Method for creating data set group for machine learning

Abstract

To provide a method capable of creating a data set group covering a wide range of process conditions while suppressing an increase in the number of data sets.SOLUTION: A method for creating a data set group including a plurality of data sets for specifying conditions of a plurality of types of processes as teacher data for machine learning for predicting performance of an object. This method includes a first process S1 of acquiring a first dataset group including first datasets of a plurality of types of process conditions different from each other, a second process S2 of performing principal component analysis on the first dataset group and converting the first dataset group into a principal component set group including a plurality of principal component sets, a third process S2 of selecting a plurality of principal component sets smaller in number than the plurality of principal component sets from the plurality of principal component sets based on a result of the second process S3, and a fourth process S3 of generating a second dataset group including a plurality of second datasets obtained by applying inverse transformation of the principal component analysis to the plurality of principal component sets selected in the third process S4.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to a method for creating a set of data sets for machine learning. [Background technology]

[0002] Patent Document 1 listed below describes a method for predicting tire performance. This method includes a first step of inputting time-series temperature data of the rubber members when an unvulcanized tire including the unvulcanized rubber members is vulcanized and molded, a second step of predicting the physical properties of the rubber members after vulcanization based on the temperature data of the rubber members, and a third step of predicting the performance of the tire after vulcanization based on the physical properties of the rubber members. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2023-073083 Summary of the Invention [Problem to be solved by the invention]

[0004] In the technology of Patent Document 1, a learning model is used to predict the physical properties of rubber members. Generating such a learning model requires training data created for machine learning. Generally, the performance of an object such as a rubber member varies depending on the conditions of the vulcanization process, such as temperature data. Therefore, to improve the prediction accuracy of the learning model, it is important to create multiple data sets that cover a wide range of process conditions as training data. However, in order to cover a wide range of process conditions, it is conceivable to create a large number of data sets by assuming every possible condition. However, as the number of data sets increases, there is a problem in that machine learning requires a long time.

[0005] The present invention has been devised in view of the above-described circumstances, and has as its main object to provide a method capable of creating a group of data sets that cover a wide range of process conditions while suppressing an increase in the number of data sets. [Means for solving the problem]

[0006] The present invention is a method for creating a dataset group for machine learning, which includes a plurality of datasets that specify conditions for a plurality of processes and that serve as training data for machine learning to predict the performance of an object that is obtained through a process and whose performance varies depending on the conditions of the process. The method includes: a first step of acquiring a first dataset group that includes first datasets for different types of conditions for the process; a second step of performing principal component analysis on the first dataset group to convert the first dataset group into a principal component set group that includes a plurality of principal component sets; a third step of selecting a smaller number of principal component sets from the plurality of principal component sets based on the results of the second step; and a fourth step of generating a second dataset group that includes a plurality of second datasets by applying an inverse transformation of the principal component analysis to the plurality of principal component sets selected in the third step. [Effects of the Invention]

[0007] By adopting the above-described steps, the method for creating datasets for machine learning of the present invention makes it possible to create datasets that cover a wide range of process conditions while suppressing an increase in the number of datasets. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a perspective view showing an example of a computer for executing a method for creating a group of datasets for machine learning and a method for creating a trained model. [Figure 2] 1 is a cross-sectional view showing an example of a rubber product and vulcanized rubber (object). [Figure 3] 1 is a graph showing an example of vulcanization temperature history. [Figure 4]1 is a flowchart illustrating an example of a processing procedure of a method for creating a group of datasets for machine learning. [Figure 5] 10 is a flowchart showing an example of a processing procedure of a first step. [Figure 6] 1A and 1B are diagrams illustrating an example of a rubber product model and a mold model. [Figure 7] 1 is a graph showing an example of multiple types of vulcanization temperature histories. [Figure 8] FIG. 2 is a diagram illustrating an example of a first data set and a first data set group. [Figure 9] FIG. 10 is a diagram illustrating an example of a principal component set group. [Figure 10] 10 is a graph showing the contribution rate of the principal component. [Figure 11] 10 is a flowchart showing an example of a processing procedure of a third step. [Figure 12] 10 is a graph showing an example of the relationship between a first principal component and a second principal component. [Figure 13] 10 is a graph showing the second selection number. [Figure 14] 10A and 10B are diagrams illustrating an example of a second data set and a second data set group. [Figure 15] FIG. 10 is a diagram showing an example of a recreated vulcanization temperature history. [Figure 16] 1 is a flowchart illustrating an example of a processing procedure for creating a trained model. [Figure 17] FIG. 1 is a conceptual diagram illustrating an example of a trained model. [Figure 18] 10 is a flowchart illustrating an example of a processing procedure of a method for predicting performance of an object. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments of the present invention will be described with reference to the drawings. It should be understood that the drawings include exaggerated representations and representations that differ from the dimensional ratios of actual structures in order to facilitate understanding of the contents of the invention. Furthermore, identical or common elements are designated by the same reference numerals throughout the embodiments, and redundant explanations will be omitted. Furthermore, the specific configurations shown in the embodiments and drawings are for the purpose of understanding the contents of the present invention, and the present invention is not limited to the specific configurations shown in the drawings.

[0010] [computer] A computer is used in the method for creating a group of datasets for machine learning and the method for creating a trained model (hereinafter, these may be referred to as "creation methods") of this embodiment. Fig. 1 is a perspective view showing an example of a computer 1 for executing the method for creating a group of datasets for machine learning and the method for creating a trained model.

[0011] The computer 1 of this embodiment includes, for example, a main body 1a, a keyboard 1b, a mouse 1c, and a display device 1d. The main body 1a is provided with, for example, a central processing unit (CPU), a read-only memory (ROM), a storage device such as a magnetic disk, and disk drives 1a1 and 1a2. The storage device also stores software for executing the creation method of this embodiment in advance. Therefore, the computer 1 is configured as a creation device 1A for creating a group of machine learning datasets and a trained model.

[0012] In the method for creating a dataset group for machine learning according to this embodiment, a dataset group consisting of a plurality of datasets is created. The plurality of datasets (dataset group) are created as training data for machine learning to predict the performance of an object. The object is obtained through a process, and its performance changes depending on the process conditions. To enable machine learning to predict the performance of such an object, a plurality of types of process conditions are specified in the plurality of datasets.

[0013] [Object] The object is not particularly limited as long as it is obtained through a process and its performance changes depending on the process conditions. An example of the object in this embodiment is vulcanized rubber.

[0014] The vulcanized rubber is not particularly limited as long as it is vulcanized rubber. An example of vulcanized rubber is a rubber material (vulcanized rubber) that constitutes a vulcanized rubber product. The rubber product of this embodiment is exemplified as a tire. Note that the rubber product is not limited to a tire, and may be, for example, a laminated rubber bearing for seismic isolation. FIG. 2 is a cross-sectional view showing an example of a rubber product 2 and vulcanized rubber 4 (object S).

[0015] The rubber product 2 is configured as a tire 2A. The tire 2A of this embodiment is configured as, for example, a pneumatic tire for a passenger car. However, the tire 2A is not limited to this form and may be configured as, for example, a pneumatic tire for heavy loads or a tire for a motorcycle. The tire 2A of this embodiment is configured to include a fibrous member 3 and vulcanized rubber 4.

[0016] The fibrous member 3 includes, for example, a carcass 3a, an inner belt 3b, and an outer belt 3c. The carcass 3a extends from the tread portion 2a through the sidewall portion 2b to the bead cores 5 of the bead portions 2c. The inner belt 3b and the outer belt 3c are disposed outside the carcass 3a in the tire radial direction and inside the tread rubber 4a.

[0017] The vulcanized rubber 4 includes a tread rubber 4a, a sidewall rubber 4b, a clinch rubber 4c, a bead apex rubber 4d, and an inner liner rubber 4e. The tread rubber 4a is disposed on the outer side of the outer belt 3c in the tread portion 2a. The sidewall rubber 4b is disposed on the outer side of the carcass 3a in the sidewall portion 2b. The clinch rubber 4c is fixed to the inner side of the sidewall rubber 4b in the tire radial direction. The bead apex rubber 4d extends from the bead core 5 outward in the tire radial direction. The inner liner rubber 4e is disposed on the inner surface of the carcass 3a.

[0018] The vulcanized rubber (rubber material) 4 of this embodiment contains, for example, a filler, a cross-linking agent, etc. Examples of the filler include silica and carbon black.

[0019] The vulcanized rubber 4 is obtained by subjecting a rubber material (not shown) before vulcanization to a process of vulcanizing the rubber material. A known mold is used for the vulcanization process. In general, the performance of the vulcanized rubber 4 varies depending on the vulcanization temperature history, which indicates the relationship between the rubber temperature and the vulcanization time when the rubber material is vulcanized. Therefore, when the vulcanized rubber 4 is the object S, as in this embodiment, the conditions of the process that change the performance of the object S include the vulcanization temperature history.

[0020] FIG. 3 is a graph showing an example of a vulcanization temperature history 11. In FIG. 3, one vulcanization temperature history 11 (process conditions 13) is shown as a representative example. The vulcanization temperature history 11 is divided into a first temperature curve 11A, a second temperature curve 11B, and a third temperature curve 11C. The first temperature curve 11A shows the temperature change of the rubber material during a temperature rise time T1, which is a time period during which the temperature of the rubber material is increased after the start of vulcanization of the rubber material. The second temperature curve 11B shows the temperature change of the rubber material during a heating time T2, which is a time period during which the temperature of the rubber material is maintained after the temperature rise time T1. The third temperature curve 11C shows the temperature change of the rubber material during a heat release time T3, which is a time period from the heating time T2 until the rubber material is removed from the mold (not shown). This vulcanization temperature history 11 can be changed as needed by adjusting the set temperature and heating time of a heat source (not shown) provided in the mold. The performance of the vulcanized rubber 4 (object S) shown in FIG. 2 changes depending on this vulcanization temperature history 11 (process conditions 13).

[0021] The data set is not particularly limited as long as it can identify the process conditions 13 (in this example, the vulcanization temperature history 11). The data set in this embodiment includes fitting parameters that configure a function that can approximate the vulcanization temperature history 11.

[0022] The function (fitting parameter) is not particularly limited as long as it can approximate the vulcanization temperature history 11. The function in this embodiment is defined by the following formulas (1) to (3).

number

[0023] The above formula (1) is for fitting to the first temperature curve 11A during temperature rise T1 shown in FIG. 3. The above formula (2) is for fitting to the second temperature curve 11B during heating T2. The above formula (3) is for fitting to the third temperature curve 11C during heat dissipation T3. By fitting the functions formed by these formulas (1) to (3) to the first temperature curve 11A to the third temperature curve 11C, the variables t2, t1, k0, k1, a, a3, a2, a1, the temperature q(t) of the rubber material, and the initial temperature q0 of the rubber material are determined. Function fitting can be easily performed, for example, based on particle swarm optimization (PSO).

[0024] The fitting parameters constituting the data set include at least one of t2, t1, k0, k1, a, a3, a2, a1, q(t), and q0 in the above formulas (1) to (3), and in this example, all of them. The data set including such fitting parameters allows the vulcanization temperature history 11 (process conditions 13), which changes from moment to moment, to be uniquely identified.

[0025] The performance of the vulcanized rubber 4 (object S) that changes depending on the vulcanization temperature history 11 includes the degree of swelling, loss tangent, and absolute value of the complex modulus. These performances affect the performance of the rubber product 2 (in this example, a tire 2A) made of the vulcanized rubber 4 shown in Fig. 2. In this embodiment, the loss tangent is predicted as the performance of the vulcanized rubber 4 (object S).

[0026] The swelling degree (toluene swelling degree) is the mass change rate of a vulcanized rubber (test piece, not shown) before and after immersion in toluene at room temperature for 24 hours (mass after immersion / mass before immersion).

[0027] The absolute values ​​of the loss tangent and complex modulus are values ​​measured using a viscoelasticity spectrometer under the conditions shown below in accordance with the provisions of JIS K6394 "Vulcanized rubber and thermoplastic rubber - Determination of dynamic properties - General guidelines." Initial distortion: 10% Amplitude: ±2% Frequency: 10Hz Deformation mode: tension Temperature: 70℃ Viscoelasticity spectrometer: GABO "Iplexar (registered trademark)"

[0028] As described above, the performance of the target object S, such as vulcanized rubber 4, varies depending on the process conditions 13, such as the vulcanization temperature history 11 shown in FIG. 3. Therefore, to improve the prediction accuracy of the learning model, it is important to create multiple data sets that cover (distribute) a wide range of the process conditions 13 (in this example, the vulcanization temperature history 11) as training data. However, in order to cover a wide range of the process conditions 13, it is possible to create a large number of data sets by assuming every possible condition, but as the number of data sets increases, there is a problem in that machine learning takes a long time.

[0029] [Method for creating a group of machine learning datasets (first embodiment)] In the creation method of this embodiment, a group of data sets that covers a wide range of process conditions 13 (in this example, vulcanization temperature history 11) is created as training data for machine learning while suppressing an increase in the number of data sets. Fig. 4 is a flowchart showing an example of the processing procedure of the method for creating a group of data sets for machine learning.

[0030] [Acquire the first set of data (first step)] In the creation method of this embodiment, first, a first data set group including first data sets of multiple different types of process conditions 13 (in this example, vulcanization temperature history 11) is acquired (first step S1). The first step S1 of this embodiment is performed by a computer 1 (shown in FIG. 1). FIG. 5 is a flowchart showing an example of the processing procedure of the first step S1.

[0031] [Acquire multiple types of vulcanization temperature history] In the first step S1 of this embodiment, first, a plurality of types of vulcanization temperature histories are acquired when a plurality of rubber materials are vulcanized under different vulcanization conditions (step S11).

[0032] The vulcanization conditions include, for example, the vulcanization temperature (temperature conditions set in the heat source (not shown) of the vulcanization mold) and the vulcanization time. Multiple types of vulcanization temperature histories can be acquired as appropriate. In step S11 of this embodiment, a vulcanization simulation step is carried out to calculate multiple types of vulcanization temperature histories.

[0033] In the vulcanization simulation process of this embodiment, the vulcanization temperature history of a rubber material (e.g., tread rubber 4a) when the rubber product 2 is vulcanized using a vulcanization mold (not shown) for vulcanizing the rubber product 2 shown in Fig. 2 is calculated by the computer 1 shown in Fig. 1. This makes it possible to accurately calculate the vulcanization temperature history when the rubber material constituting the rubber product 2 is vulcanized. Note that in the vulcanization simulation, for example, the vulcanization temperature history when the rubber material alone is vulcanized may also be calculated.

[0034] The vulcanization simulation process can be carried out as appropriate as long as the vulcanization temperature history of the rubber material can be obtained. Fig. 6 is a diagram showing an example of a rubber product model 7 and a mold model 8.

[0035] In the vulcanization simulation of this embodiment, a rubber product model 7 modeling the rubber product 2 shown in FIG. 2 and a mold model 8 modeling the vulcanization mold (not shown) are created. The rubber product model 7 and mold model 8 are discretized into a finite number of elements F(i) (i = 1, 2, ...). Next, heat transfer between the rubber product model 7 and the mold model 8 is calculated. In this heat transfer calculation, the temperature of each element F(i) is calculated for each unit step of the simulation. Then, from the elements F(i) constituting the rubber product model 7, one element F(i) constituting the rubber material to be predicted (in this example, the tread rubber model 10a) is selected. Then, based on the temperature calculated for the selected element F(i), a vulcanization temperature history 11 (shown in FIG. 3) of the rubber material (vulcanized rubber 4) constituting the rubber product 2 shown in FIG. 2 is acquired. This heat transfer calculation can be performed, for example, in a procedure similar to the first step of Patent Document (JP 2023-073083 A).

[0036] In the first step S1 of this embodiment, the vulcanization simulation step as described above makes it possible to calculate multiple types of vulcanization temperature histories 11 (shown in FIG. 3) based on different vulcanization conditions without actually vulcanizing the rubber product 2 or rubber material shown in FIG. 2. This can prevent an increase in costs required for obtaining multiple types of vulcanization temperature histories 11 and for obtaining the first data set group.

[0037] As described above, in order to improve the prediction accuracy of the learning model, it is important to create multiple data sets that cover a wide range of the process conditions 13 (in this example, the vulcanization temperature history 11) shown in Figure 3. In this embodiment, the vulcanization simulation process can be repeatedly performed with different vulcanization conditions until the desired vulcanization temperature history 11 is obtained. Therefore, multiple data sets that cover a wide range of the vulcanization temperature history 11 can be easily and reliably obtained.

[0038] The number of types of process conditions (vulcanization temperature histories 11) acquired in step S11 can be set appropriately, taking into consideration, for example, the prediction accuracy of the learning model and the calculation cost of the vulcanization simulation process. In this embodiment, 200 to 500 types (361 types in this example) of vulcanization temperature histories 11 are acquired. FIG. 7 is a graph showing an example of multiple types of vulcanization temperature histories 11. FIG. 7 shows 361 types of vulcanization temperature histories 11 (process conditions 13). Multiple types of vulcanization temperature histories 11 are stored in the computer 1 (shown in FIG. 1).

[0039] [Get the first set of data] Next, in the first step S1 of this embodiment, each of the multiple types of vulcanization temperature histories 11 (shown in FIG. 7) is approximated by a function to obtain a first data set group including a first data set consisting of multiple fitting parameter values ​​(step S12).

[0040] In step S12 of this embodiment, functions consisting of the above formulas (1) to (3) are fitted to each of a plurality of types of vulcanization temperature histories 11. As a result, a first data set consisting of values ​​of a plurality of fitting parameters is obtained. FIG. 8 is a diagram showing an example of the first data set D1 and the first data set group G1. In FIG. 8, one first data set D1 is shown as a representative. Furthermore, for each fitting parameter in FIG. 8, an appropriate value is input for reference.

[0041] In this embodiment, when 361 types of vulcanization temperature histories 11 (shown in FIG. 7) are acquired, a first data set group G1 including 361 first data sets D1 is acquired. Each first data set D1 includes a plurality of fitting parameters including t2, t1, k0, k1, a, a3, a2, a1, q(t), and q0 in the above equations (1) to (3). The first data set group G1 is stored in a computer 1 (shown in FIG. 1).

[0042] [Obtain principal component sets (step 2)] Next, in the creation method of this embodiment, the first data set group G1 shown in Fig. 8 is subjected to principal component analysis and converted into a principal component set group consisting of a plurality of principal component sets (second step S2). The second step S2 of this embodiment is performed by a computer 1 (shown in Fig. 1).

[0043] In the second step S2 of this embodiment, principal component analysis is performed on the first dataset group G1 shown in Fig. 8 based on a known procedure. In this principal component analysis, a principal component set group consisting of a plurality of principal component sets obtained by transforming a plurality of first datasets D1 (361 first datasets in this example) can be obtained based on a plurality of principal components that aggregate information contained in the plurality of first datasets D1. For the principal component analysis, for example, a library such as scikit-learn for the known programming language Python is used.

[0044] Fig. 9 shows an example of a principal component set group G3. One principal component set D3 is shown as a representative in Fig. 9. The multiple principal component sets D3 include, for each of multiple principal components, principal component scores obtained by converting the first data set D1 shown in Fig. 8. Note that appropriate values ​​have been input as reference for the principal component scores in Fig. 9.

[0045] The multiple principal components in this embodiment include principal component 1 to principal component 10. Note that the principal components are not limited to this example, and can be set appropriately depending on, for example, the results of the principal component analysis, the number of first data sets D1 (fitting parameters), etc.

[0046] The principal component scores are obtained by converting the values ​​of the fitting parameters of the first data set D1 shown in Fig. 8 for each of the multiple principal components. When 361 first data sets D1 (shown in Fig. 8) are acquired as in this embodiment, a principal component set group G3 including 361 principal component sets D3 is acquired.

[0047] In the principal component analysis, the contribution ratio of each principal component can be obtained. The contribution ratio indicates the proportion of each principal component that explains the first data set group G1 (the vulcanization temperature history 11 (shown in FIG. 7) approximated by the first data set group G1) shown in FIG. 8. The importance of each principal component can be easily understood from the contribution ratio.

[0048] Fig. 10 is a graph showing the contribution rates of the principal components. As shown in Fig. 10, the contribution rate of the first principal component is the largest. On the other hand, the contribution rate of the tenth principal component is the smallest. The principal component set group G3 and the contribution rates are stored in the computer 1 (shown in Fig. 1).

[0049] [Select multiple principal component sets (Step 3)] Next, in the creation method of this embodiment, a smaller number of principal component sets D3 than the principal component sets D3 shown in Fig. 9 is selected based on the result of the second step S2 (third step S3). The third step S3 of this embodiment is performed by the computer 1 (shown in Fig. 1).

[0050] In this embodiment, a plurality of principal component sets D3 are selected that are fewer than the plurality of principal component sets D3 (361 principal component sets in this example) converted in the second step S2. The selection of principal component sets is appropriately acquired based on the results of the second step S2. As described above, the contribution rate of the principal component analysis shown in FIG. 10 indicates the proportion of each principal component that explains the first data set group G1 shown in FIG. 8 (the vulcanization temperature history 11 shown in FIG. 7). By selecting a plurality of principal component sets D3 based on such contribution rates, it becomes possible to select a smaller number of principal component sets D3 that can comprehensively explain the conditions 13 of the plurality of processes (the vulcanization temperature history 11) than the principal component set group G3 shown in FIG. 8.

[0051] In the third step S3, it is preferable to preferentially select the principal component set D3 shown in Fig. 8 among the principal components with relatively large contribution rates among the multiple principal components obtained by the principal component analysis. This allows the principal component set D3 to be preferentially selected among the principal components with a relatively large proportion that explains the first data set group G1 shown in Fig. 8. Therefore, it becomes possible to select a small number of principal component sets D3 that can comprehensively explain the vulcanization temperature history 11 (process condition 13) shown in Fig. 7.

[0052] On the other hand, if the principal component set D3 is selected based only on the principal components with relatively large contribution rates, the selected principal component set D3 may be biased. For this reason, it is preferable to select at least one principal component set for multiple principal components without being influenced by the contribution rates of the principal component analysis. This makes it possible to select a principal component set D3 that can explain the vulcanization temperature history 11 (process condition 13) shown in Figure 7 in more detail. Figure 11 is a flowchart showing an example of the processing procedure of the third step S3.

[0053] [Identify the first choice number] In the third step S3 of this embodiment, first, a first selection number that is smaller than the number of principal component sets D3 shown in FIG. 9 is identified (step S31). The first selection number in this embodiment is the total number of the principal component sets D3 selected from the principal component set group G3 shown in FIG. 9. The first selection number can be appropriately identified as long as it is smaller than the number of the principal component sets D3 (e.g., 361). In this embodiment, the first selection number is set to, for example, 30 to 80 (50 in this example) so that a small number of principal component sets D3 that can comprehensively explain the vulcanization temperature history 11 shown in FIG. 7 are selected. The first selection number is stored in the computer 1 (shown in FIG. 1).

[0054] [Select the principal component sets with the minimum and maximum principal component scores] Next, in the third step S3 of this embodiment, for each of the multiple principal components, a principal component set D3 with the smallest principal component score and a principal component set D3 with the largest principal component score are selected (step S32). Figure 12 is a graph showing an example of the relationship between the first principal component and the second principal component. In Figure 12, only a portion of the multiple principal component set group G3 shown in Figure 9 is shown as a representative principal component set D3.

[0055] In step S32 of this embodiment, first, as shown in FIG. 12, for the first principal component, a principal component set (first principal component set) D3a with the smallest principal component score and a principal component set (second principal component set) D3b with the largest principal component score are selected. Next, in step S32, for the second principal component, a principal component set (third principal component set) D3c with the smallest principal component score and a principal component set (fourth principal component set) D3d with the largest principal component score are selected. At this time, the principal component sets already selected for the first principal component (first principal component set D3a and second principal component set D3b) are excluded from the selection candidates for the third principal component set D3c and the fourth principal component set D3d. Then, in step S32, for the third to tenth principal components, a principal component set D3 with the smallest principal component score and a principal component set D3 with the largest principal component score are selected, respectively, based on the same procedure as for the first and second principal components.

[0056] In step S32 of this embodiment, a principal component set D3 (for example, the first principal component set D3a) with the smallest principal component score and a principal component set D3 (for example, the second principal component set D3b) with the largest principal component score are selected from the first to tenth principal components. Therefore, at least one (two in this example) principal component set D3 is selected from multiple principal components, regardless of the magnitude of the contribution rate of the principal component analysis shown in Figure 10. This makes it possible to select a principal component set D3 that can explain the vulcanization temperature history 11 shown in Figure 7 in detail.

[0057] In step S32 of this embodiment, two principal component sets D3 are selected for each of the first to tenth principal components. As a result, 20 principal component sets D3 are selected in step S32. The selected principal component sets D3 are stored in the computer 1 (shown in FIG. 1).

[0058] [Specify the second choice number] Next, in the third step S3 of this embodiment, a second selection number calculated by the following formula is specified for each of the plurality of principal components (step S33). Number of second choices = Contribution rate × Number of first choices - 2

[0059] The second selection number is used to specify the number of principal component sets D3 to be selected for each of the multiple principal components (in this example, the first principal component to the tenth principal component). As described above, the contribution ratios shown in FIG. 10 indicate the proportion of each principal component that explains the first data set group G1 shown in FIG. 8. By multiplying each of these contribution ratios of each principal component by the first selection number (50 in this example), the number of principal component sets D3 to be selected based on the contribution ratios can be obtained for each principal component. By selecting each principal component set D3 based on these selection numbers, principal component sets D3 are preferentially selected for principal components with relatively large contribution ratios, and the first selection number (50 in this example) of principal component sets D3 can be selected.

[0060] In step S32, for each of the multiple principal components, the principal component set D3 with the smallest principal component score and the principal component set D3 with the largest principal component score (for example, the first principal component set D3a and the second principal component set D3b for the first principal component shown in FIG. 12) are already selected. Therefore, in the above formula, the contribution rate of each principal component is multiplied by the first selection number (50 in this example), and then the number already selected (2 in this example) is further subtracted. This determines the second selection number. By selecting a new principal component set D3 based on these second selection numbers, the principal component set D3 is preferentially selected for principal components with relatively large contribution rates, and the first selection number of principal component sets D3 (50 in this example) can be selected.

[0061] Fig. 13 is a graph showing the second selection number. Fig. 13 also shows the selection numbers (hereinafter sometimes referred to as "third selection numbers") of the principal component set D3 with the smallest principal component score and the principal component set D3 with the largest principal component score. Therefore, the sum of the second selection number and the third selection number is equal to the first selection number (50 in this example).

[0062] In Fig. 13, the first to fifth principal components have a second selection number of 1 or more, and the principal components with larger contribution rates shown in Fig. 10 have larger second selection numbers. On the other hand, the sixth to tenth principal components have a second selection number of zero. The second selection numbers are stored in computer 1 (shown in Fig. 1).

[0063] [Select the principal component set that is closest to the principal component score when divided equally by the second selection number] Next, in the third step S3 of this embodiment, when the second selection number is 1 or more, a principal component set D3 is selected for each of the plurality of principal components, the principal component set D3 being in the vicinity of the principal component score when the range between the minimum and maximum principal component scores is equally divided by the second selection number (step S34). As shown in Fig. 13, in this embodiment, the second selection number is 1 or more for the first to fifth principal components. For each of these first to fifth principal components, a principal component set D3 is selected that is in the vicinity of the principal component score when the range between the minimum and maximum principal component scores is equally divided by the second selection number.

[0064] In step S34 of this embodiment, first, as shown in FIG. 12, for the first principal component, principal component scores are identified when the distance between the minimum and maximum principal component scores is equally divided by the second selection number. As shown in FIG. 13, the second selection number for the first principal component is 13. In this case, for the first principal component, principal component scores are identified when the distance between the first principal component set D3a and the second principal component set D3b is equally divided into 13. Then, for each of the identified 13 principal component scores, a principal component set D3 that is close to the principal component score is selected. Note that if there are multiple principal component sets D3 near each identified principal component score, the principal component set D3 that is closest (having the closest principal component score) is selected. Also, if there are multiple principal component sets D3 that are close to each other, any one of the principal component sets D3 is selected. It should be noted that the principal component sets D3 that have already been selected (for example, the third principal component set D3c and the fourth principal component set D3d) are excluded from the selection targets in step S34.

[0065] Next, in step S34 of this embodiment, based on the same procedure as for the first principal component, a principal component set D3 is selected for each of the second to fifth principal components, the principal component scores of which are close to the principal component scores obtained when the range between the minimum and maximum principal component scores is equally divided by the second selection number.

[0066] In step S34 of this embodiment, the principal component set D3 is preferentially selected for principal components with relatively large contribution rates among the multiple principal components obtained by the principal component analysis. As a result, the principal component set D3 can be preferentially selected for principal components that account for a relatively large proportion of the first data set group G1 shown in FIG. 8. Furthermore, the principal component set D3 that is near the principal component score obtained when the range between the minimum and maximum principal component scores is equally divided by the second selection number is selected, so that the principal component sets D3 can be selected in a dispersed manner for each principal component. Therefore, it is possible to select a small number of principal component sets D3 that can comprehensively explain the vulcanization temperature history 11 shown in FIG. 7.

[0067] In step S34 of this embodiment, the number of principal component sets D3 selected in step S32 is subtracted from the first selection number (i.e., 50-20=30) to select the principal component sets D3. The selected principal component sets D3 are stored in the computer 1 (shown in FIG. 1).

[0068] In the third step S3 of this embodiment, a principal component set D3 with the smallest principal component score and a principal component set D3 with the largest principal component score are selected for each of the plurality of principal components in step S32. Furthermore, in the third step S3 of this embodiment, a principal component set D3 with a principal component score close to the principal component score obtained when the range between the minimum and maximum principal component scores is equally divided by the second selection number is selected in step S34. As a result, in the third step S3 of this embodiment, a principal component set D3 that can comprehensively and in detail explain the vulcanization temperature history 11 shown in FIG. 7 can be selected.

[0069] [Generate a second set of data from multiple principal component sets (Step 4)] Next, in the creation method of this embodiment, a second data set group consisting of a plurality of second data sets is generated by applying the inverse transformation of the principal component analysis to the plurality of principal component sets D3 selected in the third step S3 (fourth step S4). The fourth step S4 of this embodiment is performed by a computer 1 (shown in FIG. 1).

[0070] As described above, in the third step S3 of this embodiment, a first selection number (50 in this example) of principal component sets D3 are selected from the plurality of principal component sets D3 (361 principal component sets in this example) shown in FIG. 9. The inverse transformation of the principal component analysis is applied to the first selection number of principal component sets D3. This generates a second data set group consisting of a plurality of second data sets. For example, the above-mentioned software for principal component analysis is used for this inverse transformation.

[0071] Fig. 14 is a diagram showing an example of the second data set D2 and the second data set group G2. One second data set D2 is shown as a representative in Fig. 14. For reference, appropriate values ​​are entered for each fitting parameter in Fig. 14.

[0072] When a first selection number (50 in this example) of principal component sets D3 are selected as in this embodiment, a second dataset group G2 is generated that includes the first selection number (50 in this example) of second datasets D2. Each second dataset D2 includes multiple fitting parameters including t2, t1, k0, k1, a, a3, a2, a1, q(t), and q0 in the above formulas (1) to (3), similar to the first dataset D1 shown in FIG.

[0073] As described above, in the third step S3, a principal component set D3 (shown in FIG. 9) is selected that can comprehensively and in detail explain the multiple types of process conditions 13 (in this example, the vulcanization temperature history 11) shown in FIG. 7. By applying the inverse transformation of the principal component analysis to these principal component sets D3, a second data set group G2 can be created that includes multiple second data sets D2 that cover a wide range of the process conditions 13 (the vulcanization temperature history 11) while suppressing an increase in the number of data sets. The second data set group G2 is stored in the computer 1 (shown in FIG. 1).

[0074] [Recreate the vulcanization temperature history for each of the second data sets (step 5)] Next, in the creation method of this embodiment, a vulcanization temperature history is recreated for each of the plurality of second data sets D2 created in the fourth step S4 (fifth step S5). The fifth step S5 of this embodiment is performed by the computer 1 (shown in FIG. 1).

[0075] In the fifth step S5 of this embodiment, the values ​​of the multiple fitting parameters are substituted into the function consisting of the above formulas (1) to (3) for each of the multiple second data sets D2 shown in Fig. 14. This allows the vulcanization temperature history to be recreated.

[0076] Fig. 15 is a diagram showing an example of the recreated vulcanization temperature history 12. In this embodiment, the same number (50 in this example) of vulcanization temperature histories 12 as the number of second data sets D2 constituting the second data set group G2 shown in Fig. 14 are recreated.

[0077] The recreated vulcanization temperature history 12 is smaller in number (50 in this example) than the multiple vulcanization temperature histories 11 (361 vulcanization temperature histories in this example) shown in FIG. 7, but can cover a wide range of the multiple vulcanization temperature histories 11 shown in FIG. 7. Here, covering a wide range means that the ranges of maximum and minimum values ​​of the vulcanization time of the multiple vulcanization temperature histories 11 shown in FIG. 7 and the ranges of maximum and minimum values ​​of the temperature overlap with the vulcanization temperature history 12 shown in FIG. 15. Furthermore, in the vulcanization temperature history 12 shown in FIG. 15, a detailed vulcanization temperature history 11 in which a low-temperature heating state continues for a long period of time is created, similar to the vulcanization temperature history 11 shown in FIG. 7.

[0078] In this way, the creation method of this embodiment makes it possible to create a second data set group G2 that covers a wide range of the process conditions 13 (in this example, the vulcanization temperature history 11) shown in Fig. 7 while suppressing an increase in the number of second data sets D2 shown in Fig. 14. The recreated vulcanization temperature history 12 is stored in the computer 1 (shown in Fig. 1).

[0079] [Evaluate the second dataset] Next, in the creation method of this embodiment, the plurality of second data sets D2 shown in Fig. 14 are evaluated as to whether they are good or bad (step S6). The quality of the plurality of second data sets D2 may be evaluated by the computer 1 (shown in Fig. 1) or by an operator.

[0080] The plurality of second data sets D2 are evaluated as appropriate. In this embodiment, when it is determined that the plurality of second data sets D2 can cover the process conditions 13 (in this example, the vulcanization temperature history 11) shown in Fig. 7, the plurality of second data sets D2 are evaluated as good.

[0081] It can be determined as appropriate whether the plurality of second data sets D2 cover the process conditions 13 (in this example, the vulcanization temperature history 11) shown in Fig. 7. For example, it can be determined whether the plurality of second data sets D2 cover the process conditions 13 by comparing the vulcanization temperature history 11 (shown in Fig. 7) acquired in the first process S1 with the recreated vulcanization temperature history 12 (shown in Fig. 15).

[0082] If the plurality of second data sets D2 are evaluated as good ("Yes" in step S6), the series of processes in the creation method ends. On the other hand, if the plurality of second data sets D2 are evaluated as not good ("No" in step S6), the third step S3 to step S6 are performed again.

[0083] In the third step S3, which is performed again, a plurality of principal component sets D3 (shown in FIG. 9) are reselected so as to create a good second data set group G2 (a plurality of second data sets D2). To reselect such principal component sets D3, for example, the first selection number may be set to a large number. This increases the total number of principal component sets D3 selected from the principal component set group G3 shown in FIG. 8, making it possible to create a plurality of second data sets D2 (shown in FIG. 14) that can cover a wide range of the process conditions 13 (in this example, the vulcanization temperature history 11) shown in FIG. 7.

[0084] [Method for creating a group of machine learning datasets (second embodiment)] The first step S1 in the above-described embodiments includes a vulcanization simulation step of calculating a plurality of types of vulcanization temperature histories 11 (shown in FIG. 7), but is not limited to this. For example, the first step S1 may include a vulcanization step of actually measuring a plurality of types of vulcanization temperature histories 11.

[0085] In the vulcanization process of this embodiment, a vulcanization mold (not shown) for vulcanizing the rubber product 2 shown in Fig. 2 is used to acquire the vulcanization temperature history 11 (shown in Fig. 7) of the rubber material when the rubber product 2 is vulcanized. As a result, in this embodiment, a more realistic vulcanization temperature history 11 can be acquired compared to the vulcanization simulation processes of the previous embodiments.

[0086] [Method for creating a group of machine learning datasets (third embodiment)] In the embodiments described above, as shown in FIG. 7, the vulcanization temperature history 11 is included as a process condition 13, but the present invention is not limited to this. Generally, vulcanized rubber undergoes thermal degradation when subjected to a process in which heat is applied. The performance of such thermally deteriorated vulcanized rubber 4 changes depending on the thermal degradation temperature history (not shown), which indicates the relationship between the temperature and time when the vulcanized rubber 4 is thermally deteriorated. Therefore, the process condition 13 that changes the performance of the vulcanized rubber 4 may include the thermal degradation temperature history.

[0087] In this embodiment, the vulcanized rubber 4 is continuously heated at a predetermined temperature (for example, 70 to 120 degrees) to obtain the temperature history (thermal degradation temperature history) of the vulcanized rubber 4. By creating a group of data sets consisting of multiple data sets that specify such thermal degradation temperature history, it becomes possible to create training data for machine learning that can predict the performance of the vulcanized rubber 4 that changes due to thermal degradation.

[0088] [How to create a trained model] Next, an example of a procedure for creating a trained model (hereinafter sometimes referred to as the "creation method") will be described. A trained model is obtained through a process and is used to predict the performance of an object whose performance varies depending on the conditions of the process.

[0089] The object is not particularly limited as long as it is obtained through a process and its performance changes depending on the process conditions. The object in this embodiment is exemplified by the vulcanized rubber 4 shown in FIG. 2. Details of the vulcanized rubber 4 are as described above. FIG. 16 is a flowchart showing an example of the processing procedure of the method for creating a trained model.

[0090] [Enter multiple secondary datasets] In the creation method of this embodiment, first, a plurality of second data sets D2 shown in FIG. 14 are input to the computer 1 (shown in FIG. 1) (step S7). The second data set D2 can be created based on the processing procedure of the method for creating a group of data sets for machine learning shown in FIG. 4. This makes it possible to create a group of second data sets G2 that broadly covers the process conditions 13 (in this example, the vulcanization temperature history 11) shown in FIG. 7 while suppressing an increase in the number of second data sets D2. The plurality of second data sets D2 (second data set group G2) are input to the computer 1 (shown in FIG. 1).

[0091] [Enter multiple object performances] Next, in the creation method of this embodiment, based on a plurality of types of process conditions 13 (shown in FIG. 15), a plurality of performances of the object S obtained through a plurality of types of process conditions 13 are input to the computer 1 (shown in FIG. 1) (step S8). The plurality of types of process conditions 13 are specified by each of a plurality of second data sets D2 (shown in FIG. 14). In addition, the performance of the object S is input as the performance of the vulcanized rubber 4.

[0092] In step S8 of this embodiment, first, a plurality of unvulcanized rubber materials (not shown) are prepared. From the above-mentioned viewpoint, it is preferable that these plurality of rubber materials have the same composition. Furthermore, the dimensions of the rubber material may be determined in consideration of the measurement of the first physical property value, and the like, and the rubber material may be formed as a test piece having a predetermined size. An example of the size of the test piece is 0.5 mm thick, 10 mm long, and 5 mm wide.

[0093] Next, in step S8 of this embodiment, an unvulcanized rubber material (test piece) is vulcanized based on a plurality of vulcanization temperature histories 12 (shown in FIG. 15) recreated from a plurality of second data sets D2 (shown in FIG. 14). This produces vulcanized rubber (not shown). The rubber material is appropriately vulcanized. For example, a known dynamic rubber process analyzer (for example, the "D-RPA3000" manufactured by Montec) is used to vulcanize the rubber material of this embodiment. Such an analyzer makes it possible to easily vulcanize the rubber material while reproducing a plurality of vulcanization temperature histories 12. The rubber material may also be press-vulcanized.

[0094] Next, in step S8 of this embodiment, the performance of the vulcanized rubber 4 is measured. The performance of the vulcanized rubber 4 includes the degree of swelling, the loss tangent, and the absolute value of the complex modulus (in this example, the loss tangent). The measurement conditions for these performances are as described above.

[0095] As described above, the performance of the vulcanized rubber 4 varies depending on the vulcanization temperature history 12 (shown in FIG. 15) which indicates the relationship between the rubber temperature and the vulcanization time when the rubber material is vulcanized. Therefore, in step S8, different performances (loss tangent in this example) are input for each of the multiple vulcanization temperature histories 12.

[0096] The number of vulcanization temperature histories 12 identified in the plurality of second data sets D2 (50 in this example) is smaller than the number of vulcanization temperature histories 11 shown in Fig. 7 (361 in this example). By acquiring the performance of the vulcanized rubber 4 based on such vulcanization temperature histories 12, it is possible to suppress an increase in costs compared to, for example, acquiring the performance based on the vulcanization temperature histories 11. The performance of the vulcanized rubber 4 is stored in a computer 1 (shown in Fig. 1).

[0097] [Create a trained model] Next, in the creation method of this embodiment, a computer 1 (shown in FIG. 1) creates a trained model that has been machine-learned to estimate performance from a data set that identifies conditions 13 of an arbitrary process (step S9). To create the trained model, multiple second data sets D2 acquired in step S7 and the performance of the target object S (vulcanized rubber 4) acquired in step S8 are used as training data. FIG. 17 is a conceptual diagram showing an example of a trained model 30.

[0098] The trained model 30 of this embodiment is configured as a neural network defined based on a radial basis function (RBF). This trained model (RBF network) 30 includes an input layer 31, an output layer 32, and an intermediate layer 33, and is defined as an approximate response surface expressed by superimposing Gaussian functions.

[0099] The input layer 31 of this embodiment receives the second data set D2 shown in Fig. 14. As described above, the second data set D2 of this embodiment includes a plurality of fitting parameters including t2, t1, k0, k1, a, a3, a2, a1, q(t), and q0 in the above formulas (1) to (3).

[0100] The output layer 32 of this embodiment is capable of outputting the performance of the object S (in this example, vulcanized rubber 4) shown in Fig. 2. The performance of the object in this embodiment includes the performance (loss tangent) of the vulcanized rubber 4 vulcanized based on the vulcanization temperature history 12 regenerated from the second data set D2.

[0101] The intermediate layer 33 in this embodiment is a basis function (Gaussian function) and is generated by machine learning.

[0102] In step S9, the basis functions of the intermediate layer 33 are adjusted (trained) so as to reduce the difference between the output for the input and the true output (teaching data). In this embodiment, fitting parameters (process conditions 13) of a plurality of second data sets D2 are used as the input. As a result, the performance of the object S estimated for these inputs (in this example, the loss tangent of the vulcanized rubber 4) is output (estimated). The performance of the object S (in this example, the loss tangent of the vulcanized rubber 4) input in step S8 is used as the true output. Then, the basis functions of the intermediate layer 33 are adjusted so as to reduce the difference between the output (estimated) performance of the object S and the actual performance (true output) of the object S. For example, a gradient method or the like is used as the learning method. As a result, a trained model (approximate response function) 30 capable of estimating the performance of the object S from any process condition 13 is created.

[0103] The trained model (approximate response function) 30 is configured as an RBF network, which allows it to accurately represent even highly nonlinear relationships between input and output. Furthermore, even if abnormal data is included in the training data (multiple second data sets D2 and the performance of the target object S), the trained model 30 can generate a response function without being affected by the abnormal data because the Gaussian function is superimposed using the least squares method. Furthermore, unlike conventional neural networks, the trained model 30 does not require backpropagation, which reduces computational costs.

[0104] The trained model (approximate response function) 30 can predict, for example, the performance of an unknown object S not included in the training data by complementing it with the performance of a known object S included in the training data. Therefore, by constructing such a trained model 30 in advance, it becomes possible to predict the performance of the object S without actually manufacturing the object S.

[0105] The trained model 30 can be constructed, for example, by using commercially available computer software (for example, MATLAB manufactured by The MathWorks, Inc., or modeFRONTIER manufactured by ESTECO, Inc.).

[0106] As described above, in the method for creating a dataset group for machine learning according to this embodiment, a second dataset group G2 that covers a wide range of the process conditions 13 (in this example, the vulcanization temperature history 11) shown in FIG. 7 can be created while suppressing an increase in the number of second datasets D2 shown in FIG. 14. By using such a plurality of second datasets D2 for machine learning of the trained model 30 shown in FIG. 17, it is possible to shorten the time required to create the trained model 30 and improve the prediction accuracy of the trained model 30. The trained model 30 is stored in the computer 1 (shown in FIG. 1).

[0107] The trained model 30 can estimate the performance of the object S (in this example, the loss tangent of the vulcanized rubber 4) from a data set that identifies the conditions 13 of any process (in this example, the vulcanization temperature history 11). Then, by performing a structural analysis based on the estimated performance, the performance values ​​of the rubber product 2 shown in FIG. 2 (e.g., the rolling resistance of the tire 2A) can be calculated. Therefore, the trained model 30 is useful for calculating not only the performance of the vulcanized rubber 4, but also the performance of the rubber product 2 using the vulcanized rubber 4.

[0108] In the creation methods of the above-described embodiments, the multiple second data sets D2 acquired in step S7 and the performance of the target object S (vulcanized rubber 4) acquired in step S8 are used as training data. However, this is not a limitation. For example, in addition to the multiple second data sets D2 acquired in step S7 and the performance of the target object S (vulcanized rubber 4) acquired in step S8, additional data sets and performance, for example, from past accumulated data, may be added to the training data. This enables adjustment of the machine learning of the trained model 30. Furthermore, additional experiments may be performed based on the second data set D2, and the second data set D2 may be updated based on the results of the additional experiments. This enables creation of a trained model 30 that can estimate performance with higher accuracy.

[0109] [Method for predicting performance of an object] Next, the processing steps of a method for predicting the performance of an object (hereinafter, sometimes referred to as a "prediction method") will be shown. The object whose performance is predicted is not particularly limited as long as it is obtained through a process and its performance changes depending on the process conditions. An example of the object S in this embodiment is the vulcanized rubber 4 shown in Figure 2. Details of the vulcanized rubber 4 are as described above. Furthermore, the performance in this embodiment is not particularly limited as long as it changes depending on the process conditions (in this example, vulcanization temperature history 11). An example of the performance in this embodiment is the loss tangent of the vulcanized rubber 4. Figure 18 is a flowchart showing an example of the processing steps of a method for predicting the performance of an object.

[0110] [Enter trained model] In the creation method of this embodiment, first, the trained model 30 shown in FIG. 17 is input to the computer 1 (shown in FIG. 1) (step S10). The trained model 30 can be created based on the processing procedure of the trained model creation method shown in FIG. 16. This enables the time required to create the trained model 30 to be shortened while improving the prediction accuracy of the trained model 30. The trained model 30 is input to the computer 1 (shown in FIG. 1).

[0111] [Output the performance of the target object obtained under any process conditions] Next, in the creation method of this embodiment, the performance of the object S obtained under any process conditions is output (step S11). In step S11 of this embodiment, a data set specifying the conditions of the arbitrary process (in this example, the vulcanization temperature history 11) is input to the trained model 30. This allows the performance of the object S obtained under any process conditions (in this example, the loss tangent of the vulcanized rubber 4) to be estimated. The estimated performance of the object S is stored in the computer 1 (shown in FIG. 1). Furthermore, the estimated performance of the object S can be output (displayed), for example, on the display device 1d shown in FIG. 1.

[0112] [Determine whether the performance of the object meets the standards] Next, in the creation method of this embodiment, it is determined whether the performance of the output object S satisfies a predetermined standard (step S12). In this embodiment, whether the performance satisfies the standard is determined by a computer 1 (shown in FIG. 1), but this is not particularly limited. For example, an operator may make the determination based on the performance value output from the computer 1. The standard is set appropriately depending on the performance required of the object S.

[0113] If the performance of the object S satisfies the standard ("Yes" in step S12), the conditions of the arbitrary process used to estimate the performance (in this example, the vulcanization temperature history 11) are output (step S13). The conditions of the arbitrary process can be displayed (output) on the display device 1d shown in FIG. 1, for example. By manufacturing the object S based on such conditions of the arbitrary process, it becomes possible to manufacture an object S whose performance satisfies the standard.

[0114] On the other hand, if the performance of the object S does not satisfy the criteria ("No" in step S12), step S14 is performed to change the conditions of the arbitrary step, and steps S11 and S12 are performed again. In step S14, it is preferable to change the conditions of the arbitrary step using, for example, a known optimization algorithm. Note that in step S14, the conditions of the arbitrary step may be changed by an operator. Furthermore, in step S11, which is performed again, a data set specifying the changed conditions of the arbitrary step is input into the trained model 30. As a result, the performance of the object S obtained under the changed conditions of the arbitrary step can be output.

[0115] In this way, in the performance prediction method of this embodiment, step S14 of changing the conditions of any process is repeated until the performance of the target object S satisfies the predetermined standard. This makes it possible to predict the process conditions (appropriate manufacturing conditions) that can obtain the target object S whose performance satisfies the standard.

[0116] Although a particularly preferred embodiment of the present invention has been described in detail above, the present invention is not limited to the illustrated embodiment and can be modified and implemented in various ways. [Example]

[0117] Based on the processing procedure shown in Figure 4, a group of datasets consisting of multiple datasets specifying the conditions of multiple types of processes was created as training data for machine learning to predict the performance of an object whose performance changes depending on the process conditions (Example).

[0118] In the embodiment, a first step of acquiring a first dataset group including first datasets for multiple different types of process conditions, and a second step of performing principal component analysis on the first dataset group and converting it into a principal component set group consisting of multiple principal component sets were performed.

[0119] Furthermore, in the embodiment, a third step was performed in which a smaller number of principal component sets were selected from the plurality of principal component sets based on the results of the second step. In the third step, principal component sets were preferentially selected for principal components with relatively large contribution rates of the principal component analysis based on the processing procedure shown in Fig. 11. Then, in the embodiment, a fourth step was performed in which a second dataset group consisting of a plurality of second datasets was generated by applying the inverse transformation of the principal component analysis to the plurality of principal component sets selected in the third step.

[0120] For comparison, a set of data consisting of multiple data sets was created (Comparative Example) by performing a process of obtaining the vulcanization temperature history of tire components during tire manufacturing based on the description in a patent document (Japanese Patent No. 7331566).The common specifications are as follows:

[0121] Material: Vulcanized rubber Process conditions: Vulcanization temperature history Working Example: Number of vulcanization temperature history and first data set: 361 Number of principal component sets and second dataset groups: 50 Comparative Example: Number of vulcanization temperature histories: 361

[0122] Fig. 7 is a graph showing an example of multiple types of vulcanization temperature histories in an embodiment. Fig. 15 is a graph showing vulcanization temperature histories recreated from multiple second data sets in an embodiment. In the embodiment, a wider range of the multiple vulcanization temperature histories shown in Fig. 7 can be covered with fewer data sets than the multiple vulcanization temperature histories shown in Fig. 7. Furthermore, in the embodiment, a vulcanization temperature history in which a state of heating at a low temperature continues for a long period of time is reproduced in detail.

[0123] On the other hand, in the comparative example, it is important to create multiple data sets that cover a wide range of process conditions as training data, and a large number of data sets were created, assuming every possible condition.

[0124] In this way, in the example, it was possible to create a group of data sets that covered a wide range of process conditions while suppressing an increase in the number of data sets compared to the comparative example.

[0125] [Note] The present invention includes the following aspects.

[0126] [Invention 1] A method for creating a group of datasets, which are training data for machine learning to predict the performance of an object obtained through a process and whose performance varies depending on the conditions of the process, and which includes a plurality of datasets specifying conditions of the process, the method comprising: a first step of acquiring a first data set group including first data sets for a plurality of different types of conditions for the step; a second step of subjecting the first group of data sets to principal component analysis and converting the first group of data sets into a group of principal component sets consisting of a plurality of principal component sets; a third step of selecting a smaller number of principal component sets from the plurality of principal component sets based on the result of the second step; and a fourth step of generating a second dataset group consisting of a plurality of second datasets by applying an inverse transformation of the principal component analysis to the plurality of principal component sets selected in the third step. How to create datasets for machine learning. [Invention 2] The method for creating a group of datasets for machine learning according to aspect 1, wherein in the third step, the plurality of principal component sets are selected based on the contribution rate of the principal component analysis. [Invention 3] In the third step, the principal component set is preferentially selected from among the plurality of principal components obtained by the principal component analysis, for which the contribution rate is relatively large. [Invention 4] the plurality of principal component sets each include a principal component score obtained by transforming the first data set for each of the plurality of principal components; The method for creating a group of datasets for machine learning according to aspect 3, wherein the third step includes a step of selecting, for each of the plurality of principal components, a principal component set that minimizes the principal component score and a principal component set that maximizes the principal component score. [Invention 5] the third step includes identifying a first selection number that is less than the plurality of principal component sets; identifying a second selection number for each of the plurality of principal components, the second selection number being calculated by the following formula: and when the second selection number is 1 or more, selecting, for each of the plurality of principal components, a principal component set whose principal component score is in the neighborhood of a principal component score obtained when a range between the minimum and maximum principal component scores is equally divided by the second selection number. Number of second choices = Contribution rate × Number of first choices - 2 [Invention 6] the object is vulcanized rubber, A method for creating a group of datasets for machine learning described in any one of present inventions 1 to 5, wherein the conditions of the process include a vulcanization temperature history showing the relationship between the rubber temperature and vulcanization time when the rubber material is vulcanized for the rubber material before vulcanization of the vulcanized rubber. [Invention 7] the object is vulcanized rubber, A method for creating a group of datasets for machine learning described in any one of the present inventions 1 to 6, wherein the conditions of the process include a thermal degradation temperature history showing the relationship between temperature and time when the vulcanized rubber is thermally deteriorated. [Invention 8] the first data set includes fitting parameters constituting a function that can approximate the vulcanization temperature history; The first step is acquiring a plurality of types of vulcanization temperature histories when the plurality of rubber materials are vulcanized under different vulcanization conditions; and approximating each of the plurality of types of vulcanization temperature histories with the function to obtain the first dataset group including the first dataset consisting of values ​​of the plurality of fitting parameters. [Invention 9] The method for creating a group of datasets for machine learning according to present invention 8, further comprising a fifth step of recreating the vulcanization temperature history by substituting values ​​of the plurality of fitting parameters into the function for each of the plurality of second datasets created in the fourth step. [Invention 10] The function is defined by the following formulas (1) to (3): The method for creating a group of datasets for machine learning according to invention 8 or 9, wherein the plurality of fitting parameters include at least one of q(t), q0, a, a1, a2, a3, k0, k1, t1 and t2 in the following formulas (1) to (3).

number

[0127] S1 1st process S2 2nd process S3 3rd process S4 4th process

Claims

1. A method for creating a group of datasets, each of which specifies a plurality of types of process conditions, as training data for machine learning to predict the performance of an object that is obtained through a process and whose performance varies depending on the process conditions, comprising: a first step of acquiring a first data set group including first data sets for a plurality of different types of conditions for the step; a second step of subjecting the first group of data sets to principal component analysis and converting the first group of data sets into a group of principal component sets consisting of a plurality of principal component sets; a third step of selecting a smaller number of principal component sets from the plurality of principal component sets based on the result of the second step; and a fourth step of generating a second dataset group consisting of a plurality of second datasets by applying an inverse transformation of the principal component analysis to the plurality of principal component sets selected in the third step. How to create datasets for machine learning.

2. The method for creating a group of datasets for machine learning according to claim 1 , wherein in the third step, the plurality of principal component sets are selected based on contribution rates of the principal component analysis.

3. 3. The method for creating a dataset group for machine learning according to claim 2, wherein in the third step, the principal component set is preferentially selected for principal components having a relatively large contribution rate among a plurality of principal components obtained by the principal component analysis.

4. the plurality of principal component sets each include a principal component score obtained by transforming the first data set for each of the plurality of principal components; 4. The method for creating a dataset group for machine learning according to claim 3, wherein the third step includes a step of selecting, for each of the plurality of principal components, a principal component set that minimizes the principal component score and a principal component set that maximizes the principal component score.

5. the third step includes identifying a first selection number that is less than the plurality of principal component sets; identifying a second selection number for each of the plurality of principal components, the second selection number being calculated by the following formula: and when the second selection number is 1 or more, selecting, for each of the plurality of principal components, a principal component set whose principal component score is near a principal component score obtained when a range between a minimum value and a maximum value of the principal component scores is equally divided by the second selection number. Second selection number = contribution rate x first selection number - 2

6. the object is vulcanized rubber, 2. The method for creating a group of datasets for machine learning described in claim 1, wherein the process conditions include a vulcanization temperature history showing the relationship between the rubber temperature and vulcanization time when the rubber material is vulcanized for the rubber material before vulcanization of the vulcanized rubber.

7. the object is vulcanized rubber, The method for creating a group of datasets for machine learning according to claim 1, wherein the process conditions include a thermal degradation temperature history showing the relationship between temperature and time when the vulcanized rubber is thermally deteriorated.

8. the first data set includes fitting parameters that form a function that can approximate the vulcanization temperature history; The first step comprises: acquiring a plurality of types of vulcanization temperature histories when the plurality of rubber materials are vulcanized under different vulcanization conditions; 7. The method for creating a dataset group for machine learning according to claim 6, further comprising: a step of approximating each of the plurality of types of vulcanization temperature histories with the function to obtain the first dataset group including the first dataset consisting of values ​​of the plurality of fitting parameters.

9. 9. The method for creating a group of datasets for machine learning according to claim 8, further comprising a fifth step of recreating the vulcanization temperature history by substituting values ​​of the plurality of fitting parameters into the function for each of the plurality of second datasets created in the fourth step.

10. The functions are defined by the following formulas (1) to (3): The plurality of fitting parameters are q(t) and q 0 , a, a 1 , a 2 , a 3 , k 0 , k 1 , t 1 and t 2 The method for creating a group of machine learning datasets according to claim 8, comprising at least one of the following steps: [Equation 1] q(t): Temperature of rubber material at vulcanization time t q 0 : Initial temperature of rubber material t: vulcanization time a, a 1 , a 2 , a 3 , k 0 , k 1 , t 1 and t 2 :variable

11. The method for creating a group of datasets for machine learning according to claim 6 , wherein the first step includes a vulcanization simulation step of calculating the multiple types of vulcanization temperature histories.

12. The method for creating a group of datasets for machine learning according to claim 6 , wherein the first step includes a vulcanization step of determining the multiple types of vulcanization temperature histories by actual measurement.

13. A method for creating a trained model for predicting the performance of an object obtained through a process and whose performance changes depending on the conditions of the process, A step of inputting a plurality of second datasets created by the method for creating a group of machine learning datasets according to any one of claims 1 to 12 into a computer; inputting into the computer a plurality of performances of the object obtained through a plurality of types of conditions for the process based on a plurality of types of conditions for the process specified in each of the plurality of second data sets; and creating a trained model by machine learning the computer to estimate the performance from a data set that specifies conditions of an arbitrary process, using the plurality of second data sets and the performance as training data. How to create a trained model.

14. A method for predicting the performance of an object obtained through a process and whose performance changes depending on the conditions of the process, comprising: A step of inputting a trained model obtained by the trained model creation method according to claim 13 into a computer; and inputting a data set specifying the conditions of the arbitrary process into the trained model and outputting the performance of the object obtained under the conditions of the arbitrary process. A method for predicting the performance of an object.

15. The method for predicting performance of an object according to claim 14, further comprising repeating the step of changing conditions of the arbitrary step until the performance satisfies a predetermined standard.

16. The method for predicting performance of an object according to claim 15, further comprising the step of outputting a condition for the any process under which the performance satisfies the criterion.

Images