JP7555263B2 - Information processing method, system and program - Google Patents

Description

This disclosure relates to an information processing method, system, and program that belong to technology related to machine learning of tire images.

The use of machine learning is spreading to various industrial fields, and there are also an increasing number of cases where other companies in the same industry are applying it to tires. Attempts are being made to obtain predicted values from tire images using machine learning. For example, Patent Document 1 describes a method in which tire images showing the tire tread are input into a neural network to estimate the state or amount of wear.

If it were possible to know what changes in the tire image would cause the predicted value to change, it would be possible to consider ways to improve the predicted value. Examples of predicted values include tire performance values and wear assessment values. However, the features of tire images extracted by machine learning are numerical values whose meanings are unclear, and the features extracted from tire images cannot be used as the basis for human judgment.

JP 2019-35626 A

This disclosure provides technology that provides new data that can be used to elucidate the relationship between features and predicted values in machine learning.

The information processing method disclosed herein is a method executed by one or more processors, and includes the steps of: acquiring teacher data in which tire images and predicted values related to tires are associated with each other; a feature extraction step of inputting an input image based on the tire image to a feature extraction unit and extracting multiple feature values for one of the tire images; a learning step of machine learning a decision tree prediction model to input the extracted multiple feature values and output corresponding predicted values in the teacher data; an importance calculation step of calculating the importance of each of the multiple feature values using the trained decision tree prediction model; a step of accepting a selection of a feature value to be searched from among the multiple feature values; a step of generating multiple input patterns in which the selected feature values of the search target are different from each other; a step of inputting the generated multiple input patterns into the decision tree prediction model and predicting the corresponding predicted values; and a step of calculating a threshold value of the feature value of the search target that represents a boundary within which the predicted value fluctuates, based on the predicted multiple predicted values.

FIG. 1 is a block diagram showing a system according to a first embodiment. 1 is a diagram showing an example of a ground plane image. 3 is a diagram showing an image obtained after the ground plane image shown in FIG. 2 has been subjected to a predetermined image processing. 3 is a diagram showing an image obtained after the ground plane image shown in FIG. 2 has been subjected to a predetermined image processing. 3 is a diagram showing an image obtained after the ground plane image shown in FIG. 2 has been subjected to a predetermined image processing. FIG. 13 is a plot of feature amounts of generated input patterns and predicted values based on the input patterns. FIG. 11 is a diagram showing box plots of first and second image parameter groups for 20 types of image parameters in a comparative manner; FIG. 13 is a graph showing ranking results of 20 types of image parameters by Wilcoxon test. 4 is a flowchart executed by the system of the first embodiment. 10 is a flowchart showing a modification of the example shown in FIG. 9 .

First Embodiment
Hereinafter, a first embodiment of the present disclosure will be described with reference to the drawings.

FIG. 1 is a block diagram showing a system of a first embodiment. The system is configured to machine-learn a prediction model to output a predicted value from an image related to a tire (tire image), and to identify and output information related to the machine learning or information using the machine-learned prediction model. The system may calculate a threshold related to the machine learning of the tire image. The system may identify information related to image parameters by machine learning. In the first embodiment, the tire image is a contact patch image, and the predicted value is a tire performance value, but this is an example and is not limited to the present embodiment.

Figure 2 is a schematic diagram of a ground contact patch image 20. For ease of explanation, Figure 2 is neither in color nor in grayscale, but is a line diagram in which the ground contact patch 42 is shown hatched. In the example of Figure 2, the circumferential groove 40 extending in the tire circumferential direction is shown extending in a first direction Ax1 (vertical direction) of the image, and the lateral groove 41 extending in the tire axial direction is shown extending in a second direction Ax2 (horizontal direction) of the image, but is not limited to this. In the example of Figure 2, the ground contact patch 42 is divided into multiple blocks 43 by grooves (at least one of the circumferential groove 40 and the lateral groove 41).

As shown in FIG. 2, the ground contact surface image 20 represents the tire ground contact surface shape. The ground contact surface image 20 of the first embodiment represents the tire ground contact surface shape, and also represents the ground contact pressure Pz in a direction perpendicular to the road surface by color in the case of a color image, or by brightness in the case of a grayscale image. In order to display the ground contact pressure together with the ground contact surface shape, the ground contact surface image 20 represents the ground contact pressure distribution. The ground contact pressure distribution is associated with a pressure value for each area of the ground contact surface. The ground contact surface image 20 of the first embodiment is an image of a stationary state in which the tire is not rolling, but is not limited to this, and may be an image of the tire rolling. Regardless of whether the tire is rolling or stationary, the camber angle of the tire can be set to any angle. The ground contact surface image 20 of the first embodiment is an image of a stationary state of the tire, but when the tire is rolling, the slip angle with respect to the traveling direction may be 0 degrees or may be an angle other than 0 degrees. The ground contact surface image 20 of the first embodiment represents only the ground contact pressure Pz in a direction perpendicular to the road surface in a stationary state, but is not limited to this. For example, in a stationary state, the pressure Px along the front-rear direction perpendicular to both the tire axial direction and the direction perpendicular to the road surface, or the pressure Py along the tire axial direction may be expressed. In addition, when the tire is in a rolling state, the pressure Px along the tire traveling direction or the pressure Py along the direction perpendicular to the tire traveling direction may be expressed. Note that these coordinate systems are only examples and can be changed as appropriate. Note that, as long as accuracy can be ensured, the contact patch image 20 may represent only the tire contact patch shape. The contact patch image may be acquired by a test device (not shown) or may be acquired from the simulation results of a tire contact simulation system. The test device places the test target tire on the test road surface under a predetermined load and captures the contact patch shape with a camera through a transparent road surface on the test road surface. The contact pressure is measured using an optical method using a transparent plate or the like, or a pressure sensor. The test device generates a contact patch image by the above test.

The tire performance values are index values (actual measured values) that represent the tire performance obtained by testing the tire. The tire performance values used in the first embodiment are CP (cornering power), SAP (self-aligning power), CFmax (maximum cornering force), and SAT (self-aligning torque). Of course, any other tire performance can be used as the tire performance value. For example, rolling resistance or rolling resistance coefficient, friction coefficient, interior and exterior noise performance, and radiated sound performance of the tire alone can be mentioned.

As shown in FIG. 1, the system has a teacher data acquisition unit 10, a feature extraction unit 11, a learning unit 12, an importance calculation unit 14, a reception unit 15, an input pattern generation unit 16, a prediction unit 17, and a threshold calculation unit 18. The system may further have an image parameter calculation unit 30, a classification unit 31, a characteristic information output unit 32, a judgment unit 33, and an evaluation information output unit 34. The image parameter calculation unit 30, the classification unit 31, the characteristic information output unit 32, the judgment unit 33, and the evaluation information output unit 34 are optional and can be added as desired.

These units 10-18, 30-34 are realized by software and hardware working together when processor 1a executes the processing routines shown in FIG. 9 or 10 that are pre-stored in a computer equipped with processor 1a, memory 1b, various interfaces, etc. In this embodiment, the processor 1a in one device realizes each unit, but this is not limited to this. For example, the units may be distributed using a network, with multiple processors executing the processing of each unit. In other words, one or multiple processors execute the processing.

The teacher data acquisition unit 10 shown in FIG. 1 acquires teacher data D1 from an external source. The acquired teacher data D1 is stored in the memory 1b. The teacher data D1 is data in which a contact patch image 20 (tire image; images 1 to N) to be input to a machine learning decision tree prediction model 13 is associated with a tire performance value 21 (predicted value; X ₁ , X ₂ , ..., X _N ) to be output from the decision tree prediction model 13. The teacher data D1 is used for machine learning of the decision tree prediction model 13. For convenience of explanation, in this specification and FIG. 1, it is assumed that there are N contact patch images. N indicates the number of teacher data.

When an input image based on the ground plane image 20 is input, the feature extraction unit 11 extracts a plurality of feature amounts for one ground plane image 20. The extracted plurality of feature amounts are stored in memory in association with the ground plane image (see FIG. 1). Here, there are M feature amounts, and each feature amount is represented by (FV ₁ , FV ₂ , ..., FV _M ). M is a number that is predetermined by an algorithm for calculating the feature amounts. By sequentially inputting N pieces of teacher data D1, that is, N ground plane images 20, to the feature extraction unit 11, N x M feature amounts are output as a result. The input image means an image input to the feature extraction unit 11. The input image means the ground plane image 20 itself, or an image that is processed and input to the feature extraction unit 11 from the ground plane image 20. The feature extraction unit 11 may use any algorithm as long as it can extract feature amounts from an image. For example, examples include neural networks (e.g., SqueezeNet, Alexnet, GoogleNet, ResNet101), discrete cosine transform processing, AutoEncoder, and a configuration that combines discrete cosine transform processing and AutoEncoder.

The learning unit 12 constructs a decision tree prediction model 13 by machine learning so as to use a plurality of feature amounts ( _FV1 , _FV2 , ..., _FVM ) of the image extracted by the feature extraction unit 11 as a group of explanatory variables and output corresponding tire performance values (corresponding values among _X1 , _X2 , ..., _XN ) in the teacher data D1 as objective variables. In the first embodiment, the decision tree prediction model 13 uses a random forest, but is not limited to this. The decision tree prediction model 13 can be appropriately changed as long as it is an algorithm belonging to a decision tree.

The importance calculation unit 14 calculates the importance ( _I1 , _I2 , ..., _IM ) of each of a plurality of feature quantities ( _FV1 , _FV2 , ..., _FVM ) using the trained decision tree prediction model 13. M importance quantities are calculated from M feature quantities. The importance quantity is an index value that indicates the degree to which the feature quantity contributes to the predicted value. In other words, if a feature quantity with high importance is changed, it is highly likely that the predicted value obtained as the objective variable will change, and the change in the predicted value will be relatively large. The calculated importance quantities ( _I1 , _I2 , ..., _IM ) are stored in the memory 1b in association with the feature quantities ( _FV1 , _FV2 , ..., _FVM ).

An importance output unit (not shown) may be provided. The importance output unit may output (display) the feature with the highest importance to the outside, or may output (display) multiple feature amounts in order according to their importance to the outside. This makes it possible to notify the user of the feature with the highest importance or multiple feature amounts with relatively high importance.

The reception unit 15 receives a selection of a feature to be searched from among the multiple feature amounts. The importance calculation unit 14 may output one or multiple feature amounts to the outside according to the importance, the user may select which feature amount to search and input the selection to the reception unit 15, and the reception unit 15 may receive the selected feature amount. Alternatively, the processor 1a may select at least one feature amount from the multiple feature amounts according to the importance, the selected feature amount is input to the reception unit 15, and the reception unit 15 may receive the selected feature amount. When multiple feature amounts are selected, each of the subsequent processes may be performed for each selected feature amount.

The input pattern generation unit 16 generates a plurality of input patterns (patterns 1 to P) to be inputted as explanatory variable groups to the trained decision tree prediction model 13. The generated plurality of input patterns (patterns 1 to P) are stored in the memory 1b. P indicates the number of input patterns to be generated and can be set appropriately. The plurality of input patterns have different features of the selected search target, and the other features among the plurality of features are the same. In the example of FIG. 1, an example in which the first feature FV ₁ is selected will be described. As shown in FIG. 1, the second to Mth features (FV ₂ to FV _M ) of each of the P input patterns (patterns 1 to P) are all the same. On the other hand, the first features of each of the P input patterns (patterns 1 to P) are all different. For example, the first feature FV _{1_1} of pattern 1 is different from the first feature FV _{1_2} of pattern 2, and is also different from the first feature FV _{1_P} of pattern P. The method of generating different values of the feature may be, for example, to adopt multiple (P) random values within the range from the maximum value to the minimum value of the first feature values stored in the memory 1b and extracted by the feature extraction unit 11, or to adopt values at regular intervals within the range from the maximum value to the minimum value. If the value is between the maximum value and the minimum value, it will be a numerical value within the learning range, so that it is possible to obtain an appropriate predicted value. The multiple (M) feature values that are the basis for generating the multiple (P) input patterns may be obtained by selecting one of the N ground surface images that have already been recorded, and using a group of feature values (M) based on the selected one ground surface image.

The prediction unit 17 inputs each of the multiple input patterns (patterns 1 to P) generated by the input pattern generation unit 16 into the decision tree prediction model 13, and predicts the corresponding predicted values (tire performance values; _Y1 , _Y2 , ..., _YP ). The acquired predicted values are stored in the memory 1b. The predicted value _Y1 is a predicted value corresponding to pattern 1, the predicted value _Y2 is a predicted value corresponding to pattern 2, and the predicted value _YP is a predicted value corresponding to pattern P.

The threshold calculation unit 18 calculates a threshold FV _Th1 of the feature quantity to be searched. The threshold FV _Th1 represents a boundary where a plurality of predicted values (Y ₁ , Y ₂ , ..., Y _P ) fluctuate. Whether or not the predicted value is a boundary where the predicted value fluctuates may be determined based on whether or not a predetermined threshold condition is met. Various predetermined threshold conditions can be set. For example, a feature quantity may be determined to be a threshold when the change in the predicted value per unit change in the feature quantity exceeds a predetermined value.

A specific example of threshold determination will be described with reference to FIG. 6. In the example of FIG. 6, the 647th feature is the search target, and the ground surface images "Image 1", "Image 2", "Image 3" and "Image 4" are used as basic data, respectively, to generate P input patterns per image, and P predicted values are predicted from the generated P input patterns by the decision tree prediction model 13, and the vertical axis is the predicted value and the horizontal axis is the value of the 647th feature. Here, four images are used as basic data, but it is possible to make a judgment using at least one image. More specifically, of the M feature amounts extracted from image 1, the 1st to 646th and 648th to Mth feature amounts are left as they are, and the value of the 647th feature amount is changed in minute value increments as shown in FIG. 6 to generate P input patterns. As shown in FIG. 6, it can be seen that the predicted value of the 647th feature amount changes rapidly around "10.5". In this example, the threshold FV _Th1 of the 647th feature amount is calculated to be 10.5. As described below, the threshold value is used to classify the multiple contact patch images 20 in the training data D1 into either the first group G1 or the second group G2, so there is no need to determine a precise value.

The image parameter calculation unit 30 calculates at least one image parameter (PL ₁ , ..., PL _N ) from each of the multiple tire images (contact patch images 20; images 1 to N) in the teacher data D1. The calculated image parameters (PL ₁ , ..., PL _N ) are stored in the memory 1b. The image parameter is a value calculated from at least one of the contact patch shape and the contact pressure in the contact patch image 20 based on a predetermined calculation rule process.

Specific examples of the at least one image parameter include the following (1) to (31), and the image parameter may include at least one of these. Note that the following (1) to (31) are merely examples, and details can be changed as appropriate as long as the values are calculated by a predetermined calculation rule process.
(1) Area of the ground surface in the image:
This is the area of the hatched region in Figure 2. It can be calculated by adding up the areas of the pixels that make up the ground surface.
(2) The sum of the contact area and the groove area in the image:
The area is the area of the hatched region in Fig. 3. The hatched region in Fig. 3 includes the ground contact region and the groove region. The closed region of the lateral groove 41 and the closed region of the circumferential groove 40 shown in Fig. 2 are identified by a predetermined processing method and are defined as the groove region.
(3) Ratio of (1) to (2):
In the first embodiment, the relationship is (1)/(2), but is not limited to this. The relationship may be (2)/(1).
(4) The sum of the values obtained by multiplying the area per unit pixel by the ground contact pressure for all pixels that make up the ground contact surface:
If there are Q pixels that make up the ground surface, the ground pressure of the Qth pixel is PZ _Q , and the area of the unit pixel is A ₁ , then it can be calculated as A ₁ ×PZ ₁ +A ₁ ×PZ ₂ + ... +A ₁ ×PZ _Q-1 +A ₁ ×PZ _Q , where Q is a natural number.
(5) The sum of the values obtained by multiplying the distance from each of the positions of the pixels constituting the ground contact surface to the center of gravity of the ground contact surface by the ground contact pressure:
If there are Q pixels that make up the ground contact surface, the ground contact pressure of the Qth pixel is _PZQ , and the distance between the Qth pixel and the center of gravity is _DDQ , then it can be calculated as _DD1 × _PZ1 + _DD2 × _PZ2 +...+ _DDQ-1 × _PZQ-1 + _DDQ × _PZQ , where Q is a natural number.
The center of gravity of the contact area corresponds to the hatched area in Fig. 3. The center of gravity in the image is a point where the first moment of the surroundings (first moment of area of the contact area) is zero.
(6) The average value of the area of each of the multiple blocks 43 divided by the perimeter of the block:
This value indicates the roughness or fineness of the distribution of the contact area. The (block area/block perimeter) was calculated for each block, and the average value of these was used.
(7) Number of blocks:
The number of blocks 43 corresponds to the number of labels counted using a labeling algorithm that recognizes regions of the image. In the example of FIG. 2, 15 blocks 43 are detected.
(8) Four angles of the area including the contact surface and the groove area:
The index value is 0 and 1, and is 0 if it is an ellipse and 1 if it is a rectangle. In the first embodiment, the area shown by hatching in Fig. 3 is targeted. Specifically, a rectangle that encloses the contact surface area in Fig. 3 is calculated, and an ellipse that fits within the rectangle is calculated. If the area of the contact surface is A, the area of the rectangle is B, and the area of the ellipse is C, then
The four angles can be calculated as follows: (A-C)/(B-C).
(9) Ratio of circumferential dimension of contact patch to axial dimension of tire:
In the first embodiment, the area indicated by hatching in FIG. 3 is the subject of the analysis.
(10) Number of circumferential grooves extending in the tire circumferential direction:
As shown in Fig. 4, the area of the lateral grooves 41 is added to the ground contact surface, and image processing is performed so that only the circumferential grooves 40 that define the ground contact surface are present, and then the number of circumferential grooves 40 is counted. In the example of Fig. 4, there are four circumferential grooves 40.
(11) Width of circumferential groove extending in the tire circumferential direction:
In the first embodiment, the image after image processing shown in Fig. 4 was used as the basis for judgment. If it is a value related to the width of the circumferential groove, the maximum width or the minimum width of the circumferential groove may be adopted. Also, it may be an average value of the width in the second direction Ax2 for each pixel line.
(12) A value obtained by summing the distances from the positions of each of a plurality of pixels constituting a circumferential groove extending in the tire circumferential direction to the center of gravity of the circumferential groove:
In the first embodiment, the judgment is based on the image after image processing shown in Fig. 4. The center of gravity in the image is a point where the first moment around it is zero, as in the above.
(13) Spacing of axially extending lateral grooves:
As shown in Fig. 5, only the land lateral grooves 41 located on the outermost side in the tire axial direction were extracted. The intervals between the extracted lateral grooves 41 in the tire circumferential direction (first direction Ax1 in the image) were measured. The intervals were measured for each pixel line and averaged.
(14) Width of axially extending lateral groove:
The image shown in Fig. 5 was used as the basis for the judgment. The maximum and minimum values of the lateral grooves 41 may be adopted. Also, the average value of the width of the lateral grooves in the first direction Ax1 for each pixel line may be adopted.
(15) Area of axially extending lateral grooves:
The image shown in Figure 5 was used as the basis for the judgment. It can be calculated by adding up the areas of the pixels that make up the lateral groove 41.
(16) Angle of axially extending lateral grooves:
The image shown in Figure 5 was used as the basis for judgment. The average value of the center line angle of each lateral groove was used.
(17) Axial width of the contact patch:
In the first embodiment, the target area is the hatched area in Fig. 3. The maximum width of the contact patch in the tire axial direction, that is, in the second direction Ax2, is adopted.
(18) Maximum contact length of the center land closest to the equatorial plane of the contact surface:
The maximum contact length is the maximum dimension of the center land in the first direction Ax1. If there are multiple center lands closest to the equatorial plane, the average value of each may be obtained, or an arbitrary land may be selected as the center land from among the multiple lands according to a predetermined rule.
(19) Maximum contact length of shoulder land on the outermost side of the tire in the axial direction of the contact surface:
The maximum contact length is the maximum dimension of the shoulder land in the first direction Ax1. When there are multiple shoulder lands, the average value of each may be obtained, or an arbitrary land may be selected as the shoulder land from among the multiple lands according to a predetermined rule.
(20) Ratio of (18) to (19):
The ratio may be (18)/(19) or (19)/(18).
(21) The sum of the ground pressures of the center land closest to the equatorial plane of the contact surface divided by the area of that center land:
(22) The sum of the contact pressures of the shoulder land located at the outermost part of the tire in the axial direction of the contact surface divided by the area of the shoulder land:
(23) Ratio of (21) to (22):
The ratio may be (21)/(22) or (22)/(21).
(24) The value obtained by dividing (1) by the test load or (4) above:
The test load means the vertical load applied to the axis of the pneumatic tire when the test device generates the contact patch image 20, or the vertical load applied to the axis of the pneumatic tire model when generating the contact patch image 20 by simulation.
(25) The value obtained by dividing (2) by the test load or (4) above:
The definition of the test load is the same as above.
(26) The value obtained by dividing (3) by the test load or (4) above:
The definition of the test load is the same as above.
(27) The value obtained by dividing (4) above by the test load:
The definition of the test load is the same as above.
(28) The value obtained by dividing (5) above by the test load:
The definition of the test load is the same as above.
(29) The value obtained by dividing (21) by the test load or (4) above:
The definition of the test load is the same as above.
(30) The value obtained by dividing (22) by the test load or (4) above:
The definition of the test load is the same as above.
(31) The value obtained by dividing (23) by the test load or (4) above:
The definition of the test load is the same as above.

The classification unit 31 classifies the multiple tire images (images 1 to N) in the teacher data D1 into either a first group G1 or a second group G2 based on the threshold value FV _Th1 . The classification result is stored in the memory 1b. As an example, the classification unit 31 classifies tire images and image parameters whose first feature value FV ₁ , which is the search target, is greater than the threshold value FV _Th1 into the first group G1, and classifies tire images and image parameters whose first feature value FV ₁ is equal to or less than the threshold value FV _Th1 into the second group G2. Of course, this is an example, and is not limited to this. As shown in FIG. 1, here, the tire images (ground contact surface images 20) classified into the first group G1 are referred to as the first tire image group, and the image parameters calculated from the first tire image group are referred to as the first image parameter group. Similarly, the tire images (ground contact surface images 20) classified into the second group G2 are referred to as the second tire image group, and the image parameters calculated from the second tire image group are referred to as the second image parameter group.

1, images 1, 3, etc. are classified into a first group G1, and the corresponding first image parameter PL ₁ , third image parameter PL ₃ , etc. are classified into the second group G2, and the corresponding second image parameter PL ₂ , Nth image parameter PL _N , etc. are classified into the second group G2.

The characteristic information output unit 32 outputs characteristic information D2 relating to the characteristics of the first image parameter group and the second image parameter group. The characteristic information D2 can be changed in various ways as long as it represents the characteristics of the first image parameter group and the second image parameter group. The characteristic information D2 may include, for example, information relating to the distribution of the first image parameter group and the second image parameter group. The characteristic information D2 may include, for example, at least one of statistics of the first image parameter group, statistics of the second image parameter group, and information relating to a comparison result between the first image parameter group and the second image parameter group.
The statistics may include, for example, at least one of standard deviation, variance, minimum, first quartile, second quartile (median), third quartile, fourth quartile (maximum), average, sum, and sample size.
The information regarding the comparison results may include, for example, at least one of a comparative display of a box plot of the first group of parameters and a box plot of the second group of parameters, and a test result evaluating the overlap of the distributions of the first group of parameters and the second group of parameters.
The test result may be the result of at least one of the following test methods, for example: Wilcoxon rank sum test, one-dimensional KL divergence, one-dimensional Bhattacharyya distance, and ROC analysis using logistic regression.

The determination unit 33 determines whether the image parameter is correlated with the predicted value based on the characteristic information D2 or based on the first image parameter group and the second image parameter group. In the first embodiment, the first and third quartiles of the first image parameter group and the first and third quartiles of the second image parameter group are calculated as the characteristic information. If the quartile range of the first image parameter group and the quartile range of the second parameter group overlap, it is determined that the image parameter is not correlated with the predicted value. If the quartile range of the first image parameter group and the quartile range of the second parameter group do not overlap, it is determined that the image parameter is correlated with the predicted value.

As a specific example, as shown in FIG. 7, the calculation of image parameters is executed multiple times to calculate multiple types of image parameters. In FIG. 7, 20 types of image parameters (1 to 20) are calculated. Characteristic information D2 is displayed for each of the multiple types of image parameters. The characteristic information D2 in the example of FIG. 7 displays a box plot of the first parameter group and a box plot of the second parameter group in comparison. In FIG. 7, "true" is the first group, and "false" is the second group. When the user looks at FIG. 7, for example, it can be seen that the interquartile ranges (box shapes) of each group for image parameter 11 are separated, and it can be seen that image parameter 11 is highly likely to be correlated with the predicted value.

As shown in FIG. 7, the determination unit 33 may be configured to compare the interquartile range of the first image parameter group with the interquartile range of the second image parameter group for each of the multiple types of image parameters, and determine that the image parameter is correlated with the predicted value if there is no overlap between the interquartile ranges, and determine that the image parameter is not correlated with the predicted value if there is overlap between the interquartile ranges.

The evaluation information output unit 34 may calculate an evaluation value relating to the distribution of the first image parameter group and the second image parameter group for each of the multiple image parameters, and output at least a part of the multiple image parameters in an order according to the evaluation value. The evaluation value relating to the distribution may be the degree of separation between the interquartile range of the first image parameter group and the interquartile range of the second image parameter group. In addition, the evaluation value relating to the distribution may be a value calculated using a testing method including at least one of the Wilcoxon rank sum test, one-dimensional KL divergence, one-dimensional Bhattacharyya distance, and ROC analysis using logistic regression.

As a specific example, FIG. 8 shows the evaluation values of each image parameter using the Wilcoxon rank sum test, arranged in descending order of evaluation value. The Wilcoxon rank sum test uses a hypothesis testing method with the evaluation criterion being distribution overlap, the assumed distribution being non-Gaussian, and the function is ranksum in MATLAB (registered trademark). That is, in FIG. 8, the overlap of the distributions of the first image parameter group and the second image parameter group is evaluated, and the group with the least overlap is considered to have a higher evaluation value. That is, it can be seen that the distributions of both groups are most separated for image parameter 11, followed by image parameter 5.

In the one-dimensional KL divergence, the evaluation criterion is the overlap of distributions, the assumed distribution is a Gaussian distribution, and a hypothesis testing method is used. The function in MATLAB (registered trademark) is relativeEntropy.
The one-dimensional Bhattacharya distance uses a distance method in which the evaluation criterion is the overlap of distributions, the assumed distribution is a Gaussian distribution, and the function in MATLAB® is bhattacharyyaDistance.
In the ROC analysis by logistic regression, the evaluation criterion is the overlap of distributions, the assumed distribution is non-Gaussian, and a distance method is used. The function in MATLAB (registered trademark) is perfcurve.

[Information processing method]
A method executed by one or more processors in the system shown in FIG. 1 will be described with reference to FIG.

First, in step ST1, the teacher data acquisition unit 10 acquires teacher data D1 in which tire images (contact patch images 20) related to tires are associated with predicted values.
In the next step ST2, the feature extraction unit 11 receives an input image based on a tire image and extracts a plurality of feature amounts for one tire image (contact patch image 20).
In the next step ST3, the learning unit 12 trains the decision tree prediction model 13 through machine learning so as to input the extracted multiple feature amounts and output corresponding prediction values in the teacher data D1.
In the next step ST4, the importance calculation unit 14 calculates the importance of each of the multiple feature amounts using the trained decision tree prediction model 13.
In the next step ST5, the receiving unit 15 receives a selection of a feature quantity to be searched from among the multiple feature quantities.
In the next step ST6, the input pattern generating section 16 generates a plurality of input patterns having different feature amounts of the selected search target.
In the next step ST7, the prediction unit 17 inputs the generated input patterns to the decision tree prediction model 13, respectively, and predicts the corresponding predicted values.
In the next step ST8, the threshold calculation unit 18 calculates a threshold value FV _Th1 of the feature amount of the search target, which represents a boundary where the predicted value fluctuates, based on the multiple predicted values.
By executing the above steps ST1 to ST8, a method for calculating a threshold value related to machine learning of a tire image is executed.

In addition to the above steps ST1 to ST8, the following steps are executed to implement a method for identifying information about image parameters by machine learning.
In step ST9, the classification unit 31 classifies the multiple tire images in the teacher data D1 into either a first group or a second group based on the threshold value FV _Th1 calculated by the threshold value calculation unit 18.
In step ST10, the image parameter calculation unit 30 calculates at least one image parameter from each of the tire images in the training data D1. Steps ST9 and ST10 may be performed in any order.
In the next step ST11, the characteristic information output unit 32 outputs characteristic information regarding the characteristics of the first image parameter group calculated from the tire image group belonging to the first group and the second image parameter group calculated from the tire image group belonging to the second group.
In the next step ST12, the determination unit 33 determines whether or not the image parameters are correlated with the predicted values based on the first image parameter group and the second image parameter group. The result of the determination may be included in the characteristic information and output.
Steps ST10 to ST12 are executed for a plurality of types (31 types) of image parameters.

Alternatively, it may be as shown in Fig. 10. In Fig. 10, steps ST1 to ST10 are the same as those in Fig. 9. As shown in Fig. 10, in step ST111 following step ST10, the evaluation information output unit 34 calculates evaluation values relating to the distribution of the first image parameter group and the second image parameter group for each of a plurality of image parameters. In the example shown in Fig. 8 of the first embodiment, these are test results (Wilcoxon statistics) of the Wilcoxon rank sum test.
Steps ST10 and ST111 are executed for a plurality of types (31 types) of image parameters.
In the next step ST112, the evaluation information output unit 34 outputs at least some of the image parameters in an order according to the evaluation values. The evaluation information output unit 34 of the first embodiment displays as shown in FIG.

<Modification>
<1> In the first embodiment, the tire image is a ground contact image, but the image is not limited to a ground contact image as long as it is an image related to a tire. For example, an image showing the exterior of a tire may be used.

<2> In the first embodiment, as shown in FIG. 1, the characteristic information output unit 32, the judgment unit 33, and the evaluation information output unit 34 are all implemented, but these can be omitted as appropriate.

<3> In the first embodiment, the predicted value is a tire performance value, but is not limited to this. For example, the predicted value may be an index value representing the amount of tire wear or the wear state of the tire, an index value representing the deterioration state of the tire, etc.

As described above, although not particularly limited, as in the first embodiment, the information processing method may be a method executed by one or more processors, and may include: a step (ST1) of acquiring teacher data in which tire images related to tires are associated with predicted values; a feature extraction step (ST2) of inputting an input image based on a tire image to the feature extraction unit 11 and extracting a plurality of feature amounts for one tire image; a learning step (ST3) of machine learning the decision tree prediction model 13 so as to input the extracted plurality of feature amounts and output corresponding predicted values in the teacher data D1; an importance calculation step (ST4) of calculating the importance of each of the plurality of feature amounts using the trained decision tree prediction model 13; a step (ST5) of accepting a selection of a feature amount to be a search target among the plurality of feature amounts; a step (ST6) of generating a plurality of input patterns in which the selected search target feature amounts are different from each other; a step (ST7) of inputting the generated plurality of input patterns into the decision tree prediction model 13 and predicting the corresponding predicted values; and a step (ST8) of calculating a threshold value of the feature amount to be searched, which represents a boundary within which the predicted value fluctuates, based on the predicted plurality of predicted values.
Although not limited to this, as in the first embodiment, the system may include a teacher data acquisition unit 10 that acquires teacher data in which tire images related to tires are associated with predicted values, a feature extraction unit 11 that inputs an input image based on a tire image and extracts multiple feature amounts for one tire image, a learning unit 12 that machine-learns a decision tree prediction model 13 so as to input the extracted multiple feature amounts and output corresponding predicted values in the teacher data D1, an importance calculation unit 14 that calculates the importance of each of the multiple feature amounts using the trained decision tree prediction model 13, a reception unit 15 that receives a selection of a feature amount to be a search target from among the multiple feature amounts, an input pattern generation unit 16 that generates multiple input patterns in which the selected search target feature amounts are different from each other, a prediction unit 17 that inputs the generated multiple input patterns into the decision tree prediction model 13 and predicts the corresponding predicted values, and a threshold calculation unit 18 that calculates a threshold value for the feature amount to be the search target that represents a boundary between which the predicted value fluctuates, based on the predicted multiple predicted values.

In this way, by using the decision tree prediction model 13, since the predicted value tends to vary significantly depending on the branching of the tree structure, it is possible to obtain a threshold value FV _Th1 of the feature value to be searched, which represents the boundary at which the predicted value varies. The threshold value FV _Th1 can be used to clarify the relationship between the feature value and the predicted value in machine learning that inputs a feature value based on a tire image related to a tire and predicts the corresponding predicted value. Therefore, it is possible to provide such new data.

Although not particularly limited, as in the first embodiment, the information processing method may be a method executed by one or more processors, and may include a method of calculating a threshold FV _Th1 related to machine learning of tire images (ST1 to 8), a classification step (ST9) of classifying a plurality of tire images in the teacher data D1 into either a first group G1 or a second group G2 based on the threshold FV _Th1 , an image parameter calculation step (ST10) of calculating at least one image parameter from each of the plurality of tire images in the teacher data D1, and a step (ST11) of outputting characteristic information D2 related to the characteristics of the first image parameter group calculated from the tire image group belonging to the first group G1 and the second image parameter group calculated from the tire image group belonging to the second group G2.
Although not particularly limited, as in the first embodiment, the system may include a system for calculating a threshold FV _Th1 for machine learning of tire images, a classification unit 31 for classifying a plurality of tire images in the teacher data D1 into either a first group G1 or a second group G2 based on the threshold FV _Th1 , an image parameter calculation unit 30 for calculating at least one image parameter from each of the plurality of tire images in the teacher data D1, and a characteristic information output unit 32 for outputting characteristic information D2 related to the characteristics of a first image parameter group calculated from the tire image group belonging to the first group G1 and a second image parameter group calculated from the tire image group belonging to the second group G2.

In this way, characteristic information D2 regarding the characteristics of the first image parameter group and the second image parameter group is output. If it can be evaluated that the distributions of the two groups are separate based on characteristic information D2, it can be determined that the image parameters have a correlation with the features to be searched for, and that the image parameters are parameters related to the predicted value.

Although not particularly limited, as in the first embodiment, the information processing method may include a step (ST12) of determining whether the image parameters are correlated with the predicted values based on the first image parameter group and the second image parameter group.
Although not particularly limited, as in the first embodiment, the system may include a determination unit 33 that determines whether or not the image parameters are correlated with the predicted values based on the first group of image parameters and the second group of image parameters.

In this way, it becomes easy to know which image parameters are correlated with the predicted values.

Although not limited thereto, the characteristic information may include a first quartile and a third quartile of the first group of image parameters, and a first quartile and a third quartile of the second group of image parameters.
This makes it possible for a person or a computer to evaluate the degree of separation of the distributions of the two groups based on the interquartile range (first quartile and third quartile) of the first image parameter group and the interquartile range (first quartile and third quartile) of the second image parameter group.

Although not particularly limited, as in the first embodiment, the information processing method may include executing the image parameter calculation step (ST10) multiple times to calculate multiple types of image parameters, and the information processing method may include a step (ST11) of outputting characteristic information D2 for each of the multiple types of image parameters.
Although not particularly limited, as in the system of the first embodiment, the image parameter calculation unit 30 may be executed multiple times to calculate multiple types of image parameters, and the characteristic information output unit 32 may output characteristic information D2 for each of the multiple types of image parameters.
This makes it possible to easily know the characteristic information D2 for a plurality of types of image parameters.

Although not particularly limited, as in the first embodiment, the information processing method may include executing the image parameter calculation step (ST10) multiple times to calculate multiple types of image parameters, and the information processing method may include a step (ST111) of calculating an evaluation value related to the distribution of a first image parameter group and a second image parameter group for each of the multiple image parameters, and a step (ST112) of outputting at least a portion of the multiple image parameters in an order according to the evaluation value.
Although not particularly limited, like the system of the first embodiment, the image parameter calculation unit 30 may be executed multiple times to calculate multiple types of image parameters, and the evaluation information output unit 34 may calculate an evaluation value relating to the distribution of the first image parameter group and the second image parameter group for each of the multiple image parameters, and output at least a portion of the multiple image parameters in an order according to the evaluation value.

This allows at least a portion of the multiple image parameters to be output in an order based on the evaluation value related to the distribution, making it possible to know the multiple parameters in order of high or low correlation with the predicted value.

Although not particularly limited, as in the first embodiment, the evaluation value may be calculated using a testing method including at least one of the Wilcoxon rank sum test, one-dimensional KL divergence, one-dimensional Bhattacharyya distance, and ROC analysis using logistic regression.
This makes it possible to evaluate, by ranking, the differences between the first and second image parameter groups, i.e., image parameters that are highly correlated with the predicted values.

Although not particularly limited, as in the first embodiment, the tire image may be a contact patch image 20 representing the contact patch shape and the contact pressure, the predicted value may be a tire performance value, and the image parameter may be a value calculated from at least one of the contact patch shape and the contact pressure. This makes it possible to understand the relationship between the tire performance value and the image parameter.

As in the first embodiment, the at least one image parameter may include at least one of the above (1) to (31). The above image parameters (1) to (31) appear in the contact patch image 20 and can be used in tire design, so that it is possible to directly know the image parameters that contribute to tire performance.

The program according to this embodiment is a program for causing a computer to execute the above method.
By executing these programs, it is possible to obtain the effects of the above-mentioned methods.

Although the embodiments of the present disclosure have been described above with reference to the drawings, the specific configuration should not be considered to be limited to these embodiments. The scope of the present disclosure is indicated not only by the description of the above embodiments but also by the claims, and further includes all modifications within the meaning and scope equivalent to the claims.

The structures employed in each of the above embodiments can be employed in any other embodiment. The specific configurations of each part are not limited to the above-mentioned embodiments, and various modifications are possible without departing from the spirit of this disclosure.

For example, the order of execution of each process, such as operations, procedures, steps, and stages, in the devices, systems, programs, and methods shown in the claims, specifications, and drawings, can be realized in any order, so long as the output of a previous process is not used in a subsequent process. Even if the flow in the claims, specifications, and drawings is explained using "first," "next," etc. for convenience, this does not mean that it is necessary to execute the processes in this order.

Each unit shown in FIG. 1 is realized by executing a specific program on one or more processors, but each unit may also be configured with a dedicated memory or dedicated circuit. In the above embodiment of the system 1, each unit is implemented in the processor 1a of one computer, but each unit may be distributed and implemented on multiple computers or in the cloud. In other words, the above method may be executed on one or more processors.

System 1 includes a processor 1a. For example, processor 1a can be a central processing unit (CPU), a microprocessor, or other processing unit capable of executing computer-executable instructions. System 1 also includes memory 1b for storing data for system 1. In one example, memory 1b includes computer storage media, including RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, DVD or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other media that can be used to store desired data and that can be accessed by system 1.

11: feature extraction unit, 13: decision tree prediction model, 20: contact patch image (images 1 to N) [tire image], D1: teacher data, D2: characteristic information, FV ₁ to FV _M : feature amount, FV _Th1 : threshold value, G1: first group, G2: second group, X ₁ to X _N : tire performance value (predicted value), Y ₁ to Y _P : predicted value corresponding to input patterns 1 to P.

Claims (11)

A method executed by one or more processors, comprising:
acquiring training data in which tire images and predicted values related to the tire are associated with each other;
a feature extraction step of inputting an input image based on the tire image to a feature extraction unit and extracting a plurality of feature amounts for one tire image;
a learning step of machine learning a decision tree prediction model so as to input the extracted plurality of feature quantities and output corresponding prediction values in the training data;
an importance calculation step of calculating an importance of each of the plurality of feature amounts by using the trained decision tree prediction model;
receiving a selection of a feature quantity to be searched from among the plurality of feature quantities;
generating a plurality of input patterns each having different feature values of the selected search target and the same feature values in other respects ;
inputting the generated input patterns into the decision tree prediction model and predicting corresponding prediction values;
calculating a threshold value of the feature quantity of the search target, which represents a boundary within which the predicted value varies, based on the plurality of predicted values;
An information processing method comprising: a classification step of classifying a plurality of tire images in the training data into either a first group or a second group based on the threshold value;
an image parameter calculation step of calculating at least one image parameter from each of the plurality of tire images in the training data;
outputting characteristic information relating to characteristics of a first image parameter group calculated from the tire image group belonging to the first group and a second image parameter group calculated from the tire image group belonging to the second group;
The method of claim 1 , comprising: The method of claim 2, further comprising: determining whether the image parameters are correlated with the predicted value based on the first set of image parameters and the second set of image parameters. The method according to claim 2 or 3, wherein the characteristic information includes a first quartile and a third quartile of the first group of image parameters, and a first quartile and a third quartile of the second group of image parameters. Executing the image parameter calculation step a plurality of times to calculate a plurality of types of image parameters;
The method comprises:
The method according to any one of claims 2 to 4, further comprising the step of outputting the characteristic information for each of the plurality of types of image parameters. Executing the image parameter calculation step a plurality of times to calculate a plurality of types of image parameters;
The method comprises:
calculating an evaluation value relating to distribution of the first image parameter group and the second image parameter group as characteristic information relating to characteristics of the first image parameter group and the second image parameter group for each of the plurality of types of image parameters;
outputting at least a portion of the plurality of types of image parameters in an order according to the evaluation values;
The method according to any one of claims 2 to 4, comprising: The method according to claim 6, wherein the evaluation value is calculated using a test method including at least one of the Wilcoxon rank sum test, one-dimensional KL divergence, one-dimensional Bhattacharyya distance, and ROC analysis using logistic regression. the tire image is a contact patch image representing a contact patch shape and a contact pressure;
The predicted value is a tire performance value,
The method according to any one of claims 2 to 7, wherein the image parameter is a value calculated from at least one of the contact patch shape and the contact pressure. The at least one image parameter is
(1) The area of the ground surface in the image,
(2) The sum of the area of the contact patch and the area of the groove in the image,
(3) the ratio of (1) to (2),
(4) The sum of the values obtained by multiplying the area per unit pixel by the ground contact pressure for all pixels that make up the ground contact surface,
(5) A total value obtained by multiplying the distance from each of the positions of the pixels constituting the ground contact surface to the center of gravity of the ground contact surface by the ground contact pressure;
(6) The average value of the area of each of the multiple blocks divided by the perimeter of the block;
(7) the number of blocks;
(8) The four angles of the area including the contact surface and the groove area;
(9) Ratio of the tire circumferential dimension to the tire axial dimension of the contact patch;
(10) The number of circumferential grooves extending in the tire circumferential direction;
(11) Width of a circumferential groove extending in the tire circumferential direction,
(12) A value obtained by summing the distances from the positions of each of a plurality of pixels constituting a circumferential groove extending in the tire circumferential direction to the center of gravity of the circumferential groove;
(13) The spacing between axially extending lateral grooves of the tire;
(14) Width of a lateral groove extending in the tire axial direction;
(15) The area of a lateral groove extending in the tire axial direction;
(16) Angle of a lateral groove extending in the tire axial direction;
(17) The width of the tire in the axial direction of the contact surface; (18) The maximum contact length of the center land closest to the equatorial plane of the contact surface;
(19) The maximum contact length of the shoulder land on the outermost side of the tire in the axial direction of the contact surface;
(20) The ratio of (18) to (19),
(21) The sum of the ground pressures of the center land closest to the equatorial plane of the contact surface divided by the area of that center land;
(22) The sum of the contact pressures of the shoulder land located at the outermost part of the tire in the axial direction of the contact surface divided by the area of the shoulder land.
(23) The ratio of (21) to (22),
(24) (1) divided by the test load or (4) above,
(25) The value obtained by dividing (2) by the test load or (4) above,
(26) The value obtained by dividing (3) by the test load or (4) above,
(27) The value obtained by dividing (4) by the test load,
(28) The value obtained by dividing (5) by the test load,
(29) The value obtained by dividing (21) by the test load or (4) above,
(30) The value obtained by dividing (22) by the test load or (4) above,
(31) The value obtained by dividing (23) by the test load or (4) above,
The method of claim 8 , comprising at least one of: A system comprising one or more processors for executing the method according to any one of claims 1 to 9. A program that causes one or more processors to execute the method according to any one of claims 1 to 9.

Images