Material quality estimation model creation method, material quality estimation method, material quality estimation model creation device, material quality estimation device, and material quality estimation model creation system

Abstract

A material quality estimation model creation method according to the present invention comprises: a material quality estimation base model creation step in which a model creation unit provided to a computer performs machine learning using existing manufacturing condition data as input and existing quality data associated with the existing manufacturing condition data as output to create a material quality estimation base model; and a material quality estimation model retraining step in which the model creation unit is retrained on the material quality estimation base model using additional manufacturing condition data as input and additional quality data associated with the additional manufacturing condition data as output to create a material quality estimation model.

Description

Material quality estimation model creation method, material quality estimation method, material quality estimation model creation device, material quality estimation device, and material quality estimation model creation system

[0001] The present invention relates to a material quality estimation model creation method, a material quality estimation method, a material quality estimation model creation device, a material quality estimation device, and a material quality estimation model creation system.

[0002] Patent Literature 1 discloses a quality prediction system that predicts the quality of a target product based on the similarity between a learning prediction model and a quality prediction model derived from the actual operating conditions of the target product. This system applies a machine learning algorithm to build the prediction model and estimates the probability of defect occurrence with high accuracy.

[0003] Furthermore, Patent Document 2 discloses a prediction system that receives input data including output factors of metal plate manufacturing equipment and component values ​​of the metal plate being manufactured and includes a prediction model that predicts material property values ​​of the manufactured metal plate. This system applies a machine learning algorithm and estimates material property values ​​with high accuracy by using a model that receives input data and outputs manufacturing condition factors and a model that receives the manufacturing condition factors as input and outputs material property values.

[0004] Patent No. 6953990 JP 2022-48038 A

[0005] However, the material quality estimation models disclosed in Patent Documents 1 and 2 are models that are applied to materials that are manufactured in actual manufacturing facilities under manufacturing conditions that are the same as or similar to those of already-manufactured materials. In other words, the material quality estimation models disclosed in Patent Documents 1 and 2 are generally created on the assumption that a large amount of manufacturing condition data and quality data exists.

[0006] On the other hand, in the development of new manufacturing materials, it is common to conduct preliminary experiments in experimental equipment to verify whether the material manufactured under newly designed manufacturing conditions satisfies the desired quality. The above-mentioned preliminary experiments are usually conducted by carefully designing and controlling only certain types of manufacturing conditions that are expected to significantly contribute to determining the quality of the material. Therefore, preliminary experiments rarely take into account other manufacturing conditions, including environmental conditions and disturbance conditions that only exist when manufacturing in actual manufacturing equipment. Furthermore, because preliminary experiments are typically conducted only a small number of times, the amount of experimental data that can be obtained is also limited.

[0007] When new production is carried out using actual production equipment based on the production conditions designed in the above-mentioned preliminary experiments, there are many cases where the desired material quality cannot be obtained due to differences from the actual production that could not be fully taken into account in the preliminary experiments.In addition, when starting to produce a new material, it is not easy to create an accurate material quality estimation model because there is not enough experimental condition data and quality data from the preliminary experiments, as mentioned above.

[0008] The present invention has been made in consideration of the above, and aims to provide a material quality estimation model creation method, a material quality estimation method, a material quality estimation model creation device, a material quality estimation device, and a material quality estimation model creation system that are capable of creating a model that estimates the quality of newly manufactured materials with high accuracy.

[0009] In order to solve the above-mentioned problems and achieve the objective, the method for creating a material quality estimation model according to the present invention includes a material quality estimation base model creation step in which a model creation unit provided in a computer takes existing manufacturing condition data as input and performs machine learning using existing quality data linked to the existing manufacturing condition data as output to create a material quality estimation base model, and a material quality estimation model re-learning step in which the model creation unit takes additional manufacturing condition data as input and additional quality data linked to the additional manufacturing condition data as output to re-learn the material quality estimation base model to create a material quality estimation model.

[0010] In the method for creating a material quality estimation model according to the present invention, in the above invention, the existing manufacturing condition data is data relating to an existing manufacturing material, and the additional manufacturing condition data is data relating to a new manufacturing material.

[0011] In the method for creating a material quality estimation model according to the present invention, in the material quality estimation model re-learning step, if the types of additional manufacturing condition data are insufficient compared to the types of existing manufacturing condition data that are input to the material quality estimation base model, the model creation unit adds types of additional manufacturing condition data to match the types of existing manufacturing condition data, and re-learns the material quality estimation base model.

[0012] In the method for creating a material quality estimation model according to the present invention, in the material quality estimation model re-learning step, the model creation unit adds, for each missing type of additional manufacturing condition data that is missing from the existing manufacturing condition data, by complementing it using statistics of the corresponding type of existing manufacturing condition data.

[0013] In the method for creating a material quality estimation model according to the present invention, in the material quality estimation model re-learning step, the model creation unit adds the additional manufacturing condition data types that are missing from the existing manufacturing condition data by complementing each missing type with a sample value randomly extracted from the corresponding type of the existing manufacturing condition data.

[0014] In the method for creating a material quality estimation model according to the present invention, in the material quality estimation model re-learning step, the model creation unit locally extracts only manufacturing condition data similar to the additional manufacturing condition data from the set of existing manufacturing condition data for types of manufacturing condition data that already match between the existing manufacturing condition data and the additional manufacturing condition data, and adds types of additional manufacturing condition data that are missing from the existing manufacturing condition data by complementing them with statistics of the corresponding types of existing manufacturing condition data.

[0015] In the method for creating a material quality estimation model according to the present invention, in the material quality estimation model re-learning step, the model creation unit locally extracts only manufacturing condition data similar to the additional manufacturing condition data from the set of existing manufacturing condition data for types of manufacturing condition data that already match between the existing manufacturing condition data and the additional manufacturing condition data, and adds types of additional manufacturing condition data that are missing from the existing manufacturing condition data by complementing them with sample values ​​randomly extracted from the corresponding types of the existing manufacturing condition data.

[0016] In the method for creating a material quality estimation model according to the present invention, in the material quality estimation model re-learning step, if the types of the additional manufacturing condition data are insufficient compared to the types of the existing manufacturing condition data that are input to the material quality estimation base model, the model creation unit reduces the types of the existing manufacturing condition data so that they match the types of the existing manufacturing condition data, and creates the material quality estimation base model.

[0017] In the method for creating a material quality estimation model according to the present invention, in the above invention, the method for relearning the material quality estimation base model is transfer learning.

[0018] In the method for creating a material quality estimation model according to the present invention, in the above invention, the method for relearning the material quality estimation base model is fine tuning.

[0019] In order to solve the above-mentioned problems and achieve the objectives, the material quality estimation method of the present invention includes a quality estimation step in which a quality estimation unit provided in a computer estimates the quality of a manufactured product based on a material quality estimation model created by the above-mentioned method for creating a material quality estimation model.

[0020] In order to solve the above-mentioned problems and achieve the objectives, the material quality estimation model creation device of the present invention includes a model creation unit that takes existing manufacturing condition data as input and performs machine learning using existing quality data linked to the existing manufacturing condition data as output to create a material quality estimation base model, and the model creation unit takes additional manufacturing condition data as input and additional quality data linked to the additional manufacturing condition data as output to re-learn the material quality estimation base model to create a material quality estimation model.

[0021] In order to solve the above-mentioned problems and achieve the objectives, the material quality estimation device of the present invention includes a quality estimation unit that estimates the quality of a manufactured product based on a material quality estimation model created by the above-mentioned material quality estimation model creation device.

[0022] In order to solve the above-mentioned problems and achieve the objectives, the material quality estimation model creation system of the present invention includes a model creation device that takes existing manufacturing condition data as input and performs machine learning using existing quality data linked to the existing manufacturing condition data as output to create a material quality estimation base model, and the model creation device takes additional manufacturing condition data as input and additional quality data linked to the additional manufacturing condition data as output to re-learn the material quality estimation base model to create a material quality estimation model.

[0023] In the method, device, and creation of a material quality estimation model according to the present invention, a machine learning model based on existing manufacturing condition data is used as a base model, and re-learning is performed using preliminary experimental data (additional manufacturing condition data) for a new manufacturing material. This makes it possible to create a model that can estimate material quality with high accuracy even when sufficient manufacturing data is not available when actually manufacturing a new manufacturing material.

[0024] The material quality estimation method and material quality estimation device according to the present invention use the model created as described above to estimate the variation in material quality when variations in manufacturing conditions for a new manufacturing material are assumed in advance. Furthermore, the material quality estimation method and material quality estimation device according to the present invention enable material design of manufacturing conditions that do not deviate from the allowable range of quality variations. This allows for efficient material development.

[0025] FIG. 1 is a diagram showing the schematic configuration of a material quality estimation model creation device and an information processing device realizing the material quality estimation device according to an embodiment of the present invention. FIG. 2 is a diagram for explaining an overview of transfer learning in the material quality estimation model creation device according to an embodiment of the present invention. FIG. 3 is a diagram for explaining an example of a case in which transfer learning is performed by adding additional manufacturing condition data types when the number of additional manufacturing condition data types is insufficient compared to the number of existing manufacturing condition data types that are input to the material quality estimation base model in the material quality estimation model creation device according to an embodiment of the present invention. FIG. 4 is a flowchart showing the steps of a material quality estimation model creation method executed by the material quality estimation model creation device according to an embodiment of the present invention.

[0026] A method for creating a material quality estimation model (hereinafter referred to as a "model creation method"), a material quality estimation method, a device for creating a material quality estimation model (hereinafter referred to as a "model creation device"), and a material quality estimation device according to embodiments of the present invention will be described with reference to the drawings. Note that the components in the following embodiments include those that are easily replaceable by a person skilled in the art, or those that are substantially identical.

[0027] (Device Configuration) Fig. 1 shows an example of the configuration of an information processing device 1 for realizing a model creation device and a material quality estimation device according to an embodiment. The information processing device 1 includes an input unit 10, a storage unit 20, a calculation unit 30, and a display unit 40. The model creation device according to an embodiment is realized by the components of the information processing device 1 excluding the quality estimation unit 33 of the calculation unit 30. The material quality estimation device according to an embodiment is realized by all the components of the information processing device 1.

[0028] The input unit 10 is an input means for the calculation unit 30, and is realized by an input device such as a keyboard, a mouse pointer, or a numeric keypad.

[0029] The storage unit 20 is composed of storage media such as an erasable programmable read only memory (EPROM), a hard disk drive (HDD), and removable media. Examples of removable media include disk storage media such as a universal serial bus (USB) memory, a compact disc (CD), a digital versatile disc (DVD), and a Blu-ray (registered trademark) disc (BD). The storage unit 20 can store an operating system (OS), various programs, various tables, various databases, and the like. The storage unit 20 also stores an operation database (DB) 21.

[0030] The operation DB 21 stores existing manufacturing condition data, existing quality data, additional manufacturing condition data, and additional quality data. The existing manufacturing condition data and existing quality data are data used in actual equipment, and there is a large amount of data. On the other hand, the additional manufacturing condition data and additional quality data are data used in experimental equipment, for example, when developing new manufacturing materials, and there is generally a limited amount of data.

[0031] The existing manufacturing condition data indicates existing data obtained by measuring the manufacturing state during product manufacturing or by estimating it by some other means. This existing manufacturing condition data is preferably data that varies depending on the position within the product.

[0032] The existing quality data refers to existing data obtained by measuring the quality of a product or estimating it by some means. The method of measurement or estimation of this existing quality data varies depending on the type of quality, but it is preferable that the data is data at a specific position within the product.

[0033] The additional manufacturing condition data indicates additional data obtained by measuring the manufacturing state during product manufacturing or by estimating it by some other means. This additional manufacturing condition data is preferably data that varies depending on the position within the product.

[0034] The additional quality data refers to additional data obtained by measuring or estimating the quality of a product. The method of measurement or estimation of this additional quality data varies depending on the type of quality, but it is preferably data at a specific position within the product.

[0035] The calculation unit 30 is realized by a processor such as a CPU (Central Processing Unit) and a memory (main storage unit) such as a RAM (Random Access Memory) or a ROM (Read Only Memory).

[0036] The calculation unit 30 loads a program into the working area of ​​the main storage unit, executes it, and controls each component unit through the execution of the program, thereby realizing functions that meet a predetermined purpose. Through the execution of the program, the calculation unit 30 functions as a data creation unit 31, a model creation unit 32, and a quality estimation unit 33. Note that while Fig. 1 shows an example in which the functions of each unit are realized by a single computer, the specific method for realizing the functions of each unit is not particularly limited, and for example, the functions of each unit may be realized by multiple computers.

[0037] The data creation unit 31 creates data necessary for creating a material quality estimation base model, which will be described later. That is, the data creation unit 31 collects existing manufacturing condition data and existing quality data from the operation DB 21. Next, the data creation unit 31 links the existing manufacturing condition data with the existing quality data to create existing manufacturing data. As described above, this existing manufacturing data is data related to existing manufacturing materials.

[0038] Specifically, the data creation unit 31 aggregates and combines the collected existing manufacturing condition data and existing quality data. That is, the data creation unit 31 links data obtained by measuring or estimating the physical state of the product with data obtained by measuring or estimating the quality of the product. Here, in creating the existing manufacturing data, it is desirable to link a specific position on the product where the existing quality data was measured or estimated with the existing manufacturing condition data measured or estimated at the same position.

[0039] The data creation unit 31 also creates data necessary for creating a material quality estimation model, which will be described later. That is, the data creation unit 31 collects additional manufacturing condition data and additional quality data from the operation DB 21. Next, the data creation unit 31 links the additional manufacturing condition data with the additional quality data to create additional manufacturing data. As described above, this additional manufacturing data is data related to new manufacturing materials.

[0040] Specifically, the data creation unit 31 aggregates and combines the collected additional manufacturing condition data and additional quality data. That is, the data creation unit 31 links data obtained by measuring or estimating the physical state of the product with data obtained by measuring or estimating the quality of the product. Here, when creating the additional manufacturing data, it is desirable to link a specific position on the product where the additional quality data was measured or estimated with the additional manufacturing condition data measured or estimated at the same position.

[0041] The model creation unit 32 creates a material quality estimation base model using existing manufacturing data. That is, the model creation unit 32 performs machine learning using existing manufacturing condition data as input and existing quality data linked to the existing manufacturing condition data as output to create a material quality estimation base model that represents the relationship between manufacturing conditions and quality. A machine learning algorithm having a neural network structure is used to train this material quality estimation base model.

[0042] If existing manufacturing data is used as manufacturing data for already manufactured materials at actual manufacturing facilities, it is often possible to secure a sufficient amount of data for machine learning, which makes it possible to create a model with high estimation accuracy, i.e., high generalization performance, even for unknown data.

[0043] The model creation unit 32 also creates a material quality estimation model using the material quality estimation base model and the additional manufacturing data. That is, the model creation unit 32 receives the additional manufacturing condition data as input, receives the additional quality data linked to the additional manufacturing condition data as output, and performs re-learning of the material quality estimation base model to create a material quality estimation model for the additional manufacturing data.

[0044] For example, transfer learning is used for re-learning. Figure 2 shows an overview of transfer learning. In transfer learning, generally, a certain input X = [X １ , X ２ , …, X ｍ , …, X Ｍ ], output Y Ａ A model M that estimates with high accuracy Ａ When the model M Ａ The intermediate layer is reused from the input layer of the input matrix, and only the linear layer A close to the output layer is replaced with another linear layer B. Then, the input X = [X １ , X ２ , …, X ｍ , …, X Ｍ ], another output Y Ｂ The output is Y Ｂ A model M that estimates with high accuracy Ｂ Create a.

[0045] Here, the linear layer is also called a fully connected layer. In addition, in transfer learning, the already learned model M Ａ In order to reuse the input layer and intermediate layer of the model M Ｂ It is known that a highly accurate model can be created even if the amount of data used for learning is smaller than that of general supervised learning. In this embodiment, if the additional manufacturing data is manufacturing data from a preliminary experiment on a new manufacturing material, it may not be possible to secure a sufficient amount of data for general machine learning. Even if the additional manufacturing data is limited to a small amount, a highly accurate material quality estimation model for the additional manufacturing data can be created by using transfer learning.

[0046] Furthermore, if the types of additional manufacturing condition data are insufficient compared to the types of existing manufacturing condition data that are input to the material quality estimation base model, the model creation unit 32 adds types of additional manufacturing condition data so that the types of additional manufacturing condition data match the types of existing manufacturing condition data, and re-learns the material quality estimation base model. In this case, the model creation unit 32 adds types of additional manufacturing condition data, for example, as in the following (1) to (4).

[0047] (1) For types of additional manufacturing condition data that are missing from the existing manufacturing condition data, the data is added by complementing each missing type using statistics (e.g., average values) of the corresponding type of existing manufacturing condition data. (2) For types of additional manufacturing condition data that are missing from the existing manufacturing condition data, the data is added by complementing each missing type using sample values ​​randomly extracted from the corresponding type of existing manufacturing condition data. (3) For types of manufacturing condition data that already match between the existing manufacturing condition data and the additional manufacturing condition data, only manufacturing condition data similar to the additional manufacturing condition data is locally extracted from the set of existing manufacturing condition data. Then, for types of additional manufacturing condition data that are missing from the existing manufacturing condition data, the data is added by complementing each missing type using statistics of the corresponding type of existing manufacturing condition data locally extracted. (4) For types of manufacturing condition data that already match between the existing manufacturing condition data and the additional manufacturing condition data, only manufacturing condition data similar to the additional manufacturing condition data is locally extracted from the set of existing manufacturing condition data. Then, for each type of additional manufacturing condition data that is missing from the existing manufacturing condition data, the missing type is supplemented and added using sample values ​​randomly extracted from the corresponding type of locally extracted existing manufacturing condition data.

[0048] In the above (3) and (4), a method for locally extracting only manufacturing condition data similar to the additional manufacturing condition data from the set of existing manufacturing condition data can be, for example, a just-in-time method. In this just-in-time method, for manufacturing condition data of a type that already matches the existing manufacturing condition data and the additional manufacturing condition data, only data near the requested additional manufacturing condition data is locally extracted from the set of existing manufacturing condition data.

[0049] 3 shows a method for supplementing the types of additional manufacturing condition data. For example, the existing manufacturing data may be manufacturing data for already manufactured materials in actual manufacturing equipment, and the additional manufacturing data may be manufacturing data from preliminary experiments on new manufactured materials. In this case, for example, a material quality estimation base model trained using manufacturing data for already manufactured materials is re-trained using some types of manufacturing conditions designed in preliminary experiments on new manufactured materials. This enables re-training that takes into account other manufacturing conditions, including environmental conditions and disturbance conditions that exist only when manufacturing using actual manufacturing equipment.

[0050] In the above (1) to (4), the types of additional manufacturing condition data that are lacking in the existing manufacturing condition data are added, but conversely, the types of existing manufacturing condition data may be reduced. That is, if it is already known at the stage of creating the material quality estimation base model that the types of additional manufacturing condition data are lacking compared to the types of existing manufacturing condition data, the types of existing manufacturing condition data that are input to the material quality estimation base model may be reduced.

[0051] When the types of additional manufacturing condition data are insufficient compared to the types of existing manufacturing condition data that are input to the material quality estimation base model, the model creation unit 32 creates the material quality estimation base model by reducing the types of existing manufacturing condition data so that they match the types of existing manufacturing condition data. By reducing the types of existing manufacturing condition data that are input to the material quality estimation base model in advance in this way, it becomes possible to align the types of additional manufacturing condition data with the types of existing manufacturing condition data when relearning the material quality estimation base model.

[0052] In addition, although the present embodiment has exemplified a method using a general neural network algorithm, the specific algorithm is not limited as long as it is a machine learning algorithm capable of performing transfer learning. In addition, although the present embodiment has exemplified a method using transfer learning, fine tuning may be used instead of transfer learning.

[0053] The quality estimation unit 33 estimates the quality of the manufactured product based on the material quality estimation model created by the model creation unit 32. That is, the quality estimation unit 33 inputs manufacturing condition data of the product whose quality is estimated to the material quality estimation model, and causes the material quality estimation model to output a quality estimation result.

[0054] The display unit 40 is realized by a display device such as an LCD display, a CRT display, etc. Based on the display signal input from the calculation unit 30, the display unit 40 displays, for example, the quality estimation results by the quality estimation unit 33 in the form of characters, figures, etc.

[0055] (Model Creation Method) A model creation method using a model creation device according to an embodiment will be described with reference to FIG. 4. The model creation method includes an existing manufacturing data collection step (step S1), an existing manufacturing data creation step (step S2), and a material quality estimation base model creation step (step S3). The model creation method also includes an additional manufacturing data collection step (step S4), an additional manufacturing data creation step (step S5), and a material quality estimation model re-learning step (step S6). Note that the timing of executing steps S4 and S5 is not particularly limited, and steps S4 and S5 may be executed in parallel with steps S1 and S2.

[0056] In the existing manufacturing data collection step, the data creation unit 31 collects existing manufacturing condition data and existing quality data (step S1).

[0057] In the existing manufacturing data creation step, the data creation unit 31 links the existing manufacturing condition data and the existing quality data to create existing manufacturing data (step S2).

[0058] In the material quality estimation base model creation step, the model creation unit 32 performs machine learning using existing manufacturing condition data as input and existing quality data as output to create a material quality estimation base model (step S3).

[0059] In the additional manufacturing data collection step, the data creation unit 31 collects additional manufacturing condition data and additional quality data (step S4).

[0060] In the additional manufacturing data creation step, the data creation unit 31 links the additional manufacturing condition data and the additional quality data to create additional manufacturing data (step S5).

[0061] In the material quality estimation model re-learning step, the model creation unit 32 re-learns the material quality estimation base model using the additional manufacturing condition data as input and the additional quality data as output, thereby creating a material quality estimation model (step S6).

[0062] (Material quality estimation method) In the formulation quality estimation method according to the embodiment, the quality estimation unit 33 performs a quality estimation step of estimating the quality of the manufactured product based on the material quality estimation model created by the model creation unit 32.

[0063] (Example) An example of the method for creating a material quality estimation model according to the embodiment will be described. In this example, the method for creating a material quality estimation model according to the embodiment is applied to the manufacture of a specific product.

[0064] Specifically, in the development of new manufacturing materials for steel products, machine learning was performed using existing manufacturing condition data from actual manufacturing equipment for already manufactured materials as input and existing quality data from actual manufacturing equipment for already manufactured materials as output to create a base model for estimating material quality. Next, transfer learning was performed from the base model using additional manufacturing condition data from preliminary experiments on new manufacturing materials as input and additional quality data from preliminary experiments on new manufacturing materials as output to create a material quality estimation model for new manufacturing material data.

[0065] In the existing manufacturing data collection step, manufacturing condition data (steel composition) in the steelmaking process and manufacturing condition data (hot rolling temperature, hot rolling speed) in the hot rolling process were collected in the actual manufacturing equipment for the already-manufactured material. Also, in the existing manufacturing data collection step, manufacturing condition data (cold rolling speed) in the cold rolling process and manufacturing condition data (annealing temperature, annealing time) in the annealing process were collected. Also, in the existing manufacturing data collection step, quality data (tensile strength) of the steel plate inspected after the annealing process in the actual manufacturing equipment for the already-manufactured material was collected.

[0066] In the existing manufacturing data creation step, the manufacturing condition data of the already manufactured material and the quality data of the already manufactured material are combined to create the existing manufacturing data. In this step, the leading and trailing end positions of the product where the quality data of the already manufactured material was measured are linked to the manufacturing condition data measured at the same positions or the already manufactured material.

[0067] In the material quality estimation base model creation step, learning was performed using manufacturing condition data for each process of the already manufactured material (steel composition, hot rolling temperature, hot rolling speed, cold rolling speed, annealing temperature, annealing time) as input and quality data (tensile strength) of the already manufactured material as output. Furthermore, in the material quality estimation base model creation step, specifically, learning was performed on a neural network having an input layer, intermediate layer, linear layer, and output layer. This resulted in the creation of a material quality estimation base model.

[0068] In the additional manufacturing data collection and creation step, manufacturing condition data (components in steel) in the steelmaking process, manufacturing condition data (hot rolling temperature) in the hot rolling process, and manufacturing condition data (annealing temperature, annealing time) in the annealing process in the preliminary experiment of the new manufacturing material were collected. Also, in the additional manufacturing data collection and creation step, quality data (tensile strength) of the steel plate inspected after the annealing process in the preliminary experiment of the new manufacturing material was collected.

[0069] In the additional manufacturing data creation step, the manufacturing condition data for the new manufacturing material and the quality data for the new manufacturing material are combined to create additional manufacturing data. In this step, the position of the product where the quality data for the new manufacturing material was measured is linked to the manufacturing condition data measured at the same position or the new manufacturing material.

[0070] In the material quality estimation model re-learning step, various manufacturing condition data for each process of the newly manufactured material (steel composition, hot rolling temperature, annealing temperature, annealing time) were used as input, and the quality data (tensile strength) of the newly manufactured material was used as output, and transfer learning was performed from the material quality estimation base model. In this way, a material quality estimation model for the manufacturing data of the newly manufactured material was created.

[0071] In this example, the types of manufacturing condition data for new manufacturing materials for transfer learning were insufficient compared to the types of existing manufacturing condition data that serve as input to the material quality estimation base model. Therefore, to match the types of conditions in the manufacturing condition data for existing manufacturing materials, the missing hot rolling speeds and cold rolling speeds were added to each type by using the average values ​​of the corresponding types of existing manufacturing materials.

[0072] In this way, in this example, a material quality estimation base model was created using manufacturing data from actual manufacturing equipment for already manufactured materials, and transfer learning was performed from the material quality estimation base model using manufacturing data from a preliminary experiment on a new manufactured material. As a result, a material quality estimation model that can estimate the material quality of a product with high accuracy was created.

[0073] In the method and device for creating a material quality estimation model according to the above-described embodiment, a machine learning model based on existing manufacturing data is used as a base model, and re-learning is performed using preliminary experimental data (additional manufacturing data) for a new manufacturing material. This makes it possible to create a model that can estimate material quality with high accuracy even when sufficient manufacturing data is not available when actually manufacturing a new manufacturing material.

[0074] Furthermore, the material quality estimation method and material quality estimation device according to the embodiment use the model created as described above to estimate the variation in material quality when the variation in manufacturing conditions for a new manufacturing material is assumed in advance. Furthermore, the material quality estimation method and material quality estimation device according to the present invention enable material design of manufacturing conditions that do not deviate from the allowable range of quality variation. This allows for efficient material development.

[0075] The method, device, and system for creating a material quality estimation model according to the present invention have been specifically described above using the preferred embodiments and examples for carrying out the invention, but the scope of the present invention should not be limited to these descriptions and should be broadly interpreted based on the claims. It goes without saying that various changes and modifications based on these descriptions are also included within the scope of the present invention.

[0076] For example, a material quality estimation model creation system may be constructed by configuring each unit of the material quality estimation model creation device in the embodiment with different devices. In this case, the material quality estimation model creation system includes a data creation device having the same function as the data creation unit 31, and a model creation device having the same function as the model creation unit 32.

[0077] REFERENCE SIGNS LIST 1 Information processing device 10 Input unit 20 Storage unit 21 Operation DB 30 Calculation unit 31 Data creation unit 32 Model creation unit 33 Quality estimation unit 40 Display unit

Claims

1. A method for creating a material quality estimation model, comprising: a material quality estimation base model creation step in which a model creation unit provided in a computer performs machine learning with existing manufacturing condition data as input and existing quality data associated with the existing manufacturing condition data as output, and creates a material quality estimation base model; and a material quality estimation model re-learning step in which the model creation unit performs re-learning of the material quality estimation base model with additional manufacturing condition data as input and additional quality data associated with the additional manufacturing condition data as output, and creates a material quality estimation model.

2. The method for creating a material quality estimation model according to claim 1, wherein the existing manufacturing condition data is data related to existing manufacturing materials, and the additional manufacturing condition data is data related to new manufacturing materials.

3. The method for creating a material quality estimation model according to claim 1, wherein in the material quality estimation model re-learning step, when the types of the additional manufacturing condition data are insufficient with respect to the types of the existing manufacturing condition data that are inputs to the material quality estimation base model, the model creation unit adds the types of the additional manufacturing condition data so as to match the types of the existing manufacturing condition data, and performs re-learning of the material quality estimation base model.

4. The method for creating a material quality estimation model according to claim 3, wherein in the material quality estimation model re-learning step, for the types of the additional manufacturing condition data that are insufficient with respect to the existing manufacturing condition data, the model creation unit adds them by complementing each insufficient type using the statistic of the corresponding type of the existing manufacturing condition data.

5. The method for creating a material quality estimation model according to claim 3, wherein in the material quality estimation model re-learning step, for the types of the additional manufacturing condition data that are insufficient with respect to the existing manufacturing condition data, the model creation unit adds them by complementing each insufficient type using sample values randomly extracted from the corresponding type of the existing manufacturing condition data.

6. In the material quality estimation model re-learning step, the model creation unit locally extracts only the manufacturing condition data similar to the additional manufacturing condition data from the set of the existing manufacturing condition data in the types of manufacturing condition data that already match between the existing manufacturing condition data and the additional manufacturing condition data, and for the types of the additional manufacturing condition data that are lacking in the existing manufacturing condition data, adds them by complementing each lacking type using the corresponding statistic of the existing manufacturing condition data. The method for creating a material quality estimation model according to claim 3.

7. In the material quality estimation model re-learning step, the model creation unit locally extracts only the manufacturing condition data similar to the additional manufacturing condition data from the set of the existing manufacturing condition data in the types of manufacturing condition data that already match between the existing manufacturing condition data and the additional manufacturing condition data, and for the types of the additional manufacturing condition data that are lacking in the existing manufacturing condition data, adds them by complementing each lacking type using the sample values randomly extracted from the corresponding types of the existing manufacturing condition data. The method for creating a material quality estimation model according to claim 3.

8. In the material quality estimation model re-learning step, when the types of the additional manufacturing condition data are lacking compared to the types of the existing manufacturing condition data that are inputs to the material quality estimation base model, the model creation unit reduces the types of the existing manufacturing condition data so as to match the types of the existing manufacturing condition data, and creates the material quality estimation base model. The method for creating a material quality estimation model according to claim 1.

9. The method for re-learning the material quality estimation base model is transfer learning. The method for creating a material quality estimation model according to any one of claims 1 to 8.

10. The method for re-learning the material quality estimation base model is fine-tuning. The method for creating a material quality estimation model according to any one of claims 1 to 8.

11. A material quality estimation method including a quality estimation step in which a quality estimation unit provided in a computer estimates the quality of a manufactured product based on a material quality estimation model created by the method for creating a material quality estimation model according to any one of claims 1 to 8.

12. A model creation unit that performs machine learning with existing manufacturing condition data as input and existing quality data associated with the existing manufacturing condition data as output to create a material quality estimation base model, and the model creation unit inputs additional manufacturing condition data and outputs additional quality data associated with the additional manufacturing condition data to perform re-learning of the material quality estimation base model and create a material quality estimation model. A device for creating a material quality estimation model.

13. A material quality estimation device comprising a quality estimation unit that estimates the quality of a manufactured product based on the material quality estimation model created by the device for creating a material quality estimation model according to claim 12.

14. A model creation device that performs machine learning with existing manufacturing condition data as input and existing quality data associated with the existing manufacturing condition data as output to create a material quality estimation base model, and the model creation device inputs additional manufacturing condition data and outputs additional quality data associated with the additional manufacturing condition data to perform re-learning of the material quality estimation base model and create a material quality estimation model. A system for creating a material quality estimation model.

Images