Machine learning-based technology for classifying the properties of skincare formulations

Abstract

To classify a wide variety of skin care formulations in a short time to learn the types of the functions of the skin care formulations.SOLUTION: A method for creating a learned model for classifying the properties of skin care formulations includes the steps of: performing principal component analysis on a plurality of frictional characteristic values obtained from a frictional force between two contactors in each of which a sample of the skin care formulation is applied on a contact interface, thereby obtaining one or more principal components; and performing machine learning by using, as training data, a plurality of datasets including the principal components, the viscosity of the sample, and the classification of the properties of the skin care formulations, thereby creating the learned model.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a technique for classifying the properties of skin care preparations by machine learning. More specifically, the present invention relates to a method for generating a learned model for classifying the dosage form of a skin care preparation, a learning device for generating the learned model, a learning program, a classification method using the learned model, a classification device, and a classification program.

Background Art

[0002] For skin care preparations, the touch sensation during use is an important factor directly related to product characteristics. A skin care preparation (hereinafter also simply referred to as "preparation") is usually spread over the entire target area by rubbing it with a finger after applying an appropriate amount to the skin. At this time, the touch sensation such as the smoothness of the finger, the stickiness felt at the fingertips, and the viscosity greatly affects the consumer's preference.

[0003] The touch sensation of such a preparation is generally evaluated by a sensory evaluation by humans (expert panels). Although the touch sensation of the preparation in the actual situation becomes clear in the sensory evaluation, there is a disadvantage that it is difficult to perform a quantitative evaluation. As a quantitative evaluation method for overcoming such a sensory evaluation, analysis of the relationship between a specific sense (specifically, touch sensation) and the physical properties of the preparation has been advanced.

[0004] For example, Patent Document 1 discloses a touch characteristic evaluation method for quantitatively evaluating the touch sensation when a human applies a preparation to the skin. In the method described in Patent Document 1, a friction vibration generated when spreading a preparation with a finger is measured for each finger position by a distance meter and a vibration sensor attached to a human forearm or finger, and the touch characteristics of the preparation are evaluated or classified based on the friction vibration. Thus, in Patent Document 1, it is proposed to analyze the physical properties of a preparation by an objective method代替 the sensory evaluation, focusing on the relationship between the touch sensation of the preparation and the friction vibration (frictional force).

Prior Art Documents

Patent Documents

[0005] [Patent Document 1] Japanese Patent Publication No. 2021-139633 [Overview of the project] [Problems that the invention aims to solve]

[0006] Incidentally, the evaluation method in Patent Document 1 focuses on the frictional vibrations at different locations that occur when the formulation is spread with the fingers as a physical property of the formulation. However, there are many types of skincare formulations, such as lotions, serums, and toners, and their properties (dosage forms) are also diverse, such as liquid, paste, and gel. Therefore, it is difficult to accurately classify the functional properties of a formulation based solely on information about the frictional vibrations of the formulation and their location. On the other hand, if the number of types of physical properties of the formulation used in the analysis to classify the properties is increased, the problem arises that it becomes difficult to identify which characteristic value of the formulation affects the functional properties and to what extent.

[0007] Therefore, the main objective of the present invention is to provide a technology that can quickly classify the functional properties of various skincare formulations according to their physical characteristics. [Means for solving the problem]

[0008] The inventors of the present invention diligently studied solutions to the problems of the prior art and found that by obtaining multiple friction characteristic values ​​from the frictional force of two contacts coated with a sample of a skincare formulation, and then performing machine learning using the data containing the principal components obtained by applying principal component analysis to these multiple friction characteristic values ​​as training data, it becomes possible to predict the properties of a wide variety of skincare formulations in a short time. Based on this finding, the inventors realized that the problems of the prior art could be solved and completed the present invention. Specifically, the present invention has the following configuration or steps.

[0009] The first aspect of the present invention relates to a method for generating a trained model for classifying the properties of skincare formulations. In this specification, "skincare formulations" include basic cosmetics such as lotions, serums, toners, facial cleansers, or cleansing oils, as well as sunscreens. Furthermore, the "properties" of a skincare formulation refer to the physical state of the skincare formulation, such as liquid, paste, gel, cream, or paste, which is mainly classified by human sensory evaluation.

[0010] The method for generating a trained model according to the present invention includes a principal component analysis step and a training step. The principal component analysis determines one or more principal components by performing principal component analysis on multiple friction characteristic values ​​related to a sample of a skincare formulation. These friction characteristic values ​​are values ​​obtained from the frictional force of two contacts to which the skincare formulation sample is applied at the contact interface. The frictional force of these two contacts itself is not included in the frictional characteristic values. In this specification, the frictional force of the two contacts to which the skincare formulation sample is applied at the contact interface, or the frictional characteristic value obtained therefrom, is also referred to simply as the frictional force or frictional characteristic value of the skincare formulation. The training step is a step of generating a trained model by performing machine learning using multiple datasets, which include the principal components obtained by the principal component analysis, the viscosity of the sample, and the correct labels related to the classification of the properties of the skincare formulation, as training data. In the present invention, "machine learning" can be any known method that performs supervised learning, such as a deep learning model, and its structure and learning method are not particularly limited. Furthermore, a "trained model" is a model whose features and parameters (weights) have been optimally adjusted by loading a large dataset into the model.

[0011] It is empirically clear that the frictional force and viscosity of a formulation are related to its sensory properties (especially tactile sensation), and previous studies have also acknowledged a certain degree of correlation between the two. However, because the properties of formulations are diverse, and humans can subtly perceive differences in the feel of formulations, it has been difficult to classify the properties of formulations with a level of detail that humans can recognize based solely on their frictional force and viscosity. In contrast, this invention applies principal component analysis to multiple frictional characteristic values ​​of skincare formulations, and then performs machine learning using the resulting combination of principal components, viscosity, and correct labels as training data. The frictional characteristic values ​​of skincare formulations will be described in more detail later, but various physical properties that can affect human sensory perception during the use of the formulation can be used, such as the static friction coefficient and the dynamic friction coefficient. Furthermore, the principal components of the frictional characteristic values ​​can explain most of the frictional characteristic values ​​by adjusting the cumulative contribution rate (for example, to 80% or more or 90% or more). By using the principal components of these multiple friction characteristic values ​​as training data for machine learning, it is theoretically possible to obtain a trained model capable of classifying a wide variety of skincare formulations with a level of detail that humans can recognize. Then, by using such a trained model for inference, it becomes possible to classify the properties of skincare formulations in a short amount of time.

[0012] In the method for generating a trained model according to the present invention, the frictional force of the two contacts is preferably measured by repeatedly sliding at least one of the two contacts within the area to which the sample is applied. In this case, the frictional characteristic value of the formulation preferably includes the static friction coefficient (maximum), the dynamic friction coefficient (average), and the fluctuation value of the dynamic friction coefficient during the sliding motion period from the time the dynamic friction force is generated to a predetermined time. An example of the "predetermined time" here is the time before the sliding motion of the contacts enters the deceleration process, but is not limited to this. That is, the fluctuation value of the dynamic friction coefficient should be measured in either the forward or return path of the reciprocating contact, during the period from the time the dynamic friction force is generated to the time immediately before the sliding motion of the contacts enters the deceleration process. Furthermore, the "fluctuation value of the dynamic friction coefficient" is the value obtained by subtracting the minimum value from the maximum value of the dynamic friction coefficient measured during a predetermined sliding motion period. In order to achieve the objective of the present invention, which is to classify a wide variety of formulations according to their properties, it is considered particularly effective to use the static friction coefficient, the dynamic friction coefficient, and the fluctuation value of the dynamic friction coefficient as variables in the principal component analysis. In addition, the following values ​​can be used as friction characteristic values ​​and as variables in principal component analysis.

[0013] In the method for generating a trained model according to the present invention, the friction characteristic value may further include the difference between the static friction coefficient and the kinetic friction coefficient. The difference value here may be, for example, the value obtained by subtracting the kinetic friction coefficient from the static friction coefficient. However, it is also possible to use the value obtained by subtracting the static friction coefficient from the kinetic friction coefficient. The static friction coefficient is uniquely determined as it is the value obtained by dividing the maximum static friction force by the normal force acting on the contact surface. On the other hand, the kinetic friction coefficient may change during measurement. In that case, the difference between the average value of the kinetic friction coefficient over a predetermined period and the static friction coefficient should be obtained. As a general theory of friction, it is said that static friction force is greater than kinetic friction force (Amontons-Coulomb law). However, this law does not hold when a fluid such as a lubricant is interposed between objects. Since the present invention uses the friction force of a skincare formulation as a characteristic value, static friction force > kinetic friction force is not necessarily true. For this reason, it is meaningful to include a value that quantifies the relationship between static friction force and kinetic friction force in the friction characteristic value, such as the difference between the kinetic friction coefficient and the static friction coefficient described above.

[0014] In the method for generating a trained model according to the present invention, the sliding motion of the contactor is reciprocating motion, and the friction characteristic value may further include the average value and / or variation of the kinetic friction coefficient in each section when the forward or return process is divided into multiple sections. To determine the average value of the kinetic friction coefficient in each section, for example, when the forward process of the contactor is divided into two sections, the first half and the second half, the average value of the kinetic friction coefficient measured at numerous plots included in each section can be obtained. Also, for example, when the forward process of the contactor is divided into two sections, the first half and the second half, the variation of the kinetic friction coefficient in the first half section of a certain forward process can be calculated as (maximum value of n plots of the kinetic friction coefficient in the first half section of a certain forward process) - (minimum value of n plots of the kinetic friction coefficient in the first half section of the same forward process). Similarly, the variation of the kinetic friction coefficient in the second half section of a certain forward process can also be obtained. Note that "plot" means a point on which data is recorded, and "n" is a natural number of 2 or more representing the number of such plots. For example, if friction characteristic data is recorded every 2ms (2 milliseconds), and the entire forward journey takes 1 second (1 second), then 500 plots of data will be measured for the entire forward journey. If this forward journey is divided into two sections, the first and second halves, then each section will yield 250 plots of data in 0.5 seconds. In each section, data for the kinetic friction coefficient can be extracted from the total number of plots (e.g., 100 plots) and their average value can be calculated.

[0015] In the method for generating a learning model according to the present invention, the friction characteristic value may further include the change in the average value of the kinetic friction force in the preceding and succeeding sections. For example, if the forward path process of a contactor is divided into two sections, the first half and the second half, the change in the average value of the kinetic friction coefficient in the preceding and succeeding sections can be calculated as (average value of the n-plot of the kinetic friction coefficient in the first half section) - (average value of the n-plot of the kinetic friction coefficient in the second half section).

[0016] A second aspect of the present invention relates to a learning device for training a model for classifying the properties of skincare formulations. The learning device according to the second aspect is basically a computer device for executing the method for generating a learning model according to the first aspect described above. The learning device comprises a principal component analysis unit and a learning unit. The principal component analysis unit obtains one or more principal components by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts to which a sample of the skincare formulation is applied at the contact interface. The learning unit trains a model by performing machine learning using multiple datasets, which include the principal components of the friction characteristic values, the viscosity of the sample, and correct labels related to the classification of the properties of the skincare formulation, as training data.

[0017] A third aspect of the present invention is a program for causing a computer to function as a learning device according to the second aspect described above. This program may be pre-installed on the computer, available for download via the internet, or stored on a storage medium such as a CD-ROM.

[0018] A fourth aspect of the present invention relates to a method for classifying skincare formulations. The classification method according to the fourth aspect basically classifies skincare formulations using a trained model generated by the learning model generation method according to the first aspect described above. The classification method for skincare formulations includes a principal component analysis step and an inference step. In the principal component analysis step, one or more principal components are determined by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts to which a sample of the skincare formulation is applied at the contact interface. In the inference step, the principal components of the friction characteristic values ​​and the viscosity of the sample are input into the trained model to obtain information (estimated label) regarding the classification of the sample.

[0019] A fifth aspect of the present invention relates to a classification device for classifying skincare formulations. The classification device according to the fourth aspect is basically a computer device for executing the classification method according to the fourth aspect described above. The classification device according to the fifth aspect comprises a principal component analysis unit and a trained model. The principal component analysis unit obtains one or more principal components by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts to which a sample of the skincare formulation is applied at the contact interface. The trained model is generated by the method for generating the trained model according to the first aspect described above. When the principal components of the friction characteristic values ​​and the viscosity of the sample are input to this trained model, it outputs information regarding the classification of the sample (estimated label).

[0020] A sixth aspect of the present invention is a program for causing a computer to function as a classification device relating to the fifth aspect described above. This program may be pre-installed on the computer, available for download via the internet, or stored on a storage medium such as a CD-ROM. [Effects of the Invention]

[0021] According to the present invention, it is possible to quickly classify the sensory properties of various skincare formulations according to their physical characteristics. Specifically, by classifying skincare formulations according to their physical characteristics, it is possible to predict the sensory properties associated with those physical characteristics, or to predict which other formulations have similar properties. Furthermore, once a trained model is generated, classification can be performed quickly regardless of the number of formulation samples. For example, the present invention can be effectively used in the research and development of formulations when it is necessary to visualize the differences in sensory properties between existing products and renewed products, or between prototypes and target products, or between a company's own products and those of competitors. [Brief explanation of the drawing]

[0022] [Figure 1]FIG. 1 is a block diagram showing the functional configurations of a learning device and a classification device according to an embodiment of the present invention. [Figure 2] FIG. 2 schematically shows an example of a measuring device for measuring the friction characteristics of a sample of a skin care preparation. [Figure 3] FIG. 3 shows an example of the hardware configuration of a computer that functions as the learning device and the classification device of the present invention. MODE FOR CARRYING OUT THE INVENTION

[0023] Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. The present invention is not limited to the embodiments described below, and includes those appropriately modified by those skilled in the art within the obvious scope from the following embodiments.

[0024] FIG. 1 shows the functional configurations of a learning device 10 and a classification device 20 according to an embodiment of the present invention. As shown in FIG. 1, the learning device 10 generates a learned model 21 by executing machine learning using teacher data including correct labels. The classification device 20 includes the learned model 21 generated by the learning device 10. When the classification device 20 inputs input data 23 to this learned model 21, it outputs output data 24 corresponding to the correct label. Thereby, the input data 23 can be classified. Hereinafter, the functions of each device 10 and 20 will be specifically described.

[0025] [1. Learning device] In the present embodiment, the learning device 10 includes a data set 11, a principal component analysis unit 12, and a machine learning unit 13. The learning device 10 is configured by a general-purpose computer. The data set 11 is stored in a storage device such as a memory or a storage provided in the computer. The principal component analysis unit 12 and the machine learning unit are realized by a processor (arithmetic device) such as a CUP or a GPU provided in the computer executing a predetermined program.

[0026] Dataset 11 contains a large amount of data for each skincare formulation sample, associating multiple frictional properties, rheological properties (viscosity), and correct labels corresponding to the classification of the formulation's properties. The frictional properties of a formulation are determined from the frictional force between two contacts applied to the contact interface of the formulation sample. Details on the measurement method of frictional force and the frictional properties will be described later. The rheological properties are basically the viscosity of the formulation. This viscosity should be measured according to official standards such as "JIS Z 8803 Viscosity of liquids - Measurement method". The correct labels should be labels that classify the texture of the formulation to a level of detail that can be classified by human sensory evaluation, such as liquid, paste-like, gel-like, cream-like, or paste-like. The classification of formulations is not limited to the examples given here, and further detailed classification is possible. For example, labels corresponding to human sensory vocabulary can be assigned, such as "blends in (penetrates)," "smooth," "adheres well," "moist," "gentle on the skin," "rich," "soft and fluffy," "spreads easily," "not too heavy," "becomes lighter halfway through," and "moist." Preferably, there should be at least two correct labels, but it can also be five or more, seven or more, or ten or more.

[0027] The principal component analysis unit 12 performs principal component analysis on multiple friction characteristic values ​​included in the dataset 11 to obtain one or more principal components that summarize these values. Principal component analysis is a statistical data analysis method that reduces multidimensional quantitative data described by many explanatory variables to lower-dimensional, uncorrelated composite variables (principal components) with as little information loss as possible. As will be described in detail later, in this embodiment, a large number of values ​​are used as friction characteristic values ​​of the formulation, including static friction coefficient, dynamic friction coefficient, variation of the dynamic friction coefficient, and the difference between the static friction coefficient and the dynamic friction coefficient. If these numerous friction characteristic values ​​are used directly as training data for machine learning, the model structure will become complex, raising concerns that the accuracy of the inference results will decrease or that it will take a long time to output the inference results. For this reason, in this invention, multiple friction characteristic values ​​(explanatory variables) are subjected to principal component analysis to obtain fewer principal components (composite variables), and these principal components are used as training data for machine learning. There are no particular restrictions on the number of principal components used as training data, but it is sufficient to select them in descending order of their contribution rate so that the cumulative contribution rate is 80% or higher, preferably 90% or higher. Alternatively, all principal components with one or more eigenvalues ​​may be used as training data for machine learning. An "eigenvalue" is an indicator that represents the amount of information contained in each principal component, and principal components with one or more eigenvalues ​​are considered to be closely related to the original data. The "contribution rate" is an indicator that represents what proportion of the total information in the original data is represented by the eigenvalue of a particular principal component, and the "cumulative contribution rate" is the sum of the contribution rates of each principal component in descending order, indicating how much of the information contained in the data is explained by the principal components up to that point.

[0028] The learning unit 13 performs machine learning and trains the learning model using training data which is the viscosity and correct labels contained in the dataset 11 plus principal components obtained by principal component analysis. As for the machine learning method, any known supervised learning method used to solve classification problems can be adopted. For example, known methods such as SVM (Support Vector Machine), logistic regression, k-nearest neighbors, decision trees, and neural networks can be used. Among these methods, SVM has linear and nonlinear discriminant functions and is considered to have the best discriminant and classification ability among the many machine learning methods known today, so it is preferable to adopt SVM in this invention. SVM is a learning model mainly for solving binary classification, and the classifier is constructed using linear threshold elements. In SVM, the parameters of the linear threshold elements are learned based on the criterion of maximizing the margin that maximizes the distance to each data point in the training data.

[0029] A trained model 21 is generated by training the learning model with the learning unit 13. This trained model 21 is optimized for classifying the properties of skincare formulations based on the main components of their frictional properties and viscosity. This trained model 21 is primarily used by being incorporated into the classification device 20.

[0030] [2. Friction Characteristics Values] Next, we will explain in detail the friction characteristic values ​​(explanatory variables) of skincare formulations that are subjected to principal component analysis.

[0031] First, with reference to Figure 2, a method for measuring the frictional force of a skincare formulation will be described. Figure 2 shows an example of a friction measuring device 30. As shown in Figure 2, the friction measuring device 30 includes an artificial skin substrate 31 (first contact), a fingerprint mimic contact 32 (second contact), a load cell 33, a weight 34, and a heating plate 35, as well as a reciprocating motion mechanism (not shown) for repeatedly sliding the artificial skin substrate 31 and the heating plate 35 from side to side. The artificial skin substrate 31 mimics the surface shape of human skin and is fixed on the heating plate 35. In particular, this substrate 31 has a structure that mimics the wetting properties of human skin. The fingerprint mimic contact 32 mimics the fingerprint shape of a human fingertip and is positioned so that its tip is in constant contact with the substrate 31 when a predetermined load is applied. In this state, the sliding mechanism is driven to slide the substrate 31 and the heating plate 35, causing the substrate 31 and the contact 32 to rub against each other. The load cell 33 is a transducer that converts changes in load (axial force) into electrical signals and is attached to the contact 32. A weight 34 for applying load is also attached to the contact 32, and the load cell 33 detects the change in force when the substrate 31 and the contact 32 rub against each other. The measured value from the load cell 33 is input to a computing device (not shown), such as a computer. The computing device calculates the frictional force from the measured value of the load cell 33. The heating plate 35 heats the artificial skin substrate 31 during the measurement of the frictional force, maintaining the temperature of the substrate 31 at approximately human body temperature (for example, in the range of 30 to 40 degrees Celsius). This makes it possible to artificially reproduce the conditions when a skincare product is actually applied to human skin and spread with the fingers.

[0032] Furthermore, as shown in Figure 2, a sample S of the skincare formulation is applied to the contact interface between the artificial skin substrate 31 and the fingerprint mimicking contact 32. The sample S is either pre-applied to the entire area where the contact 32 slides on the substrate 31 (the area indicated by the arrow), or applied by sliding a bulk sample attached to the contact 32 on the substrate 31. The physical properties of the sample S change the frictional force acting between the substrate 31 and the contact 32 during the sliding motion. In other words, if the lubricity of the sample S is high, the frictional force between the substrate 31 and the contact 32 will be low, and if the lubricity of the sample S is low, the frictional force between the substrate 31 and the contact 32 will be high.

[0033] The term "friction" here includes both static friction and kinetic friction. The maximum friction force that occurs just before an object begins to slide when a force of a certain magnitude or greater is applied to a stationary object is called "maximum static friction," while the friction force that occurs while an object is moving is called "kinetic friction." Here, the maximum static friction F and the kinetic friction F' are expressed by the following equations, respectively. [Maximum static friction force] F=μN [Dynamic friction force] F´=μ´N As shown in the above equation, both the maximum static friction force F and the kinetic friction force F' are proportional to the normal force N. On the other hand, in the above equation, μ represents the coefficient of static friction and μ' represents the coefficient of kinetic friction, but both friction coefficients (μ,μ') are dimensionless quantities obtained by dividing the friction force (F,F') by the normal force (N) acting on the contact surface. Therefore, by converting the friction force (F,F') to the friction coefficient (μ,μ'), the load applied to the object can be ignored.

[0034] In this embodiment, not only the coefficient of dynamic friction but also the coefficient of static friction is measured. Therefore, the reciprocating motion of the contactor 32 on the substrate 31 is not performed continuously in the forward and return phases, but rather the contactor 32 is stopped once between the forward and return phases, and then the stopped contactor 32 is moved again.

[0035] During the sliding motion, the frictional force acting between the substrate 31 coated with the sample S and the contactor 32 is measured at numerous plots during each of the repeated forward and return movements of the contactor 32. For example, if the frictional force data is measured every 2ms (2 milliseconds) using a load cell 33, and the entire forward and return process takes 1 s (1 second), then the number of plots (data measurement points) will be 500 for the forward process and 500 for the return process. Here, it is not necessary to use the data measured at all plots in the forward and return processes; for example, the data used may be limited to only one of the forward or return processes, and the data from the other process may be ignored. In this embodiment, as an example, only the data acquired in the forward process is adopted, and the data acquired in the return process is ignored.

[0036] Furthermore, although the contactor 32 will repeatedly reciprocate on the substrate 31, it is not necessary to use the friction force data obtained for all forward (or return) cycles; data from only some of the forward (or return) cycles may be used. For example, if the contactor 32 is to reciprocate 20 times on the substrate 31, it is possible to use data only from the 1st, 2nd, 5th, and 20th forward cycles. The data from which forward cycles to use, or the number of forward cycles to use, can be adjusted as appropriate.

[0037] Alternatively, the friction force data may be organized by dividing one stroke of the contactor 32 (either the forward or return stroke) into multiple sections. For example, one stroke may be divided into two sections, the first half and the second half, or into three sections, the early, middle, and late stages, or even more. This makes it possible to, for example, determine the change in the friction coefficient when comparing one section of a given forward stroke with the same section of the next forward stroke, or to determine the change in the friction coefficient when comparing one section of a given forward stroke with the next section.

[0038] An example of a method for measuring frictional force is as follows: [Equipment Used] TL201Tt Static and Dynamic Friction Measuring Machine (Trinity Lab Co., Ltd.) [Number of measurements] Measurements are taken 3 times for each condition, and the average value is used for analysis. [Weight] A 30g weight was used. [Movement speed] 30mm / s [Traveling distance] 30mm [Number of round trips] 20 round trips per condition [Base Plate] Skin Texture Model (Made of polyurethane, Trinity Lab Co., Ltd.) [Contactor] Tactile contactor (made of polyurethane, Trinity Lab Co., Ltd.) [Measurement environment] Room temperature 25℃, humidity 50%, substrate temperature 36℃

[0039] In this embodiment, the frictional force of the skincare formulation sample S measured as described above (more precisely, the frictional force acting between the substrate 31 on which the sample S is coated and the contactor 32) is used to calculate the following 12 frictional characteristic values. Furthermore, only the data measured during the forward path of the contactor 32 is used, and each forward path is divided into three sections: early, middle, and late, and the data is organized accordingly. In addition, only the data from the 1st, 2nd, 5th, and 20th forward paths are used for each item. Therefore, the number of explanatory variables subjected to principal component analysis is 48 (12 items × 4 forward paths).

[0040] [Friction property value] 01. Static friction coefficient (maximum) 02. Coefficient of kinetic friction (average) 03. Variation of the coefficient of kinetic friction 04. Difference between static friction coefficient and kinetic friction coefficient 05. Average value of the coefficient of kinetic friction in the first half of the outbound journey (100 plots) 06. Fluctuation values ​​of the coefficient of dynamic friction in the first half of the outbound journey. 07. Average value of the coefficient of kinetic friction in the mid-term section of the outbound journey (100 plots) 08. Fluctuation values ​​of the coefficient of dynamic friction in the mid-term section of the outbound journey. 09. Average value of the coefficient of kinetic friction in the later section of the outbound journey (100 plots) 10. Fluctuation values ​​of the coefficient of dynamic friction in the later section of the outbound journey. 11. Change in the average value of the coefficient of kinetic friction in the first and middle sections of the outbound journey. 12. Change in the average value of the coefficient of kinetic friction in the mid-term and late-term sections of the outbound journey.

[0041] Of the 12 friction characteristic values ​​listed above, items 01-03 have the greatest impact on the sensory properties of the formulation and are therefore considered to have the highest priority. Item 04 is the next highest priority, followed by items 05-09, and then items 11-12. The following sections will explain each of the friction characteristic values.

[0042] (01. Static friction coefficient) As mentioned earlier, the maximum static friction force is the frictional force that occurs just before an object begins to slide when a force of a certain magnitude or greater is applied to a stationary object. This maximum static friction force F is expressed by the equation F = μN and is proportional to the normal force N. The coefficient of static friction μ is a dimensionless quantity obtained by dividing the maximum static friction force F by the normal force N acting on the contact surface.

[0043] (02. Coefficient of kinetic friction) As mentioned above, the frictional force generated during the movement of an object is called kinetic friction. This kinetic friction force F' is expressed by the equation F' = μ'N and is proportional to the normal force N. The coefficient of kinetic friction μ' is a dimensionless quantity obtained by dividing the kinetic friction force F' by the normal force N acting on the contact surface. Furthermore, in this embodiment, if the coefficient of kinetic friction changes during one stroke of the contactor (either the forward or return stroke), the coefficient of kinetic friction referred to here means the average coefficient of kinetic friction measured during one stroke.

[0044] (03. Variation in the coefficient of kinetic friction) The variation in the coefficient of dynamic friction refers to the variation in the coefficient of dynamic friction force during the sliding motion period of the contactor from the point of generation of dynamic friction to a predetermined point in time. Specifically, the variation in the coefficient of dynamic friction force can be calculated by subtracting the minimum value from the maximum value of the coefficient of dynamic friction measured during one stroke of the contactor, from the point of generation of dynamic friction to the point in time when the contactor enters the deceleration process. Generally, there is no significant variation in the dynamic friction force or coefficient of dynamic friction that occurs between solid surfaces. However, in friction measurements under the condition that a skincare formulation has been applied, the presence of a lubricating layer must be considered because the formulation acts as a lubricant. In addition, under the action of spreading the formulation, it is expected that the thickness of the lubricating layer will be uneven. For example, it is possible that the lubricating layer will be thinner the longer the application distance, or thicker the lubricating layer will be the higher the viscosity of the formulation. For this reason, when spreading the formulation, the magnitude of dynamic friction between the contactor and the substrate is not constant, and it is expected that it will gradually fluctuate as the reciprocating motion continues. It is thought that this frictional characteristic of gradually fluctuating coefficient of dynamic friction affects the sensory experience of the formulation during use. Therefore, in this embodiment, the amount by which the coefficient of dynamic friction changes during the single-coating operation (the section from the point where dynamic friction begins until the contactor enters the deceleration process) is calculated as the variation value of the coefficient of dynamic friction.

[0045] (04. The difference obtained by subtracting the coefficient of kinetic friction from the coefficient of static friction) The difference between the static friction coefficient and the kinetic friction coefficient can also be used as one of the friction characteristic values. The static friction coefficient, as mentioned above, is uniquely determined by dividing the maximum static friction force F by the normal force N acting on the contact surface. On the other hand, the kinetic friction coefficient can change during a single stroke. In this case, for example, if a stroke is divided into three sections—early, middle, and late—to make it easier to compare the static friction coefficient with the kinetic friction coefficient immediately following it, the difference between the average value of the kinetic friction coefficient data measured at each plot during the early period and the static friction coefficient can be calculated. If there are many plots during the early period, the difference between the average value of the kinetic friction coefficient data from 100 randomly selected plots and the static friction coefficient can also be used. As a general friction theory, the maximum static friction force is said to be greater than the kinetic friction force (Amontons-Coulomb law). However, this law does not hold true for fluid-lubricated surfaces. Actual measurements revealed a tendency for thick, high-viscosity skincare creams to exhibit static friction coefficients that were equal to or below the kinetic friction coefficient. On the other hand, watery, low-viscosity lotions exhibited static friction coefficients that were higher than their kinetic friction coefficients. Therefore, it is thought that the difference between the static friction coefficient and the kinetic friction coefficient also influences the sensory experience of the formulation during use.

[0046] (05-10. Average value and variation of the coefficient of kinetic friction in each section) As mentioned above, in this embodiment, each forward journey is divided into three sections: early, middle, and late, and the data is organized accordingly. Here, 100 plots are randomly selected from the early, middle, and late sections, and the average value of the kinetic friction coefficient measured in each of the 100 plots is calculated. This allows us to obtain the average value of the kinetic friction coefficient in the early section of the forward journey (100 plots), the average value of the kinetic friction coefficient in the middle section of the forward journey (100 plots), and the average value of the kinetic friction coefficient in the late section of the forward journey (100 plots). Furthermore, the variation value of the kinetic friction coefficient in each section can be calculated, for example, for the early section, by (maximum value of the kinetic friction coefficient in the early section of a given forward journey) - (minimum value of the kinetic friction coefficient in the early section of the same forward journey). The same applies to the variation value of the kinetic friction coefficient in the middle section and the variation value of the kinetic friction coefficient in the late section.

[0047] (11-12. Change in the average value of the coefficient of kinetic friction in the preceding and succeeding intervals) As mentioned above, in this embodiment, each forward journey is divided into three sections: early, middle, and late, and the data is organized accordingly. Therefore, the change in the average value of the kinetic friction coefficient between preceding and succeeding sections can also be used as one of the friction characteristic values. Specifically, the change in the average value of the kinetic friction coefficient between the early and middle sections of the forward journey is calculated as (average value of 100 plots of the kinetic friction coefficient in the early section of a given forward journey) - (average value of 100 plots of the kinetic friction coefficient in the middle section of the same forward journey). Similarly, the change in the average value of the kinetic friction coefficient between the middle and late sections of the forward journey is calculated as (average value of 100 plots of the kinetic friction coefficient in the middle section of a given forward journey) - (average value of 100 plots of the kinetic friction coefficient in the late section of the same forward journey).

[0048] As mentioned above, in this embodiment, data for the 12 items described above is collected during 20 round trips of the contactor, and only the data from the 1st, 2nd, 5th, and 20th round trips are used. Therefore, the number of explanatory variables subjected to principal component analysis is 48 (12 items × 4 round trips). When these explanatory variables are subjected to principal component analysis, they are reduced to lower-dimensional principal components (composite variables). Of the generated principal components, those with higher contribution rates are used as training data for machine learning. For example, principal components can be extracted in order of highest contribution rate so that the cumulative contribution rate is 90% or higher. There is no particular restriction on the number of principal components used as training data, but it is preferable that it be 10 or less, and particularly preferable that it be 8 or less or 6 or less.

[0049] [3. Classification device] In this embodiment, the classification device 20 comprises a trained model 21 and a principal component analysis unit 22, as shown in Figure 1. The classification device 20 is configured to output output data 24, which includes an estimated label for the classification of a sample, by inputting input data 23, which is a set of friction characteristic values ​​and viscosity for a sample of a skincare formulation, into the trained model 21 and the principal component analysis unit 22.

[0050] The trained model 21 is generated by the learning device 10 described above and is used in the classification process of skincare formulations by the classification device 20. The principal component analysis unit 22 of the classification device 20 performs the same processing as the principal component analysis unit 12 of the learning device 10 described above. That is, the principal component analysis unit 22 of the classification device 20 obtains one or more principal components that summarize these by performing principal component analysis on multiple friction characteristic values ​​included in the input data 23. In this embodiment, as described above, many values ​​such as the static friction coefficient, dynamic friction coefficient, fluctuation value of the dynamic friction coefficient, and difference value between the static friction coefficient and the dynamic friction coefficient are used as friction characteristic values ​​of the formulation. The trained model 21 is input with the viscosity of a sample of a certain skincare formulation and the principal components obtained from multiple friction characteristic values ​​related to that sample. Since this trained model 21 is trained using viscosity, principal components obtained from friction characteristic values, and correct labels related to the classification of samples for a large number of skincare formulation samples as training data, when viscosity and principal components are input, it outputs the corresponding estimated label. In this way, it is possible to configure a skincare formulation classification device 20 using the trained model 21 generated by the learning device 10.

[0051] [4. Actual Calculation Example] For 43 types of skincare formulations, frictional properties and viscosity (JIS Z 8803) were measured for 48 values ​​(12 items x 4 forward passes) according to the example described above. Furthermore, the frictional properties and viscosity of each of the 43 formulations were subjected to hierarchical cluster analysis and classified into five groups. These groups were used as the correct labels. Of the 43 formulations, 31 (70%) were used as training data, and the remaining 12 (30%) were used as test data.

[0052] First, principal component analysis was performed on the friction characteristic values ​​(48 values) of 31 types of formulations that served as training data. The cumulative contribution rate of the top 1st to 6th principal components was 91%, so these 1st to 6th principal components were adopted as training data. For the 31 types of skincare formulations, machine learning (SVM) was performed to train a model using the 1st to 6th principal components, viscosity, and the ground truth labels obtained from cluster analysis as training data, and a trained model was generated.

[0053] On the other hand, the frictional properties of 12 types of formulations (30%) that served as test data were substituted into the top 1st to 6th principal components calculated from the training data. Then, the principal components and viscosity of each formulation in the test data were input into the pre-trained model described above, and estimated labels were output. The results were then verified to see if the estimated labels of the test data output matched the correct labels classified by prior cluster analysis. As a result, it was confirmed that 10 out of the estimated labels of the 12 formulations (approximately 83%) matched the correct labels. This confirmed that the pre-trained model generated by the present invention has sufficiently practical accuracy. It is also presumed that the accuracy of the pre-trained model will improve further if the training data is increased.

[0054] [5. Computer Configuration] Figure 3 is a block diagram showing an example of the hardware configuration of computer 100. Computer 100 includes a processor 110 such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), main memory 120, storage 130, and an input / output interface 140. The learning device 10 and classification device 20 mentioned above can both be implemented in computer 100 with this configuration.

[0055] Specifically, programs for the learning device 10 and / or the classification device 20 are stored in the storage 130. The storage 130 may be an HDD or SSD installed inside the computer 100, or it may be an external storage device (HDD, SSD, CD-ROM, etc.) connected via the interface 140. The processor 110 reads each program from the storage 130, loads it into the main memory 120, and executes processing according to the program. The processor 110 also allocates memory space in the main memory 120 necessary for the execution of each process in the learning device 10 and the classification device 20, as described above, according to each program. Information input to and output from the processor 110 are performed via the interface 140. For example, input devices may be connected to the computer 100 via the interface 140. Also, display devices, communication devices, and printing devices may be connected to the computer 100 via this interface 140.

[0056] When the aforementioned learning device 10 (see Figure 1) is implemented in this computer 100, the program for executing the operations of the principal component analysis unit 12 and the learning unit 13 is mainly stored in the storage 130. The dataset 11 used as training data for machine learning is also stored in this storage 130. The processor 110 reads the program from the storage 130 and loads it into the main memory 120, and while reading the data stored in the storage 130, executes processing according to the program. The processor 110 also executes processing according to the program while temporarily storing data generated during processing in the main memory 120 or writing it to the storage 130.

[0057] Similarly, when the aforementioned classification device 20 (see Figure 1) is implemented in this computer 100, the program for executing the operation of the trained model 21 and the principal component analysis unit 22 is stored in the storage 130. The input data 23 is also stored in this storage 130. The processor 110 reads the program from the storage 130 and expands it into the main memory 120, and while reading the data stored in the storage 130, executes processing according to the program. The processor 110 also executes processing according to the program, temporarily storing data generated during processing or writing it to the storage 130. Finally, the processor 110 writes the output data 24 obtained as a result of the processing to the storage 130 or outputs it to the outside via the interface 140.

[0058] In this specification, embodiments of the present invention have been described with reference to the drawings in order to express the content of the present invention. However, the present invention is not limited to the above embodiments, and includes modifications and improvements that are obvious to those skilled in the art based on the matters described in this specification. [Explanation of symbols]

[0059] 10…Learning device 11…Dataset 12…Principal Component Analysis Department 13…Learning Department 20...Classification device 21...Trained model 22...Principal component analysis unit 23...Input data 24…Output data 30…Friction measuring device 31…Artificial skin substrate (first contact) 32…Fingerprint mimicry contact (second contact) 33...Load cell 34...Weight 35...Heating plate S...Sample 100... Computer 110... Processor 120...Main memory 130...Storage 140… Interface

Claims

1. A method for generating a trained model for classifying the dosage forms of skin care formulations, A step of determining one or more principal components by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts coated on the contact interface of a sample of a skincare formulation, The process includes generating a trained model by performing machine learning using multiple datasets, including the aforementioned main components, the viscosity of the sample, and the classification of the dosage form of the skin care formulation, as training data. Method for generating a pre-trained model.

2. The frictional force of the two contacts was measured by repeatedly sliding at least one of the two contacts within the area where the sample was applied. The friction characteristic value includes the static friction coefficient, the dynamic friction coefficient, and the variation in the dynamic friction coefficient during the sliding motion from the time the dynamic friction force is generated to a predetermined time. A method for generating a trained model according to claim 1.

3. The aforementioned friction characteristic value further includes the difference between the static friction coefficient and the kinetic friction coefficient. A method for generating a trained model according to claim 2.

4. The aforementioned sliding motion is a reciprocating motion. The friction characteristic value further includes, when the forward or return process is divided into multiple sections, the average value or variation of the kinetic friction coefficient in each section. A method for generating a trained model according to claim 2.

5. The friction characteristic value further includes the change in the average value of the kinetic friction force in the preceding and succeeding sections. A method for generating a trained model according to claim 4.

6. A learning device for training a model for classifying the dosage forms of skin care formulations, A principal component analysis unit that determines one or more principal components by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts coated on the contact interface of a sample of a skincare formulation, The system includes a learning unit that trains the model by performing machine learning using multiple datasets, including the main components, the viscosity of the sample, and the classification of the dosage form of the skin care formulation, as training data. Learning device.

7. A program for causing a computer to function as the learning device described in claim 6.

8. A classification method for classifying skin care formulations using the trained model obtained by the method of claim 1, A step of determining one or more principal components by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts coated on the contact interface of a sample of a skincare formulation, The process includes an inference step in which the main components and the viscosity of the sample are input into the trained model to obtain information regarding the classification of the sample. Classification method.

9. A classification device for classifying skincare formulations, A principal component analysis unit that determines one or more principal components by performing principal component analysis on multiple friction characteristic values ​​obtained from the frictional force of two contacts coated on the contact interface of a sample of a skincare formulation, The trained model obtained by the method of claim 1 takes the main component and the viscosity of the sample as input and outputs information regarding the classification of the sample. Classification device.

10. A program for causing a computer to function as the classification device described in claim 9.

Images