Method for generating texture prediction model, method for predicting texture, texture prediction model generation system, texture prediction system, and program

Abstract

To provide a texture prediction technique which can be realized by a simple measurement method.SOLUTION: A method for generating a texture prediction model includes the steps of: acquiring teacher data D1, of which input is a plurality of measurement units [F1, ..., F780] (or F1, ..., F180) showing change of force received from food when the food is pressed and of which output is food texture evaluation values [X1 to X85]; and generating a texture prediction model 20 for outputting the texture evaluation values [X1 to X85] by machine learning using the teacher data D1.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a food texture prediction technology using machine learning. [Background technology]

[0002] Food texture is defined by sensory evaluation based on human sensory characteristics. Sensory evaluation has the advantage of directly quantifying human sensations, but to achieve highly reproducible evaluations, it is necessary to train a large number of trained evaluators (panels), which is costly and time-consuming.

[0003] In response to this, research and development is being conducted on evaluating food texture through physical measurements using equipment. Patent Document 1 discloses that it is considered difficult in principle to evaluate human food texture by measuring a single-axis reaction force when food is pressed. Patent Document 1 also discloses that a pressure distribution image showing the pressure distribution received from food when the food is pressed by a measuring device is acquired, and a convolutional neural network is trained to output a food texture evaluation value based on the acquired pressure distribution image. Non-Patent Document 1 also mentions the use of machine learning technology to distinguish foods using the force measured for foods that break. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-207123 A [Non-patent literature]

[0005] [Non-Patent Document 1] Shunsuke Yoshida, Makoto Takemasa: Food texture analysis based on machine learning: The 34th Annual Conference of the Japanese Society for Artificial Intelligence, 2020 Summary of the Invention [Problem to be solved by the invention]

[0006] However, the method described in Patent Document 1 requires measuring the pressure distribution received from the food when the food is pressed, and there is a demand for a technology that enables prediction of food texture using a simpler measurement method.

[0007] The present disclosure provides a food texture prediction technology that can be implemented using a simple measurement method. [Means for solving the problem]

[0008] The method for generating a texture prediction model disclosed herein includes the steps of acquiring training data which takes as input a number of measurement values ​​indicating changes in force exerted by a food when the food is pressed and outputs a texture evaluation value, and performing machine learning using the training data to generate a texture prediction model which outputs a texture evaluation value. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a block diagram showing a food texture prediction model generation system 1 and a food texture prediction system 2 according to a first embodiment. [Diagram 2] FIG. 2 is a schematic explanatory diagram showing the configuration of a measurement device 3. [Diagram 3] FIG. 1 is an explanatory diagram regarding 85 types of food texture evaluation values ​​[X1, X2, ..., X85] used in the first embodiment, showing a portion selected from the 85 types. [Figure 4] 1 is a flowchart showing the processing executed by the food texture prediction model generation system 1. [Diagram 5]1 is a flowchart showing the processing executed by the food texture prediction system 2. [Figure 6] This is a plot of multiple load values ​​[F1, F2, ..., F780] measured by measuring device 3 for some foods (tofu) out of the multiple foods used to generate the food texture prediction model 20. [Figure 7] FIG. 13 is a diagram showing a comparison between the top five sensory evaluation scores for four foods and the food texture evaluation scores (estimated values) obtained by the food texture prediction model 20. [Figure 8] FIG. 13 is a diagram showing time-series waveforms of weight values ​​used in predicting the food texture evaluation scores of the above four foods. [Figure 9] FIG. 13 is a diagram showing waveforms of a distortion series of weight values ​​used in predicting the food texture evaluation scores of the above four foods. [Figure 10] FIG. 13 is a diagram showing a comparison between the top five sensory evaluation scores for four foods and the food texture evaluation scores (estimated values) obtained by the food texture prediction model 20 using the distorted series waveform data of the load value. [Figure 11] FIG. 13 shows the coefficient of determination between the food texture evaluation value (estimated value) obtained by the food texture prediction model 20 based on the waveform data of the force-strain series and the sensory evaluation value of 36 foods, and the coefficient of determination between the food texture evaluation value (estimated value) obtained by the food texture prediction model 20 based on the waveform data of the force time series and the sensory evaluation value of 36 foods. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] [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 food texture prediction model generation system 1 and a food texture prediction system 2 of the first embodiment. Fig. 2 is a schematic explanatory diagram showing the configuration of a measurement device 3.

[0011] <Texture prediction model generation system 1, Texture prediction system 2> As shown in Fig. 1, the food texture prediction model generation system 1 uses teacher data D1 to construct a food texture prediction model 20 through machine learning. The food texture prediction system 2 predicts a food texture evaluation value (estimated value) based on force change data G1 (details will be described later; sometimes referred to as force waveform data) measured when the food to be predicted is pressed uniaxially. The food texture prediction system 2 uses the food texture prediction model 20 that was generated in advance by the food texture prediction model generation system 1 through machine learning.

[0012] The food texture prediction system 2 has a prediction target measurement value acquisition unit 21 and a prediction unit 22. The food texture prediction model generation system 1 has a teacher data acquisition unit 10, teacher data D1, and a learning unit 11. The teacher data D1 is stored in a memory 1a. The teacher data acquisition unit 10 and the learning unit 11 are realized by a processor 1b. The prediction target measurement value acquisition unit 21 and the prediction unit 22 are realized by a processor 2b. In this embodiment, the processors 1b and 2b in one device realize each unit, but this is not limited to this. For example, each process may be distributed using a network, and multiple processors 1b and 2b may execute the process of each unit. In other words, one or multiple processors 1b and 2b execute the process.

[0013] The prediction target measurement value acquisition unit 21 acquires force change data G1 measured when the food to be predicted is pressed in one axis. The force change data G1 can be obtained by a measurement device 3 having a sensor 30 and a presser 31, which are shown in FIG. 2. The presser 31 is configured to be capable of a pressing operation in which the food S is compressed in a direction UD1 (pressing direction, downward), and a return operation in which the presser 31 is compressed a predetermined distance and then returned in a direction UD2 (opposite to the pressing direction, upward) from pressing the food S. The presser 31 is associated with a sensor 30 such as a load cell, and is capable of detecting the force (load value [N] or pressure value [Pa]) that the presser 31 receives from the food S during the pressing operation and the return operation of the presser 31. Thus, the force change data G1 measured by the sensor 30 is composed of a plurality of measurement values ​​indicating the change in the force received from the food S when the food S is pressed. In this embodiment, the plurality of measurement values ​​indicate the change in force at one measurement point. One measurement point may be one sensor that measures the force applied from the food S, or may include combining the measured values ​​of multiple sensors into one measured value through statistical processing such as averaging. In this embodiment, the multiple measured values ​​are values ​​measured every time a predetermined time has elapsed, and are values ​​that indicate the magnitude of the force every time a predetermined time has elapsed. The multiple measurement values ​​included in the teacher data D1 described below can be acquired in a manner similar to the force change data G1 acquired by the prediction target measurement value acquisition unit 21 described above.

[0014] FIG. 6 shows a plurality of load values ​​[F1, F2, ..., F] measured by the measuring device 3 for a part of the foods (tofu) used in generating the food texture prediction model 20. 780] is a plot of the time elapsed and the load value. The horizontal axis indicates the elapsed time, and the vertical axis indicates the load value. The positive load value is a value measured when the presser 31 presses the food S, and the negative load value is a value measured when the presser 31 returns from pressing the food S after pressing. The multiple measured values ​​(load values ​​or pressure values) in this embodiment include values ​​measured when the presser 31 returns from pressing the food, but are not limited to this as long as they include values ​​measured when the presser 31 presses the food S. Note that the positive and negative values ​​of the measured values ​​can be interchanged depending on the detection direction of the sensor 30, so the value measured when the presser 31 presses the food S may be a negative value, and the value measured when the presser 31 returns from pressing the food S may be a positive value.

[0015] Figure 6 shows the waveforms indicating the time series changes in the load values ​​when the same food is measured five times. Even for the same food, the waveform data is different each time it is measured, but it is believed that the characteristics of each food are reflected in the waveforms.

[0016] As shown in FIG. 1, the food texture prediction model 20 constituting the prediction unit 22 uses multiple measured values ​​[F1, F2, ..., F 780 ] and the texture evaluation score [X1, X2, …, X 85 ] and multiple measured values ​​F i are entered as explanatory variables, and the texture evaluation values ​​[X1, X2, …, X 85 ] as a response variable. 85 In this embodiment, there are 85 food texture evaluation values ​​as shown in FIG. 85 ] can be expressed as X jIt can also be expressed as (j=1 to M, M=85). Since one model is required for one texture evaluation value, the texture prediction model 20 in this embodiment includes 85 models, which is equal to the number of texture evaluation items to be evaluated. That is, 780 measured values ​​are input to one model, and one texture evaluation value is output. In this embodiment, multiple texture evaluation values ​​can be estimated based on one force waveform data. In this embodiment, the model uses KNIME's Gradient Boosted Decision Tree, but is not limited to this as long as it is a supervised machine learning model. For example, various models such as linear regression, regression tree, random forest, support vector machine, neural network, ensemble, etc. can be used.

[0017] Texture evaluation value [X1, X2, …, X 85 ] is 85 as shown in Figure 3. The food texture evaluation value [X1, X2, ..., X 85 ] is obtained as an evaluation value for each item expressing texture based on a sensory evaluation. It is preferable to use onomatopoeic terms as the items expressing texture. Onomatopoeic terms are onomatopoeic words or mimetic words, and can be selected by extracting terms that consist of two repeated sounds, such as "purupuru," "purun," or "purut," or terms that consist of two repeated sounds and auxiliary sounds, such as "n, ri, t," from terms used to express texture. In addition, the items expressing texture used in the above evaluation can be aggregated by grouping together similar items. The texture evaluation value is scored, for example, on a six-point scale from 0 to 5 (where "0" means no sensation and "5" means very strong sensation). The texture evaluation value may be the average of the scores of multiple subjects, but since there is a risk of individual differences depending on the subjects (panel), it is preferable to use the texture evaluation values ​​[X1, X2, ..., X 85 It is preferable to perform a normalization process (a standardization process that standardizes the average value and variance) on [] to generate a Z-score. This makes it possible to compare food texture evaluation values ​​from multiple people, reducing or eliminating individual differences, and generating and acquiring preferable teacher data D1.

[0018] The training data acquisition unit 10 acquires a plurality of measurement values ​​[F1, F2, ..., F 780 ] and obtain the food texture evaluation value [X1, X2, …, X 85 ], and acquires teacher data D1 associating the two, and stores it in the memory 1a. In this embodiment, the teacher data acquisition unit 10 generates and acquires the teacher data D1, but the teacher data acquisition unit 10 may not generate the teacher data D1, and may acquire teacher data D1 that has already been generated externally. The learning unit 11 trains the food texture prediction model 20 using the teacher data D1.

[0019] <Method of generating texture prediction model> Fig. 4 is a flowchart showing the processing executed by the food texture prediction model generation system 1. As shown in Fig. 4, in step ST1, the teacher data acquisition unit 10 acquires teacher data D1 that inputs a plurality of measurement values ​​indicating changes in force received from food when the food is pressed and outputs a food texture evaluation value. In step ST2, the learning unit 11 performs machine learning using the teacher data D1 to generate a food texture prediction model 20 that outputs a food texture evaluation value.

[0020] <Method for predicting texture> Fig. 5 is a flowchart showing the processing executed by the food texture prediction system 2. As shown in Fig. 5, in step ST101, the prediction target measurement value acquisition unit 21 acquires a plurality of measurement values ​​that indicate the change in force received from the food when the food to be predicted is pressed. In step ST102, the prediction unit 22 inputs the acquired plurality of measurement values ​​to the food texture prediction model 20, and outputs a food texture evaluation value (estimated value).

[0021] [Effects of this disclosure] Example 1 In order to demonstrate the effects of the present disclosure, Example 1 is shown below. 21 types of commercially available gel foods were used as samples. For each food, the change in force (time series change in force) due to uniaxial pressing was measured five times (measurement of the load value each time a specified time has elapsed). In addition, a sensory evaluation was performed, and 85 types of food texture evaluation values ​​[X1 to X 85In the sensory evaluation, eight subjects (six males, two females, average age 36±8.5 years) evaluated the texture using 85 onomatopoeic terms (evaluation items) shown in Figure 3, and the average of the eight subjects' sensory intensities was set as the texture evaluation value ("0" is not felt and "5" is felt very strongly). The food samples were kept at 5°C and cut into pieces of approximately 5cm2, the same size as the compression test described below. 3 The force change data for the first to fourth cuts and the 85 texture evaluation values ​​[X1 to X 85 A food texture prediction model 20 was generated by machine learning using training data D1 including the above. The fifth force change data was input to the generated food texture prediction model 20 as the measurement value to be predicted, and food texture evaluation values ​​(estimated values) of 85 items were predicted for each of the 21 types of food.

[0022] Regarding the force measurements, specifically, 780 force measurements [F1, F2, …, F 780 ]. The force measurement F i (i=1 to N, N=780). The measurement time is 0.005 seconds x 780 = approximately 3.9 seconds. A texture analyzer TA-XT2i manufactured by Stable Microsystems was used as the measurement device 3. The temperature of the food sample was adjusted to 5°C. The size of the food was the same as that of the sensory evaluation test described later, and was cylindrical with a diameter of 23 mm and a height of 12 mm. The presser 31 was a stainless steel cylindrical shape with a diameter of 3 mm. The compression speed was 10 mm / s. The clearance was 1 mm. The pressing surface of the presser 31 was smaller than the food. The clearance was the distance between the tip of the presser 31 when the food sample was most compressed and the sample stage on which the food sample was placed.

[0023] Fig. 7 is a diagram showing a comparison between the top five sensory evaluation values ​​for four foods and the texture evaluation values ​​(estimated values) obtained by the food texture prediction model 20. Fig. 8 is a diagram showing the time series waveforms of the weighting values ​​used to predict the food texture evaluation values ​​of the above four foods. Although Figs. 7 and 8 only show four foods, the top five sensory evaluation values ​​for each of 21 types of food were compared with the food texture evaluation values ​​(estimated values) obtained by the food texture prediction model 20. The coefficient of determination (R2 ) is 0.968 and the RMSE (root mean square error) is 0.121. It is considered that the corresponding food texture evaluation values ​​for the top five sensory evaluation values ​​can be predicted well based on the waveform data of the load values.

[0024] From the above results, it can be seen that it is possible to predict a food texture evaluation value using force change data G1 (force waveform data, multiple measurement values) at one measurement point, rather than a pressure distribution image at multiple measurement points arranged in a matrix.

[0025] [Second embodiment] A second embodiment of the present disclosure will be described. The difference between the second embodiment and the first embodiment is the force change data G1. That is, in the first embodiment, the force time series data is a measured value of the magnitude of the force each time a predetermined time has elapsed, whereas in the second embodiment, the force distortion series data is a measured value of the magnitude of the force each time the food is distorted by a predetermined amount. The distortion is a value obtained by dividing the moving distance of the plunger (pressure piece 31) by the height (dimension along the moving direction) of the food S. Since the moving distance of the pressure piece 31 can be calculated by multiplying the speed and time, the moving time and moving distance of the pressure piece 31 moving at a constant speed can be converted into distortion. By setting the force value to the degree that the food is distorted and deformed to a certain degree, restrictions such as the requirement that the size and shape of the food be the same are eliminated. Therefore, it becomes possible to measure food in a container or food removed from a container. In addition, since it is possible to measure food in a container, it becomes possible to predict the texture of foods that cannot retain their shape.

[0026] Example 2 In order to demonstrate the effects of the present disclosure, Example 2 is shown below. 21 types of commercially available gel foods are used as samples. For each food, the force change due to uniaxial pressing was measured 10 times or more, and the time series change in force was converted into the strain change of the force. The sensory evaluation was performed in the same manner as in Example 1, as described in the above paragraph

[0021] .

[0027] Specifically, the presser 31 is a stainless steel cylinder with a diameter of 5 mm. The compression speed is 10 mm / s. The trigger load is set to 0.005 N (the time when the load of 0.005 N is detected is set as the initial position where the presser 31 touches the sample surface, and data acquisition is started). The initial position is set to 0% strain, and the height of the food S is set to 100%. The food S is pressed from the initial position of 0% strain to 90% strain, and then pulled up at the same speed (i.e., the clearance is 10%). A total of 180 measured values ​​for every 1% strain were used as explanatory variables. The measurements were repeated 10 times or more. As a specific example, a large number of force measurements taken every 0.005 seconds were converted into corresponding representative values ​​for every 1% strain. For example, the measured value for 1% strain can be a statistical value (e.g., average value) of multiple values ​​measured within the time required for the food S to deform by 1% strain.

[0028] FIG. 10 is a diagram showing a comparison between the top five sensory evaluation values ​​for the four foods and the food texture evaluation values ​​(estimated values) obtained by the food texture prediction model 20 using the waveform data of the distortion series of weight values. FIG. 9 is a diagram showing the waveform of the distortion series of weight values ​​used to predict the food texture evaluation values ​​of the above four foods. Specifically, FIG. 9 shows the waveforms of multiple weight values ​​[F1, F2, ..., F] measured by the measuring device 3 for the four foods (tofu 1, tofu 2, jelly 1, jelly 2). 180 9 and 10 show a plot of the sensory evaluation values ​​of the food S and the food texture prediction model 20. The horizontal axis shows strain, and the vertical axis shows load value. The positive load value is the value measured when the presser 31 presses the food S (indicated as "pressing" in FIG. 9), and the negative load value is the value measured when the presser 31 returns from pressing the food S after pressing (indicated as "pulling up" in FIG. 9). Although only four foods are shown in FIG. 9 and FIG. 10, the top five sensory evaluation values ​​of 21 types of foods were compared with the food texture evaluation values ​​(estimated values) obtained by the food texture prediction model 20. The coefficient of determination (R2) was 0.967, and the RMSE (root mean square error) was 0.120. It is considered that the corresponding food texture evaluation values ​​of the top five sensory evaluation values ​​can be well predicted based on the waveform data of the strain series of load values.

[0029] FIG. 11 shows the coefficient of determination between the texture evaluation value (estimated value) obtained by the food texture prediction model 20 based on the waveform data of the force distortion series and the sensory evaluation value of the 36 foods, and the coefficient of determination between the texture evaluation value (estimated value) obtained by the food texture prediction model 20 based on the waveform data of the time series of force and the sensory evaluation value of the 36 foods, when the number of foods is further increased to 36. Six foods are randomly selected from the 36 foods as unknown samples, and the remaining 30 foods are used as training data to create the texture prediction model 20, and the texture evaluation value (estimated value) is obtained from the force change data (waveform data of the load value distortion series). By repeating this six times in cross-validation, the texture evaluation value (estimated value) is obtained for all 36 foods for the number of times the instrumental measurement was repeated, and the coefficient of determination between the sensory evaluation value and the texture evaluation value (estimated value) can be obtained. It is believed that even when the change in strain due to force is used as an explanatory variable, the food texture evaluation value can be predicted well based on the waveform data of the strain series of load values. Moreover, according to FIG. 11, it can be seen that the coefficient of determination based on the force strain series data is relatively larger than the coefficient of determination based on the force time series data, and it can be understood that the prediction accuracy is higher when the force strain series data is used than when the force time series data is used.

[0030] [1] As described above, in the present embodiment, the method for generating a food texture prediction model uses a plurality of measured values ​​[F1, ..., F 780 ] (or F1, …, F 180 ) is input, and the texture evaluation value [X1~X 85 ] is obtained as an output of training data D1, and a food texture evaluation value (estimated value) [X1 to X 85 and generating a food texture prediction model 20 that outputs the food texture prediction model 20. In this way, it is possible to predict a food texture evaluation value from the force change data G1 (multiple measured values), making it possible to provide a food texture prediction technique that can be realized using a simple measurement method.

[0031] [2] The method for generating a food texture prediction model according to the above-mentioned [1], 180 ] may be a value indicating the magnitude of force at which the food S is deformed by a specified strain. This corresponds to the second embodiment. By setting the measurement value to the degree to which the food undergoes a specified distortion as in the second embodiment, restrictions such as the requirement that the size of the food be the same shape are eliminated. This makes it possible to measure food in a container or food removed from a container. In addition, since it is possible to measure food in a container, it is possible to predict the texture of foods that cannot retain their shape.

[0032] [3] The method for generating a food texture prediction model according to the above-mentioned [1], 780 ] may be a value indicating the magnitude of the force each time a predetermined time has elapsed. This corresponds to the first embodiment.

[0033] [4] The method for generating a food texture prediction model according to any one of the above [1] to [3], 780 ] (or F1, …, F 180 ) may be a value indicating the time series change of the force at one measurement point. It can be seen that it is possible to predict a food texture evaluation value using force change data G1 (waveform data of pressure values, multiple measurement values) at one measurement point, rather than a pressure distribution image at multiple measurement points arranged in a matrix.

[0034] [5] The method for generating a food texture prediction model according to any one of the above [1] to [4], 780 ] (or F1, …, F 180 ) may include a value measured when the presser 31 presses the food S and a value measured when the presser 31 is released from pressing the food S after pressing. The values ​​measured when the presser 31 presses the food S are information related to the texture evaluation value related to elasticity, hardness, and toughness, and the values ​​measured when the presser 31 is released from the food S after pressing are thought to be information related to the texture evaluation value related to adhesiveness, so it is thought that various texture evaluation values ​​can be predicted. This is particularly suitable for evaluating gel-like foods that have adhesiveness.

[0035] [6] As in this embodiment, the food texture prediction method uses a plurality of measured values ​​[F1, ..., F 780 ] (or F1, …, F 180 ) and obtaining multiple measurements [F1, …, F 780 ] (or F1, …, F 180 ) is input and the texture evaluation value [X1~X 85 ], the obtained multiple measurements [F1, ..., F 780 ] and enter the texture evaluation value (estimated value) [X1~X 85 and outputting the result. In this way, it is possible to predict a food texture evaluation value from the force change data G1 (multiple measured values), making it possible to provide a food texture prediction technique that can be realized using a simple measurement method.

[0036] [7] The method for predicting food texture according to the above [6], 180 ] may be a value indicating the magnitude of force at which the food S is deformed by a specified strain. This corresponds to the second embodiment. By setting the measurement value to the degree to which the food undergoes a specified distortion as in the second embodiment, restrictions such as the requirement that the size of the food be the same shape are eliminated. This makes it possible to measure food in a container or food removed from a container. In addition, since it is possible to measure food in a container, it is possible to predict the texture of foods that cannot retain their shape.

[0037] [8] The method for predicting food texture according to the above [6], 780 ] may be a value indicating the magnitude of the force each time a predetermined time has elapsed. This corresponds to the first embodiment.

[0038] [9] The method for predicting food texture according to any one of the above [6] to [8], 780 ] (or F1, …, F 180 ) may be a value indicating the change in force at one measurement point. It can be seen that it is possible to predict a food texture evaluation value using pressure value change data G1 (pressure value waveform data, multiple measurement values) at one measurement point, rather than a pressure distribution image at multiple measurement points arranged in a matrix.

[0039]

[10] The method for predicting food texture according to any one of the above [6] to [9], 780 ] (or F1, …, F 180 ) may include a value measured when the presser 31 presses the food S and a value measured when the presser 31 is released from pressing the food S after pressing. The values ​​measured when the presser 31 presses the food S are information related to the texture evaluation value related to elasticity, hardness, and toughness, and the values ​​measured when the presser 31 is released from the food S after pressing are thought to be information related to the texture evaluation value related to adhesiveness, so it is thought that various texture evaluation values ​​can be predicted. This is particularly suitable for evaluating gel-like foods that have adhesiveness.

[0040]

[11] As in this embodiment, the food texture prediction model generation system 1 uses a plurality of measured values ​​[F1, ..., F 780 ] (or F1, …, F 180 ) is input, and the texture evaluation value [X1~X 85A training data acquisition unit 10 acquires training data D1 that outputs the training data D1, and performs machine learning using the training data D1 to obtain a food texture evaluation value (estimated value) [X1 to X 85 and a learning unit 11 that generates a food texture prediction model 20 that outputs the food texture prediction model.

[0041]

[12] As in this embodiment, the food texture prediction system 2 uses a plurality of measured values ​​[F1, ..., F 780 ] (or F1, …, F 180 ), and a prediction target measurement value acquisition unit 21 that acquires a plurality of measurements [F1, ..., F 780 ] (or F1, …, F 180 ) is input and the texture evaluation value [X1~X 85 ], the obtained multiple measurements [F1, ..., F 780 ] and enter the texture evaluation value (estimated value) [X1~X 85 and a prediction unit 22 that outputs the prediction result.

[0042]

[13] As in this embodiment, the program may cause one or more processors 1b, 2b to execute any one of the methods described in [1] to

[10] above.

[0043] A computer-readable temporary recording medium according to the present embodiment stores the above program.

[0044] Although the embodiments of the present disclosure have been described above based on the drawings, the specific configurations 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-mentioned embodiments but also by the claims, and further includes all modifications within the meaning and scope equivalent to the claims.

[0045] The structures employed in the above-described embodiments may be employed in any other embodiment. The specific configurations of the components are not limited to the above-described embodiments, and may be modified in various ways without departing from the spirit of the present disclosure.

[0046] (A) For example, in the first embodiment described above, the number of pressure values ​​is 780, but the length of the measurement time and the sampling frequency can be varied, and various modifications are possible.

[0047] (B) In the first embodiment, the number of types of food texture evaluation values ​​is 85, but is not limited to this. If one type of food texture evaluation value is sufficient, the number of types of food texture evaluation values ​​may be one, and the number of types of food texture evaluation values ​​can also be changed in various ways.

[0048] (C) In the first embodiment, four force waveform data (multiple measurement values) for 21 types of food are used to generate the food texture prediction model 20, but this is not limited to this. The types of food and the number of force waveform data can be changed. The more types of food and the more number of force waveform data, the more preferable it is from the viewpoint of improving prediction accuracy.

[0049] (D) In ​​measurements using the above-mentioned measuring device 3, the shape and size of the presser 31, the moving speed of the presser 31, the size of the food sample, the temperature of the sample, and the like are not limited to those mentioned above and can be changed as appropriate. The method of sensory evaluation of food texture evaluation values ​​is not limited to the above. For example, the number of evaluation values ​​is not limited to six levels, and the number of levels may be changed. The number of evaluators may also be changed as appropriate.

[0050] (F) In the above embodiment, a gel-like food product is given as an example, however, the food texture estimation technology of the present disclosure is not limited to gel-like foods and can be applied to a wide range of foods. [Explanation of symbols]

[0051] 1: Texture prediction model generation system 1b, 2b: Processor 2: Texture prediction system 10: Teacher data acquisition section 11: Learning Department 20: Texture prediction model 21: Prediction target measurement value acquisition unit 22: Prediction section 31: Presser D1: Training data F1~F 780 : Measurement value (time series change data of load value) F1~F 180 : Measurement value (load value strain series change data) S:Food X1~X 85 : Texture evaluation value

Claims

1. acquiring training data having as input a plurality of measured values ​​indicating changes in force exerted by a food product when the food product is pressed and an output of a food texture evaluation value; generating a food texture prediction model that outputs a food texture evaluation value by performing machine learning using the training data; A method for generating a food texture prediction model, comprising:

2. 2. The food texture prediction model generating method according to claim 1, wherein the plurality of measurement values ​​are values ​​indicating a magnitude of force at which the food product undergoes a predetermined distortion.

3. The food texture prediction model generating method according to claim 1 , wherein the plurality of measured values ​​are values ​​indicating the magnitude of force each time a predetermined time has elapsed.

4. The food texture prediction model generating method according to claim 1 , wherein the plurality of measured values ​​are values ​​indicating a change in force at one measurement point.

5. 4. The food texture prediction model generating method according to claim 1, wherein the plurality of measurement values ​​include values ​​measured when a presser presses the food product, and values ​​measured when the presser is released from pressing the food product after the pressing.

6. obtaining a number of measurements indicative of changes in force experienced by the food product as the food product is compressed; a step of inputting the acquired multiple measurement values ​​into a food texture prediction model trained by machine learning so as to input multiple measurement values ​​and output a food texture evaluation value, and outputting the food texture evaluation value; The method for predicting food texture comprises:

7. 7. The food texture prediction method according to claim 6, wherein the plurality of measured values ​​are values ​​indicating a magnitude of force at which the food product undergoes a specified distortion.

8. The food texture prediction method according to claim 6 , wherein the plurality of measured values ​​are values ​​indicating the magnitude of force each time a predetermined time has elapsed.

9. The food texture prediction method according to claim 6 , wherein the plurality of measured values ​​are values ​​indicating a change in force at one measurement point.

10. The food texture prediction method according to any of claims 6 to 8, wherein the plurality of measurement values ​​include values ​​measured during an action of a presser pressing the food product, and values ​​measured during an action of returning the presser from pressing the food product after the pressing.

11. a teacher data acquisition unit that acquires teacher data having as input a plurality of measurement values ​​indicating changes in force received from a food product when the food product is pressed, and as output a food texture evaluation value; a learning unit that generates a food texture prediction model that outputs a food texture evaluation value by performing machine learning using the teacher data; A food texture prediction model generation system comprising:

12. a prediction target measurement value acquisition unit that acquires a plurality of measurement values ​​that indicate a change in force received from the food when the food is pressed; a prediction unit which inputs the acquired multiple measurement values ​​into a food texture prediction model trained by machine learning so as to output a food texture evaluation value using multiple measurement values ​​as input, and outputs the food texture evaluation value; A texture prediction system comprising:

13. A program causing one or more processors to execute the method according to claim 1 or 6.

Images