JP7670959B2 - Target model generation method using a pre-trained model, evaluation data prediction method, computer-readable recording medium recording a target model generated using a pre-trained model, and a program for the target model generation method using a pre-trained model and the evaluation data prediction method - Google Patents

Description

The present invention relates to a target model generation method using a pre-trained model, an evaluation data prediction method using a target model generated by the model generation method, a computer-readable recording medium on which a target model generated using a pre-trained model is recorded, and a program for a target model generation method using a pre-trained model and a program for an evaluation data prediction method.

To improve the efficiency of tire design and development, tire characteristics are predicted using machine learning.

JP 2021-22276 A

In tire characteristic prediction using machine learning, multiple learning input data sets are prepared, including tire specification data and evaluation data related to tire performance, etc. Then, a target model is generated by performing machine learning on the target model using the multiple learning data sets. Then, the target model is used to predict prediction target evaluation data corresponding to the prediction target specification data related to the tire to be predicted.

In particular, in order to use a target model to accurately predict the prediction target evaluation data corresponding to the prediction target specification data, the target model needs to be trained using a learning input dataset that includes the prediction target specification data.

If the real specification data of an existing tire is used as the real specification data included in the real input data set for learning, and the actual measured real evaluation data of an existing tire is used as the real evaluation data included in the real input data, the range of prediction target specification data that satisfies the above conditions and can predict the prediction target evaluation data will be limited.

Furthermore, in order to expand the range, it would normally be necessary to actually create tires based on new real-world specification data that corresponds to the expanded area, and then actually measure the real-world evaluation data for those tires. In this case, in order to actually create tires based on the new real-world specification data, it would be necessary to create a new mold, which would be costly in terms of both time and money.

Patent Document 1 discloses a data processing method for efficiently creating a prediction module that accurately predicts values related to features by inputting values of multiple explanatory variables from an original data set that includes experimental data and simulation data. However, the data processing method of Patent Document 1 cannot solve the above-mentioned problems.

The present invention aims to provide a target model generation method using a pre-learning model that enables accurate prediction of prediction target evaluation data corresponding to prediction target specification data in an area that cannot be covered by real specification data included in a real data set for learning, an evaluation data prediction method, a computer-readable recording medium that records a target model generated using the pre-learning model, and a program for the target model generation method and evaluation data prediction method using the pre-learning model.

According to the present invention,
preparing a plurality of real data sets (X, Y) including real specification data (X) and real evaluation data (Y) corresponding to the real specification data (X);
a virtual data set generating step of applying a specification change by applying computer processing to real specification data (X) in any of the real data sets included in the plurality of real data sets (X, Y) to generate virtual specification data, and applying computer processing to the virtual specification data to generate virtual evaluation data corresponding to the virtual specification data, thereby repeatedly generating a plurality of virtual data sets each including the virtual specification data and the virtual evaluation data corresponding thereto;
a pre-training model generation step of generating a pre-training model (M1') by performing machine learning on an initial pre-training model (M1) using each of the multiple virtual data sets (X', Y') as a learning input data set;
a target model generation step of generating a target model (M2') by performing machine learning on an initial target model (M2) using the pre-learning model (M1') and at least each of the multiple real data sets (X, Y) as a learning input data set;
A method for generating a target model using a pre-trained model having the following structure is provided.

The present invention also provides a computer-readable recording medium that records the target model generated by the target model generation method using the above pre-trained model.

Furthermore, according to the present invention, there is provided an evaluation data prediction method for predicting evaluation data corresponding to the specification data to be predicted using the target model generated by the target model generation method using the above-mentioned pre-learning model.

Furthermore, according to the present invention, a program is provided for causing a computer to execute a target model generation method using the above-mentioned pre-trained model.

Furthermore, according to the present invention, a program for causing a computer to execute the above-mentioned evaluation data prediction method is provided.

According to the present invention, it is possible to generate a target model that enables accurate prediction of target evaluation data corresponding to target specification data that is in an area that cannot be covered by the real specification data contained in the real data set for learning.

1 is a flowchart illustrating a method including a target model generating method and a prediction target evaluation data predicting method according to an embodiment of the present invention. 2 is a flow chart for explaining details of a first aspect of the step of generating a plurality of virtual data sets shown in FIG. 1 . 2 is a flow chart for explaining details of a second aspect of the step of generating a plurality of virtual data sets shown in FIG. 1 . FIG. 2 is a conceptual diagram for explaining details of a first aspect of the step of generating a target model shown in FIG. 1 . FIG. 2 is a conceptual diagram (1/2) for explaining details of a second aspect of the step of generating a target model shown in FIG. 1 . FIG. 2 is a conceptual diagram (2/2) for explaining details of a second aspect of the step of generating a target model shown in FIG. 1 . 1 is a block diagram showing a target model generating device for executing a target model generating method and an evaluation data predicting device for executing a prediction target evaluation data predicting method according to an embodiment of the invention, as well as input and output data thereof; FIG. 1 is a block diagram showing a configuration of a computer for executing a target model generating method and a prediction target evaluation data predicting method according to an embodiment of the present invention. 9 is a diagram showing data stored in each storage unit provided in the recording medium shown in FIG. 8 . FIG. 2 is a diagram for explaining tire specifications changed in the specification change step shown in FIG. 1 . 4 is a diagram for explaining the steps of correcting the virtual specification data shown in Fig. 3. (a) shows the shape of a base model of a tire on which the virtual specification data is based, (b) shows the shape of a tire corresponding to the virtual specification data, and (c) shows the shape of a tire corresponding to the corrected virtual specification data. 1A is a graph for explaining the accuracy of the predicted values of evaluation target evaluation data predicted by the target model generation method and the predicted target evaluation data prediction method according to an embodiment of the present invention, and FIG. 1B is a comparison table for explaining the accuracy of the predicted values of evaluation target evaluation data predicted by the target model generation method and the predicted target evaluation data prediction method according to an embodiment of the present invention.

Below, the embodiment of the present invention will be described in detail with reference to the drawings.

[First embodiment]
FIG. 1 is a flowchart illustrating a target model generating method and a prediction target evaluation data predicting method according to a first embodiment of the present invention.

Referring to FIG. 1, first, multiple real-world data sets (X, Y) are prepared (step S201). Here, X is real-world specification data, Y is real-world evaluation data, and W is real-world evaluation condition data.

Next, multiple virtual data sets (X', Y') are generated (step S203), where X' is virtual specification data and Y' is virtual evaluation data.

Next, a target model M2' is generated based on multiple real data sets (X, Y) and multiple virtual data sets (X', Y') (step S205).

Next, the target model M2' is used to predict the prediction target evaluation data (YY) for the prediction target specification data (XX) of the prediction target (step S207).

Figure 2 is a flowchart showing details of one aspect of the step (step S203) of generating multiple virtual data sets (X', Y') shown in Figure 1.

Referring to FIG. 2, first, from the multiple reality data sets (X, Y) prepared in step S201, real data (X) contained in one reality data set (X, Y) is selected (step S301).

Next, the content of the specification change is determined (step S303).

Next, the specification changes determined in step S303 are applied to the real data (X) selected in step S301 to generate virtual specification data (X') (step S305).

Next, virtual evaluation data (Y') corresponding to the virtual specification data (X') generated in step S305 is generated (step S307).

Next, a virtual data set (X', Y') including the virtual specification data (X') generated in step S305 and the virtual evaluation data (Y') generated in step S307 is stored on a recording medium (step S309).

Next, if a further virtual data set is to be added (YES in step S311), the process from step S301 to step S309 is repeated.

This process continues, and steps S301 to S309 are repeated a predetermined number of times to generate a predetermined number of virtual data sets (X', Y').

The actual specification data (Xi) to which specification changes are applied in each iteration is different for each iteration, or is common between the iterations for at least some of the iterations.

In addition, the specification changes applied to the real data in each iteration may be different for each iteration, or may be common across at least some of the iterations.

Figure 3 is a flowchart showing details of another aspect of the step (step S203) of generating multiple virtual data sets (X', Y') shown in Figure 1.

Compared to the embodiment shown in FIG. 2, the other embodiment shown in FIG. 3 differs in that a step (step S313) of modifying the virtual specification data (X') is added after the step (S305) of generating the virtual specification data (X').

Referring to FIG. 3, first, from the multiple reality data sets (X, Y) prepared in step S201, real data (X) contained in one reality data set (X, Y) is selected (step S301).

Next, the content of the specification change is determined (step S303).

Next, the specification changes determined in step S303 are applied to the real data (X) selected in step S301 to generate virtual specification data (X') (step S305).

Next, the virtual specification data (X') generated in step S305 is modified (step S313).

Next, virtual evaluation data (Y') corresponding to the virtual specification data (X') modified in step S313 is generated (step S307).

Next, a virtual data set (X', Y') including the virtual specification data (X') modified in step S313 and the virtual evaluation data (Y') generated in step S307 is stored on a recording medium (step S309).

Next, if a further virtual data set is to be added (YES in step S311), the process from step S301 to step S309 is repeated.

This process continues, and steps S301 to S309 are repeated a predetermined number of times to generate a predetermined number of virtual data sets (X', Y').

Figure 4 is a conceptual diagram for explaining the contents of the first aspect of the step (step S205) of generating the target model M2' shown in Figure 1.

To generate the target model M2', a pre-trained model M1' is created in advance using the virtual data set (X', Y').

Then, based on this, a target model M2' is generated. Specifically, at least a part of the pre-training model M1' is copied to the target model M2, and then the target model M2 is subjected to machine learning to generate the target model M2'.

Referring to FIG. 4, a pre-training model M1' is obtained by performing machine learning on an initial pre-training model M1 using each of a plurality of virtual data sets (X', Y') as a training input data set (step S231).

Next, at least a portion of the pre-trained model M1' is copied to the initial target model M2 (step S233).

Here, the pre-training model and the target model are, for example, models that utilize neural networks. Also, in step S233, the intermediate layer of the pre-training model M1' is copied to the initial target model M2. In particular, layers other than the discrimination layer, which is the final layer of the pre-training model M1', may be copied to the initial target model M2.

Next, a target model M2' is obtained by performing machine learning on the initial target model M2 using each of the multiple real-world data sets (X, Y) as a learning input data set (step S235).

Figures 5 and 6 are conceptual diagrams for explaining the second aspect of the step (step S205) of generating the target model M2' shown in Figure 1.

To generate the target model M2', a pre-trained model M1' is created in advance using the virtual data set (X', Y').

Then, using the pre-learning model M1', additional reality evaluation data (Y2) corresponding to the real specification data (X) is generated, and these are combined to create an additional reality data set (X, Y2).

Furthermore, a target model M2' is generated by performing machine learning on the target model M2 using the real data set (X, Y) and the additional real data set (X, Y2).

First, referring to FIG. 5, a pre-training model M1' is obtained by performing machine learning on an initial pre-training model M1 using each of a plurality of virtual data sets (X', Y') as a learning input data set (step S241).

Next, using the pre-learning model M1', evaluation data corresponding to the real specification data (X) included in each data set (X, Y) is obtained as additional real evaluation data (Y2), and this is repeated for multiple data sets (X, Y), thereby obtaining multiple additional real evaluation data (Y2) (step S243).

Then, an additional reality data set (X, Y2) containing each reality specification data (X) and the corresponding additional reality evaluation data (Y2) is created and saved. This results in multiple additional reality data sets (X, Y2) being saved.

6, a plurality of input data sets (X, Y3) for training a target model are generated based on a plurality of real data sets (X, Y) and a plurality of additional real data sets (X, Y2) (step S251).
Y3=a1×Y+(1-a1)×Y2
Y3 is calculated by the following formula.

Next, a target model M2' is obtained by performing machine learning on the initial target model M2 using each of the multiple target model learning input data sets (X, Y3) as a learning input data set (step S253).

In addition, multiple real-world data (X, Y) and multiple additional real-world data sets (X, Y2) may be used directly as training data sets.

Machine learning techniques include neural networks and random forests, and are not particularly limited. Different techniques may be used for the pre-training model and the target model.

Figure 7 is a block diagram showing a target model generation device for executing a target model generation method according to an embodiment of the present invention, an evaluation data prediction device for executing a prediction target evaluation data prediction method, and their input and output data.

Referring to FIG. 7, the target model generation device 501 inputs multiple real data sets (X, Y) and generates multiple virtual data sets (X', Y'). Then, using the multiple real data sets (X, Y) and the multiple virtual data sets (X', Y'), a pre-training model M1 (M1') is used to generate a target model M2 (M2').

The evaluation data prediction device 503 uses the target model M2' to generate prediction target evaluation data YY corresponding to the prediction target specification data XX.

Figure 8 is a block diagram showing the configuration of a computer for executing the target model generation method and the prediction target evaluation data prediction method according to an embodiment of the present invention.

The computer includes a CPU 701, a main memory 703, a user interface 705, an external interface 707, and a recording medium 709, all of which are connected to each other by a bus.

The recording medium 709 includes a real data set (X, Y) storage unit 709a, a virtual data set (X', Y') storage unit 709b, a model storage unit 709c, a prediction target specification data (XX) storage unit 709d, a prediction target evaluation data (YY) storage unit 709e, a program storage unit 709f for the target model generation device, and a program storage unit 709g for the evaluation data prediction device.

The CPU 701 executes the program stored in the program storage unit 709f for the target model generating device, and the target model generating device 501B is constructed in the CPU 701.

When the CPU 701 executes the program stored in the program storage unit 709g for the evaluation program prediction device, an evaluation data prediction device 503B is constructed in the CPU 701.

FIG. 9 is a diagram showing data stored in each storage unit provided in the recording medium 709 shown in FIG.
9, the real data set (X, Y) storage unit 709a stores a plurality of real data sets (X, Y).

The virtual data set (X', Y') storage unit 709b stores multiple virtual data sets (X', Y').

The model memory unit 709c stores the pre-training model M1 (M1') and the target model M2 (M2').

The predicted target specification data (XX) storage unit 709d stores the predicted target specification data XX.

The prediction target evaluation data (YY) storage unit 709e stores the prediction target evaluation data (YY).

The program storage unit 709f for the target model generating device stores the program 709fp for the target model generating device.

The evaluation data prediction device program storage unit 709g stores the evaluation data prediction device program 709gp.

Next, we will explain an example of how the above method is applied to vehicle tires.

The tire specifications are represented by specification data including a set of numerical values for the dimensions of each part DA, DB, DC, RA, RB, RC, RD, WA, WB, and LA in the cross-sectional shape of the tire shown in FIG. 10.

For example, if the specification change is to increase the width of the tire, then LA, RA, WA, and WB will be increased. Also, if the specification change is to increase the diameter of the tire, then DA, DB, and DC will be increased.

The specification change may also be a change to the tire structure or the tire material.

Evaluation data includes inflated dimensions, ground contact characteristics, spring characteristics, rolling resistance, wear characteristics, durability, etc.

Evaluation condition data includes air pressure, load, speed, rim size, etc.

To generate the virtual evaluation data (Y'), a tire shape in a characteristic calculation model (e.g., a model used in the finite element method, a surrogate model) is generated by computer based on the specification data including a set of numerical values for the dimensions of each part of the tire cross-sectional shape as described above. Then, the virtual evaluation data (Y') is generated by the finite element method or surrogate modeling based on the characteristic calculation model.

Referring to Figure 11, we will explain an example of generating virtual specification data for a tire from real specification data and then modifying the virtual specification data.

Figure 11(a) shows the tire cross-sectional shape 261 corresponding to the actual specification data.

Figure 11 (b) shows a tire cross-sectional shape 263 corresponding to the virtual specification data generated by applying specification changes to the cross-sectional shape 261. In the cross-sectional shape 263, a reverse R portion exists near the boundary between the tread portion and the sidewall portion.

Figure 11(c) shows a tire cross-sectional shape 265 generated by applying shape modification to the cross-sectional shape 263 to remove the inverse R portion. The virtual specification data corresponding to the cross-sectional shape 265 replaces the virtual specification data corresponding to the cross-sectional shape 263.

This is not the only example, but for example, when the mold profile is changed due to a change in the specification data, the internal structure is also changed accordingly. In this case, the shape of the carcass or belt may be unintentionally curved or may have dimensions that are outside the design standard, which is undesirable from a design perspective. Therefore, by modifying the shape of the carcass or belt after the above changes, tire specification data that is in line with the design guidelines can be obtained.

The specification data is not limited to data relating to the cross-sectional shape of the tire. For example, it may be data relating to the shape of the mold used to manufacture the tire. The shape of the mold corresponds to the shape of the outer surface of the tire. If the shape of the mold is changed, the shapes of parts that depend on or are associated with the shape of the mold may also be changed.

When obtaining tire specification data that corresponds to the changed tire cross-sectional shape, make sure it matches the tire model.

When modifying the mold shape, the lines on the mold profile are divided into multiple regions, and a tire model is created by enlarging or reducing each of the divided lines relative to the modified shape.

In other words, this solves the problem that when a tire model constructed using the finite element method is used and elements and nodes on the base model profile are evenly projected onto the modified profile to create a new tire model, the internal structure may end up with an unnatural shape. In other words, the above-mentioned inconvenience can be solved by dividing the tire into multiple regions and projecting elements and nodes for each region to create a tire model.

After the shape change is performed, the shape or dimensions of the reinforcement may be modified to conform to design standards or guidelines.

This allows us to obtain virtual data that is in line with the design guidelines.

Fig. 12(a) is a graph for comparing the predicted values when predicting prediction target evaluation data (predicted values) corresponding to prediction target specification data for a tire between a conventional method and the method according to this embodiment. Fig. 12(b) is a table for comparing the predicted values when predicting prediction target evaluation data (predicted values) corresponding to prediction target specification data for a tire between a conventional method and the method according to this embodiment.

12A, the horizontal axis corresponds to the actual measured value, and the vertical axis corresponds to the predicted value. Points indicated by circles are points obtained by the conventional method, and points indicated by triangles are points obtained by the method according to the present embodiment.
12B is a table for comparing R2, MAE, and MSE between the conventional method and the method according to the present embodiment. Here, R2 is the coefficient of determination, MAE is the mean absolute error, and MSE is the mean square error. A larger R2 indicates superiority, and a smaller MAE and MSE indicate superiority. By comparing the numerical values, it can be seen that the method according to the present embodiment is superior to the conventional method in all of R2, MAE, and MSE.

In this embodiment, a virtual data set (X', Y') including virtual specification data (X') and virtual evaluation data (Y') can be generated by using a computer.

By applying the target model generation method and evaluation data prediction method according to this embodiment to tires, it becomes unnecessary to create a new mold corresponding to the additional specification data, create a new tire using that mold, and measure the evaluation data of that tire.

In particular, creating a new mold (or changing the mold shape) involves significant time and financial costs, and it is difficult to add a large number of tires that correspond to new real-world specification data. Therefore, the effect of generating a virtual data set that includes virtual specification data that replaces the new real-world specification data and the corresponding virtual evaluation data is significant.

The present invention can be implemented in various other forms without departing from its spirit or main features. Therefore, the above-described embodiments are merely illustrative and should not be interpreted as being limiting. The scope of the present invention is defined by the claims and is not limited to the text of the specification. Furthermore, all modifications and changes within the equivalent scope of the claims are within the scope of the present invention.

M1 Initial pre-training model M1' Pre-training model M2 Initial target model M2' Target model X Actual specification data Y Actual evaluation data (X, Y) Actual data set X' Virtual specification data Y' Virtual evaluation data (X', Y') Virtual data set XX Prediction target specification data YY Prediction target evaluation data (XX, YY) Prediction target data set

Claims (12)

preparing a plurality of real data sets (X, Y) including real specification data (X) and real evaluation data (Y) corresponding to the real specification data (X);
a virtual data set generating step of applying a specification change by applying computer processing to real specification data (X) in any of the real data sets included in the plurality of real data sets (X, Y) to generate virtual specification data, and applying computer processing to the virtual specification data to generate virtual evaluation data corresponding to the virtual specification data, thereby repeatedly generating a plurality of virtual data sets each including the virtual specification data and the virtual evaluation data corresponding thereto;
a pre-training model generation step of generating a pre-training model (M1') by performing machine learning on an initial pre-training model (M1) using each of the multiple virtual data sets (X', Y') as a learning input data set ;
a target model generation step of generating a target model (M2') by performing machine learning on an initial target model (M2) using the pre-learning model (M1') and at least each of the multiple real data sets (X, Y) as a learning input data set;
A method for generating a target model using a pre-trained model having the following structure: In the virtual specification data set generation step, the virtual specification data (X') is generated by changing at least a part of a plurality of numerical values included in the real specification data (X) in one real data set (X, Y) selected from the plurality of real data sets (X, Y).
A method for generating a target model using the pre-trained model according to claim 1. The actual specification data (Xi) to which the specification change is applied in each repetition in the virtual data set generation step is different for each repetition, or is common between the repetitions for at least some of the repetitions.
A method for generating a target model using the pre-trained model according to claim 2. The specification changes applied to the real data in each repetition in the virtual data set generation step are different for each repetition, or are common between at least some of the repetitions.
A method for generating a target model using the pre-trained model according to claim 2 or 3. copying at least a portion of said pre-trained model (M1') into an initial target model (M2);
Further comprising
A method for generating a target model using the pre-trained model according to any one of claims 1 to 4. A step of acquiring a plurality of additional real evaluation data (Y2) corresponding to a plurality of real specification data (X) using the pre-learning model (M1');
constructing a plurality of additional reality data sets (X, Y2) by combining corresponding additional reality evaluation data (Y2) for each of a plurality of reality specification data (X);
Further comprising:
In the target model generation step, machine learning is performed on an initial target model (M2) using each of a plurality of target model learning input data sets obtained by adding the plurality of additional reality data sets (X, Y2) to the plurality of reality data sets (X, Y);
A method for generating a target model using the pre-trained model according to any one of claims 1 to 4. In the virtual data set generation step,
The virtual evaluation data (Y') corresponding to each of the virtual specification data (X') is calculated using a finite element method.
A method for generating a target model using the pre-trained model according to any one of claims 1 to 6. the real data set and the virtual data set relate to a tire;
A method for generating a target model using the pre-trained model according to any one of claims 1 to 7. The target model generating method using a pre-learning model according to claim 8 , wherein the real specification data (X) and the virtual specification data include a plurality of numerical values related to a cross-sectional shape of the tire. The specification change is a change in the cross-sectional shape of the tire.
A method for generating a target model using the pre-trained model according to claim 9. When the virtual specification data is generated by the specification change, a cross-sectional shape of the tire corresponding to the virtual specification data is modified, the virtual specification data is modified so that the cross-sectional shape of the tire after the modification corresponds to the virtual specification data, and then a computer process is applied to the virtual specification data to generate the virtual evaluation data corresponding to the virtual specification data.
A method for generating a target model using the pre-trained model according to claim 10. A program for causing a computer to execute a target model generation method using the pre-training model according to any one of claims 1 to 11.

Images