JP2018018354A - Quality prediction method for food and drink using deep learning, and food and drink - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a quality prediction method for food and drink using a quality prediction model of food and drink obtained by the learning through deep learning.SOLUTION: The method includes a step in which individual samples obtained by performing pretreatment on a plurality of food and drink products with known quality are subjected to one or more analyses selected from an instrumental analysis or a function evaluation to obtain individual data, and then the data is converted to numerical data or image data to divide the same into training data and test data, a quality prediction model of food and drink obtained by learning a relation between the training data and known quality with deep learning is compared with the test data to change a parameter, for obtaining a tuned quality prediction model.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a food and beverage quality prediction method utilizing deep learning, and a food and beverage manufactured using the quality prediction method.

Foods and drinks generally have a complex composition and are evaluated by humans according to their senses and sensibilities when they enter the human mouth. The sensory evaluation is widely adopted in the food and drink field. However, it is also true that sensory evaluation is difficult to achieve reproducibility at the same level as instrumental analysis, and to suppress variations among individuals. Therefore, in the practice of quality assessment of food and drink, instrumental analysis and sensory evaluation are used together, but it cannot be said that a sufficiently satisfactory method is provided for the interpretation method that links these results. Moreover, it cannot be said that a sufficiently satisfactory method is provided for the interpretation method that links the product design quality, sensory evaluation, and instrumental analysis results.
On the other hand, with the remarkable development of computer technology in recent years, metabolome analysis, which is a research method for comprehensive analysis of low molecular weight metabolites related to life phenomena, has been rapidly and globally utilized,・ There are many applications in the industrial field. In the agriculture, forestry, fishery and food fields, metabolomic analysis is an applicable technology for agricultural products such as taste and aroma, and low molecular weight metabolites related to food quality. It is required to strongly promote the application of (Non-patent Document 1).

  Examples of reports that apply metabolomic analysis to the food field include: a step of pre-treating green tea to obtain an analysis sample; a step of subjecting the analysis sample to instrumental analysis to obtain an analysis result; and a step of converting the analysis result into numerical data A multivariate analysis step; and a green tea quality prediction method (Patent Document 1) including a step of predicting the quality from the obtained analysis result, or a pretreatment step of pretreating cheese with a known ripening quality to obtain an analysis sample Instrument analysis step of subjecting the analysis sample to instrument analysis to obtain instrument analysis data; multivariate analysis using instrument analysis data for a plurality of the cheeses and ripening quality data representing each ripening quality of the cheese; A multivariate analysis step of creating a cheese quality prediction model that represents the relationship between the device analysis data and the ripening quality predicted from the device analysis data Having a quality prediction method for cheese (Patent Document 2) are known.

  However, the interpretation method used in these reports to connect the instrumental analysis results and sensory evaluation results is a multivariate analysis method, and machine learning is not mentioned. It cannot be said that the model described in the above has been created.

  As a report that applied machine learning to design quality prediction of food and drink, high-performance liquid chromatography of RTD (Ready To Drink) coffee, regular coffee extract, and coffee extract-diode array detector (HPLC-DAD) chromatogram data There is a report on predicting information related to the design of coffee products such as coffee varieties and roasting colors by applying various chemometric techniques using Non-patent Document 2 (Non-patent Document 3). However, with regard to machine learning, only support vector machines, random forests, neural networks, bagging, and gradient boosting are illustrated, and no deep learning is mentioned.

  In addition, in the review on the case where chemometrics is applied to the correlation analysis of the results of gas chromatogram data and sensory evaluation data, only examples of neural networks are shown for machine learning. It is not mentioned (Non-Patent Document 4).

Promotion Policy for Industry-Academia-Government Collaboration Research in the Agriculture, Forestry and Fisheries / Food Field Utilizing Metabolome Analysis January 14, 2015 Industry-Academia-Government Collaboration Research Study Group in Agriculture, Forestry and Fisheries / Food Field Utilizing Metabolome Analysis Japan Soft Drinks Research Group "25th Research Presentation" Lecture, pp15-21, published on December 25, 2015 Takasago Perfume Time Report, pp 24-27, Reissue No. 177, May 18, 2016 Food Chemistry 210 (2016) 530-540

JP 2009-14700 A JP 2013-7732 A

  An object of this invention is to provide the method of predicting the quality of food / beverage products flexibly by making food / beverage products into prediction object.

The present invention relates to the following [1] to [7].
[1]
(A) A step of pretreating a plurality of foods and drinks of known quality to obtain an analysis sample;
(B) A step of obtaining data by subjecting the analysis sample obtained in (a) to one or more analyzes selected from instrumental analysis or sensory evaluation;
(C) a step of converting the data obtained in (b) into numerical values or image data;
(D) dividing the numerical value or image data obtained in (c) into training data and test data;
(E) learning the relationship between individual training data and known quality by deep learning to obtain a quality prediction model;
(F) A step of predicting quality by providing test data to the quality prediction model obtained in (e);
(G) collating the quality prediction result obtained in (f) with a known quality;
(H) changing the parameters of the quality prediction model;
(I) (e) (f) (g) (h) is repeated one or more times to tune the quality prediction model and obtain a tuned quality prediction model;
Quality prediction method for foods and drinks including
[2]
[1] The method for predicting quality of food and drink according to [1],
(J) By selecting two or more numerical data of individual analysis samples obtained in (c) and proportionally calculating the data at an arbitrarily selected ratio, virtually blending at an arbitrarily selected ratio Obtaining numerical data of the analyzed sample;
(K) a step of converting the numerical data obtained in (j) into image data as necessary;
(L) Predicting the quality of the blended food and drink at an arbitrarily selected ratio by providing the numerical value or image data obtained in (k) to a tuned quality prediction model;
Quality prediction method for foods and drinks including
[3]
[1] The method for predicting quality of food and drink according to [1],
(M) a step of pre-processing food and drink of unknown quality to obtain an analysis sample;
(N) A step of obtaining data by subjecting the analysis sample obtained in (m) to one or more types of analysis selected from instrumental analysis or sensory evaluation;
(O) converting the data obtained in (n) into numerical values or image data;
(P) predicting quality by subjecting the numerical value or image data obtained in (o) to a tuned quality prediction model;
Quality prediction method for foods and drinks including
[4]
[1] to [3] The quality described in [1] to [3] is selected from the quality sensory-evaluated using human senses and / or the quality analyzed by the instrument and / or the design quality of the product. ] The method of description.
[5]
The method according to [1] to [4], wherein the deep learning is selected from a convolutional neural network, a deep neural network, a deep belief network, and a recursive neural network.
[6]
The method according to any one of [1] to [5], wherein the step of converting the analysis result into image data is a step of converting the obtained numerical data into a graph to form image data.
[7]
A food or drink that reproduces the quality predicted by the method according to [1] to [6].

  ADVANTAGE OF THE INVENTION According to this invention, the quality prediction of the food / beverage products which was difficult conventionally can be performed with an accurate and flexible method.

FIG. 1 is a flowchart for creating a quality prediction model that has been tuned according to claim 1. FIG. 2 is a flowchart of quality prediction of food and drink related to claim 2. Note that * represents a quality prediction model that has been tuned and obtained in the same procedure as in FIG. FIG. 3 is a flowchart of quality prediction of food and drink related to claim 3. Note that * represents a quality prediction model that has been tuned and obtained in the same procedure as in FIG.

(Food and drink subject to quality prediction)
Although food / beverage products are not specifically limited, Food / beverage products etc. which include a foodstuff and processed food, and are described in a Japanese food ingredient table are mentioned.
Examples of the food include dairy products, confectionery, fat products, processed agricultural products, seasonings, soups, processed livestock products, and processed fishery products.
Dairy products include butter, cheese and cheese food.
Examples of the confectionery include frozen confectionery, Japanese confectionery, Western confectionery, dessert, baked confectionery, bakery, and candy.
Examples of oils and fats include margarine, coffee whitener, and chocolates.
Examples of processed agricultural products include noodles, processed vegetable protein products, jams, pastes, pickles, canned agricultural products, fruit juices, and processed meat products.
Seasonings include miso, soy sauce, sauce, mayonnaise, and dressing.
Examples of soups include powdered soups, retort pouch soups, and canned soups.
Processed livestock products include ham, sausage, hamburger and canned meat.
Examples of processed fishery products include fish ham, fish sausage, fish paste products, and canned fish.
Others include bakery products, candies, gums, frozen foods, retort foods, instant foods, livestock feed, fish feed, pet feed, etc. These ingredients are milk, cereals, edible fats, fruits and vegetables. Agricultural products such as animal meat, seafood, etc. are also covered.

Beverages include carbonated drinks, fruit drinks, vegetable drinks, beers, non-alcoholic beers, non-alcoholic drinks, Chuhai, other liquors, tea drinks, coffee drinks, functional drinks, cocoa drinks, cacao drinks, sugarless Beverages, sports drinks, nutrition / nutrition drinks, dairy products, lactic acid bacteria drinks, milk drinks, vinegar drinks, and the like.
Examples of teas include tea leaves and tea leaf extracts that are raw materials for green tea, black tea, oolong tea, and the like.
Coffee beverages include coffee beans, coffee liquor, instant coffee and the like.
Preferred are coffee drinks and tea drinks, and particularly preferred are coffee drinks.

In addition, the food and drink includes a fragrance-based raw material.
Perfume base materials include citrus fruits such as lemon, grapefruit and orange, fruits such as apple, melon, grape, peach and pineapple, beverages such as tea, green tea, oolong tea and coffee, and dairy products such as milk and yogurt. Mint, vanilla, peppermint, spearmint, Japanese mint, black pepper, cinnamon, clove, nutmeg, turmeric, basil, oregano, rosemary, sage, thyme, cardamom, coriander, dill, fennel , Nuts such as almonds, cashews, peanuts, walnuts and pine nuts, meats such as beef, pork and chicken, and seafoods such as salmon, bonito, seaweed, kombu, crab and scallops. As long as they are plant-derived substances, these fruits, flowers, buds, trees, bark, branches, leaves, stems, roots and extracts thereof can be targeted.
In addition to these, specific examples include the substances described in “Natural Fragrance Base Material Collection” (issued June 1, 2011, edited by the Japan Fragrance Industry Association).

(Definition of coffee beans)
In the present invention, the types of coffee beans that are listed as the quality prediction target coffee products are Arabica, Robusta, and Revelica, and may be a blend of two or more of these. The coffee beans may be green beans, and in the case of roasted beans or roasted and ground beans, the roasting conditions and roasting equipment can be performed by any known equipment and conditions, and those appropriately selected are subject to quality prediction. It becomes. The roasting degree can be expressed by brightness such as a Hunter L value in the Lab color space and an L star value in the CIE color space. For example, in the case of the Hunter L value, the range of 13 to 32 can be exemplified. As a method for measuring the roasting degree, for example, roasted beans are pulverized while removing chaff as appropriate, the pulverized beans are put into a cell, sufficiently tapped, and then measured with a color difference meter. As the color difference meter, ZE-2000, ZE-6000 (manufactured by Nippon Denshoku Industries Co., Ltd.) or the like can be used. Moreover, as a roasting apparatus, a direct fire type, a semi-hot air type, a hot air type etc. can be illustrated, for example. In the case of roasted and ground beans, the roasted beans are produced by pulverizing roasted beans by any known method, and there is no particular limitation.

(Definition of coffee liquid)
In the present invention, the coffee liquid that is subject to quality prediction includes a coffee extract obtained by bringing a roasted ground bean and an aqueous solvent into contact with each other, and a coffee extract having a higher soluble solid content than the normal drinking concentration. Extract, liquid in which instant coffee is dissolved, coffee liquid containing finely pulverized roasted coffee beans that have been finely pulverized to an average particle size of about 300 μm or less, a container that is packed in a container and refrigerated and stored at room temperature Includes stuffed beverages. Furthermore, auxiliary materials such as milk materials and sugars may be added to these coffee liquids as needed, or different types of coffee liquids may be mixed. The extraction method of the coffee extract and the coffee extract is not particularly limited, and examples thereof include a drip method, an espresso method, a fixed bed column method, a slurry method, a continuous countercurrent method, and an extractive distillation method. Extraction conditions such as extraction temperature, extraction pressure, extraction time, and bean / extraction solvent ratio are not particularly limited, and those manufactured under any conditions are targeted. As a container-packed beverage, for example, sugar-free black coffee; sweetened black coffee to which sucrose, liquid sugar, sweeteners, etc. are added; milk components such as milk, skim milk powder, and fresh cream are added to sugar-free or sweetened coffee beverages Café olea type coffee drinks etc. are included in the quality prediction targets. When producing a container-packed beverage, any component that can be used in a coffee beverage can be added to the coffee liquid as an auxiliary material. For example, a coffee extract obtained by a conventional method, an antioxidant, pH Examples thereof include regulators, emulsifiers, and fragrances. After normal pH adjustment, homogenize, heat to 80-95 ° C using a plate heat exchanger, etc., then fill and tighten, and sterilize at 120-125 ° C for 10-40 minutes And a retort sterilized coffee drink is obtained. In the case of UHT sterilization, a plate-type heat exchanger or the like is used, and after sterilization treatment at 130 to 145 ° C. for 15 to 60 seconds, it is filled into a container such as a can or a plastic bottle to obtain a container-packed beverage. The container-packed beverage is included as a quality prediction target.

(Definition of instant coffee)
In the present invention, instant coffee (including regular soluble coffee, hybrid coffee, etc.) that is a quality prediction target is a coffee extract that is a coffee extract having a higher soluble solid content concentration than the normal drinking concentration, or an average particle size of 300 μm. Includes coffee extract containing finely pulverized roasted coffee beans that have been finely pulverized to the following extent, in which water has been removed by spray drying or freeze drying, and sugars, milk materials, etc. Also includes instant coffee products with added and mixed ingredients.

<Pretreatment process>
In the present invention, the pretreatment step is a step performed in order to obtain a form suitable for instrumental analysis and sensory evaluation in order to obtain information related to the quality of the food and drink to be subjected to quality prediction. Unit operations such as fractionation, concentration, drying, and filtration are included. These unit operations may be appropriately combined. In particular, when the instrumental analysis is high-performance liquid chromatography (HPLC), an eluent of high-performance liquid chromatography (HPLC) containing an organic solvent such as methanol or acetonitrile is added to the analytical sample to precipitate insolubles, and then chromatographed. A pretreatment step of microfiltration using a disk or the like is suitable.
When the instrumental analysis is gas chromatography (GC), a distillate generated by steam distillation of food and drink is collected using a known organic solvent such as diethyl ether, isopentane, pentane, Solvent extraction using known organic solvents, pretreatment process used for headspace gas analysis in which the volatile components of coffee that have reached equilibrium in a sealed container are collected with a syringe or adsorbed to an adsorbent, nonpolar porous A pretreatment step used in a solid phase extraction method in which it is adsorbed on a polymer bead and then eluted with a solvent, a pretreatment step used in a reduced pressure steam distillation method for separating volatile components by performing steam distillation with a rotary evaporator, etc. A pretreatment process is mentioned.

<Process for obtaining data for analysis>
(Instrument analysis)
In the present invention, instrumental analysis refers to performing analysis using an analytical instrument, and analysis methods include high performance liquid chromatography (HPLC), gas chromatography (GC), ion chromatography, mass spectrometry (MS). ), Near infrared spectroscopy (NIR), Fourier transform infrared spectroscopy (FT-IR), nuclear magnetic resonance analysis (NMR), Fourier transform nuclear magnetic resonance analysis (FT-NMR), inductively coupled plasma mass spectrometer ( ICP-MS), taste recognition device with artificial lipid membrane sensor, RGB camera, image spectroscopic camera, depth camera, infrared camera, magnetic camera, microphone, multi-array microphone, electronic stethoscope, frequency analyzer, vibration sensor, acceleration A sensor, a rheology meter, a texturometer, etc. can be illustrated.
Preferably high performance liquid chromatography (HPLC), gas chromatography (GC), ion chromatography, mass spectrometry (MS), near infrared spectroscopy (NIR), Fourier transform infrared spectroscopy (FT-IR), nuclear magnetic Examples include taste analysis apparatus equipped with resonance analysis (NMR), Fourier transform nuclear magnetic co-analysis (FT-NMR), inductively coupled plasma mass spectrometer (ICP-MS), and artificial lipid membrane sensor, particularly preferably high performance liquid chromatography. Examples include chromatography (HPLC), gas chromatography (GC), and mass spectrometry (MS).

A detector for high performance liquid chromatography (HPLC) is not particularly limited, and a known one can be used as appropriate. For example, a diode array (DAD) detector can be suitably used. The detector for gas chromatography (GC) is not particularly limited, and for example, a mass analyzer (MS) or a flame ionization detector (FID) can be used.

Further, two or more of these instrumental analysis methods may be combined.
The analysis conditions can be arbitrarily selected as long as sufficient reproducibility according to the target prediction accuracy is obtained.

(sensory evaluation)
The sensory evaluation in the present invention includes a sensory test and a sensory test, and means evaluation of food and drink using human senses (sight, hearing, taste, smell, touch). As the sensory evaluation method, a method suitable for the evaluation of the target food and drink can be arbitrarily selected. Although there is no particular limitation, an analytical sensory evaluation method using a specialized panel (two-point identification method, 3 Discrimination type test such as point discrimination method, scoring method, descriptive type test such as QDA (Quantitative Descriptive Analysis) method, time intensity test, TDS (Temporal Dominance of Sensations), TCATA (Temporal Check All Time, etc.) Method), a taste-type sensory evaluation method using a general panel.
Preferably, a 2-point identification method, a 3-point identification method, and a taste-type sensory evaluation method using a general panel are used, and a 2-point identification method is particularly preferable.

<Step of converting to numerical value or image data>
Data obtained by instrumental analysis is obtained as waveform data obtained by sampling the result of the change in signal intensity converted into a normal electrical signal at regular time intervals, or as an integrated value obtained by integrating electrical signals for a certain time. Feature values such as slope, periodicity, and amplitude obtained from waveform data, and smoothing processing of waveform data to detect peaks and extracting numerical values such as area value, peak height, peak width, etc. The process of converting the numerical values into numerical data, such as ratioating the numerical values, using the signal intensity in a specific time range as it is, or taking out the quantified numerical value by paying attention to a specific compound, is arbitrary from among known methods Can be selected.
In particular, when performing an instrumental analysis in which a diode array detector is provided as a detector after the high performance liquid chromatography (HPLC), a peak is detected from the obtained chromatogram data, and the area value is obtained as numerical data. Is preferable from the viewpoint of high reproducibility of numerical values. Therefore, although it is possible to create a quality prediction model without specifying the target peak component, use the numerical data of the compound that is common to the training and prediction target samples and is relatively easy to quantify. Is preferred.

  In sensory evaluation, data is digitized and collected as continuous values such as scoring values and ratios, and discrete values such as rating values and frequencies.

  The obtained numerical data can be visualized by an arbitrary visualization method such as a chromatogram, radar chart, spider chart, heat map, contour map, etc., and can be converted into image data by storing it in an electronic file as an image. In addition, when an RGB camera, an image spectroscopic camera, or the like is used for instrument analysis, directly digitized image data can be obtained. Furthermore, image conversion, cropping, resizing, slight rotation fluctuation, resolution fluctuation, gamma correction, gray scale conversion, flattening, threshold processing, color reduction processing as needed when converting to image data and after conversion Further, processing such as mask processing, random noise may be added uniformly, or data augmentation such as filtering such as a blur filter or a Gaussian filter may be performed. There are no particular restrictions on the software or file format to be converted into image data. For example, commercially available Excel (manufactured by Microsoft), Photoshop (registered trademark) (manufactured by Adobe Systems), freeware R or OpenCV Examples of the file format include, but are not limited to, a png format, a jpg format, a bmp format, a tiff format, and the like.

<The process of learning by deep learning and obtaining a quality prediction model>
In the present invention, the step of learning by deep learning is performed using the obtained numerical values or image data. In the present invention, deep learning means machine learning using a learning device having a network structure having a hierarchical structure of at least four layers, and has a flexible number of parameters that can be tuned compared to conventional machine learning. Therefore, high performance can be exhibited for various data configurations. There are no particular restrictions on the configuration of the deep learning network that can be used, and any method can be selected, and examples include convolutional neural networks, deep neural networks, deep belief networks, and recurrent neural networks. There are no particular restrictions on the types and numerical ranges of the tuning parameters for deep learning, and appropriate tuning may be performed from the viewpoint of the data to be used, the target prediction accuracy, the calculation cost, etc. Size and number of kernels to be convolved, type of activation function, pooling method and size, number of convolution layers, number of Maxout, number of units of all connected layers, number of all connected layers, dropout ratio, mini-batch size, The number of updates, the learning rate, the presence or absence of an auto encoder, etc. can be exemplified. Further, the input variable and the response variable can be arbitrarily selected from discrete values, continuous values, or combinations thereof.

  The framework for performing these deep learnings can be appropriately selected from a large number of examples. , KERAS, MatConvNet, OpenDeep, and Matlab.

In the present invention, the quality of the food and drink is the type of food and drink used, the blend ratio of the food variety, the origin of the food, the blend ratio of the foods with different origins, the degree of processing of the raw materials for food and drink such as heat processing, etc. Design quality related to the design of food and drink, pass / fail judgment by sensory evaluation, judgment of the presence or absence of odor by sensory evaluation, flavor rating by sensory evaluation, quality related to human sensory quantity such as sensory evaluation score, etc. Includes the quality to be analyzed.

  In the present invention, a plurality of foods and drinks of known quality are pretreated to obtain individual analysis samples, and the individual analysis samples are subjected to instrument analysis or sensory evaluation to obtain individual instrument analysis data or sensory evaluation data. Further, the individual device analysis or sensory evaluation data and the known quality of each of a plurality of foods and beverages are used for learning by deep learning to be predicted from the device analysis or sensory evaluation data and the device analysis data or sensory evaluation data. Create a quality prediction model for foods and drinks that represents the relationship with quality.

<Process for predicting quality by supplying test data to a quality prediction model>
In order to evaluate the high accuracy of the quality prediction model, separate quality-equipped instrumental analysis data or sensory evaluation data for training and testing, and using only the training data to create a quality model, The test data is collated with the quality prediction model to obtain a result of predicting the quality of the test data.

<Process for comparing quality prediction results with known quality>
The holdout evaluation can be performed by comparing the quality prediction result of the test data with the known quality of the test data. Further, as a more sophisticated evaluation method, evaluation by cross-validation such as K-division cross-validation or Leave-one-out cross-validation may be performed. Evaluate the prediction accuracy of the quality prediction model and adjust the appropriate deep learning network configuration and tuning parameters.

<Process for changing quality prediction model parameters>
There are no particular restrictions on the types and numerical ranges of the parameters that can be changed in the quality prediction model learned by deep learning, and it is only necessary to perform appropriate tuning from the viewpoints of data used, target prediction accuracy, calculation cost, etc. However, the types of tuning parameters include the size and number of kernels to be convolved, the type of activation function, the method and size of pooling, the number of convolution layers, the number of Maxout, the number of units of all coupling layers, the number of all coupling layers, Examples include dropout ratio, mini-batch size, number of updates, learning rate, presence / absence of auto encoder, and the like.

<Tuning the quality prediction model and obtaining a tuned quality prediction model>
A tuned quality prediction model can be obtained by appropriately changing and tuning the changeable parameters.
The prediction accuracy of the quality prediction model can be said to be higher, but it varies depending on the type of target quality, the cost of training data collection for creating the quality prediction model, and time constraints, The range of the prediction accuracy can be arbitrarily set according to the purpose, but in order to avoid overfitting and improve the generalization performance, the quality prediction is substantially achieved unless at least one tuning is performed. It is difficult to create a model that can.

  And about food-drinks with unknown quality, it pre-processes and obtains an analysis sample, The analysis sample is subjected to instrumental analysis or sensory evaluation to obtain analysis data, and the analysis data is prepared as described above. This quality prediction model can be used to predict the quality of foods and beverages whose quality is unknown.

  Based on the predicted quality of the food and drink, obtain the information related to the food and drink design, make a prototype based on this information, and further perform processing such as extraction according to the form of the food and drink, the predicted quality Can be obtained efficiently.

  Hereinafter, the present invention will be described more specifically with reference to examples. However, the present invention is not limited to these examples.

Example 1
(Prediction of acceptance / rejection judgment by sensory evaluation of raw coffee beans)
<Step of pre-processing to obtain analysis sample>
About 725 specimens of Vietnamese Robusta seed beans, they were pulverized with a commercially available coffee mill (Fuji Royal TYPE R-440, manufactured by Fuji Koki Co., Ltd.) to prepare ground coffee beans for instrumental analysis. In addition, the sample was roasted into a medium roast to a light roast with a test roaster (Probat), pulverized with a commercially available coffee mill (Fuji Royal TYPE R-440, manufactured by Fuji Koki Co., Ltd.), and roasted for sensory evaluation. Roasted coffee beans were prepared.

<Instrument analysis process>
50 g of the ground coffee beans were steam-distilled to collect aroma components in a hexane layer of a Nikerson essential oil quantifier, then concentrated with KD (Kuderna-Danish) and made up to 1 mL with ethyl acetate. Absolute calibration curve based on GC-MS chromatographic peak area comparison using TCA (Trichlorophenol), TCP (Trichlorophenol), TBA (Tribromophenol), and TBP (Tribromophenol) standard products (obtained from Tokyo Chemical Industry) as halogen-related odor components. By the method, halogen-based off-flavor components in green coffee beans were quantified. For one sensory evaluation pass / fail determination, analysis was performed 1 to 12 times according to the sample lot size, variation in analysis value, and the like.
[GC-MS conditions]
GC device: TRACE GC ULTRA (manufactured by Thermo Fisher Scientific)
Detector: DSQII (manufactured by Thermo Fisher Scientific)
Injection volume: 1 μL
Column: Supelco SLBTM-5ms

<Sensory evaluation process>
About 130 mL of hot water was poured into about 10 g of the roasted and ground coffee beans, and the beans floating on the water surface sank to some extent. Then, the liquid scooped with a test spoon was sucked into the mouth to make a sensory evaluation. Each sample was evaluated for 10 to 20 cups, and these evaluations were combined to determine whether or not they were accepted.

<Process to convert to numerical data>
21 variables (TCA average value, TCP average value, average value of the sum of TCA and TCP) by combining the quantitative values of halogen-based off-flavor components (TCP, TCA, TBP, TBA) obtained as a result of GC-MS analysis , TCA / TCP average, TBA average, TBP average, TBA and TBP sum, TCA median, TCP median, TCA and TCP sum, TCA / TCP median, TBA Median, median TBP, median of the sum of TBA and TBP, TCA maximum, TCP maximum, maximum of TCA and TCP, TCA / TCP maximum, TBA maximum, TBP maximum, TBA The maximum value of the sum of TBP) was calculated.

<Process of converting numerical data into image>
Twenty-one variables belonging to each specimen were graphed in a radar chart format, and the resulting graph was converted into a 32 × 32 pixel jpg format image to create a total of 725 specimen image data.

<The process of learning by deep learning and obtaining a quality prediction model>
A total of 200 samples were selected from 224 samples that were accepted by sensory evaluation and 501 samples that were rejected, and a total of 200 samples were selected as a training sample group. A total of 200 samples were selected from the remaining 124 samples and 401 samples that were rejected, and a total of 200 samples were selected as a test sample group.

  The image data set of the specimen group for training was applied to a total of 6 layers of convolutional neural networks to which a convolution layer was added, and a model for predicting acceptance / rejection determination by sensory evaluation as a response variable was created.

<Process to change parameters by matching with quality prediction model>
The image data set of the test sample group was used for the created prediction model, and acceptance / rejection prediction by sensory evaluation was performed to collate with a known pass / fail judgment. Next, the parameters of the prediction model were changed and the prediction model was tuned to obtain a tuned quality prediction model. The activation function of the model was Rectified Linear Unit (ReLU), the batch size was 10, and no dropout was performed. The accuracy rate of the pass / fail judgment prediction of the test data using the tuned model was 81.0%, and high prediction accuracy was obtained.

  Pythone (version 2.7.9) and Chainer (version 1.8.2) were used for the step of learning by deep learning and the step of predicting quality (acceptance determination).

(Example 2)
(Prediction of pass / fail judgment of raw coffee beans blend product)
A change in the pass / fail judgment of the blended product was predicted when two arbitrarily selected sensory evaluation rejected samples (specimens A and B) were blended at various ratios. Numerical values obtained by instrumental analysis of the blended product were calculated by proportional calculation corresponding to the blend ratio of specimens A and B, and converted into images in the same manner as in Example 1. Next, the above-mentioned image set was used for the pass / fail judgment prediction model with the accuracy rate of 81.0% created in Example 1, and the pass / fail judgment was predicted. As a result, it was predicted that the blend ratios of A: B = 1: 9 and 1: 2 were rejected. However, the blend ratios of A: B = 1: 1, 2: 1, and 9: 1 were predicted to pass and predicted. It was done. That is, by blending two rejected samples, an appropriate blend ratio of a blended product that passed with a high probability of about 81% could be calculated.
Pythone (version 2.7.9) and Chainer (version 1.8.2) were used for the step of learning by deep learning and the step of predicting quality (acceptance determination).

(Example 3)
(Prediction of presence or absence of chlorine smell comment)
From the image data obtained in the same manner as in Example 1, a total of 300 samples were selected from the 319 samples commented as having a chlorine odor by sensory evaluation and 406 samples having no chlorine odor, each of which was randomly selected for training. A sample group of From the remaining 169 specimens with chlorine odor and 256 specimens without chlorine odor, a total of 300 specimens, each of 150 specimens, were selected as a test specimen group.

  The image data set of the sample group for training was applied to a total of 6 layers of convolutional neural networks to which a convolution layer was added, and a model for predicting the presence or absence of chlorine odor comments by sensory evaluation as a response variable was created.

The image data set of the test sample group was provided to the created prediction model, and the presence / absence of a chlorine odor comment by sensory evaluation was predicted and collated with the presence / absence of a known chlorine odor comment. Next, the parameters of the prediction model were changed and the prediction model was tuned to obtain a tuned quality prediction model. The activation function of the model used ReLU, the batch size was 10, and no dropout was performed. The accuracy rate of prediction of the presence or absence of a chlorine odor comment in the test data using the tuning-completed model was 71.0%, and high prediction accuracy was obtained.
Pythone (version 2.7.9) and Chainer (version 1.8.2) were used for the step of learning by deep learning and the step of predicting quality (presence / absence of chlorine smell comment).

Example 4
(Prediction of presence or absence of musty odor comments)
From 435 samples commented as having a musty odor by sensory evaluation and 290 samples having no musty comment from the numerical data obtained in the same manner as in Example 1, 290 samples and 193 samples were selected at random, for a total of 483 samples. A sample group for training was used. The remaining 145 specimens with a musty odor comment and 97 specimens without a musty odor comment, a total of 242 specimens were used as test specimen groups.

  A numerical data set of the sample group for training was applied to a deep neural network of all six layers, and a model for predicting the presence or absence of a musty odor comment by sensory evaluation as a response variable was created.

The numerical data set of the test sample group was applied to the created prediction model, and the presence / absence of a musty odor comment by sensory evaluation was predicted and collated with the presence / absence of a known musty odor comment. Next, the parameters of the prediction model were changed and the prediction model was tuned to obtain a tuned quality prediction model. As the activation function of the model, tanh was used, the batch size was 1000, and the learning rate was 0.03. The accuracy rate of the prediction of the presence or absence of a musty odor comment in the test data using the already-tuned model was 71.0%, and a high prediction accuracy was obtained.
R (version 3.3.0) was used for the step of learning by the deep learning and the step of predicting the quality (the presence or absence of a musty odor comment).

(Example 5)
(Prediction of variety)
<Step of pre-processing to obtain analysis sample>
Arabica roasted beans 212 specimens (Brazil 158 specimens, Colombia 33 specimens, Guatemala 21 specimens) L value 15-24, obtained from Takasago coffee) Robusta seed (Vietnam) roasted beans 58 specimens (L value 16-23, Takasago) Crushed using a commercially available dipping mill (manufactured by Ditting Machinechin AG), and then extracted by contacting 10 parts by weight with 1 to 10 parts by weight of hot water for 1 to 10 parts by weight. Then, filtration with filter paper was performed to obtain a coffee extract.
Of the obtained coffee extract, 195 samples were directly subjected to pretreatment for HPLC analysis.
Of the obtained coffee extract, 75 samples were adjusted to pH 5.9 using sodium bicarbonate water, and then the ion exchanged water was used to obtain a soluble solid content concentration of 1.2% by weight. Diluted, filled into a can and wound, and then retort sterilized under the condition of F ₀ = 10 to obtain a black-type container-packed beverage. Of the obtained container-packed beverages, 36 samples were stored as they were at room temperature until they were subjected to pretreatment for HPLC analysis (black type container-packed beverages). The remaining 39 samples were stored at 60 ° C. for 1 week, and then stored at room temperature until pretreatment for HPLC analysis (warmed black type container-packed beverage).
1.5 g of the obtained coffee liquid (195 samples of coffee extract, 36 samples of black-type container-packed beverage, 39 samples of heat-treated black-type container-packed beverage) were weighed, and HPLC eluent (liquid A: 0.05M acetic acid, 5% by volume acetonitrile aqueous solution) was added, and the mixture was subjected to HPLC analysis after fine filtration with a 0.45 μm aqueous chromatographic disk (manufactured by Kurashiki Boseki).

<Instrument analysis process>
10 μL of the coffee liquid after microfiltration was subjected to instrumental analysis (1100 series, manufactured by Agilent Technologies) in which a diode array detector was provided as a detector after the HPLC.
[HPLC conditions]
Column: Capsule pack MGIII (4.6 × 150 mm, 3 μm, manufactured by Shiseido)
Solvent: A: 0.05 M acetic acid, 5% by volume acetonitrile aqueous solution, B: acetonitrile flow rate: 1 mL / min.

<Process to convert to numerical data>
In the chromatogram obtained as a result of the HPLC analysis, a total of 29 peaks presumed to exist in common with the specimen donated to the analysis were selected from the retention time and the UV-Vis spectrum, and each area value was calculated ( Agilent OpenLAB Chromatodata System version C.01.07 (manufactured by Agilent Technologies)). In addition, the sum of these 29 peak area values was calculated. Next, a peak area ratio obtained by dividing each of the 29 peak area values by the sum of the 29 peak area values was calculated and subjected to a step of learning by deep learning.

<The process of learning by deep learning and obtaining a quality prediction model>
A total of 175 samples, 137 samples and 38 samples in total, were selected from 212 samples of coffee solution made from Arabica roasted beans and 58 samples of coffee solution made from Robusta roasted beans. It was. A total of 95 samples were prepared as a test sample group, 75 samples of coffee liquid made from the remaining roasted Arabica beans and 20 samples of coffee liquid made from Robusta roasted beans.

  The numerical data set of the sample group for training was used for the deep belief network of all five layers, and the model which predicted the raw material roasted bean variety which is a response variable was created.

<Process to change parameters by matching with quality prediction model>
The numerical data set of the test sample group was applied to the created prediction model, and the variety of the roasted beans as the raw material of the coffee liquid was predicted and compared with the known roasted bean varieties. Next, the parameters of the prediction model were changed and the prediction model was tuned to obtain a tuned quality prediction model. The activation function of the model used ReLU, the batch size was 10, and the dropout ratio of the hidden layer was 0.5. The accuracy rate of the variety prediction of roasted beans in the test data using the already tuned model was 98.9%, and high prediction accuracy was obtained.

  R (version 3.3.0) was used for the step of learning by the deep learning and the step of predicting the variety of roasted beans.

(Example 6)
(Prediction of blend ratio of varieties)
<Step of pre-processing to obtain analysis sample>
18 Arabica roasted beans (9 Brazil samples, 9 Colombia samples, L value 15-24, obtained from Takasago coffee), 19 Robusta (Vietnam) roasted beans 19 samples (obtained from L value 16-23, Takasago coffee) 34 specimens of Arabica and Robusta blend roasted beans (15%, 25%, 40%, 50%, 60%, 75% and 80% by weight of Robusta roasted beans) ) Was ground, extracted and filtered with the same method as described in Example 5 to obtain 71 coffee extracts. The obtained coffee extract was pretreated in the same manner as described in Example 5 and subjected to HPLC analysis.
Separately, 10 coffee extracts (Brix 15-55, 4 types of Arabica, 3 types of Robusta, 3 types of blends of Arabica and Robusta were obtained from Takasago Coffee) diluted with Brix 1.5 with ion-exchanged water. Thereafter, pretreatment was performed in the same manner as in the method described in Example 1, and the resultant was subjected to HPLC analysis.

<Instrument analysis process and conversion to numerical data>
In the same manner as in the method described in Example 5, the peak area ratio was calculated for 71 samples of coffee extract and 10 samples of coffee extract.

<The process of learning by deep learning and obtaining a quality prediction model>
A total of 46 samples, 10 samples, 10 samples, 26 samples, were selected for training from 18 samples of Arabica seed extract, 19 samples of Robusta seed extract, and 34 samples of blend extract of Arabica and Robusta. Furthermore, 4 samples were selected for training, 2 samples at random from 4 samples of Arabica coffee extract and 3 samples of Robusta coffee extract. A total of 46 samples of the coffee extract and 4 samples of the coffee extract were used as a sample group for training. The remaining 25 samples of coffee extract and 6 samples of coffee extract, 31 samples in total, were used as test sample groups.
The 29 HPLC analysis peak area ratios of each sample belonging to the training sample group are used as input variables, and the blend ratio of Robusta species is used as a response variable. A model for predicting the blending ratio was created.

<Process to change parameters by matching with quality prediction model>
The 29 HPLC analysis peak area ratios of each sample belonging to the test sample group are used as input variables, calculation is performed using the created quality prediction model, and the blend ratio of the Robusta species, which is the response variable, is predicted. Goed and verified with blend ratios of known roasted bean varieties. Next, the parameters of the prediction model were changed and the prediction model was tuned to obtain a tuned quality prediction model. The activation function of the model used ReLU, the batch size was 10, and the dropout ratio of the hidden layer was 0.5. The prediction accuracy of the blend ratio of roasted bean varieties in the test data using the tuned model was evaluated with the coefficient of determination. As a result, a high prediction accuracy was obtained with a determination coefficient = 0.77.

(Example 7)
(Prediction of roasting degree)
<Pretreatment process, instrument analysis process and process for converting to numerical data>
Among the specimens shown in Example 5, the data of Arabica species was used.

<The process of learning by deep learning and obtaining a quality prediction model>
137 samples were randomly selected from 212 Arabica species and used as a sample group for Arabica training, and the remaining 75 samples were used as test sample groups. The training sample group, 29 HPLC analysis peak area ratios of each sample belonging to the training sample group are used as input variables, and the roasting degree (L value) of the coffee used is a response variable. A model for predicting the L value, which is a response variable, was created for the deep belief network of layers.

<Process to change parameters by matching with quality prediction model>
The 29 HPLC analysis peak area ratios of the respective samples belonging to the test sample group are used as input variables, the calculation is performed using the created quality prediction model, and the roasting degree (L value of coffee) as the response variable ) To match the known roasting degree of coffee. Next, the parameters of the prediction model were changed and the prediction model was tuned to obtain a tuned quality prediction model. The activation function of the model used ReLU, the batch size was 10, and the dropout ratio of the hidden layer was 0.5. The prediction accuracy of the roasting degree of the coffee of the test data using the tuned model was evaluated with the coefficient of determination. As a result, a high prediction accuracy was obtained with a coefficient of determination = 0.85.

According to the present invention, in quality evaluation of food and drink, it provides a flexible interpretation method that combines device analysis and sensory evaluation results, which has been difficult in the past. Predictable. By using this method, for example, by mixing a plurality of foods and drinks, it becomes possible to efficiently design higher quality foods and drinks.

Claims (7)

(A) A step of pretreating a plurality of foods and drinks of known quality to obtain an analysis sample;
(B) A step of obtaining data by subjecting the analysis sample obtained in (a) to one or more analyzes selected from instrumental analysis or sensory evaluation;
(C) a step of converting the data obtained in (b) into numerical values or image data;
(D) dividing the numerical value or image data obtained in (c) into training data and test data;
(E) learning the relationship between individual training data and known quality by deep learning to obtain a quality prediction model;
(F) A step of predicting quality by providing test data to the quality prediction model obtained in (e);
(G) collating the quality prediction result obtained in (f) with a known quality;
(H) changing the parameters of the quality prediction model;
(I) (e) (f) (g) (h) is repeated one or more times to tune the quality prediction model and obtain a tuned quality prediction model;
Quality prediction method for foods and drinks including The method for predicting quality of food and drink according to claim 1 further comprises:
(J) By selecting two or more numerical data of individual analysis samples obtained in (c) and proportionally calculating the data at an arbitrarily selected ratio, virtually blending at an arbitrarily selected ratio Obtaining numerical data of the analyzed sample;
(K) a step of converting the numerical data obtained in (j) into image data as necessary;
(L) Predicting the quality of the blended food and drink at an arbitrarily selected ratio by providing the numerical value or image data obtained in (k) to a tuned quality prediction model;
Quality prediction method for foods and drinks including The method for predicting quality of food and drink according to claim 1 further comprises:
(M) a step of pre-processing food and drink of unknown quality to obtain an analysis sample;
(N) A step of obtaining data by subjecting the analysis sample obtained in (m) to one or more types of analysis selected from instrumental analysis or sensory evaluation;
(O) converting the data obtained in (n) into numerical values or image data;
(P) predicting quality by subjecting the numerical value or image data obtained in (o) to a tuned quality prediction model;
Quality prediction method for foods and drinks including The method according to claims 1 to 3, wherein the quality according to claims 1 to 3 is selected from the quality sensory-assessed using human sensation and / or the quality analyzed by the instrument and / or the design quality of the product. . The method according to claims 1 to 4, wherein the deep learning is selected from a convolutional neural network, a deep neural network, a deep belief network, a recursive neural network. 6. The method according to claim 1, wherein the step of converting the analysis result into image data is a step of converting the obtained numerical data into a graph to form image data. The food / beverage products which reproduced the quality estimated by the method of Claim 1 thru | or 6.

Images