JP2007139621A - Determination device, control program of determination device, and recording medium recording control program of determination device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an inspection device capable of providing to a user, information for setting a threshold capable of determining accurately whether an inspection object is normal or not. <P>SOLUTION: This inspection device is a device for acquiring a characteristic quantity from an engine, and inspecting whether the engine is a nondefective article or not according to the characteristic quantity. The device is equipped with a wrong discrimination rate measured value calculation part 25 for calculating each measured value of an over-detection rate which is a ratio of determining a nondefective engine as a defective one and an overlooking rate which is a ratio of determining a defective engine as a nondefective one from a frequency distribution of the characteristic quantity acquired from the nondefective engine and the defective engine, a wrong discrimination rate expected value calculation part 26 for calculating each expected value of the over-detection rate and the overlooking rate, and a display control part 29 for displaying a graph showing the change of each measured value of the calculated over-detection rate and overlooking rate corresponding to a change of the threshold, and a graph showing the change of each predicted value of the calculated over-detection rate and overlooking rate. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention obtains a characteristic value, which is information serving as an element for distinguishing whether or not the inspection object is normal from the inspection object, and whether or not the inspection object is normal according to the characteristic value. The present invention relates to a determination device, a control program for the determination device, and a recording medium on which the control program for the determination device is recorded.

  Conventionally, as a method for inspecting the quality of a product such as a produced engine, so-called sensory inspection has been performed, for example, in which an inspector hears a sound generated by driving the engine. However, since this sensory test is a method that depends on the five senses of the inspector, the inspector needs to have skill and a lot of empirical knowledge about the test method. Further, even inspectors having skillful techniques have a problem that the test results vary due to individual differences. For this reason, in order to prevent variation due to individual differences as described above, it is preferable that such an inspection can be performed based on a quantitative and clear standard.

  For example, Patent Document 1 discloses an apparatus for measuring sound or vibration generated by driving a manufactured engine with a sensor, extracting a feature amount from the measurement result, and analyzing the extracted feature amount to check the quality. Has been.

  By the way, when the feature quantity of the sound or vibration generated from a product such as an engine is extracted and the quality of the product is checked, for example, if this feature quantity is equal to or greater than a predetermined threshold, it is abnormal (defective product) When it is less than the threshold value, it can be determined that the product is normal (non-defective product). However, in the determination using this threshold value, there are cases where a non-defective product is erroneously determined as a defective product, or a defective product is erroneously determined as a non-defective product. In the present specification, the former case is referred to as “overdetection”, and the latter case is referred to as “missing”.

  For example, as shown in FIG. 25, when a histogram of feature values obtained from a non-defective product and a histogram of feature values obtained from a non-defective product are superimposed, There are sections where defective products are mixed. Here, when this mixed section exists in an area that is distinguished from non-defective products by a threshold value, the ratio of the number of defective products included in the section is defined as a missed rate. On the contrary, when this section exists in an area that is identified as a defective product by the threshold, the ratio of the number of non-defective products included in the section is defined as an overdetection rate. The overdetection rate and the oversight rate are collectively referred to as a misclassification rate.

  In addition, this overdetection rate can be calculated | required by the formula (1) shown below, On the other hand, an overlook rate can be calculated | required by the formula (2) shown below.

Overdetection rate = n _{OK → NG} / n _OK (1)
Missed rate = n _{NG → OK} / n _NG (2)
(N _{OK → NG} : number of good products erroneously determined as defective products n _OK : number of all good products)
(N _{NG → OK} : Number of defective products erroneously determined as good products n _NG : Number of all defective products)
In this way, products that are determined to be defective even though they are good products (over-detected products) are not shipped as inspection-defective products but discarded, or more accurate quality. If it is determined that the product is non-defective by inspection, the product is shipped. In this way, whether the good product is discarded or re-inspected, it becomes a factor that increases the manufacturing cost of the product.

  In addition, a product that is determined to be a non-defective product (a product that is overlooked) despite being a defective product is shipped as a product that has passed the inspection, and the defective product is put on the market. Thus, when defective products are circulated, complaints, returns, etc. of this product are generated, which is a cause of reducing the social credibility of the product manufacturer.

  For this reason, it is important to appropriately set the threshold value for distinguishing between the area determined to be non-defective and the area determined to be defective so that the ratio of occurrence of the overdetection rate and the oversight rate is within the allowable range. Become.

  For example, in the measurement result shown in FIG. 25, as shown in FIG. 26, the overlook rate increases as the value of the threshold t increases, and conversely, the overdetection rate increases as the value of the threshold t decreases. That is, as the value of the threshold value t increases, the interval for determining good products increases, and the number of defective products included in the interval, that is, the number of defective products regarded as non-defective products increases. For this reason, the miss rate increases as the threshold value t increases.

  Conversely, the smaller the threshold value t, the larger the section that is determined as a defective product, and the greater the number of non-defective products included in the section, that is, the number of non-defective products that are regarded as defective products. For this reason, the over-detection rate increases as the value of the threshold value t decreases.

  Therefore, conventionally, the above-described threshold value is determined within a range in which each of the miss rate and the overdetection rate falls within an allowable value.

  As a theory that can be used for the statistical processing of the sample as described above, Non-Patent Document 1 describes an actual misclassification rate and its estimation in asymmetric discrimination embodied in the Mahalanobis Taguchi system. Non-Patent Document 2 describes a study for evaluating a classifier that is truly effective under a real situation where the ratio of the number of training samples to the number of features is small. Non-Patent Document 3 describes a statistical method for determining the control limit of a multivariate control chart using a known sample.

In addition, the said sample is a product (product group) extracted in order to inspect from this all products, when all manufactured products are made into a population.
Japanese Patent Laying-Open No. 2005-121639 (released on May 12, 2005) Masami Miyagawa, Kentaro Tanaka, Tomoyuki Iwasawa, Hiroko Nakanishi, “Actual Misclassification Rate in Asymmetric Discrimination by Mahalanobis Distance”, Proc. 1-6 (2004) Yoshihiko Hamamoto, Shunji Uchimura, Yasunobu Kanaoka, Shingo Tomita, “Evaluation of Discriminators in a Small Number of Samples”, Information Processing Society of Japan Report “Graphics and CAD”, Vol. 1992. NO101, 1992 NOLA D.TRACY, JOHN C. YOUNG, ROBERT L. MASON, “Multivariate Control Charts for Individual Observations”, Journal of Duality Technology, Vol. 24, no. 2, April 88/95, (1992)

  However, in the above-described conventional configuration, a threshold value is set for distinguishing between non-defective products and defective products based on a predetermined number of limited feature quantity measurement results. The problem is not.

  That is, the number of samples that have already been measured and collected is finite, and the feature quantity distribution estimated from these samples is significantly different from the true feature quantity distribution. That is, the misclassification rate calculated from the feature amount distribution obtained from the already collected sample tends to be estimated smaller than the misclassification rate obtained from the feature amount distribution of the actual collection result, and is not accurate.

  Therefore, instead of setting a threshold value for discriminating between non-defective products and defective products based on the actually measured feature quantities of a limited number of samples, the prediction of the future misclassification rate is also taken into consideration. It is preferable that a threshold value can be set.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to provide information for setting a threshold value with which a user can accurately determine whether or not an inspection object is normal. An object of the present invention is to realize a determination device, a control program for the determination device, and a recording medium on which the control program for the determination device is recorded.

  In order to solve the above-described problem, the determination apparatus according to the present invention acquires a characteristic value that is information serving as an element for distinguishing whether or not the inspection target is normal from the inspection target. A determination device for determining whether or not an inspection object is normal according to a value, wherein the inspection object is normal from distribution information of characteristic values acquired from each of a plurality of normal inspection objects When the threshold for determining whether or not there is an arbitrary value, the above characteristic is calculated by a calculation means for calculating an overdetection rate that is a ratio for determining that a normal inspection object is abnormal, and a statistical method. A predicted value calculation means for calculating a predicted value of the overdetection rate based on the value, a graph showing a change in the overdetection rate calculated by the calculation means in accordance with a change in the threshold, and a change in response to the change in the threshold By the predicted value calculation means And an outputting means for outputting a graph showing changes in the predicted value of the issued over detection rate.

  According to the above configuration, since the output unit is provided, the graph showing the change in the actual value of the overdetection rate calculated by the calculation unit according to the change in the threshold value and the excess value calculated by the predicted value calculation unit. It is possible to output a graph showing a change in the predicted value of the detection rate.

  Here, the predicted value is a value obtained by predicting the overdetection rate of the population of inspection objects using characteristic values extracted from a limited number of inspection objects. Therefore, the user refers to the graph showing the change in the predicted value of the overdetection rate, and the inspection object is normal compared to the case where the threshold is set with reference to the graph showing the change in the overdetection rate. It is possible to set a threshold value with which it can be determined with high accuracy.

  In addition, since a graph indicating a change in the overdetection rate and a graph indicating a change in the predicted value of the overdetection rate in accordance with the change in the threshold can be output, the user can obtain the graph by comparing the two graphs. The reliability of the detected overdetection rate and the predicted value can be confirmed. That is, when the number of inspection objects used to calculate the overdetection rate is extremely small, and the reliability of the overdetection rate calculated from the characteristic value of each inspection object and the predicted value of the overdetection rate is low, Compared with the case where the reliability of the overdetection rate and the predicted value of the overdetection rate is high, the difference between the two becomes large.

  Therefore, by outputting a graph showing the change between the overdetection rate and the predicted value of the overdetection rate, the user can judge the reliability of the overdetection rate and the predicted value of the overdetection rate from the output result. Can do.

  As described above, the determination apparatus according to the present invention has an effect that it is possible to provide information for setting a threshold value with which it is possible to accurately determine whether or not the inspection object is normal to the user.

  In the above-described configuration, the determination apparatus according to the present invention includes an acquisition unit that acquires the characteristic value from the inspection object, and an acquisition instruction reception unit that receives an instruction to acquire the characteristic value from the new inspection object by the user. In addition, when the acquisition instruction reception unit receives an instruction to acquire a characteristic value from the new inspection object, the acquisition unit preferably acquires the characteristic value from the new inspection object.

  According to the above configuration, since the acquisition instruction receiving unit is provided, it is possible to receive an instruction to acquire a characteristic value from a new inspection object by the user. Moreover, since the acquisition means is provided, it is possible to acquire and add a characteristic value from a new inspection object in accordance with the acquisition instruction received by the acquisition instruction receiving unit.

  Here, when the user is instructed to acquire a characteristic value from a new inspection object, that is, to add a sample, the number of inspection objects from which the characteristic value has been acquired is small. This is a case where the reliability of the overdetection rate value and the predicted value is low.

  Here, the determination apparatus according to the present embodiment can acquire the characteristic value from the new inspection object when the reliability with respect to the overdetection rate and the predicted value of the overdetection rate is low. The number of inspection objects for which characteristic values are acquired can be added until a certain overdetection rate and a predicted value of the overdetection rate are obtained.

  In the above-described configuration, the determination apparatus according to the present invention has an acquisition unit that acquires the characteristic value from the inspection object, an overdetection rate calculated by the calculation unit, and an overdetection calculated by the prediction value calculation unit. A difference determining means for determining a difference between the calculated overdetection rate and the predicted value of the overdetection rate based on the predicted value of the rate, and determining whether the difference is a predetermined value or more, and When the difference determination unit determines that the difference is equal to or greater than a predetermined value, the difference determination unit is preferably configured to instruct the acquisition unit to acquire a characteristic value from a new inspection object.

  According to the above configuration, since the difference determination unit is provided, it can be determined whether or not the difference between the overdetection rate and the predicted value of the overdetection rate is equal to or greater than a predetermined value.

  The predetermined value is a value set within a range in which it is possible to determine that there are as many inspection objects as the number of reliable overdetection rates and the predicted value of the overdetection rate. .

  Here, the difference between the overdetection rate and the predicted value of the overdetection rate is equal to or greater than a predetermined value means that the overdetection rate and the predicted value of the overdetection rate do not have sufficient reliability. is there. That is, the number of inspection objects for calculating the overdetection rate and the predicted value of the overdetection rate is small.

  When determining that the difference is equal to or greater than a predetermined value, the determination device can instruct the acquisition unit to acquire a characteristic value from a new inspection object. For this reason, it is possible to obtain characteristic values from as many inspection objects as can obtain reliable overdetection rates and predicted values.

  Therefore, since a reliable overdetection rate and a predicted value can be obtained, information for setting a threshold value that can accurately determine whether or not the inspection object is normal is provided to the user. Can do.

  Further, the determination apparatus according to the present invention acquires a characteristic value that is information serving as an element for distinguishing whether or not the inspection target is normal from the inspection target in order to solve the above-described problem, A determination apparatus for determining whether or not an inspection object is normal according to the characteristic value, wherein the inspection object is obtained from distribution information of characteristic values acquired from a plurality of normal inspection objects. When a threshold value for determining whether or not it is normal is arbitrarily set, a calculation means for calculating an overdetection rate that is a ratio for determining that a normal inspection object is abnormal, and a statistical method, A predicted value calculating means for calculating a predicted value of the overdetection rate based on the characteristic value; and a graph showing a change in the predicted value of the overdetection rate calculated by the predicted value calculation means in accordance with a change in threshold value; Output means and inspection object Based on the acquisition means for acquiring the characteristic value, the overdetection rate calculated by the calculation means, and the predicted value of the overdetection rate calculated by the prediction value calculation means, A difference determination unit that obtains a difference from the predicted value of the overdetection rate and determines whether the difference is equal to or greater than a predetermined value, and the difference determination unit determines that the difference is equal to or greater than a predetermined value. In this case, the acquisition unit is instructed to acquire characteristic values from a new inspection object.

  According to the above configuration, since the difference determination unit is provided, it can be determined whether or not the difference between the overdetection rate and the predicted value of the overdetection rate is equal to or greater than a predetermined value. The predetermined value is a value set within a range in which it is possible to determine that there are as many inspection objects as the number of reliable overdetection rates and the predicted value of the overdetection rate. .

  Here, the difference between the overdetection rate and the predicted value of the overdetection rate is equal to or greater than a predetermined value means that the overdetection rate and the predicted value of the overdetection rate do not have sufficient reliability. is there. That is, the number of inspection objects for calculating the overdetection rate and the predicted value of the overdetection rate is small.

  When the determination device determines that the difference is equal to or greater than a predetermined value, the determination device can instruct acquisition means to acquire a characteristic value from a new inspection object. For this reason, it is possible to acquire characteristic values from the number of inspection objects for which the predicted value of the reliable overdetection rate can be calculated. As described above, since a reliable predicted value of the overdetection rate can be obtained, information for setting a threshold value that can accurately determine whether or not the inspection target is normal is provided to the user. There is an effect that can be.

  In the determination apparatus according to the present invention, in the configuration described above, the distribution information of the characteristic value further includes distribution information of the characteristic value acquired from each of the inspection objects in which a plurality of abnormalities occur, The means determines from the distribution information of the characteristic values that the inspection object in which an abnormality has occurred is normal when a threshold value for determining whether the inspection object is normal is arbitrarily set. The missed rate, which is a ratio, is further calculated, the predicted value calculating means further calculates the predicted value of the missed rate based on the characteristic value by a statistical method, and the output means is configured to detect a threshold value. A graph showing changes in the overdetection rate and the miss rate calculated by the calculation unit according to the change, and a prediction of the overdetection rate and the overlook rate calculated by the prediction value calculation unit according to the change in the threshold value It is preferably configured to output a graph showing a change.

  According to the above configuration, the output unit displays the graph indicating the change in the overdetection rate and the miss rate calculated by the calculation unit according to the change in the threshold value, and the overdetection rate and the overlook rate calculated by the prediction value calculation unit. It is possible to output a graph showing a change in the predicted value of the rate.

  Here, the predicted value is a value that predicts an overdetection rate and an overlook rate of a population by using characteristic values of a limited number of inspection objects. Therefore, the user can refer to the graph indicating the change in the predicted value of the overdetection rate and the oversight rate, and can set the threshold with reference to the graph indicating the change in the overdetection rate and the oversight rate. Thus, it is possible to set a threshold that can accurately determine whether or not the inspection object is normal.

  In addition, as described above, the user can output the graph indicating the change in the overdetection rate and the oversight rate and the graph indicating the change in the predicted value of the overdetection rate and the oversight rate in accordance with the change in the threshold value. Thus, it is possible to confirm the reliability of the overdetection rate and the oversight rate obtained by comparing these two graphs with the predicted values of the overdetection rate and the oversight rate.

  That is, the number of inspection objects used for calculating the overdetection rate and the oversight rate is extremely small, and the overdetection rate and the oversight rate calculated from the characteristic values of the inspection object, and the overdetection rate and the oversight rate When the reliability with the predicted value is low, the difference between the two is larger than when the reliability between the overdetection rate and the missed rate and the predicted value of the overdetection rate and the missed rate is high.

  Therefore, by outputting a graph showing the change of the overdetection rate / missing rate and the predicted value of the overdetection rate / missing rate together, the user can detect the overdetection rate / missing rate / overdetection rate and overlooked rate from the output result. The reliability with the predicted value of the rate can be determined.

  As described above, the determination device according to the present invention can provide information for setting a threshold value with which it is possible to accurately determine whether or not the inspection object is normal to the user.

  In the above-described configuration, the determination device according to the present invention has a change unit that changes the type of characteristic value for calculating the overdetection rate and the overlooked rate, and the overdetection rate and the predicted value of the overlooked rate, A change instruction receiving unit that receives information for instructing a change in the type of the characteristic value from the user, and when the change instruction receiving unit receives information instructing a change in the type of the characteristic value, the changing unit includes: It is preferable that the type of the characteristic value to be specified is changed.

  According to the above configuration, since the change unit is provided, when information indicating the change of the type of the characteristic value by the user is received, the overdetection rate, the overlook rate, the overdetection rate, and the overlook rate are predicted. The kind of characteristic value for calculating the value can be changed.

  For this reason, for example, when the distribution of the characteristic value acquired from the normal inspection object and the distribution of the characteristic value acquired from the inspection object having an abnormality cannot be determined by the threshold value, the characteristic value A type change can be indicated. In other words, when the type of the characteristic value is inappropriate as the characteristic value for determining whether or not the inspection object is normal, it can be changed to another type of characteristic value. For this reason, the determination apparatus can make appropriate the characteristic value used for determining whether or not the inspection object is normal.

  In the above-described configuration, the determination apparatus according to the present invention includes an acquisition unit that acquires the characteristic value from the inspection object, and an acquisition instruction reception unit that receives an instruction to acquire the characteristic value from the new inspection object by the user. In addition, when the acquisition instruction reception unit receives an instruction to acquire a characteristic value from a new inspection object, the acquisition unit preferably acquires the characteristic value from the new inspection object.

  According to the above configuration, since the acquisition instruction receiving unit is provided, it is possible to receive an instruction to acquire a characteristic value from a new inspection object from the user. Moreover, since the acquisition unit is provided, the characteristic value can be acquired from the new inspection object in accordance with the acquisition instruction received by the acquisition instruction receiving unit.

  Here, when the user is instructed to acquire a characteristic value from a new inspection object, since the number of inspection objects for acquiring the characteristic value is small, the overdetection rate and prediction calculated from the characteristic value This is when the reliability of the value is low. For this reason, the determination device according to the present embodiment can acquire and add a characteristic value from a new inspection object when the reliability of the overdetection rate and the predicted value of the overdetection rate is low. The number of inspection objects for which characteristic values are acquired can be increased until a reliable overdetection rate and a predicted value of the overdetection rate are obtained.

  In this way, since a reliable overdetection rate and a predicted value of the overdetection rate can be obtained, a threshold value that can accurately determine whether or not the inspection object is normal is set for the user. Information can be provided.

  In the above-described configuration, the determination device according to the present invention includes a changing unit that changes types of characteristic values for calculating the overdetection rate and the missed rate, and the predicted value of the overdetection rate and the overlooked rate. A storage device for holding allowable information indicating an allowable overdetection rate value and an allowable overlook rate value, the allowable information, and the overdetection rate and the overlooked rate prediction calculated by the predicted value calculation unit Based on the value, threshold value information that is a range in which a threshold value can be set, in which the predicted value of the over-detection rate is equal to or less than the allowable value and the predicted value of the oversight rate is equal to or less than the allowable value, is obtained. According to the range information, it is possible to set a threshold value for discriminating between the distribution of characteristic values acquired from each normal inspection object and the distribution of characteristic values acquired from each inspection object in which an abnormality has occurred. The threshold value setting determining means for determining whether or not the threshold value setting determining means acquires the characteristic value distribution acquired from each normal inspection object and the characteristic value acquired from each inspection object in which an abnormality has occurred by the threshold value. When the output unit determines that the distribution of the detection rate is indistinguishable, the output unit displays a graph indicating the change in the overdetection rate and the overlook rate calculated by the calculation unit according to the change in the threshold value, and the change in the threshold value. And a graph indicating a change in the predicted value of the overdetection rate and the missed rate calculated by the predicted value calculation unit, and the threshold setting determination unit obtains from each normal inspection object by the threshold value. If it is determined that the distribution of characteristic values obtained and the distribution of characteristic values obtained from each inspection object in which an abnormality has occurred cannot be discriminated, the type of characteristic values to be acquired is changed. It may be configured as indicated in the change means.

  According to the above configuration, since the threshold setting determination means is provided, the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object in which an abnormality has occurred due to the threshold value. It is possible to determine whether or not it is possible to determine that the ratio of occurrence of the overdetection rate and the oversight rate is within an allowable range.

  It should be noted that here, the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object in which an abnormality has occurred can be determined by the threshold value from the characteristic value. That is, it is possible to set a threshold value for determining whether an object is normal.

  Further, in the determination apparatus, the threshold setting determination unit determines, based on the threshold, a distribution of characteristic values acquired from each normal inspection object and a distribution of characteristic values acquired from each inspection object having an abnormality. If it is determined that it is impossible, the change means can be instructed to change the type of characteristic value to be acquired.

  For this reason, the determination apparatus can make the kind of characteristic value utilized in order to determine whether the test target object is normal or not to be appropriate.

  In addition, since the graph indicating the change in the overdetection rate calculated from the appropriate characteristic value and the graph indicating the change in the predicted value of the overdetection rate can be output, the determination apparatus according to the present invention provides the user with On the other hand, it is possible to provide information for setting a threshold value that can accurately determine whether or not the inspection object is normal.

  Further, the determination device according to the present invention is a change in which the type of characteristic value for calculating the overdetection rate and the missed rate and the predicted value of the overdetection rate and the missed rate is changed in the above-described configuration. The square of the Mahalanobis distance of each characteristic value acquired from each of the inspection objects having a plurality of abnormalities can be obtained with respect to the average of the means and the distribution of the set of characteristic values acquired from each of the plurality of normal inspection objects. From the distribution of characteristic values obtained from each normal inspection object and from each inspection object in which an abnormality has occurred, according to the value obtained by obtaining the desired value as the desired characteristic and the SN ratio of the desired characteristic. Threshold setting determination means for determining whether or not a threshold value for determining the distribution of the acquired characteristic value can be set, and the threshold setting determination means each of the normal inspection objects by the threshold value. When it is determined that the distribution of the characteristic values acquired from the distribution of the characteristic values acquired from each of the inspection objects in which an abnormality has occurred can be discriminated, the output means calculates the calculation means according to the change in the threshold value. A graph showing changes in the overdetection rate and the missed rate calculated by the above-mentioned values, and a graph showing changes in the overdetection rate and the predicted value of the overlooked rate calculated by the predicted value calculating means according to the change in the threshold value are output. However, the threshold value setting determination means cannot determine the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object having an abnormality by the threshold value. If it is determined, the change unit may be instructed to change the type of characteristic value to be acquired.

  According to the above configuration, since the threshold setting determination means is provided, the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object in which an abnormality has occurred due to the threshold value. Can be determined so that the ratio of occurrence of the overdetection rate and / or the oversight rate falls within an allowable range.

  It should be noted that here, the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object in which an abnormality has occurred can be determined by the threshold value from the characteristic value. That is, it is possible to set a threshold value for determining whether an object is normal.

  Further, in the determination apparatus, the threshold setting determination unit determines, based on the threshold, a distribution of characteristic values acquired from each normal inspection object and a distribution of characteristic values acquired from each inspection object having an abnormality. If it is determined that it is impossible, the change means can be instructed to change the type of characteristic value to be acquired.

  For this reason, the determination apparatus can make the kind of characteristic value utilized in order to determine whether the test target object is normal or not to be appropriate.

  In addition, since the graph indicating the change in the overdetection rate calculated from the appropriate characteristic value and the graph indicating the change in the predicted value of the overdetection rate can be output, the determination apparatus according to the present invention provides the user with On the other hand, it is possible to provide information for setting a threshold value that can accurately determine whether or not the inspection object is normal.

  In the above-described configuration, the determination apparatus according to the present invention includes an acquisition unit that acquires the characteristic value from the inspection target, an overdetection rate and an oversight rate calculated by the calculation unit, and the predicted value calculation unit. Based on the calculated overdetection rate and the predicted value of the overlooked rate, find the difference between the overdetection rate and the predicted value of the overdetection rate, and the difference between the overlooked rate and the predicted value of the overlooked rate, and combine these differences Difference determining means for determining whether or not the difference is greater than or equal to a predetermined value, and when the difference determination means determines that the difference is greater than or equal to a predetermined value, You may be comprised so that acquisition of a characteristic value may be instruct | indicated.

  According to the above configuration, since the difference determination unit is provided, whether or not the difference between the overdetection rate and the predicted value of the overdetection rate, and the difference between the overlook rate and the predicted value of the overlook rate is equal to or greater than a predetermined value. Can be determined. The predetermined value is within a range in which it is possible to determine that there are inspection objects as many as the number of reliable overdetection rates and missed rates and the predicted values of the overdetection rates and missed rates. It is a value set by.

  Here, the difference between the overdetection rate and the predicted value of the overdetection rate, and the difference between the overlooked rate and the predicted value of the overlooked rate are equal to or greater than a predetermined value means that the overdetection rate, the overlooked rate, and the overdetection rate And the predicted value of the miss rate is not sufficiently reliable. That is, the number of inspection objects from which characteristic values used to calculate the overdetection rate and the predicted value of the overdetection rate, and the missed rate and the predicted value of the overlooked rate are small.

  When the determination device determines that the difference is greater than or equal to a predetermined value, the determination device can instruct acquisition means to acquire further characteristic values. For this reason, it is possible to obtain the characteristic values from the inspection objects for the number of inspection objects that can calculate the reliable overdetection rate and the predicted value of the detection rate, and the miss rate and the predicted value of the miss rate.

  Therefore, since it is possible to obtain a reliable overdetection rate, a predicted value of the overdetection rate, and an overlooked rate and a predicted value of the overlooked rate, the accuracy of whether or not the inspection object is normal to the user Information for setting a threshold value that can be well determined can be provided.

  Moreover, the determination apparatus according to the present invention further includes a threshold value specification receiving unit that receives information specifying the threshold value from the user in the configuration described above, and the output unit sets the threshold value received by the threshold value specification reception unit. Accordingly, it is preferable that the overdetection rate and the miss rate calculated by the calculation unit and the predicted value of the overdetection rate and the miss rate calculated by the prediction value calculation unit are output.

  According to the above configuration, the overdetection rate and the overlook rate calculated by the calculation unit and the overdetection rate and the overlooked rate prediction value calculated by the prediction value calculation unit can be output. Appropriateness of the threshold can be confirmed.

  In the determination apparatus according to the present invention, in the configuration described above, the predicted value calculation unit may include each characteristic value with respect to an average of the distribution of the characteristic values with respect to a set of characteristic values acquired from each of a plurality of normal inspection objects. For the set of values obtained by obtaining the square of the Mahalanobis distance, the predicted value may be calculated by obtaining the upper probability of the F distribution.

  The predicted value calculation means uses a plurality of sets generated by combining characteristic values acquired from a plurality of inspection objects as training samples, and overdetects the overdetection rate and the oversight rate in each set of the training samples. It may be configured to calculate the predicted value of the rate and the miss rate.

  The adjusting device may be realized by a computer. In this case, the computer can be read by recording a control program of the adjusting device that causes the adjusting device to be realized by the computer by operating the computer as the respective means. Various recording media also fall within the scope of the present invention.

  As described above, the determination apparatus according to the present invention acquires a characteristic value that is information serving as an element for distinguishing whether or not the inspection object is normal from the inspection object, and according to the characteristic value A determination device for determining whether or not an inspection object is normal, and whether or not the inspection object is normal from distribution information of characteristic values acquired from each of a plurality of normal inspection objects Based on the above characteristic value by a calculating means for calculating an overdetection rate that is a ratio for determining that a normal inspection object is abnormal, and a statistical method A predicted value calculating means for calculating a predicted value of the overdetection rate; a graph showing a change in the overdetection rate calculated by the calculating means in accordance with a change in the threshold; and the prediction in accordance with a change in the threshold. Overdetection calculated by value calculation means And an outputting means for outputting a graph showing changes in the estimated value of.

  Therefore, the determination apparatus according to the present invention has an effect that it is possible to provide information for setting a threshold value that can accurately determine whether or not the inspection object is normal to the user.

  Further, as described above, the determination apparatus according to the present invention acquires a characteristic value that is information serving as an element for distinguishing whether or not the inspection target is normal from the inspection target, and uses the characteristic value as the characteristic value. A determination device for determining whether or not the inspection object is normal according to whether or not the inspection object is normal from distribution information of characteristic values obtained from each of the plurality of normal inspection objects When the threshold value for determining whether or not is arbitrarily set, the characteristic value is obtained by a calculation means for calculating an overdetection rate that is a ratio for determining that a normal inspection object is abnormal, and a statistical method. An output that outputs a predicted value calculation unit that calculates a predicted value of the overdetection rate based on the above, and a graph that shows a change in the predicted value of the overdetection rate calculated by the predicted value calculation unit according to a change in the threshold value And the above characteristic values from the inspection object Based on the acquisition means to be obtained, the overdetection rate calculated by the calculation means, and the predicted value of the overdetection rate calculated by the predicted value calculation means, the calculated overdetection rate and the predicted value of the overdetection rate Difference determining means for determining whether or not the difference is greater than or equal to a predetermined value, and when the difference determining means determines that the difference is greater than or equal to a predetermined value, On the other hand, the acquisition of the characteristic value from the new inspection object is instructed.

  Therefore, since a reliable predicted value of the overdetection rate can be obtained, it is possible to provide information for setting a threshold value with which it is possible to accurately determine whether or not the inspection object is normal for the user. There is an effect that can be done.

  An embodiment of the present invention will be described with reference to FIGS. 1 to 24 as follows. That is, the inspection system 100 according to the present invention checks whether or not the manufactured engine 1 is a non-defective product. As shown in FIG. 2, the microphone 2, the acceleration pickup 3, the amplifier 4, and the AD converter 5 are used. , And the inspection device 10. FIG. 2 shows an embodiment of the present invention, and is a block diagram showing a main configuration of the inspection system 100.

  That is, in the inspection system 100 according to the present embodiment, the inspection object (sample) is the manufactured engine 1. The non-defective product is the engine 1 that is normally driven, and the sound and vibration that are generated when the engine 1 is normally driven are different from the sound and vibration that are generated when the engine 1 is defective. . Therefore, it is possible to determine whether or not the engine 1 is a non-defective product based on the generated sound and / or vibration.

  The microphone 2 is arranged so as to be in contact with or close to the engine 1 and collects sound data (sound data) generated from the driven engine 1. The microphone 2 transmits the collected sound data to the amplifier 4.

  The acceleration pickup 3 is arranged so as to be in contact with or close to the engine 1 and collects vibration data (vibration data) generated from the driven engine 1. The acceleration pickup 3 transmits the collected vibration data to the amplifier 4.

  The amplifier 4 amplifies the sound data received from the microphone 2 and the vibration data received from the acceleration pickup 3. The amplifier 4 transmits the amplified sound data and vibration data to the AD converter 5.

  The AD converter 5 converts the sound data and vibration data amplified by the amplifier 4 into digital data, and transmits the converted digital data of the sound data and vibration data to the inspection apparatus 10.

  The inspection device 10 acquires waveform data from each of the received sound data and / or vibration data, extracts feature amounts (characteristic values) from these waveform data, and uses the extracted feature amounts to be an inspection target. It is determined whether or not the engine 1 is a non-defective product. The inspection apparatus 10 also sets a threshold value t for determining whether or not the engine 1 is a non-defective product based on the extracted feature amount.

  The feature amount is a value obtained from sound or vibration generated when the engine is driven, and is an element (measurement item) that can distinguish between a case where the engine is a non-defective product and a case where the engine is a defective product. Is the value of That is, the feature amount is a physical characteristic used for determining whether or not the engine 1 is a non-defective product, and can be regarded as a characteristic (quality characteristic) that well represents the quality of the product.

  More specifically, the feature amount is, for example, as shown in FIG. 3, a threshold value (for example, PT = 2) in a predetermined time range (for example, TL1.0 to TH4.0) in a waveform obtained from sound data or vibration data. -It may be the number of peaks of the waveform that is 0) or more. Or the peak value which becomes a magnitude | size of a predetermined order among the peaks of the waveform in a predetermined time range may be sufficient.

  Note that FIG. 3 is a diagram illustrating an example of the feature amount used in the inspection apparatus 10 according to the present embodiment. For example, as shown in FIG. 3, in the predetermined time range TL1.0 to TH4.0, the peak value of the rank in which the largest value is designated as the predetermined rank, for example, RP = 1, S (t) = It may be 4.0.

  The upper limit value (time) TH, the lower limit value (time) TL, the threshold value PT, and the predetermined rank RP that define the predetermined time range are input and set as parameters in the inspection apparatus 10.

  In addition, as described above, the feature amount is not limited to the number of waveform data peaks in a predetermined time range and the peak value of waveform data in a specific order within a predetermined time range. It is sufficient that the value of the element is such that it can be distinguished from a defective product. Further, at least one type of feature amount may be extracted from each of the sound data or the vibration data.

  As described above, the inspection system 100 according to the present embodiment can input the sound data detected by the microphone 2 and the vibration data detected by the acceleration pickup 3 to the inspection apparatus 10. Further, the inspection apparatus 10 extracts feature amounts by using the waveform data of the input sound data and vibration data, and determines whether or not the engine 1 to be inspected is a good product based on the feature amounts. be able to.

  Next, the configuration of the inspection apparatus 10 according to the present embodiment will be described with reference to FIG. FIG. 4 shows an embodiment of the present invention, and is a block diagram showing a main configuration of the inspection apparatus 10.

  First, as illustrated in FIG. 4, the inspection apparatus 10 according to the present embodiment includes an inspection apparatus control unit 11, a data storage unit 30, a display unit 40, and an input unit 50.

  The data storage unit 30 is a readable / writable recording medium, and is realized by, for example, a hard disk. As shown in FIG. 5, the data storage unit 30 has an ID number assigned for each extraction order of feature amounts extracted from sound data and / or vibration data by a threshold setting control unit 20 described later, Information (OK / NG) indicating whether or not the engine 1 having the feature quantity corresponding to the ID number is a non-defective product, and the value for each type (examination item) of the extracted feature quantity, that is, the feature quantity X and the feature quantity Feature quantity table information 31 indicating a correspondence relationship between the values of Y and feature quantity Z is recorded. FIG. 5 is a diagram showing an example of the feature amount table information 31 according to the present embodiment.

  In the inspection apparatus 10 according to the present embodiment, as shown in FIG. 5, the feature amount table information 31 in which the feature amounts of the non-defective product sample and the defective product sample are recorded together is stored in the data storage unit 30. there were. However, as shown in FIGS. 19 (a) and 19 (b), the inspection apparatus 10 uses the feature amount extracted from the sound data or vibration data of the non-defective engine 1, and the sound data or vibration data of the defective engine 1. The feature quantity extracted from each may be recorded in the data storage unit as a separate table.

  Note that the feature amount extracted from the sound data and / or vibration data is not limited to the three types of feature amount X, feature amount Y, and feature amount Z, but may be one type or three or more types. It may be.

  In addition to the feature amount table information 31, the data storage unit 30 further stores threshold setting data 32, allowable misclassification rate data 33, and internal management data 34. The threshold value setting data 32 is a table storing a plurality of threshold value candidates (t = 1.0, 2.0,... 100.0) as shown in FIG. This is information used to calculate the overdetection rate. FIG. 6 is a diagram showing an example of the threshold setting data 32 according to the present embodiment.

  The allowable misclassification rate data 33 is used to determine whether or not the threshold value setting control unit 20 can sufficiently separate the expected value (predicted value) of the overdetection rate and the expected value of the overlook rate. It is information used for. Here, the expected values of the overdetection rate and the overlook rate are the predicted values of the overdetection rate and the overlook rate of the population of the engine 1 using the feature values extracted from the engine 1 that is a limited number of inspection objects. It is the value.

  As shown in FIG. 7, the permissible misclassification rate data 33 records the permissible values (the permissible overdetection rate and the permissible miss rate) for the expected value of the overdetection rate and the expected value of the overlook rate. Note that the allowable overdetection rate is information indicating a rate that is allowed even if overdetection occurs, and the allowable oversight rate is information indicating a rate that is allowable even if overlooking occurs. FIG. 7 is a diagram showing an example of allowable misclassification rate data 33 according to the present embodiment.

  As shown in FIG. 12, the internal management data 34 is calculated by a threshold setting control unit 20 to be described later, an actual value of an overlook rate corresponding to each threshold candidate, an expected value of an overlook rate, and an actual value of an overdetection rate. It is a table which shows the correspondence of a value and the expected value of an overdetection rate. FIG. 12 is a diagram showing an example of the internal management data 34 according to the present embodiment. The actual measurement values of the detection rate and the miss rate are the over-detection rate and the miss rate obtained by using the feature amounts acquired from the plurality of engines 1 that have actually been determined to be good or bad.

  The display unit 40 displays various data or a graph based on the data in accordance with an instruction from the inspection apparatus control unit 11. The display unit 40 can be realized by, for example, an LCD (Liquid Crystal Display).

  The input unit 50 is an external input unit for a user to input an operation instruction or the like to the inspection apparatus 10 according to the present embodiment. This operation instruction includes, for example, an input instruction or determination instruction for the threshold value t. The input unit 50 can be realized by, for example, a keyboard, a mouse, a touch panel, or the like as an external input unit.

  The inspection device control unit 11 controls each unit included in the inspection device according to the present embodiment. As shown in FIG. 4, the inspection apparatus control unit 11 includes a data collection control unit 12, a pass / fail determination unit 13, and a threshold setting control unit 20 as functional blocks. When the inspection apparatus control unit 11 is realized by a CPU or the like, this functional block can be realized by the CPU reading a program from a ROM (not shown) and executing it with a RAM (not shown).

  The data collection control unit 12 extracts and collects feature amounts from the sound data detected by the microphone 2 and the vibration data detected by the acceleration pickup 3. In addition, the extracted and collected feature quantities are notified to the quality determination unit 13 to instruct to determine whether or not the engine 1 having the feature quantities is a non-defective product.

  That is, when the data collection control unit 12 receives a feature amount collection instruction from the input unit 50 or the threshold setting control unit 20, the data collection control unit 12 determines whether or not the engine 1 having the feature amount is a non-defective product. The information is stored in the data storage unit 30 as the feature amount table information 31 in association with the information (OK / NG) and the data ID.

  The information indicating the non-defective product (OK) or the defective product (NG) is associated with the feature amount extracted as described below. That is, the feature amount table information 31 is recorded with the feature amount extracted from the sound data and / or vibration data of the engine 1 that has been previously determined to be a non-defective product or a defective product. Then, when the data collection control unit 12 extracts the feature quantity from the sound data and / or vibration data of the engine 1, whether or not the engine 1 having the feature quantity currently extracted by the input unit 50 is a non-defective product. Is sent to the data collection control unit 12. Therefore, the data collection control unit 12 records information indicating whether the product is a non-defective product notified from the input unit 50 and the extracted feature quantity in the data storage unit 30 as the feature quantity table information 31.

  Note that the method of associating the information indicating the non-defective product (OK) or the defective product (NG) with the extracted feature value is not limited to this, and for example, the input unit 50 indicates the extraction order of the feature value. Information (OK / NG) indicating whether the engine 1 is a non-defective product along with the ID number is transmitted to the data collection control unit 12 after all the feature values are extracted. The data collection control unit 12 associates information indicating a non-defective product (OK) or a defective product (NG) with the feature amount acquired according to the ID number.

  On the other hand, when receiving an instruction to determine whether or not the engine 1 that extracts the feature value from the input unit 50 is a non-defective product, the data collection control unit 12 transmits the extracted feature value to the quality determination unit 13. It is instructed to determine whether or not the engine 1 having this feature amount is a non-defective product.

  The quality determination unit 13 uses a threshold value t received from a threshold setting control unit 20 to be described later with respect to the feature value received from the data collection control unit 12, and determines whether the engine 1 having the feature value is good or bad. This is a judgment. The pass / fail determination unit 13 outputs the determined result to the display unit 40 for display.

  The threshold value setting control unit 20 refers to the feature amount table information 31 stored in the data storage unit 30 and sets a threshold value t for discriminating non-defective products or defective products. The set threshold value t is transmitted to the pass / fail determination unit 13.

  Here, a detailed configuration of the threshold setting control unit 20 will be described with reference to FIG. FIG. 1 illustrates an embodiment of the present invention, and is a block diagram illustrating an example of a main configuration of a threshold setting control unit 20 of an inspection apparatus 10.

  As shown in FIG. 1, the threshold setting control unit 20 includes, as functional blocks, a target feature amount specifying unit 23, a feature amount data reading unit 24, a misclassification rate actual measurement calculation unit 25, a misclassification rate expected value calculation unit 26, This configuration includes a margin evaluation unit 27, a divergence evaluation unit 28, a display control unit 29, and an actual measurement / expected value output unit 21.

  In response to an instruction from the input unit 50, the target feature amount specifying unit 23 selects which of the feature amount X, the feature amount Y, and the feature amount Z included in the feature amount table information 31 stored in the data storage unit 30. The feature amount is used or the feature amount type (inspection item) is designated. The target feature amount specifying unit 23 notifies the feature amount data reading unit 24 of information indicating the type of the specified feature amount, and instructs to collect data for the specified type of feature amount.

  In other words, when the target feature value specifying unit 23 receives an instruction to start setting the threshold value t from the input unit 50 together with the information related to the specification of the feature value, the target feature value specifying unit 23 collects the specified type of feature value in accordance with this instruction. An instruction is given to the data reading unit 24.

  In addition, the target feature amount specifying unit 23 changes the specification of the type of feature amount to be used in accordance with an instruction from a margin evaluation unit described later.

  The feature amount data reading unit 24 refers to the feature amount table information 31 stored in the data storage unit 30 in response to an instruction from the target feature amount specifying unit 23, and a predetermined number of engines 1 (inspection objects). Information indicating a correspondence relationship between the feature amount extracted from each and the information indicating whether each engine 1 indicating the feature amount is a non-defective product is read out. The feature amount data reading unit 24 transmits the read information to the misclassification rate actual measurement value calculation unit 25 and the misclassification rate expected value calculation unit 26 and instructs to perform a calculation based on the information.

  The misclassification rate actual value calculation unit 25 calculates the misclassification rate based on the received information in response to an instruction from the feature data read unit 24. Based on the information received from the feature data reading unit 24, the misclassification rate actual measurement value calculation unit 25 is a non-defective engine in the engine 1 (sample) from which the feature values are extracted, for example, as shown in FIG. 1 (non-defective product sample) has a feature amount histogram and a non-defective engine 1 (defective product sample) has a feature amount histogram, and these histograms are referred to to determine a misclassification rate (overdetection rate and missed rate). ) Is calculated.

  More specifically, the misclassification rate actual measurement calculation unit 25 approximates the value of the histogram shown by dividing the range of the feature amount shown in FIG. 25 into a plurality of predetermined ranges, for example, FIG. The probability density function as shown in Then, in the obtained established density function, an approximate value of the misclassification rate (overdetection rate and missed rate) is obtained from the value obtained when the threshold value t is set, and a cumulative density function as shown in FIG. 9 is obtained.

  FIG. 8 is a graph showing an example of a probability density function indicating the distribution of each of the non-defective product sample and the defective product sample, where the vertical axis is the number of appearance frequencies of samples that are good products or defective products, and the horizontal axis is the feature quantity. It is said. FIG. 9 is a graph showing an example of a change between the approximate value of the overdetection rate and the approximate value of the overlook rate in accordance with the change in the threshold, and the vertical axis represents the ratio of the overdetection rate or the overlook rate. The misclassification rate, which is a ratio of

  Here, as the value of the threshold value t, a plurality of threshold value candidates (t = 1.0, 2.0,... 100.0) are stored in the data storage unit 30 in advance as the threshold setting data 32 as described above. ing. Therefore, the misclassification rate actual measurement value calculation unit 25 calculates approximate values of the missed rate and overdetection rate actual values for the respective threshold candidate values recorded as the threshold setting data 32 in the probability density function shown in FIG. Then, the cumulative density function shown in FIG. 9 can be obtained.

  That is, as shown in FIG. 9, the approximate value of the overdetection rate is the upper probability at the percentage point t of the probability density function of the feature quantity obtained from the plurality of non-defective engines 1, and the approximate value of the overlook rate is The lower probability at the percentage point t of the probability density function of the feature quantity obtained from the sound data of the plurality of engines 1 that are defective products. In this way, the misclassification rate actual measurement calculation unit 25 obtains, for example, a cumulative density function as shown in FIG. 9 as the misclassification rate.

  The misclassification rate obtained by the misclassification rate actual measurement calculation unit 25 is not limited to the value shown as the cumulative density function described above. For example, the misclassification rate calculated directly from the histogram shown in FIG. The value shown in such a relationship may be used.

  However, as in the inspection apparatus 10 according to the present embodiment, the configuration for obtaining the value of the cumulative density function as shown in FIG. 9 as the misclassification rate changes the threshold value t as shown in FIG. 26, for example. This is advantageous in that it is easier to determine a threshold value t that satisfies a desired overdetection rate and an oversight rate than using a relationship in which the erroneous discrimination rate changes stepwise.

  That is, when the value shown by the relationship shown in FIG. 26 is used as the misclassification rate, the misclassification rate when the threshold value t is changed changes stepwise. In this way, when changing in a stepped manner, there are a plurality of thresholds that take the same misclassification rate rate, and they are not uniquely determined, so it is difficult to determine a threshold value t that satisfies the desired overdetection rate and oversight rate. It becomes.

  On the other hand, when the threshold value t is changed in the configuration in which the value indicated by the cumulative density function as shown in FIG. 9 is used as the misclassification rate as in the inspection apparatus 10 according to the present embodiment, the misclassification is performed. Since the approximate value of the rate changes continuously, there is an advantage that it is easy to determine t that satisfies the desired overdetection rate and oversight rate. The misclassification rate calculated by the misclassification rate actual measurement calculation unit 25 is referred to as an actual measurement value of the misclassification rate.

  The erroneous determination rate actual value calculation unit 25 records the actual determination value of the erroneous determination rate corresponding to each of the plurality of threshold candidates of the threshold setting data 32 as the internal management data 34 in the data storage unit 30.

  The misclassification rate expected value calculation unit 26 calculates a predicted value of the misclassification rate based on the information received from the feature amount data reading unit 24. Note that the misclassification rate expected value calculation unit 26 calculates a predicted value of the misclassification rate by, for example, the Leave-one-out method or the bootstrap method, using the sample read from the feature amount table information 31 as a training sample. In addition, this training sample is a sample used for the learning for calculating the predicted value of a misclassification rate by the Leave-one-out method etc., for example.

  For example, in the inspection apparatus 10 according to the present embodiment, the misclassification rate expected value calculation unit 26 can obtain the expected value of the misclassification rate using the bootstrap method as follows.

  That is, first, B sets of bootstrap samples (m) restored and extracted from m samples read from the feature amount table information 31 are prepared. For each set, the misclassification rate is calculated using the threshold candidates (threshold values t = 1.0, 2.0,..., 100.0) in the threshold setting data 32. The average of the misclassification rates obtained by performing this calculation operation B times is obtained as the expected value of the misclassification rate at each threshold t.

  Alternatively, when the feature value to be used is a feature value whose feature value varies depending on how the sample is selected (for example, Mahalanobis distance, linear discriminant function, distance from the SVM discrimination length plane, etc.), the misclassification rate expected value calculation The unit 26 may obtain the expected value of the misclassification rate using the Leave-one-out method as follows.

  That is, first, L sets of sets obtained by dividing m samples read from the feature amount table information 31 into evaluation samples (1) and training samples (m-1) are prepared. Then, feature amounts of all samples are calculated for each set based on the training samples, and threshold candidates (threshold values t = 1.0, 2.0,... 100) in the threshold setting data 32 for the evaluation samples for each set. .0) is calculated. The ratio of the misclassification rate obtained by performing this calculation operation L times is assumed to be an expected value.

  The misclassification rate expected value calculation unit 26 records the expected value of the misclassification rate corresponding to each of the plurality of threshold candidates of the threshold setting data 32 as internal management data 34 in the data storage unit 30. That is, the actual value of the misclassification rate calculated by the misclassification rate actual value calculation unit 25 and the expected value of the misclassification rate calculated by the misclassification rate expected value calculation unit 26 correspond to each threshold value candidate. It is recorded so that it can be taken.

  When the expected misclassification rate value calculator 26 calculates the expected misclassification rate for each threshold candidate, the margin evaluation unit 28 evaluates the margin between the overdetection rate and the miss rate in the expected misclassification rate. To instruct.

  The margin evaluation unit 27 refers to the internal management data 34 stored in the data storage unit 30 and the allowable misclassification rate data 33 in response to an instruction from the misclassification rate actual measurement value calculation unit 25. It is evaluated whether or not the expected value of the detection rate and the expected value of the miss rate are sufficiently separable.

  In other words, as described above, in the inspection apparatus 10 according to the present embodiment, as the allowable misclassification rate data 33, an overdetection rate (allowable overdetection rate) and an overlook rate (allowable oversight rate) that are allowed in advance are set. Has been.

  By the way, when the expected value of the overdetection rate and the expected value of the overlook rate are sufficiently separable, for example, as shown in FIG. In this state, there is a sufficient interval between the feature value that is equal to or greater than the allowable miss rate. Therefore, the margin between the feature value when the expected value of the overdetection rate matches the allowable overdetection rate and the feature value when the expected value of the oversight rate and the allowable oversight rate match And

  As shown in FIG. 10A, when the margin is large, a threshold value t can be provided between a feature value that is equal to or greater than the allowable overdetection rate and a feature value that is equal to or greater than the allowable oversight rate. The expected value of the detection rate and the expected value of the miss rate can be sufficiently discriminated. Therefore, in the inspection apparatus 10 according to the present embodiment, the threshold value t can be set so as to satisfy the allowable overdetection rate and the allowable oversight rate, and it is possible to accurately determine the engine 1 that is a non-defective product or a defective product. It will be possible.

  On the other hand, as shown in FIG. 10B, there is no sufficient interval between the feature value that is equal to or greater than the allowable overdetection rate and the feature value that is equal to or greater than the allowable oversight rate, that is, a margin. In a state where is small, the expected value of the overdetection rate and the expected value of the missed rate cannot be sufficiently determined, and it becomes difficult to set the threshold value t so as to satisfy the allowable overdetection rate and the allowable missed rate. .

  FIG. 10 is a graph showing an example of the change in the expected value of the overdetection rate and the change in the expected value of the overlook rate in accordance with the change in the threshold value. FIG. This is a case where the range becomes less than the allowable overmissing rate and the margin is large, and FIG. 4B shows the range where the allowable overdetection rate is lower than the allowable overmissing rate. This shows a case where the range becomes smaller and the margin is small.

  The inspection apparatus 10 according to the present embodiment is configured to determine that the margin is large when the margin is equal to or greater than a predetermined value, and to determine that the margin is small when the margin is less than the predetermined value. ing. When the margin evaluation unit 27 evaluates that the margin is large and a sufficient margin is ensured, the deviation evaluation unit 28 informs the deviance evaluation unit 28 of the actual value of the misclassification rate and the expected value of the misclassification rate. It is instructed to evaluate the degree of divergence, which is the difference between the two. On the other hand, when the margin evaluation unit 27 evaluates that the margin is small and a sufficient margin is not secured, the margin evaluation unit 27 instructs the target feature amount designation unit 23 to change the type of feature amount to be used. .

  The deviance evaluation unit 28 is calculated by the misclassification rate actual value calculated by the misclassification rate actual value calculation unit 25 and the misclassification rate expected value calculation unit 26 in response to an instruction from the margin evaluation unit 27. The degree of divergence, which is the difference from the expected value of the misclassification rate, is calculated.

  Usually, when the number of known samples used for calculating the expected value of the misclassification rate is extremely small, the estimation accuracy of the expected value of the misclassification rate calculated by the misclassification rate expected value calculation unit 26 is lowered. For this reason, when the threshold value t is determined from the expected value of the misclassification rate with low accuracy and it is determined whether or not the manufactured engine 1 is a non-defective product, it is not possible to accurately determine whether or not the engine 1 is good. Note that the known sample is a sample in which an inspection has already been performed in the sample and the determination result of the feature amount and the quality of the sample is known. On the other hand, among samples, a product that is going to be subjected to a pass / fail judgment test and whose pass / fail judgment result is not yet known is referred to as an unknown sample. The unknown sample may include a sample that has not been manufactured yet.

  By the way, when the number of known samples is small, the actually measured value of the misclassification rate obtained from the known samples is estimated to be smaller than the true misclassification rate obtained when the number of known samples is assumed to be infinite. The tendency to be marked becomes remarkable. Similarly, since the expected value of the misclassification rate is closer to the true misclassification rate than the actual value of the misclassification rate, the actual value of the misclassification rate is estimated to be smaller than the expected value of the misclassification rate. For example, as shown in FIG. 11A, when there are few known samples, there is a large difference between the expected value of the overdetection rate indicated by the solid line and the actual value of the overdetection rate indicated by the broken line. Further, there is a large difference between the expected value of the miss rate indicated by the solid line and the actual value of the miss rate indicated by the broken line.

  On the other hand, when the number of known samples is sufficient, as shown in FIG. 11 (b), between the expected value of the overdetection rate indicated by the solid line and the actual measurement value of the overdetection rate indicated by the broken line, There is no significant difference compared to the case of a). Also, there is no significant difference between the expected value of the miss rate indicated by the solid line and the actually measured value of the miss rate indicated by the broken line as compared with the case of FIG.

  FIG. 11 is a graph showing an example of the change between the actual value and the expected value of the overdetection rate according to the change in the threshold value, and the change between the actual value and the expected value of the miss rate. Shows the case where the difference between the actual value and the expected value of the overdetection rate and the difference between the actual value and the expected value of the missed rate are large, and FIG. The case where the difference between the expected value and the difference between the actually measured value of the miss rate and the expected value is small is shown.

  That is, if there is a large difference between the true misclassification rate and the actual measurement value of the misclassification rate because the number of known samples is too small, the actual measurement value of the misclassification rate obtained from this small number of known samples and the known It can be said that the reliability of the expected value obtained from the sample is low. In addition, since the number of known samples is small and the reliability of the actual measurement value of the obtained misclassification rate is low, there is a difference between the actual measurement value of the misclassification rate and the expected value of the misclassification rate. growing.

  Therefore, when the threshold value t is determined by the threshold setting control unit 20 according to the present embodiment, not only the expected value of the misclassification rate but also the degree of divergence between the actual value of the misclassification rate and the expected value of the misclassification rate. In addition, the reliability of the expected value is shown.

  The divergence degree evaluation unit 28 is based on the actual value of the misclassification rate calculated by the misclassification rate actual value calculation unit 25 and the expected value of the misclassification rate calculated by the misclassification rate expected value calculation unit 26 as follows. The degree of divergence can be obtained by obtaining the above calculation.

That is, let d be the quantitative index indicating the degree of divergence, the actual value of the overdetection rate is O _{OK → NG} (x), the expected value of the overdetection rate is E _{OK → NG} (x), and the actual value of the missed rate Is represented as O _{NG → OK} (x), and the expected value of the overlook rate is E _{NG → OK} (x), the relationship represented by the following formula (3) is established.

That is, the divergence degree evaluation unit 28 uses the difference between the actual value O _{OK → NG} (x) of the overdetection rate and the expected value E _{OK → NG} (x) of the overdetection rate as a quantitative index d indicating the degree of divergence. And a maximum difference value between the actual value O _{NG → OK} (x) of the missed rate and the expected value E _{NG → OK} (x) of the missed rate.

  When the divergence degree evaluation unit 28 obtains the divergence degree d, the divergence degree d determines whether the divergence degree d is equal to or larger than a predetermined value. The unit 12 is instructed to add and collect feature amounts for recording as feature amount table information.

  On the other hand, when the divergence degree evaluation unit 28 determines that the divergence degree d is smaller than a predetermined value, the divergence degree evaluation unit 28 outputs a graph of expected values of misclassification rates calculated from the feature amount table information 31 to the display control unit 29, respectively. To instruct.

  The display control unit 29 acquires the internal management data 34 from the data storage unit 30 in response to an instruction from the divergence degree evaluation unit 28, and displays a graph indicating a change in the actual value of the misclassification rate according to the characteristic value, The display unit 40 is controlled to display a graph corresponding to the characteristic value of the expected value of the separate rate.

  Thus, since the display unit 40 can display the expected value of the misclassification rate calculated based on the feature amount table information 31, the user can grasp how much the threshold value t should be set. Therefore, when the user sets an appropriate threshold value t with reference to the display of the graph indicating the expected value of the misclassification rate on the display unit 40, the input unit 50 receives information indicating the set threshold value t. Then, the input unit 50 transmits information indicating the received threshold value t to the measured value / expected value output unit 21 and instructs to output the measured value and expected value for each misclassification rate according to the threshold value t. To do.

  When the measured value / expected value output unit 21 receives the threshold value t from the input unit 50, the measured value / expected value output unit 21 refers to the internal management data 34 and calculates the measured value and expected value of the misclassification rate according to the threshold value t. Then, the display control unit 29 is instructed to output the calculated result. In response to an instruction from the actual measurement / expected value output unit 21, the display control unit 29 controls the display unit 40 to display the calculation result. Note that the information displayed on the display unit 40 at this time includes, for example, as shown in FIG. 13, a determined threshold value t, an actual value and an expected value of an overlook rate corresponding to the threshold value t, and an actual value of an overdetection rate And information indicating each expected value. FIG. 13 is a diagram illustrating an example of data to be displayed after the threshold is set in the inspection apparatus 10 according to the present embodiment.

  Next, a process (threshold setting process) related to threshold setting for determining whether or not the engine 1 is non-defective in the inspection apparatus 10 according to the present embodiment having the above-described configuration will be described with reference to FIG. To do. FIG. 14 is a flowchart for explaining threshold setting processing in the inspection apparatus 10 according to the present embodiment.

  First, when the user is instructed by the input unit 50 to set the threshold value t, the target feature value designating unit 23 in the threshold value setting control unit 20 determines the type of feature value for which the threshold value t is set in response to this instruction. Set (step S11, hereinafter referred to as S11). Then, the target feature amount specifying unit 23 instructs the feature amount reading unit 24 to collect the set types of feature amounts.

  In response to an instruction from the target feature amount specifying unit 23, the feature amount reading unit 24 reads out the feature amount of the set type, that is, the feature amount to be set with the threshold value t, from the feature amount table information 31 (S12). ).

  When the feature amount reading unit 24 reads the feature amount, the misclassification rate actual measurement value calculation unit 25 uses the read feature amount to determine the misclassification rate corresponding to the threshold candidates stored in the threshold setting data 32. (S13), the misclassification rate expected value calculation unit 26 calculates the expected value of the misclassification rate according to the threshold candidates stored in the threshold setting data 32 (S14). Then, the actually measured value and the expected value of the calculated misclassification rate are stored in the data storage unit 30 as internal management data 34.

  When the actually measured value and the expected value of the misclassification rate are calculated in steps S13 and S14, the margin evaluation unit 27 calculates the calculated actual value and expected value of the misclassification rate from the internal management data 30. The margin is calculated by referring to the read and allowable misclassification rate data 33 (S15). Then, the margin evaluation unit 27 evaluates the calculated margin (S16), and when it is determined that there is a sufficient margin in this evaluation (“YES” in S16), the divergence is calculated so as to calculate the divergence. The degree evaluation unit 28 is instructed. On the other hand, when the margin evaluation unit 27 determines that there is not a sufficient margin in the calculated margin evaluation (“NO” in S16), the target feature amount is specified so as to specify another type of feature amount. The designation unit 23 is instructed (S21).

  When the divergence degree evaluation unit 28 is instructed by the margin degree evaluation unit 27 to calculate the divergence degree, the divergence degree evaluation unit 28 refers to the internal management data 34 and calculates the divergence degree from the actual measurement value and the expected value of the calculated overdetection rate. (S17). When calculating the divergence degree, the divergence degree evaluation unit 28 determines whether or not the divergence degree is smaller than a predetermined value from the calculation result (S18). If it is determined that the degree of divergence is small (“YES” in S18), the display control unit 29 is instructed to display the expected value and actual measurement value of the misclassification rate calculated on the display unit 40. In response to this instruction, the display control unit 29 causes the display unit 40 to display a graph indicating the expected value and actual measurement value of the misclassification rate.

  On the other hand, when it is determined that the divergence degree is equal to or greater than the predetermined value, the divergence degree evaluation unit 28 instructs the data collection control unit 12 to collect an additional sample (S22). In response to this instruction, the data collection control unit 12 extracts a feature amount from sound data and / or vibration data obtained from another engine 1. When all manufactured engines 1 are already used as samples, a new engine 1 is manufactured, and a known sample is added by performing measurement and inspection.

  When the graph showing the expected value and the actual measurement value of the misclassification rate is displayed on the display unit 40 as described above, the user determines a threshold value with reference to these displays, and inputs the determined threshold value t through the input unit 50. To do. And the input part 40 receives the threshold value t input by the user in this way (S19). When the input unit 19 accepts the threshold value t, the actual value / expected value output unit 21 calculates the actual value and the expected value of the misclassification rate corresponding to the threshold value t, and outputs the calculation result to the display control unit 29. Then, the display control unit 29 controls to display the received calculation result on the display unit 40 (S20).

  In addition, the order of the process of the said step S13 and step S14 is not limited to this order, You may perform simultaneously and step S13 may be performed after step S14.

  As described above, since the inspection apparatus 10 according to the present embodiment includes the margin evaluation unit 27, the margin can be evaluated. That the margin can be evaluated, that is, the distribution of the feature amount acquired from the non-defective engine 1 and the distribution of the feature amount acquired from the engine 1 in which an abnormality has occurred are determined by the threshold value t. This means that it can be determined whether or not it can be determined within the miss rate.

  Further, in the inspection apparatus 10, when the margin evaluation unit 27 evaluates that the margin is small, that is, the distribution of the feature amount acquired from the non-defective engine 1 and the defective engine 1 by the threshold value t. When it is determined that the acquired feature value distribution cannot be distinguished, the target feature value specifying unit can be instructed to change the type of the acquired feature value.

  For this reason, the inspection apparatus 10 can select appropriate information as a feature amount used for determining whether or not the engine 1 is a non-defective product.

  Further, since the inspection apparatus 10 includes the divergence degree evaluation unit 28, the difference between the actual value of the overdetection rate and the expected value of the overdetection rate, and the difference between the actual measurement of the missed rate and the expected value of the missed rate are predetermined. It is possible to determine whether or not the value is smaller than the value, in other words, whether or not the value is equal to or greater than a predetermined value.

  The predetermined value is a value set within a range in which it is possible to determine that there are as many characteristic values as the number of reliable overdetection rates and actual measurement values and expected values of the miss rate. .

  Here, the difference between the actual value of the overdetection rate and the expected value of the overdetection rate, and the difference between the actual value of the oversight rate and the expected value of the oversight rate are greater than or equal to a predetermined value. The actual measurement value and the expected value are not sufficiently reliable. That is, the number of samples for calculating the actual value of the overdetection rate and the expected value of the overdetection rate, and the actual value of the oversight rate and the expected value of the oversight rate are small.

  Moreover, in the said inspection apparatus 10, when the deviation evaluation part 29 determines with the said difference being more than predetermined value, it can instruct | indicate acquisition of the further sample to the data collection control part 12. FIG. For this reason, it is possible to collect as many samples as can be used to calculate the reliable actual value of the overdetection rate and the expected value of the overdetection rate, and the actual value of the oversight rate and the expected value of the oversight rate.

  In addition, since the actual value / expected value output unit 21 is provided, a graph showing a change in the actual value of the misclassification rate calculated from the feature amount extracted from an appropriate number of samples and a change in the expected value of the misclassification rate are displayed. The graph shown can be output.

  Therefore, the inspection apparatus 10 can obtain the reliable measured value of the overdetection rate and the expected value of the overdetection rate, the value of the missed rate, and the expected value of the missed rate. It is possible to provide information for setting a threshold t with which it is possible to accurately determine whether or not the product is a non-defective product.

  As described above, in the inspection apparatus 10 according to the present embodiment, the display control unit 29 causes the display unit 40 to display the expected value and the actual measurement value of the misclassification rate, and refers the display to the user to set the threshold value t. It was a configuration to be set. Therefore, a display style on the display unit 40 when the user sets a threshold will be described.

  First, in the display unit 40, as shown in FIG. 15A, another window is started up as a threshold setting screen. In this threshold value setting screen, an area α for displaying a graph indicating an expected value and an actual value of the misclassification rate, a threshold value t, an actual value of the overdetection rate, an expected value of the overdetection rate, and an actual value of the missed rate In addition, an area β that allows the expected value, margin, and divergence of the miss rate to be displayed and input numerically is provided. In addition, a button for indicating an instruction for information on the display screen, more specifically, a threshold value determining button and a button for instructing cancellation of input information are provided.

  When the user sets a threshold value, a graph indicating the expected value of the misclassification rate in the region α (solid line in FIG. 15A) and a graph indicating the actual measurement value (broken line in FIG. 15A) are displayed. Is done. In the region β, the threshold value can be input numerically. At this time, values indicating the margin and the divergence may be displayed in the region β.

  The display at the time of setting the threshold value t is not limited to the above-described FIG. 15A, and may be a more devised display as shown in FIG. 15B, for example. That is, in the area α, a bar indicating the threshold is displayed over the display of the graph of the expected value of the misclassification rate and the actual measurement value, and this bar can be dragged to move in the horizontal axis direction within the α area. It has become.

  In addition, the threshold value in the region β is expressed by a numerical value in conjunction with the moving position of the bar indicating the threshold value. That is, by moving the bar indicating the threshold to a desired position, the threshold t corresponding to this position can be set. For this reason, the user can easily set the threshold value t.

  Further, the degree of divergence is displayed in the region β, and when the degree of divergence exceeds a predetermined value, it is displayed in a display color different from the case where it does not satisfy the predetermined value, or a warning text is displayed. It may come to be.

  Alternatively, the margin is displayed in the region β, and when the margin is less than a predetermined value, a display color different from the case where the margin is equal to or greater than the predetermined value or a warning text is displayed. It may be.

  Furthermore, the display control unit 29 holds information on the allowable overdetection rate and the allowable miss rate in advance, and the overdetection rate and the overlook rate corresponding to the threshold value t set by the user are the above-described allowable overdetection rate and If the allowable miss rate is exceeded, the over-detection rate and the miss rate are displayed in different display colors from the case where the excess detection rate and the allowable miss rate are below the above, or a warning text is displayed. It may be.

  Furthermore, in the graph indicating the expected value of the misclassification rate displayed in the region α, the display control unit 29 holds information on the allowable overdetection rate and the allowable oversight rate in advance. The section that exceeds the allowable oversight rate may be set to be displayed in a display color different from that of the other sections.

  FIGS. 15A and 15B are diagrams showing examples of display on the display unit 40 when prompting the user to set and input a threshold value. (A) of the figure shows an example of displaying a graph showing a change in the actual value of the misclassification rate according to the conversion of the threshold value t and a graph showing a change in the expected value, and (b) of the figure shows the threshold value t. An example is shown in which a graph showing a change in an actual measurement value of a misclassification rate according to the conversion, a graph showing a change in an expected value, and a threshold bar indicating a desired threshold position are displayed.

  The above-described misclassification rate actual value calculation unit 25 is configured to obtain the actual value of the misclassification rate using the threshold setting data 32 stored in advance in the data storage unit 30, and calculates the misclassification rate expected value. The unit 26 is configured to obtain the expected value of the misclassification rate using the threshold setting data 32 stored in advance in the data storage unit 30. However, instead of using the threshold candidates provided in advance as described above, the maximum value and the minimum value of the sample feature value read by the feature value data reading unit 24 are set as the upper limit and the lower limit, respectively. A configuration in which the threshold setting data 32 is automatically obtained by dividing the section between the lower limit by 100 may be possible.

Moreover, in the inspection apparatus 10 according to the present embodiment, the margin evaluation unit 27 uses the feature value when the expected value of the overdetection rate matches the allowable overdetection rate, and the expected value of the oversight rate. In this configuration, the difference between the value of the feature amount when the allowable miss rate matches and the margin is determined as the margin, and the magnitude of the margin is determined. However, the determination of the magnitude of the margin is not limited to this, and it can also be determined using the correlation ratio η ² shown in the following formula (4). This correlation ratio η ² is obtained by dividing the sum of the mean square of feature values extracted from each non-defective sample and the mean square of feature values extracted from each defective sample by the variance of the feature values of all known samples. It can ask for.

  In addition, as described above, the margin is not calculated by the margin evaluation unit 27, but is calculated by the misclassification rate expected value calculation unit 26 as shown in FIGS. 16 (a) and 16 (b). The display unit 40 may be configured to display a graph indicating the expected value of the overdetection rate and a graph indicating the expected value of the oversight rate. FIG. 16 shows a display example of a graph indicating the change in the expected value of the misclassification rate in accordance with the change in the threshold value t. FIG. 5A is a graph showing an example of a change in the expected value of the misclassification rate when the margin is large, and FIG. 10B is a graph showing a change in the expected value of the misclassification rate when the margin is small. It is a graph which shows an example.

  In this way, by setting the graph of the expected value of the overdetection rate and the expected value of the miss rate on the display unit 40, the threshold value t is set so that the good engine 1 and the defective engine 1 can be accurately distinguished. It is possible to make the user intuitively determine whether or not it is possible.

  In the case of the configuration in which the graph indicating the overdetection rate and the expected value of the overlook rate calculated by the misclassification rate expected value calculation unit 26 is displayed on the display unit 40 without calculating the margin as described above, the misjudgment is performed. The separate rate expected value calculation unit 26 transmits the calculated expected value to the display control unit 29 and instructs the display unit 40 to display it. The display control unit 29 can be realized by causing the display unit 40 to display the expected value received in response to an instruction from the misclassification rate expected value calculation unit 26. In the case of this configuration, it is not necessary to calculate the margin, and therefore it is not necessary to provide the margin evaluation unit 27. Therefore, the apparatus configuration of the inspection apparatus 10 can be further simplified.

  Further, in the threshold value setting control unit 20 according to the present embodiment, the divergence degree evaluation unit 28 includes an actually measured misclassification rate value calculated by the misclassification rate actual value calculation unit 25 and an erroneous determination rate expected value calculation unit 26. Based on the expected value of the misclassification rate calculated by the above, the divergence degree d is obtained. However, without calculating the divergence degree d in this way, as shown in FIG. 11A or FIG. 11B, the actual value of the misclassification rate calculated by the misclassification rate actual value calculation unit 25 is shown. The configuration may be such that the graph and the graph indicating the expected value of the misclassification rate calculated by the misclassification rate expected value calculation unit 26 are displayed together on the display unit 40. When comprised in this way, the test | inspection apparatus 10 can be set as the following structures.

  That is, when the misclassification rate actual value calculation unit 25 calculates the actual value of the misclassification rate, the misclassification rate actual value is transmitted to the display control unit 29, and the misclassification rate expected value calculation unit 26 calculates the expected value of the misclassification rate. Is calculated, the expected value is transmitted to the display control unit 29. When the display control unit 29 receives the actual value of the misclassification rate and the expected value of the misclassification rate, the display control unit 29 shows a graph of a probability density function as shown in FIG. Then, a cumulative density function graph as shown in FIG. 9 showing the relationship between the expected value of the misclassification rate and the feature amount is displayed on the display unit 40.

  Thus, by displaying both the graph of the actual value of the misclassification rate and the graph of the expected value of the misclassification rate on the display unit 40, the user can intuitively determine whether the expected value of the misclassification rate is reliable. Can be judged.

  Further, the inspection apparatus 10 according to the present embodiment is configured to calculate the misclassification rate for the type of feature quantity instructed from the target feature quantity designation unit 23. Then, there may be a configuration in which a plurality of types of feature amounts instructed by the target feature amount specifying unit 23 exist.

  Further, in the case of a configuration in which a misclassification rate is obtained from a plurality of types of feature amounts, a threshold value t is provided for each type of feature amount as shown in FIG. 17, and it is determined whether or not the engine 1 having the feature amount is a good product. To do. And the engine 1 from which the feature-value which exists in the area | region determined to be non-defective by all the threshold values t is determined to be non-defective.

FIG. 17 shows the distribution of feature quantities extracted from the non-defective engine 1 and the distribution of feature quantities extracted from the defective engine 1 when there are two types of feature quantities to be read out. is a diagram showing an example of a threshold t ₁ · t ₂ that is set, the vertical axis is the value of the feature x _2, the horizontal axis is the value of the feature x _1.

  However, in such a determination method, the misclassification rate cannot be calculated unless all the threshold values t set for each type of feature amount are determined. For this reason, there is a problem that it is difficult to determine how the threshold value t should be set for each type of feature amount.

Therefore, in the case of the configuration in which the misclassification rate is obtained from a plurality of types of feature amounts as described above, a plurality of types of feature amounts extracted from an arbitrary engine 1 are respectively synthesized as shown in FIG. Provide. And it is preferable to set the threshold t with respect to this newly provided feature amount and calculate the misclassification rate. FIG. 18 shows a combination of a feature quantity distribution extracted from a non-defective engine 1 and a feature quantity distribution extracted from a defective engine 1 and two types of feature quantities when there are two types of feature quantities to be read. and the vector _{f (x 1, x 2)} , is a diagram showing an example of the set threshold value t with respect to the vector.

  In the inspection apparatus 10 according to the present embodiment, the acquisition of a new type of feature value is realized by reading a plurality of feature values specified by the target feature value specifying unit 23 and performing a synthesis operation. Alternatively, the calculation result may be registered in advance in the feature amount table information illustrated in FIG. 5 and may be specified and read by the target feature amount specifying unit 23.

  As a method of calculating the misclassification rate using the feature values synthesized as described above, for example, discrimination in “Fisher line discriminant function” and “Support Vector Machine (SVM)” in discriminant analysis The distance from the hyperplane can be used.

  Further, as shown in FIG. 5, the inspection apparatus 10 according to the present embodiment refers to the feature amount table information 31 in which the feature amounts of the non-defective product sample and the defective product sample are recorded, and the misclassification rate actual measurement calculation unit. 25 is a configuration in which an actual value of the misclassification rate is calculated, and an expected misclassification rate value calculator 26 calculates an expected value of the misclassification rate. That is, both the engine 1 determined to be a non-defective product and the engine 1 determined to be a defective product are used as known samples.

  However, the inspection apparatus 10 calculates the actual value and the expected value of the misclassification rate using only the feature amount extracted from the sound data or vibration data acquired from the non-defective engine 1 as shown in FIG. The threshold value t may be determined.

In the case of such a configuration, for example, the threshold value t can be set as shown in FIG. 20 (a) or FIG. 20 (b). For example, as shown in FIG. 20 (a), set two kinds of two from the distribution of non-defective is defined as the threshold t ₁ and the threshold t ₂ by the feature quantity. In FIG. 20A, the case where the feature quantity x ₁ is equal to or less than the threshold value t ₁ and the feature quantity x ₂ is equal to or less than the threshold value t ₁ is regarded as a non-defective product.

Alternatively, for example, as shown in FIG. 20B, a sample determined to be non-defective and a non-defective product are determined by the square D ² of the Mahalanobis distance from the average distribution of non-defective samples used in the Mahalanobis Taguchi system. A sample can be distinguished. That is, as shown in FIG. 20B, the Mahalanobis distance is formed so as to form a discrimination boundary including the non-defective sample from the distribution of the non-defective sample indicated by the two types of feature amount x ₁ and feature amount x _2. square by setting a threshold t in ^{D 2} of the two types of features _{x 1,} square ^D 2 _(x 1, _{x 2)} of the Mahalanobis distance between the feature quantity _{x 2} as good a case where equal to or less than the threshold value t determined To do.

  More specifically, the misclassification actual measurement value calculation unit 25 calculates the actual value of the misclassification rate, that is, the actual value of the overdetection rate, from the distribution of only the non-defective samples as follows.

The sample average when the number of types of feature quantities is p, the number of samples of the engine 1 determined to be non-defective m is m, and the original p-dimensional feature vector of each non-defective sample is X _i , The variance-covariance matrix can be expressed as the following formula (6).

  In this case, the square of the Mahalanobis distance of each sample is defined by the following formula (7).

By the way, when the feature quantity of the engine 1 determined to be a non-defective product is a known sample, D _i for each known sample follows a chi-square (χ ² ) distribution of degrees of freedom p. Therefore, for example, as shown in FIG. 21, when the Mahalanobis distance square D ² _i exceeds the threshold t (= χ ² (p; t)) is determined to be a defective product, The actual measurement value can be obtained by obtaining the upper probability at t of the chi-square distribution with the degree of freedom p. In order to set the measured value of the overdetection rate to α, the threshold value t = χ ² (p; α) may be set.

Further, the expected value of the misclassification rate with respect to the actual value of the misclassification rate (overdetection rate) obtained as described above can be obtained by the miscalculation rate expected value calculation unit 26 performing the following calculation. More specifically, for example, when a p-dimensional feature vector X _{i of} an unknown sample j (j = 1,..., M) is given, the distribution of the Mahalanobis distance D ² _j is not a chi-square distribution, As shown in (8), it follows a distribution obtained by multiplying the F distribution with degrees of freedom p and mp by a constant.

Here, when it is determined that a sample having the Mahalanobis distance D ² _j larger than the threshold value t is determined as a defective product, the expected value of the overdetection rate is the threshold value t in the F distribution with degrees of freedom p and mp. Can be obtained by multiplying the upper probability at p (m−1) (m + 1) / m (m−p).

  For example, the relationship between the expected value of the overdetection rate and the actual measurement value is such that the former includes the latter in the mapping in a two-dimensional space as shown in FIG. Further, in the square distribution of the Mahalanobis distance of the non-defective sample, the distribution of the known sample and the distribution of the unknown sample have a relationship as shown in FIG. That is, the distribution of known samples follows a chi-square distribution, while the distribution of unknown samples follows an F distribution. From the relationship between the two, as shown in FIG. 24, when the measured value of the overdetection rate at the threshold value t is compared with the expected value, it can be seen that the measured value is estimated to be smaller than the expected value.

  In addition, since the relationship shown in the following mathematical formula (9) is established between the chi-square distribution and the F distribution, the smaller the number m of training samples, the larger the difference between the actual value and the expected value of the overdetection rate, and conversely If the number of training samples m is sufficiently large, both are approximated.

  Therefore, the actual value and the expected value of the overdetection rate are approximated as the sample quantity of the non-defective engine 1 for which the feature value is acquired in advance is larger.

  Further, in the configuration in which the misclassification rate is obtained based only on the distribution of the engine 1 determined to be non-defective as described above, the threshold value t is set so that the expected value of the overdetection rate becomes, for example, α within the allowable value range. In this case, it can be set as shown in the following formula (10).

  In addition, this configuration has an advantage that it is possible to easily collect samples because it is not necessary to prepare defective samples. In particular, since the engine 1 is manufactured under the premise that the engine 1 that satisfies the standard as a non-defective product is manufactured, it is usually difficult to collect the feature values from each of the engines 1 determined to be defective. is there. For this reason, the structure which can set the threshold value t from the sample of only good products is effective.

  In this way, in the configuration in which the misclassification rate is obtained based only on the distribution of non-defective samples, the misclassification rate is only the overdetection rate, and the margin is not defined. However, if a defective sample can be obtained even a little, for example, by setting the SN ratio γ of the desired characteristic of the Mahalanobis distance from the average of the distribution of the feature amount of the non-defective sample as the margin, the margin is increased. Can be judged. If this method is used, the size of the margin can be determined even if the number of defective products is small and there is not enough samples to ensure the reliability of the predicted value of the miss rate.

Here, in the case of the configuration using the SN ratio γ of the Mahalanobis distance telescopic characteristic as the margin, the SN ratio γ of the telescopic characteristic can be obtained as follows. First, the margin evaluation unit 27 obtains a reference space from the distribution of expected values of feature values collected from each engine 1 determined to be non-defective. That is, the Mahalanobis distance from the average of non-defective samples is obtained as follows. That is, when the type (measurement item) of the feature quantity is p, the number of samples obtained from the non-defective engine 1 is m, and the original p-dimensional feature vector of the expected value of the feature quantity obtained from each good product is X _i. The sample average can be expressed by the following formula (11). The variance-covariance matrix can be expressed by the following mathematical formula (12).

  At this time, the square of the Mahalanobis distance of each sample has the relationship shown in Equation (13).

Here, if the number of defective samples is n and the square of the Mahalanobis distance of each defective sample is D ² ₁ ,... D ² _n , the SN ratio value γ of the desired size characteristic is expressed by the following equation (14): Become.

  At this time, as the SN ratio value γ of the desired characteristic increases, the expected value group of feature values collected from the defective engine 1 (population of defective product samples) is the feature value collected from the non-defective engine 1. It shows that it is far from the expected value group (population of good sample). Therefore, it is possible to determine whether or not the expected value of the overdetection rate and the expected value of the overlook rate are sufficiently separable from the obtained value of γ.

  Thus, in the case of a configuration in which the square of the Mahalanobis distance of the defective sample from the average of the non-defective sample distribution is used as the desired characteristic, and the SN ratio γ of this desired characteristic is obtained as a margin, the allowable excess as described above. There is an advantage in that the detection rate and the allowable miss rate are not set in advance.

  As described above, in the inspection apparatus 10 according to the present embodiment, when there is a sufficient margin and the divergence is less than a predetermined value, the actually measured value and the expected value of the misclassification rate calculated by the display unit 40. This is a configuration in which graphs indicating the above are displayed together to prompt the user to set a threshold value. However, as in the inspection apparatus 10 according to the present embodiment, in the case of a configuration in which a margin and a divergence degree are evaluated in advance from the relationship between the actual value and the expected value of the misclassification rate, the display unit 40 displays The information to be made may be only a graph indicating the expected value of the misclassification rate.

  In addition, as described above, the inspection apparatus 10 according to the present embodiment has a configuration in which the display unit 40 displays the graph of the actual value and the expected value of the misclassification rate when the user sets a threshold value. Further, the display unit 40 is configured to display the data of the actual measurement value and the expected value of the misclassification rate calculated based on the threshold set by the user. However, the information of the actual measurement value and the expected value of the misclassification rate described above is not limited to the configuration displayed on the display unit 40 as described above, and is connected to the data storage unit 30 or the inspection apparatus 10 so as to be communicable, for example. It may be configured to output to another storage device. Alternatively, when the inspection apparatus 10 is connected to be able to communicate with the printing apparatus, the configuration may be such that the actual measurement value and the expected value of the misclassification rate are transmitted to the printing apparatus, and are printed and output.

  Finally, each block (data collection control unit 12, pass / fail judgment unit 13, and threshold setting control unit 20) included in the inspection device control unit 11 of the inspection apparatus 10, in particular, each block (target) included in the threshold setting control unit 20 Feature amount designation unit 23, feature amount data reading unit 24, misclassification rate actual value calculation unit 25, misclassification rate expected value calculation unit 26, margin evaluation unit 27, divergence evaluation unit 28, display control unit 29, and actual measurement The value / expected value output unit 21) may be configured by hardware logic, or may be realized by software using a CPU as follows.

  That is, the inspection apparatus 10 includes a CPU (central processing unit) that executes instructions of a control program that realizes each function, a ROM (read only memory) that stores the program, a RAM (random access memory) that expands the program, A storage device (recording medium) such as a memory for storing the program and various data is provided. An object of the present invention is a recording medium on which a program code (execution format program, intermediate code program, source program) of a control program of the inspection apparatus 10 which is software that realizes the above-described functions is recorded so as to be readable by a computer. This can also be achieved by supplying the inspection apparatus 10 and reading and executing the program code recorded on the recording medium by the computer (or CPU or MPU).

  Examples of the recording medium include a tape system such as a magnetic tape and a cassette tape, a magnetic disk such as a floppy (registered trademark) disk / hard disk, and an optical disk such as a CD-ROM / MO / MD / DVD / CD-R. Card system such as IC card, IC card (including memory card) / optical card, or semiconductor memory system such as mask ROM / EPROM / EEPROM / flash ROM.

  Further, the inspection apparatus 10 may be configured to be connectable to a communication network, and the program code may be supplied via the communication network. The communication network is not particularly limited. For example, the Internet, intranet, extranet, LAN, ISDN, VAN, CATV communication network, virtual private network, telephone line network, mobile communication network, satellite communication. A net or the like is available. Also, the transmission medium constituting the communication network is not particularly limited. For example, even in the case of wired such as IEEE 1394, USB, power line carrier, cable TV line, telephone line, ADSL line, etc., infrared rays such as IrDA and remote control, Bluetooth ( (Registered trademark), 802.11 wireless, HDR, mobile phone network, satellite line, terrestrial digital network, and the like can also be used. The present invention can also be realized in the form of a computer data signal embedded in a carrier wave in which the program code is embodied by electronic transmission.

  The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the claims. That is, embodiments obtained by combining technical means appropriately changed within the scope of the claims are also included in the technical scope of the present invention.

  For example, the feature amount used in the inspection apparatus 10 according to the present embodiment is extracted from sound data and vibration data generated from the engine 1, but is not limited to these data, and is converted into a feature amount and a threshold value. Any data can be used as long as the data can be provided. For example, as an example of this data, the rotational speed, torque, temperature, or the like of the engine 1 is increased.

  In the inspection system 100 according to the present embodiment, the inspection target is the engine 1, but the inspection target is not limited to this, and it is necessary to determine whether the product is good. Anything that can be extracted as a physical quantity such as data may be used. For example, a system for inspecting the quality of joining by soldering as described below may be used. That is, the area where the solder adheres is extracted as image data, and the area of the area where the solder adheres is obtained from the image data. And it is a system which judges the quality of joining by soldering by making this calculated | required area into a feature-value.

  The inspection apparatus 10 according to the present embodiment shows changes in the actual value and the predicted value of the misclassification rate calculated based on the distribution information of the feature amount extracted from the inspection object that is a non-defective product and the inspection object that is a defective product. Since the graph can be displayed, it is possible to provide information for setting a threshold value that enables quality inspection with high accuracy. Therefore, it can be widely applied to quality inspection of inspection objects.

1, showing an embodiment of the present invention, is a block diagram illustrating an example of a main configuration of a threshold setting control unit of an inspection apparatus. FIG. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1, showing an embodiment of the present invention, is a block diagram showing a main configuration of an inspection system. It is a figure which shows an example of the feature-value utilized in the inspection apparatus which concerns on this Embodiment. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1, showing an embodiment of the present invention, is a block diagram showing a main configuration of an inspection apparatus. It is a figure which shows an example of the feature-value table information which concerns on this Embodiment. It is a figure which shows an example of the threshold value setting data concerning this Embodiment. It is a figure which shows an example of the permissible misclassification rate data which concerns on this Embodiment. It is a graph which shows an example of the probability density function which shows distribution of the sample used as a good article and inferior goods. It is a graph which shows an example of the change of the approximate value of an overdetection rate, and the approximate value of a miss rate according to the change of a threshold value. It is a graph showing an example of the change in the expected value of the overdetection rate and the change in the expected value of the overlook rate according to the change in the threshold, and FIG. The figure shows a case where the range below the allowable oversight rate is large and the margin is large, and FIG. 8B shows a range where the range is below the allowable overdetection rate and the range below the allowable oversight rate is small and a margin. The case where the degree is small is shown. It is a graph which shows an example of the change of the actual value of an over-detection rate according to the change of a threshold value, and an expected value, and an example of the change of the actual value of an oversight rate, and an expected value. The difference between the actual measurement value and the expected value, and the difference between the actual measurement value and the expected value of the miss rate increase, and FIG. 5B shows the difference between the actual measurement value of the overdetection rate and the expected value. The case where the difference between the actually measured value and the expected value of the miss rate is small is shown. It is a figure which shows an example of the internal management data concerning this Embodiment. It is a figure which shows an example of the data displayed after the threshold value setting in the inspection apparatus which concerns on this Embodiment. It is a flowchart explaining the threshold value setting process in the test | inspection apparatus 10 which concerns on this Embodiment. It is a figure which shows the example of a display in a display part in the case of prompting a user to set and input a threshold value, and FIG. An example of displaying a graph showing a change in value is shown, and FIG. 8B shows a graph showing a change in an actual measurement value of a misclassification rate according to conversion of a threshold value t, a graph showing a change in an expected value, An example of displaying a threshold bar indicating a desired threshold position is shown. The example of a graph which shows the change of the expected value of a misclassification rate according to the change of a threshold is shown. FIG. 5A is a graph showing an example of a change in the expected value of the misclassification rate when the margin is large, and FIG. 10B is a graph showing a change in the expected value of the misclassification rate when the margin is small. It is a graph which shows an example. An example of a distribution of feature quantities extracted from a non-defective engine, a distribution of feature quantities extracted from a defective engine, and a threshold set for each feature quantity when there are two types of feature quantities to be read is shown. FIG. When there are two types of feature quantities to be read, the distribution of feature quantities extracted from a non-defective engine, the distribution of feature quantities extracted from a defective engine, a vector obtained by combining two types of feature quantities, It is a figure which shows an example with the threshold value set with respect to. It is a figure which shows an example of the table which recorded the feature-value acquired from the engine, The figure (a) shows an example of the table which records the feature-value acquired from the quality engine, and the figure (b) An example of the table which records the feature-value acquired from the quality engine is shown. It is a figure which shows an example of distribution of the feature-value extracted from the engine of the quality goods in case there are two kinds of feature-values to read, and the threshold value set with respect to this distribution. An example in the case where a threshold value is provided for each quantity is shown, and FIG. 5B shows an example in which a threshold value is set for the square of the Haranobis distance. It is a graph which shows distribution in the Mahalanobis distance of the feature-value extracted from the quality engine. It is a figure which shows the relationship between the distribution of the feature-value extracted from the quality engine, and the measured value and expected value of the overdetection rate calculated from the feature-value. It is a graph which shows distribution in the Mahalanobis distance of the feature-value extracted from the quality engine. It is a graph which shows an example of the change of the actual value of an overdetection rate according to the change of a threshold, and an expected value. It is a histogram which shows distribution of the feature-value extracted from the engine of the quality goods and the inferior goods. It is a graph which shows an example of the change of an overdetection rate when a threshold value is changed with respect to the histogram shown in FIG.

Explanation of symbols

1 Engine (inspection object)
10 Inspection device (judgment device)
12 Data collection controller (acquisition means)
21 Measured value / expected value output section (output means)
23 Target feature value specifying unit (changing means)
25 Misclassification rate actual value calculation unit (calculation means)
26 misclassification rate expected value calculation unit (predicted value calculation means)
27. Margin evaluation unit (threshold setting judgment means)
28 Deviation degree evaluation unit (difference determination means)
29 Display control unit (output means)
30 Data storage (storage device)
40 display unit 50 input unit (acquisition instruction receiving unit / change instruction receiving unit / threshold designation receiving unit)

Claims (15)

A characteristic value that is information serving as an element for distinguishing whether or not the inspection object is normal is acquired from the inspection object, and whether or not the inspection object is normal is determined according to the characteristic value. A determination device for
When the threshold value for determining whether or not the inspection object is normal is arbitrarily set from the distribution information of the characteristic values acquired from each of the plurality of normal inspection objects, the normal inspection object is abnormal. A calculating means for calculating an overdetection rate that is a ratio to determine that
Predicted value calculation means for calculating a predicted value of the overdetection rate based on the characteristic value by a statistical method;
A graph showing a change in the overdetection rate calculated by the calculation unit according to the change in the threshold, and a change in the predicted value of the overdetection rate calculated by the prediction value calculation unit in accordance with the change in the threshold. An output device for outputting a graph to be displayed. An acquisition means for acquiring the characteristic value from the inspection object;
An acquisition instruction receiving unit that receives an acquisition instruction of a characteristic value from a new inspection object by a user;
2. The acquisition unit according to claim 1, wherein when the acquisition instruction receiving unit receives an instruction to acquire a characteristic value from the new inspection object, the acquisition unit acquires the characteristic value from the new inspection object. Judgment device. An acquisition means for acquiring the characteristic value from the inspection object;
A difference between the calculated overdetection rate and the predicted value of the overdetection rate is obtained based on the overdetection rate calculated by the calculation unit and the predicted value of the overdetection rate calculated by the predicted value calculation unit. Difference determining means for determining whether or not the difference is equal to or greater than a predetermined value,
The difference determination unit, when determining that the difference is equal to or greater than a predetermined value, instructs the acquisition unit to acquire a characteristic value from a new inspection object. Judgment device. A characteristic value that is information serving as an element for distinguishing whether or not the inspection object is normal is acquired from the inspection object, and whether or not the inspection object is normal is determined according to the characteristic value. A determination device for
When the threshold value for determining whether or not the inspection object is normal is arbitrarily set from the distribution information of the characteristic values acquired from each of the plurality of normal inspection objects, the normal inspection object is abnormal. A calculating means for calculating an overdetection rate that is a ratio to determine that
Predicted value calculation means for calculating a predicted value of the overdetection rate based on the characteristic value by a statistical method;
An output unit that outputs a graph indicating a change in the predicted value of the overdetection rate calculated by the predicted value calculation unit according to a change in the threshold;
An acquisition means for acquiring the characteristic value from the inspection object;
A difference between the calculated overdetection rate and the predicted value of the overdetection rate is obtained based on the overdetection rate calculated by the calculation unit and the predicted value of the overdetection rate calculated by the predicted value calculation unit. Difference determining means for determining whether or not the difference is equal to or greater than a predetermined value,
The determination apparatus characterized by instructing the acquisition unit to acquire a characteristic value from a new inspection object when the difference determination unit determines that the difference is equal to or greater than a predetermined value. The characteristic value distribution information further includes characteristic value distribution information acquired from each of the inspection objects in which a plurality of abnormalities occur,
The calculation means that the inspection object in which an abnormality has occurred is normal when the threshold value for determining whether the inspection object is normal is arbitrarily set from the distribution information of the characteristic value. We ’re calculating the missed rate, which is the percentage we ’re judging,
The predicted value calculation means further calculates a predicted value of the miss rate based on the characteristic value by a statistical method,
The output means is a graph showing changes in the overdetection rate and the overlook rate calculated by the calculation means according to a change in the threshold value, and an overdetection calculated by the predicted value calculation means according to the change in the threshold value. The determination apparatus according to claim 1, wherein the determination device outputs a graph indicating a change in a predicted value of a rate and a miss rate. Change means for changing the type of characteristic value for calculating the overdetection rate and the missed rate and the predicted value of the overdetection rate and the overlooked rate;
A change instruction receiving unit that receives information from the user for instructing to change the type of the characteristic value,
6. The determination apparatus according to claim 5, wherein when the change instruction receiving unit receives information instructing to change the type of the characteristic value, the changing unit changes the type of the characteristic value to be designated. An acquisition means for acquiring the characteristic value from the inspection object;
An acquisition instruction receiving unit that receives an acquisition instruction of a characteristic value from a new inspection object by a user;
The acquisition unit acquires a characteristic value from a new inspection object when the acquisition instruction receiving unit receives an instruction to acquire the characteristic value from the new inspection object. Judgment device. Changing means for changing the type of characteristic value for calculating the overdetection rate and the overlooked rate, and the predicted value of the overdetection rate and the overlooked rate,
A storage device for holding tolerance information indicating an allowable overdetection rate value and an allowable oversight rate value;
Based on the tolerance information and the predicted value of the overdetection rate and the missed rate calculated by the predicted value calculation means, the predicted value of the overdetection rate is less than the allowed value, and the predicted value of the missed rate is the allowed value Obtain threshold range information, which is the range where thresholds can be set, and the distribution of characteristic values obtained from each normal inspection object according to the obtained threshold range information and the inspection target in which an abnormality has occurred Threshold setting determination means for determining whether or not a threshold for determining the distribution of characteristic values acquired from each object can be set;
The threshold setting determination means determines that the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object in which an abnormality has occurred can be determined by the threshold. In this case, the output means is calculated by the predictive value calculation means according to the change in the threshold and the graph indicating the change in the overdetection rate and the miss rate calculated by the calculation means. Output a graph showing the change in the predicted value of the overdetection rate and the oversight rate,
The threshold setting determination means determines that the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object having an abnormality cannot be determined by the threshold. The determination apparatus according to claim 5, wherein the determination unit instructs the change unit to change the type of the characteristic value to be acquired. Changing means for changing the type of characteristic value for calculating the overdetection rate and the overlooked rate, and the predicted value of the overdetection rate and the overlooked rate,
The value obtained by calculating the square of the Mahalanobis distance of each characteristic value acquired from each of the test objects in which a plurality of abnormalities are obtained relative to the average distribution of the set of characteristic values acquired from each of the plurality of normal test objects. The distribution of characteristic values acquired from each normal inspection object and the characteristics acquired from each inspection object in which an abnormality has occurred according to the value obtained by obtaining the SN ratio of the desired large characteristic Threshold setting determination means for determining whether or not a threshold for determining a distribution of values can be set;
The threshold setting determination means determines that the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object in which an abnormality has occurred can be determined by the threshold. In this case, the output means is calculated by the predictive value calculation means according to the change in the threshold and the graph indicating the change in the overdetection rate and the miss rate calculated by the calculation means. Output a graph showing the change in the predicted value of the overdetection rate and the oversight rate,
The threshold setting determination means determines that the distribution of the characteristic value acquired from each normal inspection object and the distribution of the characteristic value acquired from each inspection object having an abnormality cannot be determined by the threshold. The determination apparatus according to claim 5, wherein the determination unit instructs the change unit to change the type of the characteristic value to be acquired. An acquisition means for acquiring the characteristic value from the inspection object;
The difference between the overdetection rate and the overdetection rate prediction value based on the overdetection rate and the oversight rate calculated by the calculation unit and the overdetection rate and the overlookment rate prediction value calculated by the prediction value calculation unit. And a difference determination means for determining a difference between the missed rate and the predicted value of the missed rate, and determining whether a combination of these differences is a predetermined value or more,
The difference determination unit, when determining that the difference is equal to or greater than a predetermined value, instructs the acquisition unit to acquire a characteristic value from a further inspection object. 10. The determination device according to any one of 9 above. A threshold value specification receiving unit for receiving information specifying the threshold value from the user;
According to the threshold value received by the threshold value specification receiving unit, the output means includes an overdetection rate and an overlook rate calculated by the calculation means and an overdetection rate and an overlooked rate prediction value calculated by the prediction value calculation means. It outputs, The determination apparatus of any one of Claims 1-10 characterized by the above-mentioned.   The predicted value calculation means calculates the square of the Mahalanobis distance of each characteristic value with respect to the average of the distribution of the characteristic values for the set of characteristic values acquired from each of the plurality of normal inspection objects. The determination apparatus according to claim 1, wherein the prediction value is calculated by obtaining an upper probability of the F distribution.   The predicted value calculation means uses a plurality of sets generated by combining characteristic values acquired from a plurality of inspection objects as training samples, and overdetects the overdetection rate and the oversight rate in each set of the training samples. The determination device according to claim 1, wherein the determination device calculates the predicted value of the rate and the miss rate.   A control program for operating the determination apparatus according to claim 1, wherein the control program is for causing the computer to function as each means.   The computer-readable recording medium with which the control program of the determination apparatus of Claim 14 was recorded.

Images