Machine learning device, inspection device, machine learning method and program

Abstract

To provide a machine learning device for more satisfactorily performing inspection of a membrane electrode assembly, an inspection device, a machine learning method and a program.SOLUTION: In an inspection system SYS, a machine learning device 100 includes: a data set acquisition unit 108 for acquiring foreign matter data 14 including at least one of area data 16, brightness data 18, and shape data 20 of a foreign material 42 grasped from an image of the foreign matter mixed in a membrane electrode assembly, and teacher data 40 including data showing the volume of the foreign matter; and a learning model generation unit 112 for generating a learning model 10 with the foreign matter data as an input and with the volume of the foreign matter as an output by executing supervised learning by using the teacher data acquired by the data set acquisition unit.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a machine learning device, an inspection device, a machine learning method, and a program. [Background technology]

[0002] In recent years, research and development has been conducted on fuel cells, which contribute to energy efficiency, to ensure that more people have access to affordable, reliable, sustainable, and advanced energy. Japanese Patent Application Laid-Open Publication No. 2021-135125 discloses a method for inspecting membrane electrode assemblies. According to the disclosure, the inspection device determines the presence or absence of foreign matter in a membrane electrode assembly based on the amount of decrease in brightness in an X-ray transmission image. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-135125 Summary of the Invention [Problem to be solved by the invention]

[0004] Recently, there has been a demand for a technique for better inspecting membrane electrode assemblies.

[0005] The present disclosure aims to solve the above-mentioned problems. [Means for solving the problem]

[0006] A first aspect of the present disclosure is a machine learning device comprising: an acquisition unit that acquires supervised data including foreign matter data including at least one of area data indicating the area of ​​a foreign matter contaminated in a membrane electrode assembly ascertained from an image of the foreign matter, brightness data indicating the brightness of the foreign matter ascertained from the image, and shape data related to the shape of the foreign matter ascertained from the image; and data indicating the volume of the foreign matter; and a learning model generation unit that performs supervised learning using the supervised data acquired by the acquisition unit to generate a learning model that uses the foreign matter data as an input and outputs the volume of the foreign matter.

[0007] A second aspect of the present disclosure is an inspection device comprising: a foreign matter data acquisition unit that acquires foreign matter data including at least one of area data indicating the area of ​​a foreign matter contaminated in a membrane electrode assembly ascertained from an image of the foreign matter, brightness data indicating the brightness of the foreign matter ascertained from the image, and shape data relating to the shape of the foreign matter ascertained from the image; and an estimation unit that estimates the volume of the foreign matter using the foreign matter data acquired by the foreign matter data acquisition unit and the learning model generated by the machine learning device according to the first aspect of the present disclosure.

[0008] A third aspect of the present disclosure is a machine learning method executed by a computer, comprising: an acquisition step of acquiring supervised data including foreign matter data including at least one of area data indicating the area of ​​a foreign matter contaminated in a membrane electrode assembly ascertained from an image of the foreign matter, brightness data indicating the brightness of the foreign matter ascertained from the image, and shape data related to the shape of the foreign matter ascertained from the image; and data indicating the volume of the foreign matter; and a learning model generation step of performing supervised learning using the supervised data acquired by the acquisition step to generate a learning model that uses the foreign matter data as an input and outputs the volume of the foreign matter.

[0009] A fourth aspect of the present disclosure is a program for causing a computer to execute the machine learning method according to the third aspect of the present disclosure. [Effects of the Invention]

[0010] According to the present disclosure, a machine learning device, an inspection device, a machine learning method, and a program are provided for better inspecting membrane electrode assemblies. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a configuration diagram of an inspection system according to an embodiment. [Figure 2] 2A and 2B are schematic diagrams of foreign matter contaminating a membrane electrode assembly. [Figure 3] FIG. 3 is a diagram showing the frequency distribution of the diameter ratio. [Figure 4] FIG. 4 is a diagram showing the frequency distribution of the area. [Figure 5] FIG. 5 is a flowchart of a machine learning method according to one embodiment. [Figure 6] FIG. 6 is a flowchart of an inspection method according to one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] A membrane electrode assembly (MEA) is a structure comprising a solid polymer membrane and a catalyst layer laminated on the solid polymer membrane. The membrane electrode assembly is used, for example, in a fuel cell. Foreign matter may be mixed into the membrane electrode assembly. For example, during the manufacturing process of the membrane electrode assembly, iron particles or the like from the manufacturing equipment may be mixed into the membrane electrode assembly. The foreign matter may be mixed between the solid polymer membrane and the catalyst layer, for example.

[0013] If a relatively large amount of foreign matter is mixed into a membrane electrode assembly, the membrane electrode assembly should be determined to be defective. However, it is not easy to accurately determine the volume of the foreign matter through non-destructive testing. If the inspection standards are set too strictly to prevent defective products from being shipped as non-defective products, non-defective products will be determined to be defective, resulting in a decrease in yield.

[0014] Based on the above preliminary explanation, one embodiment will be described below.

[0015] In the following description, computer programs (computer software) are also referred to as computer program products. Computer program products are not limited to computer programs recorded on a recording medium, but also include computer programs transmitted, distributed, or downloaded via an information and communication network such as the Internet.

[0016] (One embodiment) FIG. 1 is a configuration diagram of an inspection system SYS according to an embodiment.

[0017] The inspection system SYS is a system that performs quality inspection of a membrane electrode assembly provided in a fuel cell, etc. The inspection system SYS includes a machine learning device 100 and an inspection device 200.

[0018] The machine learning device 100 is a device that performs machine learning to generate a learning model (trained model) 10. A more detailed explanation of the learning model 10 will be given later. The machine learning device 100 is, for example, an electronic device such as a computer. The machine learning device 100 includes an operation unit 102, a memory unit 104, and a calculation unit 106.

[0019] The operation unit 102 includes an input device (not shown), which includes a pointing device, a keyboard, etc. The pointing device includes, for example, a mouse, a trackpad, etc., but is not limited to these.

[0020] The storage unit 104 includes one or more memories. The one or more memories include non-volatile memories. The non-volatile memories are recording media that non-temporarily store programs (computer programs), tables, maps, etc. For example, non-volatile memories include ROM (Read Only Memory), flash memory, etc. The storage unit 104 (one or more memories) may further include volatile memories. For example, volatile memories include RAM (Random Access Memory). Furthermore, at least a part of the storage unit 104 may be realized by recording media such as USB (Universal Serial Bus) memory, memory cards, optical discs, etc.

[0021] As shown in FIG. 1, the memory unit 104 stores a learning model 10. The learning model 10 can be realized by, for example, a machine learning algorithm applicable to a regression problem. The machine learning algorithm may be, but is not limited to, a support vector machine (e.g., a support vector regression model) or a neural network (e.g., a deep neural network). When foreign matter data 14 is input, the learning model 10 outputs volume estimation data corresponding to the foreign matter data 14. The foreign matter data 14 is data grasped from an image of a foreign matter 42 (see also FIGS. 2A and 2B) mixed in a membrane electrode assembly. A more detailed explanation of the foreign matter data 14 will be given later. The volume estimation data is data indicating the volume (estimated volume value) of the foreign matter 42.

[0022] The arithmetic unit 106 includes a predetermined processing circuit (not shown). The processing circuit includes one or more processors, such as a central processing unit (CPU) or a graphics processing unit (GPU). The processing circuit may include a predetermined integrated circuit, such as an application specific integrated circuit (ASIC) or a field-programmable gate array (FPGA).

[0023] The calculation unit 106 includes a dataset acquisition unit (acquisition unit) 108, a teacher data selection unit 110, and a learning model generation unit 112. The dataset acquisition unit 108, the teacher data selection unit 110, and the learning model generation unit 112 are realized by the calculation unit 106 (processor) executing a program stored in the storage unit 104 (memory). At least some of the dataset acquisition unit 108, the teacher data selection unit 110, and the learning model generation unit 112 may be realized by an integrated circuit such as the ASIC or FPGA described above.

[0024] The data set acquisition unit 108 acquires the data set 12. The data set 12 includes foreign substance data 14 for learning and volume data 38 corresponding to the foreign substance data 14.

[0025] As described above, the foreign substance data 14 is data obtained from an image of the foreign substance 42. The image includes, for example, an X-ray transmission image of the foreign substance 42. The foreign substance 42 (membrane electrode assembly), which is the subject of the image, is obtained, for example, by actually operating the manufacturing equipment described above. The foreign substance data 14 includes, for example, at least one of area data 16, brightness data 18, and shape data 20. The area data 16 is data indicating the area (projected area) of the foreign substance 42. The brightness data 18 is data indicating the brightness (brightness value) of the foreign substance 42. The shape data 20 is data related to the shape of the foreign substance 42. The area data 16 and the brightness data 18 are obtained by performing a predetermined image analysis process on the image of the foreign substance 42.

[0026] The shape data 20 includes at least one of circumference data 24, skewness data 26, kurtosis data 28, and diameter data 30. The circumference data 24 is data relating to the circumference of a foreign object 42. The circumference of the foreign object 42 is, for example, the length of a contour line BL of the foreign object 42 in a planar view of the image, but may also be the circumference of a virtual ring VR (described later) or the circumference of a virtual rectangle VQ (described later) (see also Figures 2A and 2B). The skewness data 26 is data relating to the skewness of the foreign object 42. The kurtosis data 28 is data relating to the kurtosis of the foreign object 42. The diameter data 30 is data relating to the diameter of a virtual ring VR, which is a virtual ring circumscribing (approximating) the foreign object 42 in a planar view, the ferret diameter of the foreign object 42 in a planar view, etc. The virtual ring VR is a virtual ellipse (virtual ellipse) VE, but may also be a virtual perfect circle (virtual circle). The ferret diameter may be the maximum ferret diameter or the minimum ferret diameter. The diameter data 30 may indicate the ferret diameter of the foreign object 42 along any direction in a planar view of the image.

[0027] The skewness data 26, kurtosis data 28, and diameter data 30, like the area data 16, brightness data 18, etc., are obtained by applying a predetermined image analysis process to the image of the foreign matter 42.

[0028] Fig. 2A is a schematic diagram of a foreign object 42 that has become mixed into a membrane electrode assembly. Fig. 2B is a schematic diagram of a foreign object 42 that has become mixed into a membrane electrode assembly. Fig. 2A shows a virtual ellipse VE that circumscribes the foreign object 42. In contrast, Fig. 2B shows a virtual rectangle (virtual rectangle VQ) that circumscribes the foreign object 42.

[0029] The diameter data 30 may include first diameter data 32 and second diameter data 34. The first diameter data 32 is data indicating the major axis A1 of the imaginary ellipse VE. In contrast, the second diameter data 34 is data indicating the minor axis A2 of the imaginary ellipse VE. The diameter data 30 may further include ratio data 36. The ratio data 36 is data indicating the ratio (ellipticity) between the major axis A1 and the minor axis A2.

[0030] The first diameter data 32 may be data indicating a first ferret diameter F1. In that case, the second diameter data 34 may be data indicating a second ferret diameter F2. The first ferret diameter F1 is the ferret diameter of the foreign object 42 in a first direction D1. The first direction D1 is a predetermined direction in a planar view of the image. In contrast, the second ferret diameter F2 is the ferret diameter of the foreign object 42 in a second direction D2. The second direction D2 is a direction intersecting the first direction D1 in a planar view of the image. In this case, the ratio data 36 may indicate the ratio (aspect ratio) between the first ferret diameter F1 and the second ferret diameter F2.

[0031] The second direction D2 is, for example, perpendicular to the first direction D1. In this case, the first ferret diameter F1 may match the length of one side of the rectangular virtual quadrangle VQ (FIG. 2B). The second ferret diameter F2 may match the length of the other side of the rectangular virtual quadrangle VQ that is perpendicular to the one side. Note that the virtual quadrangle VQ is not limited to a rectangle.

[0032] In the following description, the ellipticity or aspect ratio indicated by the ratio data 36 is also collectively referred to as the diameter ratio. In other words, the diameter ratio in the following description is information indicated by the ratio data 36.

[0033] The volume data 38 is data indicating the volume of the foreign matter 42 mixed in the membrane electrode assembly. The volume data 38 is obtained by actually measuring the volume of the foreign matter 42, which is the subject of the image described above. Therefore, the volume data 38 is obtained, for example, by experiment.

[0034] The dataset 12 including the above foreign substance data 14 and volume data 38 is provided to the machine learning device 100 by, for example, a user of the machine learning device 100. The dataset acquisition unit 108 may acquire multiple datasets 12. The dataset acquisition unit 108 may store the acquired datasets 12 in the storage unit 104.

[0035] The teacher data selection unit 110 selects teacher data 40 to be used for machine learning from among the multiple datasets 12 acquired by the dataset acquisition unit 108. The teacher data selection unit 110 may select multiple pieces of teacher data 40 from among the multiple datasets 12. The criteria for selecting the teacher data 40 by the teacher data selection unit 110 may be determined in advance as appropriate. The teacher data selection unit 110 selects the teacher data 40 based on, for example, the conditions (condition 1 and condition 2) described below. The teacher data selection unit 110 may select a dataset 12 that satisfies at least one of condition 1 and condition 2 as the teacher data 40.

[0036] (Condition 1) The diameter ratio indicated by the ratio data 36 is within a predetermined range.

[0037] FIG. 3 is a diagram showing the frequency distribution of the diameter ratio.

[0038] In the following description, the above-mentioned range according to Condition 1 is also referred to as the allowable ratio range RR. The allowable ratio range RR may be determined in advance as appropriate, or may be determined based on statistics of multiple data sets 12 acquired by the data set acquisition unit 108. More specifically, a distribution (frequency distribution) of the occurrence frequency (occurrence probability) of diameter ratios may be derived based on the multiple data sets 12. A sample group GR of diameter ratios may be extracted based on this distribution. This sample group GR is composed of diameter ratios whose occurrence probability is equal to or greater than a predetermined probability. The sample group GR includes the most frequent value Rmode of the diameter ratios. The allowable ratio range RR may be determined based on the lower limit Rmin of the diameter ratios and the upper limit Rmax of the diameter ratios in this sample group GR. The predetermined probability may be determined as appropriate, for example, 99.7%. The diameter ratio distribution may be derived including the ellipticity of a virtual circle by defining the ellipticity of the virtual circle. The ellipticity of the virtual circle may be defined as 1.00 (1.000...), for example.

[0039] (Condition 2) The area indicated by the area data 16 is within a predetermined range.

[0040] FIG. 4 is a diagram showing the frequency distribution of the area.

[0041] In the following description, the above-mentioned range according to condition 2 will also be referred to as the allowable area range RA. The lower limit Amin of the allowable area range RA may be determined in advance as appropriate, or may be determined based on the lower limit of the resolution of the inspection apparatus 200 included in the inspection system SYS. The upper limit Amax of the allowable area range RA may be determined in advance as appropriate, or may be determined based on statistics of multiple data sets 12 acquired by the data set acquisition unit 108. For example, the upper limit Amax of the allowable area range RA may be determined based on the maximum value of the area of ​​the foreign matter 42 identified based on the multiple data sets 12. In this case, the maximum value of the area of ​​the foreign matter 42 may be the maximum value of the area of ​​the foreign matter 42 within the sample group GR described above.

[0042] The learning model generation unit 112 generates the learning model 10 by performing supervised learning using the teacher data 40. The learning model generation unit 112 may perform supervised learning using multiple pieces of teacher data 40 selected by the teacher data selection unit 110. The learning model generation unit 112 performs machine learning with the goal of obtaining a learning model 10 that represents the relationship between the foreign object data 14 included in the teacher data 40 and the volume data 38 included in the teacher data 40. In this case, the foreign object data 14 may be used as input data indicating features for learning. Furthermore, the volume data 38 may be used as a label corresponding to the input data. In this way, the learning model generation unit 112 generates the learning model 10 that receives the foreign object data 14 as input and outputs volume estimation data (the volume of the foreign object 42).

[0043] The learning model generation unit 112 stores the generated learning model (trained model) 10 in the above-mentioned storage unit 104. The learning model 10, which is a trained model stored in the storage unit 104, is used by the inspection device 200 for quality inspection of membrane electrode assemblies.

[0044] FIG. 5 is a flowchart of a machine learning method according to one embodiment.

[0045] The above-described machine learning device 100 can execute the machine learning method shown in Fig. 5. This machine learning method can be realized by, for example, having a program stored in a storage unit 104 (memory) of the machine learning device 100, which is a computer, executed by a calculation unit 106 (processor) of the machine learning device 100. As shown in Fig. 5, the machine learning method includes a dataset acquisition step (acquisition step) S11, a training data selection step S12, and a learning model generation step S13.

[0046] In the dataset acquisition step S11, the dataset acquisition unit 108 acquires a dataset 12. The dataset 12 includes foreign substance data 14 for learning and volume data 38 corresponding to the foreign substance data 14. The dataset acquisition unit 108 can acquire a plurality of datasets 12.

[0047] In the teacher data selection step S12, the teacher data selection unit 110 selects a dataset 12 to be used as teacher data 40 from among a plurality of datasets 12. The teacher data selection unit 110 can select a plurality of datasets 12 (a plurality of teacher data 40).

[0048] In the learning model generation step S13, the learning model generation unit 112 performs machine learning using the plurality of training data 40 selected in the training data selection step S12. As a result, the learning model generation unit 112 generates a learning model 10. The generated learning model (trained model) 10 is stored in the storage unit 104.

[0049] The inspection device 200 (FIG. 1) is a device that inspects the quality of a membrane electrode assembly using a learning model 10 generated by a machine learning device 100. The inspection device 200 is, for example, an electronic device such as a computer. The inspection device 200 may be a server device. The inspection device 200 includes an operation unit 202, a memory unit 204, and a calculation unit 206.

[0050] The operation unit 202 includes an input device (not shown), which includes a pointing device, a keyboard, etc. The pointing device includes, but is not limited to, a mouse, a trackpad, etc.

[0051] The storage unit 204 includes one or more memories. The one or more memories include non-volatile memories. Non-volatile memories are recording media that non-temporarily store programs, tables, maps, etc. For example, ROM, flash memory, etc. are included in non-volatile memories. The storage unit 204 (one or more memories) may further include volatile memories. For example, RAM is included in volatile memories. Furthermore, at least a part of the storage unit 204 may be realized by recording media such as USB memory, memory cards, optical discs, etc.

[0052] The calculation unit 206 includes a predetermined processing circuit (not shown). The processing circuit includes, for example, one or more processors such as a CPU, a GPU, etc. The processing circuit may include a predetermined integrated circuit such as an ASIC, an FPGA, etc.

[0053] The calculation unit 206 includes a foreign substance data acquisition unit 208, an estimation unit 210, and an inspection unit 212. The foreign substance data acquisition unit 208, the estimation unit 210, and the inspection unit 212 are realized by the calculation unit 206 (processor) executing a program stored in the storage unit 204 (memory). If the calculation unit 206 includes an integrated circuit such as an ASIC or FPGA, at least a part of the foreign substance data acquisition unit 208, the estimation unit 210, and the inspection unit 212 may be realized by the integrated circuit.

[0054] The foreign substance data acquisition unit 208 acquires foreign substance data 14 identified from an image of the foreign substance 42. The image includes, for example, an X-ray transmission image of the foreign substance 42. The foreign substance data 14 may be provided to the inspection device 200 by a user of the inspection device 200 via the operation unit 202, but is not limited to this. The foreign substance data acquisition unit 208 may execute a predetermined image analysis process when an image (image data) of the foreign substance 42 is provided to the inspection device 200. The predetermined image analysis process may include, for example, image recognition. In this way, the foreign substance data acquisition unit 208 may acquire (generate) the foreign substance data 14 identified from the image by executing the predetermined image analysis process.

[0055] The estimation unit 210 estimates the volume of the foreign substance 42 using the foreign substance data 14 acquired by the foreign substance data acquisition unit 208 and the learning model 10 generated by the machine learning device 100. That is, the estimation unit 210 estimates the volume of the foreign substance 42 based on volume estimation data output from the learning model 10 in accordance with the foreign substance data 14. The estimation unit 210 may estimate the volume of the foreign substance 42 using the learning model 10 stored in advance in the storage unit 204, but is not limited to this. The estimation unit 210 may access the learning model 10 stored in a memory external to the inspection device 200 (e.g., the storage unit 104) to estimate the volume of the foreign substance 42. In this case, the inspection device 200 may be provided with a communication module (not shown) that enables access to the learning model 10 stored in a memory external to the inspection device 200.

[0056] The inspection unit 212 performs a quality inspection of the membrane electrode assembly based on the result of estimation by the estimation unit 210. For example, if the volume of the foreign matter 42 estimated by the estimation unit 210 (learning model 10) is equal to or greater than a predetermined volume threshold, the inspection unit 212 may determine that the membrane electrode assembly containing the foreign matter 42 is defective. The inspection unit 212 may control a notification device (not shown) to notify the user of the result of the quality inspection. For example, if the membrane electrode assembly is defective, the inspection unit 212 may control an alarm (not shown) to notify the user that the membrane electrode assembly is defective.

[0057] FIG. 6 is a flowchart of an inspection method according to one embodiment.

[0058] The inspection device 200 described above can execute the inspection method shown in Fig. 6. This inspection method can be realized by, for example, having a program stored in a storage unit 204 (memory) of the inspection device 200, which is a computer, executed by a calculation unit 206 (processor) of the inspection device 200. As shown in Fig. 6, the inspection method includes a foreign matter data acquisition step S21, an estimation step S22, and an inspection step S23.

[0059] In the foreign substance data acquisition step S21, the foreign substance data acquisition unit 208 acquires the foreign substance data 14.

[0060] In the estimation step S22, the estimation unit 210 estimates the volume of the foreign matter 42 mixed in the membrane electrode assembly using the foreign matter data 14 acquired in the foreign matter data acquisition step S21 and the learning model 10 generated by the machine learning device 100 (the machine learning method of Figure 5).

[0061] In the inspection step (determination step) S23, the inspection unit 212 performs a quality inspection of the membrane electrode assembly. The quality inspection includes a process in which the inspection unit 212 determines the quality of the membrane electrode assembly based on the result of estimation in the estimation step S22. Therefore, in the inspection step S23, the quality can be determined based on the volume of the foreign matter 42.

[0062] According to this embodiment, the machine learning device 100 (machine learning method) and the inspection device 200 (inspection method) provide the following advantageous effects, for example.

[0063] The machine learning device 100 includes a dataset acquisition unit 108 and a learning model generation unit 112. The dataset acquisition unit 108 acquires training data 40. The learning model generation unit 112 generates a learning model 10 using the training data 40. The training data 40 includes foreign matter data 14 and data indicating the volume of a foreign matter 42. The foreign matter data 14 is data identified from an image of a foreign matter 42 mixed in a membrane electrode assembly. The foreign matter data 14 includes at least one of area data 16 of the foreign matter 42, brightness data 18 of the foreign matter 42, and shape data 20 of the foreign matter 42. The learning model 10 receives the foreign matter data 14 as an input and outputs the volume (estimated value) of the foreign matter 42. This allows the machine learning device 10 to provide a user with a learning model 10 that estimates the volume of the foreign matter 42 based on data identified from the image of the foreign matter 42. A user can effectively perform quality inspection of a membrane electrode assembly by using the learning model 10. The image is, for example, an X-ray transmission image.

[0064] The shape data 20 includes at least one of circumference data 24, skewness data 26, kurtosis data 28, and diameter data 30. The learning model generation unit 112 performs machine learning using the shape data 20 to generate a learning model 10 that can accurately estimate the volume of a foreign object 42 according to the shape of the foreign object 42. It is preferable that the shape data 20 include all of the circumference data 24, skewness data 26, kurtosis data 28, and diameter data 30, but is not limited to this.

[0065] The diameter data 30 includes ratio data 36. The ratio data 36 indicates, for example, the ratio between a first ferret diameter F1 and a second ferret diameter F2. The ratio data 36 may indicate the ellipticity of a virtual ellipse VE (virtual circle). The learning model generation unit 112 performs machine learning using the ratio data 36 to generate a learning model 10 that can accurately estimate the volume of a foreign object 42 according to the shape of the foreign object 42. The ellipticity of the virtual circle is defined as, for example, 1.00.

[0066] The machine learning device 100 further includes a teacher data selection unit 110. The teacher data selection unit 110 selects teacher data 40 to be used in supervised learning by the learning model generation unit 112 from among a plurality of data sets 12. The teacher data selection unit 110 selects the teacher data 40 based on, for example, at least one of an allowable ratio range RR and an allowable area range RA. In this way, the teacher data selection unit 110 prevents the learning model generation unit 112 from performing supervised learning using a data set 12 that includes outliers. As a result, the learning model generation unit 112 can generate a learning model 10 that can accurately estimate the volume of a foreign object 42 according to the shape of the foreign object 42.

[0067] The inspection device 200 includes a foreign matter data acquisition unit 208 and an estimation unit 210. The foreign matter data acquisition unit 208 acquires foreign matter data 14. The estimation unit 210 estimates the volume of the foreign matter 42 using the foreign matter data 14 and the learning model 10. This allows the inspection device 200 to accurately determine the quality of the membrane electrode assembly based on the estimated volume.

[0068] The embodiment may be modified as in the modified examples described below. Note that descriptions that overlap with the embodiment will be omitted as appropriate. Furthermore, among the configurations described below, the same configurations as those described in the embodiment will be assigned the same reference numerals as those in the embodiment.

[0069] (Variation 1) The inspection system SYS may further include a display device (not shown). The display device may include, for example, a display panel or display element that forms a display screen. The display device may appropriately display information stored in the memory unit 104 of the machine learning device 100 or the memory unit 204 of the inspection device 200. For example, the display device included in the inspection device 200 may display the results of a quality inspection performed by the inspection unit 212. In other words, the display device included in the inspection device 200 may function as the notification device described above. Note that at least one of the machine learning device 100 and the inspection device 200 may include a display device.

[0070] (Variation 2) A single electronic device (computer) may implement both the machine learning device 100 and the inspection device 200. In this case, the machine learning device 100 and the inspection device 200 may share hardware resources such as a processor and memory.

[0071] (Variation 3) The inspection system SYS may further include a terminal device (one or more terminal devices) not shown. The inspection system SYS may include multiple terminal devices. The terminal device is realized by an electronic device such as a PC (Personal Computer). The terminal device may be a portable smart device (such as a tablet terminal). The terminal device may remotely operate at least one of the machine learning device 100 and the inspection device 200 via a network. In this case, the network may include the Internet.

[0072] (Combination of multiple modifications) The above-described multiple modifications may be combined as appropriate within a range that does not contradict each other.

[0073] According to the above-described embodiment and modification, the machine learning device 100, the inspection device 200, the machine learning method, and the program for better inspecting membrane electrode assemblies are provided.

[0074] The following additional notes are further disclosed regarding the above embodiment.

[0075] (Appendix 1) The machine learning device according to the present disclosure is a machine learning device (100) including: an acquisition unit (108) that acquires supervised learning using the supervised learning data acquired by the acquisition unit to generate a learning model (10) that uses the supervised learning data acquired by the acquisition unit to input the foreign object data and output the volume of the foreign object. This configuration allows the generation of a learning model that can accurately estimate the volume of the foreign object. Using this learning model, the volume of the foreign object can be accurately estimated, thereby enabling successful inspection of the membrane electrode assembly.

[0076] (Appendix 2) The machine learning device described in Appendix 1 may be a machine learning device, wherein the shape data includes at least one of perimeter data (24) relating to the perimeter of the foreign object, skewness data (26) relating to the skewness of the foreign object, kurtosis data (28) relating to the kurtosis of the foreign object, and diameter data (30) relating to the diameter of a virtual circle approximating the foreign object, the diameter of a virtual ellipse (VE) approximating the foreign object, or the ferret diameter of the foreign object.

[0077] (Appendix 3) The machine learning device described in Appendix 2 may be a machine learning device, wherein the diameter data includes ratio data (36) indicating the ratio between a first ferret diameter (F1) that is the ferret diameter in a first direction (D1) and a second ferret diameter (F2) that is the ferret diameter in a second direction (D2) that intersects with the first direction, or the ellipticity of the virtual ellipse.

[0078] (Appendix 4) In the machine learning device described in Supplementary Note 3, the learning model generation unit may be a machine learning device that performs the supervised learning using the foreign substance data in which the ratio or the ellipticity is within a predetermined range.

[0079] (Appendix 5) In the machine learning device according to Supplementary Note 1, the learning model generation unit may perform the supervised learning using the foreign object data whose area is within a predetermined range.

[0080] (Appendix 6) The machine learning device according to Supplementary Note 1 may be a machine learning device in which the image is an X-ray transmission image.

[0081] (Appendix 7) The inspection device according to the present disclosure is an inspection device (200) including: a foreign matter data acquisition unit (208) that acquires foreign matter data (14) including at least one of area data (16) indicating the area of ​​a foreign matter (42) contaminated in a membrane electrode assembly ascertained from an image of the foreign matter, brightness data (18) indicating the brightness of the foreign matter ascertained from the image, and shape data (20) relating to the shape of the foreign matter ascertained from the image; and an estimation unit (210) that estimates the volume of the foreign matter using the foreign matter data acquired by the foreign matter data acquisition unit and the learning model generated by the machine learning device described in any one of Supplementary Notes 1 to 5. This configuration allows the volume of the foreign matter to be accurately estimated, thereby enabling successful inspection of the membrane electrode assembly.

[0082] (Appendix 8) The machine learning method according to the present disclosure is a computer-executed machine learning method, comprising: an acquisition step (S11) of acquiring foreign matter data (14) including at least one of area data (16) indicating the area of ​​a foreign matter (42) contaminated in a membrane electrode assembly as determined from an image of the foreign matter, brightness data (18) indicating the brightness of the foreign matter as determined from the image, and shape data (20) relating to the shape of the foreign matter as determined from the image; and training data (40) including data indicating the volume of the foreign matter; and a training model generation step (S13) of performing supervised learning using the training data acquired in the acquisition step to generate a training model (10) that uses the foreign matter data as an input and outputs the volume of the foreign matter.

[0083] (Appendix 9) The program according to the present disclosure is a program for causing the computer to execute the machine learning method described in Appendix 8.

[0084] The present disclosure is not limited to the above disclosure, and various configurations may be adopted without departing from the gist of the present disclosure. [Explanation of symbols]

[0085] 10...Learning Model 14...Foreign matter data 16...Area data 18...Brightness data 20...Shape data 24...Circumference data 26...Skewness data 28...Kurtosis data 30...Diameter data 36...Ratio data 40...Teacher data 42...Foreign object 100...Machine learning device 108...Dataset acquisition unit (acquisition unit) 112...Learning model generation unit 200...Inspection equipment 208...Foreign matter data acquisition unit 210...Estimation part F1...first ferret diameter F2: Second ferret diameter VE...Virtual ellipse

Claims

1. an acquisition unit that acquires foreign matter data including at least one of area data indicating an area of ​​a foreign matter mixed in a membrane electrode assembly ascertained from an image of the foreign matter, brightness data indicating a brightness of the foreign matter ascertained from the image, and shape data relating to a shape of the foreign matter ascertained from the image, and teacher data including data indicating a volume of the foreign matter; a learning model generation unit that performs supervised learning using the training data acquired by the acquisition unit to generate a learning model that receives the foreign object data as an input and outputs the volume of the foreign object; A machine learning device comprising:

2. The machine learning device according to claim 1 , A machine learning device, wherein the shape data includes at least one of perimeter data relating to the perimeter of the foreign object, skewness data relating to the skewness of the foreign object, kurtosis data relating to the kurtosis of the foreign object, and diameter data relating to the diameter of a virtual circle approximating the foreign object, the diameter of a virtual ellipse approximating the foreign object, or the ferret diameter of the foreign object.

3. The machine learning device according to claim 2, The diameter data includes ratio data indicating a ratio between a first ferret diameter that is the ferret diameter in a first direction and a second ferret diameter that is the ferret diameter in a second direction intersecting the first direction, or ratio data indicating the ellipticity of the virtual ellipse.

4. The machine learning device according to claim 3, The learning model generation unit performs the supervised learning using the foreign substance data in which the ratio or the ellipticity is within a predetermined range.

5. The machine learning device according to claim 1 , The learning model generation unit performs the supervised learning using the foreign object data whose area is within a predetermined range.

6. The machine learning device according to claim 1 , The machine learning device, wherein the image is an X-ray transmission image.

7. a foreign matter data acquiring unit that acquires foreign matter data including at least one of area data indicating an area of ​​a foreign matter contaminated in a membrane electrode assembly ascertained from an image of the foreign matter, brightness data indicating a brightness of the foreign matter ascertained from the image, and shape data relating to a shape of the foreign matter ascertained from the image; an estimation unit that estimates the volume of the foreign matter using the foreign matter data acquired by the foreign matter data acquisition unit and the learning model generated by the machine learning device according to any one of claims 1 to 6; An inspection device comprising:

8. an acquiring step of acquiring training data including foreign matter data including at least one of area data indicating the area of ​​the foreign matter contaminated in the membrane electrode assembly ascertained from an image of the foreign matter, brightness data indicating the brightness of the foreign matter ascertained from the image, and shape data relating to the shape of the foreign matter ascertained from the image, and data indicating the volume of the foreign matter; a learning model generation step of performing supervised learning using the training data acquired in the acquisition step to generate a learning model in which the foreign object data is used as an input and the volume of the foreign object is used as an output; 1. A machine learning method comprising:

9. A program for causing a computer to execute the machine learning method according to claim 8.

Images