test term JIS Z8144.

  By the way, in the sensory test that relies on the five senses of such an inspector, not only skill is required, but also the determination results vary widely such as individual differences and changes with time. Furthermore, there is a problem that it is difficult to convert the determination results into data and numerical values and to manage them. Therefore, in order to solve such a problem, there is an abnormal sound inspection device for the purpose of a stable inspection based on a quantitative and clear standard as an inspection device for inspecting an abnormality of a product including a drive system component.

  As an abnormal sound inspection apparatus that automatically performs an inspection (so-called abnormal sound inspection) for determining normality / abnormality from a vibration waveform obtained from an inspection object as described above, there is one disclosed in Patent Document 1 in the past. The invention disclosed in Patent Document 1 comprehensively determines normality / abnormality of an inspection object using a feature amount obtained from a time axis waveform and a feature amount obtained from a frequency waveform.

  The reason why the abnormal sound inspection is comprehensively performed based on waveforms obtained from different axes such as a time axis waveform and a frequency axis waveform is as follows. In other words, it is difficult to detect all abnormal sounds using the abnormal sound inspection only for the feature amount obtained from the time-axis waveform and the abnormal sound inspection only for the feature amount obtained from the frequency axis waveform. . This is because each feature quantity is good and bad. The abnormal sound inspection using a plurality of feature amounts has a higher discrimination ability than the abnormal sound inspection using a single feature amount.

  In other words, the drive system component is composed of a mechanism that repeats rotation and reciprocation in the first place, and if there is a slight mechanical abnormality in the mechanism, the abnormal component (what is different from the normal component emitted from a good product) is different. Component) is always transmitted to the surroundings as vibration and sound. However, abnormal components in abnormal noise tests are only slight differences in vibration and sound waveforms compared to normal components, even if there are differences that can be discerned by skilled human ears. When I analyzed the waveform, it was buried in noise and could not be detected well. This is because the conventional abnormal sound inspection was performed only on the feature amount obtained from the time axis waveform, or only on the feature amount obtained from the frequency axis waveform, and was performed based only on a single feature amount. It is. Therefore, in Patent Document 1 described above, normal / abnormal is determined based on a plurality of feature amounts obtained from a plurality of axes. In the invention disclosed in Patent Document 1, a fuzzy rule is used as a discrimination rule, and normality / abnormality is determined based on a plurality of feature amounts by fuzzy inference.

  By the way, the fuzzy reasoning used as a discrimination rule for the abnormal sound inspection disclosed in Patent Document 1 has an advantage that a person can easily understand the discrimination rule as compared with other discrimination models such as a neural network. For example, a neural network is a network in which many neuron models are connected to each other, and it is difficult to understand and understand the basis of what kind of discrimination was made. It's hard. It is difficult for people to trust things that they cannot understand sensuously. Especially if it is an inspection device that is the key to quality.

  On the other hand, fuzzy inference uses a membership function that expresses ambiguity, and the discrimination rule using fuzzy inference associates the basis of discrimination with the discrimination result, and “IF feature A = large THEN anomaly Can be expressed in an easy-to-understand manner. Those that can be understood sensuously are easy to explain, and when the quality solution is a business, it is easy to explain the discrimination rules as the inspection logic of the inspection device, so the customer who received the explanation is also highly convinced So there is an advantage that can be adopted with confidence.

  In addition, many customers who intend to introduce a new abnormal sound inspection device have been performing sensory inspections by the ears of experts (sensory inspectors) until then. The description already has its own criteria, know-how and knowledge for general inspection standards. In such a case, since the abnormal sound inspection device replaces the sensory inspection that the sensory inspector has performed so far, the consistency with the judgment criteria, know-how and knowledge possessed by the sensory inspector is naturally required. Currently. Even in such cases, the fact that it is easy to explain the consistency between the created discrimination rules and the knowledge (inspection criteria) that the sensory inspector had so far is a fuzzy for solution providers who are accountable to customers. Ease of explanation by reasoning is a great advantage for business.

  By the way, as the number of feature quantities to be used increases, the determination rules for determining pass / fail judgment become more complicated or more numerous. Therefore, in order to perform an abnormal noise inspection with high accuracy, it is necessary to create a discrimination rule with high accuracy. There is a technique disclosed in Non-Patent Document 1 as a technique for reducing the number of steps for creating a discrimination rule in abnormal noise inspection. This Non-Patent Document 1 discloses a technique that uses a genetic algorithm for feature quantity selection and parameter search used for a discrimination rule in automatic generation of a discrimination rule (inspection logic). In other words, by using genetic algorithms to select feature values and parameter search used for discrimination rules, automate / semi-determining the creation of discrimination rules, which until now could only be done by trial and error based on human intuition and experience. Added automation.

  Further, as a technique for automatically creating a discrimination rule in abnormal noise inspection, there is an invention disclosed in Non-Patent Document 2. In this non-patent document 2, in automatic generation of discrimination rules, an appropriate number of normal data and abnormal data are selected from normal data and abnormal data collected for generation of discrimination rules, and a gene algorithm is used from the selected data. Thus, there is disclosed a technique for performing feature selection and parameter adjustment used for a discrimination rule, and generating a discrimination rule from a selected feature amount and the adjusted parameter using fuzzy inference. Such a technique for generating a discrimination rule that most separates normal data and abnormal data from each other is generally called defect identification.

However, in defect identification, normal data and abnormal data are required as sample data in advance to determine normality / abnormality. Since abnormal data is difficult to obtain compared to normal data, a determination rule is generated if there is no abnormal data. There is a problem that you can not.
Japanese Patent No. 3484665 OMRON Technics Vol. 43 No. 1 pp. 99-105 (2003) OMRON Technics Vol. 44 no. 1 pp. 48-53 (2004)

  As described above, development of various abnormal sound inspection apparatuses has been attempted so far. However, in all cases, the defective product (abnormal product) is erroneously judged as a good product (normal product), while avoiding an overlook rate (it is necessary to reliably prevent defective products from being shipped). Is a high-performance pass / fail judgment algorithm designed to reduce the over-detection rate that erroneously determines a product as a defective product (to prevent waste / yield reduction where good products are not shipped and discarded) Created and improved. Therefore, the current situation is that the number of feature quantities to be used increases or the number of samples required for creating a better determination rule increases.

  For example, by adjusting the parameters so that the optimum feature value can be output for each inspection target data, it is possible to determine the subtlety of normal or abnormal. Here, for example, when the feature amount A is a certain value (threshold value) or more, if the parameter value is a defective product, the threshold value related to the feature amount A is a parameter. Adjustment of whether or not is parameter adjustment (learning). However, if the parameter adjustment is insufficient or unadjusted, there is a possibility that correct inspection cannot be performed.

  Therefore, as described above, as the amount of features and the number of samples increase, the decision rules that are generated and confirmed (judgment of rule suitability) are varied along with learning. It takes time to create rules.

  Non-Patent Document 2 discloses that an appropriate number of normal data and abnormal data are selected. Specifically, “select a data set used for parameter adjustment. Forty types of feature values are selected. Use the OK data to calculate the reference space, calculate the Mahalanobis distance for each data, and balance the total number of OK data and NG data from the specified data against the entire distribution. Is selected. ”Is described. That is, if the balance between the number of OK data (normal data) and NG data (abnormal data) is poor, there is a possibility that appropriate and efficient learning cannot be performed.

  By the way, it is difficult to prepare abnormal data based on defective products that occur during actual judgment processing in advance, and most of the sample data that can be prepared at the time of learning is normal data based on non-defective products, and the number of samples of normal data and abnormal data A large variation occurs. Therefore, in order to perform learning based on an appropriate number of normal / abnormal data, normal data is deleted (data thinning out). However, if it is deleted in the dark clouds, even if the calculation time for parameter adjustment can be shortened, overdetection increases, and the performance as an inspection device may not be achieved.

  As an example, it is assumed that a scatter diagram with two feature amounts PV and PN in normal data and abnormal data as axes is as shown in FIG. Here, the scatter diagram is a diagram in which a plurality of indexes that quantitatively represent the characteristics of data are determined, and each data is arranged according to the index values in a space configured with these indexes as axes. Here, it is an example of a scatter diagram when two types of feature quantities are used as axes.

In this scatter diagram, white circles are normal data, and black circles are abnormal data. When a determination rule is created in consideration of all data, the threshold values of the feature quantities are separated based on the boundary line as indicated by a wavy line in the figure. An example of the rule is as follows.
IF PV> A AND PN> B Then Abnormal data

A part of normal data is deleted from the data shown in FIG. 1A, and learning is performed based on the remaining normal data (representative data) and abnormal data to create a determination rule. At this time, for example, as shown in FIG. 1B, when the selection is biased (the dotted white circle is deleted and learning is performed based on the normal data of the solid white circle and the abnormal data of the black circle), there are originally two There is a possibility that erroneous inspection knowledge is created to try to separate what cannot be separated without using the feature quantity with only one feature quantity PV. An example of the rules at this time is as follows.
IF PV> A The abnormal data

Similarly, when a biased selection is made as shown in FIG. 1C (the dotted white circle is deleted and learning is performed based on the normal data of the solid white circle and the abnormal data of the black circle), normal data and abnormal data are displayed. There is a possibility that the boundary line with (threshold at the time of inspection) cannot be set well. An example of the rules at this time is as follows.
IF PV> C AND PN> D Then Abnormal data

  If an inappropriate rule is generated as described above, normal data that is a dotted white circle is erroneously determined as abnormal data, and overdetection increases. In addition, since it is an inspection device, it must be avoided that the abnormal data is misjudged (missed) as a non-defective product and shipped. Therefore, if appropriate data is not selected, there is a problem that over-detection increases as described above, and a lot of data is wasted.

  An object of the present invention is to provide a knowledge creating apparatus and a knowledge creating method capable of reducing the time required for parameter adjustment without deteriorating the optimum parameter adjustment result by appropriately reducing the data used for knowledge creation. And

In order to achieve the above-described object, the knowledge creating apparatus of the present invention extracts a feature amount from input measurement object measurement data, and determines whether or not the inspection target is good based on the extracted feature amount. A knowledge creation device that creates the pass / fail judgment knowledge used when the pass / fail judgment is performed on the basis of waveform data belonging to the same type and waveform data that does not belong to the same type, and a plurality of acquired knowledge belonging to the same type Using the data selection means for selecting predetermined waveform data from the waveform data, selection data that is the waveform data selected by the data selection means, and waveform data that do not belong to the same type, Knowledge creating means for creating, wherein the data selecting means is a predetermined number of boundary data belonging to a boundary area of a group to which the plurality of waveform data to be selected belongs. A function of selecting, and a selection function of selecting a part of the waveform data as a representative data from the waveform data that does not belong to the boundary region, the selection together with their selected the boundary data and the representative data The selection function performs a temporary feature amount calculation on all waveform data belonging to the same type, and executes the temporary feature amount calculation to obtain a feature amount calculation result. Based on the plurality of feature values obtained for each waveform data, the Mahalanobis distance is obtained, the boundary data is selected from the group having the larger Mahalanobis distance, and the representative data is selected from the group having the smaller Mahalanobis distance. Configured to select .

  In the embodiment, the knowledge creation means is realized by the “parameter optimization unit” and the “rule creation unit”. However, if the knowledge creation means creates the pass / fail judgment knowledge based on the selected selection data, Of course, it is not restricted to what was shown in the form.

  Further, as a selection method in the data selection means, a method in which the Mahalanobis distance is equal to or more than a set threshold value is selected as the boundary data, or a predetermined amount (specifically, from a large Mahalanobis distance) is selected. A number / ratio) is selected as the boundary data, each waveform data is sorted in the order of Mahalanobis distance, and the waveform data is selected every predetermined number, so that the representative data is selected, The representative data can be selected by sorting each waveform data in the order of Mahalanobis distance and selecting the waveform data at predetermined distance intervals. Of course, selection by other methods is not prevented.

  Furthermore, on the premise of the various inventions described above, determination means for determining whether or not the result of the pass / fail judgment executed using the pass / fail judgment knowledge created by the knowledge creating means has reached the target value of the discrimination performance ( In the embodiment, the data selection unit is also used, but it may be configured separately.) If the determination result of the determination unit does not reach the target value, the data selection unit selects the data to be selected by the data selection unit. It is desirable to provide a function of increasing and selecting data again. The present invention is realized by the second embodiment.

  The process of increasing the data selected by the data selection means may include at least one of the following (1) to (3). Of course, taking other methods is not precluded. (1) When the waveform data belonging to the same type is erroneously determined as a different group, the erroneously determined waveform data is added as boundary data. (2) If it is erroneously determined that the waveform data that does not belong to the same type belongs to the same type, the total number of data selection is increased. (3) When the waveform data not belonging to the same type is erroneously determined as belonging to the same type, and when the waveform data belonging to the same type is erroneously determined as a different group, the amount to be selected as boundary data is increased.

On the other hand, the knowledge creation method according to the present invention extracts the feature quantity from the input measurement target measurement data, and performs the pass / fail determination in the inspection apparatus that determines the pass / fail of the inspection target based on the extracted feature quantity. multiple belonging quality determination knowledge to be used in performing the waveform data belonging to the same type, the same kind a knowledge generating method obtained in the knowledge creation apparatus for creating, based on the waveform data that does not belong to the same type Data selection processing for selecting predetermined waveform data from among the waveform data, selection data that is waveform data selected by executing the data selection processing, and waveform data that do not belong to the same type, And knowledge creation processing for creating pass / fail judgment knowledge. The data selection process includes a selection process for selecting a predetermined number of boundary data belonging to a boundary area of a group to which the plurality of waveform data to be selected belong, and a part of the waveform data not belonging to the boundary area. Including selecting the waveform data as representative data, and combining the selected boundary data and the representative data into the selection data, wherein the selection processing is performed on all waveform data belonging to the same type. Based on the feature value calculation result obtained by performing the temporary feature value calculation and standardizing multiple feature value values obtained for each waveform data, the Mahalanobis distance is calculated. The boundary data is selected from a group having a large Mahalanobis distance, and the representative data is selected from a group having a small Mahalanobis distance. It was to be selected.

  In the present invention, the time required for parameter adjustment can be reduced without deteriorating the optimality of the parameter adjustment result by successfully reducing the data used for knowledge creation.

  Prior to describing an apparatus according to an embodiment of the present invention, an inspection system that performs pass / fail determination (sensory inspection) using a determination rule set by the apparatus will be described. As shown in FIG. 2, a signal from the microphone 2 and the acceleration pickup 3 that are in contact with and close to the inspection object 1 is amplified by an amplifier 4, converted into digital data by an AD converter 5, and then supplied to the inspection apparatus 10. It is like that. And the inspection apparatus 10 acquires the waveform data based on the sound data collected by the microphone 2 and the vibration data collected by the acceleration pickup 3, extracts the feature amount, and performs abnormality determination. As is apparent from FIG. 2, the inspection apparatus 10 is configured by a computer, and includes a CPU main body 10a, an input device 10b such as a keyboard and a mouse, and a display 10c. Further, if necessary, an external storage device can be provided, or a communication function can be provided to communicate with an external database to obtain necessary information.

  As illustrated in FIG. 3, the inspection apparatus 10 includes a feature amount extraction unit 11 that extracts a feature amount from waveform data acquired via the A / D converter 5, and a feature amount extracted by the feature amount extraction unit 11. A determination unit 12 that determines whether the data is normal data or abnormal data based on the value; a feature amount calculation parameter storage unit 13 that stores a feature amount extracted by the feature amount extraction unit 11 and its parameters; A fuzzy rule storage unit 14 that stores a fuzzy rule used when the determination unit 12 performs the pass / fail determination process. The determination unit 12 performs fuzzy inference on the given feature amount by the fuzzy inference unit 12 a according to the rules stored in the fuzzy rule storage unit 14. Then, the obtained degree of conformity is given to the threshold processing unit 12b, where threshold processing is performed, and pass / fail determination is performed. The determination result of the determination unit 12 can be displayed on the display 10c in real time or stored in a storage device, for example.

  As the specific configuration of the inspection apparatus 10 described above (detailed configuration of each processing unit), various conventionally known devices can be applied, and thus detailed description thereof is omitted. In the embodiment of the knowledge creation device according to the present invention, the feature quantity stored in the feature quantity calculation parameter storage unit 13 described above, the parameter of the feature quantity, or the rule stored in the fuzzy rule storage unit 14 is created. To do. FIG. 4 shows a schematic configuration of an embodiment of the knowledge creating apparatus according to the present invention.

  As shown in FIG. 4, the knowledge creation device 20 determines whether the inspection device 10 is good or bad based on the waveform database 21 storing sample data and the sample data (waveform data of normal data and abnormal data) stored in the waveform database. A pass / fail judgment algorithm generation unit 22 that generates the above-described feature amounts, rules, and the like used when making the determination is provided.

  The waveform data stored in the waveform database 21 is obtained by, for example, the sound and vibration generated by operating the sample product or the like with the sensor 3 (not shown in the figure, as necessary), as in the case of actual sensory testing. May be stored, or may be downloaded from another database prepared separately and stored. The waveform database 21 stores actual waveform data and its type (distinguishment between normal data and abnormal data). In other words, each waveform data may be stored in association with the type, or a holder for normal data and a holder for abnormal data may be separated, and waveform data corresponding to each holder may be stored. In short, it is only necessary to be able to understand the type of stored waveform data. Note that the above-mentioned type is determined by, for example, the inspector listening to the sound generated from the inspection object (sample product) at the same time when the inspector actually acquires the data by the sensor 3, or the waveform data once stored. And the result of the inspector listening to and judging the reproduced sound may be stored, and the inspection object for obtaining the sample data is preliminarily distinguished as good or defective. If it is, the type and waveform data may be automatically associated by specifying the type in advance and taking in the waveform data.

  The pass / fail judgment algorithm generation unit 22 includes a data selection unit 23, a parameter optimization unit 24, and a rule creation unit 25. Where the data selection unit 23 is the main part of the present invention, data used for generating a pass / fail judgment algorithm is selected from the sample data stored in the waveform database 21. In general, the number of sample data stored in the waveform database 21, in other words, the number of sample data that can be prepared, is typically the number of samples of normal data. Basically, it is possible to generate a pass / fail judgment algorithm that can perform a more accurate judgment as the number of samples increases. However, if the number of samples is large, for example, the processing time required for executing the optimization process performed by the parameter optimization unit 24 in the subsequent stage and determining the feature amount and the parameter of the feature amount becomes long. Therefore, the data selection unit 23 thins out and selects normal data that can be predicted to increase especially among the sample data stored in the waveform database 21. Thereby, there is also sample data of normal data that is not used (not reflected) for generation of the pass / fail judgment algorithm, so that the optimization processing time performed in the parameter optimization unit 24 at the subsequent stage can be shortened. And, as will be described later, when selecting normal data sample data, it is extracted under certain conditions, so the accuracy is almost the same as the pass / fail judgment algorithm obtained using all sample data before selection You can get things.

  The parameter optimizing unit 24 determines a feature amount (effective feature amount) used when the quality determination is performed and a parameter for the feature amount. In other words, an example of the feature amount includes a frequency component, an average value, a variance, a maximum value, a minimum value, a peak number exceeding a threshold, a value of the Nth peak, and the like. For the frequency component, a low-pass filter, a band-pass filter, or other various types of filters may be used, or waveform conversion such as FFT processing may be performed. The feature parameter at that time is a cut-off frequency value for the low-pass filter, or a value that divides the passband in the case of the band-pass filter. In the case of FFT processing, since it has been converted to the frequency axis, it is possible to quantify the state in which an abnormal component is included by using a component in an arbitrary frequency band as a feature amount. That is, the frequency band to be extracted after the FFT processing is a parameter. The feature values such as the average value and the maximum value are usually obtained based on the waveform data after filtering. The parameter when the number of peaks exceeding the threshold is used as the feature amount is the threshold value. When the feature amount is Nth from the top of each peak value, the value of N (single or plural) is a parameter. Furthermore, in the case of the number of peaks or peak value, the detection target time is set as a parameter, and not only one threshold value but also two threshold values such as a lower limit value and an upper limit value are set. Outside can be set as a parameter. Of course, it goes without saying that the feature amount is not limited to the above-described ones, and the parameters for the feature amounts exemplified and enumerated are not limited thereto.

  This parameter optimizing unit 24 searches for various parameters for feature amount calculation (feature amount calculation parameter and weight of evaluation value of each feature amount) that can best separate non-defective products (normal data) and defective products (abnormal data). In the present embodiment, for example, as in the invention disclosed in Japanese Patent Application Laid-Open No. 2004-079211, an effective feature amount and its parameters can be determined using GA (genetic algorithm). That is, when a genetic algorithm is used as a parameter search algorithm, individual parameters are regarded as genes, and combinations of all parameters are regarded as individuals. Therefore, a new individual is created by crossover / mutation of the individual (generation change), and the low-evaluated individual is replaced with a new individual. In this way, a more optimal individual (parameter setting) is obtained by leaving a higher-rated individual with higher evaluation.

  Then, finally obtained parameters and the like are stored in the feature amount calculation parameter storage unit 13. In the example shown in the figure, the pass / fail judgment algorithm generation unit 22 (parameter optimization unit 24) directly accesses and stores the feature amount calculation parameter storage unit 13 of the inspection apparatus 10; The feature quantity, parameters, etc. generated by the parameter optimization unit 24 are stored in the storage device of the knowledge creation device 20 (not shown), and the data is stored by a predetermined method (online, communication, via a recording medium). That's fine.

  Note that the optimization processing in the parameter optimization unit 24, that is, the search method for parameters and the like is not limited to the above-described GA (genetic algorithm), and for example, NN (neural network), SVM (support vector machine), Various methods such as brute force can be taken.

  Further, the feature quantity to be used by the inspection apparatus 10 obtained by the parameter optimization unit 24 and its parameters are given to the rule creation unit 25. The rule creation unit 25 creates a fuzzy rule based on the given feature amount and its parameters. Since the feature quantity to be used and the parameters for determining pass / fail are known, an “IF THEN” rule can be easily created by a known method, and a membership function can also be created based on the rule. Then, the created rules and the like are stored in the fuzzy rule storage unit 14. In the example shown in the figure, the pass / fail judgment algorithm generation unit 22 (rule creation unit 25) directly accesses and stores the fuzzy rule storage unit 14 of the inspection apparatus 10, but does not necessarily store directly. A fuzzy rule or the like generated by the rule storage unit 14 may be stored in a storage device of the knowledge creation device 20 (not shown), and data may be stored by a predetermined method (online, communication, via a recording medium).

  Here, the function of the data selection part 23 which becomes the principal part of this invention is demonstrated. As described above, when parameter search using a genetic algorithm is used, each individual parameter is regarded as a gene, and a combination of all parameters is regarded as an individual. By leaving an individual, it is possible to obtain an individual (parameter setting) close to the optimum. Even if a genetic algorithm is used, the time required for the search becomes enormous as the number of data used for parameter search and the number of search repetitions increase. Of course, as the number of data increases and the number of search iterations increases, it becomes easier to obtain optimal parameter settings. However, the data selection unit 23 of the present embodiment appropriately selects data used for search. By reducing the time, the time required for parameter adjustment can be reduced without impairing the optimality of the parameter adjustment result. That is, a parameter adjustment result almost equal to the feature amount calculation parameter adjustment result when data selection is not performed is obtained.

  Here, “good” parameter adjustment result means that good judgment performance (both overdetection rate and oversight rate are closer to 0% is better) can be obtained by using the feature value calculated under the parameter setting conditions. It is. However, since the value of judgment performance cannot be obtained unless the final inspection rule is finalized, at the time of parameter adjustment and at the time of the result, the degree of feature and its operational parameters are evaluated using the degree of separation. .

  Here, when the degree of separation is low, as shown in FIG. 5A, for example, the data scatter diagram shows that the good data group and the defective product regardless of how the threshold values are set for the feature values that are the axes. A state in which the area of the data group cannot be divided. In other words, the calculation result of the feature amount as it is indicates that it is difficult to discriminate between a good product and a defective product. In this scatter diagram, white circles are normal data and black circles are abnormal data.

  On the other hand, when the degree of separation is high, as shown in FIG. 5 (b), each of the non-defective product data group and the defective product data group exists by appropriately setting a threshold value for each feature amount serving as an axis. A state in which the area to be divided can be divided. This characteristic amount calculation result indicates that there is a high possibility that good products and defective products can be discriminated.

  There is a possibility that the state shown in FIG. 5A can be changed to the state shown in FIG. 5B by changing the feature amount to be used or adjusting the parameter even if the feature amount is the same. Of course, some adjustments have a low degree of separation, but they are feature quantities, parameters, etc. that are not suitable for analysis. Here, the data selection unit 23 results in poor detection as a result of selecting data, and, as shown in FIG. 1, unlike the threshold (parameter) set using all original data, erroneous detection occurs. It is necessary to avoid becoming.

  Under the above conditions, in this embodiment, data selection is performed as follows. That is, when shown as a conceptual diagram, a scatter diagram with the feature amounts PV and PN as axes is such that all normal data based on non-defective products is indicated by white circles as shown in FIG. It is assumed that the abnormal data based on the data is indicated by a black circle. As is clear from the figure, the number of samples of normal data (white circles) is larger than that of abnormal data (black circles), the balance is lacking, and normal data is similar (positions on the scatter diagram ( Although there are many things that are close to the coordinate value), the number of samples is large, and it takes time to generate the final pass / fail judgment algorithm, the existence of useless normal data samples cannot be denied, and learning is repeated repeatedly. It can be said that it is done.

  In such a case, as shown in FIG. 6 (b), the data selection unit 23, as shown in FIG. 6 (b), represents normal data (white circles) located in the central portion and normal data (between normal data ( The hatched circle) is selected, and the selected normal data (representative data) and all abnormal data are sent to the parameter optimization unit 24 in the next stage to perform learning. In this way, by leaving the normal data in the boundary portion, as shown in FIG. 6B, the threshold value for distinguishing between normal data and abnormal data can be made the same as in FIG. 6A. Since normal data indicated by wavy circles is not used for learning, the number of samples is reduced, and it is possible to generate a quality judgment algorithm with the same quality in a short time.

  In selecting the normal data at the boundary portion, in the present embodiment, data selection is performed based on the Mahalanobis distance. That is, according to the example shown in FIG. 7, the Euclidean distance is longer at point A than at point B, but the distribution area of the data enclosed by the ellipse has a larger variance in the A direction, so that B has a larger Mahalanobis distance. . Therefore, normal data having a larger Mahalanobis distance (hatched portion in FIG. 7) is selected in order. This part becomes the normal data of the boundary part. In addition, the white center region (here, the region other than the boundary portion) is selected to be uniform. This portion is representative normal data.

Here, the Mahalanobis distance is a distance scale that takes into account data variations. Even if the Euclidean distance on the scatter diagram is the same distance, the Mahalanobis distance is small in the direction where the variance is large, and the Mahalanobis distance is large in the direction where the variance is small. In other words, the Mahalanobis distance is a measure representing the distance between the data population and each data.
For example, the average of the feature amounts PV and PN is μPV and μPN, and the population of the feature amounts PV and PN is XPV and XPN. And

で求められる。 If the variance covariance matrix of the operation values of the feature values PV and PN is Σ and the inverse matrix is Σ-1, the Mahalanobis distance d between the population and each data is

Is required.

  Actually, since the span is different for each feature amount, it is necessary to normalize (standardize). That is, first, in order to match the units of each variable, for example, the average = 0 and the variance = 1 are standardized. Thereafter, the Mahalanobis distance is obtained from the center of the data based on the above formula. Then, select a certain number from the far side of Mahalanobis distance. As a result, normal data at the boundary is extracted. Then, a predetermined number of normal data (representative data) is extracted from the remaining data according to a predetermined condition. When a specific example of such processing is shown, the flowchart shown in FIG. 8 is executed.

  First, data is classified and necessary data is acquired (S1). Here, a good product (normal data) and a defective product (abnormal data) are classified. Actually, when the sample data is stored in the waveform database 21, the type is associated and registered. Therefore, here, necessary data is acquired based on the type registered in the sample data. . In the present embodiment, a predetermined number is selected from predetermined data (normal data is thinned out by a predetermined number).

  Next, the setting of the number of data selections is accepted (S2). That is, the user operates the input device or the like to input the number of normal data at the boundary and the number of typical normal data. Therefore, the data selection unit 23 acquires the number of data to be selected received via the input device or the like. The number of data specified at this time may be an absolute value or a relative value such as what percentage of the whole. Also, the normal data designation method for the boundary portion and the typical normal data designation method may be the same method or different methods.

  The feature amount calculation is performed on all the data stored in the waveform database 21 (S3). The waveform data to be subjected to the feature amount calculation is at least sample data (target) to be selected (type). Then, it becomes all normal data. In addition, when the feature-value is already calculated | required, it can be utilized.

  Next, the feature amount calculation results of all the data in the class (good product in this embodiment) obtained in the processing step S3 are standardized (S4). More specifically, the standardization is performed such that each feature value has an average of 0 and a variance of 1. Next, the Mahalanobis distance is calculated from the standardized feature data of all the data in the class obtained by executing the processing step S4 (S5). The Mahalanobis distance calculated here is the “reference space” calculated from the standardized feature data of all the data in the class (here, the normal data class) and the individual data (standardized feature data of the class) And the distance.

  Thereby, as shown in FIG. 9, it is possible to obtain a one-dimensionally aggregated distance from tens of dimensional feature quantity data. The closer to the class (non-defective), the closer the calculated Mahalanobis distance is to 1, and the farther from the non-defective features, the greater the Mahalanobis distance. Therefore, in this case, all of the data is normal data based on non-defective products. Therefore, the longer the Mahalanobis distance, the more data that exists in the region portion.

  Next, the one far from the class center is selected based on the Mahalanobis distance (S6). Specifically, the non-defective product data is sorted in descending order of the Mahalanobis distance, and a predetermined ratio or number is selected from the top of the sorted good product data. Thereby, data (normal data) located in a so-called boundary region is selected. In the present embodiment, all of the top N (or X%) parts designated in process step S2 are selected. As described above, in the present embodiment, normal data belonging to the boundary region is continuously extracted from the upper level (the largest Mahalanobis distance), but the extraction rule is not limited to this. However, by extracting a certain number of normal data having a large Mahalanobis distance, for example, as shown in FIG. 6B, normal data existing in the boundary region of the normal data existing region can be reliably extracted.

  In particular, whether or not the quality of the data in the vicinity of the boundary region is appropriate will have a severe effect on performance, so that normal data in the boundary region can be selected and used for learning to distinguish abnormal data from normal data. Rules, thresholds, etc. can be created with high accuracy.

  Next, based on the Mahalanobis distance, normal data that matches a predetermined condition (for example, equidistant intervals, every predetermined number, etc.) is selected from the remaining normal data (S7). In other words, in order to generate a pass / fail judgment algorithm for accurately discriminating between abnormal data (defective product) and normal data (good product), as described above, data of the boundary region is also necessary, but data belonging to each region It is necessary to appropriately select data scattered in areas other than the boundary area. Therefore, normal data that matches a predetermined condition that does not cause bias is selected. By selecting normal data without any bias in this way, the selected representative data becomes data that can represent the distribution characteristics of the entire class with a smaller number of cases. The predetermined condition will be described later.

  By appropriately extracting representative data in this way, parameters can be obtained by determining the degree of separation by looking at the distance between the classes, that is, the distribution of normal data (good product) and abnormal data (defective product). It is possible to evaluate the quality of adjustment.

  Then, the result obtained by executing each process described above is output to the parameter optimization unit 24 in the next stage (S8). In the above-described embodiment, only normal data is a target for data selection, but only abnormal data may be a target for data selection. In that case, the data acquired in the processing step S1 is abnormal data. In order to provide versatility, normal data and abnormal data are extracted separately in processing step S1, and the number of selections for each of normal data and abnormal data is set in processing step S2. It will be.

  As a specific method for selecting the representative data, for example, as shown in FIG. 10 (a), a selection method based on the ranking of distances, and as shown in FIG. 10 (b), based on the distance itself. There is a selection method. Of course, other methods may be used.

  The selection method based on the rank order shown in FIG. 10A is performed as follows. First, it is assumed that each value set in processing step S2 is selected from all 20 non-defective product data, and the boundary data ratio in the selected data is 40%. Therefore, the normal data selected as the boundary region is four, and six data are selected as representative data from the remaining 16 data.

  After executing the processing steps S3 to S5, normal data is sorted based on the calculated Mahalanobis distance. Thereby, as shown in FIG. 10A, each data is arranged on one axis with the Mahalanobis distance as the horizontal axis. In this state, four items having the largest distance (circles indicated by hatching in the figure) are selected as normal data (boundary data) in the boundary region. This is a specific example of the execution of the processing step S6. Next, normal data is selected from the non-defective products excluding the boundary data described above at every equally spaced number (every two in this example).

More specifically, the flowchart shown in FIG. 11 is executed. That is, first, the boundary data selection number Nsb and the representative data selection number Nsr are acquired (S11). For example, in the case where each value set in the above-described processing step 2 is a specific numerical value, the acquisition process may be performed by acquiring such a value. It is necessary to obtain from the number N by calculation. For example, as described above, when the data selection number Ns and the boundary data ratio rb are set, the boundary data selection number Nsb and the representative data selection number Nsr are calculated based on the following equations, respectively.
Nsb = Ns × rb / 100
Nsr = Ns−Nsb

Therefore, if Ns = 10 and rb = 40% as described above,
Boundary data selection number Nsb = 10 × 40/100 = 4,
The number of representative data selections is Nsr = 10−4 = 6.

  Next, normal data is sorted in descending order of Mahalanobis distance (S12). As a result, for example, as shown in FIG. 12A, each data is ranked in descending order of Mahalanobis distance. Of course, the execution order of the processing step S12 and the processing step S11 described above may be reversed.

  Then, the first to Nsb positions are selected as boundary data (S13). Thereby, in the case of the specific example described above, four data from the first place to the fourth place are selected (see FIG. 12B).

Next, the representative data selection interval Isr is calculated (S14). In the present embodiment, the following formula Isr = f [(Ns−Nsb) / Nsr]
Calculate based on Here, the function f [X] can be, for example, a maximum integer not exceeding X, or an integer obtained by rounding off the first decimal place. In either case, the data selection is equally spaced (Y units), but assuming that the function f [X] is rounded up as a result of rounding, the maximum integer that does not exceed the former X. In this case, the smaller Mahalanobis distance is selected as representative data, and in the latter rounding (result rounding up), the larger Mahalanobis distance is selected as representative data.

For example, in the case of the specific example described above,
Representative data selection interval Isr = (20−4) /6=2.666……
Therefore, when the former method is adopted, Isr = 2, and when the latter method is adopted, Isr = 3, and the values are different.

Then, for i = 0,1, ..., Nsr-1,
The (Ns-Isr × i) rank data is selected as representative data (S15). In this way, by selecting the rank of i = 0, the normal of the highest rank that exists at the center (near the center) of the normal data area that has the shortest Mahalanobis distance and that best represents the characteristics of normal data. Data is selected. In the case of the specific example described above, if Isr = 2, six data of 20th, 18th, 16th, 14th, 12th, and 10th are selected as representative data, and if Isr = 3, 20th, Six data of 17th, 14th, 11th, 8th and 5th are selected as representative data (see FIG. 12C).

  In this way, the above-described processing steps S13 and S15 are executed, and the selected boundary data and representative data are combined and determined as selection data (S16) (see FIG. 12C).

In the specific example described above, the case where four pieces of boundary data are selected according to FIG. 10A has been described as an example. However, for example, when the number of boundary data is two, the boundary data selection number Nsb is representative. The data selection number Nsr is respectively
Nsb = 10 × 20/100 = 2 pieces Nsr = 10−2 = 8 pieces

  Therefore, by executing the processing step S12, it becomes as shown in FIG. 13A, and by executing the processing step 13, the top two normal data are selected as boundary data (see FIG. 13B). ).

The representative data selection interval Isr is
(20-2) / 8 = 2
Therefore, Isr = 2 is obtained by either rule described above (execution of S14).

  Therefore, when the processing step S15 is executed, the (20-2 × i) th rank is selected as representative data for i = 0, 1, 2,... (Nsr−1). According to the specific example described above, the ranks corresponding to 0, 1, 2,..., 7, that is, the 20th, 18th,..., 6th are used as representative data (see FIG. 13C).

  Next, a selection method based on the distance shown in the conceptual diagram in FIG. 10B will be described. As one method, as described above, the Mahalanobis distance is calculated, and normal data is sorted based on the Mahalanobis distance. Then, the data whose Mahalanobis distance exceeds the threshold is selected as boundary data, and the representative data is selected by dividing the remaining data at equal intervals.

  Specifically, the flowchart shown in FIG. 14 is executed. That is, first, the selection number Ns and the boundary distance threshold th are acquired (S11). This acquisition process may be set based on, for example, the information set in the above-described processing step 2. If there is missing information (for example, boundary distance threshold), the missing information And wait for input from the user. Further, the boundary distance threshold is not limited to that based on the input from the user as described above, but may be set as an initial value on the apparatus side and used.

  Next, normal data is sorted in descending order of Mahalanobis distance (S22). Of course, the execution order of the processing step S22 and the processing step S21 described above may be reversed. Further, when the processing step S22 is executed first, the maximum value of the Mahalanobis distance, the average / dispersion, etc. are known, and the boundary distance threshold value may be calculated based on the maximum value.

  Then, normal data whose Mahalanobis distance exceeds the boundary distance threshold th is selected as boundary data, and the number of selected normal data is set to Nsb (S23). Thereby, in FIG. 10B, three normal data indicated by hatching are selected as boundary data. For example, assuming that the normal data is shown in FIG. 12A and the boundary distance threshold th is 3.0, the three normal data from the first to the third are selected as the boundary data.

  Next, among the normal data excluding the data selected as the boundary data, the maximum value Dmax and the minimum value Dmin of the Mahalanobis distance are obtained (S24). As an example, the normal data is as shown in FIG. 12A. As described above, when the boundary distance threshold th is 3.0, the normal data remaining without being selected as the boundary data is 4 Since the data is less than or equal to the order, the maximum value Dmax = 2.9 and the minimum value Dmin = 0.5.

The representative data selection number Nsr and the representative data interval distance Dsr are
Nsr = Ns−Nsb
Dsr = (Dmax−Dmin) / Nsr
(S25).

  Next, (Dmin + i × Dsr) is obtained for i = 0, 1,..., Nsr−1. Thereby, the remaining normal and other positions for dividing the other into equal distances are set. Then, normal data closest to each division position is selected as representative data (S26). By using the above arithmetic expression, the division position when i = 0 is the position of Dmin, so that the Mahalanobis distance is the shortest and the center of the normal data area that can be said to best represent the characteristics of normal data. The normal data having the highest rank existing near the center is selected. Then, the above-described processing steps S23 and S26 are executed, and the selected boundary data and representative data are combined and determined as selection data (S27).

Of course, in the present invention, it is not always necessary to use the minimum value Dmin as described above when obtaining the representative data interval distance Dsr.
Dsr = Dmax / Nsr
In addition to calculation based on the above, it can be obtained by various calculations.

  The selection method based on the distance itself is not limited to the above-described method (boundary data is determined based on a threshold value). As an example, for example, regardless of whether it is an absolute value or a relative value, first, the amount to be selected as boundary data is determined, and the higher order normal data is selected as boundary data in descending order of the Mahalanobis distance. . Then, based on the maximum value of the remaining data (also using the minimum value if necessary) and the number to be selected, the distance to be divided at equal intervals is calculated, and based on this, the data close to the position to be divided is used as representative data. You may make it select.

  FIG. 15 shows a main part of the second embodiment of the present invention. In other words, the pass / fail judgment algorithm created using the apparatus of the first embodiment described above does not always provide performance that is sufficiently satisfactory for the user. Therefore, in this embodiment, a function for performing reselection when sufficient performance is not obtained is provided.

  First, a pass / fail determination is actually performed based on a pass / fail determination algorithm created by the present apparatus, an overlook rate and an overdetection rate are obtained, and it is determined whether or not the target of discrimination performance has been achieved (S31). Here, “missing” means that a product is erroneously determined to be a non-defective product because it must be judged as “defective product” and should not be shipped after being discarded. Since such a situation must be avoided, the miss rate must be 0% (0% is preferred). “Overdetection” is to determine that a product is “defective” if it can be shipped as it is. If the occurrence rate of this “overdetection” is high, there is a risk that the yield will be reduced meaninglessly, leading to an increase in the cost of the product, and the possibility of pressing profits. Therefore, it is preferable that this overdetection rate is as small as possible. However, since it is practically difficult to set both the “missing rate” and the “overdetection rate” to 0%, it is common to allow overdetection to occur to some extent. Therefore, in this embodiment, the miss rate is 0% and the overdetection rate is 5%. This numerical value is arbitrary.

  Whether or not the discrimination performance has been achieved as a target is determined based on a pass / fail determination algorithm based on, for example, a pass / fail determination based on sample data stored in the waveform database 21 (having known pass / fail results). It is determined whether or not the result is correct, and the overdetection rate and the oversight rate are calculated based on the result. This determination is made by the data selection unit in the course of this embodiment, but it is of course possible to provide a separate determination means.

  In addition, the actual pass / fail judgment result is obtained from the output of the inspection apparatus 10 shown in FIG. 2, but when an example of a specific processing procedure is shown, all sample data brainned in the waveform database 21 Is passed to the inspection apparatus 10. The inspection device 10 performs a pass / fail determination on the acquired sample data, and passes the determination result to the knowledge creating device. Since the sample data is provided with a code for identifying the sample data, the inspection apparatus 10 associates the code with the determination result and passes the code to the knowledge creating apparatus 20. As a result, the knowledge creating device 20 can know the type of the corresponding sample data (non-defective product / defective product) stored in the waveform database 21 based on the code.

  Note that data transmission / reception between the inspection apparatus 10 and the knowledge creation apparatus 20 may be performed by communication or may be performed offline using a predetermined recording medium. If they are incorporated in the same personal computer, data can be easily transmitted and received by accessing the storage devices of the personal computers.

  Then, when the performance has achieved the target (Yes in the branch determination of the processing step S31), the process is terminated, but when the performance does not achieve the target (No in the branch determination of the processing step S31). Determines whether or not the overdetection / missing rate has converged (S32). That is, when the branch determination at processing step 31 is No, the overdetection rate and the oversight rate obtained every time are stored in the memory (temporary storage means). And it is judged whether it is in the convergence direction from the past history. That is, it is determined that convergence has occurred even when the overdetection rate and the overlook rate are gradually decreasing (when it is small even if there is some variation, it is determined that it has converged).

  If it has converged, there is a high possibility that the learning performance target will be achieved by continuing the learning as it is, so that the number of generations is increased and the discrimination rule creation is re-executed (S32). In other words, as described in the first embodiment, the representative data selected by the data selection unit 23 is sent to the parameter optimization unit 24 at the next stage, where the parameters are optimized using the GA. Therefore, the GA is continuously continued (re-execution of learning), and it is determined whether the discrimination performance has reached the target at an appropriate timing (S31).

  On the other hand, when the over-detection rate and / or the oversight rate have not converged (No in the branching determination in processing step S32), the representative data selected by the data selection unit 23 may not be appropriate. 23 performs data selection again. Specifically, first, it is determined whether or not the miss rate is larger than the target (S34). When the miss rate has reached the target (below the target) (the branch determination in processing step S34 is No), it is necessary to suppress the over detection rate, so data selection for adding over-detected data is performed. Is re-executed (S35).

  Specifically, as shown in the flowchart of FIG. 16, first, the previous data selection result is read (S41), and the over-detected non-defective product data is extracted with reference to the inspection result (S42). In such extraction processing, for example, the determination result of the inspection apparatus based on the pass / fail determination algorithm is “defective product” (the waveform data corresponding to the determination results of all inspection apparatuses (information specifying the sample data of the waveform data is also acceptable)) It is possible to automatically extract by storing or holding at least information specifying the sample data of the waveform data determined to be defective, and the determination result registered in the waveform database 21 is “non-defective”. It can be done automatically by picking up things.

  The extracted good product data is added to the previous data selection result (obtained by execution of processing step S41) (S43), and the data selection result after the addition is output as new representative data (S44). That is, it is passed to the parameter optimization unit 24 at the next stage.

  When the processing steps S41 to S44 described above are executed and the data selection unit 23 selects new representative data, the process proceeds to the processing step S36 of FIG. 15, the parameter optimization unit 24 re-executes parameter adjustment, The rule creation unit 25 re-executes the discrimination rule creation (S37). Through these processes, a pass / fail judgment algorithm based on new representative data is generated and stored in each database. Then, returning to the processing step 31, the inspection apparatus is executed using the pass / fail judgment algorithm based on the new representative data, and it is judged whether or not the discrimination performance has reached the target.

  On the other hand, if the miss rate does not reach the target value, the branch determination in processing step S34 is Yes, so the process jumps to processing step S38, and the number of selections is increased and representative data is selected again (S38). In other words, if the miss rate does not reach the target, the defective product (data above) is within the distribution range of the non-defective product (normal data), which is also located near the center of the distribution, or the distribution of the entire failure is wide. In many cases, separation is not possible at the feature level. That is, it means that the data selection is not appropriate, and the number of normal data when learning is selected by the data selection unit 23 is increased in order to separate normal data and abnormal data based on more sample data. .

  At this time, a method of simply increasing is conceivable, but if the number of data is simply increased, there is a problem that the learning time increases. Therefore, for example, by increasing in accordance with the flowchart shown in FIG. 17, an appropriate increase in the number of data in accordance with the situation at that time is aimed at, and learning performance is improved in a short time while improving the performance of the learning result. To be able to do

Specifically, first, it is determined which of the miss rate and the overdetection rate is worse (S51). In other words, it can be assumed both when the miss rate has not reached the target value, when the overdetection rate has not reached the target value, and when the overdetection rate has reached the target value, and both have reached the target value. If not, it can be divided into cases such as which one of the overlook rate and overdetection rate is worse. Therefore, in order to judge them collectively, in this embodiment,
“Overlook rate> overdetection rate?”
Judging.

  If the miss rate is high (the branch determination in processing step S51 is Yes), the boundary data ratio is increased because it is considered that the non-defective boundary data is insufficient (S52). Alternatively, the selection number of boundary data is increased. On the other hand, when the over-detection rate is larger (the branch determination in processing step S51 is No), the number of data selections is increased in order to broadly recognize the non-defective product distribution (S53).

  Then, after each of the above-described processes is executed, selection of non-defective product data (selection of new representative data) is executed in accordance with the selection conditions as newly set representative data (S54). In the flowchart shown in FIG. 17, either one of the processing steps S52 and S53 is executed based on the result of the branch determination in the processing step 51. For example, when the processing step S52 is executed, Subsequently, both processes may be executed such that the processing step S53 is also executed.

  When the above-described processing steps S51 to S54 are executed and the data selection unit 23 selects new representative data, the process proceeds to processing step S39 in FIG. 15, the parameter optimization unit 24 re-executes parameter adjustment, The rule creation unit 25 re-executes the discrimination rule creation (S40). Through these processes, a pass / fail judgment algorithm based on new representative data is generated and stored in each database. Then, returning to the processing step 31, the inspection apparatus is executed using the pass / fail judgment algorithm based on the new representative data, and it is judged whether or not the discrimination performance has reached the target.

  By repeatedly executing each processing step described above through an appropriate route, learning is performed based on representative data of an appropriate number of data, and a pass / fail judgment algorithm having a target quality can be generated.

It is a figure explaining the conventional problem. It is a figure showing an example of an inspection system. It is a block diagram which mainly shows an example of the internal structure of an inspection apparatus. It is a block diagram which shows one Embodiment of the knowledge preparation apparatus which concerns on this invention. It is a figure explaining a separation degree. It is a figure explaining the outline | summary of the operation principle and effect | action of this Embodiment. It is a figure explaining the outline | summary of the operation principle and effect | action of this Embodiment. It is a flowchart which shows an example of the function of a data selection part. It is a figure explaining each feature-value and Mahalanobis distance in this Embodiment. It is a figure explaining the outline | summary of the operation principle and effect | action of this Embodiment. It is a flowchart which shows a detailed example of the function of a data selection part. It is a figure which shows an example of the operation principle and effect | action of this Embodiment. It is a figure which shows an example of the operation principle and effect | action of this Embodiment. It is a flowchart which shows a detailed example of the function of a data selection part. It is a figure which shows an example of the flowchart which shows 2nd Embodiment. It is a flowchart which shows an example of the specific process sequence of process step S33. It is a flowchart which shows an example of the specific process sequence of process step S35.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Inspection apparatus 20 Knowledge preparation apparatus 21 Waveform database 22 Pass / fail judgment algorithm generation part 23 Data selection part 24 Parameter optimization part 25 Rule preparation part

Claims (8)

The same pass / fail judgment knowledge is used when performing the pass / fail determination in the inspection apparatus that extracts the feature amount from the input measurement target measurement data and makes the pass / fail determination of the test target based on the extracted feature amount. A knowledge creation device that creates waveform data belonging to a type and waveform data that does not belong to the same type ,
Data selection means for selecting a predetermined waveform data from among a plurality of waveform data belonging to the same type obtained,
Using the selection data that is the waveform data selected by the data selection means, and the knowledge creation means for creating the pass / fail judgment knowledge using the waveform data that does not belong to the same type ,
The data selection means includes
Selection for selecting a function for selecting a predetermined number of boundary data belonging to the border area of the group in which a plurality of waveform data of the selected object belongs, a part of the waveform data from the waveform data that does not belong to the boundary region as the representative data With functions,
The selected boundary data and the representative data are combined as the selection data ,
The selection function performs a tentative feature amount calculation on all waveform data belonging to the same type, and each waveform data based on a feature amount calculation result obtained by executing the tentative feature amount calculation. Standardizing a plurality of obtained feature values, obtaining a Mahalanobis distance, selecting the boundary data from a group having a large Mahalanobis distance, and selecting the representative data from a group having a small Mahalanobis distance. Knowledge creation device characterized by   2. The knowledge creating apparatus according to claim 1, wherein the data selection means selects the boundary data that corresponds to the Mahalanobis distance equal to or greater than a set threshold value.   The knowledge creation device according to claim 1, wherein the data selection means selects a predetermined amount as the boundary data from the one having the large Mahalanobis distance.   4. The data selection unit according to claim 1, wherein the data selection means sorts each waveform data in order of Mahalanobis distance, and selects the representative data by selecting waveform data every predetermined number. The knowledge creation device according to claim 1.   4. The data selection unit according to claim 1, wherein the data selection means sorts each waveform data in order of Mahalanobis distance, and selects the representative data by selecting waveform data at a predetermined distance interval. The knowledge creation device according to claim 1. The knowledge generating means quality determination that is performed using the quality determination knowledge created results, and determination means to determine whether the reached the target value of the discrimination performance,
6. The method according to claim 1, further comprising a function of increasing data to be selected by the data selection means and selecting data again when the judgment result of the judgment means does not reach a target value. The knowledge creation device according to item 1. 7. The knowledge creating apparatus according to claim 6, wherein the process of increasing data to be selected by the data selection means includes at least one of the following (1) to (3).
(1) When the waveform data belonging to the same type is erroneously determined as a different group, the erroneously determined waveform data is added as boundary data.
(2) If it is erroneously determined that the waveform data that does not belong to the same type belongs to the same type, the total number of data selection is increased.
(3) When the waveform data not belonging to the same type is erroneously determined as belonging to the same type, and when the waveform data belonging to the same type is erroneously determined as a different group, the amount to be selected as boundary data is increased. Extracting feature values from the input measurement target measurement data, and using the same pass / fail judgment knowledge for use in the pass / fail judgment in the inspection device that performs pass / fail judgment of the test target based on the extracted feature quantities A knowledge creation method in a knowledge creation device that creates waveform data belonging to a type and waveform data that does not belong to the same type ,
And data selection processing for selecting a predetermined waveform data from among a plurality of waveform data belonging to the same type obtained,
Including selection data which is waveform data selected by executing the data selection processing, and processing for creating knowledge using the waveform data not belonging to the same type to create the pass / fail judgment knowledge,
The data selection process includes:
A selection process for selecting a predetermined number of boundary data belonging to the boundary region of the group to which the plurality of waveform data to be selected belong, and selecting a part of the waveform data from the waveform data not belonging to the boundary region as representative data And a process of combining the selected boundary data and the representative data together as the selection data ,
The selection process performs a temporary feature amount calculation on all the waveform data belonging to the same type, and based on the feature amount calculation result obtained by executing the temporary feature amount calculation, for each waveform data Standardizing a plurality of obtained feature values, obtaining a Mahalanobis distance, selecting the boundary data from a group having a large Mahalanobis distance, and selecting the representative data from a group having a small Mahalanobis distance. Knowledge creation method.

Images