JP7400639B2 - Method for creating a defective parts prediction model, defective parts prediction method, defective parts prediction program, and defective parts prediction device - Google Patents

Description

The present disclosure relates to a method for creating a defective parts prediction model, a defective parts prediction method, a defective parts prediction program, and a defective parts prediction device, and in particular, a measuring device that is determined to be defective in an environmental test based on the result data of an environmental test. Related to technology for identifying defective parts.

A spectrophotometer is a measuring instrument that measures the color of an object based on the reflection or transmittance of each wavelength, and like other electronic components and devices, various tests are conducted for quality control purposes before product shipment. be done. A spectrophotometer consists of many parts such as electronic parts and optical parts, and if the final product is inspected and determined to be defective, the parts that are presumed to be the cause of the problem will be removed. will be replaced and inspected again.

However, since each part constituting an assembly is inspected beforehand and assembled, it is not always easy to estimate which part is causing the problem. Therefore, if a part that is not the cause of the problem is incorrectly assumed to be the cause of the problem, the part will be wasted, the effort to replace the part will be wasted, and the construction period will be extended. As a result, cost losses occur.

In addition, since it is necessary to evaluate all the final products, the final product may be determined to be defective, especially if the final product has a large number of parts and the accuracy of estimating the part that is causing the defect is low. If parts are replaced repeatedly through trial and error, the cost loss will be enormous.

To solve this problem, for example, it is possible to determine mechanically whether the parts that make up the assembly are good or defective based on the operating sounds generated when the assembly, such as a steering device, is operated mechanically. A technique for making predictions using a learning model has been proposed (see, for example, Patent Document 1).

In addition, a technology has been proposed that uses a machine learning model to estimate the presence or absence of assembly defects and the type of assembly defects in the spindle and spindle motor that make up a machine tool, based on the spindle motor's current value, vibration value, audio value, etc. ( For example, see Patent Document 2).

By using these conventional technologies, it is expected that it will be possible to reduce unnecessary parts replacement because it will be possible to estimate which parts are defective based on the inspection results obtained by mechanically operating the product. be done.

JP2019-204232A JP2019-204232A

Environmental testing is one of the product inspections performed on spectrophotometers. Among environmental tests, for example, a temperature and humidity test confirms how resistant a spectrophotometer is to atmospheric temperature and humidity and changes thereof, and whether it can maintain its performance as a measuring instrument. Spectrophotometers that are found to be of good quality through temperature and humidity tests are shipped with a certificate of acceptance.

Even if a product is determined to be defective in the temperature/humidity test, similar to cases in which the product is determined to be defective in other tests, corrections such as parts replacement are attempted. Similar to other inspections, it is necessary to reduce the cost of this modification as much as possible.

However, since a spectrophotometer performs measurement operations without mechanical operations, the above-mentioned conventional method uses mechanical operations to estimate which parts are defective based on vibration conditions, etc. Technology cannot be applied as is. Furthermore, although electronic components are incorporated, the amount of variation in current and voltage values is limited, so it is difficult to identify defective components from current values and the like.

Such problems are common not only to spectrophotometers but also to a wide range of measurement devices that perform environmental tests.

The present disclosure has been made in view of the above-mentioned problems, and is a method for predicting defective parts for predicting defective parts constituting a measuring instrument that has been determined to be defective in an environmental test based on data on the results of an environmental test. The object of the present invention is to provide a model creation method, a defective parts prediction method, a defective parts prediction program, and a defective parts prediction device.

In order to achieve the above object, a method for creating a defective parts prediction model according to an embodiment of the present disclosure includes a method for creating a defective parts prediction model according to an embodiment of the present disclosure. , a recording step in which the names of replacement parts are recorded in association with each other, and the measurement data recorded for each measuring device is acquired, and feature quantities are extracted from the measurement data corresponding to each region in which the measurement range of the measuring device is divided. a creation step in which a feature vector is created using the feature vector as a component; a classification step in which the feature vector is classified into defect types based on the similarity between the feature vectors; The method is characterized in that it includes a labeling step of associating replacement part names associated with measurement data used to create feature vectors classified into types.

In this case, the measurement data may include measurement data obtained by using a measuring instrument under a plurality of environmental conditions, and the creation step combines measurement data corresponding to different environmental conditions to Quantities may be generated.

Further, in the classification step, the feature vectors may be classified by cluster analysis.

Furthermore, after creating the defective parts prediction model, if new measurement data and defective part names are recorded in the recording step, the creation step, the classification step, and the labeling step are performed using the new measurement data. It may also include a relearning step of updating the defective parts prediction model.

Further, the recording step may further include recording at least one of measurement data obtained in an environmental test after parts replacement and a result of the environmental test in association with the measurement data.

Further, it is preferable that the environmental test is a temperature/humidity test.

Further, the measuring device may be a spectrophotometer, and the measurement range may be a wavelength range in which the spectrophotometer detects the reflectance of the measurement target.

In addition, the defective parts prediction method according to one embodiment of the present disclosure includes an acquisition step of acquiring measurement data obtained in the environmental test for a measuring device determined to be defective in the environmental test, and a step of acquiring measurement data obtained from the acquired measurement data. A creation step of calculating a feature amount for each area into which the measurement range of is divided and creating a feature vector having the feature amount as a component; and a defect created by the creation method according to any one of claims 1 to 8. a type identification step of identifying a defect type to which the feature vector belongs using a component prediction model; and a component name prediction step of predicting a defective component name corresponding to the defect type using the defective component prediction model. It is characterized by containing.

In this case, an output step may be provided to output the name of the identified defective part. The method may also include a probability obtaining step of obtaining a probability of failure occurrence for each part of the measuring instrument, and the output step may further output the probability of failure occurrence for each part.

Further, the defective parts prediction program according to one embodiment of the present disclosure includes an acquisition step of acquiring measurement data obtained in the environmental test for a measuring device determined to be defective in the environmental test, and a step of acquiring measurement data obtained from the acquired measurement data. A creation step of calculating a feature amount for each area into which the measurement range of is divided and creating a feature vector having the feature amount as a component; and a defect created by the creation method according to any one of claims 1 to 8. a type identification step of identifying a defect type to which the feature vector belongs using a component prediction model; and a component name prediction step of predicting a defective component name corresponding to the defect type using the defective component prediction model. It is characterized by being executed by a computer.

Further, the defective parts prediction device according to one embodiment of the present disclosure includes an acquisition means for acquiring measurement data obtained in the environmental test, and a method for predicting the measuring device based on the acquired measurement data. 9. A creation means for calculating a feature amount for each area into which the measurement range of is divided and creating a feature amount vector having the feature amount as a component; and a defect created by the creation method according to any one of claims 1 to 8. A type specifying means for specifying a defect type to which the feature vector belongs using a parts prediction model; and a part name prediction means for predicting a defective part name corresponding to the defect type using the defective parts prediction model. It is characterized by being prepared.

If the measurement data of a measuring instrument that has been determined to be defective in an environmental test deviates from the standard tolerance range and does not achieve the desired accuracy, it is difficult to determine which area of measurement data within the measuring range of the measuring instrument meets the standard. Depending on whether the conditions are met, the parts that should be replaced may differ. In such a case, if you associate the parts that passed the environmental test after replacement with the defect types classified by the feature vectors, you can easily determine which defect type the feature vectors created from the measurement data correspond to. This makes it possible to create a predictive model that accurately predicts which parts should be replaced.

By using such a predictive model, it is possible to reduce parts loss and man-hours for identifying defective parts that constitute a measuring instrument that has been determined to be defective in an environmental test. In addition, if parts are replaced many times until the environmental test passes, the environmental test will also be repeated, but the number of retests can also be reduced, reducing the trouble of retesting. Costs can also be reduced.

FIG. 1 is a diagram showing the main configuration of a defective parts prediction system. (a) is a graph illustrating temperature and humidity settings and measurement timing by the spectrophotometer 111 in a temperature/humidity test, and (b) is a flowchart showing the flow of work performed by a person in charge of the temperature/humidity test. The work flow according to this embodiment is compared with the conventional work flow. (a) is a graph illustrating the relationship between σ reflectance and wavelength in the repeatability test, and (b) is a graph illustrating the relationship between Δ reflectance and wavelength in the difference test between preliminary phase A and high temperature phase B. This is an illustrative graph. (a) is a graph illustrating the relationship between the standard value in the repeatability test and 3σ of the σ reflectance of a non-defective product, and (b) is a graph illustrating the relationship between the standard value and the non-defective product in the difference test between preliminary phase A and high temperature phase B. It is a graph illustrating the relationship between Δ reflectance and ±3σ. This is a table that summarizes explanatory variables, focusing on combinations of each phase of the temperature/humidity test and the wavelength region into which the measurement range of the spectrophotometer 111 is divided. (a) is a diagram illustrating a leakage current generation mechanism in a high temperature and high humidity environment, and (b) is a graph illustrating the relationship between relative humidity and leakage current amount. This is a table in which parts to be replaced for each explanatory variable are estimated by considering the characteristics of each part of the spectrophotometer 111 from the operating principle. FIG. 3 is a conceptual diagram illustrating clusters in which feature vectors are classified by defect type. 1 is a block diagram showing the main hardware configuration of a defective parts prediction device 100. FIG. (a) is a block diagram showing the functional configuration of the defective parts prediction device 100 for creating a predictive model of defective parts, and (b) is a block diagram showing the functional configuration of the defective parts prediction device 100 for predicting defective parts. FIG. 2 is a block diagram showing a functional configuration. (a) is a flowchart showing the operation of the defective parts prediction device 100 for creating a predictive model of defective parts, and (b) is a flowchart showing the operation of the defective parts prediction device 100 for predicting defective products. It is. (a) is a table explaining example 1 in which a single component satisfies the standard, but the product as a whole does not meet the standard, and (b) is a table that explains example 1 where a part that satisfies the standard no longer satisfies the standard due to environmental conditions. It is a table explaining Example 2 in which the total product does not satisfy the standards.

DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of a defective parts prediction device, a defective parts prediction method, and a defective parts prediction program according to the present disclosure will be described with reference to the drawings.
[1] Configuration of defective parts prediction system First, the configuration of the defective parts prediction system according to the present embodiment will be described.

As shown in FIG. 1, the defective parts prediction system 1 predicts which parts of the spectrophotometer 111 are defective parts using the result data of the environmental test (temperature and humidity test) regarding the spectrophotometer 111. This system includes a defective parts prediction device 100 and an external storage device 101.

The spectrophotometer 111 is a device that measures the spectral reflectance of the measurement target 123 by irradiating the measurement target 123 with illumination light and detecting the reflected light. The wavelength range in which the spectrophotometer 111 detects spectral reflectance often includes the ultraviolet region and the visible light region, and some spectrophotometers 111 are capable of detecting spectral reflectance in the infrared region.

The environmental test of the spectrophotometer 111 is performed using the environmental test device 121. The environmental test device 121 is equipped with a temperature and humidity control device 122, and can adjust the interior to a desired temperature and humidity. In an environmental test of the spectrocolorimeter 111, the spectrocolorimeter 111 and the measurement object 123 are housed in the environmental test device 121, the interior is adjusted to a predetermined temperature and humidity, and the spectrocolorimeter 111 is The color of the measurement object 123 is measured.

In this embodiment, a member whose surface to be measured is white is used as the measurement target object 123. Being white means that the variation in reflectance for each wavelength is within a predetermined range.

Measurement data obtained as a result of colorimetry is collected by connecting the terminal device 102 to the spectrophotometer 111. The terminal device 102 may be connected to the spectrophotometer 111 using, for example, a serial interface such as RS-232C, or may use another interface. The person in charge of the environmental test refers to the measurement data and determines whether the spectrophotometer 111 is a good product or a defective product.

The terminal device 102 is connected to the defective parts prediction device 100 for mutual communication via a communication network 130 such as a LAN (Local Area Network) or the Internet. When the person in charge of the environmental test determines that the spectrophotometer 111 is a defective product, the person in charge of the environmental test transmits the measurement data to the defective parts prediction device 100. Note that measurement data may be written directly from the terminal device 102 to the external storage device 101 and referred to by the defective parts prediction device 100.

The defective parts prediction device 100 refers to the measurement data received from the terminal device 102 and the measurement data read from the external storage device 101 to predict defective parts of the spectrophotometer 111 .
[2] Environmental Test Next, an environmental test performed on the spectrophotometer 111 will be described.

As described above, as an environmental test for the spectrophotometer 111, a temperature and humidity test is performed in this embodiment. Temperature and humidity tests are conducted to determine whether (1) the effects of temperature and humidity and their changes on the measured data are within the standard range, and (2) whether the fluctuations in the measured data mainly due to the temperature characteristics of optical components and circuit components are within the standard range. (3) If a leak path occurs due to humidity, the measurement data variation due to leakage current flowing through the leak path is within the standard range. . Moreover, the temperature and humidity test of the spectrophotometer 111 is a 100% evaluation, and all the spectrocolorimeter 111 are tested.

The spectrophotometer 111 photoelectrically converts reflected light from an object to be measured into an electrical signal and detects the amount of reflected light, so the current handled in the first stage of the detection circuit is several femtoamperes. In some cases, the current is minute. Another feature is the use of high-performance optical components. These are easily affected by environmental temperature and humidity and cause errors in measurement data, so a temperature and humidity test is essential to ensure normal operation of the spectrophotometer 111.

In the temperature and humidity test according to the present embodiment, as illustrated in FIG. 121 is maintained at normal temperature and normal humidity for a predetermined period of time, the measurement target object 123 is measured by the spectrophotometer 111 at timing T1. The spectrophotometer 111 records spectral reflectance as measurement data.

In the next high temperature phase B, the interior of the environmental test device 121 is kept in a high temperature and normal humidity state for a predetermined period of time, and then at timing T2, the measurement object 123 is measured with the spectrophotometer 111, and the measurement data is recorded. . Furthermore, in the high temperature and high humidity phase C, after maintaining the high temperature and high humidity state inside the environmental test device 121 for a predetermined period of time, the measurement object 123 is measured with the spectrophotometer 111 at timing T3, and the measurement data is stored. Record.

Thereafter, the inside of the environmental test device 121 is left at the highest temperature state higher than that in Phases B and C for a predetermined period of time, and then in post-phase D, the inside of the environmental test device 121 is left at normal temperature and normal humidity for a predetermined period of time. After this is maintained, the measurement target object 123 is measured with the spectrophotometer 111 at timing T4, and the measurement data is recorded.

Note that the time for maintaining the inside of the environmental test device 121 at a predetermined temperature and humidity may be, for example, a time sufficient for the entire spectrophotometer 111, including the inside, to reach the temperature and humidity. It is desirable to set the time according to the type of defect to be detected for each phase. If it takes a long time to complete the post-phase D after starting the pre-phase A, the environmental test may be automatically executed at night. In other words, the testing procedure is such that the person in charge of the test prepares for the test, starts automatic execution of the environmental test in the evening, and collects the test results the next morning.

As shown in FIG. 2(b), after carrying out the temperature and humidity test as described above (S201), the person in charge of the temperature and humidity test refers to the measurement data of the spectrophotometer 111 and performs the spectrocolorimetry. A total of 111 samples are judged to be good or bad. If the test result is defective (S202: NG), the person in charge of the test inputs the measurement data into the defective parts prediction device 100 using the terminal device 102 (S203), and the defective parts prediction device 100 uses the measurement data. A defective component is predicted using the defective component prediction model (S204).

The person in charge of the test replaces the component predicted to be a defective component by the defective component prediction device 100 (S205), and performs the temperature/humidity test again (S201). If the test result is good (S202: OK), the name of the part replaced immediately before the temperature/humidity test that was determined to be good, and the measurement data obtained from the temperature/humidity test conducted immediately before the replacement. It is associated and recorded in the external storage device 101 (S206).

If measurement data and replacement parts are recorded in association with each other in this way, trends in measurement data corresponding to parts to be replaced can be grasped by analyzing the measurement data accumulated in the external storage device 101. . Thereafter, the spectrophotometer 111 can be shipped (S207). If it is determined to be defective again (S202: NG), the above procedure is repeated.

Conventionally, when the test result was defective (S202: NG), the person in charge of the test contacted the person in charge of the model (engineer, etc.) of the spectrophotometer 111 (S203'), and the person in charge of the model who received the contact The person in charge of the test confirms the measurement data and instructs the person in charge of the test to take appropriate action (S204').Following these instructions, the person in charge of the test replaces the parts of the spectrophotometer 111 (S205'), and performs the temperature and humidity test again. was being carried out (S201).

However, even if the person in charge of the model analyzes the measurement data and replaces the defective parts estimated using past knowledge, the test results may not improve easily, or even if the measurement data is analyzed in detail, problems may occur. In many cases, the cause cannot be identified. For example, the graph illustrated in FIG. 3A shows the standard deviation of reflectance for each wavelength (hereinafter referred to as "σ reflectance") when measurement data is repeatedly recorded for each spectrophotometer 111. , where the horizontal axis is the wavelength.

In FIG. 3(a), the dashed line graph 300 is the standard value of σ reflectance, whereas the measurement data represented by graph 301 satisfies the standard, while graph 302 shows that the standard is met in the short wavelength range. Graph 303 indicates that the standard is not met in the medium wavelength range. Furthermore, the graph 304 does not meet the standards in the long wavelength range.

Furthermore, since the spectrophotometer 111 is required to be able to accurately measure color at both room temperature and high temperature, it is desirable that the difference in measurement data between preliminary phase A and high temperature phase B be small. . In FIG. 3(b), the vertical axis represents the difference in reflectance (hereinafter referred to as "Δ reflectance"), the horizontal axis represents wavelength, and the broken lines 310+ and 310- represent the upper and lower limits of the standard for Δ reflectance. represents a value.

Graph 311 is within the standard value over the entire wavelength range, but graph 312 has a portion where the lower limit of the Δ reflectance deviates from the standard value in the long wavelength range. In this way, by referring to the measurement data, it is easy to understand why the spectrophotometer 111 was determined to be defective, but it is up to the person in charge of the model to determine which part is the cause of the defective determination. Even if there is, it is often not immediately obvious.

For these reasons, parts must be replaced an average of four times before a spectrocolorimeter 111 determined to be defective is made into a non-defective product. In these four parts replacements, there is a high possibility that parts other than the defective parts have also been replaced, and it also takes time and effort to redo the temperature and humidity test. It has become.
[3] Defective Parts Prediction Model Next, a description will be given of a defective parts prediction model that predicts defective parts from the measurement data of the temperature and humidity test of the spectrophotometer 111.

Whether the measurement data obtained using the spectrophotometer 111 satisfies the standard values is important for controlling product quality, but it is not necessarily appropriate as an explanatory variable for predicting defective parts. do not have. Therefore, in this embodiment, as a measure for determining the quality of σ reflectance and Δ reflectance, 3σ (standard deviation σ The explanatory variable is whether the value is within the range of 3 times .

In FIG. 4A, a broken line graph 300 represents the standard value of σ reflectance as a product of the spectrophotometer 111, and a solid line graph 400 represents a threshold value for identifying defective parts. Since the graph 400 is a graph obtained by adding 3σ to the average value of the σ reflectance of the spectrophotometer 111 that has been determined to be a good product, it is not necessarily smooth, but it naturally follows the standard value (graph 300) over the entire wavelength range. are satisfied.

In addition, in FIG. 4(b), broken line graphs 310+ and 310- represent the upper and lower standard values of the Δ reflectance as a product of the spectrophotometer 111, and solid line graphs 410+ and 410- represent defective parts. represents the upper and lower thresholds for specifying. The graph 410+ is a graph in which 3σ is added to the average value of the Δ reflectance of the spectrophotometers 111 that are determined to be non-defective, and the graph 410- is the average value of the Δ reflectance of the spectrophotometers 111 that are determined to be non-defective. Since this is a graph obtained by subtracting 3σ from , it naturally satisfies the standard values (graphs 310+, 310-) over the entire wavelength range.

Whether the measurement data of the spectrophotometer 111 obtained in the temperature/humidity test satisfies these threshold values is used as an explanatory variable and hereinafter referred to as "high temperature ΔRef."
Furthermore, σΔE/ΔE is also used as an explanatory variable. ΔE is the color difference between two colors in the color space Lab. In this embodiment, the color difference ΔE is calculated from the Lab coordinate value of the color of the measurement object 123 and the Lab coordinate value of the color represented by the measurement data obtained using the spectrophotometer 111 in the temperature and humidity test. . This calculation uses the calculation method specified by the Commission international de l'eclairage (CIE) in 1976.

σΔE is a value representing the variation in ΔE when repeatedly measured data is recorded using the spectrophotometer 111, and is expressed by the following formula (1).
σΔE={(σL) ² + (σa) ² + (σb) ² } ^1/2 …(1)
Calculated using

Similarly, the color difference ΔE between the preliminary phase A and the high temperature phase B is determined, and whether this color difference ΔE is smaller than a predetermined threshold value is also used as an explanatory variable. This explanatory variable will be referred to as "high temperature ΔE."

Also, the explanatory variable is the wavelength range in which the σ reflectance does not satisfy the standard value. In this embodiment, the wavelength region of 400 nm or more and 460 nm or less is referred to as a short wavelength region, and a situation where the σ reflectance does not satisfy the standard value in the short wavelength region is referred to as "short wavelength NG". Similarly, 470 nm or more and 640 nm or less are called medium wavelength range, and 670 nm or more and 700 nm or less are called long wavelength range, and failure to meet the standard values in these wavelength ranges is called ``medium wavelength NG'' and ``long wavelength NG,'' respectively. NG.”

It goes without saying that the number of partial regions when dividing a wavelength region is not limited to three, and may be two, four or more. Furthermore, as described above, the wavelength regions do not necessarily have to be adjacent to each other, and may be spaced apart, such as the upper limit of 460 nm in the short wavelength region and the lower limit of 470 nm in the middle wavelength region. Moreover, the range of the wavelength region is not limited to the above.

It is desirable to divide the wavelength range according to the final replacement parts. For now, it may be possible to divide the wavelength range into small sections and use clustering, which will be described later, to group together different wavelength ranges by defect type, but if the number of explanatory variables becomes too large, it may become difficult to create a predictive model. However, it is not necessarily effective.

FIG. 5 is a table summarizing explanatory variables. Based on the past results of temperature and humidity tests regarding the spectrophotometer 111, as shown in FIG. It was found that there is a high temperature ΔRef NG and a high temperature ΔE NG.

Note that humidity leak is a phenomenon that can occur when the environment in which the spectrophotometer 111 is placed is in a high temperature and high humidity state, and as illustrated in FIG. Dust or dirt 602 that has adhered across the pattern 601 and signal pattern 603 becomes electrically conductive under the influence of a high-temperature, high-humidity environment, resulting in a short circuit between the power supply pattern 601 and signal pattern 603, and a leakage current flows. be.

The graph shown in FIG. 6(b) is a graph showing the relationship between leakage current and relative humidity. For example, when the environmental temperature is 35°C, the leakage current tends to increase as the relative humidity RH increases. This tendency is particularly noticeable in high humidity environments. Note that the humidity leak does not necessarily affect the quality of the spectrophotometer 111, so it is not necessarily a problem in quality tests.

Components constituting the spectrocolorimeter 111 include, for example, a light emitting board, a main board, a xenon tube, a mirror, an analog board, a digital board, a cell block, and the like. The xenon tube is a light source for illuminating the measurement target 123, and the light emitting board is a circuit board on which a light emitting circuit for causing the xenon tube to emit light is mounted. The mirror is an optical member (integrating sphere) used to uniformly illuminate the measurement object 123 from all directions. The main board is a circuit board on which the main circuit of the spectrophotometer 111 is mounted.

The analog board is a circuit board on which analog circuits such as a light receiving circuit for detecting reflected light from the measurement target 123 are mounted, and the digital board performs digital signal processing on digitized detection signals. This is a circuit board on which digital circuits to be executed are mounted. A cell block is a unit of a spectroscopic optical system.

Considering the characteristics of each component of the spectrophotometer 111 from its operating principle, it is inferred that if the component is defective, defects as shown in the table shown in FIG. 7 will appear in the explanatory variables. That is, if a problem occurs in the light emitting substrate and the light emission of the xenon tube becomes unstable, the amount of light is particularly small in the short wavelength range, so repeatability deteriorates and short wavelength NG occurs. As a result, σΔE for the entire wavelength range also deteriorates. The same applies if the xenon tube is defective.

It is known that analog boards have circuit parts that, in principle, cause a humidity leak phenomenon.

The mirror generates a high temperature ΔRef NG because its reflectance changes as the environmental temperature changes.

The relationship between the cell block and long wavelength NG, and the relationship between the mirror and short wavelength NG, is not necessarily a theoretical consideration, but by comparing measurement data before and after replacement, etc., the relationship between parts and explanatory variables can be confirmed. Suggested.

However, it is not easy to identify defective parts from actual measurement data only by considering the operating principle, and conventionally, parts must be replaced by trial and error.

On the other hand, there are parts that were replaced immediately before the spectrophotometer 111 was determined to be a good product, or in other words, parts that are presumed to be defective parts that caused the spectrocolorimeter 111 to be determined to be a defective product. A clustering algorithm (in this embodiment, k Analysis using the average method) revealed that using six explanatory variables: short wavelength NG, long wavelength NG, humidity leak, σΔE NG, high temperature ΔRef NG, and high temperature ΔE NG, cell blocks, mirrors, analog boards, and light emitting boards It has been found that it is possible to predict which part should be replaced to determine whether the spectrocolorimeter 111 is a defective part or not.

In other words, by checking which cluster the feature values such as σ reflectance and Δ reflectance obtained from measurement data belong to, and replacing the parts corresponding to the cluster to which the feature values belong, the next step can be performed. The spectrophotometer 111 can be determined to be a good product in the temperature and humidity test.

In the example of FIG. 8, the spectrophotometer 111 is classified into three clusters 811, 812, and 813 based on a feature vector having feature amounts A, B, and C obtained from measurement data as components. Cluster 811 is a spectrophotometer 111 that was determined to be a good product in the temperature and humidity test by replacing part □, and cluster 812 was determined to be a good product in the temperature and humidity test by replacing part ○. This is a spectrophotometer 111 that has come to be used as a spectrophotometer. Further, the cluster 813 is the spectrophotometer 111 that was determined to be a good product in the temperature and humidity test by replacing the part Δ.

If the feature vector corresponding to the measurement data of the spectrophotometer 111 that newly conducted a temperature/humidity test is located at coordinates such as the star mark 800, the star mark 800 belongs to the cluster 813, so the spectrocolorimeter 111 in question In order for the total 111 to be determined to be a good product in the temperature and humidity test, it is only necessary to replace the part Δ corresponding to the cluster 813.
[4] Configuration of defective parts prediction device 100 Next, using the above-described defective parts prediction model, a device that predicts defective parts from the measurement data of the spectrophotometer 111 that was determined to be defective in the temperature and humidity test. The configuration of the defective parts prediction device 100 will be described.

As shown in FIG. 9, the defective component prediction device 100 includes a CPU (Central Processing Unit) 911, a ROM (Read Only Memory) 912, a RAM (Random Access Memory) 913, and the like. When the defective parts prediction device 100 is reset by turning on the power, etc., the CPU 911 starts by reading the boot program from the ROM 912, uses the RAM 913 as a working storage area, and reads the program from the HDD (Hard Disk Drive) 914. A predictive model creation program 900, a defective parts prediction program 901, an OS (Operating System) 902, etc. are executed.

The NIC (Network Interface Card) 915 executes communication processing for the defective component prediction device 100 to receive measurement data from the terminal device 102 via the communication network 130 . The CPU 911, ROM 912, RAM 913, HDD 914, and NIC 915 mutually transmit and receive data, control signals, etc. via an internal bus 916.

In FIG. 9, the external storage device 101 is also connected to the internal bus 916, but the external storage device 101 may be connected to the communication network 130 to communicate with the defective parts prediction device 100.

The defective parts prediction device 100 has a functional configuration as shown in FIG. 10(a) in order to create a prediction model of defective parts. First, in the external storage device 101, the replacement part name 1002 of the spectrophotometer 111 that was determined to be good in the temperature and humidity test by replacing the part is associated with the measurement data 1001 in the temperature and humidity test immediately before the part replacement. recorded.

The feature amount creation unit 1003 reads out the measurement data 1001 and the replacement part name 1002 from the external storage device 101, and calculates feature amounts such as σ reflectance, Δ reflectance, and ΔE for each wavelength region. Hereinafter, a vector having this feature amount as a component will be referred to as a "feature amount vector."

The classification unit 1004 classifies the feature vectors. In this embodiment, feature vectors are classified into four clusters, each corresponding to a defective part that should be replaced, by clustering using the k-means method. Commercially available statistical analysis software may be used to create feature amounts and classify feature vectors, such as JMP (registered trademark of SAS Institute Inc., USA).

The label assigning unit 1005 assigns a label representing the name of a defective part to be replaced to each cluster of feature vectors. Examples include "cell block", "mirror", "analog board", and "light emitting board".

The defective parts prediction device 100 further includes a functional configuration as shown in FIG. 10(b) in order to predict defective parts. First, the external storage device 101 records measurement data of the spectrophotometer 111 that was determined to be a defective product in the temperature and humidity test.

The input unit 1011 reads measurement data from the spectrophotometer 111 from the external storage device 101 and creates a feature amount vector.

The processing unit 1012 predicts a defective part from the feature vector created by the input unit 1011 using a prediction model of the defective part. As described above, when a feature vector belongs to any cluster, a defective part to be replaced is predicted based on the label given to the cluster.

Furthermore, if the feature vector is located at the coordinates illustrated by the star 801 in FIG. 8 and does not belong to any cluster, the distance D1 from the feature vector 801 to the clusters 811, 812, and 813 is , D2, and D3. This distance may be, for example, a Euclidean distance from the feature vector 801 to the center of gravity of each of the clusters 811, 812, and 813 in the feature space, or may be weighted for each feature component. Various distances other than Euclidean distance can be used.

Using the distances D1, D2, and D3 calculated in this way, the probability that each part corresponds to a defective part that should be replaced (the probability that the feature vector belongs to the clusters 811, 812, and 813) may be calculated. . For example, the probability P1 that the part corresponding to the label given to the cluster 811 should be replaced is:
P1 = D1/(D1+D2+D3)...(2)
You can also use it as Regarding parts related to clusters 812 and 813, probabilities P2 and P3 that the parts should be replaced can be similarly calculated.

Further, if the distances D1, D2, and D3 are all larger than a predetermined threshold value, it may be determined that the part to be replaced could not be predicted. This is because the reason why the spectrophotometer 111 was determined to be a defective product in the temperature and humidity test is likely to be due to a cause other than the component defect assumed in the prediction model. As such a threshold value, for example, a range of 3σ can be used for each cluster, taking into consideration the distribution of feature vectors in the cluster.

The output unit 1013 records the prediction result of the defective part by the processing unit 1012 in the external storage device 101 as information 1015 regarding the defective part. In this case, the terminal device 102 may be notified that the information 1015 regarding the defective parts has been recorded in the external storage device 101, or the information 1015 regarding the defective components may be transmitted to the terminal device 102.
[5] Operation of defective parts prediction device 100 Next, the operation of defective parts prediction device 100 will be explained.
(5-1) Operation of creating a predictive model for defective parts When creating a predictive model for defective parts, the defective parts prediction device 100 stores data recorded in the external storage device 101 as shown in FIG. 11(a). In order to create a predictive model for defective parts, steps S1101 to S1104 are executed for each temperature/humidity test on the recorded measurement data. Specifically, measurement data is acquired from the external storage device 101 (S1102). Steps S1101 and S1104 are symbols representing a loop process that repeats the process for each temperature/humidity test described above.

Here, the measurement data for predictive model creation refers to the temperature and humidity test conducted immediately before the last parts replacement for the spectrophotometer 111, which was determined to be a good product in the temperature and humidity test by replacing parts in the past, in other words, it is a good product. This is measurement data obtained using the spectrophotometer 111 in a temperature and humidity test immediately before the determined temperature and humidity test. As described above, the external storage device 101 records the part name of the last replaced part in association with the measurement data.

Next, a feature amount vector is created from the measurement data (S1103). For this reason, for example, a feature amount that becomes a component of a feature amount vector is calculated from the reflectance for each combination of environmental conditions (phases in temperature and humidity tests) and wavelength regions. Among the feature quantities, the σ reflectance is the standard deviation of the reflectance R for each wavelength when the reflectance R is repeatedly measured for each wavelength a plurality of times, and is calculated using the following equation (3).

(σ reflectance) = [Σ {(reflectance R) - (average value of reflectance R)} ² ]/N … (3)
Note that the number of samples of the reflectance R is N, and Σ is the sum of N reflectances. Needless to say,
(Average value of reflectance R) = Σ (reflectance R)/N (4)
It is.
Further, the Δ reflectance is the difference between the reflectance R measured in the preliminary phase A and the reflectance R measured in the high temperature phase B for each wavelength. That is,
(ΔReflectance) = (Reflectance R of preliminary phase A) - (Reflectance R of high temperature phase B) ... (5)
It is.

The calculation method specified by the International Commission on Illumination CIE in 1976 is used to calculate the color difference ΔE. σΔE is the standard deviation of ΔE, and when reflectance R is measured N times,
(σΔE) = [Σ{(ΔE) - (average value of ΔE)} ² ]/N … (6)
It is.

When creating a prediction model, all such measurement data may be used, or a portion thereof may be selected and used. Furthermore, once the prediction model is created, other measurement data may be added. In this case, the processes from step S1101 to step S1104 may be executed only for the added measurement data.

In addition, when calculating feature quantities, data processing may be performed based on rules derived from a theoretical perspective associated with defective parts, or based on rules derived from a data group of the same defective part. You may also process the data.

Next, the feature vectors are classified into defect types using a k-means clustering algorithm (S1105). In addition, the replacement parts information is acquired from the external storage device 101, and for each defect type, the replacement parts associated with the measurement data used to calculate the feature vector belonging to the defect type are identified (S1106), and the By assigning a label to the type, the defect type and replacement part name are associated with each other (S1108).

Finally, as a prediction model, information for determining which defect type cluster the feature vector created from the measurement data is included in, the distance between the feature vector and each cluster, and the defect type. The correspondence with the replacement part name is recorded in the external storage device 101 (S1109), and the process ends.
(5-2) Operation for predicting defective parts When predicting a defective part, the defective part prediction device 100 accepts it from the terminal device 102 or reads it from the external storage device 101, as shown in FIG. 11(b). Measurement data of the temperature and humidity test is acquired (S1111), and a feature vector is created from the measurement data (S1112). Next, the prediction model is read from the external storage device 101, the cluster to which the feature vector belongs is specified, the probability that the feature vector belongs to each cluster is calculated, and defective parts are predicted (S1113). .

If the cluster to which the feature vector belongs can be identified, the defective part can be predicted from the label given to the defect type of the cluster. Furthermore, when the probability to which the feature vector belongs is calculated, a defective part may be predicted from the label given to the cluster defect type with the highest probability.

Finally, the prediction result is output (S1114). The prediction result may be recorded in the external storage device 101 or transmitted to the terminal device 102. Alternatively, both may be performed.
[6] Modifications Although the present disclosure has been described above based on the embodiments, it goes without saying that the present disclosure is not limited to the above-described embodiments, and the following modifications can be implemented. .
(6-1) Since the temperature and humidity test is a test that is conducted after the product is completed, all parts of the spectrophotometer 111 are determined to be non-defective in tests that are conducted in the upstream process than the temperature and humidity test. ing. Regardless of this, there may be cases where defective parts remain at the temperature/humidity test stage as follows.

For example, consider a case where a repeatability test is performed on a product made up of parts A to E. In the acceptance inspection of parts, it is confirmed whether the supplier of each part satisfies specified tolerances. For example, if the standard value of the tolerance of part A is 1.15, it is confirmed in the acceptance inspection whether the tolerance of 1.15 is satisfied.

However, when adopting parts, if the tolerances of the parts are too strict, the cost of the parts will increase, so not only the tolerances but also the actual performance values of the parts are used as a reference. For example, even if the standard value of the tolerance of part A is 1.15, if almost all parts A satisfy the tolerance of 0.94, then the part A is adopted with the tolerance of 0.94 as the actual value. There are cases where This is because even if there are parts A that do not satisfy the actual performance value, they are few, so if the parts A themselves are inexpensive, there will be cost benefits for both the supplier and the supplier.

Therefore, as illustrated in the table of FIG. 12(a), the product is designed to satisfy the standard value of 1.20 for the total tolerance of the product, including the tolerances of other parts B to E.

However, depending on the lot of part A, the tolerance is 1.1, which satisfies the standard value of 1.15, so although it is determined to be a good product in the acceptance inspection, it does not satisfy the actual value of 0.94. There may be cases where it is not. In such a case, even if all parts A to E are determined to be good in the acceptance inspection, the total product tolerance will be 1.30, which does not satisfy the standard value of 1.20, so they will be considered defective. It will be judged.

Further, even if the tolerance of part A satisfies the actual value of 0.94, the following case may occur. That is, this is a case where it is only specified that the tolerance of part A should satisfy the standard value of 0.94 at room temperature. Although such a part A usually has a tolerance that satisfies the standard value at room temperature even in a high temperature and high humidity environment, depending on the lot or a single item, the tolerance under a high temperature and high humidity environment may exceed the standard value at room temperature. The standard value below may not be satisfied (for example, 1.2 in the table of FIG. 12(b)).

Even in such a case, although all parts A to E are determined to be good in the acceptance inspection, the product as a whole cannot pass the temperature and humidity test.

In this way, even if all parts pass the acceptance inspection, the product as a whole may not pass the inspection, and in such a case, it becomes difficult to identify defective parts.
(6-2) In the above embodiment, the case where the defective parts prediction device 100 and the terminal device 102 are separate units has been described as an example, but it goes without saying that the present disclosure is not limited to this. You can also do the following.

For example, the defective part prediction process may be executed by the terminal device 102. In this case, the defective parts prediction device 100 receives measurement data and replacement part data from the terminal device 102, only creates a prediction model for predicting defective products, and transfers the created prediction model to the terminal device 102. may be provided to The prediction model may be provided by directly transmitting it from the defective parts prediction device 100 to the terminal device 102.

Alternatively, the defective parts prediction device 100 may temporarily record the prediction model in the external storage device 101, and the terminal device 102 may acquire the prediction model from the external storage device 101. In this way, the processing load on the defective parts prediction device 100 can be reduced, especially when a plurality of terminal devices 102 use a common prediction model.
(6-3) In the above embodiment, the case where the environmental test is a temperature/humidity test is explained as an example, but it goes without saying that the present disclosure is not limited to this. It may be. Further, tests other than environmental tests may be used, and the effects of the present disclosure can be obtained regardless of the type of test.
(6-4) In the above embodiment, the case where the measuring instrument to be tested is a spectrophotometer is explained as an example, but it goes without saying that the present disclosure is not limited to this. The test object may be a measuring instrument. Furthermore, when a measuring instrument other than a spectrophotometer is to be tested, it is desirable to generate feature amounts for each region into which the measurement range of the measuring instrument is divided, instead of dividing the wavelength range.
(6-5) In the above embodiment, the measurement data of the spectrophotometer 111 obtained in the temperature and humidity test immediately before the spectrocolorimeter 111 is determined to be good in the temperature and humidity test by replacing parts is stored in external storage. Although the description has been given by taking the case of recording in the device 101 as an example, it goes without saying that the present disclosure is not limited to this. Temperature and humidity test result data (OK, NG, etc.) may also be recorded in the external storage device 101. In this way, it becomes possible to analyze the influence that component replacement has on the measurement data of the spectrophotometer 111 and the results of the temperature and humidity test.

The defective parts prediction device, the defective parts prediction method, and the defective parts prediction program according to the present disclosure are useful as a technology for reducing the effort and cost of identifying defective parts that should be replaced for measuring instruments determined to be defective in environmental tests. It is.

1......Failure parts prediction system 100...Failure parts prediction device 101......External storage device 102......Terminal device 111... ......Spectrocolorimeter 121......Environmental test device 122......Temperature control and disappearance device 123...Measurement objects 1001, 1014...Measurement data 1002 ……………Replacement part name 1015………………Information about defective parts

Claims (13)

a recording step of associating and recording the measurement data obtained in the environmental test before replacing the part and the name of the replacement part for each measuring instrument that was determined to be good in the environmental test due to the replacement of the part;
Obtain the measurement data recorded for each measuring device, calculate the feature amount from the measurement data corresponding to each region of the measurement range of the measuring device, and create a feature vector with the feature amount as a component. a creation step that creates
a classification step of classifying the feature vectors into defect types based on the similarity between the feature vectors;
A method for creating a defective parts prediction model, comprising: a labeling step of associating, for each defect type, a replacement part name associated with measurement data used to create a feature vector classified into the defect type. . 2. The method for creating a defective parts prediction model according to claim 1, wherein the measurement data includes measurement data obtained by using a measuring instrument under a plurality of environmental conditions. 3. The method for creating a defective parts prediction model according to claim 2, wherein in the creation step, feature quantities are created by combining measurement data corresponding to mutually different environmental conditions. 4. The method for creating a defective parts prediction model according to claim 1, wherein the classification step classifies the feature vectors by cluster analysis. After creating the defective parts prediction model, if new measurement data and defective part names are recorded in the recording step, the creation step, the classification step, and the labeling step are executed using the new measurement data. 5. The method for creating a defective parts prediction model according to claim 1, further comprising a relearning step of updating the defective parts prediction model. 6. The recording step further comprises recording at least one of measurement data obtained in an environmental test after parts replacement and a result of the environmental test in association with the measurement data. A method for creating a defective parts prediction model as described in any of the above. 7. The method for creating a defective parts prediction model according to claim 1, wherein the environmental test is a temperature/humidity test. The measuring instrument is a spectrophotometer,
8. The method for creating a defective parts prediction model according to claim 1, wherein the measurement range is a wavelength range in which the spectrophotometer detects the reflectance of the object to be measured. an acquisition step of acquiring measurement data obtained in the environmental test for the measuring device determined to be defective in the environmental test;
a creation step of calculating a feature amount for each region into which the measurement range of the measuring instrument is divided from the acquired measurement data, and creating a feature vector having the feature amount as a component;
A type identifying step of identifying a defect type to which the feature vector belongs using a defective part prediction model created by the creation method according to any one of claims 1 to 8;
A method for predicting a defective part, comprising: predicting a name of a defective part corresponding to a type of defect using the defective part prediction model. 10. The method for predicting a defective part according to claim 9, further comprising an output step of outputting the name of the specified defective part. including a probability acquisition step of acquiring a defect occurrence probability for each part of the measuring instrument;
11. The method of predicting defective parts according to claim 10, wherein the output step further outputs a probability of occurrence of a defect for each part. an acquisition step of acquiring measurement data obtained in the environmental test for the measuring device determined to be defective in the environmental test;
a creation step of calculating a feature amount for each region into which the measurement range of the measuring instrument is divided from the acquired measurement data, and creating a feature vector having the feature amount as a component;
A type identifying step of identifying a defect type to which the feature vector belongs using a defective part prediction model created by the creation method according to any one of claims 1 to 8;
A defective part prediction program that causes a computer to execute a part name prediction step of predicting a defective part name corresponding to a defect type using the defective part prediction model. an acquisition means for acquiring measurement data obtained in the environmental test for a measuring instrument that has been determined to be defective in the environmental test;
A creation means that calculates a feature amount for each region into which the measurement range of the measuring instrument is divided from the acquired measurement data, and creates a feature vector having the feature amount as a component;
Type identifying means for identifying a defect type to which the feature vector belongs using a defective part prediction model created by the creation method according to any one of claims 1 to 8;
A defective component prediction device comprising: a component name prediction means for predicting a defective component name corresponding to a defect type using the defective component prediction model.

Images