Calculation method, product manufacturing method, product management method, calculation device, product manufacturing apparatus, measurement method, measurement system, measurement device, teach data creation method, teach data, model creation method, program, and storage medium

Abstract

A calculation method for manufacturing or using a product is provided with: a step (S2) for calculating a feature quantity using one or more input values selected from a predetermined input value group and one or more first models; and a step (S3) in which, using one or more input values selected from the group of input values and one or more second models, the amount of separation, which is the amount of shift from the first model, with respect to an input value predetermined from the group of input values is calculated, and the amount of separation is calculated using the one or more input values selected from the group of input values and the one or more second models. The second model and the first model are machine learning models generated using one or more pieces of teacher data selected from predetermined teacher data sets, respectively.

Description

Technical Field

[0001] The present disclosure relates to an operation method, a product manufacturing method, a product management method, an operation device, a product manufacturing equipment, a measurement method, a measurement system, a measurement device, a teacher data production method, teacher data, a model generation method, a program and a storage medium. Background Art

[0002] Conventionally, attempts have been made to measure or evaluate mechanical properties or states of a substance that cannot be directly determined via physical quantities that are correlated with the mechanical properties or states of the substance. For example, models using machine learning are sometimes used in such measurements or evaluations.

[0003] For example, in the steel industry, steel products need to meet the specifications of mechanical properties required by users. Therefore, in the manufacturing process of steel products, sampling inspection is carried out to determine whether the required mechanical properties are met. Sampling inspection is a so-called destructive test in which the inspection part is cut out from the steel product, processed into a test piece for investigating the mechanical properties, and then tested. However, in recent years, it has been required to directly measure or evaluate the mechanical properties of the steel product itself in a non-destructive manner rather than sampling inspection to ensure quality.

[0004] Non-patent document 1 describes a method for evaluating the hardness, tensile strength, or yield strength of a steel product after quenching using electromagnetic measurement as a non-destructive method and machine learning or multivariate regression. In addition, in non-patent document 1, a model for evaluating mechanical properties is constructed based on regression analysis of a steel product and multiple different electromagnetic quantities and mechanical property values ​​measured by an eddy current sensor or a magnetic sensor, and based on the model, mechanical property values ​​are evaluated based on multiple different electromagnetic quantities measured by the sensor.

[0005] In addition, not limited to steel materials, the method of evaluating the mechanical properties of materials through machine learning is used in various fields. For example, in Patent Document 1, the thermal conductivity of semiconductor crystal products is calculated based on the regression model generated in advance through machine learning and the surface temperature distribution and heating conditions of the measured material. For example, in Patent Document 2, the state of the road surface on which the vehicle is traveling is determined based on a naive Bayesian inference device or a neural network as machine learning and a set of multi-point reflection intensity values ​​obtained by laser irradiation. For example, in Patent Document 3, machine learning and the surface strain state obtained from the surface strain image are used to evaluate the deterioration state of the concrete surface.

[0006] Patent Document 1: International Publication No. 2020 / 110796

[0007] Patent Document 2: Japanese Patent Application Publication No. 2014-228300

[0008] Patent Document 3: Japanese Patent Application Publication No. 2021-018233

[0009] Non-patent literature 1: Wolter Bernd et al., "Nondestructive Testing with3MA-AnOverview of Principles and Applications" Applied Sciences 2019, 9(6), 1068

[0010] In the methods described in Non-Patent Document 1 and Patent Documents 1 to 3, the mechanical property values ​​of the object are evaluated based on the information measured by the sensor and according to the model constructed in advance by regression analysis or machine learning. However, in these methods, when there is no machine learning model suitable for the input value in the constructed model group, the output value becomes a value that greatly deviates from the actual mechanical property value, and there is a problem of reduced reliability for quality assurance of steel materials.

[0011] For example, in non-patent document 1, electromagnetic measurement is used to obtain electromagnetic quantities, but the interference effects of the metal structure, oxide scale, etc. on the surface of the steel plate are not considered, so there is a situation where the electromagnetic quantity deviates from the machine learning model due to the interference, and the output is a predicted value that deviates greatly from the actual hardness value. In addition, in patent document 1, the material of the evaluation object sample is not considered when the model is generated. Therefore, when evaluating a sample of a material different from the sample used when the machine learning model was generated, there is a possibility of outputting a deviated predicted value. In patent document 2, when measuring a road surface composed of an unlearned material, especially a material with a reflectivity different from the teacher data, it is impossible to accurately predict the road surface state. Patent document 3 also has the same problems as patent document 2.

[0012] As described above, methods for evaluating the mechanical properties of materials by machine learning are used in various fields, but on the other hand, there is no research on whether the machine learning model used is suitable for the input value. Here, the machine learning model is suitable for the input value, which means that when the predicted value of the feature quantity is obtained using the machine learning model for a certain input value, the predicted value of the feature quantity that deviates greatly from the actual value is not output. Summary of the invention

[0013] The present disclosure is completed in consideration of the above points. The purpose of the present disclosure is to provide a method for calculating whether a machine learning model applied to input values ​​pre-selected from the same input value group is appropriate in parallel with the output of a feature quantity calculated by a machine learning model based on one or more input values ​​selected from a predetermined input value group, a method for manufacturing a product, a method for managing a product, a computing device, a manufacturing device for a product, a measuring method, a measuring system, a measuring device, a method for preparing teacher data, teacher data, a method for generating a model, a program, and a storage medium.

[0014] (1) An operation method according to one embodiment of the present disclosure is an operation method for manufacturing or using a product, comprising:

[0015] a feature quantity calculation step in which the feature quantity is calculated using one or more input values ​​selected from a predetermined input value group and one or more first models; and

[0016] a deviation amount calculation step, in which the deviation amount for an input value predetermined from the input value group is calculated using one or more input values ​​selected from the input value group and one or more second models, the deviation amount being a deviation from the first model,

[0017] The second model and the first model are machine learning models generated using one or more teacher data selected from a predetermined teacher data group.

[0018] (2) As one embodiment of the present disclosure, based on (1),

[0019] A determination step is provided in which the separation amount is compared with a preset threshold value to determine whether the first model to be used is acceptable or not or the level of the separation amount.

[0020] (3) A method for manufacturing a product according to one embodiment of the present disclosure includes a manufacturing step of manufacturing a product.

[0021] In the above manufacturing steps, the calculation method (1) or (2) is performed.

[0022] (4) A product management method according to one embodiment of the present disclosure comprises:

[0023] a manufacturing step in which a product is manufactured and the operation method of (1) or (2) is performed; and

[0024] A management step, in which the products are classified based on at least one of the feature amount and the amount of separation.

[0025] (5) A computing device according to one embodiment of the present disclosure is a computing device for manufacturing or using a product, comprising:

[0026] a feature quantity calculation unit that calculates the feature quantity using one or more input values ​​selected from a predetermined input value group and one or more first models; and

[0027] a deviation amount calculation unit, wherein the deviation amount calculation unit calculates a deviation amount for an input value predetermined from the input value group using one or more input values ​​selected from the input value group and one or more second models, wherein the deviation amount is a deviation amount from the first model,

[0028] The second model and the first model are machine learning models generated using one or more teacher data selected from a predetermined teacher data group.

[0029] (6) As one embodiment of the present disclosure, based on (5),

[0030] A determination unit is provided for comparing the deviation amount with a preset threshold value to determine whether the first model to be used is acceptable or not or the level of the deviation amount.

[0031] (7) As one embodiment of the present disclosure, based on (5) or (6),

[0032] A data acquisition unit is provided for acquiring one or two of the input value of the first model and the signal measured to obtain the input value of the first model.

[0033] (8) A manufacturing facility for a product according to one embodiment of the present disclosure includes:

[0034] Manufacturing equipment where products are manufactured; and

[0035] The computing device according to any one of (5) to (7).

[0036] (9) A measurement method according to one embodiment of the present disclosure includes a measurement step of measuring a signal related to a physical quantity of a measurement object.

[0037] Execute the operation method of (1) or (2), in which the physical quantity obtained based on the measured signal is used as the one or more input values, and the characteristic quantity calculated in the characteristic quantity calculation step is used as the measurement value.

[0038] (10) A method for manufacturing a product according to one embodiment of the present disclosure includes a manufacturing step of manufacturing a product.

[0039] The above-mentioned manufacturing steps include the measuring method of (9).

[0040] (11) A product management method according to one embodiment of the present disclosure comprises:

[0041] a manufacturing step in which a product is manufactured and includes the measuring method of (9); and

[0042] A management step, in which the products are classified based on at least one of the characteristic amount and the deviation amount of the manufactured products.

[0043] (12) A measurement system according to one embodiment of the present disclosure includes:

[0044] a physical quantity measuring unit that measures a signal related to a physical quantity of a measurement object; and

[0045] The computing device of (5) or (6), wherein the computing device uses the physical quantity obtained based on the measured signal as the one or more input values, and uses the feature quantity calculated by the feature quantity calculation unit as the measurement value.

[0046] (13) A manufacturing facility for a product according to one embodiment of the present disclosure includes:

[0047] Manufacturing equipment where products are manufactured; and

[0048] (12) Measurement system.

[0049] (14) A measuring device according to one embodiment of the present disclosure includes:

[0050] Data output unit;

[0051] a physical quantity measuring unit that measures a signal related to a physical quantity of a measurement object; and

[0052] a physical quantity acquisition unit control unit, the physical quantity acquisition unit control unit executing a process of outputting the signal measured to obtain one or more input values ​​to an external computing device,

[0053] The computing device comprises:

[0054] a feature quantity calculation unit that calculates a feature quantity using one or more input values ​​and one or more first models; and

[0055] A deviation amount calculation unit calculates a deviation amount for the input value using one or more input values ​​selected from the input values ​​and one or more second models, wherein the deviation amount is a deviation amount from the first model.

[0056] (15) As an embodiment of the present disclosure, based on (14),

[0057] A physical quantity calculation unit is provided, wherein the physical quantity calculation unit calculates the physical quantity as the measured physical quantity based on the measured signal,

[0058] The control unit for the physical quantity acquisition unit executes a process of outputting the measured physical quantity as the one or more input values ​​instead of the signal measured in order to obtain the one or more input values.

[0059] (16) As one embodiment of the present disclosure, based on (14) or (15),

[0060] The physical quantity acquisition unit control unit executes a process of acquiring at least one of the information related to the calculated characteristic quantity and the information related to the calculated deviation amount from the calculation device.

[0061] (17) A method for preparing teacher data in one embodiment of the present disclosure selects as performance input values ​​a predetermined number or more of input values ​​of the level of the above-mentioned deviation amount that are judged as unqualified or worse than a predetermined standard based on the pass / fail of the first model or the level of the deviation amount calculated by the calculation method of (2),

[0062] The selected actual input value is assigned a feature value actually measured as an actual output value and a label indicating the level of the deviation amount indicating failure or being worse than a predetermined standard, thereby forming teacher data for generating the first model.

[0063] (18) The teacher data of one embodiment of the present disclosure is used to generate a first model, wherein:

[0064] a predetermined number or more of actual input values ​​of the first model;

[0065] For each of the above-mentioned performance input values, a characteristic value actually measured as a performance output value and a label indicating whether the first model is qualified or not or the level of deviation is calculated by the operation method of (2), and the level of the above-mentioned deviation is known to be unqualified or worse than a predetermined benchmark.

[0066] (19) A method for generating a model according to an embodiment of the present invention uses the teacher data of (18) to generate a learned model through machine learning, wherein the learned model takes one or more input values ​​selected from a predetermined input value group as input values ​​and takes a feature value having a correlation that can be derived based on the input values ​​as output values.

[0067] (20) A program according to one embodiment of the present disclosure is a program for a computing device used in any one of the computing devices of (5) to (7), wherein:

[0068] The computer is caused to execute the functions of the feature amount calculation unit and the deviation amount calculation unit.

[0069] (21) A storage medium according to one embodiment of the present disclosure stores a program for use in any one of the computing devices of (5) to (7), wherein:

[0070] A program for a computing device is stored which causes a computer to execute the functions of the feature amount calculation unit and the deviation amount calculation unit.

[0071] (22) A program according to one embodiment of the present disclosure is used for any one of the measuring devices (14) to (16), wherein:

[0072] Make the computer execute the functions of the above-mentioned computing device and data output unit.

[0073] (23) A storage medium according to one embodiment of the present disclosure stores a program for use in any one of the measuring devices of (14) to (16), wherein:

[0074] A program for causing a computer to execute the functions of the arithmetic device and the data output unit is stored.

[0075] According to an operation method, a product manufacturing method, a product management method, an operation device, a product manufacturing equipment, a measurement method, a measurement system, a measurement device, a teacher data production method, teacher data, a model generation method, a program and a storage medium of one embodiment of the present disclosure, it is possible to calculate whether the applied machine learning model is suitable for input values ​​predetermined from the input value group in parallel with the calculation based on one or more input values ​​selected from the predetermined input value group and the characteristic quantity of the machine learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a block diagram showing a configuration example of a measurement system including a computing device according to an embodiment of the present disclosure.

[0077] Figure 2 It is the overall processing flow chart.

[0078] Figure 3 This is a block diagram of the physical quantity measurement unit.

[0079] Figure 4 It is a diagram showing a specific configuration example of a sensor.

[0080] Figure 5This is a diagram showing an example of a signal applied to an exciting coil in order to generate an AC magnetic field.

[0081] Figure 6 is a flowchart for generating the first model.

[0082] Figure 7 It is a flowchart for calculating feature quantities.

[0083] Figure 8 is a flowchart for generating the second model.

[0084] Fig. 9 is a flow chart for calculating the amount of separation.

[0085] Fig.10 It is a flowchart for explaining the determination process.

[0086] Fig.11A It is a figure which shows an example of color mapping.

[0087] Fig. 11B It is a figure which shows an example of color mapping.

[0088] Fig. 11C It is a figure which shows an example of color mapping.

[0089] Fig.12 It is a figure which shows an example of a graph.

[0090] Fig.13 This is a block diagram showing another configuration example of the measurement system.

[0091] Fig.14 This is a block diagram showing another configuration example of the measurement system.

[0092] Fig.15 This is a block diagram showing another configuration example of the measurement system.

[0093] Fig.16A This is a diagram illustrating the relationship between input values ​​and feature quantities of a measurement object.

[0094] Fig. 16B This is a diagram illustrating the relationship between input values ​​and feature quantities when input values ​​are combined.

[0095] Fig.17 This is a flowchart for determining the level of the amount of separation.

[0096] Fig.18 This is a diagram showing an example of a portion where a difference of 20 HV or more occurs in a normal plate.

[0097] Fig.19 This is a diagram showing an example of a portion where a difference of 20 HV or more occurs in an abnormal plate.

[0098] Fig. 20 This is a block diagram showing another configuration example of the measurement system. DETAILED DESCRIPTION

[0099] Hereinafter, with reference to the accompanying drawings, the calculation method, the manufacturing method of the product, the management method of the product, the calculation device, the manufacturing equipment of the product, the measurement method, the measurement system, the measurement device, the method for preparing teacher data, the teacher data, the method for generating the model, the program and the storage medium of the embodiment of the present disclosure are described. In each figure, the same figure mark is marked for the same or corresponding parts. The description of the same or corresponding parts is appropriately omitted or simplified.

[0100] [Implementation Method 1]

[0101] (System Structure)

[0102] Figure 1 1 is a block diagram showing a structural example of a measurement system 20 including a computing device 12 of an embodiment of the present disclosure. The computing device 12 of the present embodiment includes a display unit 1, a control unit 3, and a storage unit 9. In addition, in order to constitute the measurement system 20, a physical quantity acquisition unit 2 and a scanning unit 25 are provided. The present embodiment is suitable for a case where a device having a control unit 3 is moved manually like a trolley type and the entire plate is measured. The details will be described later, but the sensor 14, the power supply, and the computing device 12 are moved by manpower through the trolley (scanning unit 25), and the sensor 14 receives a signal, and the control unit 3 of the computing device 12 acquires the physical quantity, infers the characteristic quantity, and infers the separation quantity based on the reception result.

[0103] The display unit 1 displays information appropriately selected by the control unit 3 from among the information related to the physical quantity, characteristic quantity, and separation quantity calculated by the control unit 3. In the present embodiment, the display unit 1 displays at least one of the graphs and maps described later generated by the control unit 3. The display unit 1 can be composed of a display capable of displaying text, images, etc. The display can be a display device such as a liquid crystal display (LCD: Liquid Crystal Display) or an organic EL display (OELD: Organic Electro-Luminescence Display).

[0104] The physical quantity acquisition unit 2 includes a physical quantity measurement unit 2a that measures a physical quantity of a measurement object. A specific example of the physical quantity measurement unit 2a will be described later.

[0105] Here, the measurement based on the physical quantity measuring unit 2a is performed to measure the characteristic quantity of the product and to evaluate the physical characteristics and quality. A physical quantity is a quantity that can be objectively measured, such as temperature, mass, and electromagnetic quantity. Of course, they can also be combined. A physical quantity is not limited to a specific physical quantity as long as it has a correlation relationship that can derive the characteristic quantity by some method. In this embodiment, due to the measurement system 20, the characteristic quantity becomes a value related to the measurement result for the product. In addition, in order to non-destructively measure the characteristic quantity of the product and use it for the evaluation of the physical characteristics and quality of the product, the measurement based on the physical quantity measuring unit 2a is performed. More specifically, the characteristic quantity is, for example, a mechanical property including hardness, but is not limited to this. Here, "product" of course includes a completed "final product" and also includes a state called "intermediate" or "semi-finished product" obtained in the process of manufacturing the final product. In addition, "substance" is "a kind of product". In this embodiment, as an example of a method for calculating the mechanical properties of a substance (a kind of product) by machine learning, the surface hardness of steel in the steel process is described as a characteristic quantity, but the application example of this method is not limited to this. In addition, in this embodiment, as long as the physical quantity is an electromagnetic quantity, the relationship between the characteristic quantity to be obtained and the physical quantity as the input value can be modeled, functionalized or digitized. Other examples may be temperature information or light information based on laser irradiation, etc.

[0106] The control unit 3 uses the measured physical quantity as one or more input values ​​and one or more first models to calculate the characteristic quantity. In addition, the control unit 3 uses one or more input values ​​selected from the input values ​​and one or more second models to calculate the deviation amount for the input value, and the above-mentioned deviation amount is the deviation amount from the first model. In addition, the control unit 3 compares the deviation amount with a preset threshold value to determine whether the first model used is qualified or not or the level of the deviation amount. The details of the calculation of the characteristic quantity, the calculation and determination of the deviation amount will be described later.

[0107] In this embodiment, the control unit 3 includes: a classification processing unit 4 for the first model, a classification processing unit 5 for the second model, a feature quantity calculation unit 6, a deviation quantity calculation unit 7 and a judgment unit 8. The classification processing unit 4 for the first model and the classification processing unit 5 for the second model perform processing to classify the input value and select the appropriate first model and second model respectively. The feature quantity calculation unit 6 performs processing to calculate the feature quantity. The deviation quantity calculation unit 7 performs processing to calculate the deviation quantity. In addition, the judgment unit 8 performs a qualification or a level judgment of the deviation quantity of the first model applied to the input value.

[0108] In addition, the determination unit 8 generates information related to the feature quantity based on the calculated feature quantity in a manner that can be output to the display unit 1. In addition, the determination unit 8 generates information related to the separation amount based on the calculated separation amount or the executed determination result. In addition, the determination unit 8 outputs the generated information to the display unit 1 in a predetermined format or a selection made as needed.

[0109] In addition, when the measurement system 20 is configured as in the first embodiment, the control unit 3 may also have a function of controlling the physical quantity acquisition unit 2. The control unit 3 functions as a physical quantity calculation unit 29a, which extracts a physical quantity according to a predetermined method based on a signal acquired by the physical quantity measurement unit 2a described later. In addition, the control unit 3 controls the physical quantity measurement unit 2a to generate a signal that can extract the physical quantity, and then acquires it.

[0110] The control unit 3 is configured to include one or more processors. The processor is, for example, a general-purpose processor or a dedicated processor dedicated to a specific process, but is not limited thereto and can be any processor. The control unit 3 not only controls the above-mentioned process, but also controls the overall operation of the computing device 12. The control unit 3 functions as a first model classification processing unit 4, a second model classification processing unit 5, a feature quantity calculation unit 6, a separation quantity calculation unit 7, and a determination unit 8 according to a program read from the storage unit 9 or an accessible storage device. In addition, in the case of configuring the measurement system 20 as in the present embodiment 1, the control unit 3 can also function as a control unit 3 including a physical quantity calculation unit 29a according to a program read from the storage unit 9 or an accessible storage device. In addition, other functions required by the computing device 12 can also be executed.

[0111] The storage unit 9 stores data used in the control unit 3. The storage unit 9 may also store programs. The storage unit 9 includes at least a first model group 10 and a second model group 11. The first model group 10 is a plurality of first models (in Figure 1 In the example, the first model _1 to the first model _n are included, but n may be 1 instead of being 2 or more. In addition, the second model group 11 is a plurality of second models (in Figure 1 In the example, the first model group 10 includes the second model_1 to the second model_k), but k may not be greater than 2 but may be 1. In addition, n, which is the number of the first models, and k, which is the number of the second models, may be the same or different. In addition, the first model group 10 may be supplemented with the first model later. The second model group 11 may be supplemented with the second model later.

[0112] In addition, in the present embodiment, the storage unit 9 can also store one or more of the measured physical quantity group 26, the calculated characteristic quantity group 27, or the calculated separation quantity group 28. The measured physical quantity group 26 represents a group of one or more physical quantities measured by the physical quantity acquisition unit 2 and the control unit 3. In addition, the physical quantities selected from these measured physical quantity groups 26 become one or more input values ​​of the first model and the second model. Next, the calculated characteristic quantity group 27 represents a group of one or more characteristic quantities calculated by the characteristic quantity calculation unit 6. Finally, the calculated separation quantity group 28 represents a group of one or more separation quantities calculated by the separation quantity calculation unit 7.

[0113] The storage unit 9 is one or more memories. The memory is, for example, a semiconductor memory, a magnetic memory, or an optical memory, but is not limited thereto and can be any memory. The storage unit 9 is, for example, built into the computing device 12, but can also be configured to be accessed by the control unit 3 via any interface.

[0114] The scanning unit 25 is a structure that moves the calculation device 12 and the physical quantity acquisition unit 2. Figure 1 In the case of , the calculation device 12 and the physical quantity acquisition unit 2 can be mounted and moved. For example, a vehicle that moves by manpower or power other than manpower can be used.

[0115] In the present embodiment, the computing device 12 is combined with the physical quantity acquisition unit 2 as a measuring device, so that the device can function as a measuring system 20 independently. In addition, by combining the scanning unit 25, the computing device 12 and the physical quantity acquisition unit 2, the device can function as a movable measuring system 20 independently. However, the structure of the computing device 12 is arbitrary. For example, in the present embodiment, the computing device 12 is composed of a device including a processor such as a computer, but it can also be composed of a plurality of computers that can send and receive data through a network connection. As an example, Figure 1 As shown, the physical quantity acquisition unit 2 and the computing device 12 including the control unit 3 and the storage unit 9 can be electrically connected, physically connected, or connected via a network to form a measurement system 20. Figure 1 , described later Figure 13 to Figure 15In the illustrated structure, the computing device 12 may also be configured so that a part of the display unit 1, the control unit 3, and the storage unit 9 are included in one computer, while the other parts are included in different computers. For example, the storage unit 9 may be present in another computer connected to the computer including the control unit 3 through a network. In addition, the processing performed by the computing device 12 may be performed by distributed processing of multiple computers. In this case, the functions of the control unit 3 may be realized by the processors of the multiple computers that cooperate. For example, the processing of the determination performed by the determination unit 8 may be performed by the computer that performs the processing of the calculation performed by the feature quantity calculation unit 6 and the separation quantity calculation unit 7 and other computers.

[0116] In the following description, the connection through the network may be a wired connection, a wireless connection, or a combination of these. In addition, a part of a global information communication network such as the Internet may be used. It is sufficient to select and use it appropriately according to the needs.

[0117] (Overall processing)

[0118] Figure 2 The flowchart of the overall processing performed by the computing device 12 is represented. The overall processing flowchart represents the outline of the processing, and the details of each processing will be described later. The control unit 3 executes an acquisition step (step S1) for acquiring an input value. The control unit 3 executes a characteristic quantity calculation step (step S2) for calculating a characteristic quantity. The control unit 3 executes a detachment quantity calculation step (step S3) for calculating a detachment quantity. Here, the characteristic quantity calculation step and the detachment quantity calculation step can be executed in parallel as in the present embodiment, or can be executed sequentially. In the case of sequential execution, it does not matter which of the characteristic quantity calculation step and the detachment quantity calculation step is executed first. The control unit 3 executes a determination step (step S4) for determining whether the first model is qualified or not or the level of the detachment quantity. The control unit 3 executes a display step (step S5) for outputting appropriately selected information from the information respectively related to the above-mentioned physical quantity, characteristic quantity and detachment quantity and displaying it on the display unit 1. In the present embodiment, the control unit 3 outputs at least one of a graph and a mapping of the characteristic quantity and the detachment quantity superimposed or displayed in parallel to the display unit 1. The following describes the details of the matters associated with these steps respectively.

[0119] (Get the input value of the step)

[0120] The physical quantity measuring unit 2a measures a signal related to the physical quantity, which is the input value obtained by the control unit 3. In this embodiment, the physical quantity is a plurality of types of electromagnetic quantities. In this embodiment, the signal related to the physical quantity is a signal of a change in the electromagnetic quantity obtained from the physical quantity measuring unit 2a. Figure 32 is a block diagram of a physical quantity measuring unit 2a. The physical quantity measuring unit 2a includes a signal transceiver 13 and a sensor 14. The sensor 14 measures a signal related to the physical quantity of the measured object. Then, the control unit 3 calculates an electromagnetic quantity that becomes an input value or an effective input value as a physical quantity based on the measured signal related to the physical quantity.

[0121] When the electromagnetic quantity is set as a physical quantity, the electromagnetic quantity obtained from a BH curve measured by a certain excitation voltage or excitation current is equivalent to a predetermined input value group. The electromagnetic quantity is calculated by performing incremental permeability analysis, high-order harmonic analysis, and eddy current impedance analysis on the measured BH curve. As the types of electromagnetic quantities calculated, for example, the maximum and minimum values ​​of incremental permeability, the maximum and minimum values ​​of impedance, the amplitude of magnetizing voltage, etc. can be cited.

[0122] Here, as an example of a "predetermined input value group", it is preferred that when measuring one or more feature quantities required to output a feature quantity from the first model, a group of feature quantities obtained for each measurement is used. The feature quantities contained in the group of feature quantities can be feature quantities that recognize interaction with the output feature quantities. For example, in the above example, in order to output a hardness in the first model, a group of electromagnetic quantities obtained from a BH curve is required, but the group of electromagnetic quantities obtained from the one BH curve is equivalent to a "predetermined input value group". In this case, a BH curve is measured at a prescribed time (which is referred to as the same time). In the case of using the first model to output a feature quantity based on multiple physical quantities measured at the same time, the multiple physical quantities measured at the same time are equivalent to a "predetermined input value group". Therefore, it is most preferred to set the multiple physical quantities measured at the same time as a "predetermined input value group". Here, most of the measurements require a certain degree of time (time range). Therefore, the same time does not need to be strictly simultaneous, and can also be a certain degree of time range required for the measurement of a BH curve.

[0123] In the present embodiment, the object to be measured includes a substance 16 and a film 15 formed on the surface of the substance 16 .

[0124] For example, when the substance 16 is a steel material, an iron oxide film called scale or black scale is formed on the surface of the steel material during the manufacture of the steel material. There are various types of iron oxide films, but magnetite (ferroferric oxide, Fe3O4), wüstite (ferrous oxide, FeO) and rust (hematite, Fe2O3) are generally known. These scales not only have different compositions of oxygen and iron, but also different electromagnetic characteristics. For example, magnetite is magnetic, but wüstite is not magnetic. Here, in order to measure the mechanical properties (especially the surface layer) of the substance 16 as a steel material, the physical quantity is measured from the surface. That is, the substance 16 as a steel material and the scale as a film 15 are taken as measurement objects together to measure the physical quantity.

[0125] Therefore, the film 15 as the oxide scale affects the measurement of the substance 16 as the steel material. In addition, the type and composition of the oxide scale vary depending on the state of the steel material during manufacturing. In addition, sometimes the magnetic properties are anisotropic due to the structure of the steel material itself, and the electromagnetic characteristics are different depending on the object to be measured. Therefore, as described later, an appropriate first model is selected from the first model group 10, and the selected first model is used to infer the mechanical properties of the substance 16 as a characteristic quantity. In general, the inferred characteristic quantity (mechanical property) is processed as a non-destructive measurement result based on the measurement system 20 using the first model.

[0126] Figure 4 1 is a diagram showing a specific structural example of the sensor 14. The sensor 14 is, for example, a magnetic sensor and may include an excitation coil 17 and a magnetizing yoke 18. The sensor 14 applies an alternating magnetic field to the object of measurement while moving relative to the object of measurement. Figure 4 In the sensor 14 shown, one coil is used as both an excitation coil and a coil for measuring electromagnetic changes. The sensor 14 measures the effects of eddy currents and the like caused by the alternating magnetic field in the measured object as changes in electromagnetic quantities. As another example, the sensor for measuring electromagnetic quantities may be constructed by winding an excitation coil around a magnetizing yoke, and separately winding an excitation coil and a coil for receiving signals. As yet another example, the sensor for measuring electromagnetic quantities may be constructed by winding an excitation coil around a magnetizing yoke, and independently arranging a coil for measuring electromagnetic changes between the magnetizing yokes. The sensor for measuring electromagnetic quantities is not limited to any structure as long as it has an excitation coil, a coil for measuring electromagnetic changes, and a magnetizing yoke. Figure 4 The structure shown.

[0127] The signal transceiver 13 is a signal applying unit connected to the excitation coil 17 of the sensor 14 and used to apply a signal for generating an AC magnetic field in the object to be measured. The signal applied to the excitation coil 17 can be either a current or a voltage. The manager can select the desired one appropriately. In addition, the control of the timing of the on-off of the signal or the signal strength is performed by the control unit 3.

[0128] On the other hand, the received signal from the coil for measuring electromagnetic changes is obtained by using the excitation coil 17 and the signal transceiver 13 connected to the excitation coil 17. Specifically, the signal received by the excitation coil 17 is filtered and then sent to the control unit 3 by means of the connection between the physical quantity acquisition unit 2 and the computing device 12. The physical quantity acquisition control unit 2c in the control unit 3 processes the sent signal and extracts the physical quantity. The extracted physical quantity is stored in a predetermined area of ​​the storage unit 9 as a measured physical quantity. Thereafter, the stored physical quantity is used as an input value of the first model or the second model, or as an actual input value for teacher data for generating these models, as needed.

[0129] Figure 5 An example of a signal applied to the exciting coil 17 in order to generate an AC magnetic field is shown. Figure 5 The signal is a signal in which a high-frequency signal is superimposed on a low-frequency signal. By using such a signal, the sensor 14 can efficiently measure the electromagnetic quantity based on the low-frequency signal and the electromagnetic quantity based on the high-frequency signal. Therefore, many types of electromagnetic quantities can be measured. As an example, the low-frequency signal is a sine wave of 150Hz. As an example, the high-frequency signal is a sine wave of 1kHz. By superimposing the high-frequency signal with the low-frequency signal, it is easy to measure the electromagnetic quantity up to the surface layer of the substance 16 even when the substance 16 is formed with a film 15. Here, for example, when the film 15 is thin, when the relative magnetic permeability of the film 15 (the ratio of the magnetic permeability of the substance to the magnetic permeability of the vacuum) is low, magnetism is easy to penetrate. When magnetism is easy to penetrate, the electromagnetic quantity can be measured using only appropriate high frequencies. In addition, for example, when the film 15 is thick, when the relative magnetic permeability of the substance constituting the film 15 is high, magnetism is difficult to penetrate, and the signal is difficult to reach the substance 16. When magnetism is difficult to penetrate, by superimposing the high-frequency signal on the low-frequency signal, magnetism can reach deeper. At this time, the low-frequency signal can be a DC signal. As another example, the low-frequency signal may be a sinusoidal wave signal or a rectangular signal.

[0130] (Generation of the first model)

[0131] Figure 6An example of a flowchart for generating a first model is shown. The generation process of the first model is performed before executing the feature quantity calculation step using teacher data having performance input values ​​and performance output values, so that the first model group 10 is already stored in the storage unit when the feature quantity calculation step is executed. In addition, in the present invention, the teacher data is one or more selected from a predetermined teacher data group, and the teacher data group is also used in the generation of the second model described later. In the present embodiment, the control unit 3 of the computing device 12 generates the first model, but it may also be another device or processor that can access the storage unit 9 to perform the generation process of the first model and store the generated first model group 10 in the storage unit 9.

[0132] First, the evaluation site to be measured is set on the steel surface (step S11). Figure 3 The physical quantity measuring unit 2a and Figure 1 The control unit 3 is used to measure to obtain the input value of one or more electromagnetic quantities (step S12). At this time, the object to be measured is a steel material with oxide scale on the surface. That is, the substance 16 is equivalent to the steel material, and the film 15 is equivalent to the oxide scale. After that, a pre-processing is performed on the measured part (step S13), and the surface hardness of the steel material as a characteristic quantity (mechanical property) is actually measured (step S14). Here, the pre-processing is a process of removing the film 15 as the oxide scale. In step S14, the surface hardness of the steel material is destructively measured. In addition, the surface hardness of the steel material can also be obtained using any measurement method including known techniques. As a highly reliable and effective measurement method, for example, the indentation test method and the dynamic hardness test method can be cited. Among them, in order to measure the local hardness with high precision, the indentation test method is particularly preferably the micro Vickers hardness test method, and the dynamic hardness test method is particularly preferably the rebound hardness test method. The process of steps S11 to S14 is performed again until the teacher data is fully collected (step S15 No). If the teacher data is collected to a predetermined number or more and the collection is completed (yes in step S15), the obtained electromagnetic quantity (actual input value) and steel surface hardness (actual output value) are used as teacher data to generate the first model (step S16). In this embodiment, a plurality of first models are generated as the first model group 10. In this embodiment, the first model is generated by machine learning of a generalized linear model (GLM), but the method of machine learning is not limited to this.

[0133] Here, the state where the collection is completed after the predetermined number of data is collected (Yes in step S15) is equivalent to the "predetermined teacher data group". In addition, the actual input value of the teacher data has the same acquisition conditions as the "one or more input values ​​selected from the predetermined input value group" described later. In addition, the actual input value and the input value may be selected from the same input value group, or may not be selected from the same input value group.

[0134] In addition, in the present embodiment, after the collection of the obtained electromagnetic quantity (actual input value) and the steel surface hardness (actual output value) is completed (Yes in step S15), the first model is generated in step S16, but it is not limited to this. For example, it is also possible to prepare only the teacher data in advance through steps S11 to S15, and then use the pre-prepared teacher data to generate the first model only in step S16.

[0135] In addition, in the collection of teacher data, the physical quantity measuring unit 2a and the control unit 3 used when measuring in the state of the film 15 in the process of manufacturing the substance 16 may also use other physical quantity measuring units 2a and control units 3. In order to improve the accuracy of the first model, the other physical quantity measuring unit 2a and the control unit 3 are most preferably provided with the same structure as the physical quantity measuring unit 2a and the control unit 3 used for measurement during manufacturing.

[0136] (Calculation of feature quantity)

[0137] Figure 7 A flowchart for calculating feature quantities is shown. Figure 6 First, the surface of the steel material with oxide scale as the measurement object was used to calculate the feature quantity. Figure 3 The physical quantity measuring unit 2a and Figure 1 The control unit 3 performs the measurement with the oxide scale as the film 15, and obtains one or more electromagnetic quantities, that is, input values. At this time, the physical quantity measuring unit 2a and Figure 1 The step of measuring the physical quantity of the measurement object having the substance 16 and the film 15 located on the surface of the substance 16 by the control unit 3 corresponds to the measurement step. In the present embodiment, the measurement method executed by the measurement system 20 also includes the measurement step.

[0138] Afterwards, the first model classification processing unit 4 selects a first model suitable for the measured object from the first model group 10 based on the obtained input value (step S21). That is, the first model classification processing unit 4 selects an appropriate first model corresponding to the difference in the amount of scale adhesion and organization on the surface of the measured object. The first model classification processing unit 4 uses, for example, the selection method described in International Publication No. 2021 / 256442 (hereinafter referred to as a reference). Specifically, a support vector machine (SVM) as a type of machine learning that is mainly used in classification problems is used. SVM has the following characteristics: although the dimension of the data is large, the recognition accuracy is high, the risk of over-learning is low, and the number of hyperparameters that need to be optimized is small. Therefore, it is suitable for processing multiple (i.e., multidimensional) physical quantities (such as electromagnetic quantities) as in the present embodiment. The explanatory variables used when selecting the first model in the first model classification processing unit 4 can be appropriately selected according to the relationship between the input value and the output value. In the case of the present embodiment, multiple or one electromagnetic quantity is used as the input value.

[0139] Here, use Fig.16A and Fig. 16B , the case where it is difficult to simply associate the characteristic quantities of the measurement object with a plurality of input values ​​for measurement is described. Fig.16A The figure shows the relationship between a certain input value as and the characteristic value of the measured object (e.g., surface hardness). When a mathematical model (e.g., model As) can be constructed that links an arbitrary input value as (e.g., one of the electromagnetic information) with the characteristic value of the measured object in a one-to-one relationship, the model can be used to calculate the characteristic value (e.g., surface hardness) based on the input value as. However, when the measured object is, for example, steel, there are actually factors such as differences in metal structure and the formation of an oxide layer on the surface. Therefore, Fig.16A As shown, the correlation between any input value as and the feature value exists in multiple relationships depending on the combination of metal structure and oxide scale thickness. Fig.16A In addition to model As, the relationships in model Bs are also shown. Fig.16A As shown in FIG. 1 , even if the measured input value as has the same value, there is a possibility that two hardnesses will be calculated, and the calculation accuracy of the characteristic quantity will be reduced. Here, by selecting an appropriate model, the reduction in the calculation accuracy of the characteristic quantity can be avoided. However, for example, in the case of multiple models having similar output mechanical characteristics for a certain value of an input value as, it is possible to identify these multiple models as one model. In response to this problem, as Fig. 16B As shown, each model can be separately identified by using multiple parameters. Fig. 16BIn the example of , by using the combination of input value as and input value bs, model As and model Bs are separately identified. By using multiple input values ​​in this way, the data group of each model can be determined. Then, an appropriate model is selected from the multiple determined models to be used, and the mechanical feature value can be measured or evaluated with high accuracy. In this embodiment, a classifier is used when using multiple input values ​​to determine the data group of each model.

[0140] In the present embodiment, the first model classification processing unit 4 uses a support vector machine (SVM) as a method for selecting the first model. The support vector machine is a type of machine learning for classification and regression problems, but is not limited to this. In addition, the selected first model is not limited to one. That is, the first model classification processing unit 4 can select one or more first models. In addition, the first model classification processing unit 4 can also select the first model determined to be the most appropriate from one or more first model groups 10. In this case, it can be expected that the difference between the calculated feature value and the actual measured value is minimized, which is the most preferred. In addition, the selection result of the above-mentioned first model can also be stored in a specific area of ​​the storage unit 9 by the first model classification processing unit 4. In this way, in addition to the feature value calculation unit 6, information on which first model is selected can also be used.

[0141] The characteristic quantity calculation unit 6 calculates the characteristic quantity using one or more input values ​​and the selected one or more first models (step S22). In the present embodiment, the characteristic quantity calculation unit 6 can obtain the desired characteristic quantity, i.e., the surface hardness, as an output value by inputting the electromagnetic quantity to the selected appropriate first model.

[0142] The feature quantity calculation unit 6 stores the calculated feature quantity in the storage unit 9 as the calculated feature quantity (step S23). Here, the feature quantity calculated by the feature quantity calculation unit 6 is set as the measured value of the feature quantity measured by the measurement system 20. For example, in the case where the feature quantity is the surface hardness of the steel material, the value calculated by the feature quantity calculation unit 6 is processed as the measured value of the surface hardness of the steel material. The feature quantity group 27 stored in the storage unit is read from the storage unit 9 by the determination unit 8 in step S4 and used for the determination process described later.

[0143] (Generation 1 of the Second Model)

[0144] Figure 81 shows a flowchart for generating the second model. The second model generation process is performed before the separation amount calculation step is performed using one or more teacher data selected from a predetermined teacher data group, so that the second model group 11 is already stored in the storage unit 9 when the separation amount calculation step is performed. In the present embodiment, the control unit 3 of the computing device 12 generates the second model, but it may be another device or processor that can access the storage unit 9 to perform the second model generation process and store the generated second model group 11 in the storage unit 9.

[0145] Through the second model, the relationship between the input values ​​contained in the predetermined input value group (in this embodiment, the electromagnetic quantity group) (in this embodiment, the electromagnetic quantities) can be evaluated. In the generation of the second model, the teacher data selected from the same teacher data group used for the generation process of the first model is used. Therefore, through the second model, the difference between the relationship between the input values ​​when the first model is generated can be evaluated. The deviation is an offset from the first model that represents the difference in the relationship between such input values. Therefore, when the deviation is large, the offset from the first model is large, so there is a concern that the feature quantity estimated using the first model is greatly deviated from the value actually measured. On the contrary, when the deviation is small, it can be judged that the feature quantity can be accurately estimated using the first model. Here, each of the second models can be associated with a specific first model based on the commonality of the teacher data. In addition, the deviation can be obtained for each input value.

[0146] As in the present embodiment, in the case of a selection process S21 of the first model based on the first model classification processing unit 4, any one of the first models must be selected for the physical quantity (input value) of the measured object. Even in the event that there is no appropriate first model, the first model determined to be appropriate by the first model classification processing unit 4 must be selected. In this case, an inappropriate feature quantity (output value) is calculated. In particular, in the middle of manufacturing any product, various unforeseen situations may occur and unexpected input values ​​may be obtained. In this case, it is considered that there is a high possibility that the unforeseen input value cannot be dealt with by only combining the first model classification processing unit 4 with one or more existing first models.

[0147] Therefore, when calculating the characteristic value based on the first model using the classification processing unit 4 and a combination of one or more first models, as in the present invention, the above-mentioned second model calculates the "deviation amount from the first model, that is, the deviation amount" and prompts it to the manager, which has a particularly great effect.

[0148] In the generation process of the second model, first, teacher data is prepared (step S31). As described above, since the teacher data is common to the first model, the teacher data is obtained as Figure 6 The operation of acquiring the teacher data obtained as a result of steps S11 to S15 corresponds to this step. As a further specific example, the control unit 3 may acquire the teacher data stored in the storage unit 9 in the generation process of the first model.

[0149] The control unit 3 analyzes the teacher data and selects the performance input values ​​related to each performance input value as the performance input value for model making (step S32). For each performance input value (each electromagnetic quantity) of the teacher data, one or more performance input values ​​(electromagnetic quantities) with high correlation other than the input value itself are selected. High correlation means, for example, that a change in a certain input value is likely to affect other input values. On the contrary, for example, even if a certain input value is greatly changed, other input values ​​remain unchanged, then the correlation between these input values ​​is low. In the present embodiment, the control unit 3 selects input values ​​based on the height of linearity for each input value, but is not limited to the height of linearity.

[0150] In the present embodiment, selection of teacher data from a predetermined teacher data group is performed by first selecting a performance input value and then selecting teacher data having the performance input value.

[0151] For example, a set of actual input value and actual output value is selected from the teacher data set, which uses a specific input value a1 as the actual input value and a deviation D of the input value a1 as the actual output value D1. One or more selected sets can be used as teacher data.

[0152] In addition, in this method, there is a case where it is difficult to obtain the amount of deviation D with high accuracy. This is because, in machine learning, usually, if a model is made to predict the input value itself based on a certain input value, it becomes a linear regression model close to y=x (x is the input value, y is the predicted value). Therefore, it is known that a value close to a certain input value x is directly output as a predicted value y. In the second model of the present invention, it is also very likely that the input value is directly calculated as a predicted value, so there is a case where the value of the amount of deviation D as the output value becomes a value that is almost the same as the input value, and the difference between the predicted value and the input value is close to 0. Therefore, when selecting the actual input value that becomes the teacher data of the second model, it is preferred that the amount of deviation D of the actual input value itself does not become the actual output value. In other words, the accuracy of the calculated amount of deviation for the actual input value for which the deviation is desired as the actual output value is high when the input value itself does not include the input value itself for which the deviation is desired, and the accuracy is also high when the input value with a high relationship with the input value of the actual output value is selected as the actual input value. Specifically, when one wants to obtain the deviation D1 of a certain input value x1 as an output value with high precision, it is preferable to select an input value x2 different from the input value x1 from which one wants to obtain the deviation D1 as the actual input value, select the deviation D1 of the input value x1 from which one wants to obtain the deviation D1 as the actual output value, and create a second model.

[0153] On the other hand, if another input value x3 having a low relationship with the input value x1 for which the deviation amount D1 is desired as the actual output value is selected as the actual input value, the accuracy of the deviation amount D may decrease. This is because, in machine learning, it is usually difficult to obtain the output value with high accuracy even if the model is made using input values ​​having a low relationship with the output value. Therefore, when selecting the actual input value of the teacher data to be the second model, if the input value x4 having a high relationship with the input value x1 for which the deviation amount D1 is desired as the actual output value is selected as the actual input value, the deviation amount can be obtained with high accuracy and is preferred.

[0154] This time, it was newly found that when the input values ​​used in the present embodiment are a plurality of electromagnetic quantities, the relationship between the electromagnetic quantities has many linear relationships. Therefore, if the linearity between the electromagnetic quantities is high, it is determined that the relationship is high.

[0155] Therefore, in this embodiment, the selection of the actual input value uses the determination coefficient R 2 , the coefficient of determination R 2It is an indicator for evaluating linearity in regression analysis. The coefficient of determination takes a value from 0 to 1, and the closer it is to 1, the higher the linearity. Consider a group of actual input values ​​grouped under a certain condition in the teacher data group. As an example, the above-mentioned multiple electromagnetic quantities obtained from a BH curve measured by a certain excitation voltage or excitation current are set as a group of actual input values. In order to obtain the actual input value group, N measurements were performed. That is, N BH curves were obtained. The electromagnetic quantity a that becomes the actual input value according to one of the BH curves is measured as 10 types a1 to a10. Then, a1_1, a2_1, a3_1, ... a10_1 represent the group of electromagnetic quantities a1 obtained from a BH curve obtained in the first measurement. Then, the group of electromagnetic quantities aN obtained in the Nth measurement is set as a1_N, a2_N, a3_N, ... a10_N. For electromagnetic quantity a, the second suffix indicates the number of BH curves, and the first suffix indicates the number of electromagnetic quantity in the same group as the second suffix.

[0156] Then, the determination coefficient R is calculated for each of the electromagnetic quantities a1_1 to a10_N corresponding to the grouped actual input value groups. 2 When there are multiple actual input value groups, for example, the determination coefficient R is calculated between the electromagnetic quantities a1_1 to a1_N and the electromagnetic quantities a2_1 to a2_N. 2 In addition, the determination coefficient R is calculated between the electromagnetic quantities a3_1 to a3_N and the electromagnetic quantities a5_1 to a5_N. 2 .

[0157] In this embodiment, the determination coefficient R 2 The threshold value is set to 0.5, which is a general standard. 2 When it is 0.5 or more, the linearity is high.

[0158] Therefore, as the actual input value, select one or more input values ​​for which the separation amount D is to be obtained as the actual output value and the determination coefficient R 2 An input value of 0.5 or more is used as an actual input value for generating the second model.

[0159] In addition, the control unit 3 performs pre-processing (step S33) on all input values ​​of the teacher data (electromagnetic quantities in this embodiment). Through pre-processing, different types of input values ​​can be compared with the same benchmark and calculated in the same way. In this embodiment, the pre-processing is a normalization in which the average value is set to 0 and the variance is set to 1 for all types of electromagnetic quantities. Through this processing, the detachment amount output from the second model also outputs a normalized value, which can simply represent how much σ (standard deviation) it is. Sometimes the normalized value in the detachment amount is specifically described as a detachment degree. That is, the detachment degree is an example of a detachment amount. Here, the average value based on normalization is not limited to 0. In addition, the variance based on normalization is not limited to 1. As another example, the pre-processing can be a scaling method, i.e., normalization, in which the minimum value is set to 0 and the maximum value is set to 1 for all types of electromagnetic quantities.

[0160] Then, the control unit 3 generates a second model using the pre-processed input values ​​(step S34). In the present embodiment, an example in which least squares ridge regression is used to generate the second model is described.

[0161] Least squares ridge regression is a method that can enhance the influence of highly correlated input values ​​and weaken the influence of less correlated input values, perform weighting, and obtain (infer) a certain input value from multiple input values ​​by considering the weighting.

[0162] In order to calculate the degree of separation, two main calculation processes are performed in the second model. The first is to calculate the predicted value of each input value based on the teacher data. The second is to take the difference between the predicted value of each input value and the input value. The degree of separation D is calculated in this way. That is, the second model is a model in which equations 4 and 5 summarize the quantities D1 to D10 (the quantities of the types of degree of separation D to be calculated).

[0163] An example of model creation that satisfies the following conditions is described. As actual input values, 10 electromagnetic quantities a1 to a10 calculated from a BH curve are used. In addition, if input values ​​x1 to x10 calculated from a measured BH curve are input to the second model, 10 degrees of separation D1 to D10 are output. Here, the degree of separation of the input value x1 is set to D1, the degree of separation of the input value x1 is set to D2, ..., and the degree of separation of the input value x10 is set to D10.

[0164] In addition, the determination coefficient R was used to create the second model. 2etc. to predetermine which input value x is used to calculate an arbitrary degree of separation D. For example, what is desired to be obtained through the second model is the degree of separation D1 of the input value x1. The actual input values ​​a5 and a7 that are closely related to the actual input value a1 as the input value x1 are selected as the actual input values ​​for the second model making. Here, a1 to a10 represent the actual input values. A predetermined input value group is given as x1 to x10. Here, when obtaining the degree of separation D1 of the input value x1 in the input value group x1 to x10, the predicted value y1 of x1 is obtained based on x5 and x7 in the same input value group, and the degree of separation D5 is obtained by subtracting the value of x1 from the predicted value y1 of x1.

[0165] In addition, as described above, the second model outputs 10 kinds of detachment degrees D1 to D10 when input values ​​x1 to x10 selected from the measured input value group are input (for example, 10 kinds of electromagnetic quantities calculated from one measured BH curve). However, this is a value obtained by combining the calculation model of the detachment degree D1 of the actual output value, the calculation model of the detachment degree D2 of the actual output value, and the calculation model of the detachment degree D10 of the actual output value. Therefore, the method of making the calculation model of the detachment degree of the actual output value x1 is shown as an example. In addition, the method of making the other detachment degree calculation models of x2 to x10 is the same as that of x1.

[0166] Prepare the teacher data for the second model you want to create and the teacher data that has been pre-processed as described above. In order to calculate the degree of separation D1 of x1, you first need to find the relationship for calculating the predicted value y1 of a1 based on a5 and a7. The relationship is a regression equation, as shown in the following formula 1. The coefficient ω of formula 1 needs to be found.

[0167] [Formula 1]

[0168]

[0169] The following equation 2 expands equation 1 into N BH curves. The coefficient ω is obtained by least squares ridge regression based on the teacher data. Assuming that the teacher data is in an ideal state with no measurement error and noise, if the teacher data corresponding to the actual input value of equation 1 is substituted into equation 1, it becomes equation 2.

[0170] [Formula 2]

[0171]

[0172] If the matrix on the left is set to B, the first matrix on the right containing input value information is set to A, and the second matrix containing ω information is set to Ω, then Equation 2 is expressed as Equation 3.

[0173] [Formula 3]

[0174] B=AΩ…(Formula 3)

[0175] However, in reality, the teacher data contains measurement errors or noise, so Formula 2 is not equal to Formula 3, and a deviation of the value occurs on the left side. Therefore, a combination of coefficients ω that minimizes the deviation is obtained. If the ridge regression coefficient is set to α and the unit matrix is ​​set to I, the matrix Ω that is desired to be obtained by the least squares ridge regression is expressed by Formula 4. Here, the ridge regression coefficient α is a high parameter for suppressing over-learning, and is set to 0.1 in this embodiment.

[0176] [Formula 4]

[0177]

[0178] The coefficient ω can be obtained by equation 4, so the degree of separation can be obtained. Let a predetermined input value group be x1 to x10. The degree of separation D1 of the desired input value x1 can be obtained from x5 and x7 of the same input value group by equation 5. Among them, the predicted value of x1 is obtained by ω0+ω5x5+ω7x7, and the difference is taken from the input value x1.

[0179] [Formula 5]

[0180] D=ω0+ω5x5+ω7x7-x1…(Formula 5)

[0181] The other actual output values ​​D2 to D10 can also be obtained by the same method. It is not necessary to calculate the separation degree D for all input values, and in particular, the separation degree D that seems to have a greater influence on the product and the manufacturing method of the product can be selected for calculation.

[0182] (Calculation of Degree of Separation 1)

[0183] Fig. 9 A flowchart for calculating the degree of separation is shown. Figure 8 The degree of separation is calculated by using the second model generated in the process of . The degree of separation represents the deviation (offset) from the first model selected for the input value obtained in the above-mentioned feature quantity calculation process. The degree of separation can be used to evaluate the suitability of the first model selected in the feature quantity calculation process from the perspective of the input value.

[0184] The second model classification processing unit 5 selects from the second model group 11 based on one or more input values ​​used in the feature quantity calculation process (step S41). Specifically, the second model classification processing unit 5 selects the second model_n corresponding to the first model_n selected by the first model classification processing unit 4. In the case where there are multiple first models selected, the second model classification processing unit 5 selects multiple second models corresponding to each first model. That is, the second model classification processing unit 5 can select one or more second models.

[0185] The deviation amount calculation unit 7 calculates the deviation degree D using one or more input values ​​as input values ​​used when calculating the feature amount using the first model and the selected one or more second models (step S42 ).

[0186] (Determination Process)

[0187] The present invention can also perform determination processing as needed. The determination unit 8 determines whether the first model is qualified or not or the level of the degree of separation based on the feature quantity calculated by the feature quantity calculation unit 6 and the degree of separation calculated by the degree of separation calculation unit 7. In addition, the determination result is processed into a state that is easy for the administrator to see as needed, and then output to the display unit 1.

[0188] Fig.10 An example of a flowchart for explaining the determination process of pass or fail is shown. First, when the degree of separation is greater than a predetermined threshold value (first threshold value) (No in step S51), the determination unit 8 determines that the selected first model is inappropriate because the feature quantity inferred by the first model selected for the measured input value has a large deviation from the teacher data of the first model, and therefore the selected first model is inappropriate, and therefore the result is determined as "failed". Here, although the degree of separation is calculated for each input value, when the degree of separation for any input value is greater than the first threshold value, the determination unit 8 may determine that the result is "failed".

[0189] Then, the determination unit 8 labels the input value determined as "failed" with a label indicating that it is "failed" and accumulates it as the actual input value for new teacher data (step S53). The actual input value for new teacher data can be stored in the storage unit 9 by addition.

[0190] The determination unit 8 proceeds to the process of step S52 when all the degrees of separation are below the first threshold. This is because, in this case, the output value inferred by the first model selected for the measured input value is small relative to the teacher data deviation of the first model, and it can be determined that the selected first model is appropriate. Here, although the degree of separation is calculated for each input value, the determination unit 8 can proceed to the process of step S52 when all the degrees of separation for all input values ​​are below the first threshold.

[0191] The determination unit 8 determines whether the quality is acceptable or unacceptable by determining whether the feature value is below a predetermined threshold value (second threshold value) (step S52). Fig.10 In the example, if the feature value is less than the second threshold value (Yes in step S52), it is determined to be qualified, and if the feature value is greater than the second threshold value (No in step S52), it is determined to be unqualified. In addition, the qualified and unqualified judgment criteria of the branch in step S52 can be changed according to the content of the feature value. For example, it can also be changed with Fig.10 On the contrary, if the feature quantity is greater than the second threshold value (Yes in step S52), it is determined to be qualified, and if the feature quantity is less than the second threshold value (No in step S52), it is determined to be unqualified. In addition, in the determination of the feature quantity, the number of thresholds is not limited to one, and multi-level determination can be performed based on multiple determination criteria.

[0192] In addition, the following processing can also be performed for the input value judged as "unqualified" based on the size of the detachment degree in step S53. For a certain measurement position of the input value judged as "unqualified", the surface hardness (feature quantity) of the steel material is actually measured separately. Then, a label of "unqualified" and the feature quantity actually measured as the actual output value are assigned to each input value. Then, after the input values ​​judged as "unqualified" are fully accumulated, as the actual input values, a predetermined number or more of input values ​​that are assigned the label of "unqualified" and the feature quantity actually measured are selected as new teacher data. The control unit 3 can generate an additional first model based on these new teacher data. In other words, the control unit 3 can also assign the actually measured feature quantity to each input value for the input value that is marked with a label that knows the "unqualified" judgment based on the size of the detachment degree, and select a predetermined number or more as new teacher data for making the first model. Using the new teacher data thus produced, a learning model can be generated through machine learning, where the learning model takes one or more input values ​​as input values ​​and takes a feature quantity having a correlation that can be derived from the input values ​​as output values. The specific method for generating the first model can be in accordance with the above Figure 6 By adding the first model, the accuracy of feature quantity calculation can be improved.

[0193] (Display Processing)

[0194] The determination unit 8 generates information related to the characteristic quantity based on the calculated characteristic quantity so as to be output to the display unit 1. In addition, the determination unit 8 generates information related to the degree of separation based on the degree of separation calculated in step S3 or the determination result obtained in step S4 so as to be displayed on the display unit 1. In addition, information related to the physical quantity may be generated based on the measured physical quantity.

[0195] The information related to the characteristic quantity, the information related to the degree of separation, and the information related to the physical quantity displayed by the determination unit 8 may be in any form. Among them, a two-dimensional image is particularly preferred because it has good visibility for managers of computing devices or manufacturing equipment. As a two-dimensional image, an image called a chart or a map can be exemplified.

[0196] Specifically, as information related to the characteristic quantity, a map of the characteristic quantity or a chart of the characteristic quantity can be exemplified. As information related to the degree of separation, a map of the degree of separation, a chart of the degree of separation, a map of the determination result, or a chart of the determination result can be exemplified. The determination unit 8 can generate one or more information from these information in a predetermined form or according to a selection made as needed. In addition, these maps or charts can be generated in a superimposed manner or in parallel. In addition, as information related to the physical quantity generated by the determination unit 8, it can also be set to at least one of a map and a chart.

[0197] In the present embodiment, the determination unit 8 generates at least one of a graph and a map in which the characteristic amount and the degree of separation are superimposed or displayed in parallel, and outputs it to the display unit 1. Based on at least one of the graph and the map displayed on the display unit 1, a manager of a steel manufacturing facility, for example, can take measures such as optimizing manufacturing parameters. Figures 11A to 11C is a diagram showing an example of color mapping. Fig.12 It is a figure which shows an example of a graph.

[0198] Figures 11A to 11C This is an example of color mapping that displays the characteristic quantity of steel in a color corresponding to the value corresponding to the measurement position when the steel is viewed from above. Here, the characteristic quantity is the surface hardness of the steel. Each input value corresponds to each electromagnetic quantity. The coordinates of the measurement position in the horizontal direction of the paper represent the length direction of the steel, and the vertical direction represents the width direction of the steel. Fig.11A represents the feature quantity calculated by the first model, Fig. 11B Indicates the degree of separation of the input value 1 as the first input value, Fig. 11C Indicates the degree of separation of input value n, which is the nth input value. In addition, Figures 11A to 11C The vertical intensity on the right side of represents the scale of the color mapping. Fig.11AIn the figure, the surface hardness of the steel is shown as between 180 and 240 HL (measured using a rebound hardness tester). Here, 180-240 HL is an example of the value of the surface hardness, and the conversion value may be HV (Vickers hardness conversion) or other hardness conversion values. Fig. 11B and Fig. 11C In the figure, the degree of separation for each input value is shown by the standard deviation σ. Of course, it can also be a value other than the standard deviation. By displaying such a color map on the display unit 1, for example, a manager of a steel computing device or manufacturing equipment can confirm which electromagnetic quantity has a high degree of separation at which part, and can also confirm the hardness at each place.

[0199] Fig.12 This is a graph that displays the data of the degree of separation of each input value in a row in the vertical direction of the display screen without associating it with the position of the steel material. Fig.12 In the case where the degree of separation is greater than a certain threshold (e.g. Fig.10 In the case of a first threshold value of ), the data is displayed in black, etc. By displaying such a graph on the display unit 1, for example, a manager of a steel manufacturing facility can immediately understand which input value has a large degree of deviation.

[0200] [Implementation Method 2]

[0201] In the first embodiment, the second model classification processing unit 5 directly uses the result selected by the first model classification processing unit 4, but the present invention is not limited to this. Therefore, as a second embodiment, a case where there is no clear correspondence between the first model and the second model, and an example of the second model classification processing unit 5 that can be used in this case are described.

[0202] In the present invention, the first model is a model that models the mutual relationship between the input value and the feature quantity having a correlation that can be derived from the input quantity. Therefore, the first model classification processing unit 4 is similarly derived from the input value and the feature quantity having a correlation that can be derived from the input quantity. On the other hand, the second model is a model that reflects the mutual relationship between the input values ​​used when generating the first model, and basically the second model classification processing unit 5 also reflects the mutual relationship between the input values ​​used to generate the first model.

[0203] Therefore, in Embodiment 2, the second model generation process (step S34) and the selection process (step S41) of the second model classification processing unit 5 from the second model group 11 are different from Embodiment 1. These processes are described below. In addition, detailed descriptions of the devices and processes common to Embodiment 1 are omitted.

[0204] (Generation of the Second Model 2)

[0205] The control unit 3 generates a second model using the pre-processed input values ​​(step S34). In this embodiment, the generation of the second model is described using an example using least squares ridge regression as in the first embodiment.

[0206] Regarding the generation process of the second model in the second embodiment, Figure 8 The process is the same as the process shown in the figure. The difference is which teacher data is used from the teacher data group to generate the first or second model. In the case of correspondence, in the first model_m which is the m-th model of the first model and the second model_m which is the m-th model of the second model, exactly the same teacher data is used to generate the same number of second models as the first model. On the other hand, in the case of no correspondence, teacher data is appropriately selected from the teacher data group for generating the first model to make the second model. At this time, in the first model_m which is the m-th model of the first model and the second model_m which is the m-th model of the second model, different teacher data are used to generate one or more second models.

[0207] The second model classification processing unit 5 selects a second model suitable for the measurement object from the second model group 11 based on one or more input values ​​used in the calculation process of the feature quantity (step S41). The second model classification processing unit 5 can use the selection method described in the above-mentioned reference in the selection of the second model. In the present embodiment, the second model classification processing unit 5 uses a support vector machine, which is a type of machine learning for classification and regression problems, as a method for selecting the second model, but is not limited to this. In addition, the selected second model is not limited to one. That is, the second model classification processing unit 5 can select one or more second models.

[0208] (When judging and processing by level)

[0209] In implementation modes 1 and 2, except Fig.10 In addition to the pass / fail determination process shown above, products can also be classified by grade. This method can be applied to product quality management, for example.

[0210] The following describes a product, taking a steel plate as an example, and a characteristic quantity of the product, taking the surface hardness of the steel plate as an example. In addition, the degree of separation is output as a standard deviation σ.

[0211] When determining the grade of the surface hardness of a steel plate, the grade is determined for a predetermined region based on a plurality of pieces of detachment information, rather than determining the detachment degree of each predetermined input value. In the present embodiment, the predetermined region is described as one steel plate. That is, the detachment grade is determined for each steel plate.

[0212] Fig.17 Indicates the judgment process using the level of detachment. The judgment process uses conditions (a) to (c). Condition (a) is that the detachment degree of each input value contained in the area of ​​1 steel plate satisfies the conditions of 1σ being 95% or more and 2σ being 5% or less. Condition (b) is that the detachment degree of each input value contained in the area of ​​1 steel plate satisfies the conditions of 1σ being 80% or more and 3σ being 1% or less. Condition (c) is that the detachment degree of each input value contained in the area of ​​1 steel plate satisfies the conditions of 1σ being 10% or more and 3σ being 60% or less. Fig.17 The judgment process determines whether the conditions are met in the order of condition (c), condition (b), and condition (a). First, if condition (c) is not met (condition (c) is no), the judgment unit 8 judges it as "unqualified" and marks the steel plate with a label that indicates that it is "unqualified". When condition (c) is met (condition (c) is yes), if condition (b) is not met (condition (b) is no), the judgment unit 8 judges it as "level 3", and marks the steel plate with a label that indicates that it is "level 3". When condition (b) is met (condition (b) is yes), if condition (a) is not met (condition (a) is no), the judgment unit 8 judges it as "level 2", and marks the steel plate with a label that indicates that it is "level 2". When condition (a) is met (condition (a) is yes), the judgment unit 8 judges it as "level 1", and marks the steel plate with a label that indicates that it is "level 1".

[0213] If the determination unit 8 completes the grading based on conditions (a) to (c) for all steel plates to be determined, the determination process ends. In addition, in the present embodiment, the conditions for the special grade classification are set to three, but it can be one or more than two. The present invention is not limited to this. The special grade classification can be appropriately changed according to the specifications of the product and the requirements of the customer.

[0214] In addition, when abnormality determination is performed in relation to the level, the output of which physical quantity (electromagnetic quantity) deviates from the teacher data by how much for one measurement position, so data that deviates in the same way for the same physical quantity (electromagnetic quantity) can be simply collected as a new classification. In addition, all the degrees of separation from each model can be used to create a model in a multidimensional manner. In this case, all models can be considered, including the models that were not selected in the classification.

[0215] (Application example)

[0216] The calculation method and calculation device described above can be used in the manufacturing method, management method and manufacturing equipment of the product. In addition, the measurement object in the middle of manufacturing may also include substances that do not remain in the finished product, such as the oxide scale (corresponding to the film 15) for the steel (corresponding to the substance 16). In other words, the measurement object is equivalent to the product in or after manufacturing. The manufacturing method or manufacturing equipment for manufacturing the product may be equipped with the above-mentioned calculation method or calculation device as a part of measurement, control of the manufacturing process, abnormality detection for the manufacturing process, etc. on the basis of the manufacturing steps or manufacturing equipment. Here, the manufacturing steps or manufacturing equipment may be a known manufacturing step or manufacturing equipment. In addition, the manufacturing steps or manufacturing equipment may be an unknown manufacturing step or manufacturing equipment. For example, the manufacturing steps or manufacturing equipment may be the content recorded in the above-mentioned reference.

[0217] The manufacturing steps described above can be performed by adding a control step, which uses the characteristic quantity (e.g., size, mass, temperature, mechanical characteristics, or a combination of these) or the degree of separation of the manufactured product obtained from the above-mentioned calculation method or calculation device to control the manufacturing step or the manufacturing equipment. That is, a manufacturing method in which a control step is added to the above-mentioned manufacturing step can be realized.

[0218] In addition, the manufacturing steps described above can be performed by adding a management step, and the management step is classified based on the characteristic quantity (for example, size, mass, temperature, mechanical characteristics or a combination of these) or the degree of separation of the manufactured product obtained from the above-mentioned calculation method or calculation device. That is, a management method in which a management step is added to the above-mentioned manufacturing steps can be realized.

[0219] In addition, the calculation method, calculation device, measurement method, measurement device and measurement system described above can be used in the manufacturing method, management method and manufacturing equipment of the material 16 such as steel. For example, the manufacturing method or manufacturing equipment for manufacturing the material 16 such as steel can be equipped with the above-mentioned measurement method or measurement device as part of the inspection on the basis of the known manufacturing steps or manufacturing equipment. Here, the known manufacturing steps or manufacturing equipment can be, for example, the manufacturing steps or manufacturing equipment recorded in the above-mentioned reference. In addition, a management step can be performed to classify the material 16 manufactured by such a manufacturing method based on the characteristic quantity of the material 16 (for example, mechanical properties such as surface hardness). That is, the management method can be implemented by a manufacturing method that incorporates a management step.

[0220] In addition, it can be an application example in which the characteristic quantity is used as the control of a certain device and the degree of separation is used as the abnormality of the device. For example, when the characteristic quantity is set to the rotation speed of a certain motor, the input value 1 is set to the voltage, the input value 2 is set to the pressure of the shaft, and the input value 3 is set to the temperature of the motor, the rotation speed of the motor can be obtained using the input values ​​1 to 3 and the first model as the characteristic quantity. In addition, the degree of separation from the first model for the input values ​​1 to 3 is obtained using the second model, so that it can be determined whether the value of the obtained characteristic quantity is a normal value. For example, when it is not separated from the first model and the rotation speed of the motor is too large, it is possible to perform a control such as reducing the rotation speed. This is because it can be determined that the characteristic quantity predicted based on the measured input value, that is, the rotation speed of the motor, is the correct value. In addition, although the rotation speed is calculated to be a normal value, the input value is originally separated from the first model. Since the measured input value is not based on the learning model, it is determined to be an abnormal state, so that it is known that the predicted rotation speed of the motor is originally a wrong value. There is a high possibility, so the control of temporarily stopping the motor can be performed.

[0221] At this time, the condition of the data used for the first model and the second model is the condition that the respective characteristic quantities are associated with the predetermined input value groups measured at the same time. This is because, for example, even if a model is generated that associates the rotational speed of the motor as a characteristic quantity with the voltage of input values ​​1 to 3 measured the previous day, the pressure of the shaft, and the temperature of the motor and performs machine learning, it is impossible to correctly evaluate the relationship between the characteristic quantity and the input value group. The data used to constitute the first model and the second model are data that obtain and associate the characteristic quantity and the input value group at the same time or the most relevant time based on the same time. Therefore, as described above, it is most preferred to set multiple physical quantities measured at the same time as a "predetermined input value group". In this case, the same time also refers to a specified time. Here, most of the measurements require a certain degree of time (time range). Therefore, the same time does not need to be strictly simultaneous, for example, it can be a certain degree of time range spent on measuring the input values ​​1 to 3 of a group of quantities.

[0222] The present disclosure is described based on the drawings and embodiments, but it should be noted that it is easy for those skilled in the art to make various modifications and corrections based on the present disclosure. Therefore, it should be noted that these modifications and corrections are included in the scope of the present disclosure. For example, the functions included in each unit, each step, etc. can be reconfigured in a logically non-contradictory manner, and multiple units and steps can be combined into one, or divided.

[0223] For example, the present disclosure can also be implemented as a program that describes the processing contents of each function of the computing device 12 or a storage medium that records the program. In addition, the present disclosure can also be implemented as a program that describes the processing contents of each function of the measuring device 19 or a storage medium that records the program. It should be understood that the scope of the present disclosure also includes these contents.

[0224] The configuration of the computing device 12 described in the above embodiment is an example, and may not include all the components. For example, the computing device 12 may be configured without the display unit 1. Some modified examples will be described below with reference to the drawings.

[0225] [Implementation method 3]

[0226] In this method, it is indicated that the deviation amount output from the second model becomes the same physical quantity as the physical quantity input to the first model instead of the standard deviation. For example, when the physical quantity input to the first model is a plurality of electromagnetic quantities (input value 1, input value 2, input value 3, input value 4, ..., input value n), the deviation amount output from the second model is also output as an offset from any one of input value 1 to input value n. Hereinafter, as in Embodiment 1, the object is set to a steel plate with oxide scale, the physical quantity is set to an electromagnetic quantity, and the characteristic quantity output from the first model is set to the surface hardness for a more specific description. However, the following description is a specific example, and any method that is formulated by a characteristic quantity and one or more input values ​​is sufficient, and is not limited to the case of hardness prediction.

[0227] Electromagnetic sensors are often used in the prediction technology of the surface hardness of thick plate steel in steelmaking. Reasons for using electromagnetic sensors include that it is not a method that requires contact with a medium like ultrasonic waves, that multiple explanatory variables can be obtained based on the shape of the BH curve, and that the oxide film on the surface of the thick plate steel is less affected. A prediction model for predicting the hardness of the thick plate surface using multiple input values ​​(electromagnetic quantities) obtained by the electromagnetic sensor is represented by the following formula (6).

[0228] [Formula 6]

[0229] hv=ω0+ω1x1+ω2x2+…+ω n x n …(Formula 6)

[0230] Here, hv is the predicted hardness of the thick plate surface, that is, the "hardness value". ω0 is a constant term. ω1~ωn are coefficients. x1~xn are input values, which are multiple electromagnetic quantities in this example. The prediction model can be a linear regression model as shown in formula (6), but is not limited to this. For example, it can also be a quadratic regression model, as long as it is a method of weighting the input value to calculate the characteristic quantity. The coefficient ω represents the influence value on the hardness value when the corresponding electromagnetic quantity changes by 1 (unit amount), and is predetermined by the teacher data. For example, if the electromagnetic quantity x1 increases by 1, the calculated hardness value becomes the value of the increase coefficient ω1. In this way, the regression coefficient of the prediction model for predicting the hardness of the thick plate surface represents the weight of each electromagnetic quantity as an input value on the hardness value.

[0231] In addition, a reference value (hereinafter referred to as "reference offset") indicating the hardness of the thick plate surface predicted based on the separation amount of each electromagnetic quantity is expressed by, for example, the following equation (7). The reference offset means a reference value indicating how much the predicted hardness value deviates from the actual hardness value.

[0232] [Formula 7]

[0233] Deq=ω1D1+ω2D2+…+ω n D n …(Formula 7)

[0234] Here, Deq is the reference offset. D1~Dn are the deviations, corresponding to the input values ​​x1~xn respectively. In addition, ω is the same as formula (6). For example, the coefficient ω1, as mentioned above, represents the influence on the hardness value when the input value x1 changes by 1. "ω1D1" multiplied by the deviation from the first model, that is, D1, represents the influence on the hardness value caused by the deviation of the input value x1. Similarly, "ω2D2" represents the influence on the hardness value caused by the deviation of the input value x2. Therefore, formula (7) obtained by adding together all the terms representing the influence on the hardness value caused by these input values ​​is a calculation formula that takes into account the reference offset (Deq) of all input values.

[0235] In this method, the calculation method of the reference offset also basically follows Figure 2 However, the weight coefficient may be obtained instead of the calculation of the feature amount (step S2). In addition, the level of the deviation amount is determined (step S4) by calculating the reference offset amount.

[0236] First, the input value (electromagnetic quantity at the measuring point) is obtained from the sensor. The sensor can be Figure 4The electromagnetic sensor shown is a type commonly used in eddy current flaw detection tests, etc. Next, the weight coefficients of each electromagnetic quantity are obtained from the machine learning model that predicts the hardness value. In addition, the detachment amount is calculated for the measured electromagnetic quantity. For example, a least squares ridge regression model can be used to obtain a predicted value, and the detachment amount is calculated based on the difference between the predicted value and the measured value, but it is not limited to such a method. For example, the LOF method that can calculate the distance between the predicted value and the measured value in N-times space can be used. The reference offset is calculated based on the obtained weight coefficient and the calculated detachment amount, and is output as the obtained information (calculation result).

[0237] Here, for example, the operator can judge whether the steel plate is qualified or unqualified based on the output reference offset. For example, the operator can refer to the pre-obtained specification value, customer required value, standard value, etc. to determine how much the hardness prediction value deviates before judging the steel plate as unqualified. That is, a method of setting a threshold value for the reference offset and judging as unqualified even if only one measurement location exceeds the threshold value can be adopted.

[0238] (Application example of embodiment 3)

[0239] Fig.18 and Fig.19 This is a graph that maps the results of calculating the hardness prediction reference offset for two thick steel plates and performs binarization processing with 20HV as the threshold. The analyzed thick steel plates are all steel plates with a width of about 3m, a length of about 12m, and a thickness of about 3cm. After calculating the predicted value using the method of this method, the actual hardness value was measured using the Echo Chip measurement.

[0240] Fig.18 Indicates the analysis result of a normal board. "Normal" means that except for the electromagnetic sensor dead zone at the end of the board ( Fig.18 Except for D), the difference between the measured value and the predicted value in the parts where the hardness was measured was within 20HV. The hardness prediction reference offset was within the range of -5HV to +5HV for the entire plate surface, and there was no part that deviated by more than 20HV except for the electromagnetic sensor dead zone.

[0241] Fig.19 Indicates the analysis results of abnormal boards. "Abnormal" means that there is at least one point where the difference between the measured value and the predicted value is 20HV or more in the area where the hardness is measured, except for the electromagnetic sensor dead zone at the end of the board. The hardness prediction reference offset is about 15HV on the entire board surface, and there are also areas near the center of the board that deviate by more than 20HV in addition to the sensor dead zone.

[0242] [Modification 1]

[0243] As including Figure 1 A variation of the measurement system 20 of the computing device 12 is Fig.13 An example is shown in FIG. In addition, it corresponds to both Implementation Example 1 and Implementation Example 2. In addition, the function is the same as Figure 1 The same parts are marked with the same reference numerals, and detailed descriptions are omitted. Fig.13 In the embodiment, the display unit 1a is arranged in the computing device 12, the display unit 1b is arranged in the physical quantity acquisition unit 2, and the display unit 1c is arranged in another place. The distance between the computing device 12 and the physical quantity acquisition unit 2 is larger than that of the modified example 2 described later. In the case where the operation manager of the physical quantity acquisition unit 2 is located near the physical quantity acquisition unit 2, and the process manager of the computing device 12 is located near the computing device 12 in a separated position, such a structure is preferred. In this case, the display unit c can be arranged in the manufacturing equipment of the product. For example, the physical quantity acquisition unit 2 is assembled into the manufacturing equipment of the product, and the computing device 12 is arranged at a position away from the physical quantity acquisition unit 2, and the physical quantity measurement, characteristic quantity calculation and separation degree calculation of the manufactured product are almost fully automatically performed.

[0244] The display units 1a, 1b, and 1c function as Figure 1 The same function as the display unit 1. Figure 1 The same is true for the case of display units 1a, 1b, and 1c. The information appropriately selected by the control unit 3 from the information related to the physical quantity, characteristic quantity, and degree of separation calculated by the control unit 3 is displayed. As in the above-mentioned embodiment 1 or embodiment 2, the measured physical quantity may also be output. The specific structure of the display units 1a, 1b, and 1c and the information displayed are similar to Figure 1 The same situation.

[0245] In addition, when there are two or more display units, for example, the display content may be appropriately changed according to the business content of the manager who monitors the display units 1a, 1b, and 1c. For example, the display unit 1a in the computing device 12 displays information related to the degree of separation calculated using one or more second models output from the control unit 3 to the process manager. Figure 1 Similarly, information related to the characteristic quantity or information related to the measured physical quantity may be displayed together on the basis of appropriate selection. Next, the display unit 1b located in the physical quantity acquisition unit 2 outputs the characteristic quantity calculated using one or more first models output from the control unit 3 to the operation manager. In addition, information related to the physical quantity measured by the physical quantity measurement unit 2a and the control unit 3 may be displayed together on the basis of appropriate selection. The display unit 1c located in any part of the manufacturing equipment outputs the characteristic quantity calculated using one or more first models output from the control unit 3 to other operation managers.

[0246] The scanning unit 25 moves the sensor 14 to a predetermined measurement position relative to the product manufactured by the manufacturing equipment under the control of the control unit 3. By moving the scanning unit 25 to the predetermined measurement position relative to the product, it is possible to automatically measure the physical quantity at the required measurement position, calculate the characteristic quantity, and calculate the degree of separation.

[0247] [Modification 2]

[0248] As including Figure 1 Another modification of the measurement system of the computing device 12 is shown. Fig.14 In addition, the function Figure 1 , Fig.13 The same parts are marked with the same reference numerals, and detailed descriptions are omitted. Fig.14 In the embodiment, the computing device 12 is connected to the measuring device 19 for measuring a specific physical quantity of the product via a network, forming a Figure 1 , Fig.13 Different measurement systems. This example is particularly preferred when the distance between the computing device 12 and the measurement device 19 is very far and when the computing device 12 and the measurement device 19 have different managers. In addition, this example is further preferred when the computing device 12 and the measurement device 19 are configured in different countries and used in different companies.

[0249] The calculation device 12 includes the control unit 3 , the storage unit 9 , the display unit 1 a , and the data acquisition unit 21 . The data acquisition unit 21 is controlled by the control unit 3 .

[0250] The measuring device 19 includes a display unit 1b, a physical quantity acquisition unit 2, a data output unit 22, and a physical quantity acquisition unit control unit 24. The physical quantity acquisition unit 2 includes a physical quantity measurement unit 2a and a scanning unit 25. The physical quantity acquisition unit control unit 24 includes the data output unit 22 and the physical quantity acquisition unit 2.

[0251] Furthermore, the data acquisition unit 21, the data output unit 22, and the data output unit are connected via a network.

[0252] The sensor 14 of the physical quantity measuring unit 2a acquires a signal for the object to be measured. This signal corresponds to a signal related to the physical quantity of the object to be measured.

[0253] The scanning unit 25 of the physical quantity acquisition unit 2 moves the sensor 14 to a measurement position relative to the object to be measured. A variety of mobile technologies can be used in conjunction with the manufacturing equipment used. The mobile technology used can use wheels, arms, or mechanical arms, etc. In addition, it can also be operated by human or mechanical force. In addition, the control of the scanning unit 25 is performed by the physical quantity acquisition unit control unit 24.

[0254] That is, the physical quantity acquisition unit control unit 24 controls the physical quantity measurement unit 2 a and the scanning unit 25 to measure the physical quantity.

[0255] The physical quantity acquisition unit control unit 24 functions as a physical quantity calculation unit 29 b that calculates a physical quantity as a measurement value based on a signal acquired by the physical quantity measurement unit 2 a .

[0256] The physical quantity acquisition unit control unit 24 outputs the physical quantity calculated based on the measured signal as an input value from the data output unit 22 to the computing device 12 via a network including the Internet. The control unit 3 of the computing device 12 acquires the output physical quantity via the data acquisition unit 21 and stores it in a specific area of ​​the storage unit of the computing device 12. Then, it is stored as a measured physical quantity group 26.

[0257] The storage location of the physical quantity extracted as the measurement value is not limited to the storage unit 9 of the computing device 12. For example, the physical quantity acquisition unit control unit 24 may store and save the physical quantity extracted as the measurement value in a storage unit (not shown) provided in the measuring device 19.

[0258] The physical quantity acquisition unit control unit 24 is similar to the control unit 3 and is configured to include one or more processors in the measuring device 19. The processor is, for example, a general-purpose processor or a dedicated processor dedicated to a specific process, but is not limited thereto and can be any processor. The physical quantity acquisition unit control unit 24 controls not only the above-mentioned processes but also the overall operation of the measuring device 19. The physical quantity acquisition unit control unit 24 executes control of the physical quantity acquisition unit 2, control of the data acquisition unit 21, and control of the data acquisition unit 23 according to a program read from a storage unit, storage unit 9, or an accessible storage device in the measuring device 19 not shown in the figure. In addition, it also functions as a physical quantity calculation unit 29.

[0259] Then, the control unit 3 calculates information related to the characteristic quantity and information related to the degree of separation based on the acquired physical quantity. If necessary, information related to the measured physical quantity can also be calculated. These calculation methods are similar to Figure 1 The same is true for the case of , so the detailed description is omitted. In addition, the storage of the calculated feature quantity group 27 and the calculated separation quantity group 28 to the storage unit 9 is also the same as Figure 1 The situation is the same, so the detailed description is omitted.

[0260] After that, the control unit 3 transmits information related to the characteristic quantity calculated by one or more first models output by the control unit 3 from the data acquisition unit 21 to the measuring device 19 via the network. The physical quantity acquisition unit control unit 24 acquires the information related to the physical quantity sent from the data acquisition unit 23 in the measuring device 19 and displays it on the display unit 1b. At this time, information related to the measured physical quantity can also be displayed together.

[0261] Accordingly, the control unit 3 displays information related to the degree of separation calculated by one or more second models output by the control unit 3 on the display unit 1a of the computing device 12. At this time, at least one of information related to the characteristic quantity calculated by one or more first models and information related to the measured physical quantity may also be displayed.

[0262] Like this, in Fig.14 In the example of , when the computing device 12 and the measuring device 19 placed at remote locations have different managers, information suitable for the business of each manager can be appropriately selected and displayed on the respective display units 1a, 1b. Of course, in the case where there are two or more display units as in the second modification, it is sufficient as long as the information related to the feature quantity, the information related to the degree of separation, or the information related to the physical quantity can be appropriately selected and displayed on the respective display units 1a, 1b.

[0263] Here, the structure of the data acquisition unit 21, the data output unit 22 or the data acquisition unit 23 is arbitrary. As for the function, it is not particularly limited as long as it is a technology that can exchange information via a network. For example, anything can be used as long as it is an interface for connecting to the Internet. For example, a virtual server on the network represented by a physical server, a cloud service, etc. can also be used. In addition, an email service can also be used to exchange data.

[0264] The control of the data acquisition unit 21 is performed by the control unit 3. The control of the data output unit 22 is performed by the physical quantity acquisition unit control unit 24.

[0265] The data output unit 22 and the data acquisition unit 21 may be an integrated unit having both functions.

[0266] Here, the configurations of the physical quantity acquisition unit control unit 24, the storage unit (not shown), and the display unit 1b of the measuring device 19 are arbitrary. Fig.14 In the figure, the physical quantity acquisition unit is composed of a device including a processor such as a computer, a control unit 24, a storage unit (not shown), a display unit 1b, a data output unit 22 and a data acquisition unit 23, but can also be composed of multiple computers that can send and receive data via a network connection. Fig.14The physical quantity acquisition part control unit 24, the storage unit (not shown), the display unit 1b, the data output unit 22, and the data acquisition unit 23 of the measuring device 19 in the figure can be electrically connected, physically connected, or connected via a network to constitute the measuring device 19. The physical quantity acquisition part control unit 24, the storage unit (not shown), the display unit 1b, the data output unit 22, and the data acquisition unit 23 of the measuring device 19 can be configured so that part of it is included in one computer, and the rest is included in different computers. For example, the storage unit of the measuring device 19 can be located in another computer connected to the computer including the physical quantity acquisition part control unit 24, the display unit 1b, the data output unit 22, and the data acquisition unit 23 through a network. In addition, the processing performed by the physical quantity acquisition part control unit 24 can be performed by distributed processing of multiple computers. At this time, the function of the control unit 3 can be realized by the processors of the multiple computers that cooperate. For example, the processing of the determination performed by the determination unit 8 can be performed by the computer that performs the calculation processing performed by the feature quantity calculation unit 6 and the separation amount calculation unit 7 and another computer.

[0267] The physical quantity acquisition unit control unit 24 , the data output unit 22 , the physical quantity calculation unit 29 b , and the display unit 1 b may be configured by a general-purpose computer, a desktop PC, or the like.

[0268] In addition, the storage unit (not shown) of the measuring device 19 can store programs. In addition, the physical quantity acquisition unit control unit 24 can also store and keep the physical quantity extracted as the measured value in the storage unit. The storage unit of the measuring device is one or more memories. The memory is, for example, a semiconductor memory, a magnetic memory, or an optical memory, but is not limited to them, and can be set to any memory. The storage unit of the measuring device is, for example, built into the same device as the physical quantity acquisition unit control unit 24, but can also be configured to be accessed by the physical quantity acquisition unit control unit 24 via any interface.

[0269] [Variation 3]

[0270] As including Fig.14 Another modification of the measurement system of the computing device 12 and the measuring device 19 is shown. Fig.15 In addition, the function Figure 1 , Fig.13 , Fig.14 The same parts are marked with the same reference numerals, and detailed descriptions are omitted. Fig.15 The difference is that the control unit 3 in the arithmetic device 12 further includes a physical quantity calculation unit 29b that calculates the physical quantity as a measurement value based on the signal acquired by the physical quantity measurement unit 2a.

[0271] In this case, the signal related to the physical quantity measured by the physical quantity measuring unit 2a is sent from the data output unit 22 to the data acquisition unit 21 of the computing device 12 via the network through the physical quantity acquisition unit control unit 24. The control unit 3 temporarily stores the sent signal from the data acquisition unit 21 in the storage unit 9. In addition, the control unit 3 calculates the measured physical quantity based on the sent signal and stores it in the measured physical quantity group 26 of the storage unit 9. In addition, the control unit 3 calculates the characteristic quantity and the degree of separation using the measured physical quantity.

[0272] Furthermore, a signal related to the physical quantity measured by the physical quantity measuring unit 2 a may be stored in a storage unit (not shown) in the measuring device 19 via the physical quantity acquisition unit control unit 24 .

[0273] The physical quantity calculation unit 29a of the measuring device 19 may not be provided. In this case, the control unit 3 of the arithmetic device 12 calculates the measured physical quantity as an input value based on the signal related to the physical quantity measured by the physical quantity measurement unit 2a.

[0274] In the case of variation 3, the distance between the computing device 12 and the measuring device 19 is also very far, and this example is particularly preferred when there are different managers for the computing device 12 and the measuring device 19. In addition, when the computing device 12 and the measuring device 19 are configured in different countries and used in different companies, this example is further preferred.

[0275] (In the case where there is no physical quantity acquisition unit 2)

[0276] In addition, Figure 1 In the physical quantity acquisition unit 2, a physical quantity input unit (not shown) may be provided. As a unit capable of manually inputting a physical quantity (input value), any device that can be used as an input terminal device such as a keyboard, a mouse, a tablet computer or a smart phone may be used. The physical quantity input unit is electrically connected, physically connected or connected via a network to the computing device 12. If the physical quantity is input from the physical quantity input unit, the control unit 3 stores and preserves them as measured physical quantities in a specific area of ​​the storage unit 9. Thereafter, the control unit 3 calculates the characteristic quantity and the degree of separation using the input physical quantity as the input value. In this way, the present invention not only uses the computing device 12 as a part of the measurement system 20, but can also be used as a simple characteristic quantity calculation device and a separation quantity calculation device.

[0277] [Variation 4]

[0278] As Fig.15 A further modification example (modification example 4) of Fig. 20In the case of variant example 3, the distance between the computing device 12 and the measuring device 19 is also very far, but when the computing device 12 and the measuring device 19 have different managers, the configuration example of variant example 4 is preferred. In particular, when the computing device 12 and the measuring device 19 are configured in different countries or companies, the configuration example of variant example 4 is preferred. Here, with respect to variant example 4, in order to avoid repeated description, only the parts that are different from variant example 3 are described. In addition, in order to distinguish it from the second computing device 12s, in the following description, the computing device 12 is sometimes referred to as the first computing device.

[0279] A second computing device 12s is provided inside or near the measuring device 19. The second computing device 12s includes a control unit 3s and a storage unit 9s. The control unit 3s includes a data acquisition unit 21s, a first model classification processing unit 4s, and a feature quantity calculation unit 6s. In addition, the storage unit 9s includes a first model group 10s, a measured physical quantity group 26s, and a calculated feature quantity group 27s. These components have the same functions as the components of the same name of the first computing device.

[0280] The physical quantity is sent from the data output unit 22 of the control unit 24 of the physical quantity acquisition unit to the second computing device 12s and the computing device 12 in the measuring device 19. Since the same data as that of the second computing device 12s is also sent to the feature quantity calculation unit 6 of the computing device 12, the feature quantity calculation unit 6s of the second computing device 12s and the feature quantity calculation unit 6 of the computing device 12 perform the same output. In addition, the degree of separation is output by the separation quantity calculation unit 7 and the second model group 11 of the computing device 12. Therefore, for example, the manager who manages the second computing device 12s can simulate the appropriateness of the first model in the feature quantity calculation unit 6s. Then, the manager of one side can create new teacher data based on the input value with a high degree of separation and the corresponding output value, and use the teacher data to create a new first model. After creating the new first model, the new first model can be provided to other managers via the network. Here, in the above-mentioned embodiment 3, an example of outputting a hardness prediction reference offset is shown. For example, based on the hardness prediction reference offset, it is possible to carry out a work collaboration such as obtaining the information required by the manager of one side from other managers (for example, obtaining the true value by measurement) while promoting the creation of the required first model.

[0281] (Other methods for generating the second model)

[0282] As a method for generating the second model, the least squares ridge regression is used in the above-mentioned embodiment, but it is not limited to this. For example, multivariate regression analysis such as least squares Lasso regression can also be used. As a method for generating the second model, in addition to this, any method can be used as long as it is a method of predicting and calculating one variable based on multiple variables. For example, the naive Bayes method (naive Bayes method) and the Kullback-Leibler information volume that perform binary judgment based on probability calculation can be used. For example, the KNN / LOF method used in anomaly detection, etc. to calculate the distance of the data from the first model can be used. In addition, for example, a neural network as one of deep learning can be used. However, it is more preferable to have a method that can perform grade division in order to judge one measurement point and one steel plate with higher accuracy, that is, a method that is not a binary judgment but can output a continuous value.

[0283] Here, as the output value (degree of detachment) from the second model, a value can be used directly instead of the standard deviation. In this case, the model is made without performing standardization processing at the stage of pre-processing. Here, in the case of binarization, the part where the degree of detachment becomes above a certain threshold can be displayed in black, for example, and the part where it becomes less than the threshold can be displayed in white, for example. Through binarization, it is easy for the operator to see and manage. Whether the output value from the second model is the standard deviation or the value is used directly, binarization can be performed.

[0284] In addition, with respect to the structure of the measurement system 20, although Figure 1 , Figures 13-15 The examples are given in the following, but the following modifications and usages are also possible. The modifications and usages listed below can be combined for application. For example, Figure 1 In the embodiment, the computing device 12 and the physical quantity acquisition unit 2 can be arranged in the scanning unit 25. The computing device 12 and the physical quantity acquisition unit 2 can be integrated and both controlled by the control unit 3. The display unit 1 can display information about the physical quantity, characteristic quantity, and degree of separation to the operator. For example, Fig.13In the embodiment, the operation device 12 can be set at a distance relative to the physical quantity acquisition unit 2. The display unit 1b and the scanning unit 25 of the physical quantity acquisition unit 2 can be integrated. In addition, the operation manager can be near the physical quantity acquisition unit 2. The connection between the operation device 12 and the physical quantity acquisition unit 2 can include a network. The measurement system 20 can be implemented by an online device assembled on a production line. The physical quantity acquisition unit 2 including the scanning unit 25 can be controlled by the control unit 3 of the operation device 12. The process manager can be near the operation device 12. The operation device 12 and the physical quantity acquisition unit 2 can be integrated and both controlled by the control unit 3. The display unit 1a can display at least information related to the degree of separation (such as physical quantities, characteristic quantities, and information on the degree of separation) to the operator. The display unit 1b can display at least information related to the degree of separation (such as physical quantities, characteristic quantities, and information on the degree of separation) to the operator. The display unit 1c can be set in a place different from the display unit 1a and the display unit 1b, and display the required information according to the purpose of the setting. For example, in Fig.14 In the embodiment, the physical quantity acquisition unit 2 and the computing device 12 can be connected via a network. The implementation country of the physical quantity acquisition unit 2 and the implementation country of the computing device 12 may be different. The implementer of the physical quantity acquisition unit 2 and the implementer of the computing device 12 may be different. The data acquisition unit 21 may have an output function of outputting data. The data output unit 22 may have an acquisition function of acquiring data. The physical quantity acquisition unit control unit 24 and the display unit 1b and other components of the measuring device 19 may be implemented by a computer having a CPU, a memory, an HDD, a storage device, a monitor, a keyboard, a mouse, etc. The functions of each unit (including parts not shown in the figure) can be implemented by executing processing by the physical quantity acquisition unit control unit 24. The display unit 1a and the display unit 1b can be configured according to the configuration or work content of the operator or process manager. The measurement system 20 can be implemented by an online device assembled on a production line. The scanning unit 25 and the physical quantity measurement unit 2a can be controlled by the physical quantity acquisition unit control unit 24. The physical quantity calculated by the physical quantity acquisition control unit 24 can be output from the data output unit 22, sent to the data acquisition unit 21 via the network, and stored in the computing device 12. Fig.15 In the process, the signal can be measured by the measuring device 19, and the signal can be sent to the computing device 12, and the physical quantity can be calculated by the physical quantity calculation unit 29a of the control unit 3. Alternatively, the signal can be measured by the measuring device 19, and the physical quantity can be calculated by the physical quantity calculation unit 29a, and the physical quantity can be sent to the computing device 12.

[0285] In addition, as described in the above-mentioned embodiment, the calculation method and the calculation device 12 are used for the manufacture of products, and are not limited to steel materials. Here, as a method for the manufacture of products, the application in the manufacturing equipment for manufacturing materials is mainly described, but the calculation method and the calculation device 12 can be used in the inspection and maintenance related to the manufacturing. Therefore, the calculation method and the calculation device 12 are used for the manufacture or use (including inspection, maintenance, etc.) of products.

[0286] Here, the model addition in the first model group 10 and the second model group 11 is supplemented. In the present embodiment, the following method is adopted: after determining whether the input value is appropriate relative to the machine learning model used, a new teacher data group is made based on the input value determined to be inappropriate, and a new first model is made. By performing such a process, data that cannot be accurately predicted by the existing first model can also be predicted with high accuracy. For example, in the prior art (refer to Japanese Patent Publication No. 2017-187820 as a reference), the active model used for the current abnormal diagnosis is covered for updating. The method of the prior art is that the number of abnormal diagnosis models at any time is a fixed number (no increase or decrease), so it is useful from the perspective of suppressing the increase in the amount of data. However, covering the used model makes it difficult to reproducibly determine the same conditions in the past. In particular, in the manufacture of products, from the perspective of specific causes such as operational stability and defective generation, it is hoped that the models used in the past will be retained as much as possible. In the calculation method and calculation device of the present embodiment, in addition to the first model that can function as an abnormal diagnosis model, a second model that outputs the deviation of the first model from the measured value (input value) as the amount of separation is also used. In the case where the input value is not suitable for the first model group 10 currently in use, a new teacher data group is created based on the input value determined to be inappropriate, and a new first model is created. The newly created first model is added to the first model group 10. The number of first models constituting the first model group 10 changes over time, and the first model group 10 is not covered as a whole. Here, in order to suppress the increase in the amount of data, the first model with less frequency of use can be deleted from the first model group 10. In this way, it is possible to cope with the increase in the amount of data.

[0287] Description of Reference Numerals

[0288] 1…display unit; 1a…display unit (inside the computing device); 1b…display unit (inside the physical quantity measuring unit or in the measuring device); 1c…display unit (other); 2…physical quantity acquiring unit; 2a…physical quantity measuring unit; 3…control unit; 3s…control unit (inside the measuring device 19); 4…classification processing unit for the first model; 4s…classification processing unit for the first model (inside the measuring device 19); 5…classification processing unit for the second model; 6…feature quantity calculating unit; 6s…feature quantity calculating unit (inside the measuring device 19); 7…separation amount calculating unit; 8…judgment unit; 9…storage unit; 9s…storage unit (inside the measuring device 19); 10…first model group; 10s…first model group (inside the measuring device 19); 11…second model group; 12…computing device (first computing device); 12s…second computing device ; 13…signal transceiver; 14…sensor; 15…film; 16…substance; 17…excitation coil; 18…magnetization yoke; 19…measuring device; 20…measuring system; 21…data acquisition unit (inside the computing device 21); 21s…data acquisition unit (inside the measuring device 19); 22…data output unit (inside the measuring device 19); 24…control unit for physical quantity acquisition unit (inside the measuring device 19); 25…scanning unit; 26…measured physical quantity group; 26s…measured physical quantity group (inside the measuring device 19); 27…calculated characteristic quantity group; 27s…calculated characteristic quantity group (inside the measuring device 19); 28…calculated separation quantity group; 29a…physical quantity calculation unit (inside the control unit 3 in the computing device 12); 29b…physical quantity calculation unit (inside the control unit 24 for the physical quantity acquisition unit).

Claims

1. A method of operation for use in the manufacture or use of a product, The computing method is characterized by comprising: a feature quantity calculation step in which the feature quantity is calculated using one or more input values ​​selected from a predetermined input value group and one or more first models; and a deviation amount calculation step, in which a deviation amount for an input value predetermined from the input value group is calculated using one or more input values ​​selected from the input value group and one or more second models, the deviation amount being a deviation from the first model, The second model and the first model are machine learning models generated using one or more teacher data selected from a predetermined teacher data group.

2. The calculation method according to claim 1, characterized in that: A determination step is provided in which the separation amount is compared with a preset threshold value to determine whether the first model to be used is acceptable or not or the level of the separation amount.

3. A method for manufacturing a product, characterized in that: Have the manufacturing steps to manufacture the product, In the manufacturing step, the calculation method according to claim 1 or 2 is performed.

4. A product management method, characterized in that: have: a manufacturing step in which a product is manufactured and the operation method according to claim 1 or 2 is performed; and A management step in which the products are classified based on at least one of the feature amount and the amount of separation.

5. A computing device used in the manufacture or use of a product, The computing device is characterized in that it has: a feature quantity calculation unit that calculates a feature quantity using one or more input values ​​selected from a predetermined input value group and one or more first models; and a deviation amount calculation unit that calculates a deviation amount for an input value predetermined from the input value group using one or more input values ​​selected from the input value group and one or more second models, the deviation amount being a deviation amount from the first model, The second model and the first model are machine learning models generated using one or more teacher data selected from a predetermined teacher data group.

6. The computing device according to claim 5, characterized in that: A determination unit is provided for comparing the deviation amount with a preset threshold value to determine whether the first model to be used is acceptable or not or the level of the deviation amount.

7. The computing device according to claim 5 or 6, characterized in that: A data acquisition unit is provided for acquiring one or two of an input value of the first model and a signal measured to obtain the input value of the first model.

8. A manufacturing device for a product, characterized in that: have: Manufacturing equipment where products are manufactured; and The computing device according to any one of claims 5 to 7.

9. A measurement method, characterized in that: A measuring step of measuring a signal related to a physical quantity of a measurement object is provided. Execute the operation method described in claim 1 or 2, in which the physical quantity obtained based on the measured signal is used as the one or more input values, and the characteristic quantity calculated in the characteristic quantity calculation step is used as the measurement value.

10. A method for manufacturing a product, characterized in that: Have the manufacturing steps to manufacture the product, The measuring method according to claim 9 is included in the manufacturing steps.

11. A product management method, characterized in that: have: a manufacturing step in which a product is manufactured and includes the measuring method of claim 9; and A management step of classifying the manufactured products based on at least one of the feature amount and the separation amount.

12. A measurement system, characterized in that: have: a physical quantity measuring unit that measures a signal related to a physical quantity of a measurement object; and The computing device according to claim 5 or 6, wherein the computing device uses the physical quantity obtained based on the measured signal as the one or more input values, and uses the feature quantity calculated by the feature quantity calculation unit as the measurement value.

13. A manufacturing device for a product, characterized in that: have: Manufacturing equipment where products are manufactured; and The measurement system of claim 12.

14. A measuring device, characterized in that: have: Data output unit; a physical quantity measuring unit that measures a signal related to a physical quantity of a measurement object; and a physical quantity acquisition unit control unit, the physical quantity acquisition unit control unit executing a process of outputting the signal measured to obtain one or more input values ​​to an external computing device, The computing device comprises: a feature quantity calculation unit that calculates a feature quantity using one or more input values ​​and one or more first models; and A deviation amount calculation unit calculates a deviation amount for the input value using one or more input values ​​selected from the input values ​​and one or more second models, the deviation amount being a deviation amount from the first model.

15. The measuring device according to claim 14, characterized in that A physical quantity calculation unit is provided, wherein the physical quantity calculation unit calculates the physical quantity as the measured physical quantity based on the measured signal, The physical quantity acquisition unit control unit executes a process of outputting the measured physical quantity as the one or more input values ​​instead of the signal measured to obtain the one or more input values.

16. The measuring device according to claim 14 or 15, characterized in that The physical quantity acquisition unit control unit executes a process of acquiring at least one of the information related to the calculated feature quantity and the information related to the calculated separation amount from the calculation device.

17. A method for producing teacher data, characterized in that: As the performance input values, a predetermined number or more of the input values ​​of the level of the separation amount that are judged as unqualified or worse than a predetermined standard based on the pass / fail of the first model or the level of the separation amount calculated by the calculation method according to claim 2 are selected; The selected actual performance input value is assigned a characteristic value actually measured as an actual performance output value and a label indicating the level of the deviation amount that is unacceptable or worse than a predetermined standard, This becomes the teacher data for generating the first model.

18. A teacher data, which is used to generate a first model, The teacher data is characterized in that having a predetermined number or more of actual input values ​​of the first model, For each of the performance input values, there is a characteristic value actually measured as a performance output value, and a label indicating whether the first model is qualified or not or the level of deviation is marked according to the calculation method described in claim 2, indicating that the level of deviation is unqualified or worse than a predetermined benchmark.

19. A method for generating a model, characterized in that: Using the teacher data according to claim 18, A learned model is generated by machine learning, wherein the learned model takes one or more input values ​​selected from a predetermined input value group as input values ​​and takes a feature quantity having a correlation that can be derived from the input values ​​as an output value.

20. A program for a computing device used in the computing device according to any one of claims 5 to 7, The program is characterized in that A computer is caused to execute the functions of the feature amount calculation unit and the deviation amount calculation unit.

21. A storage medium storing a program for the computing device according to any one of claims 5 to 7, The storage medium is characterized in that A program for a computing device is stored which causes a computer to execute the functions of the feature amount calculation unit and the deviation amount calculation unit.

22. A program for use in the measuring device according to any one of claims 14 to 16, The program is characterized in that The computer is made to execute the functions of the arithmetic device and the data output unit.

23. A storage medium storing a program for the measuring device according to any one of claims 14 to 16, The storage medium is characterized in that A program for causing a computer to execute the functions of the arithmetic device and the data output unit is stored.

Images