Tire performance prediction device, performance prediction method, and data set creation method

Abstract

A device is provided that can predict tire performance based on at least one of a skeleton image and a pattern image of a tread portion in a meridian cross section of the tire. [Solution] A tire performance prediction device includes a trained model storage unit that extracts feature quantities (27) that can be restored to data (21) from data (21) that includes at least one of a skeleton image (21A) of a tire meridian cross section and a pattern image of the tread portion, and stores trained model (24B) that has been machine-learned to output tire performance (22) from the feature quantities (27).

Description

[Technical Field]

[0001] The present invention relates to a tire performance prediction device, a performance prediction method, and a data set creation method. [Background technology]

[0002] Patent Document 1 listed below describes a tire image recognition method. This method involves inputting an image of the contact patch of a target tire to be recognized into a machine learning model that has learned the relationship between an image representing the tire contact patch and the contour of the tire contact patch, deriving an output, and estimating the contour of the contact patch of the target tire based on the derived output. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2024-048936 Summary of the Invention [Problem to be solved by the invention]

[0004] It is generally known that tire performance varies depending on the tire's framework shape in meridian cross section and its tread pattern. Therefore, if performance could be predicted from these framework shapes and tread patterns, it would be advantageous to be able to reduce the costs of tire prototypes and experiments. However, no specific method for realizing such predictions has been proposed.

[0005] The present invention has been devised in view of the above-described circumstances, and its main object is to provide a device capable of predicting tire performance based on at least one of a skeleton image of the tire at a meridian cross section and a pattern image of the tread portion. [Means for solving the problem]

[0006] The present invention is a tire performance prediction device that includes a trained model storage unit that extracts features that can be restored from data including at least one of a skeleton image of a tire at a meridian cross section and a pattern image of a tread portion, and stores a trained model that has been machine-learned to output the tire performance from the features. [Effects of the Invention]

[0007] By adopting the above-described configuration, the tire performance prediction device of the present invention can predict tire performance based on at least one of a skeleton image of the tire at a meridian cross section and a pattern image of the tread portion. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram illustrating an example of a tire performance prediction device, a trained model generation device, and a dataset creation device. [Figure 2] 1 is a tire meridian cross-sectional view showing an example of a tire. [Figure 3] 1 is a development view showing an example of a tread portion of a tire. [Figure 4] 1 is a flowchart illustrating an example of a processing procedure for a method for generating a trained model. [Figure 5] FIG. 10 is a diagram illustrating an example of a training dataset. [Figure 6] 10 is a flowchart showing an example of a processing procedure of a preparation step. [Figure 7] FIG. 1 is a conceptual diagram illustrating an example of a machine learning model (trained model). [Figure 8] 10 is a flowchart showing an example of a processing procedure of a learning step. [Figure 9] 1 is a flowchart showing an example of a processing procedure of a tire performance prediction method. [Figure 10] FIG. 10 is a diagram illustrating an example of a training dataset according to another embodiment of the present invention. [Figure 11] FIG. 10 is a conceptual diagram illustrating an example of a machine learning model (trained model) according to another embodiment of the present invention. [Figure 12] FIG. 10 is a diagram showing an example of a training dataset according to yet another embodiment of the present invention. [Figure 13] FIG. 10 is a conceptual diagram illustrating an example of a machine learning model (trained model) according to another embodiment of the present invention. [Figure 14] 10 is a flowchart illustrating an example of a processing procedure of a data set creation method. [Figure 15] FIG. 10 is a diagram illustrating an example of a data set. [Figure 16] 10 is a flowchart showing an example of a processing procedure of a tire performance prediction method according to another embodiment of the present invention. [Figure 17] 10 is a flowchart illustrating an example of a processing procedure of an optimal solution acquisition step. [Figure 18] FIG. 10 is a diagram illustrating an example of a processing procedure of an optimal solution acquisition step. 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] [Tire performance prediction device] 1 is a block diagram showing an example of a tire performance prediction device 1A, a trained model generation device 1B, and a dataset creation device 1C. The tire performance prediction device (hereinafter sometimes referred to as the "prediction device") 1A is a device for predicting tire performance. This prediction device 1A executes a tire performance prediction method (hereinafter sometimes referred to as the "prediction method") described below.

[0011] The prediction device 1A of this embodiment is also configured as a trained model generation device (hereinafter sometimes referred to as a "generation device") 1B, and further configured as a dataset creation device (hereinafter sometimes referred to as a "creation device") 1C. Note that the prediction device 1A, the generation device 1B, and the creation device 1C may each be configured independently.

[0012] The generation device 1B of this embodiment is a device for generating a trained model. The generation device 1B executes a trained model generation method (hereinafter, sometimes referred to as a "generation method") described below.

[0013] The creating device 1C of this embodiment is a device for creating a dataset to be input to a trained model. This creating device 1C executes a dataset creation method (hereinafter, sometimes referred to as a "creation method") described below.

[0014] The prediction device 1A (the generation device 1B and the creation device 1C) of this embodiment is configured as a general-purpose computer 1. Examples of the computer 1 include a desktop personal computer, a laptop personal computer, a tablet, a smartphone, and a cloud server.

[0015] The prediction device 1A (the generation device 1B and the creation device 1C) of this embodiment includes an input device 3, an output device 4, and a calculation processing device 5.

[0016] [Input section, output section, arithmetic processing unit] The input device 3 is configured as an input device. Examples of this input device include a keyboard and a mouse. The output device 4 is configured as an output device. Examples of this output device include a display device and a printer. The arithmetic processing device 5 of this embodiment is configured to include a processor (arithmetic unit) 5A that performs various calculations, a storage unit 5B that stores data, programs, etc., and a working memory 5C.

[0017] [Processor] The processor 5A in this embodiment is configured as a central processing unit (CPU), but is not particularly limited thereto and may be configured as, for example, a microprocessor or other processing unit. In this embodiment, the tire performance prediction method, the trained model generation method, and the dataset creation method are executed by one processor 5A, but may also be executed (parallel processing) by multiple processors 5A (not shown).

[0018] [Storage] The storage unit 5B is a non-volatile information storage device formed of, for example, a magnetic disk, an optical disk, an SSD, etc. The storage unit 5B is provided with a data unit 7 and a program unit 8.

[0019] [Data section] The data unit 7 of this embodiment is for storing data and the like necessary to execute the tire performance prediction method, the trained model generation method, and the dataset creation method. The data unit 7 of this embodiment includes a training dataset storage unit 7a, a trained model storage unit 7b, a data storage unit 7c, a feature storage unit 7d, a performance storage unit 7e, an input dataset storage unit 7f, and an optimal solution storage unit 7g. The data unit 7 is not limited to this configuration, and may include storage units capable of storing other data as needed, or some of these may be omitted. The data stored in each storage unit will be described in each step of the prediction method, generation method, and creation method, which will be described later.

[0020] [Program section] The program unit 8 of this embodiment is a program (application) required to execute the tire performance prediction method, the trained model generation method, and the dataset creation method. When such program unit 8 is executed by the processor (computation unit) 5A, it can cause the computer 1 (the prediction device 1A, the generation device 1B, and the creation device 1C) to function as a specific means. The program unit 8 of this embodiment includes an input unit 8a, an extraction unit 8b, an output unit 8c, a preparation unit 8d, a learning unit 8e, an evaluation unit 8f, a creation unit 8g, an optimal solution acquisition unit 8h, and a restoration unit 8i. Note that the program unit 8 is not limited to this configuration, and may include other programs as needed, or some of these programs may be omitted. The functions of each program unit 8 will be described in the respective steps of the prediction method, generation method, and creation method described below.

[0021] [tire] 2 is a tire meridian cross-sectional view showing an example of a tire 11. The tire 11 of this embodiment is exemplified as a pneumatic tire mounted on an SUV or a passenger car, but is not limited to this. The tire 11 may also be a tire of another category, such as a heavy-duty tire for trucks, buses, etc.

[0022] The tire 11 of this embodiment includes a tread portion 12, a pair of sidewall portions 13, and a pair of bead portions 14. The bead portions 14 have, for example, an annularly extending bead core 15. The tire 11 of this embodiment includes a carcass 16 and a belt layer 17.

[0023] [Carcass] The carcass 16 extends between the pair of bead portions 14. The carcass 16 includes at least one carcass ply 16A (one in this example).

[0024] The carcass ply 16A includes, for example, a main body portion 16a and a turned-up portion 16b. The main body portion 16a extends, for example, between the two bead portions 14. The turned-up portion 16b is continuous with the main body portion 16a and is turned up around the bead core 15 from the axially inner side to the outer side. A bead apex 18 extending from the bead core 15 to the radially outer side of the tire is disposed between the main body portion 16a and the turned-up portion 16b. The carcass ply 16A has a plurality of carcass cords (not shown) arranged in parallel. These carcass cords may be made of organic fiber cords such as aramid or rayon. The framework of the tire 11 may be formed by such a carcass 16 (carcass ply 16A).

[0025] [Belt layer] The belt layer 17 is disposed in the tread portion 12. The belt layer 17 includes at least one belt ply 17A, 17B, two in this embodiment. These belt plies 17A, 17B have a plurality of belt cords (not shown) arranged in parallel. Steel cords are used for these belt cords, but highly elastic organic fiber cords such as aramid and rayon can also be used as needed.

[0026] [Tire frame shape] As described above, the skeleton of the tire 11 is formed by the carcass 16 (carcass ply 16A). Therefore, in the tire meridian cross section, the contour line 19a of the carcass 16 represents the skeleton shape 20A. Furthermore, the contour line 19b of the tire outer surface 11o, the contour line 19c of the tire cavity surface 11i, and the contour line 19d of the belt layer 17 (belt plies 17A, 17B), which are formed based on the carcass 16, also represent the skeleton shape 20A.

[0027] [Tread section] Fig. 3 is a development view showing an example of the tread portion 12 of the tire 11. As shown in Fig. 2 and Fig. 3, the outer surface 12o of the tread portion 12 of this embodiment includes a contact patch 31 that contacts the road surface (not shown) and grooves 32 that are recessed radially inward from the contact patch 31. Furthermore, the outer surface 12o of the tread portion 12 may include sipes 33 that are narrower than the grooves 32 (for example, 1.5 mm or less).

[0028] As shown in FIG. 3 , the groove 32 of this embodiment includes a longitudinal groove 32A and a lateral groove 32B. The longitudinal groove 32A of this embodiment extends continuously in a zigzag pattern in the tire circumferential direction, but may extend continuously in a straight line. In this embodiment, four longitudinal grooves 32A are provided to divide the land portion 34 into five portions, but this is not limited to this. Furthermore, the lateral groove 32B extends in the land portion 34 in the tire axial direction. This divides the land portion 34 into a plurality of blocks 35.

[0029] In the tread portion 12 of this embodiment, a tread pattern 20B is formed in which a plurality of pitches (not shown), which are pattern constituent units, are arranged in the tire circumferential direction by the longitudinal grooves 32A and the lateral grooves 32B. Each pitch may be of only one type having the same circumferential length, or may be of multiple types having different circumferential lengths.

[0030] It is generally known that the performance of a tire 11 varies depending on the skeleton shape 20A (shown in FIG. 2) and the tread pattern 20B (shown in FIG. 3) of the tire 11 in a meridian cross section. The skeleton shape 20A and the tread pattern 20B can be specified at the design stage of the tire 11 (i.e., the stage before prototyping). Therefore, if the performance of the tire 11 can be predicted from the skeleton shape 20A and the tread pattern 20B, there is an advantage in that costs for prototyping and testing the tire 11 can be reduced. However, no specific method for realizing such prediction has been proposed.

[0031] In the prediction method of this embodiment, feature amounts that can be restored to data including at least one of a skeleton image (sometimes referred to as a "skeleton image") of a meridian cross section of the tire 11 and a pattern image (sometimes referred to as a "pattern image") of the tread portion 12 are extracted. The skeleton image is image data that can identify the skeleton shape 20A shown in FIG. 2. The pattern image is image data that can identify the tread pattern 20B shown in FIG. 3. Then, the performance of the tire 11 corresponding to the extracted feature amounts is predicted. This reduces the costs of tire prototyping and experiments, and can improve development efficiency.

[0032] A trained model is used to predict the performance of the tire 11. Therefore, in this embodiment, a trained model generation method is performed prior to the execution of the prediction method. Note that if a trained model has already been generated when the prediction method is executed, the generation method may be omitted.

[0033] [Method for generating trained model (first embodiment)] In the generation method of this embodiment, a trained model is generated. The trained model of this embodiment is a model trained by machine learning to extract, from data including at least one of a skeleton image and a pattern image of the tire 11, feature amounts that can be restored from the data, and output the performance of the tire 11 from the extracted feature amounts. In order to generate such a trained model, it is effective to train a machine learning model using multiple types of training datasets.

[0034] 4 is a flowchart showing an example of a processing procedure of a method for generating a trained model. Each step of the generation method of this embodiment is executed by a processor 5A included in the generation device 1B (computer 1) shown in FIG.

[0035] [Prepare the training dataset (preparation step)] In the generation method of this embodiment, first, a training dataset is prepared (preparation step S1). In the preparation step S1 of this embodiment, first, a preparation unit 8d included in the program unit 8 shown in FIG. 1 is loaded into the working memory 5C. The preparation unit 8d is a program for preparing the training dataset. When the preparation unit 8d is executed by the processor 5A, the computer 1 (generation device 1B) can function as a means for preparing the training dataset.

[0036] It is preferable that the learning dataset is acquired based on multiple types of tires 11. The categories of the multiple types of tires 11 are not particularly limited. For example, from the viewpoint of improving the prediction accuracy of the trained model (machine learning model), the categories of the multiple types of tires 11 may be the same. Furthermore, from the viewpoint of improving the versatility of the trained model (machine learning model), some of the categories of the multiple types of tires 11 may be different. It is preferable that the multiple types of tires 11 differ from each other in at least one of, for example, category, tire size, tread pattern, and internal structure (including carcass and belt layer).

[0037] FIG. 5 is a diagram showing an example of the training dataset 23. FIG. 5 shows a representative example of the training dataset 23 acquired from one tire 11 (shown in FIGS. 2 and 3). The training dataset 23 of this embodiment is used for training the above-mentioned machine learning model. Therefore, the training dataset 23 of this embodiment includes data 21 including a skeleton image 21A of the tire 11 and performance 22 of the tire 11. FIG. 6 is a flowchart showing an example of the processing procedure of the preparation step S1.

[0038] [Acquire data including at least one of the tire skeleton image and pattern image] In the preparation step S1 of this embodiment, first, as shown in Fig. 5, data 21 including at least one of a skeleton image 21A of a meridian cross section of the tire 11 and a pattern image (not shown) of the tread portion 12 is acquired (step S11). In step S11 of this embodiment, the data 21 (at least one of the skeleton image 21A and the pattern image) is acquired for each of a plurality of types of tires 11.

[0039] The data 21 may include only the skeleton image 21A, only the pattern image (not shown), or both the skeleton image 21A and the pattern image, or may further include other data. The data 21 of this embodiment includes only the skeleton image 21A.

[0040] The skeleton image 21A can be acquired as appropriate. For example, the skeleton image 21A can be acquired by actually photographing a meridian cross section of the tire 11 shown in FIG. 2, or it can be acquired from design data (CAD data) of the tire 11, or it can be a hand-drawn image by an operator. The skeleton image 21A can also be acquired from a tire model (not shown) modeled based on the design data of the tire 11. This tire model can be modeled as appropriate based on, for example, a known method (for example, the method described in Japanese Patent Application Laid-Open No. 2023-084539). The skeleton image 21A of the present embodiment can be acquired from the tire model. This allows the skeleton image 21A to be acquired at low cost and in a short time without actually manufacturing multiple types of tires 11.

[0041] The skeleton image 21A may be configured as, for example, a color image, a grayscale image, or a black-and-white image. The skeleton image 21A of this embodiment is configured as a grayscale image. Such a grayscale image has the advantage of being able to identify the skeleton shape 20A of the tire 11 with high resolution compared to a black-and-white image. On the other hand, a grayscale image has the advantage of having a smaller data size compared to a color image.

[0042] The skeleton image 21A can be acquired as appropriate. The skeleton image 21A of this embodiment can be acquired of an uninflated tire 11 (shown in FIG. 2). This allows the skeleton image 21A to be acquired in a shorter time than when the tire is inflated. The acquisition of the skeleton image 21A is not limited to a mode in which the skeleton image 21A is acquired from an uninflated tire 11. For example, the skeleton image 21A may be acquired of a tire 11 mounted on a rim and inflated to a predetermined internal pressure, or may be acquired of a tire 11 to which a predetermined load is applied. This stabilizes the shape of the tire 11, and further allows skeleton images 21A of multiple types of tires 11 to be acquired under the same conditions.

[0043] The rim (not shown) is exemplified as a genuine rim. A "genuine rim" is a rim defined for each tire by a standard system that includes the standard on which the tire 11 is based. Therefore, a genuine rim is, for example, a "standard rim" for JATMA, a "design rim" for TRA, or a "measuring rim" for ETRTO.

[0044] The internal pressure is exemplified as a normal internal pressure. The "normal internal pressure" is the air pressure determined for each tire by each standard in a standard system including the standard on which the tire 11 is based. Therefore, the normal internal pressure is, for example, the "maximum air pressure" in the case of JATMA, the maximum value listed in the table "TIRE LOAD LIMITS AT VARIOUS COLD INFLATION PRESSURES" in the case of TRA, and the "INFLATION PRESSURE" in the case of ETRTO.

[0045] The load is exemplified as a normal load. The "normal load" is the load determined for each tire by each standard in a standard system including the standard on which the tire 11 is based. Therefore, the normal load is, for example, "maximum load capacity" in the case of JATMA, the maximum value listed in the table "TIRE LOAD LIMITS AT VARIOUS COLD INFLATION PRESSURES" in the case of TRA, or "LOAD CAPACITY" in the case of ETRTO.

[0046] As shown in FIG. 5, the skeleton image 21A may include a contour line 19b of the tire outer surface 11o. This tire outer surface 11o comes into contact with the road surface (not shown) during running. Therefore, such contour line 19b has a strong correlation with the performance 22 of the tire 11. The performance 22 correlated with this contour line 19b includes, for example, the rolling resistance coefficient, vertical spring constant, lateral spring constant, air resistance value, and distribution data of the tire 11. Note that the distribution data in this embodiment indicates the distribution of physical quantities acting on the tread portion 12 when the tread portion 12 comes into contact with the road surface.

[0047] The skeleton image 21A may include a contour line 19c of the tire cavity surface 11i. The shape of the tire cavity surface 11i can identify, for example, the tire thickness between the tire outer surface 11o and the presence or absence of reinforcing rubber in a run-flat tire. Therefore, the contour line 19c of the tire cavity surface 11i has a strong correlation with performance 22 of the tire 11, including, for example, the rolling resistance coefficient, vertical spring constant, and lateral spring constant.

[0048] The skeleton image 21A may include a contour line 19a of the carcass 16. As described above, the carcass 16 forms the skeleton of the tire 11. Therefore, the contour line 19a of the carcass 16 has a strong correlation with the performance 22 including, for example, the rolling resistance coefficient, the vertical spring constant, the lateral spring constant, the air resistance value, and distribution data of the tire 11.

[0049] The skeleton image 21A may include a contour line 19d of the belt layer 17. This belt layer 17 tightens the carcass 16 to increase the rigidity of the tread portion 12. Therefore, the contour line 19d of the belt layer 17 has a strong correlation with the performance 22 including, for example, the rolling resistance coefficient of the tire 11, the contact shape of the tire 11, the vertical spring constant, the lateral spring constant, the air resistance value, and distribution data.

[0050] The skeleton image 21A of this embodiment includes a contour line 19b of the tire outer surface 11o, a contour line 19c of the tire cavity surface 11i, a contour line 19a of the carcass 16, and a contour line 19d of the belt layer 17. The skeleton image 21A may include at least one contour line selected from these contour lines 19a to 19d (for example, only the contour line 19a of the carcass 16). The skeleton image 21A may also include other contour lines (not shown), such as the bead core 15 shown in FIG. 2. As described above, the skeleton image 21A including at least one of these contour lines 19a to 19d has a strong correlation with tire performance 22.

[0051] The skeleton image 21A in this embodiment represents the skeleton shape 20A on both sides in the tire axial direction with respect to the tire equator C, but is not particularly limited thereto. The skeleton image 21A may represent, for example, the skeleton shape 20A on either side in the tire axial direction. The data 21 including the skeleton image 21A is stored in the learning dataset storage unit 7a shown in FIG. 1.

[0052] [Get tire performance] Next, in the preparation step S1 of this embodiment, the performance 22 of the tire 11 is acquired (step S12), as shown in Fig. 5. In step S12 of this embodiment, the performance 22 of each of a plurality of types of tires 11 is acquired.

[0053] The performance 22 is not particularly limited as long as it can be used to evaluate the tire 11. The performance 22 in this embodiment includes at least one of the rolling resistance coefficient, vertical spring constant, lateral spring constant, and air resistance value of the tire 11. These performances 22 affect the shapes (contours) of the tire outer surface 11o, the tire cavity surface 11i, the carcass 16, and the belt layer 17, and therefore have a strong correlation with the skeleton image 21A that can identify the skeleton shape 20A (contour lines 19a to 19d) of the tire 11.

[0054] The rolling resistance coefficient, longitudinal spring constant, and lateral spring constant can be obtained, for example, by using a known testing machine or the like. The air resistance value can be obtained, for example, by a wind tunnel test of the tire 11. These performances 22 may also be obtained by a simulation using a computer 1 (shown in FIG. 1). For such a simulation, it is preferable to use the tire model described above.

[0055] The performance 22 may also include distribution data (not shown) of physical quantities acting on the tread portion 12 when the tread portion 12 comes into contact with the road surface. The distribution data includes physical quantities acting on the outer surface 12o of the tread portion 12 shown in FIG. 3 (hereinafter, sometimes referred to as "outer surface physical quantities") and physical quantities acting on the inside of the tread radially inward of the outer surface 12o (hereinafter, sometimes referred to as "internal physical quantities"). Examples of the outer surface physical quantities include wear energy, shear force, and slip amount. Examples of the internal physical quantities include strain and stress. These physical quantities affect the shapes (contours) of the tire outer surface 11o, the tire cavity surface 11i, the carcass 16, and the belt layer 17 shown in FIG. 2, and therefore have a strong correlation with a skeleton image 21A (shown in FIG. 5) that can identify the skeleton shape 20A of the tire 11.

[0056] The distribution data (not shown) may be measured using a known measuring device (not shown), or may be determined by performing a simulation using a computer 1 (shown in FIG. 1). In the simulation, the distribution data may be calculated by bringing a tire model (not shown) that models the tire 11 into contact with a road surface model (not shown) that models the road surface (not shown) and running the tire model. Such a simulation may be performed as appropriate based on a known method (for example, the method described in JP 2023-084539 A).

[0057] The performance 22 may be at least one of the rolling resistance coefficient, vertical spring constant, lateral spring constant, air resistance value, and distribution data of the tire 11, or all of these may be acquired, or other performance data may be acquired in addition. In this embodiment, the rolling resistance coefficient 22A of the tire 11 is acquired as the performance 22. FIG. 5 shows a case where the rolling resistance coefficient 22A is "6.0". The performance 22 is stored in the learning dataset storage unit 7a shown in FIG. 1.

[0058] [Create a training dataset] Next, in the preparation step S1 of this embodiment, a learning dataset 23 (shown in FIG. 5) is created (step S13). In step S13 of this embodiment, data 21 (skeleton image 21A in this example) and performance 22 (rolling resistance coefficient 22A in this example) are combined for each of a plurality of types of tires 11. In this way, a plurality of learning datasets 23 are created.

[0059] In this embodiment, training datasets 23 of multiple types of tires 11 are used to train a machine learning model in a training step S2, which will be described later. This makes it possible to extract feature amounts that can be restored to data 21 from data 21 (skeleton image 21A in this example) of a tire to be evaluated, and to generate a trained model that can output performance 22 (rolling resistance coefficient 22A in this example) from the feature amounts. The multiple training datasets 23 are stored in the training dataset storage unit 7a shown in FIG. 1.

[0060] [Training the machine learning model (training step)] Next, in the generation method of this embodiment, learning of a machine learning model is performed using the learning dataset 23 shown in Fig. 5 (learning step S2). In learning step S2, feature amounts that can be restored to data 21 are extracted from data 21 (skeleton image 21A in this example) including at least one of a tire skeleton image 21A and a pattern image, and the machine learning model is trained so that tire performance 22 is output from the feature amounts.

[0061] In the learning step S2 of this embodiment, first, the learning dataset 23 (shown in FIG. 5) stored in the learning dataset storage unit 7a shown in FIG. 1 and a learning unit 8e included in the program unit 8 are loaded into the working memory 5C. The learning unit 8e is a program for training the machine learning model using the learning dataset 23. When the learning unit 8e is executed by the processor 5A, the computer 1 (generation device 1B) can function as a means for training the machine learning model.

[0062] The machine learning model is not particularly limited as long as it can extract features that can be restored from data 21 (skeleton image 21A in this example) including at least one of a tire skeleton image 21A and a pattern image of the tire, and output tire performance 22 from the features. Fig. 7 is a conceptual diagram showing an example of a machine learning model 24A (trained model 24B).

[0063] The machine learning model 24A of this embodiment includes a first model 25 and a second model 26. The first model 25 of this embodiment is for extracting feature quantities 27 that can be restored to data 21 from data 21 that includes at least one of a tire skeleton image 21A and a pattern image (not shown). On the other hand, the second model 26 is for outputting performance 22 from the feature quantities 27 extracted by the first model 25. Note that the machine learning model 24A is not limited to an aspect that includes the first model 25 and the second model 26, and may be configured with only one of these models, or may further include other models, depending on, for example, the algorithm employed, etc.

[0064] As described above, the first model 25 is used to extract, from data 21 including at least one of a tire skeleton image 21A and a pattern image (not shown), feature quantities 27 that can be restored to the data 21. Such a first model 25 is not particularly limited as long as it can extract feature quantities 27 that can be restored to the data 21. However, it is preferable that the feature quantities 27 are extracted using at least one of an autoencoder, a principal component analysis, an independent component analysis, a singular value decomposition, and a Fourier transform. The autoencoder, the principal component analysis, the independent component analysis, the singular value decomposition, and the Fourier transform are all well known, and the first model 25 can be easily generated.

[0065] In this embodiment, the first model 25 uses an autoencoder 41. The autoencoder 41 is one of unsupervised machine learning techniques based on a neural network. Various types of autoencoders can be used, such as a convolutional autoencoder or a variational autoencoder.

[0066] The autoencoder 41 includes an encoder 41A and a decoder 41B. The encoder 41A is for compressing the dimensions of the data 21 to extract the feature 27. On the other hand, the decoder 41B is for restoring the feature 27 to the original dimensions to obtain the data 21. The encoder 41A and the decoder 41B enable the first model 25 (autoencoder 41) to compress the dimensions of the data 21 and extract the feature 27 so that the original data 21 can be restored.

[0067] Furthermore, the feature 27 extracted by the first model 25 is composed of at least one (in this example, a plurality of) numerical data 28 that can be handled (for example, edited) by the computer 1 shown in FIG. 1. Therefore, the feature 27 can be extracted so that it can be edited as desired. By editing this feature 27, it is possible to obtain a pseudo feature 27 of a tire 11 that has a skeleton shape 20A, etc. that is different from the tire 11 from which the data 21 was obtained.

[0068] As described above, the second model 26 is for outputting the performance 22 from the feature 27 extracted by the first model 25. Such a second model 26 is not particularly limited as long as it can output the performance 22 from the feature 27, but it is preferable that the performance 22 be output based on at least one of a neural network, multiple regression, random forest regression, and deep learning, for example. These neural network, multiple regression, random forest regression, and deep learning are all well known, and the above-mentioned second model 26 can be easily generated.

[0069] The second model 26 of this embodiment is based on a neural network and outputs performance 22. The neural network can be defined as appropriate. The neural network (second model 26) of this embodiment is defined based on a radial basis function (RBF).

[0070] The second model (RBF network) 26 includes an input layer 43, an output layer 44, and an intermediate layer 45, and is defined as an approximate response surface expressed by superimposing Gaussian functions.

[0071] The input layer 43 of this embodiment receives input of the feature quantity 27. As described above, the feature quantity 27 of this embodiment is composed of at least one piece of numerical data 28. The input layer 43 of this embodiment receives input of the numerical data 28 constituting the feature quantity 27. The output layer 44 of this embodiment is capable of outputting the performance 22 of the tire 11 (in this example, the rolling resistance coefficient 22A). The intermediate layer 45 of this embodiment is a basis function (Gaussian function) and is generated by machine learning. By generating such a second model 26, it becomes possible to output the performance 22 from the feature quantity 27 extracted by the first model 25.

[0072] Next, a description will be given of an example of a machine learning procedure for the trained model 24B (the first model 25 and the second model 26). Fig. 8 is a flowchart showing an example of a procedure for the training step S2.

[0073] [Machine learning of the first model] 7, in the learning step S2 of the present embodiment, first, machine learning of the first model 25 is performed (step S21). In this step S21, machine learning of the first model 25 is performed so that feature quantities 27 that can be restored to the data 21 are extracted from the data 21 including at least one of the skeleton image 21A and the pattern image of the tire 11.

[0074] In step S21 of this embodiment, first, data 21 (skeleton image 21A in this example) included in one training dataset 23 shown in FIG. 5 out of multiple training datasets 23 (not shown) is input to an encoder 41A of a first model 25 shown in FIG. 7. As a result, features 27 obtained by dimensionally compressing the data 21 are extracted. Furthermore, the extracted features 27 are input to a decoder 41B of the first model 25. As a result, data 21 (skeleton image 21A in this example) in which the features 27 have been restored to their original dimensions is output. Then, various parameters of the first model 25 (e.g., weighting coefficients, biases, etc.) are updated so as to minimize the error between the data (estimated data) 21 output from the decoder 41B and the data (ground truth data) 21 of the training dataset 23 shown in FIG. 5.

[0075] In step S21 of this embodiment, a series of processes from inputting data 21 to updating parameters is performed using each of the data 21 in a plurality of training datasets 23 (not shown), thereby optimizing the first model 25. This allows the first model 25 to be generated that can extract features 27 that can be restored to the data 21 while compressing the dimensions of the data 21 including at least one of a skeleton image 21A and a pattern image (not shown). The first model 25 is stored in the trained model storage unit 7b shown in FIG. 1.

[0076] [Implement machine learning for the second model] Next, in the learning step S2 of this embodiment, as shown in FIG. 7, machine learning of the second model 26 is performed so as to output the tire performance 22 from the feature amount 27 (step S22).

[0077] In step S22 of this embodiment, first, data 21 (skeleton image 21A in this example) included in one training data set 23 shown in FIG. 5 among a plurality of training data sets 23 (not shown) is input to the encoder 41A of the first model 25 shown in FIG. 7. In this step S22, the first model 25 trained in the previous step S21 is used. Therefore, while the dimensionality of the data 21 is compressed, feature quantities 27 that can be restored to the data 21 can be extracted with high accuracy.

[0078] Next, in step S22 of this embodiment, the extracted feature quantity 27 is input to the second model 26. As a result, the performance 22 (in this example, the tire rolling resistance coefficient 22A) estimated by the second model 26 can be output. Next, in step S22 of this embodiment, the basis functions of the intermediate layer 45 of the second model 26 are adjusted so as to minimize the error between the output performance (estimated data) 22 and the performance 22 (ground truth data) 22 included in the training dataset 23 shown in FIG. 5. For example, a gradient method or the like is used for this adjustment.

[0079] In step S22 of this embodiment, a series of processes from inputting the data 21 to adjusting the basis functions is performed using each of a plurality of training data sets 23 (not shown), thereby optimizing the second model 26. This generates the second model 26 that can output the tire performance 22 from the feature quantities 27 extracted by the first model 25. The second model 26 is stored in the trained model storage unit 7b shown in FIG. 1.

[0080] As described above, the performance of the tire 11 varies depending on the skeleton shape 20A of the tire 11 shown in FIG. 2. Therefore, the performance 22 shown in FIG. 7 has a strong correlation with the skeleton image 21A that can identify the skeleton shape 20A. By inputting data 21 including such skeleton image 21A (in this example, data including only the skeleton image 21A) into the first model 25 of the trained model 24B, feature quantities 27 for estimating the performance 22 can be effectively extracted. Then, by inputting the feature quantities 27 extracted by the first model 25 into the second model 26 of the trained model 24B, the performance 22 can be accurately estimated. The trained model 24B including the first model 25 and the second model 26 makes it possible to estimate the performance 22 from the data 21 without requiring the experience or intuition of an expert.

[0081] Furthermore, the first model 25 of this embodiment can extract feature quantities 27 that can be restored to the data 21, and therefore, for example, the feature quantities 27 from which good performance 22 is output by the second model 26 can be restored to data 21 (skeleton image 21A) that can demonstrate the performance 22. This allows an operator or the like to easily grasp the shape (skeleton shape 20A) of the tire 11 that can demonstrate the good performance 22, which can be useful in the design and development of the tire 11.

[0082] Furthermore, the first model 25 of this embodiment can extract feature quantities 27 obtained by dimensionally compressing the data 21 (skeleton image 21A in this example) of the tire 11. In this embodiment, the feature quantities 27 are input to the second model 26, and the number of machine learning parameters of the second model 26 is reduced compared to when, for example, the data 21 (skeleton image 21A) is input directly to the second model 26. This allows the second model 26 to be made smaller.

[0083] [Tire performance prediction method (first embodiment)] Next, an example of a processing procedure for a tire performance prediction method is shown. The prediction method of this embodiment uses a trained model 24B (including a first model 25 and a second model 26 in this example) generated based on the generation method shown in Figures 4, 6, and 8.

[0084] 9 is a flowchart showing an example of a processing procedure of a tire performance prediction method. Each step of the prediction method of this embodiment is executed by a processor 5A included in the prediction device 1A (computer 1) shown in FIG.

[0085] [Enter data including either a tire skeleton image or a pattern image] In the prediction method of this embodiment, first, as shown in Fig. 7, data 21 including at least one of a skeleton image 21A and a pattern image (not shown) of the tire 11 is input (step S3). In step S3, the data 21 including at least one of the skeleton image 21A and the pattern image for the tire 11 to be evaluated is input to the prediction device 1A (computer 1) shown in Fig. 1.

[0086] In step S3 of this embodiment, first, the input unit 8a included in the program unit 8 shown in Fig. 1 is loaded into the working memory 5C. The input unit 8a is a program for inputting data 21 including at least one of a skeleton image 21A and a pattern image (not shown) shown in Fig. 7 for the tire 11 to be evaluated shown in Figs. 2 and 3. Execution of this input unit 8a by the processor 5A allows the computer 1 (prediction device 1A) to function as a means for inputting the data 21.

[0087] As shown in FIG. 7, the trained model 24B (first model 25) of this embodiment can receive data 21 including only a skeleton image 21A. Therefore, in step S3 of this embodiment, the prediction device 1A (input unit 8a) shown in FIG. 1 acquires only the skeleton image 21A as data 21 of the tire 11 to be evaluated. In this embodiment, the skeleton image 21A (shown in FIG. 7) of the tire 11 to be evaluated is acquired based on the processing procedure in step S11 of the preparation step S1 shown in FIG. 6. The acquired data 21 including the skeleton image 21A is stored in the data storage unit 7c shown in FIG. 1.

[0088] [Extracting features from data containing either a tire skeleton image or a pattern image] 7, in the prediction method of this embodiment, feature quantities 27 that can be restored to the data 21 are extracted from the data 21 including at least one of a skeleton image 21A and a pattern image (not shown) of the tire 11 (step S4). The feature quantities 27 are extracted using the first model 25 included in the trained model 24B.

[0089] In step S4 of the present embodiment, first, data 21 including at least one of a skeleton image 21A and a pattern image (in this example, only the skeleton image 21A) of the tire 11 to be evaluated, which is stored in the data storage unit 7c shown in FIG. 1, is loaded into the working memory 5C. Furthermore, a first model 25 (shown in FIG. 7) stored in the trained model storage unit 7b and an extraction unit 8b included in the program unit 8 are loaded into the working memory 5C. The extraction unit 8b is a program for inputting the data 21 including at least one of the skeleton image 21A and the pattern image into the trained model 24B (in this example, the first model 25) and extracting feature quantities 27 that can be restored to the data 21. Execution of this extraction unit 8b by the processor 5A causes the computer 1 (prediction device 1A) to function as a means for extracting feature quantities 27.

[0090] In step S4 of the present embodiment, as shown in FIG. 7, the prediction device 1A (extraction unit 8b) inputs data 21 including at least one of a skeleton image 21A and a pattern image (not shown) of the tire 11 to be evaluated to the encoder 41A of the first model 25. In the present embodiment, only the skeleton image 21A is input to the encoder 41A. This allows the data 21 (skeleton image 21A) to be dimensionally compressed, while extracting feature quantities 27 that can be restored to the original data 21. The extracted feature quantities 27 are stored in the feature quantity storage unit 7d shown in FIG. 1.

[0091] [Output tire performance corresponding to feature values] Next, in the prediction method of this embodiment, the performance 22 of the tire 11 corresponding to the feature 27 is output (step S5), as shown in Fig. 7. To output the performance 22, a trained model 24B (in this example, the second model 26) that has been machine-trained to output the performance 22 of the tire 11 from the feature 27 is used.

[0092] In step S5 of this embodiment, first, the feature 27 stored in the feature storage unit 7d shown in FIG. 1 and the second model 26 (shown in FIG. 7) stored in the trained model storage unit 7b are loaded into the working memory 5C. Furthermore, the output unit 8c included in the program unit 8 is loaded into the working memory 5C. The output unit 8c is a program for outputting the performance 22 of the tire 11 corresponding to the feature 27. When the output unit 8c is executed by the processor 5A, the computer 1 (prediction device 1A) can function as a means for outputting the performance 22 of the tire 11 to be evaluated.

[0093] In step S5 of this embodiment, as shown in FIG. 7, the prediction device 1A (output unit 8c) inputs feature quantities 27 extracted from data 21 of the tire 11 to be evaluated to the input layer 43 of the second model 26. As a result, the performance 22 of the tire 11 to be evaluated (in this example, the rolling resistance coefficient 22A of the tire 11) is output from the output layer 44 of the second model 26. Therefore, the prediction method of this embodiment makes it possible to estimate the performance 22 of the tire 11 to be evaluated without requiring the experience or intuition of a skilled person.

[0094] The performance 22 output from the second model 26 is stored in the performance storage unit 7e shown in Fig. 1. In step S5, the output performance 22 may be output (displayed) to the output device 4 including, for example, a display device. This allows an operator or the like to understand the tire performance 22, and further enables evaluation of the performance 22 of the tire 11 to be evaluated.

[0095] [Evaluate the performance data of the target tires] Next, in the prediction method of this embodiment, the quality of tire performance 22 (shown in FIG. 7) is evaluated (step S6). The quality of performance 22 may be evaluated by the prediction device 1A (computer 1) shown in FIG. 1, or may be evaluated by an operator or the like.

[0096] In step S6 of this embodiment, first, the performance 22 of the tire 11 (in this example, the tire rolling resistance coefficient 22A) stored in the performance storage unit 7e shown in FIG. 1 and an evaluation unit 8f included in the program unit 8 are read into the working memory 5C. The evaluation unit 8f is a program for determining whether the output performance 22 is good or not. Execution of the evaluation unit 8f by the processor 5A allows the computer 1 to function as a means for evaluating the performance 22.

[0097] The quality of the performance 22 can be evaluated as appropriate. In the case where the performance 22 is the rolling resistance coefficient 22A of the tire as in the present embodiment, for example, if the rolling resistance coefficient 22A is less than a predetermined threshold, it can be determined that the rolling resistance performance is excellent and good. The threshold can be set as appropriate depending on the performance (rolling resistance performance, etc.) required of the tire 11 to be evaluated shown in FIGS. 2 and 3, for example.

[0098] If the performance 22 of the tire 11 to be evaluated is determined to be good ("Yes" in step S6), the tire 11 is manufactured based on the data 21 (skeleton image 21A in this example) input to the trained model 24B (first model 25) shown in FIG. 7 (step S7). In this embodiment, the feature 27 is extracted so that it can be restored to the data 21 (skeleton image 21A in this example). Therefore, the original data 21 (skeleton image 21A) may be restored from the feature 27 input to the second model 26, and the tire 11 may be manufactured based on the restored data 21.

[0099] On the other hand, if it is determined that the performance 22 of the tire 11 to be evaluated is not good ("No" in step S6), at least a part of the shape etc. of the tire 11 to be evaluated (in this example, the skeleton shape 20A shown in FIG. 2) is changed (step S8), and steps S3 to S6 are then performed again.

[0100] When the rolling resistance coefficient 22A is output as the tire performance 22 as in this embodiment, the skeleton shape 20A of the tire 11 can be changed in step S8 so that the rolling resistance coefficient becomes smaller (so that the performance 22 becomes better). Such changes can be performed, for example, by an operator. Then, in steps S3 to S7, which are performed again, new feature quantities 27 can be extracted based on the data 21 of the changed tire 11, and the performance 22 corresponding to the feature quantities 27 can be output, as shown in FIG. 7. Therefore, a tire 11 having desired performance (for example, rolling resistance performance) can be reliably designed and manufactured.

[0101] As described above, the feature quantities 27 extracted from the data 21 are extracted so as to be arbitrarily editable. Therefore, in step S8, only the feature quantities 27 may be edited, and steps S5 to S6 may be performed again. This allows evaluation of the performance 22 of a tire whose skeleton shape 20A has been pseudo-changed, without making any specific changes to the tire skeleton shape 20A or producing a prototype.

[0102] [Method for generating trained model (second embodiment)] In the above embodiments, as shown in Fig. 7, the data 21 from which the feature amount 27 is extracted includes only the skeleton image 21A, but the present invention is not limited to this. The data 21 may further include a pattern image (not shown) of the tread portion 12 of the tire 11 shown in Fig. 3.

[0103] In the generation method of this embodiment, in step S11 (shown in FIG. 6) of acquiring data 21, a skeleton image 21A shown in FIG. 5 and a pattern image (not shown) are acquired. The skeleton image 21A can be acquired in the same procedure as in the previous embodiments.

[0104] Fig. 10 is a diagram showing an example of a training data set 23 according to another embodiment of the present invention. Fig. 10 shows a representative training data set 23 acquired from one tire 11. The pattern image 21B is configured as a grayscale image.

[0105] The pattern image 21B may be identified based on, for example, a tire 11 (shown in FIG. 3) that is not in contact with a road surface (not shown). In this case, unlike the contact patch image of Patent Document 1, the pattern image 21B does not require the tire 11 to be inflated with internal pressure and in contact with the road surface. Therefore, the pattern image 21B can be acquired in a shorter time than the contact patch image. However, the pattern image 21B is not limited to this embodiment, and may be acquired, for example, from a tire 11 that is in contact with a road surface (not shown).

[0106] The pattern image 21B is an image for identifying a pattern (tread pattern 20B) engraved on the outer surface 12o of the tread portion 12 shown in Fig. 3. Such a pattern is composed of, for example, grooves 32 (longitudinal grooves 32A and lateral grooves 32B) and sipes 33 provided on the outer surface 12o.

[0107] In the case where the tread pattern 20B is formed by arranging a plurality of pitches (not shown) in the tire circumferential direction, as in the tread portion 12, a pattern image 21B at a specific (one or more) pitches may be acquired. Alternatively, pattern images 21B at all pitches may be acquired. In this embodiment, pattern images 21B at a plurality of pitches (e.g., three pitches) are acquired. In this case, the pitches from which the pattern images 21B are acquired are appropriately selected according to the performance 22 (shown in FIG. 10 ) estimated by the machine learning model, and preferably include, for example, the pitch with the highest ground contact pressure among all pitches. This allows pattern images of a plurality of types of tires 11 to be acquired under the same conditions. The pitch with the highest ground contact pressure can be easily identified, for example, by measuring or calculating the ground contact pressure of all pitches.

[0108] As shown in Fig. 10, in the pattern image 21B of this embodiment, the contour of the tread pattern 20B (grooves 32, sipes 33, etc.) is identified for each position (pixel) on the outer surface 12o of the plurality of pitches. The contour of the tread pattern 20B is identified based on, for example, color information or shading information. The color information is used to identify the color. The shading information is used to identify the grayscale gradation.

[0109] The pattern image 21B in this embodiment is configured as a grayscale image. As a result, the contours of the tread pattern 20B (i.e., the contours of the grooves 32 and sipes 33, etc.) can be distinguished by shading in the pattern image 21B. Furthermore, compared to color images, grayscale images have the advantage of suppressing an increase in the amount of data. Note that the pattern image 21B is not limited to a grayscale image, and may be configured as, for example, a color image or a black and white image.

[0110] In the pattern image 21B of this embodiment, contours including the contact patch 31, grooves 32 (including longitudinal grooves 32A and lateral grooves 32B), and sipes 33 can be identified. These contours make it possible to identify the uneven shape (longitudinal grooves 32A, lateral grooves 32B, and sipes 33) of the tread portion 12. Such uneven shape affects the rigidity and contact shape of the tread portion 12, and therefore has a strong correlation with performance 22 (in this example, the tire rolling resistance coefficient 22A) of the tire 11, including the rolling resistance coefficient, longitudinal spring constant, lateral spring constant, air resistance value, and distribution data.

[0111] The pattern image 21B can be acquired as appropriate. For example, the pattern image 21B can be acquired by actually photographing the tread portion 12 of the tire 11 shown in FIG. 3, or can be acquired from design data (CAD data) of the tire 11, or can be a hand-drawn image by an operator. The pattern image 21B can also be acquired from a tire model (not shown) modeled based on the design data of the tire 11. This tire model can be modeled as appropriate based on, for example, the known methods described above.

[0112] The pattern image 21B in this embodiment represents the outline of the tread pattern 20B of the tread portion 12 of the tire 11 on both sides in the tire axial direction with respect to the tire equator C shown in Fig. 3, but is not particularly limited thereto. For example, the pattern image 21B may represent the outline of the tread pattern 20B of the tread portion 12 on either side in the tire axial direction with respect to the tire equator C.

[0113] In step S11 of this embodiment, the skeleton image 21A and the pattern image 21B shown in Fig. 10 are acquired for each of a plurality of types of tires 11. The data 21 including the skeleton image 21A and the pattern image 21B is stored in the learning dataset storage unit 7a shown in Fig. 1.

[0114] Next, in the generation method of this embodiment, in step S13 of the preparation step S1 shown in Fig. 6, a training data set 23 is created as shown in Fig. 10. In this step S13, the training data set 23 is created for each of a plurality of types of tires 11, combining data 21 (in this example, a skeleton image 21A and a pattern image 21B) and performance 22 (in this example, a tire rolling resistance coefficient 22A).

[0115] FIG. 11 is a conceptual diagram showing an example of a machine learning model 24A (trained model 24B) according to another embodiment of the present invention.

[0116] Next, in the generation method of this embodiment, in step S21 of the learning step S2 shown in Fig. 8, machine learning of the first model 25 shown in Fig. 11 is performed. Machine learning is performed on the first model 25 so as to extract, from the data 21 including the skeleton image 21A and the pattern image 21B of the tire 11, feature quantities 27 that can be restored to the data 21. As a result, the first model 25 is generated that can extract feature quantities 27 that can be restored to each of the skeleton image 21A and the pattern image 21B while compressing the dimensions of both the skeleton image 21A and the pattern image 21B.

[0117] Next, in the generation method of this embodiment, in step S22 of the learning step S2 shown in Fig. 8, machine learning of the second model 26 shown in Fig. 11 is performed. Machine learning is performed on the second model 26 so that the tire performance 22 is output from the feature quantities 27 extracted from the data 21 including the skeleton image 21A and the pattern image 21B. As a result, the second model 26 capable of outputting the tire performance 22 from the feature quantities 27 extracted by the first model 25 is generated.

[0118] In this embodiment, in addition to the skeleton image 21A, a pattern image 21B is input to the trained model 24B (first model 25). The skeleton image 21A and the pattern image 21B are related to each other in that they affect the rigidity and contact shape of the tread portion 12 shown in FIGS. 2 and 3. Furthermore, the skeleton image 21A and the pattern image 21B are strongly correlated with the performance 22 (tire rolling resistance coefficient 22A) output from the trained model 24B (second model 26) shown in FIG. 11. Therefore, a trained model 24B (second model 26) capable of accurately estimating the performance 22 of the tire 11 can be generated based on the feature amount 27 extracted from the skeleton image 21A and the pattern image 21B.

[0119] Furthermore, the first model 25 of this embodiment can extract feature quantities 27 that can be restored to the skeleton image 21A and the pattern image 21B. As a result, for example, from the feature quantities 27 for which good performance 22 has been output by the second model 26, the shape of the tire 11 that can exhibit the performance 22 (skeleton image 21A and pattern image 21B) can be restored. This allows an operator or the like to easily grasp the skeleton shape 20A and tread pattern 20B of the tire 11 that can exhibit the good performance 22, which can be useful in the design and development of the tire 11.

[0120] Furthermore, the first model 25 of this embodiment can extract feature quantities 27 obtained by dimensionally compressing the data 21 (in this example, the skeleton image 21A and the pattern image 21B) of the tire 11. In this embodiment, in which such feature quantities 27 are input to the second model 26, the number of machine learning parameters of the second model 26 is reduced compared to when, for example, the data 21 (the skeleton image 21A and the pattern image 21B) is input directly to the second model 26. This allows the second model 26 to be made smaller.

[0121] [Tire performance prediction method (second embodiment)] The prediction method of this embodiment uses the trained model 24B (first model 25 and second model 26) generated based on the generation method shown in FIGS.

[0122] In the prediction method of this embodiment, in step S3 shown in Fig. 9, data 21 including a skeleton image 21A and a pattern image 21B of the tire 11 to be evaluated is input to the prediction device 1A (computer 1) shown in Fig. 1, as shown in Fig. 11. In this embodiment, the skeleton image 21A and the pattern image 21B of the tire 11 to be evaluated are acquired based on the processing procedure in step S11 of the preparation step S1 shown in Fig. 6.

[0123] Next, in the prediction method of this embodiment, in step S4 shown in Fig. 9, feature quantities 27 that can be restored to the data 21 are extracted from the data 21 including the skeleton image 21A and pattern image 21B of the tire 11, as shown in Fig. 11. In this step S4, the prediction device 1A (extraction unit 8b) shown in Fig. 1 inputs the data 21 including the skeleton image 21A and pattern image 21B of the tire 11 to be evaluated to the encoder 41A of the first model 25. As a result, the feature quantities 27 that can be restored to the original data 21 can be extracted while compressing the dimensions of the data 21.

[0124] Next, in the prediction method of this embodiment, in step S5 shown in FIG. 9, the performance 22 of the tire 11 corresponding to the feature 27 is output as shown in FIG. 11. In step S5 of this embodiment, the feature 27 extracted from the data 21 of the tire 11 to be evaluated by the prediction device 1A (output unit 8c) shown in FIG. 1 is input to the input layer 43 of the second model 26. As a result, the performance 22 of the tire 11 to be evaluated (in this example, the tire rolling resistance coefficient 22A) is output from the output layer 44 of the second model 26. Therefore, in the prediction method of this embodiment, as in the previous embodiments, it is possible to estimate the performance 22 of the tire to be evaluated without requiring the experience or intuition of a skilled person.

[0125] Next, as shown in FIG. 9, if it is determined that the performance 22 of the tire 11 to be evaluated is good ("Yes" in step S6), the tire 11 is manufactured based on the skeleton image 21A and pattern image 21B input to the first model 25 (step S7). On the other hand, if it is determined that the performance 22 of the tire 11 to be evaluated is not good ("No" in step S6), at least a part of the shape (skeleton shape 20A and tread pattern 20B) of the tire 11 to be evaluated shown in FIGS. 2 and 3 is changed in step S8. Then, steps S3 to S6 are performed again. This allows the tire 11 to be reliably designed and manufactured with desired performance (for example, rolling resistance performance).

[0126] As described above, the feature quantities 27 extracted from the data 21 shown in Fig. 11 are extracted so as to be editable as desired. Therefore, in step S8, only the feature quantities 27 may be edited, and steps S5 to S6 may be performed again. This allows the performance 22 of a tire whose skeleton shape 20A or tread pattern 20B has been pseudo-changed to be evaluated without making any specific changes or prototyping of the skeleton shape 20A or tread pattern 20B.

[0127] [Method for generating trained model (third embodiment)] 7 and 11, the second model 26 is exemplified as outputting the performance 22 from only the feature quantities 27 extracted from the data 21. However, the present invention is not limited to such an embodiment. The second model 26 may output the performance 22 from, for example, the feature quantities 27 and design factors of the tire 11 (hereinafter, sometimes referred to as "design factors").

[0128] In the generation method of this embodiment, in step S11 (shown in FIG. 6) of acquiring data 21, data 21 including a skeleton image 21A shown in FIG. 5 and design factors (not shown) are acquired. The skeleton image 21A can be acquired in the same procedure as in the previous embodiments.

[0129] Fig. 12 is a diagram showing an example of a training data set 23 according to yet another embodiment of the present invention. In Fig. 12, the training data set 23 acquired from one tire 11 is shown as a representative example.

[0130] The design factors 46 are not particularly limited as long as they relate to the design conditions and / or evaluation conditions of the tire. The design factors 46 in this embodiment include, for example, at least one of the tread gauge, category, load, internal pressure, aspect ratio, tire width, outer diameter, complex modulus of elasticity E* of the tread portion, rim width, running speed, and camber angle. These design factors affect the rigidity and contact shape of the tread portion 12. Therefore, the design factors 46 have a strong correlation with the performance 22 (in this example, the tire rolling resistance coefficient 22A) of the tire 11, such as the rolling resistance coefficient, vertical spring constant, lateral spring constant, air resistance value, and distribution data.

[0131] The design factors 46 are composed of numerical data 47 that specify the tread gauge, category, load, internal pressure, aspect ratio, tire width, outer diameter, and complex modulus of elasticity E* of the tread portion. Therefore, the numerical data 47 of the design factors 46 can be treated in the same way as the numerical data 28 of the feature quantities 27 shown in Fig. 11. Therefore, unlike the skeleton image 21A and the pattern image 21B, the design factors 46 do not require extraction of restorable feature quantities 27, and therefore the time required to create the training dataset 23 can be shortened.

[0132] In step S11 of this embodiment, data 21 including skeleton images 21A and design factors 46 are acquired for multiple types of tires 11. The data 21 and design factors 46 are stored in the learning dataset storage unit 7a shown in FIG.

[0133] Next, in the generation method of this embodiment, in step S13 of the preparation step S1 shown in Fig. 6, the training data set 23 shown in Fig. 5 is created. In this step S13, the training data set 23 is created for each of a plurality of types of tires 11, as shown in Fig. 12, by combining data 21 including a skeleton image 21A, design factors 46, and performance 22 (tire rolling resistance coefficient 22A).

[0134] FIG. 13 is a conceptual diagram showing an example of a machine learning model 24A (trained model 24B) according to another embodiment of the present invention.

[0135] Next, in the generation method of this embodiment, in step S21 of the learning step S2 shown in Fig. 8, machine learning of a first model 25 is performed so as to extract, from data 21 including a skeleton image 21A, feature amounts that can be restored to the data 21, as shown in Fig. 13. In step S21 of this embodiment, the skeleton image 21A is subjected to dimensional compression, and a first model 25 that can extract feature amounts 27 that can be restored to the skeleton image 21A is generated.

[0136] Next, in the generation method of this embodiment, in step S22 of the learning step S2 shown in Fig. 8, machine learning of the second model 26 is performed so as to output the tire performance 22 from the feature amounts 27 extracted from the data 21 including the skeleton image 21A and the design factors 46. In this step S22, the second model 26 capable of outputting the tire performance 22 is generated from the feature amounts 27 extracted by the first model 25 and the design factors 46.

[0137] In this embodiment, numerical data including numerical data 28 of the feature quantities 27 extracted from the skeleton image 21A and numerical data 47 of the design factors 46 are input to the input layer 43 of the trained model 24B (second model 26). The skeleton image 21A and the design factors 46 are related to each other in that they affect the rigidity, contact shape, etc. of the tread portion 12, and further have a strong correlation with the performance 22 (in this example, the rolling resistance coefficient 22A of the tire) output from the second model 26. Therefore, a trained model 24B (second model 26) capable of accurately estimating the performance 22 of the tire 11 can be generated based on the feature quantities 27 extracted from the skeleton image 21A and the design factors 46.

[0138] The first model 25 of this embodiment can extract feature quantities 27 that can be restored to the skeleton image 21A. As a result, for example, from feature quantities 27 for which good performance 22 has been output by the second model 26, the shape of the tire 11 (skeleton image 21A) that can demonstrate that performance 22 can be restored, and therefore the shape can be easily grasped by an operator or the like. Furthermore, the first model 25 of this embodiment can extract feature quantities 27 obtained by dimensionally compressing the data 21 of the tire 11 (skeleton image 21A in this example), and therefore the second model 26 can be made smaller.

[0139] Furthermore, because the design variables 46 are composed of numerical data 47, like the features 27, there is no need to extract reconstructable features 27, as is the case with the skeleton image 21A and the pattern image 21B. This can reduce the time required to create the training dataset 23 and the time required to generate the trained model 24B (the first model 25 and the second model 26).

[0140] [Tire performance prediction method (third embodiment)] The prediction method of this embodiment uses the trained model 24B (first model 25 and second model 26) generated based on the generation method shown in FIGS.

[0141] In the prediction method of this embodiment, in step S3 shown in Fig. 9, data 21 including a skeleton image 21A of the tire 11 to be evaluated and design factors 46 are input to the prediction device 1A (computer 1) shown in Fig. 1, as shown in Fig. 13. In this embodiment, the skeleton image 21A of the tire 11 to be evaluated and the design factors 46 are acquired based on the processing procedure in step S11 of the preparation step S1 shown in Fig. 6.

[0142] Next, in the prediction method of this embodiment, in step S4 shown in Fig. 9, feature quantities 27 that can be restored to the data 21 are extracted from the data 21 including the skeleton image 21A of the tire 11, as shown in Fig. 13. In this step S4, the prediction device 1A (extraction unit 8b) shown in Fig. 1 inputs the data 21 including the skeleton image 21A of the tire 11 to be evaluated to the encoder 41A of the first model 25. As a result, the feature quantities 27 that can be restored to the original data 21 can be extracted while compressing the dimensions of the data 21.

[0143] Next, in the prediction method of this embodiment, in step S5 shown in FIG. 9, the performance 22 of the tire 11 corresponding to the feature 27 and the design factor 46 is output, as shown in FIG. 13. In step S5 of this embodiment, the feature 27 extracted from the data 21 of the tire 11 to be evaluated and the design factor 46 are input to the input layer 43 of the second model 26 by the prediction device 1A (output unit 8c). As a result, the performance 22 of the tire 11 to be evaluated (in this example, the tire rolling resistance coefficient 22A) is output from the output layer 44 of the second model 26. Therefore, in the prediction method of this embodiment, it is possible to estimate the performance 22 of the tire to be evaluated without requiring the experience or intuition of a skilled person, as in the previous embodiments.

[0144] Next, as shown in FIG. 9, if it is determined that the performance 22 of the tire 11 to be evaluated is good ("Yes" in step S6), the tire 11 is manufactured based on the skeleton image 21A and design factors 46 input into the first model 25 shown in FIG. 13 (step S7). On the other hand, if it is determined that the performance 22 of the tire 11 to be evaluated is not good ("No" in step S6), at least a part of the shape (skeleton shape 20A) of the tire 11 to be evaluated shown in FIG. 2 and the design factors 46 shown in FIG. 12 are changed in step S8. Then, steps S3 to S6 are performed again. This allows the tire 11 to be reliably designed and manufactured with desired performance (for example, rolling resistance performance).

[0145] As described above, the feature quantities 27 extracted from the data 21 shown in Fig. 13 are extracted so as to be arbitrarily editable. Therefore, in step S8, only the feature quantities 27 and the design factors 46 may be edited, and steps S5 to S6 may be performed again. This allows the performance 22 of a tire in which the skeleton shape 20A and the design factors 46 have been changed in a simulated manner to be evaluated, without the need to specifically change the skeleton shape 20A of the tire 11 or to produce a prototype.

[0146] Furthermore, similar to the data 21, the design variables 46 may have features that can be restored to the design variables 46 by the first model 25. This reduces the number of machine learning parameters of the second model 26, making it possible to miniaturize the second model 26.

[0147] [Dataset creation method] In the prediction method (prediction device 1A) of the above embodiment, the tire 11 (skeleton shape 20A, pattern image 21B, and design factors 46) is changed and new feature quantities 27 are extracted so as to improve the performance 22, but the present invention is not limited to such an embodiment. For example, an optimal solution of the feature quantities 27 may be obtained so that the performance 22 satisfies a predetermined standard. By restoring such an optimal solution of the feature quantities 27 to the data 21 (skeleton shape 20A), it is possible to easily obtain data 21 whose performance 22 satisfies the standard.

[0148] The optimal solution of the feature quantity 27 is obtained based on, for example, a known optimization method. Examples of the optimization method include a genetic algorithm (GA), Bayesian optimization (BO), and particle swarm optimization (PSO). The optimization method of this embodiment uses a genetic algorithm.

[0149] In genetic algorithms, genetic operations such as crossover are repeatedly performed on multiple chromosome information expressed by genes. This makes it possible to evolve chromosome information over time from a small number of samples, making it possible to efficiently obtain optimal solutions.

[0150] In this embodiment, the feature 27 shown in Fig. 11 is treated as chromosome information, and the numerical data 28 constituting the feature 27 is treated as genes. As described above, in a genetic algorithm, genetic manipulation is performed on a plurality of chromosome information (in this example, feature 27), so it is preferable to create a dataset having a plurality of mutually different feature 27. To create such a dataset, the following dataset creation method is carried out.

[0151] In the dataset creation method (hereinafter sometimes referred to as the "creation method") of this embodiment, a dataset is created for input that is necessary for a trained model 24B (shown in FIG. 11) that has been machine-learned to output performance 22 of a tire 11 to output performance 22. This dataset is made up of a plurality of feature quantities 27 (one feature quantity 27 is shown in FIG. 11), and is used to find an optimal solution for the feature quantity 27.

[0152] Fig. 14 is a flowchart showing an example of the processing procedure of a data set creation method. Each step of the creation method of this embodiment is executed by a processor 5A included in the creation device 1C (computer 1) shown in Fig. 1. Fig. 15 is a diagram showing an example of a data set 29.

[0153] [Enter data including either a tire skeleton image or a pattern image] In the prediction method of this embodiment, first, data 21 including at least one of a skeleton image 21A and a pattern image 21B of the tire 11 is input (step S31). In step S31, data 21 including at least one of a skeleton image 21A and a pattern image 21B (in this example, only the skeleton image 21A) for the tire 11 to be evaluated is input.

[0154] In step S31 of this embodiment, first, the input unit 8a included in the program unit 8 shown in Fig. 1 is loaded into the working memory 5C. The input unit 8a is a program for inputting data 21 including at least one of a skeleton image 21A and a pattern image 21B for the tire 11 to be evaluated. When the input unit 8a is executed by the processor 5A, the computer 1 (prediction device 1A) can function as a means for inputting the data 21.

[0155] In the trained model 24B (first model 25) of this embodiment, only the skeleton image 21A can be input as the data 21. Therefore, in step S31 of this embodiment, the prediction device 1A (input unit 8a) shown in FIG. 1 acquires the skeleton image 21A of the tire 11 to be evaluated. In this embodiment, the skeleton image 21A of the tire 11 to be evaluated is acquired based on the processing procedure in step S11 of the preparation step S1 shown in FIG. 6. The acquired data 21 including the skeleton image 21A is stored in the data storage unit 7c shown in FIG. 1.

[0156] [Extracting features from data containing at least one of skeleton images and pattern images] 15, in the creation method of this embodiment, feature quantities 27 that can be restored to the data 21 and can be arbitrarily edited are extracted from the data 21 including at least one of a skeleton image 21A and a pattern image 21B of the tire 11 (step S32). A first model 25 is used to extract the feature quantities 27.

[0157] In step S32 of the present embodiment, first, data 21 including at least one of a skeleton image 21A and a pattern image 21B (in this example, only the skeleton image 21A) of the tire 11 to be evaluated, which is stored in the data storage unit 7c shown in FIG. 1, is loaded into the working memory 5C. Furthermore, a first model 25 (shown in FIG. 15) stored in the trained model storage unit 7b and an extraction unit 8b included in the program unit 8 are loaded into the working memory 5C. The extraction unit 8b is a program for inputting the data 21 including at least one of the skeleton image 21A and the pattern image 21B into the trained model 24B (in this example, the first model 25) and extracting feature quantities 27 that can be restored to the data 21 and that can be arbitrarily edited. Execution of the extraction unit 8b by the processor 5A causes the computer 1 (the prediction device 1A) to function as a means for extracting feature quantities 27.

[0158] In step S32 of this embodiment, the prediction device 1A (extraction unit 8b) inputs data 21 including at least one of the skeleton image 21A and the pattern image 21B of the tire 11 to be evaluated (in this example, only the skeleton image 21A) to the encoder 41A of the first model 25. This allows the data 21 to be dimensionally compressed and the feature amount 27 that can be restored to the original data 21 to be extracted.

[0159] As described above, the feature 27 extracted by the first model 25 is composed of at least one piece of numerical data 28 that can be handled (including edited) by the computer 1 (shown in FIG. 1). Therefore, the feature 27 can be extracted in an editable manner.

[0160] In step S32 of this embodiment, one piece of data 21 (skeleton image 21A) is input to the first model 25, thereby extracting one feature amount 27. The extracted feature amount 27 is stored in the feature amount storage unit 7d shown in FIG.

[0161] [Creating a dataset with multiple distinct features] Next, in the creation method of this embodiment, as shown in Fig. 15, the extracted feature 27 is edited to create a data set 29 having a plurality of feature quantities 27 that are different from one another (step S33). In step S33 of this embodiment, one extracted feature quantity 27 is copied multiple times, and each of the copied feature quantities 27 is edited to create a data set 29 having a plurality of feature quantities 27 that are different from one another.

[0162] In step S33 of this embodiment, first, the features 27 (shown in FIG. 15) stored in the feature storage unit 7d shown in FIG. 1 and a creation unit 8g included in the program unit 8 are loaded into the working memory 5C. The creation unit 8g is a program for creating a data set 29 having a plurality of different features 27 by editing the features 27 shown in FIG. 15. When the creation unit 8g is executed by the processor 5A, the computer 1 (creation device 1C) can function as a means for creating the data set 29.

[0163] In step S33 of this embodiment, first, one feature 27 extracted in step S32 is copied multiple times. Next, for the copied feature 27, at least one piece of numerical data 28 constituting each feature 27 is changed. This creates a data set 29 having multiple feature 27 that are different from one another. These multiple feature 27 indicate characteristics of tires 11 whose shapes (skeleton shapes 20A) shown in FIG. 2 are different from one another.

[0164] The number of feature quantities 27 constituting the data set 29 can be set appropriately in consideration of, for example, the efficiency of optimization, etc. The number of feature quantities 27 can be set to, for example, 30 to 100.

[0165] The numerical data 28 constituting the feature 27 may be edited as appropriate. For example, the numerical data 28 constituting the feature 27 may be edited based on the genetic manipulation (crossover or mutation) of a genetic algorithm. This can prevent convergence to a local optimum solution.

[0166] In this way, in the creation method of this embodiment, the feature values ​​27 extracted from the shape (skeleton image 21A) of one tire 11 are edited to create a dataset 29 including the feature values ​​27 of multiple tires 11 having different shapes (skeleton shapes 20A). This eliminates the need to design and prototype multiple tires 11 and extract multiple feature values ​​27 from the data 21 of those tires 11, for example. Therefore, the dataset 29 including multiple feature values ​​27 can be created easily and in a short time. The dataset 29 is stored in the input dataset storage unit 7f shown in FIG. 1.

[0167] [Tire performance prediction method (fourth embodiment)] The prediction method of this embodiment uses the trained model 24B (first model 25 and second model 26) generated based on the generation method shown in FIGS.

[0168] 16 is a flowchart showing an example of a processing procedure of a tire performance prediction method according to another embodiment of the present invention. Each step of the prediction method according to this embodiment is executed by a processor 5A included in the prediction device 1A (computer 1) shown in FIG.

[0169] [Create dataset (creation step)] In the prediction method of this embodiment, first, an input dataset 29 necessary for causing the trained model 24B (second model 26) to output the performance 22 is created (creation step S41). In the creation step S41 of this embodiment, a dataset 29 having a plurality of mutually different feature quantities is created based on the creation method shown in FIG. 14, as shown in FIG. 15. Note that if the dataset 29 has already been created, the creation step S41 may be omitted. The dataset 29 is stored in the input dataset storage unit 7f shown in FIG. 1.

[0170] [Find the optimal solution for features that meet the performance criteria (optimal solution acquisition step)] Next, in the generation method of this embodiment, an optimal solution of the feature 27 is obtained so that the performance 22 satisfies a predetermined standard (optimal solution obtaining step S42). In the optimal solution obtaining step S42 of this embodiment, the optimal solution of the feature 27 can be obtained based on the known optimization method described above (in this example, a genetic algorithm).

[0171] In the optimal solution acquisition step S42 of this embodiment, first, the dataset 29 (shown in FIG. 15) stored in the input dataset storage unit 7f shown in FIG. 1 is loaded into the working memory 5C. Furthermore, in the optimal solution acquisition step S42, the second model 26 (shown in FIG. 13) stored in the trained model storage unit 7b and an optimal solution acquisition unit 8h included in the program unit 8 are loaded into the working memory 5C. The optimal solution acquisition unit 8h is a program for obtaining an optimal solution of the feature quantity 27 so that the performance 22 satisfies a predetermined standard. When the optimal solution acquisition unit 8h is executed by the processor 5A, the computer 1 (prediction device 1A) can function as a means for obtaining an optimal solution of the feature quantity 27. FIG. 17 is a flowchart showing an example of the processing procedure of the optimal solution acquisition step S42. FIG. 18 is a diagram for explaining an example of the processing procedure of the optimal solution acquisition step S42.

[0172] [Output tire performance for each of multiple features] In the optimal solution acquisition step S42 of this embodiment, first, as shown in Fig. 18, for each of a plurality of feature quantities 27 included in the data set 29, the performance 22 of the tire 11 corresponding to each feature quantity 27 is output (step S51). The second model 26 included in the trained model 24B is used to output the performance 22.

[0173] In step S51 of this embodiment, the prediction device 1A (output unit 8c) shown in FIG. 1 inputs the plurality of feature quantities 27 included in the data set 29 to the input layer 43 of the second model 26, as shown in FIG. 18. As a result, the performance 22 of the tire 11 (in this example, the rolling resistance coefficient 22A of the tire 11) estimated for each of the plurality of feature quantities 27 is output from the output layer 44 of the second model 26. The output performance 22 is stored in the performance storage unit 7e shown in FIG. 1.

[0174] [Determine whether the tire performance output for each feature meets the criteria] Next, in the optimal solution acquisition step S42 of this embodiment, it is determined whether or not at least one of the performances 22 of the tire 11 output for each of the plurality of feature quantities 27 satisfies a predetermined criterion (step S52). The criterion is determined as appropriate depending on the performance 22 required of the tire 11. In the case where the performance 22 is the rolling resistance coefficient 22A of the tire as in this embodiment, for example, the rolling resistance coefficient 22A of the tire is determined to satisfy the criterion when it is less than a predetermined threshold value.

[0175] If it is determined that at least one of the performances 22 of the tire 11 output for each of the plurality of feature amounts 27 satisfies the criteria ("Yes" in step S52), the feature amount 27 that outputs the best performance 22 is selected as the optimum solution (step S53). The optimum solution (feature amount 27) is stored in the optimum solution storage unit 7g shown in Fig. 1. Then, the series of processes in the optimum solution acquisition step S42 is completed.

[0176] If it is determined that all of the performances 22 output for each of the plurality of feature quantities 27 do not satisfy the criteria ("No" in step S52), step S54 is performed to perform genetic manipulation on the plurality of feature quantities 27 included in the dataset 29. Then, steps S51 and S52 are performed again.

[0177] [Genetic manipulation of multiple features] Next, in the optimal solution acquisition step S42 of this embodiment, genetic manipulation is performed on the plurality of feature quantities 27 included in the data set 29 shown in Fig. 18 (step S54). In this embodiment, in step S54, genetic manipulation including selection, crossover, and mutation is performed on the plurality of feature quantities 27 included in the data set 29 based on a known genetic algorithm.

[0178] Selection is a genetic operation that selects high-performance features 27 or a specific number of features 27 from multiple features 27 included in a dataset 29. The selected features 27 are included in a new next-generation dataset 30. Crossover is a genetic operation that combines numerical data 28 of multiple features 27 included in a dataset 29 to generate new features 27. The new features 27 are included in a new next-generation dataset 30. Mutation is a genetic operation that randomly changes the numerical data 28 of features 27 included in a dataset 29 to generate new features 27. The new features 27 are included in a new next-generation dataset 30.

[0179] In step S54, genetic manipulation is performed on a plurality of features 27 included in the dataset 29, thereby creating a new dataset 30 including features 27 obtained by evolving the numerical data 28 so that the performance 22 satisfies a standard. This new dataset 30 is then replaced with the dataset 29 for outputting the performance 22 in step S51, which is performed again. The new dataset 30 is stored in the input dataset storage unit 7f shown in FIG. 1.

[0180] In the optimal solution acquisition step S42 of this embodiment, steps S51 to S52 are performed again after step S54 of performing genetic manipulation is performed. In step S51, which is performed again, tire performance 22 corresponding to a plurality of feature quantities 27 is output for the dataset 29 replaced from the new dataset 30. Then, genetic manipulation (optimal solution acquisition step S42) is performed until the performance 22 satisfies the criterion, and therefore an optimal solution for the feature quantities 27 whose performance 22 satisfies the criterion is reliably found.

[0181] [Restore the optimal solution to the data] Next, in the prediction method of this embodiment, the optimal solution of the feature 27 is restored to the data 21 (skeleton image 21A in this example) as shown in Fig. 15 (step S43). The first model 25 included in the trained model 24B is used for the restoration to the data 21.

[0182] In step S43 of this embodiment, first, the optimal solution of the feature 27 stored in the optimal solution storage unit 7g shown in FIG. 1 and the first model 25 (shown in FIG. 15) stored in the trained model storage unit 7b are loaded into the working memory 5C. Furthermore, a restoration unit 8i included in the program unit 8 is loaded into the working memory 5C. The restoration unit 8i is a program for restoring the optimal solution of the feature 27 to the data 21. When the restoration unit 8i is executed by the processor 5A, the computer 1 (prediction device 1A) can function as a means for restoring the data 21.

[0183] In step S43, the feature 27 selected as the optimal solution is input to the decoder 41B of the first model 25. This allows data 21 (in this example, a skeleton image 21A) in which the feature 27 has been restored to its original dimension to be acquired. Therefore, a skeleton image 21A whose performance 22 satisfies the standard is acquired.

[0184] [Manufacturing tires] Next, in the prediction method of this embodiment, the tire 11 is manufactured based on the restored data 21 (skeleton image 21A) (step S44). The data 21 is the optimal solution of the feature quantity 27 restored to its original dimension. By manufacturing the tire 11 based on such data 21, a new tire 11 having the desired performance 22 (e.g., rolling resistance coefficient 22A) can be reliably designed and manufactured without being influenced by the experience or intuition of a skilled person.

[0185] 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.

[0186] [Note] The present invention includes the following aspects.

[0187] [Invention 1] A tire performance prediction device, and a trained model storage unit that extracts feature amounts that can be restored from data including at least one of a skeleton image of a tire at a meridian cross section and a pattern image of a tread portion, and stores a trained model that has been machine-learned to output tire performance from the feature amounts. Tire performance prediction device. [Invention 2] The tire performance prediction device according to the first aspect of the present invention, wherein the feature amount is extracted in an editable manner. [Invention 3] 3. The tire performance prediction device according to claim 1, wherein the trained model includes a first model that extracts the feature amount by dimensionally compressing the data. [Invention 4] The tire performance prediction device according to aspect 3, wherein the first model extracts the feature quantities using at least one of an autoencoder, a principal component analysis, an independent component analysis, a singular value decomposition, and a Fourier transform. [Invention 5] 5. The tire performance prediction device according to any one of claims 1 to 4, wherein the trained model includes a second model that outputs the performance from the feature amount. [Invention 6] The tire performance prediction device according to aspect 5, wherein the second model outputs the performance based on at least one of a neural network, multiple regression, random forest regression, and deep learning. [Invention 7] 7. The tire performance prediction device according to claim 5 or 6, wherein the second model outputs the performance from the feature amount and design factors of the tire. [Invention 8] 8. The tire performance prediction device according to any one of claims 1 to 7, wherein the performance includes distribution data of physical quantities acting on the tread portion when the tread portion comes into contact with the road surface. [Invention 9] 9. The tire performance prediction device according to any one of claims 1 to 8, wherein the performance includes at least one of the rolling resistance coefficient, vertical spring constant, lateral spring constant, and air resistance value of the tire. [Invention 10] an optimal solution acquisition unit that obtains an optimal solution for the feature quantity so that the performance satisfies a predetermined standard; 10. The tire performance prediction device according to any one of claims 1 to 9, further comprising: a restoration unit that restores the optimal solution to the data. [Invention 11] 1. A method for predicting tire performance executed by a processor, comprising: extracting, from data including at least one of a skeleton image and a pattern image of a tread portion in a meridian cross section of the tire, a feature amount that can be restored to the data; and outputting the performance corresponding to the feature quantity using a trained model that has been machine-learned to output the tire performance from the feature quantity. A method for predicting tire performance. [Invention 12] A method for creating an input dataset necessary for outputting tire performance to a trained model trained by machine learning to output the performance, comprising: extracting, from data including at least one of a skeleton image and a pattern image of a tread portion at a meridian cross section of the tire, a feature amount that can be restored to the data and can be arbitrarily edited; and editing the feature values ​​to create a data set having a plurality of feature values ​​that are different from each other. How to create the dataset. [Explanation of symbols]

[0188] 21 Data 21A Skeleton Image 22 Performance 24B trained model 27 Features

Claims

1. A tire performance prediction device, and a trained model storage unit that extracts feature amounts that can be restored from data including at least one of a skeleton image of a tire at a meridian cross section and a pattern image of a tread portion, and stores a trained model that has been machine-learned to output tire performance from the feature amounts. Tire performance prediction device.

2. The tire performance prediction device according to claim 1 , wherein the extracted feature amount is editable as desired.

3. The tire performance prediction device according to claim 1 , wherein the trained model includes a first model that extracts the feature amount by dimensionally compressing the data.

4. The tire performance prediction device according to claim 3 , wherein the first model extracts the feature quantities using at least one of an autoencoder, a principal component analysis, an independent component analysis, a singular value decomposition, and a Fourier transform.

5. The tire performance prediction device according to claim 1 , wherein the trained model includes a second model that outputs the performance from the feature amount.

6. The tire performance prediction device according to claim 5 , wherein the second model outputs the performance based on at least one of a neural network, a multiple regression, a random forest regression, and a deep learning.

7. The tire performance prediction device according to claim 5 , wherein the second model outputs the performance from the feature amount and design factors of the tire.

8. The tire performance prediction device according to claim 1 , wherein the performance includes distribution data of physical quantities acting on the tread portion when the tread portion comes into contact with the road surface.

9. The tire performance prediction device according to claim 1 , wherein the performance includes at least one of a rolling resistance coefficient, a longitudinal spring constant, a lateral spring constant, and an air resistance value of the tire.

10. an optimal solution acquisition unit that obtains an optimal solution for the feature quantity so that the performance satisfies a predetermined standard; The tire performance prediction device according to claim 1 , further comprising: a restoration unit that restores the optimal solution to the data.

11. 1. A method for predicting tire performance executed by a processor, comprising: extracting, from data including at least one of a skeleton image and a pattern image of a tread portion in a meridian cross section of the tire, a feature amount that can be restored to the data; and outputting the performance corresponding to the feature quantity using a trained model that has been machine-learned to output the tire performance from the feature quantity. A method for predicting tire performance.

12. A method for creating an input dataset necessary for a trained model that has been machine-learned to output tire performance, the method comprising: extracting, from data including at least one of a skeleton image and a pattern image of a tread portion at a meridian cross section of the tire, a feature amount that can be restored to the data and can be arbitrarily edited; and editing the feature values ​​to create a data set having a plurality of feature values ​​that are different from each other. How to create the dataset.

Images