Physical property prediction model construction method, physical property prediction method, and physical property prediction system

Abstract

To provide a physical property prediction model construction method for constructing a prediction model that can predict physical properties and physical property values of a composition from material blend information of the composition even when the composition uses materials with unknown material properties, and to provide a physical property prediction method and a physical property prediction system using the physical property prediction model.SOLUTION: A physical property prediction system provided herein comprises a prediction device and a database. In the material prediction system, a control unit is configured to: search for multiple pieces of similar material identification information corresponding to multiple pieces of physical property information with mutually similar physical property values in each set of composition property data of a reference dataset (S12); assign an identification code common to the retrieved multiple pieces of similar material identification information (S14); generate a learning dataset for training a prediction model (S16); and input the generated learning dataset to the prediction model as teacher data to train the prediction model (S16).SELECTED DRAWING: Figure 7

Description

Technical Field

[0001] The present disclosure relates to a physical property prediction model construction method for constructing a prediction model for predicting physical property values of a composition manufactured using a plurality of materials, a physical property prediction method for predicting the physical property values, and a physical property prediction system.

Background Art

[0002] Conventionally, a prediction method for obtaining specific performance (physical properties) of a rubber composition, which is a material such as a tire mounted on an automobile, is known. Further, using the names of a plurality of raw materials used in the production of the rubber composition (for example, the product names or model numbers specific to the raw materials) and their blending ratios as explanatory variables, and using the physical property data of the rubber composition as an objective function, a physical property data prediction method is known in which a computer is machine-learned to construct a prediction model and the physical property data is predicted using the prediction model.

[0003] Further, as another physical property data prediction method, a method using an identification name in which a plurality of raw materials with similar material characteristics are classified into one without using the product name or model number specific to the raw material as the identification information of the raw material is disclosed (see Patent Document 1).

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, the conventional physical property data prediction method described in Patent Document 1 cannot classify raw materials with unknown material properties. Therefore, for raw materials whose material properties are not publicly available, it is necessary to analyze the material properties of the raw materials in advance. In addition, since the material properties of natural rubber vary depending on the growing environment and the like, it is also necessary to analyze the material properties in advance when using natural rubber as a raw material. Further, when using recycled rubber produced by recycling discarded tires and rubber products as a raw material, since the recycled rubber is composed of various rubber materials, it is impossible to specify unique material properties and classification based on material properties cannot be performed.

[0006] An object of the present disclosure is to provide a physical property prediction model construction method for constructing a prediction model capable of predicting the physical properties and physical property values of a composition using a material with unknown material properties from the material blending information of the composition, a physical property prediction method using the physical property prediction model, and a physical property prediction system.

Means for Solving the Problems

[0007] A physical property prediction model construction method according to an aspect of the present disclosure is a physical property prediction model construction method for constructing a prediction model for predicting physical property values of a composition manufactured using a plurality of materials, including physical property information including physical property values of each of the plurality of compositions, material identification information indicating the materials used in the manufacture of the composition, and blending information regarding the blending of the plurality of materials. A search step of searching for a plurality of similar material identification information corresponding to a plurality of the physical property information approximated to each other by a predetermined optimization process with reference to a first data set; An assignment step of assigning an identifier common to the plurality of searched similar material identification information; A learning data set generation step of replacing the similar material identification information in the first data set with the identifier to generate a second data set for learning for training the prediction model; And a learning step of performing machine learning on the physical property prediction model using the second data set.

Effects of the Invention

[0008] According to the present disclosure, even for a composition using a material with unknown material properties, it is possible to construct a prediction model capable of predicting the physical properties and their values from the material composition information of the composition. Further, by performing physical property prediction using the prediction model thus constructed, it is possible to efficiently and reliably predict the physical property values of the composition.

Brief Description of the Drawings

[0009]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Embodiments for Carrying Out the Invention

[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. Note that the following embodiments are an example of embodying the present disclosure and do not limit the technical scope of the present disclosure.

[0011] FIG. 1 is a diagram showing the configuration of a physical property prediction system 100 according to an embodiment of the present disclosure. The physical property prediction system 100 of the present embodiment mainly includes a prediction device 10 and a database 30. Note that the physical property prediction system 100 is an example of the physical property prediction system of the present disclosure and is also a system for realizing the physical property prediction method of the present disclosure.

[0012] The physical property prediction system 100 is configured to be able to predict the physical properties and their physical property values of a polymer composition (an example of the composition of the present disclosure) composed of a plurality of materials including polymers and additives. The physical properties are the properties of the polymer composition, and the physical property values are physical quantities obtained by quantifying the degree of those properties. In the present embodiment, the physical property prediction system 100 uses material blending data including material identification information indicating each material required for the production of the polymer composition and blending information indicating the blending ratios of a plurality of materials required for the production of the polymer composition to predict the physical properties and their physical property values of the polymer composition.

[0013] Hereinafter, as a polymer composition to be predicted by the physical property prediction system 100, a rubber composition (rubbery elastomer) obtained by vulcanizing a compounded rubber (kneaded material) in which a plurality of materials including a polymer such as raw rubber and an additive such as silica are kneaded will be exemplified, and various processes executed by the physical property prediction system 100 will be described. Needless to say, the rubber composition is merely an example of the composition to be predicted, and the physical property prediction system 100 may be configured to predict the physical properties and their physical property values of a composition (an example of the composition of the present disclosure) as a rubbery elastomer formed without going through a vulcanization process.

[0014] Here, as described above, the rubber composition is obtained by vulcanizing the compounded rubber, and specifically, it is a rubber material used for the production of tire products such as pneumatic tires mounted on vehicles such as automobiles. That is, the physical property prediction system 100 of the present embodiment can predict the physical properties of the rubber composition constituting the tire product.

[0015] The prediction results by the physical property prediction system 100 are used, for example, in the development of the rubber composition and the development of the tire product manufactured using this rubber composition. If the predicted values of the physical properties of the rubber composition are obtained, the developers of the rubber composition and the tire product can grasp the general physical properties and their physical property values of the test products without performing physical property measurement inspections on the test products of the rubber composition and the tire product. Further, if the type of the material constituting the rubber composition and the tire product with unknown characteristics and the compounding information such as the compounding ratio thereof are obtained, the developers can grasp in advance the general physical properties and their physical property values of the rubber composition and the tire product without manufacturing the test products of the rubber composition and the tire product.

[0016] In addition, in the present embodiment, as an example of the rubber composition, a rubber material used for manufacturing the tire product is exemplified. However, for example, the rubber composition may be the tire product itself. In this case, the physical property prediction system 100 predicts the tire performance (an example of physical property) of the tire product and its performance value (an example of physical property value). Further, the rubber composition may be the industrial rubber product itself such as a vibration-proof rubber or a rubber material used for manufacturing the industrial rubber product.

[0017] As described above, the compounded rubber is an unvulcanized rubber obtained by kneading a plurality of materials including one or more polymers and one or more additives. The rubber composition is obtained by vulcanizing the compounded rubber.

[0018] The polymer is, for example, an unvulcanized raw rubber compounded into the compounded rubber. Examples of the raw rubber include natural rubber (NR), isoprene rubber (IR), butadiene rubber (BR), styrene-butadiene rubber (SBR), and the like.

[0019] The additives are, for example, fillers (filler agents) such as carbon black and silica, coupling agents, anti-aging agents, vulcanization accelerators, oils, zinc oxide, stearic acid, sulfur, processing aids, and the like.

[0020] [Configuration of Physical Property Prediction System 100] As shown in FIG. 1, the physical property prediction system 100 includes a prediction device 10 and a database 30, which are connected by a network N1 so as to be capable of data communication. The network N1 is, for example, a wired communication network connected by a LAN or the like, or a wireless communication network such as a dedicated line or a public line.

[0021] In this embodiment, a configuration in which the database 30 is connected to the network N1 is illustrated. However, for example, the database 30 may be provided in the prediction device 10. That is, a storage device that replaces the database 30 may be provided in the prediction device 10. In this case, the physical property prediction system of the present disclosure is realized by the prediction device 10 alone.

[0022] The prediction device 10 is an element constituting the physical property prediction system 100. The prediction device 10 uses various information such as input information input to the prediction device 10 and a prediction model 123 (see FIG. 4) described later that has been constructed in advance to predict the physical properties of the rubber composition to be predicted and the physical property values thereof. Further, the prediction device 10 outputs the prediction result to the information terminal 20 connected to the network N1.

[0023] Further, the prediction device 10 generates a material classification list 32 (see FIG. 3) described later using various information such as input information input to the prediction device 10 and an optimization model 124 (see FIG. 4) described later, and further generates a learning data set 33 (see FIG. 6) as teacher data necessary for machine learning of the prediction model 123. The learning data set 33 is an example of the second data set of the present disclosure.

[0024] The prediction device 10 is an information processing device capable of executing various arithmetic processes, and is, for example, a server computer, a cloud server, or a personal computer connected to the network N1. Note that the prediction device 10 is not limited to a single computer, and may be a computer system in which a plurality of computers operate in cooperation, or a cloud computing system. Further, various processes executed by the prediction device 10 may be distributed and executed by one or a plurality of processors. A program or computer software for operating the physical property prediction system 100 is installed in the prediction device 10.

[0025] The information terminal 20 is an information processing device or a terminal device used by a user. The information terminal 20 is a so-called desktop personal computer, a notebook personal computer, or a portable terminal such as a smartphone or a tablet terminal that can be carried around. The user operates the information terminal 20 to input various information necessary for various processes executed in the prediction device 10 into the prediction device 10. Further, the information terminal 20 displays the prediction result and the like transmitted from the prediction device 10 on the display screen. Therefore, a program or computer software for transmitting the various information to the prediction device 10 in cooperation with the physical property prediction system 100 and displaying the prediction result on the display screen is installed in the information terminal 20.

[0026] The database 30 is a data group in which various data handled in the physical property prediction system 100 are stored in a storage medium based on a predetermined data management method. The database 30 is managed in various forms such as a storage device, an information processing device, a cloud server, and a data server that are data communicably connected to the network N1. The database 30 stores a reference data set 31 (an example of the first data set of the present disclosure) and a material classification list 32.

[0027] The reference dataset 31 is a dataset used for constructing (generating) a prediction model 123 (see FIG. 4) used in the prediction process by the prediction device 10. In other words, the reference dataset 31 is a dataset used to generate a learning dataset 33 (see FIG. 6) used for machine learning of the prediction model 123 described later.

[0028] Specifically, the reference dataset 31 includes material identification information, compounding information, and property information for each of a large number of the rubber compositions whose physical properties and physical property values are known. Hereinafter, the rubber composition whose physical properties and physical property values are known may be referred to as a sample rubber.

[0029] The sample rubber is, for example, a rubber product (such as a tire product or an industrial rubber product) manufactured as a product so far, a prototype rubber product manufactured during research and experiments for the development of the rubber product, or a test piece made of the rubber composition manufactured during the research.

[0030] The reference dataset 31 is an aggregate of data in which a large number of composition characteristic data 31A corresponding to various sample rubbers are collected. Each composition characteristic data 31A includes the material identification information, the compounding information, and the property information regarding the corresponding sample rubber. This reference dataset 31 is stored in the database 30.

[0031] Here, the material identification information is information for identifying a plurality of materials used in the manufacture of the sample rubber, and is, for example, a material name for specifying a material, a product name, a product model number, a serial number, and the like.

[0032] The compounding information is information indicating the compounding ratio of a plurality of materials used in the manufacture of the sample rubber.

[0033] The property information is information indicating the physical properties and physical property values of the sample rubber.

[0034] FIG. 2 is a diagram illustrating one piece of composition characteristic data 31A included in the reference dataset 31. As shown in FIG. 2, the composition characteristic data 31A includes the material name and its blending ratio for each material, and also includes the physical property name and physical property value as the physical property information of the sample rubber.

[0035] In the composition characteristic data 31A of FIG. 2, "TSR20", "SBR1502", and "BR10B" are material names for identifying polymers (raw rubbers), "N220" is a material name for identifying carbon black as a filler (filler / reinforcing agent), "VN3" is a material name for identifying silica as a filler (filler / reinforcing agent), and "Si69" is a material name for identifying a coupling agent as an additive. Also, as material names for identifying other additives, "5% oil sulfur", "zinc oxide", "stearic acid", "vulcanization accelerator", etc. are shown in the composition characteristic data 31A.

[0036] Also, in the composition characteristic data 31A, as the blending information of each material, numerical values representing the blending ratio of each material in parts by mass are shown. Specifically, the blending ratio of each material indicates the ratio of the blending amount (parts by mass) of each material when the total mass parts of each polymer contained in the sample rubber is 100. The unit used for the blending ratio is phr (: per hundred rubber).

[0037] Also, in the composition characteristic data 31A, as various physical properties of the sample tire, for example, Mooney viscosity, scorch time measured by a Mooney viscometer, vulcanization characteristic values (induction time tC(10), 50% vulcanization time tC(50), and 90% vulcanization time tC(90), etc.) measured by a vulcanization tester (curastometer (registered trademark)), hardness, tensile strength, glass transition temperature, high-temperature loss tangent tanδ, low-temperature loss tangent tanδ, elastic modulus (storage elastic modulus E′, loss elastic modulus E″, storage shear elastic modulus G′, loss shear elastic modulus G″, etc.), breaking strength, wear performance, each physical property name and their physical property values are shown.

[0038] Note that the composition characteristic data 31A included in the reference dataset 31 is not limited to that shown in FIG. 2. For example, kneading conditions when kneading a plurality of materials of the sample rubber, vulcanization conditions for vulcanizing the compounded rubber after kneading, etc. may be included in the composition characteristic data 31A. The kneading conditions are, for example, the volume of the chamber of the kneader, the filling amount of the polymer composition during kneading, the kneading time, the kneading temperature, etc. Also, the vulcanization conditions are, for example, the vulcanization time, vulcanization temperature, etc. set for the mixed rubber before vulcanization of each sample rubber. Further, the composition characteristic data 31A may include all identification information used in the market (e.g., product name, product model number, serial number, manufacturer information, etc.) as identification information of each material, and function information indicating features such as the role and function of each material.

[0039] The material classification list 32 (see FIG. 1) is data (material classification data) referred to when generating a learning dataset 33 (see FIG. 6) used for machine learning of the prediction model 123 described later. The material classification list 32 is stored in the database 30 together with the reference dataset 31. Note that the material classification list 32 may be stored in a storage device different from the reference dataset 31.

[0040] FIG. 3 is a diagram showing an example of the material classification list 32. As shown in FIG. 3, the material classification list 32 is table data (material classification data) representing the material type divided by the type of each material for each material category, the material name (material identification information) belonging to the material type, and the identification code (an example of the identifier of the present disclosure) assigned to the material name. In the material classification list 32 of FIG. 3, two categories of polymer and filler roughly classified based on their uses are defined as the material categories. Also, the material category "polymer" is classified into four material types (NR, IR, SBR, BR), and the material category "filler" is classified into two material types (carbon, silica). Note that in FIG. 3, in the column of the material name, some actual names are shown and some fictional names are shown.

[0041] In this embodiment, the material classification list 32 includes all the material names included in the reference dataset 31. For example, the material classification list 32 is generated by the control unit 11 (see FIG. 4) of the prediction device 10. For example, the control unit 11 extracts all the material names from the reference dataset 31, identifies the material type to which each material belongs based on each material name, and further identifies the material category to which each material type belongs based on the type name of each material type. Then, table data (material classification list 32) in which the material category, the material type, and the material name are summarized in a list format is created.

[0042] On the other hand, the identification code included in the material classification list 32 is issued when the prediction model construction process (see FIGS. 7 and 8) described later is executed and is assigned to each material name. The method of assigning the identification code will be described later.

[0043] Note that the material classification list 32 is not limited to that shown in FIG. 3. For example, not only the two categories of polymer and filler for the material category, but also the material types and material categories corresponding to all the materials that can be selected as the materials (such as polymers and additives) of the rubber composition may be defined.

[0044] [Prediction device 10] Hereinafter, with reference to FIG. 4, the configuration of the prediction device 10 will be described. FIG. 4 is a block diagram showing the configuration of the prediction device 10.

[0045] The prediction device 10 is for realizing the physical property prediction system 100 of this embodiment, and uses the prediction model 123 described later to perform a process of predicting the physical properties and physical property values that a rubber composition (hereinafter sometimes referred to as a prediction target composition) having unknown physical properties will have.

[0046] In addition, the prediction device 10 generates a learning dataset 33 (see FIG. 6) necessary for machine learning of the prediction model 123, and performs a process of constructing (generating) the prediction model 123 using the learning data.

[0047] As shown in FIG. 4, the prediction device 10 includes a control unit 11, a storage unit 12, a communication unit 13, a display unit 14, and an operation unit 15.

[0048] The communication unit 13 is a communication interface for connecting the prediction device 10 to the network N1 and performing data communication with each device connected to the network N1 according to a predetermined communication protocol. Specifically, the communication unit 13 performs data communication with the information terminal 20 and the database 30 through the network N1.

[0049] The display unit 14 is a display device such as a liquid crystal display or an organic EL display that displays various types of information.

[0050] The operation unit 15 is an input device such as a mouse, a keyboard, or a touch panel that receives an operation by an operator.

[0051] The storage unit 12 is a non-volatile storage medium such as an HDD, an SSD, or a flash memory that stores various types of information and data. The storage unit 12 stores a control program 121, a prediction model 123, and an optimization model 124. Note that the prediction model 123 and the optimization model 124 may be realized as electronic circuits each having a memory in which they are stored.

[0052] The control program 121 is non-temporarily recorded on a computer-readable recording medium such as a CD or a DVD, and may be read from the recording medium by a reading device (not shown) such as a CD drive or a DVD drive that is electrically connected to the prediction device 10 and copied to the storage unit 12. Further, the control program 121 may be read from an external storage connected to the network N1, input through the communication unit 13, and copied to the storage unit 12.

[0053] The control program 121 is a program for executing prediction processing using the prediction model 123 or for executing the optimization processing (see FIGS. 7 and 8) described later using the optimization model 124. Note that instead of the control program 121, a program for causing the control unit 11 to execute the prediction processing and a program for causing the control unit 11 to execute the optimization processing may be separately stored in the storage unit 12.

[0054] The prediction model 123 is a learned model used for prediction processing to predict the physical properties and physical property values of the prediction target composition, and derives the physical properties and predicted values of the physical property values of the rubber composition (prediction target composition) that is the prediction target. In the present embodiment, when the material identification information and the compounding information of the prediction target composition to be predicted are input to the input unit of the prediction model 123 as explanatory variables, the prediction model 123 predicts the physical property values of each characteristic of the prediction target composition and outputs the predicted values from the output unit of the prediction model 123.

[0055] In the present embodiment, predicted values of each of a plurality of characteristics are predicted and output by the prediction model 123. Note that the prediction model 123 may include a well-known function (prediction function) that returns a predicted value (output data) for an input value (input data).

[0056] The prediction model 123 is constructed (generated) by a prediction model construction unit 200 described later. In the present embodiment, the prediction model 123 is constructed (generated) by machine learning by the control unit 11 based on the learning data set 33 (see FIG. 6) generated by the learning data set generation unit 204 (see FIG. 5) and a predetermined algorithm. Further, when the learning data set 33 is updated, the control unit 11 re-learns and updates the prediction model 123 each time. Details of the method for constructing the prediction model 123 will be described later.

[0057] As algorithms necessary for constructing the prediction model 123, for example, algorithms such as multiple regression, generalized linear regression, principal component regression, ridge regression, lasso regression, kernel regression, random forest regression, Gaussian process regression, multi-layer neural network, clustering, support vector machine, and RBF network defined by radial basis functions are preferably used. Further, the prediction model 123 may be constructed by deep learning (deep learning) that multiplies the intermediate layer of the neural network. Note that the prediction model 123 may use one of the above-described algorithms, or may use a plurality of algorithms. In the present embodiment, the generalized linear regression, which is excellent in the balance between accuracy and cost, is used as the algorithm of the prediction model 123.

[0058] The optimization model 124 is used for the optimization process (see FIGS. 7 and 8) by the optimization processing unit 202 (see FIG. 5) described later. In the present embodiment, the optimization model 124 refers to the reference data set 31 including the physical property information, the material identification information, and the compounding information, and searches for a plurality of the material identification information (hereinafter referred to as similar material identification information) corresponding to a plurality of physical property information whose physical property values are approximate to each other. The details of the optimization process by the optimization processing unit 202 (see FIG. 5) will be described later.

[0059] The optimization model 124 preferably uses algorithms such as a genetic algorithm (GA), list search, particle swarm optimization (PSO), response surface methodology (RSM), simulated annealing (SA), multi-objective optimization algorithm, and Bayesian optimization. In the present embodiment, the optimization model 124 may use one of the above-described algorithms or may use a plurality of algorithms. In the present embodiment, from the viewpoint of computational efficiency, the Bayesian optimization is used as the algorithm of the optimization model 124.

[0060] The control unit 11 controls the operations of each part of the prediction device 10. The control unit 11 includes control devices such as a CPU, a ROM, and a RAM. The CPU is a processor that executes various arithmetic processes. The ROM is a non-volatile storage medium in which control programs such as BIOS and OS for causing the CPU to execute various arithmetic processes are stored in advance. The RAM is a volatile or non-volatile storage medium that stores various information and is used as a temporary storage memory (working area) for various arithmetic processes executed by the CPU. The control unit 11 controls the prediction device 10 by causing the CPU to execute various control programs stored in advance in the ROM or the storage unit 12.

[0061] As shown in FIG. 4, the control unit 11 includes various processing units such as a compounding data acquisition processing unit 111 (an example of the data acquisition unit of the present disclosure), an extraction processing unit 112 (an example of the extraction processing unit of the present disclosure), a data generation processing unit 113, a prediction processing unit 114 (an example of the prediction processing unit of the present disclosure), and a prediction model construction unit 200.

[0062] The control unit 11 functions as the various processing units when the CPU executes various arithmetic processes according to the control program. The control unit 11 or the CPU is an example of a computer that executes the control program. Note that some or all of the processing units included in the control unit 11 may be configured by electronic circuits. Further, the control program may be a program for causing a plurality of processors to function as the various processing units.

[0063] The composition data acquisition processing unit 111 performs a process of acquiring material composition data including the material identification information indicating the materials used for manufacturing the prediction target composition with unknown physical properties and physical property values, and the composition information indicating the composition ratios regarding the composition of a plurality of materials required for manufacturing the prediction target composition.

[0064] The material composition data of the prediction target composition is input by the user on the information terminal 20 when obtaining a predicted value of the physical property value of the physical property that the prediction target composition has. For example, when the material composition data of the prediction target composition input by the user on the information terminal 20 is transferred from the information terminal 20 to the prediction device 10 through the network N1, the composition data acquisition processing unit 111 acquires the material composition data.

[0065] Note that the composition data acquisition processing unit 111 may acquire the material composition data of the prediction target composition, for example, by accessing an external storage device indicated by the address information input from the information terminal 20, reading out the material composition data of the prediction target composition stored in the external storage device, and storing it in the storage unit 12.

[0066] The extraction processing unit 112 refers to the material classification list 32 in the database 30, and performs a process of extracting the identification code (see FIG. 3) corresponding to the material identification information included in the material formulation data acquired by the formulation data acquisition processing unit 111 from the material classification list 32. For example, when the material name "NR1101" is included in the material formulation data of the composition to be predicted, the extraction processing unit 112 checks whether the material name "NR1101" is included in the material classification list 32. When the material name "NR1101" is included, the extraction processing unit 112 extracts the identification code "NR01" corresponding to the material name "NR1101" from the material classification list 32.

[0067] The data generation processing unit 113 performs a process of generating input data for inputting to the prediction model 123 as explanatory variables by replacing the material identification information in the material formulation data with the identification code extracted by the extraction processing unit 112. For example, when the material name "NR1101" and its blending ratio are included in the material formulation data of the composition to be predicted, the data generation processing unit 113 replaces the material name "NR1101" in the material formulation data with the identification code "NR01", and generates the replaced material formulation data as the input data.

[0068] The prediction processing unit 114 performs a process of predicting the physical properties and their physical property values of the composition to be predicted based on the input data generated by the data generation processing unit 113.

[0069] Specifically, the prediction processing unit 114 inputs the input data generated by the data generation processing unit 113 to the input unit of the prediction model 123, causes the prediction model 123 to predict the physical property values of the composition to be predicted, and outputs the prediction results of the physical properties and their physical property values of the composition to be predicted from the output unit of the prediction model 123.

[0070] The physical properties to be predicted are, for example, Mooney viscosity, scorch time measured by a Mooney viscometer, vulcanization characteristic values (induction time tC(10), 50% vulcanization time tC(50), 90% vulcanization time tC(90), etc.) measured by a vulcanization tester (curastometer (registered trademark)), hardness, tensile strength, glass transition temperature, high-temperature loss tangent tanδ, low-temperature loss tangent tanδ, elastic modulus (storage elastic modulus E′, loss elastic modulus E″, storage shear elastic modulus G′, loss shear elastic modulus G″, etc.), fracture strength, and wear performance, which are any plurality or all of them. Note that the prediction processing unit 114 may predict the physical property values of at least one physical property.

[0071] The predicted values predicted by the prediction processing unit 114 are transferred to the information terminal 20 or other external devices by the communication unit 13 in order to be displayed on the display unit of the information terminal 20 or other external devices.

[0072] The prediction model construction unit 200 executes a process of constructing (generating) a learned prediction model 123 by learning the prediction model 123 by machine learning based on a pre-generated learning data set 33 (see FIG. 6) and a predetermined algorithm.

[0073] FIG. 5 is a block diagram showing the configuration of the prediction model construction unit 200. As shown in FIG. 5, the prediction model construction unit 200 includes various processing units such as a condition setting unit 201, an optimization processing unit 202, an identification code assignment unit 203, a learning data set generation unit 204, and an evaluation determination unit 205.

[0074] The condition setting unit 201 sets setting conditions such as constraint conditions for the optimization target value by the optimization processing unit 202 and inputs them to the optimization processing unit 202. The constraint conditions are, for example, constraint conditions for determining the common identification code assigned to each material, and are, for example, the maximum value of the number of types of materials belonging to a specific material category. For example, when the number of types of SBR is 42, the possible values of the identification code are at most 42 and at least 1. In this case, the condition setting unit 201 sets, as the constraint condition, the maximum value of the number of searches for the plurality of similar material identification information, in other words, the upper limit value of the possible number of the common identification code to 42. Note that the constraint conditions may be conditions for restricting the blending ratio of each material within a predetermined range, conditions for restricting the physical property value of a specific physical property within a predetermined range, and the like.

[0075] The optimization processing unit 202 refers to the reference data set 31 in the database 30 and performs a process of searching for a plurality of the similar material identification information corresponding to a plurality of physical property information in which the physical property values approximate each other in each of the plurality of composition property data 31A.

[0076] In the present embodiment, when an execution instruction for executing the optimization process is input, the optimization processing unit 202 enables the optimization model 124, inputs the constraint conditions and the reference data set 31 to the optimization model 124, and causes the optimization model 124 to execute a search process for an optimal solution. By repeating this search process, design variables for which the calculated value of the above-described objective function becomes the minimum value are specified. That is, the similar material identification information for achieving the target value is specified.

[0077] The identification code assignment unit 203 performs an assignment process of assigning a common identification code to the plurality of searched similar material identification information. For example, when the optimization processing unit 202 obtains a solution that the physical property values of the material indicated by the material name "NR1101" and the physical property values of the material indicated by the material name "NR1102" approximate each other, in the material classification list 32, a common identification code "NR01" is assigned as the identification code corresponding to these material names and registered in the material classification list 32.

[0078] The learning dataset generation unit 204 performs data generation processing for generating a learning dataset 33 for training the prediction model 123 by replacing the similar material identification information in the reference dataset 31 with the identification code assigned by the identification code assignment unit 203. FIG. 6 is a diagram showing an example of the learning dataset 33, in which the material names of the polymer and the filler in each composition property data 31A of the reference dataset 31 shown in FIG. 2 are replaced with the identification codes. That is, the learning dataset 33 includes a plurality of composition property data 33A in which the material names are converted into the identification codes. FIG. 6 shows a plurality of composition property data 33A in which the material names are converted into the identification codes.

[0079] The learning dataset 33 generated by the learning dataset generation unit 204 is used as teacher data for machine learning of the prediction model 123. That is, when it is necessary to machine-learn the prediction model 123, the control unit 11 inputs the learning dataset 33 to the prediction model 123 to train the prediction model 123.

[0080] The evaluation determination unit 205 inputs the identification code and the compounding information included in the pre-prepared verification data into the learned prediction model 123 as explanatory variables, and based on the predicted values output from the prediction model 123, performs processing for evaluating the prediction result by the prediction model 123. Here, the verification data is data including the physical property information of the verification rubber composition prepared for verification, the identification code corresponding to the verification rubber composition, and the compounding information.

[0081] An example of the prediction evaluation value indicating the evaluation of the prediction result is, for example, the correlation coefficient (:R) between the predicted value obtained by inputting the identification code and the formulation information included in the verification data into the prediction model 123 and the known physical property value (measured value) included in the verification data, and the root mean square error (:RMSE) between the predicted value and the known physical property value (measured value). The evaluation determination unit 205 calculates the correlation coefficient and the root mean square error between the predicted value and the known physical property value.

[0082] Here, the correlation coefficient is an index indicating the linear relationship between the predicted value and the known physical property value, and takes a value from -1 to +1. The larger the correlation coefficient, the higher the prediction accuracy by the prediction model 123 is evaluated. Let the correlation coefficient be R, then the correlation function is calculated by the following calculation formula.

[0083]

Equation

[0084] In addition, the root mean square error is an index indicating the magnitude of the error of the predicted value, and is expressed as the square root of the average value of the squares of the differences between the predicted value and the known physical property value. The smaller the root mean square error, the higher the prediction accuracy by the prediction model 123 is evaluated. Let the root mean square error be RMSE, then the root mean square error is calculated by the following calculation formula.

[0085]

Equation

[0086] In this embodiment, based on the prediction evaluation value, the search process for the optimal solution by the optimization model 124, the subsequent assignment process by the identification code assignment unit 203, the data generation process by the learning data set generation unit 204, and the learning process of the prediction model 123 using the learning data set 33 are repeatedly executed.

[0087] Specifically, since the higher the correlation coefficient, the higher the prediction accuracy by the prediction model 123, the control unit 11 repeatedly executes the search process, the allocation process, the data generation process, and the learning process until the calculated value of the correlation coefficient becomes equal to or greater than a predetermined first set value. In this case, the prediction model 123 is learned using the learning data set 33 when the calculated value of the correlation coefficient becomes equal to or greater than the first set value, and the construction of the learned prediction model 123 is completed.

[0088] Alternatively, the control unit 11 may repeatedly execute the search process, the allocation process, the data generation process, and the learning process until the number of times the maximum value of the calculated value of the correlation coefficient has not been updated reaches a predetermined first set number of times. For example, even if the calculation of the first set number of times is performed after the previous update, if a calculation result larger than the calculated value at the previous update cannot be obtained, it is determined that the calculated value of the correlation coefficient has converged. In this case, the prediction model 123 is learned using the learning data set 33 at the time of the previous update, and the construction of the learned prediction model 123 is completed.

[0089] Also, since the smaller the root mean square error, the higher the prediction accuracy by the prediction model 123, the control unit 11 repeatedly executes the search process, the allocation process, the data generation process, and the learning process until the calculated value of the root mean square error becomes less than a predetermined second set value. In this case, the prediction model 123 is learned using the learning data set 33 when the calculated value of the root mean square error becomes less than the second set value, and the construction of the learned prediction model 123 is completed.

[0090] Alternatively, the control unit 11 may repeatedly execute the search process, the allocation process, the data generation process, and the learning process until the number of times the minimum value of the calculated root mean square error has not been updated reaches a second set number of times determined in advance. For example, even if the calculation of the second set number of times is performed after the previous update, if a calculation result smaller than the calculated value at the previous update cannot be obtained, it is determined that the calculated value of the root mean square error has converged. In this case, the prediction model 123 is learned using the learning data set 33 at the previous update, and the construction of the learned prediction model 123 is completed.

[0091] Regarding the correlation coefficient, a range equal to or greater than the first set value is an example of the allowable range of the present disclosure. Also, regarding the root mean square error, a range less than the second set value is an example of the allowable range of the present disclosure. The first set value and the second set value can be arbitrarily set according to the prediction accuracy required for the prediction model 123.

[0092] In the present embodiment, the control unit 11 repeatedly executes the search process, the allocation process, the data generation process, and the learning process until the correlation coefficient is equal to or greater than the first set value and the root mean square error is less than the second set value.

[0093] Note that the control unit 11 may repeatedly execute the search process, the allocation process, the data generation process, and the learning process until the number of times the maximum value of the calculated correlation coefficient has not been updated reaches the first set number of times and the number of times the minimum value of the calculated root mean square error has not been updated reaches the second set number of times.

[0094] Further, in the present embodiment, even if the predicted evaluation value is outside the allowable range, when the number of executions of the search process reaches a preset upper limit number of times, the control unit 11 may end the construction of the prediction model 123 without executing the search process and subsequent processes any further. For example, even if the correlation coefficient is less than the first set value or the root mean square error is greater than or equal to the second set value, when the number of executions of the search process reaches the set number of times, the prediction model 123 is learned using the learning data set 33 at that time, and the construction of the prediction model 123 ends.

[0095] Note that the predicted evaluation value may be, for example, a difference value between the predicted value and the physical property value included in the known material information. In this case, the search process and subsequent processes are executed until the difference value becomes less than a preset threshold value.

[0096] By the way, conventional physical property data prediction methods cannot classify raw materials with unknown material properties. Therefore, for raw materials whose material properties are not publicly available, it is necessary to analyze the material properties of the raw materials in advance. In addition, since the material properties of natural rubber vary depending on the growing environment, etc., it is also necessary to analyze the material properties in advance when using natural rubber as a raw material. Further, when using recycled rubber manufactured by recycling discarded tires and rubber products as a raw material, since the recycled rubber is composed of various rubber materials, it is impossible to specify unique material properties and classification based on the material properties cannot be performed.

[0097] On the other hand, in the present embodiment, since the prediction model 123 is constructed as described above, even for a rubber composition using a material with unknown material properties, it is possible to predict the physical properties and physical property values of the rubber composition from the blending information of each material of the rubber composition.

[0098] [Prediction Model Construction Process] Hereinafter, with reference to the flowchart of FIG. 7, an example of the procedure of the prediction model construction process executed in the physical property prediction system 100 will be described together with the physical property prediction model construction method of the present disclosure. Note that one or more steps included in the prediction model construction process described below may be appropriately omitted, and the execution order of each step may be different within the range that produces the same operational effects. Further, in the following description, the execution subject of each step is described as the control unit 11, but the execution subject of each step may be the prediction model construction unit 200 or one or more processors included in the prediction model construction unit 200.

[0099] When an execution instruction for executing the optimization process is input, the control unit 11 inputs the reference data set 31 read from the database 30 and the constraint conditions set by the condition setting unit 201 into the optimization model 124 of the optimization processing unit 202 (S11).

[0100] Then, in the next step S12, the control unit 11 activates the optimization model 124 and executes a search process for an optimal solution by the optimization model 124 based on the reference data set 31 and the constraint conditions. The optimization model 124 performs a process of searching for a plurality of the similar material identification information corresponding to a plurality of physical property information in which the physical property values are approximated to each other in each of the composition property data 31A of the reference data set 31 based on the input reference data set 31 and the constraint conditions. Note that step S12 is an example of the search step of the present disclosure.

[0101] When a solution is output by the search process by the optimization model 124 (S13), the control unit 11 assigns an identification code common to the plurality of the searched similar material identification information (S14). Note that step S14 is an example of the assignment step of the present disclosure.

[0102] In the next step S15, the control unit 11 replaces the similar material identification information in the reference data set 31 with the common identification code, and generates a learning data set 33 for training the prediction model 123. Note that step S15 is an example of the learning data set generation step of the present disclosure.

[0103] In step S16, the control unit 11 inputs the generated learning data set 33 as teacher data into the prediction model 123, and trains the prediction model 123. Note that step S16 is an example of the learning step of the present disclosure.

[0104] Thereafter, the trained prediction model 123 is stored in the storage unit 12 to enable prediction processing by the prediction model 123 (S17). Thereby, a series of prediction model construction processes are completed.

[0105] [Other processing examples of prediction model construction processing] Hereinafter, with reference to the flowchart of FIG. 8, other processing examples of the prediction model construction processing will be described. For the same processing procedures as those shown in the processing procedure of FIG. 7, the same step numbers are given in FIG. 8, and the detailed description thereof is omitted.

[0106] As shown in FIG. 8, when the processes shown in steps S11 to S16 are sequentially performed, then, in step S201, the control unit inputs the identification code corresponding to the rubber composition with known physical properties and the compounding information as explanatory variables to the input unit of the trained prediction model 123 learned in step S16, and causes the prediction model 123 to perform prediction processing. Thereafter, in step S202, the control unit 11 evaluates the prediction result by the prediction model 123 based on the predicted value output from the prediction model 123.

[0107] In the next step S203, the control unit 11 determines whether the predicted evaluation value is within a predetermined allowable range. In the present embodiment, the control unit 11 determines whether the correlation coefficient is equal to or greater than the first set value and whether the root mean square error is less than the second set value. That is, when the correlation coefficient is equal to or greater than the first set value and the root mean square error is less than the second set value, the control unit 11 determines that the predicted evaluation value is within the allowable range.

[0108] Here, when it is determined that the predicted evaluation value is within the allowable range, the control unit 11 determines that the prediction model 123 has been learned to the extent that it can execute high-precision prediction processing, stores the learned prediction model 123 in the storage unit 12, and enables the prediction processing by the prediction model 123 (S17).

[0109] On the other hand, in step S203, when it is determined that the predicted evaluation value is outside the allowable range, the control unit 11 determines in the next step S204 whether the number of executions of the optimization process in step S12 has reached a predetermined set number. In step S204, when it is determined that the number of executions is less than the set number, the control unit 11 returns to step S12 and repeatedly executes the processes after step S12. That is, the control unit 11 repeatedly executes the processes after step S12 until it is determined in step S203 that the predicted evaluation value is within the allowable range.

[0110] On the other hand, in step S204, when it is determined that the number of executions is equal to or greater than the set number, the control unit 11 determines that the prediction model 123 has been learned to the extent that it can execute high-precision prediction processing, stores the learned prediction model 123 in the storage unit 12, and enables the prediction processing by the prediction model 123 (S17). That is, the control unit 11 repeatedly executes the processes after step S12 until it is determined in step S204 that the number of executions is equal to or greater than the set number.

[0111] [Physical property prediction processing] Hereinafter, with reference to the flowchart of FIG. 9, an example of the procedure of the physical property prediction process executed in the physical property prediction system 100 will be described together with the physical property prediction method of the present disclosure. Note that one or more steps included in the physical property prediction process described below may be appropriately omitted, and the execution order of each step may be different as long as the same operational effects are achieved. In the following description, the execution subject of each step is described as the control unit 11, but the execution subject of each step may be the prediction processing unit 114 or one or more processors included in the prediction processing unit 114.

[0112] When an execution instruction for executing the physical property prediction process is input, in step S31, the control unit 11 acquires the material composition data including the material identification information indicating the material used in the production of the prediction target composition whose physical properties and physical property values are unknown, and the blending information indicating the blending ratio regarding the blending of a plurality of materials required for the production of the prediction target composition. Note that step S31 is an example of the data acquisition step of the present disclosure.

[0113] In the next step S32, the control unit 11 refers to the material classification list 32 in the database 30 and extracts the identification code (see FIG. 3) corresponding to the material identification information included in the material composition data from the material classification list 32. Note that step S32 is an example of the extraction step of the present disclosure.

[0114] Thereafter, the control unit 11 replaces the material identification information in the material composition data with the identification code extracted in step S32 to generate input data for inputting to the prediction model 123 as an explanatory variable (S33).

[0115] In the next step S34, the control unit 11 performs a process of predicting the physical properties and their physical property values of the composition to be predicted based on the input data generated in step S33. Specifically, the control unit 11 inputs the generated input data into the input unit of the learned prediction model 123, causes the prediction model 123 to predict the physical property values of the composition to be predicted, and outputs the prediction results of the physical properties and their physical property values of the composition to be predicted from the output unit of the prediction model 123. Note that step S34 is an example of the prediction step of the present disclosure.

[0116] After that, the control unit 11 transfers the prediction result to the information terminal 20 or other external devices (S35) in order to display the predicted predicted value on the display unit of the information terminal 20 or other external devices.

[0117] As described above, in the present embodiment, since the prediction model construction process (see FIGS. 7 and 8) described above is executed in the physical property prediction system 100, even for a rubber composition using a material with unknown material properties, a prediction model 123 capable of predicting the physical properties and physical property values of the rubber composition from the material identification information and the compounding information of the rubber composition can be constructed.

[0118] In addition, since the processes after the optimization process (S12) are repeatedly executed until the prediction evaluation value is within the allowable range, a high-quality learning dataset 33 for improving the prediction accuracy of the prediction model 123 is generated.

[0119] Furthermore, since the processes after the optimization process (S12) are repeatedly executed until the number of executions is equal to or greater than the set number, a high-quality learning dataset 33 for further improving the prediction accuracy of the prediction model 123 is generated.

[0120] In addition, since the physical property prediction process (see FIG. 9) described above is executed in the physical property prediction system 100, even for a rubber composition using a material with unknown material properties, by inputting the material blending data including the material identification information of each material of the rubber composition and their blending information into the prediction model 123, a highly accurate prediction process by the prediction model 123 is executed. As a result, the physical properties and their physical property values of the rubber composition are predicted with high accuracy from the material identification information of each material of the rubber composition and their blending information.

[0121] In the above-described embodiment, the configuration in which the prediction device 10 is provided with the blending data acquisition processing unit 111, the extraction processing unit 112, the data generation processing unit 113, the prediction processing unit 114, and the prediction model construction unit 200 has been illustrated. However, the physical property prediction system 100 of the present disclosure is not limited to such a configuration. For example, the prediction device 10 may not be provided with the prediction model construction unit 200. In this case, the present disclosure may be configured as a prediction model construction device or a prediction model construction system including each processing unit of the prediction model construction unit 200. Further, the present disclosure may be configured as a learning data set generation device or a learning data set generation system including each processing unit of the prediction model construction unit 200.

[0122] The embodiments of the present disclosure described above include the following disclosed matters (1) to (15).

[0123] The method for constructing a physical property prediction model according to the disclosed matter (1) is a method for constructing a prediction model for predicting the physical property value of a composition manufactured using a plurality of materials, a search step of searching for a plurality of similar material identification information corresponding to a plurality of mutually approximate physical property information by a predetermined optimization process with reference to a first data set including the physical property information including the physical property values of each of the plurality of compositions, the material identification information indicating the materials used in the manufacture of the composition, and the blending information regarding the blending of the plurality of materials; an assignment step of assigning an identifier common to the plurality of searched similar material identification information; A learning dataset generation step of generating a second dataset for learning to train the prediction model by replacing the similar material identification information in the first dataset with the identifier; A learning step of performing machine learning on the prediction model using the second dataset.

[0124] Disclosure item (2) is in the physical property prediction model construction method described in the disclosure item (1), An evaluation step of inputting the identifier and the formulation information corresponding to the composition with known physical property information into the prediction model learned in the learning step as explanatory variables, and evaluating the prediction result by the prediction model based on the predicted value output from the prediction model; Based on the evaluation of the prediction result, the search step, the allocation step, the learning dataset generation step, and the learning step are repeatedly executed.

[0125] Disclosure item (3) is in the physical property prediction model construction method described in the disclosure item (2), and the search step, the allocation step, the learning dataset generation step, and the learning step are repeatedly executed until the evaluation of the prediction result is within a predetermined allowable range.

[0126] Disclosure item (4) is in the physical property prediction model construction method described in the disclosure item (3), When the number of executions of the search step, the allocation step, the learning dataset generation step, and the learning step reaches a predetermined set number, even if the evaluation of the prediction result is outside the allowable range, the construction of the prediction model is terminated.

[0127] Disclosure item (5) is a physical property prediction method for predicting the physical property value of a prediction target composed of a composition using the prediction model constructed by the physical property prediction model construction method described in any one of the disclosure items (1) to (4), A data acquisition step of acquiring material composition data including material identification information indicating a material used in manufacturing the prediction target whose physical property values are unknown and composition information regarding the composition of the plurality of materials; An extraction step of extracting the identifier corresponding to the material identification information included in the material composition data with reference to a material classification list in which the plurality of similar material identification information searched in the search step and the identifier assigned in the assignment step are associated; A prediction step of inputting input data obtained by replacing the material identification information in the material composition data with the identifier into the prediction model and outputting a prediction result of the physical property value of the prediction target from the prediction model.

[0128] Disclosure item (6) is a physical property prediction system for predicting the physical property value of a prediction target composed of a composition using the prediction model constructed by the physical property prediction model construction method according to any one of disclosure items (1) to (4), A data acquisition unit that acquires material composition data including material identification information indicating a material used in manufacturing the prediction target whose physical property values are unknown and composition information regarding the composition of the plurality of materials; An extraction processing unit that extracts the identifier corresponding to the material identification information included in the material composition data with reference to a material classification list in which the plurality of similar material identification information searched in the search step and the identifier assigned in the assignment step are associated; A prediction processing unit that inputs input data obtained by replacing the material identification information in the material composition data with the identifier into the prediction model and outputs a prediction result of the physical property value of the prediction target from the prediction model.

Explanation of Signs

[0129] 10: Prediction device 11: Control unit 12: Storage unit 20: Information terminal 30: Database 31: Reference dataset 32: Material Classification List 33: Learning Dataset 100: Physical Property Prediction System 111: Blending Data Acquisition Processing Unit 112: Extraction Processing Unit 113: Data Generation Processing Unit 114: Prediction Processing Unit 121: Control Program 123: Prediction Model 124: Optimization Model 200: Prediction Model Construction Unit 201: Condition Setting Unit 202: Optimization Processing Unit 203: Identification Code Assignment Unit 204: Learning Dataset Generation Unit 205: Evaluation Judgment Unit

Claims

1. A physical property prediction model construction method for constructing a prediction model for predicting physical property values of a composition manufactured using a plurality of materials, comprising: With reference to a first dataset including physical property information including physical property values of a plurality of compositions, material identification information indicating materials used in the manufacture of the compositions, and compounding information regarding compounding of the plurality of materials, a search step of searching for a plurality of similar material identification information corresponding to the plurality of physical property information approximating each other by a predetermined optimization process; An assignment step of assigning a common identifier to the plurality of searched similar material identification information; A learning dataset generation step of replacing the similar material identification information in the first dataset with the identifier to generate a second dataset for learning for learning the prediction model; A learning step of performing machine learning on the prediction model using the second dataset. A physical property prediction model construction method comprising.

2. Further comprising an evaluation step of inputting the identifier and the compounding information corresponding to the composition with known physical property information into the prediction model learned in the learning step as explanatory variables, and evaluating the prediction result by the prediction model based on the predicted value output from the prediction model; The physical property prediction model construction method according to claim 1, wherein the search step, the assignment step, the learning dataset generation step, and the learning step are repeatedly executed based on the evaluation of the prediction result.

3. The physical property prediction model construction method according to claim 2, wherein the search step, the assignment step, the learning dataset generation step, and the learning step are repeatedly executed until the evaluation of the prediction result is within a predetermined allowable range.

4. If the number of executions of the search step, the allocation step, the learning data set generation step, and the learning step reaches a preset number of settings, even if the evaluation of the prediction result is outside the allowable range, the construction of the prediction model is terminated. The physical property prediction model construction method according to claim 3.

5. A physical property prediction method for predicting a physical property value of a prediction target composed of a composition using the prediction model constructed by the physical property prediction model construction method according to claim 1, A data acquisition step of acquiring material composition data including material identification information indicating a material used in the production of the prediction target whose physical property value is unknown and composition information regarding the composition of the plurality of materials, An extraction step of extracting the identifier corresponding to the material identification information included in the material composition data with reference to a material classification list in which a plurality of the similar material identification information searched by the search step and the identifier assigned by the assignment step are associated, A prediction step of inputting input data obtained by replacing the material identification information in the material composition data with the identifier into the prediction model and outputting a prediction result of the physical property value of the prediction target from the prediction model. The physical property prediction method includes the above steps.

6. A physical property prediction system for predicting a physical property value of a prediction target composed of a composition using the prediction model constructed by the physical property prediction model construction method according to claim 1, A data acquisition unit that acquires material composition data including material identification information indicating a material used in the production of the prediction target whose physical property value is unknown and composition information regarding the composition of the plurality of materials, An extraction processing unit that extracts the identifier corresponding to the material identification information included in the material composition data with reference to a material classification list in which a plurality of the similar material identification information searched by the search step and the identifier assigned by the assignment step are associated, A physical property prediction system comprising: a prediction processing unit that inputs input data obtained by replacing the material identification information in the material composition data with the identifier into the prediction model, and outputs a prediction result of the physical property value of the prediction target from the prediction model.

Images