Condiment evaluation device, evaluation method, and evaluation program

Abstract

To provide an evaluation device, an evaluation method and an evaluation program capable of evaluating seasonings while suppressing the influence of human subjectivity.SOLUTION: An evaluation device, an evaluation method and an evaluation program acquire data of a first spectral spectrum of a first seasoning, and calculate a first feature vector using an evaluation model that outputs a first feature vector that characterizes the first seasoning in response to an input based on the first spectral spectrum; and store a second feature vector that characterizes a second seasoning, and specify a specific second seasoning associated with the first seasoning from among the second seasonings based on the first feature vector and the second feature vector. The evaluation model is a learning model obtained by changing parameters of a learning model in a direction of increasing the variance of the second feature vector output from the learning model with respect to an input based on a second spectral spectrum of the second seasoning.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to a seasoning evaluation device, an evaluation method, and an evaluation program. [Background technology]

[0002] Patent Document 1 discloses a technology for learning the relationship between instrumental analysis data or sensory evaluation data of a plurality of foods and beverages whose quality is known and the quality of each of the plurality of foods and beverages through deep learning. According to this technology, a food and beverage quality prediction model for predicting the quality of a food and beverage whose quality is unknown is created based on the instrumental analysis data or sensory evaluation data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2018-18354 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, according to the technology disclosed in Patent Document 1, it is necessary to provide equipment analysis data or sensory evaluation data for a plurality of foods and beverages whose quality is known as teacher data. Here, when acquiring the characteristics of foods and beverages using equipment analysis data, human subjectivity is likely to be involved in the selection of the characteristics. Also, when acquiring the characteristics of foods and beverages using sensory evaluation data, human subjectivity is likely to be involved in the determination of the evaluation value. In this way, when the equipment analysis data or sensory evaluation data provided as teacher data is biased by human subjectivity, the prediction accuracy of the quality prediction model obtained after learning is deteriorated, and there has been a problem in that foods and beverages such as seasonings cannot be evaluated with high accuracy.

[0005] The present disclosure has been made in consideration of the above-mentioned circumstances. That is, an object of the present disclosure is to provide an evaluation device, an evaluation method, and an evaluation program that can accurately evaluate seasonings while suppressing the influence of human subjectivity. [Means for solving the problem]

[0006] The evaluation device according to the present disclosure includes an acquisition unit, a calculation unit, a storage unit, and an identification unit. Here, the acquisition unit acquires data of a first spectroscopic spectrum of a first seasoning, and the calculation unit calculates the first feature vector using an evaluation model that outputs a first feature vector characterizing the first seasoning for an input based on the first spectroscopic spectrum. The storage unit stores a second feature vector characterizing the second seasoning. The identification unit identifies a specific second seasoning associated with the first seasoning from among the second seasonings based on the first feature vector and the second feature vector. The evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance of the second feature vector output from the learning model for an input based on the second spectroscopic spectrum of the second seasoning.

[0007] The identification unit may calculate an index value based on the first feature vector and the second feature vector, and identify a second seasoning having a second feature vector whose index value is greater than a predetermined threshold value as a identified second seasoning.

[0008] The index value may be larger as the Euclidean length of a difference vector obtained by subtracting the second feature vector from the first feature vector is smaller.

[0009] The index value may be a cosine similarity between the first feature vector and the second feature vector.

[0010] The identification unit may identify a second seasoning having a second feature vector with a large index value as the identified second seasoning in preference to a second seasoning having a second feature vector with a small index value.

[0011] The evaluation model may output a first feature vector in response to an input based on a chemical analysis value of the first seasoning.

[0012] The evaluation model may output a first feature vector in response to an input based on the optical spectrum of the raw materials of the first seasoning.

[0013] The evaluation model may be a learning model obtained by changing parameters in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the chemical analysis value of the second seasoning.

[0014] The evaluation model may be a learning model obtained by changing parameters in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the optical spectrum of the raw materials of the second seasoning.

[0015] The evaluation model may be represented by a neural network or a support vector machine.

[0016] The evaluation model may output a first feature vector in response to an input obtained by performing principal component analysis or factor analysis on the first spectrum.

[0017] The evaluation model may be a learning model obtained by varying a parameter until the amount of change in variance per unit change in the parameter becomes equal to or less than a predetermined amount.

[0018] The evaluation device may further include a presentation unit that presents to a user the first feature vector and the second feature vector characterizing the specified second seasoning.

[0019] The evaluation method according to the present disclosure acquires data on a first spectral spectrum of a first seasoning, and calculates a first feature vector using an evaluation model that outputs a first feature vector characterizing the first seasoning in response to an input based on the first spectral spectrum. A second feature vector characterizing a second seasoning is stored, and a specific second seasoning associated with the first seasoning is identified from among the second seasonings based on the first feature vector and the second feature vector. The evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the second spectral spectrum of the second seasoning.

[0020] The evaluation program according to the present disclosure causes a computer to execute a step of acquiring data of a first spectral spectrum of a first seasoning. Then, the computer executes a step of calculating a first feature vector using an evaluation model that outputs a first feature vector characterizing the first seasoning for an input based on the first spectral spectrum. The evaluation model further executes a step of storing a second feature vector characterizing a second seasoning, and a step of identifying a specific second seasoning associated with the first seasoning from among the second seasonings based on the first feature vector and the second feature vector. The evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance of the second feature vector output from the learning model for an input based on the second spectral spectrum of the second seasoning. Effect of the Invention

[0021] According to the present disclosure, seasonings can be evaluated with high accuracy while suppressing the influence of human subjectivity. [Brief description of the drawings]

[0022] [Figure 1] 1 is a conceptual diagram of an evaluation system including an evaluation device according to an embodiment of the present disclosure. [Diagram 2] FIG. 2 is a block diagram showing a configuration of an evaluation device. [Diagram 3]13 is a flowchart showing the procedure of a learning process of the evaluation device. [Figure 4] 10 is a flowchart showing a procedure of an evaluation process of the evaluation device. [Diagram 5] FIG. 13 is a diagram showing an example of feature vectors represented by a radar chart. [Figure 6] FIG. 13 is a diagram showing an example in which the characteristics of a sample are expressed by points on a two-dimensional plane. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] Hereinafter, some exemplary embodiments will be described with reference to the drawings. In addition, common parts in each drawing are given the same reference numerals, and duplicated explanations will be omitted.

[0024] [Evaluation system configuration] Fig. 1 is a conceptual diagram of an evaluation system including an evaluation device according to an embodiment of the present disclosure. As shown in Fig. 1, the evaluation system includes a light source 5, a spectroscopic device 10, an evaluation device 20, and a sample mass 40. The spectroscopic device 10 is connected to the evaluation device 20 via a wireless or wired network.

[0025] The sample mass 40 holds a sample SP (a seasoning to be evaluated). A light source 5 is provided to irradiate the sample mass 40 with incident light R1, and a spectrometer 10 is provided to separate scattered light R2 scattered from the sample SP upon receiving the incident light R1.

[0026] An optical fiber may be provided between the light source 5 and the sample SP in order to guide the incident light R1 from the light source 5 to the sample SP. Also, an optical fiber may be provided between the sample SP and the spectroscopic device 10 in order to guide the scattered light R2 from the sample SP to the spectroscopic device 10. In the following description, the spectroscopic device 10 will be described as dispersing the scattered light R2 from the sample SP, but instead of the scattered light R2, the spectroscopic device 10 may disperse the transmitted light after the incident light R1 passes through the sample SP.

[0027] The spectroscopic device 10 separates scattered light R2 from a sample SP (first seasoning, second seasoning) illuminated by incident light R1 from a light source 5, and measures the light intensity for each wavelength of light contained in the scattered light R2. For example, the light source 5 irradiates the sample SP with near-infrared light (light containing continuous wavelength components with wavelengths of approximately 800 nm to 2600 nm) as the incident light R1.

[0028] The spectroscopic device 10 then calculates the common logarithm of the value obtained by dividing the light intensity detected for each wavelength by the reference light intensity (for example, the intensity of the incident light R1 for each wavelength), and gives a negative sign to the calculation result to make the case where there is absorption a positive value, thereby obtaining the absorbance for each wavelength. Data in which the absorbance is arranged for each wavelength is called a spectroscopic spectrum. The spectroscopic spectrum is transmitted from the spectroscopic device 10 to the evaluation device 20.

[0029] The light emitted by the light source 5 and the light dispersed by the spectroscopic device 10 are not limited to near-infrared light, but may be, for example, ultraviolet light, visible light, infrared light, submillimeter waves, millimeter waves, microwaves, or the like.

[0030] The evaluation device 20 acquires the spectrum obtained by the spectroscopic device 10, and determines a sample (registered sample, i.e., second seasoning) that is similar to the sample SP and has been registered in the database based on the acquired spectrum. The evaluation device 20 will be described in detail later.

[0031] The reason why the evaluation device 20 determines a registered sample similar to the sample SP based on the spectroscopic spectrum is as follows.

[0032] When a manufacturer develops a seasoning product, a registered sample similar to sample SP is selected based on the designer's experience. The registered sample similar to sample SP is used as the starting point for product development. However, depending on the designer's level of skill, an inappropriate registered sample may be selected, which may result in more trial and error steps in product development. If there are more trial and error steps in product development, the time required to complete the product will increase.

[0033] In addition, one method for objectively identifying the taste of a seasoning is to use sensor technology that mimics human taste cues such as sweetness, sourness, saltiness, bitterness, and umami, but there are some seasoning characteristics that cannot be identified using these sensor technologies. For example, it can be difficult to identify seasoning characteristics such as initial taste, middle taste, aftertaste, whether full-bodied, and whether sharpness is present or not using sensor technology.

[0034] Therefore, it is necessary to determine registered samples similar to the sample SP using a method that is as unlikely to be influenced by human subjectivity or taste as possible. For this reason, the evaluation device 20 determines registered samples similar to the sample SP based on the spectroscopic spectrum.

[0035] [Evaluation device configuration] Fig. 2 is a block diagram showing the configuration of the evaluation device. As shown in Fig. 2, the evaluation device 20 includes an acquisition unit 21, a database 23 (storage unit), a controller 25, an operation unit 27, and a presentation unit 29. The controller 25 is connected to the acquisition unit 21, the database 23, the operation unit 27, and the presentation unit 29 so as to be able to communicate with them.

[0036] In addition, the operation unit 27 and the presentation unit 29 may be provided in the evaluation device 20 itself, or may be installed outside the evaluation device 20 and connected to the evaluation device 20.

[0037] The acquiring unit 21 is connected to the spectroscopic device 10 wirelessly or by wire so as to be able to communicate with the spectroscopic device 10. The acquiring unit 21 acquires data of the spectroscopic spectrum (first spectroscopic spectrum, second spectroscopic spectrum) of the sample SP (first seasoning, second seasoning) from the spectroscopic device 10. The acquiring unit 21 may also acquire identification information for identifying the sample SP together with the spectroscopic data. The acquiring unit 21 may also acquire a timestamp indicating the date and time when the spectroscopic spectrum was acquired together with the spectroscopic data.

[0038] Alternatively, the acquiring unit 21 may acquire the spectral data from a storage medium or the like that stores the spectral data. In this case, the spectral data stored in the storage medium or the like may be measured using a spectroscopic device 10 that is not connected to the evaluation device 20.

[0039] The database 23 (storage unit) stores an evaluation model set by the evaluation model setting unit 253 described later. The database 23 also stores acquired spectral data. Furthermore, the database 23 stores feature vectors (first feature vector, second feature vector) characterizing the sample SP (first seasoning, second seasoning). Here, the feature vector characterizing the sample SP may be calculated using the evaluation model based on the spectral spectrum. Each component of the feature vector characterizing the sample SP may be calculated by linearly combining the absorbance at multiple wavelengths.

[0040] The database 23 may store the optical spectrum (second optical spectrum) of the raw material of the second seasoning and the chemical analysis value of the second seasoning. For example, the chemical analysis value is Brix value, sugar content, total acid value, salinity value, acidity, viscosity, etc. The chemical analysis value is not limited to the examples given here.

[0041] The database 23 may store the similarity of the ingredients of the second seasoning. Here, the similarity of the ingredients of the second seasoning represents an index value calculated based on a feature vector based on the optical spectrum of a predetermined ingredient and a feature vector based on the optical spectrum of the ingredients of the second seasoning. As the predetermined ingredients, ingredients that are generally available in the product development of seasonings are selected in advance.

[0042] The index value between feature vectors may be set to be larger as the absolute value of the difference between the feature vectors is smaller, or may be set to be larger as the cosine similarity between the feature vectors is larger. The feature vector may be calculated using an evaluation model based on the optical spectrum. Each component of the feature vector may be calculated by linearly combining absorbances at multiple wavelengths.

[0043] Additionally, the database 23 may store identification information for identifying the first seasoning or the second seasoning, a time stamp, and the like, which are acquired together with the spectroscopic data.

[0044] The operation unit 27 is an input device that can be operated by a user, such as a monitor or maintenance person of the evaluation system. For example, the operation unit 27 is a keyboard, a mouse, a trackball, a touch panel, or the like. The operation unit 27 is not limited to the examples given here. The content of the user's operation input via the operation unit 27 is transmitted to the controller 25.

[0045] The presentation unit 29 displays the information received from the controller 25. Furthermore, the presentation unit 29 may present the second seasoning (specified second seasoning) specified by the controller 25 to the user. For example, the presentation unit 29 may be a display that displays figures and characters by combining a plurality of display pixels, or may be a rotating light, a buzzer, or the like. The presentation unit 29 is not limited to the examples given here.

[0046] Furthermore, when multiple specific second seasonings are identified, the presentation unit 29 may present the specific second seasonings to the user in descending order of priority based on the priority set for each specific second seasoning. For example, the priority may be set to a larger value as the index value determined corresponding to the first seasoning and the specific second seasoning is larger.

[0047] Alternatively, the presentation unit 29 may present to the user a first feature vector characterizing the first seasoning and a second feature vector characterizing the specified second seasoning. For example, the presentation unit 29 may present the first feature vector and the second feature vector by a radar chart having an axis representing the magnitude of each component of the feature vector. Fig. 5 shows how the magnitudes of each component C1, C2,..., C10 of a 10-dimensional feature vector are represented by a polygonal shape PS.

[0048] Furthermore, the presenting unit 29 may map the first feature vector and the second feature vector to points on a two-dimensional plane using a predetermined mapping relationship, and present the first feature vector and the second feature vector as a set of points on the two-dimensional plane. The presenting unit 29 may use a mapping relationship in which the larger the index value determined corresponding to the sample SP and the registered sample, the closer the mapped points are located. For example, the mapping relationship may be one in which the value obtained by linearly combining the components of the feature vectors is set as the axis value of the two-dimensional plane.

[0049] 6 shows how the features of samples A to G are expressed by points mapped onto a two-dimensional plane (Z1-Z2 plane). It can be seen that points close to each other on the Z1-Z2 plane have large index values ​​for the feature vectors corresponding to the points.

[0050] The controller 25 (control unit) is a general-purpose microcomputer equipped with a CPU (Central Processing Unit), a memory, and an input / output unit. A computer program (evaluation program) for functioning as the evaluation device 20 is installed in the controller 25. By executing the computer program, the controller 25 functions as a plurality of information processing circuits (251, 253, 255, 257) included in the evaluation device 20. The computer program (evaluation program) may be stored in a storage medium that can be read and written by a computer.

[0051] In this disclosure, an example is shown in which multiple information processing circuits (251, 253, 255, 257) are realized by software. However, it is also possible to configure the information processing circuits (251, 253, 255, 257) by preparing dedicated hardware for executing each information process shown below. Also, the multiple information processing circuits (251, 253, 255, 257) may be configured by individual hardware. Furthermore, the information processing circuits (251, 253, 255, 257) may also be used as a control unit used for monitoring or controlling the spectroscopic device 10.

[0052] As shown in FIG. 2, the controller 25 includes a preprocessing unit 251, an evaluation model setting unit 253, a calculation unit 255, and an identification unit 257 as a plurality of information processing circuits (251, 253, 255, 257).

[0053] The preprocessing unit 251 performs preprocessing on the spectrum acquired by the spectroscopic device 10. Specifically, in order to clarify the characteristics of the absorption peaks due to the components contained in the sample SP, the preprocessing unit 251 performs smoothing processing using a moving average, first-order differential processing, second-order differential processing, SNV (Standard Normal Variate) conversion processing, and the like.

[0054] Smoothing is used to reduce noise by replacing the absorbance at each wavelength in the spectrum with the average of the absorbance at the wavelengths before and after it. The more absorbance values ​​at the wavelengths before and after it are used for smoothing, the greater the noise reduction effect.

[0055] First derivative processing is used to remove the effects of baselines. Second derivative processing is used to remove the effects of baseline changes that can be expressed as a linear function of wavelength, and to make the differences in absorption peaks clearer. When second derivative processing is performed on the wavelength of a spectrum, the relative relationship between the peak positions and peak heights is maintained, making it possible to perform quantitative analysis based on the spectrum after second derivative processing.

[0056] SNV (Standard Normal Variate) conversion is used to suppress baseline shift in the spectrum. The spectrum is treated as a set of absorbance values, and the spectrum is converted so that the average of the set of absorbance values ​​is 0 and the variance is 1.

[0057] In addition, the pre-processing unit 251 may remove outliers using a Hampel filter in order to remove spike noise contained in the spectrum, or may perform linear interpolation to fill in missing parts of the data.

[0058] Furthermore, the preprocessing unit 251 may perform preprocessing on the spectrum using multiple scattering correction processing, offset processing, trend removal, resampling, FFT (Fast Fourier Transform), Wavelet transformation, and the like.

[0059] Additionally, the preprocessing unit 251 may perform principal component analysis or factor analysis on the spectral data to reduce the amount of spectral data.

[0060] The spectroscopic data after preprocessing by the preprocessing unit 251 may be stored in the database 23.

[0061] The evaluation model setting unit 253 adjusts a learning model by performing machine learning based on the data of the second spectral spectrum stored in the database 23, and sets the adjusted learning model as an evaluation model. In addition, the evaluation model setting unit 253 stores the obtained evaluation model and a second feature vector calculated based on the second spectral spectrum in the database 23.

[0062] Examples of methods for creating a learning model using machine learning include neural networks, support vector machines, Random Forest, XGBoost, LightGBM, PLS regression, Ridge regression, and Lasso regression. Methods that use one or a combination of two or more of the methods listed here can also be used. Methods for creating a learning model using machine learning are not limited to the examples listed here.

[0063] For example, the evaluation model setting unit 253 reads out the data of the second optical spectrum stored in the database 23. Then, the evaluation model setting unit 253 performs preprocessing on the second optical spectrum via the preprocessing unit 251.

[0064] The evaluation model setting unit 253 inputs the second spectrum after the preprocessing to the learning model, and sets the feature amount output from the learning model as a second feature vector. Then, the evaluation model setting unit 253 calculates the variance of the obtained second feature vector. Then, in order to maximize the calculated variance, the evaluation model setting unit 253 changes the parameters of the learning model in a direction to increase the variance of the second feature vector.

[0065] In addition, when there are multiple parameters related to the learning model, the evaluation model setting unit 253 may change only some of the multiple parameters, or may change all of the multiple parameters. The amount of change in the parameter is set to an amount of a suitable magnitude for searching for the maximum value of the variance for the second feature vector.

[0066] For example, the evaluation model setting unit 253 calculates the variance for the second feature vector, then changes the parameters for the learning model, and calculates the variance for the second feature vector using the learning model after the change. Note that the magnitude of the second feature vector is normalized to, for example, 1.

[0067] A difference occurs between the variance for the second feature vector before the parameter is changed (variance before parameter change) and the variance for the second feature vector after the parameter is changed (variance after parameter change). Therefore, the amount of change obtained by subtracting the variance before parameter change from the variance after parameter change is divided by the amount of change in the parameter. The evaluation model setting unit 253 sequentially changes the parameters and updates the learning model until the amount of change in variance per unit amount of change in the parameter obtained by the division becomes equal to or less than a predetermined amount.

[0068] In addition, the evaluation model setting unit 253 may update the learning model by changing the parameters from the initial values ​​of multiple parameters set randomly in order to suppress the variance from remaining at a local maximum value different from the maximum value (local solution convergence). In other words, the learning model may start updating from the initial values ​​of multiple different parameters and check whether the same parameters are reached.

[0069] The evaluation model setting unit 253 ends the update of the learning model when the amount of change in the variance becomes equal to or less than a predetermined amount, and sets the learning model at that time as the evaluation model.

[0070] The evaluation model setting unit 253 stores in the database 23 a second feature vector obtained when the second spectrum is input to the obtained evaluation model.

[0071] Note that the second feature vector may vary depending on the change in the parameters of the learning model. If the size of the second feature vector is normalized to 1, and the feature vector space is N-dimensional, each end point of the second feature vector moves on the surface of an N-1-dimensional sphere embedded in the N-dimensional space according to the change in the parameters.

[0072] Here, since the learning model is updated and the evaluation model is set so as to maximize the variance regarding the second feature vector, each end point of the second feature vector calculated based on the evaluation model is arranged as widely and evenly as possible on the spherical surface of the N-1-dimensional sphere. In other words, it can be said that the evaluation model setting unit 253 performs a process of selecting an axis of the feature vector space such that each end point of the second feature vector is arranged as widely and evenly as possible on the spherical surface of the N-1-dimensional sphere. This reduces the possibility that the evaluation model is set biased toward one of the second seasonings stored in the database 23.

[0073] The evaluation model setting unit 253 may provide the learning model with an input based on the chemical analysis value of the second seasoning in addition to the input based on the second spectral spectrum of the second seasoning. The evaluation model setting unit 253 may also provide the learning model with an input based on the spectral spectrum of the raw materials of the second seasoning. Furthermore, the evaluation model setting unit 253 may provide the learning model with an input based on the similarity of the raw materials of the second seasoning.

[0074] The evaluation model setting unit 253 may adjust the learning model by changing parameters in a direction that increases the variance for the second feature vector output from the learning model based on the above-mentioned input.

[0075] Next, a note will be made on the parameters related to the learning model. It is assumed that the learning model is a neural network. The neural network includes an input layer to which spectral data is input, an output layer to which output values ​​are output, and at least one or more hidden layers provided between the input layer and the output layer, and signals are propagated in the order of the input layer, hidden layer, and output layer. Each of the input layer, hidden layer, and output layer is composed of one or more units. The units between the layers are connected to each other, and each unit has an activation function (e.g., a sigmoid function, a normalized linear function, a softmax function, etc.). A weighted sum is calculated based on multiple inputs to the unit, and the value of the activation function with the sum value as a variable becomes the output of the unit.

[0076] For example, the evaluation model setting unit 253 adjusts the weights used to calculate the weighted sum in each unit of the neural network, regarding them as parameters related to the learning model. The evaluation model setting unit 253 adjusts the weights used to calculate the weighted sum in each unit to maximize the variance of the output of the neural network.

[0077] In order to maximize the variance of the output of the neural network, for example, the evaluation model setting unit 253 may use a gradient descent method, a stochastic gradient descent method, etc. The evaluation model setting unit 253 may use an error backpropagation method for gradient calculation in the gradient descent method or the stochastic gradient descent method.

[0078] In machine learning using neural networks, generalization performance (the ability to discriminate against unknown data) and overfitting (the phenomenon in which the learning model fits the data used to create it but the generalization performance does not improve) can become problems.

[0079] Therefore, in creating a learning model in the evaluation model setting unit 253, a method such as regularization that restricts the degree of freedom of weights during learning may be used to alleviate overfitting. In addition, a method such as dropout that probabilistically selects units in a neural network and disables other units may be used. Furthermore, in order to improve generalization performance, the pre-processing unit 251 may use methods such as data regularization, data standardization, and data expansion that eliminate bias in the data.

[0080] Alternatively, the evaluation model setting unit 253 may divide a plurality of second spectral data into a plurality of groups and create a learning model for each group. When the second spectral data is classified into several seasoning groups, a learning model may be created for each seasoning group.

[0081] The evaluation model setting unit 253 may use a support vector machine instead of a neural network. In this case, the learning model is expressed by a support vector machine. Machine learning using a support vector machine does not have the problem of local solution convergence and tends to improve generalization performance.

[0082] The evaluation model set by the evaluation model setting unit 253 is transmitted to the calculation unit 255 or the database 23 and used for calculation in the calculation unit 255. The evaluation model may be represented by a neural network or a support vector machine.

[0083] The calculation unit 255 calculates a first feature vector characterizing the first seasoning based on the first spectrum obtained by measuring the first seasoning and the evaluation model. More specifically, the first spectrum after preprocessing by the preprocessing unit 251 is input to the evaluation model set by the evaluation model setting unit 253. The calculation unit 255 calculates the output of the evaluation model. Then, the output of the evaluation model is obtained as the first feature vector characterizing the first seasoning.

[0084] The calculation unit 255 may provide the evaluation model with an input based on the chemical analysis value of the first seasoning in addition to the input based on the first spectral spectrum of the first seasoning. The calculation unit 255 may also provide the evaluation model with an input based on the spectral spectrum of the raw materials of the first seasoning. Furthermore, the calculation unit 255 may provide the evaluation model with an input based on the similarity of the raw materials of the first seasoning.

[0085] The calculation unit 255, the evaluation model setting unit 253 may obtain an output from the evaluation model based on the above-mentioned input as a first feature vector characterizing the first seasoning.

[0086] Alternatively, the calculation unit 255 may acquire a created evaluation model from the database 23 instead of acquiring the evaluation model from the evaluation model setting unit 253 .

[0087] The identifying unit 257 identifies a specific second seasoning associated with the first seasoning from among the second seasonings stored in the database 23 based on the first feature vector and the second feature vector.

[0088] More specifically, the identification unit 257 reads out the second feature vector related to the second seasoning stored in the database 23, and calculates an index value based on the first feature vector and the read out second feature vector. After that, the identification unit 257 identifies the second feature amount having the second feature vector whose calculated index value is larger than a predetermined threshold as a specific second seasoning associated with the first seasoning.

[0089] Here, the index value may be calculated so that the smaller the Euclidean length of the difference vector obtained by subtracting the second feature vector from the first feature vector, the larger the index value becomes. For example, the index value may be a value of a strictly monotonically decreasing function that uses the Euclidean length of the difference vector as an argument.

[0090] The index value may also be the cosine similarity between the first feature vector and the second feature vector. In other words, the index value may be a value obtained by dividing the inner product of the first feature vector and the second feature vector by the absolute value of the first feature vector and the absolute value of the second feature vector. Note that, when the output from the evaluation model is normalized to 1, the calculated absolute values ​​of the first feature vector and the second feature vector are 1. Therefore, the index value may be the inner product of the first feature vector and the second feature vector.

[0091] Furthermore, the identification unit 257 may be configured to identify a second seasoning having a second feature vector with a large index value as the identified second seasoning, in preference to a second seasoning having a second feature vector with a small index value. For example, the identification unit 257 may calculate index values ​​for all second feature vectors stored in the database 23, and then select a predetermined number of second feature vectors in descending order of index value. Then, the identification unit 257 may be configured to identify a second seasoning associated with the selected second feature vector as the identified second seasoning.

[0092] Alternatively, the identification unit 257 may set a higher priority for a second seasoning having a second feature vector with a large index value than a second seasoning having a second feature vector with a small index value. Also, the identification unit 257 may prepare a plurality of second feature vectors (e.g., a feature vector based on a spectroscopic spectrum and a feature vector based on a physicochemical analysis value) and evaluate the overall similarity by taking a weighted average of the similarity for each vector.

[0093] [Learning process procedure for evaluation device] Next, the procedure of the learning process of the evaluation device according to the present disclosure will be described with reference to the flowchart of FIG.

[0094] The process of the flowchart shown in FIG. 3 may be started based on an instruction from a user, or may be started when a predetermined number or more pieces of data of the second optical spectrum are stored in the database 23.

[0095] In step S101, the preprocessing unit 251 acquires the data of the second optical spectrum stored in the database .

[0096] In step S103, the preprocessing unit 251 performs preprocessing on the acquired data of the second optical spectrum.

[0097] In step S105, the evaluation model setting unit 253 inputs the preprocessed second spectrum to the learning model in which the parameters are set to initial values, and calculates the second feature vector and the variance related to the second feature vector.

[0098] In step S107, the evaluation model setting unit 253 changes the parameters of the learning model to update the learning model. Note that the variance for the second feature vector before changing the parameters of the learning model is held as the variance before parameter change.

[0099] In step S109, the evaluation model setting unit 253 inputs the preprocessed second spectrum to the updated learning model, calculates the second feature vector and the variance related to the second feature vector, and holds the variance related to the second feature vector after changing the parameters of the learning model as the variance after parameter change.

[0100] In step S111, the evaluation model setting unit 253 calculates the amount of change in variance obtained by subtracting the variance before the parameter change from the variance after the parameter change, and determines whether the amount of change in variance is equal to or less than a predetermined amount.

[0101] The evaluation model setting unit 253 may determine whether or not the amount of change in variance per unit change in parameter, obtained by dividing the amount of change in variance by the amount of change in the parameter, is equal to or less than a predetermined amount.

[0102] If the amount of change in variance is not equal to or less than the predetermined amount (NO in step S111), the process returns to step S107, and the evaluation model setting unit 253 sequentially changes the parameters and updates the learning model.

[0103] On the other hand, if the amount of change in variance is equal to or less than the predetermined amount (YES in step S111), in step S113, the evaluation model setting unit 253 ends the update of the learning model, and sets the learning model at that time as the evaluation model.

[0104] In step S115, the evaluation model setting unit 253 stores in the database 23 a second feature vector obtained when the preprocessed second optical spectrum data is input to the obtained evaluation model.

[0105] In this way, by performing the process of the flowchart shown in Fig. 3, it is possible to set an evaluation model that maximizes the variance of the second feature vector. That is, the axes of the feature vector space are selected so that the end points of the second feature vector calculated based on the second spectral data stored in the database 23 are widely distributed within the feature vector space.

[0106] [Evaluation process procedure for evaluation device] Next, the procedure of the evaluation process of the evaluation device according to the present disclosure will be described with reference to the flowchart of FIG.

[0107] 4 may be started based on a user's instruction, or may be started when data of the first optical spectrum is input to the evaluation device 20. It is assumed that an evaluation model has been set by the evaluation model setting unit 253 before the processing of the flowchart shown in FIG.

[0108] In step S201, the preprocessing unit 251 acquires data of the first optical spectrum.

[0109] In step S203, the preprocessing unit 251 performs preprocessing on the acquired data of the first optical spectrum.

[0110] In step S205, the calculation unit 255 inputs the preprocessed first spectrum to the evaluation model, and calculates a first feature vector.

[0111] In step S207, the identification unit 257 acquires a second feature vector related to the second seasoning stored in the database 23.

[0112] In step S209, the identification unit 257 calculates an index value based on the first feature vector and the acquired second feature vector.

[0113] In step S211, the identification unit 257 determines whether the calculated index value is greater than a predetermined threshold value. If the calculated index value is greater than the predetermined threshold value (YES in step S211), the process proceeds to step S213. On the other hand, if the calculated index value is equal to or less than the predetermined threshold value (NO in step S211), the process proceeds to step S215.

[0114] In step S213, the identifying unit 257 identifies a second feature amount having a second feature vector in which the calculated index value is greater than a predetermined threshold value as a identified second seasoning associated with the first seasoning.

[0115] In step S215, the identification unit 257 determines whether or not all the second feature vectors stored in the database 23 have been acquired. If all the second feature vectors have not been acquired (NO in step S215), the process returns to step S207.

[0116] On the other hand, if all the second feature vectors have been acquired (YES in step S215), in step S217, the identification unit 257 outputs information related to the identified second seasoning. The output information related to the identified second seasoning is presented to the user via the presentation unit 29, for example.

[0117] [Effects of the embodiment] As described above in detail, the evaluation device, evaluation method, and evaluation program according to the present disclosure acquire data on a first spectroscopic spectrum of a first seasoning. Then, the first feature vector is calculated using an evaluation model that outputs a first feature vector characterizing the first seasoning in response to an input based on the first spectroscopic spectrum. A second feature vector characterizing a second seasoning is stored, and a specific second seasoning associated with the first seasoning is identified from among the second seasonings based on the first feature vector and the second feature vector. The evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the second spectroscopic spectrum of the second seasoning.

[0118] This allows seasonings to be evaluated with high accuracy while suppressing the influence of human subjectivity. In particular, it is possible to identify a second seasoning similar to a first seasoning from among the second seasonings stored in the database while suppressing the influence of human subjectivity. Furthermore, it is possible to identify a second seasoning similar to a first seasoning by taking into account the characteristics of seasonings that are difficult to identify using sensor technology that mimics human taste.

[0119] In addition, in the process of developing a seasoning similar to a first seasoning, an appropriate second seasoning can be selected as the starting point for development. As a result, the trial and error step in development can be reduced, and the development period can be shortened. Furthermore, by automating the task of selecting an appropriate second seasoning as the starting point for development, the dependency on experienced developers can be reduced, and development costs such as labor costs can be suppressed.

[0120] The index value may be calculated based on the first feature vector and the second feature vector. Alternatively, a second seasoning having a second feature vector with an index value greater than a predetermined threshold may be identified as a specific second seasoning. This makes it possible to identify a second seasoning similar to a first seasoning from among the second seasonings stored in the database using the index value, which is a standard that is not influenced by human subjectivity.

[0121] The index value may be a value that is larger as the Euclidean length of the difference vector obtained by subtracting the second feature vector from the first feature vector is smaller. The index value may also be the cosine similarity between the first feature vector and the second feature vector. With these, it is possible to identify a second seasoning that is similar to the first seasoning from among the second seasonings stored in the database using the index value, which is a standard that is not influenced by human subjectivity.

[0122] A second seasoning having a second feature vector with a large index value may be specified as the specific second seasoning in preference to a second seasoning having a second feature vector with a small index value. This makes it possible to specify a second seasoning that tends to be more similar to the first seasoning from among a plurality of second seasonings similar to the first seasoning.

[0123] The evaluation model may output a first feature vector in response to an input based on a chemical analysis value of a first seasoning, thereby making it possible to identify a second seasoning similar to the first seasoning, taking into account the chemical analysis value of the first seasoning.

[0124] The evaluation model may output a first feature vector in response to an input based on the spectral spectrum of the raw material of the first seasoning, thereby making it possible to identify a second seasoning similar to the first seasoning, taking into account the raw material of the first seasoning.

[0125] The evaluation model may be a learning model obtained by changing parameters in a direction that increases the variance of the second feature vector output from the learning model for an input based on the chemical analysis value of the second seasoning. This makes it possible to identify the second seasoning that is similar to the first seasoning, taking into account the chemical analysis value of the second seasoning.

[0126] The evaluation model may be a learning model obtained by changing parameters in a direction that increases the variance of the second feature vector output from the learning model for an input based on the spectral spectrum of the raw materials of the second seasoning. This makes it possible to identify a second seasoning similar to the first seasoning, taking into account the raw materials of the second seasoning.

[0127] The evaluation model may be expressed by a neural network or a support vector machine. This makes it possible to identify a second seasoning similar to a first seasoning while suppressing the influence of human subjectivity. In addition, it becomes easy to improve generalization performance (discrimination ability for unknown data). Furthermore, it becomes easy to suppress overfitting (a phenomenon in which the generalization performance does not improve while the learning model is adapted to the data used to create the learning model).

[0128] The evaluation model may output a first feature vector for an input obtained by performing principal component analysis or factor analysis on the first spectrum, thereby reducing the amount of data input to the evaluation model and reducing calculation costs.

[0129] The evaluation model may be a learning model obtained by changing parameters until the amount of change in variance per unit change in the parameters becomes equal to or less than a predetermined amount. This reduces the possibility that the evaluation model is set biased toward one of the second seasonings stored in the database. In particular, by setting the predetermined amount to a small amount in advance, it is possible to ensure that biased learning is not performed in the evaluation model.

[0130] The first feature vector and the second feature vector characterizing the specified second seasoning may be presented to the user via the presentation unit. This allows the user to recognize the characteristics of the first seasoning and the characteristics of the second seasoning similar to the first seasoning without relying on an expert. Furthermore, it is possible to reduce the dependency on experts in development and suppress development costs such as labor costs.

[0131] Each of the functions described in the above embodiments may be implemented by one or more processing circuits, including programmed processors, electrical circuits, and even devices such as application specific integrated circuits (ASICs), or circuit components arranged to perform the described functions.

[0132] Although several embodiments have been described, the embodiments can be modified or modified based on the above disclosure. All components of the above embodiments and all features described in the claims may be individually extracted and combined as long as they are not mutually inconsistent. [Explanation of symbols]

[0133] 10 Spectroscopic equipment 20 Evaluation Equipment 21 Acquisition Department 23 Database 25 Controller 251 Pretreatment section 253 Evaluation Model Setting Department 255 Calculation Unit 257 Specific part 29 Presentation section SP Sample

Claims

1. an acquisition unit that acquires data of a first spectroscopic spectrum of a first seasoning; a calculation unit that calculates the first feature vector by using an evaluation model that outputs a first feature vector that characterizes the first seasoning in response to an input based on the first spectral spectrum; a storage unit that stores a second feature vector that characterizes the second seasoning; an identification unit that identifies a specific second seasoning associated with the first seasoning from among the second seasonings based on the first feature vector and the second feature vector; An evaluation device comprising: The evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the second spectroscopic spectrum of the second seasoning, in an evaluation device.

2. The identification unit is Calculating an index value based on the first feature vector and the second feature vector; The evaluation device according to claim 1 , wherein the second seasoning having the second feature vector whose index value is greater than a predetermined threshold value is identified as the identified second seasoning.

3. The evaluation device according to claim 2 , wherein the index value is larger as a Euclidean length of a difference vector obtained by subtracting the second feature vector from the first feature vector is smaller.

4. The evaluation device according to claim 2 , wherein the index value is a cosine similarity between the first feature vector and the second feature vector.

5. The evaluation device according to any one of claims 2 to 4, wherein the identification unit identifies the second seasoning having the second feature vector with a large index value as the identified second seasoning in preference to the second seasoning having the second feature vector with a small index value.

6. The evaluation device according to claim 1 , wherein the evaluation model outputs the first feature vector in response to an input based on a chemical analysis value of the first seasoning.

7. The evaluation device according to claim 1 , wherein the evaluation model outputs the first feature vector in response to an input based on a spectroscopic spectrum of a raw material of the first seasoning.

8. The evaluation device according to any one of claims 1 to 7, wherein the evaluation model is a learning model obtained by changing the parameters in a direction that increases the variance for the second feature vector output from the learning model in response to an input based on a chemical analysis value of the second seasoning.

9. The evaluation device according to any one of claims 1 to 8, wherein the evaluation model is a learning model obtained by changing the parameters in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the spectroscopic spectrum of the raw materials of the second seasoning.

10. The evaluation device according to any one of claims 1 to 9, wherein the evaluation model is represented by a neural network or a support vector machine.

11. The evaluation device according to claim 1 , wherein the evaluation model outputs the first feature vector in response to an input obtained by performing a principal component analysis or a factor analysis on the first spectrum.

12. The evaluation device according to any one of claims 1 to 11, wherein the evaluation model is a learning model obtained by changing the parameters until a change in the variance per unit change in the parameters becomes equal to or less than a predetermined amount.

13. The evaluation device according to claim 1 , further comprising a presentation unit that presents to a user the first feature vector and the second feature vector characterizing the specified second seasoning.

14. acquiring first spectroscopic data of a first seasoning; calculating the first feature vector using an evaluation model that outputs a first feature vector characterizing the first seasoning in response to an input based on the first spectral spectrum; storing a second feature vector characterizing the second seasoning; Identifying a specific second seasoning associated with the first seasoning from among the second seasonings based on the first feature vector and the second feature vector.

1. An evaluation method comprising: An evaluation method, wherein the evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance of the second feature vector output from the learning model in response to an input based on the second spectroscopic spectrum of the second seasoning.

15. On the computer, acquiring first spectral data of a first seasoning; calculating the first feature vector using an evaluation model that outputs a first feature vector characterizing the first seasoning in response to an input based on the first spectral spectrum; storing a second feature vector characterizing the second seasoning; Identifying a specific second seasoning associated with the first seasoning from among the second seasonings based on the first feature vector and the second feature vector; An evaluation program for executing An evaluation program, wherein the evaluation model is a learning model obtained by changing parameters related to the learning model in a direction that increases the variance for the second feature vector output from the learning model in response to an input based on the second spectroscopic spectrum of the second seasoning.

Images