Koji production support system and program

Abstract

To use a learned model which is constituted by causing, a learning model being a method of machine learning to execute machine learning, for outputting a determination result with preferable reproduction ability to finished malted rice, for supporting malted rice production.SOLUTION: Malted rice 1 grain crushing data for teacher, and malted rice 1 grain class labels for teacher which are classified into a plurality of classes corresponding to the malted rice 1 grain crushing data, are used as teacher data, then, with respect to a learned malted rice 1 grain class estimation model which is constructed by causing a learning model being a method of the machine learning to execute machine learning, any one of the malted rice 1 grain crushing data is input for estimating a malted rice 1 grain class corresponding to any malted rice 1 grain crushing data, for outputting a determination result with preferable reproduction ability to finished malted rice, and for supporting malted rice production.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a koji-making assistance system and a koji-making assistance program. [Background technology]

[0002] Koji-making conditions vary widely depending on the desired quality of koji. Experts adjust the conditions depending on the quality of the koji, and the quality of the koji is judged based on analytical values ​​such as enzyme activity, as well as the expert's experience and intuition. However, differences in koji-making conditions naturally result in differences in the amount of ha-seed and the amount of ha-seed mixed in. Even when comparing individual koji grains, there is variation in the amount of ha-seed mixed in and the amount of ha-seed mixed in. Modifying the koji-making conditions for the next koji-making run based on these differences in koji-making conditions and the variation in the amount of ha-seed mixed in and the amount of ha-seed mixed in for each individual koji grain is an extremely difficult task. Furthermore, the enzyme activity values ​​of the koji after koji production is complete are also important factors in judging the quality of the koji. In addition to the amount of ha-seed mixed in and the amount of ha-seed mixed in, determining the koji-making conditions to obtain koji with the desired enzyme activity is even more difficult.

[0003] As a prior art, Patent Document 1 proposes a method (method and device for quantitative analysis of the breaking apart of koji substrate) in which image information from a test piece made of koji or koji slices is converted into an electrical signal and an image processing device is used to analyze koji production. The method of analyzing koji production disclosed in Patent Document 1 is said to be able to analyze the degree of breaking apart by calculating the area of ​​the entire koji substrate, the area around the breaking apart, and the area including the breaking apart, obtained from the image information of the test piece, and to determine the quality of the koji or the state of production of the koji. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 63-225151 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the method of analyzing koji production disclosed in Patent Document 1 simply calculates the area of ​​the entire koji substrate, the area around the broken-strain area, and the area including the broken-strain area, obtained from image information of the test piece, making it difficult to accurately judge the quality of koji using this method to the same extent as an expert.

[0006] The koji-making support system and program of the present invention have been developed to solve the above-mentioned problems, and aim to support koji-making by using a trained model constructed by running a machine learning model, which is a machine learning technique, to provide reproducible judgment results on the quality of koji. The koji-making support system and program of the present invention also aim to support koji-making by using a database for extracting candidate combinations created using the trained model to predict and provide koji-making conditions for obtaining koji of a target quality. [Means for solving the problem]

[0007] The koji-making support system of the present invention is characterized in that it inputs arbitrary koji-grain-breaking data into a trained koji-grain class estimation model constructed by running a learning model, which is a machine learning technique, using training data consisting of training data for single-grain koji breakage and training koji-grain class labels classified into multiple classes corresponding to the training data for single-grain koji breakage, and estimates the koji-grain class corresponding to the arbitrary single-grain koji breakage data.

[0008] The koji-making support system of the present invention is characterized in that it inputs teacher koji-making conditions and single-grain koji-breaking data obtained from koji made under the teacher koji-making conditions, and predicts the breakage distribution using a trained breakage distribution prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, using as training data teacher koji-making conditions and a single-grain koji-breaking distribution generated by converting the multiple single-grain koji classes output by the trained single-grain koji class estimation model into the ratio of each class.

[0009] The koji-making support system of the present invention is characterized in that it predicts koji multiple grain breakage data by inputting arbitrary koji-making conditions into a learned koji multiple grain breakage data prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, using teacher koji-making conditions and teacher koji multiple grain breakage data obtained from koji made under the teacher koji-making conditions as teacher data.

[0010] The koji-making support system of the present invention is characterized in that it predicts enzyme titers by inputting arbitrary koji-making conditions and a koji-breaking distribution corresponding to the arbitrary koji-making conditions into a first learned enzyme titer prediction model constructed by performing machine learning using teacher koji-making conditions and single-grain koji breakage data obtained from koji made under the teacher koji-making conditions as input, a teacher breakage distribution generated by converting multiple single-grain koji classes output by a learned single-grain koji class estimation model into the ratio of each class, and teacher enzyme titer analysis values ​​obtained by analyzing koji made under the teacher koji-making conditions.

[0011] The koji-making support system of the present invention is characterized in that it predicts enzyme titer by inputting arbitrary koji-making conditions and multiple koji-grain-breaking data corresponding to the arbitrary koji-making conditions into a second trained enzyme titer prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, using as training data the training koji-making conditions, the training koji multiple-grain-breaking data of koji made under the training koji-making conditions, and the training enzyme titer analysis value obtained by analyzing the koji made under the training koji-making conditions.

[0012] The koji-making support system of the present invention is characterized by creating a database for extracting candidate combinations of koji-making conditions and breaking distributions, which is composed of koji-making conditions for database configuration and breaking distributions for database configuration that are output by inputting the koji-making conditions for database configuration into a learned breaking distribution prediction model, searching the database for extracting candidate combinations of koji-making conditions and breaking distributions using one or more target data from the ratio data of any koji class that constitutes the breaking distribution as search conditions, and extracting candidate combinations of koji-making conditions and breaking distributions that satisfy the search conditions.

[0013] The koji-making support system of the present invention is characterized in that it creates a database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data of koji, which is composed of koji-making conditions for database configuration and multiple-grain-breaking data of koji for database configuration output by inputting the koji-making conditions for database configuration into a learned multiple-grain-breaking data prediction model, and searches the database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data of koji using one or more target data from any of the breakage data that make up the multiple-grain-breaking data as search conditions, and extracts candidate combinations of koji-making conditions and multiple-grain-breaking data of koji that satisfy the search conditions.

[0014] The koji-making support system of the present invention is characterized in that it creates a database for extracting candidate combinations of first koji-making conditions and enzyme titers, which is composed of koji-making conditions for database construction and enzyme titers for database construction output by inputting the koji-making conditions for database construction and the breakage distribution output by the learned breakage distribution prediction model using the koji-making conditions for database construction as input to a first trained enzyme titer prediction model, and searches the database for extracting candidate combinations of first koji-making conditions and enzyme titers using one or more target data from any of the enzyme titer data that make up the enzyme titers as search conditions, and extracts candidate combinations of koji-making conditions and enzyme titers that satisfy the search conditions.

[0015] The koji-making support system of the present invention is characterized in that it creates a database for extracting candidate combinations of second koji-making conditions and enzyme potencies, which is composed of koji-making conditions for database construction and enzyme potencies for database construction output by inputting the koji-making conditions for database construction and the multiple koji grain breaking data output by the learned koji multiple grain breaking data prediction model using the koji-making conditions for database construction as input to a second trained enzyme potency prediction model, and searches the database for extracting candidate combinations of second koji-making conditions and enzyme potencies using one or more target data from any of the enzyme potency data that make up the enzyme potency as search conditions, and extracts candidate combinations of koji-making conditions and enzyme potencies that satisfy the search conditions.

[0016] The koji-making assistance system according to the present invention is preferably characterized by increasing the number of target data items.

[0017] The koji-making support program of the present invention is characterized in that it uses teacher single-grain koji breaking data and teacher single-grain koji class labels classified into multiple classes corresponding to the teacher single-grain koji breaking data as training data, and inputs arbitrary single-grain koji breaking data into a trained single-grain koji class estimation model constructed using machine learning techniques, and causes a computer to execute a process to estimate the single-grain koji class corresponding to the arbitrary single-grain koji breaking data.

[0018] The koji-making support program of the present invention is characterized in that it inputs teacher koji-making conditions and a teacher-use breakage distribution generated by converting multiple single-grain koji classes output by a learned single-grain koji class estimation model into ratios of each class as training data, and causes a computer to input arbitrary koji-making conditions into a learned breakage distribution prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, and execute a process to predict the breakage distribution.

[0019] The koji-making support program of the present invention is characterized in that it inputs arbitrary koji-making conditions into a learned koji multiple grain breakage data prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, using teacher koji-making conditions and teacher koji multiple grain breakage data obtained from koji made under the teacher koji-making conditions as teacher data, and causes a computer to execute a process of predicting koji multiple grain breakage data.

[0020] The koji-making support program of the present invention is characterized in that it inputs teacher koji-making conditions and single-grain koji-breaking data obtained from koji made under the teacher koji-making conditions as input, and converts the multiple single-grain koji classes output by the learned single-grain koji class estimation model into ratios of each class using the teacher koji-making conditions and the teacher enzyme titer analysis values ​​obtained by analyzing koji made under the teacher koji-making conditions as teacher data.The program inputs arbitrary koji-making conditions and a single-grain koji-breaking distribution corresponding to the arbitrary koji-making conditions into a first learned enzyme titer prediction model constructed by performing machine learning on a learning model, which is a machine learning method, and causes a computer to execute a process of predicting the enzyme titer.

[0021] The koji-making support program of the present invention is characterized in that it inputs arbitrary koji-making conditions and data on multiple koji grains broken that correspond to the arbitrary koji-making conditions into a second learned enzyme potency prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, using as training data the training koji-making conditions, the training koji multiple grain breakage data of koji made under the training koji-making conditions, and the training enzyme potency analysis value obtained by analyzing the koji made under the training koji-making conditions, and causes a computer to execute a process of predicting the enzyme potency.

[0022] The koji-making support program of the present invention is characterized in that it creates a database for extracting candidate combinations of koji-making conditions and breaking distributions, which is composed of koji-making conditions for database configuration and breaking distributions for database configuration that are output by inputting the koji-making conditions for database configuration into a learned breaking distribution prediction model, and searches the database for extracting candidate combinations of koji-making conditions and breaking distributions using one or more target data from the ratio data of any koji class that constitutes the breaking distribution as search conditions, and causes a computer to execute a process to extract candidate combinations of koji-making conditions and breaking distributions that satisfy the search conditions.

[0023] The koji-making support program of the present invention is characterized in that it creates a database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data of koji, which is composed of koji-making conditions for database configuration and multiple-grain-breaking data of koji for database configuration output by inputting the koji-making conditions for database configuration into a learned multiple-grain-breaking data prediction model, and searches the database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data of koji using one or more target data from any of the breakage data that make up the multiple-grain-breaking data as search conditions, and extracts candidate combinations of koji-making conditions and multiple-grain-breaking data of koji that satisfy the search conditions.

[0024] The koji-making support program of the present invention is characterized in that it creates a database for extracting candidate combinations of first koji-making conditions and enzyme titers, which is composed of koji-making conditions for database construction and enzyme titers for database construction output by inputting the koji-making conditions for database construction and the breakage distribution output by the learned breakage distribution prediction model using the koji-making conditions for database construction as input to a first learned enzyme titer prediction model, and searches the database for extracting candidate combinations of first koji-making conditions and enzyme titers using one or more target data from any of the enzyme titer data that make up the enzyme titers as search conditions, and extracts candidate combinations of koji-making conditions and enzyme titers that satisfy the search conditions.

[0025] The koji-making support program of the present invention is characterized in that it creates a database for extracting candidate combinations of second koji-making conditions and enzyme potencies, which is composed of koji-making conditions for database construction and enzyme potencies for database construction output by inputting the koji-making conditions for database construction and the multiple-grain-breaking-of-koji data output by the learned multiple-grain-breaking-of-koji data prediction model using the koji-making conditions for database construction as input to a second trained enzyme potency prediction model, and searches the database for extracting candidate combinations of second koji-making conditions and enzyme potencies using one or more target data from any of the enzyme potency data that make up the enzyme potency as search conditions, and extracts candidate combinations of koji-making conditions and enzyme potencies that satisfy the search conditions.

[0026] The koji-making assistance program according to the present invention is preferably characterized in that the number of target data items is increased. [Effects of the Invention]

[0027] The koji-making support system and program according to the present invention use a trained model constructed by running a machine learning model, which is a machine learning technique, to provide reproducible judgment results on the quality of koji and support koji-making. Furthermore, the koji-making support system and program according to the present invention can support koji-making by using a database for extracting candidate combinations created using the trained model to predict and provide koji-making conditions for obtaining koji of a target quality. [Brief explanation of the drawings]

[0028] [Figure 1] FIG. 1 is a configuration diagram showing an example of a koji-making support system. [Figure 2] FIG. 10 is a diagram showing an example of data on the destruction of multiple grains. [Figure 3] FIG. 1 is a diagram showing an example of data on breaking apart one grain of koji. [Figure 4] FIG. 1 shows an example of the results of classification of data on the breaking down of one grain of koji. [Figure 5] FIG. 10 is a diagram showing an example of a method for classifying data on the breaking apart of one grain of koji. [Figure 6] FIG. 1 is a configuration diagram showing an example of a koji-making support system equipped with a trained model construction means. [Figure 7] FIG. 10 is a diagram illustrating an example of a test result according to the first embodiment. [Figure 8] FIG. 2 is a flow chart showing an example of a koji-making assistance program. [Figure 9] FIG. 10 is a diagram illustrating an example of contour detection data. [Figure 10] FIG. 10 is a diagram illustrating an example of pixel distribution data. [Figure 11] FIG. 10 is a diagram illustrating an example of a test result of the second embodiment. [Figure 12]FIG. 2 is a flow chart showing an example of a koji-making assistance program. DETAILED DESCRIPTION OF THE INVENTION

[0029] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The following describes the present invention with reference to the accompanying drawings, taking 11 embodiments (first to eleventh embodiments) as examples.

[0030] The koji-making support system of each embodiment is configured using a computer.

[0031] As shown in FIG. 1, the koji-making assistance system 100 of each embodiment includes a storage unit 101, an input unit 102, a control unit 103, and an output unit 104.

[0032] In the koji-making assistance system 100 of each embodiment, the storage means 101 is made up of a storage device such as a hard disk that records various data and various programs. The storage means 101 stores a trained model or database constructed by running a learning model, which is a machine learning method, through machine learning. In the koji-making assistance system of the present invention, the storage means may be a storage device such as a main memory that temporarily records various data and various programs, and is not limited to the configuration of each embodiment.

[0033] In the koji-making assistance system 100 of each embodiment, the input means 102 comprises an input device such as a camera or keyboard for inputting data. In the koji-making assistance system of the present invention, the input means is not limited to the configuration of each embodiment as long as it is a device for inputting data, and may also be a receiving device for receiving data from another device.

[0034] In the koji-making assistance system 100 of each embodiment, the control means 103 is composed of an arithmetic device such as a processor that executes commands from various programs and a control device. In the koji-making assistance system of the present invention, the control means is not limited to the configuration of each embodiment, as long as it is a device that executes commands from various programs and performs arithmetic processing, and may be, for example, a CPU, DSP, GPU, etc.

[0035] In the koji-making assistance system 100 of each embodiment, the output means 104 is composed of an output device such as a display or printer for outputting data. In the koji-making assistance system of the present invention, the output means is not limited to the configuration of each embodiment as long as it is a device for outputting data, and may also be a transmitting device for transmitting data to another device.

[0036] In the koji-making assistance system 100 of each embodiment, the storage means 101, input means 102, control means 103, and output means 104 may be configured as separate computers, or may be configured as a single integrated computer.

[0037] In each embodiment, the production of rice koji used in sake will be described as an example of koji production. In the present invention, the use of koji, the type of koji mold, or the culture substrate are not limited to those in each embodiment.

[0038] In each embodiment, the koji will be described as an example of koji that has been produced through each step of the koji-making process (hereinafter referred to as "finished koji"). However, the present invention is not limited to the configuration of each embodiment, and koji collected at an intermediate stage of each step of the koji-making process may also be used.

[0039] 1. First embodiment The koji-making support system of the first embodiment, which estimates the koji grain class from the data on the breaking down of one koji grain, will be described.

[0040] In the present invention, "single koji grain breaking data" refers to data obtainable from a single koji grain, including image data of the koji obtained by photographing the koji with a camera or the like and in a format that can be processed by a computer, data obtained by processing the image data of the koji, reflectance spectrum analysis values, absorption spectrum analysis values, and other various data obtainable from a single koji grain. In the present invention, examples of data obtained by processing image data of the koji include image data obtained by changing the size of the image data of the koji, contour detection data, and pixel distribution data. In the present invention, examples of contour detection data include the area of ​​a single koji grain, the area of ​​the broken part of the koji, the black ratio, perimeter, and number of colonies. In the present invention, examples of pixel distribution data include frequency data of pixel divisions and frequency distribution of pixel values.

[0041] In the first embodiment, a transillumination device having a flat plate and a light source, and a camera are used to obtain data on the breaking of the koji. The flat plate is a light-transmitting member. Multiple grains of koji are placed on the flat plate of the transillumination device, and light is irradiated onto the koji from a light source positioned below the plate. The koji are then photographed with a camera positioned above the transillumination device, thereby obtaining image data of the multiple grains of koji using transmitted light. The image data of the multiple grains of koji photographed by the camera in a format that can be processed by a computer is processed by a computer, and the processed image data of the multiple grains of koji is stored in a computer storage device consisting of a hard disk.

[0042] In the first embodiment, the data used for the breaking-sperm data is image data of koji obtained by photographing the koji with a camera using transmitted light and converted into a format that can be processed by a computer. However, in the present invention, an image using reflected light may also be used as the breaking-sperm data. Note that when an image using transmitted light is used, the areas where the koji fungus hyphae have penetrated into the koji can be detected as a dark black color. Therefore, in the present invention, it is preferable to use image data of koji obtained by photographing the koji with a camera using transmitted light and converted into a format that can be processed by a computer, or data obtained by processing the image data of the koji. Furthermore, the image of the koji may be a monochrome image, a grayscale image, or a color image.

[0043] Furthermore, in the present invention, the light used to acquire the breaking data is not limited to the configuration of the first embodiment, and may be any type such as visible light or ultraviolet light, and the wavelength band is not limited.

[0044] In the first embodiment, in order to obtain data obtained by processing image data of koji, which is the data of the broken koji, only koji with no cracked or chipped grains is used. However, the koji in the present invention is not limited to the configuration of the first embodiment.

[0045] As shown in Figure 2, image data of multiple grains of koji is saved in a storage device, which is the storage means 101 of a computer. The image data of multiple grains of koji is cut out for each grain of koji by the control means 103, and is separated into multiple pieces of processed image data of koji, which is data on the broken down single grains of koji. As shown in Figure 3, the processed image data of koji, which is data on the broken down single grains of koji, is processed data that has been adjusted by processes such as resizing, taking into account the calculation costs required for machine learning and the accuracy of the estimation results using the trained model.

[0046] In the first embodiment, the size of the data for breaking down a single grain of koji is changed to 32 x 32 pixels in length and width. However, in the present invention, the data for breaking down a single grain of koji may also be changed to 8 x 8 pixels in length and width, or 64 x 64 pixels in length and width, and is not limited to the configuration of the first embodiment. Note that if the size is too small, highly accurate estimation results cannot be obtained, and if the size is too large, the analysis time becomes long, the calculation cost increases, and highly accurate estimation results cannot be obtained. Therefore, the size of the data for breaking down a single grain of koji can be determined appropriately within an appropriate range.

[0047] In the first embodiment, the data for breaking down a single grain of koji is cut out as a rectangular image from the image data for multiple grains of koji, and then resized to a size of 32 x 32 pixels. However, in the present invention, the shape of the cutout process is not limited to the configuration in the first embodiment.

[0048] The koji-per-grain breaking data is evaluated by an expert based on the state of the koji mold hyphae spreading out (hereinafter referred to as "breaking around") and the state of the koji mold hyphae penetrating into the koji (hereinafter referred to as "breaking in"), and the koji is classified into one of four classes. In the first embodiment, the quality of the koji is evaluated from the perspective of whether it is suitable for use as koji for ginjo sake. In the following, the results of classifying the koji-per-grain breaking data into one of the four classes will be explained as the "one-grain koji class."

[0049] As shown in Figure 4, koji with a low degree of Hasei-makiri and Hasei-komi is classified as Class 0, koji with a moderate degree of Hasei-makiri and Hasei-komi is classified as Class 1, koji with a high degree of Hasei-makiri and Hasei-komi is classified as Class 2, and koji with an excessive degree of Hasei-makiri and Hasei-komi is classified as Class 3.

[0050] As shown in Figure 5, the single-grain koji class is classified taking into consideration the spread of a cluster of areas with a good degree of breakage (hereinafter referred to as a "colony"). In the first embodiment, a skilled person can evaluate the quality of koji by assuming that a colony that does not reach a certain area is not included in the breakage surrounding the evaluated grain.

[0051] In the example shown in Figure 5, koji a and koji b are considered to have the same image area per koji grain and the same total area of ​​the broken grain periphery. In koji a, areas with a good degree of broken grain periphery are relatively concentrated and localized. Each colony of koji a, where areas with a good degree of broken grain periphery are relatively concentrated and localized, meets a certain area and is therefore evaluated as being included in the broken grain periphery to be evaluated. As a result, koji a is classified into Class 1, which has a moderate degree of broken grain periphery and broken grain periphery. On the other hand, koji b has small broken grain peripheries scattered throughout. Each colony of koji b, where small broken grain peripheries are scattered throughout, does not meet a certain area and is therefore evaluated as not being included in the broken grain periphery to be evaluated. As a result, koji b is classified into Class 0, which has a low degree of broken grain periphery and broken grain periphery. In this way, even if the area of ​​a single koji grain and the total area of ​​the broken-seed surrounding area are the same, koji in which areas with a good degree of broken-seed are relatively concentrated and localized is evaluated as being more suitable for use as koji for ginjo sake, compared to koji in which small broken-seed surrounding areas are scattered throughout.

[0052] In the first embodiment, in addition to the above, the expert uses the ratio of the broken-stalk area to the other area, the color depth of the broken-stalk area, etc. as criteria for evaluating the quality of koji.

[0053] In the first embodiment, data on the breaking down of one grain of koji is classified into one of four classes. However, in the present invention, the number of classes may be any number as long as it is plural, and is not limited to the configuration of the first embodiment. Note that if the number of classes is too small, highly accurate estimation results cannot be obtained, and if the number of classes is too large, it becomes difficult to evaluate the quality of the koji; therefore, the number of classes is preferably about four.

[0054] In the first embodiment, data obtained by processing image data of koji, which is single-grain-breaking data obtained by the above-mentioned method, is used as one piece of teacher single-grain-breaking data, and the single-grain-breaking data classified by an expert using the above-mentioned method is used as one teacher single-grain-breaking class label. A set of teacher data is created from this teacher single-grain-breaking data and teacher single-grain-breaking class label. In this set of teacher data, the teacher single-grain-breaking data functions as an explanatory variable, and the teacher single-grain-breaking class label functions as a target variable. By repeating the creation of teacher data using the above procedure, a teacher dataset is created in which multiple pieces of teacher data are accumulated.

[0055] As shown in Figure 6, the koji-making assistance system 100 includes a trained model construction means 200. The trained model construction means 200 uses the above-mentioned teacher dataset to have a learning model, which is a machine learning technique, perform machine learning to construct a trained single-grain-koji class estimation model, which is a trained model. The trained single-grain-koji class estimation model constructed by the trained model construction means 200 is stored in the storage means 101.

[0056] In the first embodiment, the trained model construction means 200 is configured as an integrated unit using the same computer as the computer that configures the koji-making assistance system 100. However, in the present invention, the trained model construction means may be configured using a computer separate from the computer that configures the koji-making assistance system, and the trained koji single-grain class estimation model constructed by the trained model construction means may be configured to be transmitted to the koji-making assistance system via a network and stored in storage means.

[0057] In the first embodiment, a convolutional neural network (CNN) is used as a learning model, which is a machine learning technique. In the first embodiment, the CNN is configured with three layers: an input layer, an intermediate layer, and an output layer. The input layer passes input data to the intermediate layer. The intermediate layer has a convolution layer and a pooling layer, extracts features from data, and passes the extracted features to the output layer. The output layer determines the data to be output based on the extracted features.

[0058] In the first embodiment, the trained model construction means 200 inputs a training dataset into the CNN. The training single-grain koji breakage data input to the input layer is processed in the middle layer and output as an estimation result by the output layer. The CNN adjusts the parameters of the middle layer so that the estimation result output by the output layer approaches the training single-grain koji class label corresponding to the training single-grain koji breakage data input to the input layer. By repeatedly adjusting the parameters according to the above procedure using a large number of training datasets, a trained single-grain koji class estimation model is constructed that can output a single-grain koji class close to the corresponding training single-grain koji class label based on the input single-grain koji breakage data.

[0059] The first embodiment is configured to use CNN as a learning model, which is a machine learning technique. However, the present invention is not limited to the configuration of the first embodiment, and may be configured to use other algorithms that can perform class classification from image data of koji, for example.

[0060] In the first embodiment, the accuracy of the estimation results of the trained single-grain-koji class estimation model is evaluated using a test dataset. The test dataset is a combination of single-grain-koji crushed data and single-grain-koji class labels, just like the training dataset, but it was not used to build the trained single-grain-koji class estimation model.

[0061] In the first embodiment, the training dataset used to construct the trained single-grain-koji class estimation model contains 1600 items for each of classes 0 to 3, for a total of 6400 items. The accuracy of the estimation results from the trained single-grain-koji class estimation model is evaluated using a test dataset containing a total of 1600 items.

[0062] As shown in Figure 7, the accuracy of the trained single-grain koji class estimation model is evaluated using a test dataset. When 399 pieces of single-grain koji breakage data classified as class 0 are input into the trained single-grain koji class estimation model, the following results are output as the single-grain koji class: class 0: 393, class 1: 6, class 2: 0, and class 3: 0. When 407 pieces of single-grain koji breakage data classified as class 1 are input into the trained single-grain koji class estimation model, the following results are output as the single-grain koji class: class 0: 4, class 1: 367, class 2: 36, and class 3: 0. When 398 pieces of single-grain koji breakage data classified as class 2 are input into the trained single-grain koji class estimation model, the following results are output as the single-grain koji class: class 0: 0, class 1: 26, class 2: 357, and class 3: 15. When 396 pieces of data on broken koji grains classified as class 3 are input into the trained koji grain class estimation model, the output is the koji grain class: class 0: 0, class 1: 0, class 2: 26, and class 3: 370.

[0063] As shown in the above test results, the trained single-grain koji class estimation model is able to output the correct single-grain koji class with high accuracy based on the input single-grain koji breakage data. The accuracy rates for the above test dataset were approximately 98% for Class 0, approximately 90% for Class 1, approximately 90% for Class 2, and approximately 93% for Class 3, for an overall accuracy rate of approximately 93%. Furthermore, no errors were observed in the estimations made by the trained single-grain koji class estimation model in the first embodiment that were significantly different from the correct single-grain koji class. For example, in the above test results, no errors were observed in which koji breakage data that should have been determined to be Class 0 was determined to be Class 2 or 3.

[0064] As shown in Fig. 1, the koji-making assistance system 100 of the first embodiment includes a storage means 101, an input means 102, a control means 103, and an output means 104. The storage means 101 is a hard disk. The input means 102 is a camera. The control means 103 is a processor. The output means 104 is a display.

[0065] The storage means 101 stores a trained single-grain koji class estimation model constructed by running the learning model, which is the above-mentioned machine learning method, through machine learning. The input means 102 photographs the koji using the method using the above-mentioned transmitted illumination device, and acquires image data of the koji, which is single-grain koji destruction data. The control means 103 controls the calculations performed by the trained single-grain koji class estimation model. The output means 104 displays the single-grain koji class output by the trained single-grain koji class estimation model.

[0066] In the koji-making assistance system 100 of the first embodiment, the storage means 101 stores a koji-making assistance program 300. Furthermore, the control means 103 executes commands according to the koji-making assistance program 300.

[0067] As shown in Figure 8, the koji-making assistance program 300 of the first embodiment includes an input step 301, a control step 302, and an output step 303. The input step 301 inputs the single-grain-breaking data of koji acquired by the input means 102 to the trained single-grain-koji class estimation model stored in the storage means 101. The control step 302 controls the calculations by the trained single-grain-koji class estimation model, and causes the control means 103 to execute a process of outputting a single-grain-koji class based on the input single-grain-koji breakage data. The output step 303 acquires the single-grain-koji class output by the trained single-grain-koji class estimation model and displays it on the output means 104. By including these steps, the koji-making assistance program 300 can cause the koji-making assistance system 100 to execute a process of estimating a single-grain-koji class from the single-grain-koji breakage data.

[0068] The koji-making support system 100 of the first embodiment is configured to estimate the koji grain class from the koji grain breakage data using one trained koji grain class estimation model. However, the koji-making support system of the present invention may have multiple trained koji grain class estimation models and is not limited to the configuration of the first embodiment. For example, the koji-making support system of the present invention may be configured to use multiple trained koji grain class estimation models, each constructed with different hyperparameters. Here, the same koji grain breakage data may be input to the multiple trained koji grain class estimation models, and the most frequently output koji grain class from each output may be determined as the final estimation result.

[0069] The koji-making support system 100 of the first embodiment can input any single-grain koji breakage data into a trained single-grain koji class estimation model constructed by running a learning model, which is a machine learning technique, and can accurately estimate the single-grain koji class corresponding to the any single-grain koji breakage data, enabling stable judgment equivalent to the method used by an expert to evaluate the quality of koji.

[0070] In the koji-making support system 100 of the first embodiment, the trained koji single-grain class estimation model is constructed by running a learning model, which is a machine learning technique, on a training koji single-grain class label that has been classified by an expert using evaluation criteria such as the area around the broken grain, the area where the broken grain is packed in, the extent of the colony, and the color depth of the area where the broken grain is packed in, etc. Therefore, compared to conventional methods that determine the quality of koji by only calculating the area of ​​the entire koji substrate, the area around the broken grain, and the area where the broken grain is packed in, the quality of koji can be evaluated more accurately.

[0071] 2. Second embodiment A koji-making support system according to a second embodiment will be described, which estimates the koji grain class from data on the destruction of a single koji grain, which differs from the first embodiment.

[0072] In the second embodiment, the koji is photographed in the same manner as in the first embodiment, saved as image data of the koji, and adjusted by processes such as resizing to become processed data. In addition, a koji single grain class label, which is paired with data obtained by processing the image data of the koji, which is data on the destruction of the koji, as a set of training data, is determined by an expert in the same manner as in the first embodiment.

[0073] In the second embodiment, the computer constituting the koji-making assistance system 100 includes a contour detection unit and a pixel distribution data acquisition unit.

[0074] In the second embodiment, the contour detection unit acquires contour detection data from processed image data of koji. As shown in Fig. 9, in the second embodiment, the contour detection data consists of the area T of one koji grain, the area A of the broken part, the black ratio B, the perimeter L, and the number of colonies C.

[0075] In the second embodiment, a pixel value will be described as a value that represents the shade of a pixel's color as an integer between 0 and 255. In the second embodiment, when a pixel value is 0, the pixel is black, and when the pixel value is 255, the pixel is white.

[0076] The contour detection unit calculates the total number of pixels below a certain pixel value in the processed image data of the koji as the area T of one grain of koji. In the processed image data of the koji in the second embodiment, the pixel values ​​of the background are high and the pixel values ​​of the koji are low, and the two clearly differ in pixel value. Therefore, by regarding the parts with values ​​greater than a certain pixel value as the background and the parts with values ​​below the certain pixel value as the koji, it is possible to distinguish between the background and the koji. In the second embodiment, the threshold value of the pixel value for distinguishing between the background and the koji is set to 240, and binarization processing is performed.

[0077] As shown in Figure 9, the contour detection unit performs binarization processing on the processed koji image data, using a pixel value threshold of 60. The contour detection unit regards the black areas detected by the binarization processing as broken areas and draws a contour line around the contour of the broken areas. The contour detection unit calculates the total number of pixels inside the contour line as the broken area A. The contour detection unit calculates the black rate B by dividing the broken area A by the area T of one koji grain. The contour detection unit calculates the total number of pixels on the contour line as the perimeter L. The contour detection unit calculates the total number of colonies drawn by the contour line as the number of colonies C.

[0078] In the example shown in Figure 9, the contour detection unit obtains contour detection data consisting of the area of ​​one koji grain T: 769, the area of ​​the broken part A: 217, the black ratio B: 0.28, the circumference L: 82, and the number of colonies C: 2 based on data processed from the image data of the koji.

[0079] In the second embodiment, in order to draw the outline of the broken portion, the pixel value threshold is set to 60. However, the pixel value threshold in the present invention is not limited to the configuration of the second embodiment, and can be set to an appropriate value such as 40, 60, 80, 100, or 120.

[0080] In the second embodiment, the contour detection unit can determine that black portions detected by the binarization process that do not meet a certain area are not included in the broken portion.

[0081] In the second embodiment, the pixel distribution data acquisition unit acquires pixel distribution data from data obtained by processing image data of koji.

[0082] In the second embodiment, as shown in FIG. 10 , the pixel distribution data acquisition unit acquires pixel values ​​from each pixel included in the processed data of the koji image data, and creates a frequency distribution of the pixel values. Specifically, first, the pixel distribution data acquisition unit acquires pixel values ​​from 0 to 255 for each pixel included in the processed data of the koji image data. To prevent the color of the background part of the koji image from affecting the pixel distribution data, the pixel distribution data acquisition unit acquires data excluding pixel values ​​in a specific range from 241 to 255. Based on the pixel values ​​acquired from the pixels, the pixel distribution data acquisition unit classifies the values ​​from 0 to 240 into one of eight pixel divisions. Next, the pixel distribution data acquisition unit acquires the number of pixels classified into each pixel division as frequency data for that pixel division. The pixel distribution data acquisition unit creates a frequency distribution of pixel values ​​using the frequency data for these pixel divisions. The pixel distribution data in the present invention is not limited to that in the second embodiment and may be set as appropriate. For example, pixel values ​​from 0 to 255 may be used, or pixel segments may be selected and used according to the purpose. In the pixel distribution data acquisition unit of the present invention, if the background color of the koji image data is black, data may be acquired that excludes pixel values ​​in a specific range from 0 to 15 to prevent the color of the background color in the koji image data from affecting the pixel distribution data.

[0083] In the second embodiment, pixel distribution data is a frequency distribution of pixel values ​​created from frequency data of pixel segments. In the second embodiment, the pixel segments divided into eight segments will be described as H_1 to H_8. In the example shown in Fig. 10, the pixel distribution data acquisition unit acquires the following as pixel distribution data: [H_1: 76, H_2: 107, H_3: 99, H_4: 91, H_5: 155, H_6: 165, H_7: 39, H_8: 37].

[0084] In the second embodiment, each pixel corresponding to the image of koji in the processed koji image data is classified into one of eight pixel segments. However, in the present invention, each pixel corresponding to the image of koji may also be classified into one of sixteen pixel segments, and the number of pixel segments can be set to an appropriate value.

[0085] In the second embodiment, the contour detection unit and pixel distribution data acquisition unit are configured as an integrated unit using the same computer as the computer that configures the koji-making assistance system 100. In the present invention, the contour detection unit and pixel distribution data acquisition unit may also be configured using a computer that is separate from the computer that configures the koji-making assistance system 100.

[0086] In the second embodiment, the single-grain koji data is contour detection data and pixel distribution data obtained from a single grain of koji. In the second embodiment, the single-grain koji data obtained using the above-described method is used as one piece of training single-grain koji data, and the single-grain koji class classified by the expert corresponding to the single-grain koji data is used as one training single-grain koji class label. A set of training data is created using this training single-grain koji data and training single-grain koji class label. In this set of training data, the training single-grain koji data functions as an explanatory variable, and the training single-grain koji class label functions as a target variable. As shown in Table 1, by repeatedly creating training data using the above procedure, a training dataset is created that accumulates multiple pieces of training data.

[0087] [Table 1]

[0088] In the second embodiment, a trained koji grain class estimation model that estimates the koji grain class from the koji grain crushing data is constructed by the trained model construction means 200 using the above-mentioned training data set, as in the first embodiment, and is stored in the memory means 101 of the koji production support system 100.

[0089] The second embodiment is configured to use logistic regression (LGR) as a learning model, which is a machine learning technique. However, the present invention is not limited to the configuration of the second embodiment, and may be configured to use other algorithms that can perform class classification from numerical data, such as neural networks (NN), random forest classifiers (RFC), support vector machines (SVM), decision trees (DT), K-nearest neighbor algorithms (KNN), and naive Bayes (NB).

[0090] In the second embodiment, the training dataset used to build the trained single-grain-koji class estimation model uses 1600 items for each of classes 0 to 3, for a total of 6400 items. The accuracy of the estimation results by the trained single-grain-koji class estimation model is evaluated using a total of 1600 test datasets. In the second embodiment, the test datasets are not used to build the trained single-grain-koji class estimation model, as in the first embodiment.

[0091] As shown in Figure 11, the accuracy of the estimation results of the trained single-grain koji class estimation model is evaluated using a test dataset. When 399 pieces of data on single-grain koji breakage classified as class 0 are input into the trained single-grain koji class estimation model, the following results are output as the single-grain koji class: class 0: 395, class 1: 4, class 2: 0, and class 3: 0. When 407 pieces of data on single-grain koji breakage classified as class 1 are input into the trained single-grain koji class estimation model, the following results are output as the single-grain koji class: class 0: 5, class 1: 332, class 2: 70, and class 3: 0. When 398 pieces of data on single-grain koji breakage classified as class 2 are input into the trained single-grain koji class estimation model, the following results are output as the single-grain koji class: class 0: 0, class 1: 67, class 2: 291, and class 3: 40. When 396 pieces of data on broken koji grains classified as class 3 are input into the trained koji grain class estimation model, the output is the koji grain class: class 0: 0, class 1: 1, class 2: 18, and class 3: 377.

[0092] As shown in the above test results, the trained single-grain-koji class estimation model is able to output the correct single-grain koji class with high accuracy based on the input data on broken single-grain koji. The accuracy rates for the above test dataset were approximately 99% for Class 0, approximately 82% for Class 1, approximately 73% for Class 2, and approximately 95% for Class 3, for an overall accuracy rate of approximately 87%. The estimations made by the trained single-grain-koji class estimation model in the second embodiment rarely showed errors that were significantly different from the correct single-grain koji class.

[0093] As shown in Fig. 1, the koji-making assistance system 100 of the second embodiment, like the first embodiment, comprises a storage means 101, an input means 102, a control means 103, and an output means 104. Also, as shown in Fig. 8, the koji-making assistance program 300 of the second embodiment, like the first embodiment, comprises an input step 301, a control step 302, and an output step 303. The koji-making assistance program 300 is stored in the storage means 101 of the koji-making assistance system 100, and can cause the koji-making assistance system 100 to execute a process of estimating the koji grain class from the koji grain breaking data.

[0094] In the second embodiment, when the input means 102 is a camera, the koji-making assistance system 100 includes the above-mentioned contour detection unit and pixel distribution data acquisition unit. In the second embodiment, the koji-making assistance system 100 can acquire koji-grain breaking data consisting of contour detection data and pixel distribution data based on image data of koji photographed by the input means 102.

[0095] The present invention is not limited to the second embodiment. When the input means 102 is a keyboard, the koji-making assistance system 100 can acquire data on breaking down one grain of koji, which consists of contour detection data and pixel distribution data input as numerical values ​​by the input means 102.

[0096] The koji-making support system 100 of the present invention may be configured to estimate the koji grain class from the single-grain koji breakage data using multiple trained single-grain koji class estimation models constructed using different methods for each class, such as a trained single-grain koji class estimation model constructed using LGR, a trained single-grain koji class estimation model constructed using NN, or a trained single-grain koji class estimation model constructed using RFC. Alternatively, multiple trained single-grain koji class estimation models constructed using the same method with different hyperparameters may be used. Here, the multiple trained single-grain koji class estimation models may be configured to input single-grain koji breakage data obtained from a common koji, and the most frequently output koji grain class from each output may be determined as the final estimation result.

[0097] 3. Third embodiment A koji-making support system for predicting the distribution of broken spermatozoa from koji-making conditions according to the third embodiment will be described.

[0098] In the third embodiment, the hazei distribution is expressed by the proportion of each class in a plurality of koji grains classified into one of the four classes described above. For example, if 100 grains of koji contain 10 grains of koji classified as class 0, 40 grains of koji classified as class 1, 30 grains of koji classified as class 2, and 20 grains of koji classified as class 3, the hazei distribution in the 100 grains of koji is [class 0: 10%, class 1: 40%, class 2: 30%, class 3: 20%]. The hazei distribution of the final koji is one of the indicators in koji production and indicates the quality of the koji.

[0099] In the third embodiment, the koji-making conditions include the seed amount, moisture, and product temperature.

[0100] In the third embodiment, the seeding amount is a value measured in the process of sprinkling seed koji on steamed rice. The seeding amount is calculated by converting the weight of seed koji per weight of original white rice. The seeding amount in koji production is one of the factors that affect the circulation of koji.

[0101] In the third embodiment, the seeding amount is a quantified weight of seed koji per weight of original white rice. However, the seeding amount in the present invention is not limited to the configuration of the third embodiment, and the weight of seed koji may be the weight of koji mold spores only, or the weight including a bulking agent.

[0102] In the third embodiment, the moisture content is the value measured immediately after sprinkling seed koji onto the steamed rice. The moisture content is calculated by comparing the weight of the steamed rice with the weight of the steamed rice in a dried state. Since the activity of koji mold differs depending on the amount of water contained in the steamed rice, which is the culture substrate, moisture content in koji production is one of the factors that affect the distribution of the haze and the inclusion of the haze.

[0103] In the third embodiment, the moisture content is determined by a value measured immediately after seed koji is sprinkled on the steamed rice. However, the present invention is not limited to the configuration of the third embodiment, and may alternatively be determined by a value measured during the soaking of polished rice, or by the moisture content of steamed rice cooled after steaming. Note that the moisture content of steamed rice varies depending on the steaming conditions, cooling conditions, and ambient air conditions, and therefore, in the present invention, it is preferable to use the moisture content measured immediately after seed koji is sprinkled on the steamed rice, i.e., at the start of koji production.

[0104] In the third embodiment, the product temperature is a value measured at each step of the koji-making process. The product temperature is measured by a temperature sensor inserted into the steamed rice. The product temperature measured by the temperature sensor is automatically recorded in a storage device, which is the computer's storage means 101, every time a predetermined time has passed. The activity of koji mold varies depending on the temperature and humidity of the environment, and the product temperature during koji-making is one of the factors that ultimately affect the distribution of the haze and the inclusion of the haze.

[0105] The koji-making processes given as examples in the third embodiment can be divided into a bed process, which is the first half, and a shelf process, which is the second half. The bed process is a process that takes a total of about 22 hours, and the shelf process is a process that takes a total of about 24 hours. In the third embodiment, the product temperatures set as koji-making conditions are the product temperatures at 1, 5, 15, and 22 hours in the bed process, and the product temperatures at 1, 5, 10, 15, 20, and 24 hours in the shelf process, respectively.

[0106] In the third embodiment, the product temperature set as a koji-making condition is configured to use the product temperature over the elapsed time described above. However, the present invention is not limited to the configuration of the third embodiment, and the product temperature set as a koji-making condition may be configured to use product temperatures measured every hour or every few minutes, and the measurement interval can be set appropriately within an appropriate range. Note that if the measurement interval is shortened, the data size of the product temperature set as a koji-making condition increases, and if the measurement interval is lengthened, there will be problems such as less product temperature data set as a koji-making condition. Therefore, it is preferable to set an appropriate interval according to the environment in which the trained model is constructed.

[0107] In the third embodiment, the seed amount, moisture, and product temperature are used as the koji-making conditions. However, the present invention is not limited to the configuration of the third embodiment, and the koji-making conditions may also be configured to use information such as the type and total weight of raw materials, information such as the type of seed koji, information on the production process up to koji-making, the temperature of the koji-making chamber, the humidity of the koji-making chamber, the number of maintenance sessions, the timing of maintenance, etc. For example, if the raw material is white rice, the production year, variety, rice polishing ratio, etc. may be used as information on the type of raw material.

[0108] In the koji-making process of the third embodiment, the results of the koji-making conditions, including the seed amount, moisture, and product temperature, are saved as a series of data in a storage device, which is the computer's storage means 101. Furthermore, a portion of the koji produced under these koji-making conditions is collected and converted into multiple single-grain koji-breaking data using a method similar to that of the first embodiment. Furthermore, each of these multiple single-grain koji-breaking data is input into a trained single-grain koji class estimation model, and multiple single-grain koji classes are output. By converting the multiple single-grain koji classes obtained using the above procedure into the ratios of each class, a distribution of the breakage data of the koji produced under these koji-making conditions is generated. In the present invention, the koji produced under the koji-making conditions is not limited to the third embodiment, and may be converted into multiple single-grain koji-breaking data using a method similar to that of the second embodiment, and multiple single-grain koji classes may be output and converted into the ratios of each class.

[0109] In the third embodiment, 100 unbroken grains are collected from the koji produced, and the above-mentioned broken grain distribution is generated. However, the present invention is not limited to the configuration of the third embodiment, and the number of koji to be collected can be set as appropriate. Note that if a small number of koji are collected, the influence of sampling error becomes large, and if a large number of koji are collected, the amount of work increases. Therefore, the number of koji to be collected can be set as appropriate taking these factors into consideration.

[0110] In the third embodiment, the koji-making conditions obtained by the above-described method are used as one teacher koji-making condition, and the breakage distribution obtained by the above-described method from the koji produced under the koji-making conditions is used as one teacher breakage distribution. A set of teacher data is created using these teacher koji-making conditions and teacher breakage distribution. In this set of teacher data, the teacher koji-making conditions function as explanatory variables, and the teacher breakage distribution functions as a target variable. By repeating the creation of teacher data using the above procedure, a teacher dataset containing multiple pieces of teacher data is created.

[0111] In the third embodiment, a trained breakage distribution prediction model that predicts the breakage distribution from the koji-making conditions is constructed by the trained model construction means 200 using the above-mentioned training data set, as in the first embodiment, and is stored in the memory means 101 of the koji-making support system 100.

[0112] The third embodiment is configured to use a neural network (NN) or linear regression (LR) as a learning model, which is a machine learning technique. However, the present invention is not limited to the configuration of the third embodiment, and may be configured to use other regression algorithms, such as a random forest regressor (RFR), support vector regression (SVR), or decision tree (DT).

[0113] In the third embodiment, since there are four types of breakage distributions, the learning model, which is a machine learning method, has four objective variables. When configured using NN, the trained model construction means 200 will construct one trained breakage distribution prediction model for the four objective variables. When configured using LR, the trained model construction means 200 will construct four trained breakage distribution prediction models for the four objective variables, which is the same number.

[0114] In the third embodiment, the training data set used to build the trained breakage distribution prediction model uses 180 pieces of training data. In addition, the accuracy of the trained breakage distribution prediction model is evaluated using 20 test data sets. In the third embodiment, the test data set, like the first embodiment, is made up of data that was not used to build the trained breakage distribution prediction model.

[0115] As shown in Table 2, the accuracy of the prediction results of the trained breakage distribution prediction model is evaluated using a test data set. In the third embodiment, the coefficient of determination R is used as an evaluation index for the prediction results. 2 is used.

[0116] [Table 2]

[0117] As shown in the above test results, the trained breakage distribution prediction model constructed using NN is capable of outputting highly accurate breakage distribution prediction values ​​based on the input koji-making conditions. Furthermore, as shown in Table 2, when the training data set in the third embodiment is used, the prediction accuracy of the trained breakage distribution prediction model constructed using NN is higher than the prediction accuracy of the trained breakage distribution prediction model constructed using LR in all four classes. If an additional training data set is added to the training data set used in the third embodiment and machine learning is performed, the prediction accuracy of the trained breakage distribution prediction model may be improved. The present invention is not limited to the third embodiment, and a trained breakage distribution prediction model with high prediction accuracy for the four classes may be used, or a trained breakage distribution prediction model with high prediction accuracy for each class may be appropriately selected and used.

[0118] As shown in FIG. 1, the koji-making assistance system 100 of the third embodiment includes a storage unit 101, an input unit 102, a control unit 103, and an output unit 104, similar to the first embodiment. Also, as shown in FIG. 8, the koji-making assistance program 300 of the third embodiment includes an input step 301, a control step 302, and an output step 303, similar to the first embodiment. The koji-making assistance program 300 is stored in the storage unit 101 of the koji-making assistance system 100. The input step 301 inputs the koji-making conditions acquired by the input unit 102 into the trained breakage distribution prediction model stored in the storage unit 101. The control step 302 controls the calculation by the trained breakage distribution prediction model and causes the control unit 103 to execute a process of outputting the breakage distribution based on the input koji-making conditions. The output step 303 acquires the breakage distribution output by the trained breakage distribution prediction model and displays it on the output unit 104. By providing these steps, the koji-making assistance program 300 can cause the koji-making assistance system 100 to execute the process of predicting the distribution of broken grains from the koji-making conditions.

[0119] In the third embodiment, the input means 102 is a keyboard, and the koji-making assistance system 100 can acquire the koji-making conditions input as numerical values ​​by the input means 102.

[0120] The koji-making support system 100 of the present invention may be configured to predict the breakage distribution from koji-making conditions using multiple trained breakage distribution prediction models constructed using different methods, such as a trained breakage distribution prediction model constructed using NN, a trained breakage distribution prediction model constructed using LR, or a trained breakage distribution prediction model constructed using RFR. It may also be configured to use multiple trained breakage distribution prediction models constructed using the same method, each with different hyperparameters. Here, common koji-making conditions may be input to multiple trained breakage distribution prediction models, and the average value may be calculated from the output breakage distributions, and the average value may be determined as the final prediction result.

[0121] In the koji-making support system 100 of the third embodiment, the teacher-use breakage distribution, which is part of the teacher data used to construct the trained breakage distribution prediction model, is generated by converting the multiple koji single-grain classes output by the trained koji single-grain class estimation model into the ratio of each class. The trained breakage distribution prediction model constructed using the teacher-use breakage distribution generated by converting the multiple koji single-grain classes output by the trained koji single-grain class estimation model into the ratio of each class, can predict an accurate breakage distribution based on the input koji-making conditions.

[0122] The third embodiment of the koji-making support system 100 can predict the distribution of broken spores from any koji-making conditions using a learned broken spore distribution prediction model constructed by performing machine learning on a learning model, which is a machine learning technique.This allows the user to simulate the results of koji-making without actually making koji.

[0123] 4. Fourth embodiment A koji-making support system for predicting data on multiple koji grains breaking from koji-making conditions according to the fourth embodiment will be described.

[0124] In the fourth embodiment, the data on the destruction of multiple grains of koji is the average value of multiple data sets of destruction data that can be obtained from one grain of koji, and is the average value of the contour detection data of multiple grains of koji. In the fourth embodiment, the contour detection data of koji is obtained by the same method as in the second embodiment.

[0125] In the fourth embodiment, the average values ​​of the contour detection data for multiple koji grains consist of four types: the average value of the area of ​​the broken portion of the multiple koji grains, the average value of the black percentage of the multiple koji grains, the average value of the circumference of the multiple koji grains, and the average value of the number of colonies of the multiple koji grains. For example, as shown in Table 1, if the black percentage data for 100 koji grains, which is the contour detection data for koji that can be obtained from a single koji grain, is collected and added together to obtain a value of "40 (=0.28 + 0.39 + 0.07 + )", the average value of the black percentage of the multiple koji grains is calculated by dividing this value by the number of koji grains (100). In the present invention, the multiple koji grain breaking data is not limited to that of the fourth embodiment and may be the average value of various data that can be obtained from multiple koji grains, such as pixel distribution data for multiple koji grains, reflectance spectrum analysis values ​​for multiple koji grains, or absorption spectrum analysis values ​​for multiple koji grains.

[0126] In the present invention, the various data of the koji multiple grain breaking data are not limited to those in the fourth embodiment, and can be set appropriately as the average value, mode, median, etc. of multiple grains. The data of the multiple koji breaking data of the final koji is one of the indicators in koji production and represents the quality of the koji. However, in the present invention, the data of the multiple koji breaking data does not include the distribution of the breaking data.

[0127] In the fourth embodiment, the koji-making conditions are the seed amount, moisture, and product temperature, as in the third embodiment.

[0128] In the koji-making process of the fourth embodiment, the koji-making conditions are stored as a series of data in a storage device, which is the computer's storage means 101, as in the third embodiment. A portion of the koji made under the koji-making conditions is collected and converted into koji multiple grain breaking data that can be obtained from multiple grains of koji using the same method as in the second embodiment.

[0129] In the fourth embodiment, 100 unbroken grains of koji are collected from the koji produced and converted into the multiple grain broken koji data that can be obtained from the multiple grains of koji described above. However, the present invention is not limited to the configuration of the fourth embodiment, and the number of koji to be collected can be set to an appropriate number as needed.

[0130] In the fourth embodiment, the koji-making conditions obtained by the above-described method are used as one teacher koji-making condition, and the data on multiple koji grains broken obtained by the above-described method from koji made under those koji-making conditions are used as one teacher koji multiple grain broken data. A set of teacher data is created from these teacher koji-making conditions and teacher multiple koji grain broken data. In this set of teacher data, the teacher koji-making conditions function as explanatory variables, and the teacher multiple koji grain broken data function as a target variable. By repeating the creation of teacher data using the above procedure, a teacher dataset containing multiple teacher data is created.

[0131] In the fourth embodiment, a trained koji multiple grain breaking data prediction model that predicts koji multiple grain breaking data from koji-making conditions is constructed by the trained model construction means 200 using the above-mentioned training data set, as in the first embodiment, and is stored in the memory means 101 of the koji-making support system 100.

[0132] The fourth embodiment is configured to use linear regression (LR) as a learning model, which is a machine learning technique. However, the present invention is not limited to the configuration of the fourth embodiment, and may be configured to use other regression algorithms, such as neural network (NN), random forest regressor (RFR), support vector regression (SVR), and decision tree (DT).

[0133] In the fourth embodiment, there are four types of koji multiple grain breakage data, so there are four LR objective variables. Because LR is used, the trained model construction means 200 will construct four trained koji multiple grain breakage data prediction models for the same number of objective variables. The present invention is not limited to the fourth embodiment, and if a configuration is used in which NN is used as a training model, which is a machine learning method, the trained model construction means 200 will construct one trained koji multiple grain breakage data prediction model for the four objective variables.

[0134] In the fourth embodiment, 180 training datasets are used to build a trained koji multiple grain breakage data prediction model. The accuracy of the trained koji multiple grain breakage data prediction model is evaluated using 20 test datasets. In the fourth embodiment, the test datasets are not used to build the trained koji multiple grain breakage data prediction model, as in the first embodiment.

[0135] As shown in Table 3, the accuracy of the prediction results of the trained koji multiple grain breaking data prediction model is evaluated using a test data set. In the fourth embodiment, the coefficient of determination R is used as an evaluation index for the prediction results. 2 is used.

[0136] [Table 3]

[0137] As shown in the test results above, the trained koji multiple grain breaking data prediction model is able to accurately predict the koji multiple grain breaking data, including the broken area A, black rate B, circumference L, and number of colonies C, based on the input koji-making conditions.

[0138] As shown in FIG. 1, the koji-making assistance system 100 of the fourth embodiment includes a storage unit 101, an input unit 102, a control unit 103, and an output unit 104, similar to the first embodiment. Also, as shown in FIG. 8, the koji-making assistance program 300 of the fourth embodiment includes an input step 301, a control step 302, and an output step 303, similar to the first embodiment. The koji-making assistance program 300 is stored in the storage unit 101 of the koji-making assistance system 100. In the input step 301, the koji-making conditions acquired by the input unit 102 are input to the trained koji multiple grain breakage data prediction model stored in the storage unit 101. In the control step 302, the trained koji multiple grain breakage data prediction model is controlled, and the control unit 103 is caused to execute a process of outputting koji multiple grain breakage data based on the input koji-making conditions. In the output step 303, the trained koji multiple grain breakage data output by the trained koji multiple grain breakage data prediction model is acquired and displayed on the output unit 104. By providing these steps, the koji-making assistance program 300 can cause the koji-making assistance system 100 to execute a process for predicting data on multiple koji grains breaking from the koji-making conditions.

[0139] In the fourth embodiment, the input means 102 is a keyboard, and the koji-making assistance system 100 can acquire the koji-making conditions input as numerical values ​​by the input means 102.

[0140] The koji-making support system 100 of the present invention may be configured to predict koji multiple grain breakage data from koji-making conditions using multiple trained koji multiple grain breakage data prediction models constructed using different techniques, such as a trained koji multiple grain breakage data prediction model constructed using NN, a trained koji multiple grain breakage data prediction model constructed using LR, or a trained koji multiple grain breakage data prediction model constructed using RFR. Alternatively, multiple trained koji multiple grain breakage data prediction models constructed using the same technique with different hyperparameters may be used. Here, common koji-making conditions may be input to the multiple trained koji multiple grain breakage data prediction models, and the average value may be calculated from the output koji multiple grain breakage data, and this average value may be determined as the final prediction result.

[0141] The koji-making support system 100 of the fourth embodiment can predict multiple koji grain breakage data from any koji-making conditions using a trained koji multiple grain breakage data prediction model constructed by running machine learning on a learning model, which is a machine learning technique. This allows the user to simulate the results of koji-making without actually making koji.

[0142] 5. Fifth embodiment A koji-making support system according to a fifth embodiment that predicts the enzyme activity of koji from the koji-making conditions and the distribution of broken spermatozoa will be described.

[0143] In the fifth embodiment, the enzyme activity is determined by measuring the enzyme activity of α-amylase and glucoamylase from the koji produced by a predetermined analytical method. The enzyme activity of the final koji is one of the indicators in koji production and indicates the quality of the koji.

[0144] The fifth embodiment is configured to use α-amylase and glucoamylase as enzyme titers, however, the present invention is not limited to the configuration of the fifth embodiment, and may be configured to use enzyme titers of acid protease, acid carboxypeptidase, etc. in addition to the above enzyme titers.

[0145] In the koji-making process of the fifth embodiment, the koji-making conditions are stored as a series of data in a storage device, which is the computer's storage means 101, as in the third embodiment. A portion of the koji produced under these koji-making conditions is sampled and converted into a plurality of single-grain koji-breaking data using the same method as in the first embodiment. Each of these single-grain koji-breaking data is then input into a trained single-grain koji class estimation model, which outputs a plurality of single-grain koji classes. The multiple single-grain koji classes obtained using the above procedure are converted into a ratio of each class, generating a distribution of koji-breaking data for the koji produced under these koji-making conditions. At the same time, a portion of the koji produced under these koji-making conditions is sampled, and the enzyme activity is measured using a predetermined analytical method.

[0146] In the fifth embodiment, the koji-making conditions obtained by the above-described method are used as one teacher koji-making condition, the breakage distribution obtained by the above-described method from koji produced under those koji-making conditions is used as one teacher breakage distribution, and the enzyme titer obtained by the above-described method from koji produced under those koji-making conditions is used as one teacher enzyme titer analysis value. A set of teacher data is created from these teacher koji-making conditions, the teacher breakage distribution, and the teacher enzyme titer analysis value. In this set of teacher data, the data consisting of the teacher koji-making conditions and the teacher breakage distribution function as explanatory variables, and the teacher enzyme titer analysis value functions as a target variable. By repeating the creation of teacher data using the above procedure, a teacher dataset containing multiple teacher data is created.

[0147] In the fifth embodiment, a first trained enzyme potency prediction model that predicts the enzyme potency of koji from the koji-making conditions and the distribution of broken spermatozoa is constructed by the trained model construction means 200 using the above-mentioned training data set, as in the first embodiment, and is stored in the memory means 101 of the koji-making support system 100.

[0148] The fifth embodiment is configured to use LR as a learning model, which is a machine learning technique. However, the present invention is not limited to the configuration of the fifth embodiment, and may be configured to use other regression algorithms, such as NN, RFR, SVR, and DT.

[0149] In the first trained enzyme potency prediction model of the fifth embodiment, data on koji-making conditions and breakage distribution are explanatory variables, and the enzyme potency is the objective variable. In the fifth embodiment, since there are two types of enzyme potency, α-amylase and glucoamylase, there are two objective variables for LR. Furthermore, since LR is used, the trained model construction means 200 will construct two first trained enzyme potency prediction models, the same number of variables, for the two objective variables. The present invention is not limited to the fifth embodiment, and if a configuration is adopted in which NN is used as a learning model, which is a machine learning method, the trained model construction means 200 will construct one first trained enzyme potency prediction model for the two objective variables.

[0150] In the fifth embodiment, 180 training datasets are used to construct the first trained enzyme titer prediction model. The accuracy of the first trained enzyme titer prediction model is evaluated using 20 test datasets. In the fifth embodiment, the test datasets are not used to construct the first trained enzyme titer prediction model, as in the first embodiment.

[0151] As shown in Table 4, the accuracy of the prediction results of the first trained enzyme titer prediction model is evaluated using a test data set. In the fifth embodiment, the coefficient of determination R 2 is used.

[0152] [Table 4]

[0153] As shown in the above test results, the first trained enzyme potency prediction model is capable of outputting highly accurate predicted values ​​of enzyme potency based on the input koji-making conditions and breaking-seed distribution.

[0154] As shown in FIG. 1, the koji-making assistance system 100 of the fifth embodiment includes a storage unit 101, an input unit 102, a control unit 103, and an output unit 104, similar to the first embodiment. Also, as shown in FIG. 8, the koji-making assistance program 300 of the fifth embodiment includes an input step 301, a control step 302, and an output step 303, similar to the first embodiment. The koji-making assistance program 300 is stored in the storage unit 101 of the koji-making assistance system 100. The control step 302 controls the processing of image data of koji, which is data on the single-grain koji breakage acquired by the input unit 102, and causes the control unit 103 to execute a process of processing the output image data of koji and storing the processed data in the storage unit 101. The input step 301 inputs the processed image data of koji stored in the storage unit 101 into the trained single-grain koji class estimation model stored in the storage unit 101. The control step 302 controls the calculation using the trained koji-grain class estimation model, converts the output koji-grain classes into the ratios of each class, and causes the control means 103 to execute a process of storing the broken-grain distribution generated by the conversion in the storage means 101. The input step 301 inputs the koji-making conditions acquired by the input means 102 and the broken-grain distribution stored in the storage means 101 into the first trained enzyme potency prediction model stored in the storage means 101. The control step 302 controls the calculation using the first trained enzyme potency prediction model, and causes the control means 103 to execute a process of outputting the enzyme potency based on the input koji-making conditions and broken-grain distribution. The output step 303 acquires the enzyme potency output by the first trained enzyme potency prediction model and displays it on the output means 104. By including each of these steps, the koji-making assistance program 300 can cause the koji-making assistance system 100 to execute a process of predicting the enzyme potency from the koji-making conditions and broken-grain distribution.

[0155] In the fifth embodiment, the input means 102 for acquiring image data of koji, which is data on the breaking of a single grain of koji, is a camera, and the koji-making assistance system 100 can acquire the breaking distribution generated by processing the image data of multiple koji grains acquired by the input means 102. Furthermore, the input means 102 for acquiring koji-making conditions is a keyboard, and the koji-making assistance system 100 can acquire the koji-making conditions input as numerical values ​​by the input means 102.

[0156] The koji-making support system 100 of the present invention may be configured to predict enzyme titers from koji-making conditions and the breakage distribution using multiple first trained enzyme titer prediction models constructed using different techniques, such as a first trained enzyme titer prediction model constructed using neural networks, a first trained enzyme titer prediction model constructed using LR, or a first trained enzyme titer prediction model constructed using RFR. Alternatively, multiple first trained enzyme titer prediction models constructed using the same technique but with different hyperparameters may be used. Here, the system may be configured to input the same koji-making conditions and the breakage distribution to the multiple first trained enzyme titer prediction models, calculate an average value from the enzyme titers output from each, and determine the average value as the final prediction result.

[0157] The koji-making support system 100 of the fifth embodiment can predict enzyme titer from any koji-making condition and the distribution of broken spermatozoa corresponding to that condition using a first trained enzyme titer prediction model constructed by running machine learning on a learning model, which is a machine learning technique. This allows the user to predict the enzyme titer without measuring it with a specified analytical method immediately after koji-making is completed, and enables fine-tuning of mash management based on the predicted results.

[0158] In the fifth embodiment, the breakage distribution corresponding to any koji-making condition is the actual measured value of the breakage distribution generated by inputting any koji-making condition into the learned koji-grain class estimation model and converting the output multiple koji-grain classes into the ratio of each class. In the present invention, the breakage distribution corresponding to any koji-making condition is not limited to the fifth embodiment, and may be the predicted value of the breakage distribution output from the learned breakage distribution prediction model that predicts the breakage distribution from the koji-making conditions of the third embodiment. However, when using the predicted value of the breakage distribution output from the learned breakage distribution prediction model that predicts the breakage distribution from the koji-making conditions, the accuracy of predicting the enzyme titer of the koji may be inferior to when using the actual measured value of the breakage distribution in the fifth embodiment, since it depends on the prediction accuracy of the learned breakage distribution prediction model.

[0159] Instead of the enzyme titer as the response variable, it is also possible to predict other physical properties obtained from koji, such as the amount of bacterial cells, as the response variable.

[0160] 6. Sixth embodiment A sixth embodiment of a koji-making support system for predicting the enzyme activity of koji based on koji-making conditions and data on the breaking down of multiple koji grains will be described.

[0161] In the sixth embodiment, the data on the destruction of multiple grains of koji is the average value of multiple data sets of destruction data that can be obtained from a single grain of koji, obtained using a method similar to that of the fourth embodiment, and is the average value of the contour detection data for multiple grains of koji.

[0162] In the sixth embodiment, the enzyme activity is the activity of each of α-amylase and glucoamylase measured from the produced koji by a predetermined analytical method, as in the fifth embodiment. The enzyme activity of the final koji is one of the indicators in koji production and indicates the quality of the koji.

[0163] The sixth embodiment is configured to use α-amylase and glucoamylase as enzyme titers, however, the present invention is not limited to the configuration of the sixth embodiment, and may be configured to use enzyme titers of acid protease, acid carboxypeptidase, etc. in addition to the enzyme titers described above.

[0164] In the koji-making process of the sixth embodiment, the koji-making conditions are stored as a series of data in a storage device, which is the computer's storage means 101, as in the fourth embodiment. A portion of the koji made under the koji-making conditions is sampled and converted into data on the breaking down of multiple koji grains using the same method as in the second embodiment. At the same time, a portion of the koji made under the koji-making conditions is sampled and the enzyme activity is measured using a predetermined analytical method.

[0165] In the sixth embodiment, the koji-making conditions obtained using the above-described method are used as one teacher koji-making condition, the data on multiple koji grains broken obtained using the above-described method from koji produced under the koji-making conditions are used as one teacher koji multiple grain broken data, and the enzyme titer obtained using the above-described method from koji produced under the koji-making conditions is used as one teacher enzyme titer analysis value. A set of teacher data is created using these teacher koji-making conditions, teacher multiple koji broken data, and teacher enzyme titer analysis value. In this set of teacher data, the data consisting of the teacher koji-making conditions and teacher multiple koji broken data function as explanatory variables, and the teacher enzyme titer analysis value functions as a target variable. By repeating the creation of teacher data using the above procedure, a teacher dataset containing multiple teacher data is created.

[0166] In the sixth embodiment, a second trained enzyme potency prediction model that predicts the enzyme potency of koji from the koji-making conditions and data on the breaking down of multiple koji grains is constructed by the trained model construction means 200 using the above-mentioned training data set, as in the first embodiment, and is stored in the memory means 101 of the koji-making support system 100.

[0167] The sixth embodiment is configured to use LR as a learning model, which is a machine learning technique. However, the present invention is not limited to the configuration of the sixth embodiment, and may be configured to use other regression algorithms, such as NN, RFR, SVR, and DT.

[0168] In the second trained enzyme potency prediction model of the sixth embodiment, the koji-making conditions and the data on breaking down multiple koji grains are explanatory variables, and the enzyme potency is the objective variable. In the sixth embodiment, since there are two types of enzyme potency, α-amylase and glucoamylase, there are two objective variables for LR. Furthermore, since LR is used, the trained model construction means 200 will construct two second trained enzyme potency prediction models, the same number of variables, for the two objective variables. The present invention is not limited to the sixth embodiment, and if a configuration is adopted in which NN is used as the learning model, which is a machine learning method, the trained model construction means 200 will construct one second trained enzyme potency prediction model for the two objective variables.

[0169] In the sixth embodiment, 180 training datasets are used to construct the second trained enzyme titer prediction model. The accuracy of the second trained enzyme titer prediction model is evaluated using 20 test datasets. In the sixth embodiment, the test datasets are not used to construct the second trained enzyme titer prediction model, as in the first embodiment.

[0170] As shown in Table 5, the accuracy of the prediction results of the second trained enzyme titer prediction model is evaluated using a test data set. In the sixth embodiment, the coefficient of determination R 2 is used.

[0171] [Table 5]

[0172] As shown in the above test results, the second trained enzyme potency prediction model is capable of outputting highly accurate predicted values ​​of enzyme potency based on the input koji-making conditions and data on breaking down multiple koji grains.

[0173] As shown in FIG. 1, the koji-making assistance system 100 of the sixth embodiment includes a storage unit 101, an input unit 102, a control unit 103, and an output unit 104, similar to the first embodiment. Also, as shown in FIG. 8, the koji-making assistance program 300 of the sixth embodiment includes an input step 301, a control step 302, and an output step 303, similar to the first embodiment. The koji-making assistance program 300 is stored in the storage unit 101 of the koji-making assistance system 100. The control step 302 controls the processing of koji image data acquired by the input unit 102, and causes the control unit 103 to execute the processing of storing the output processed data in the storage unit 101. The input step 301 inputs the processed data stored in the storage unit 101 into the contour detection unit. The control step 302 controls the processing by the contour detection unit, and causes the control unit 103 to execute the processing of storing the output multiple contour detection data in the storage unit 101. The control step 302 causes the control means 103 to execute a process of averaging multiple contour detection data stored in the storage means 101 and storing the averaged multiple contour detection data in the storage means 101 as multiple-grain-breaking-of-koji data. The input step 301 inputs the koji-making conditions acquired by the input means 102 and the multiple-grain-breaking-of-koji data stored in the storage means 101 into the second trained enzyme potency prediction model stored in the storage means 101. The control step 302 controls calculations using the second trained enzyme potency prediction model and causes the control means 103 to execute a process of outputting the enzyme potency based on the input koji-making conditions and multiple-grain-breaking-of-koji data. The output step 303 acquires the enzyme potency output by the second trained enzyme potency prediction model and displays it on the output means 104. By including each of these steps, the koji-making assistance program 300 causes the koji-making assistance system 100 to execute a process of predicting the enzyme potency from the koji-making conditions and multiple-grain-breaking-of-koji data.

[0174] In the sixth embodiment, the input means 102 for acquiring image data of koji, which is data for breaking down a single grain of koji, is a camera, and the koji-making assistance system 100 can acquire data for breaking down multiple grains of koji, which is generated by processing the image data of multiple pieces of koji acquired by the input means 102. Furthermore, the input means 102 for acquiring koji-making conditions is a keyboard, and the koji-making assistance system 100 can acquire the koji-making conditions input as numerical values ​​by the input means 102.

[0175] The koji-making support system 100 of the present invention may be configured to predict enzyme titers from koji-making conditions and multiple koji grain breakage data using multiple second trained enzyme potency prediction models constructed using different techniques, such as a second trained enzyme potency prediction model constructed using neural networks, a second trained enzyme potency prediction model constructed using LR, or a second trained enzyme potency prediction model constructed using RFR. Alternatively, multiple second trained enzyme potency prediction models constructed using the same technique but with different hyperparameters may be used. Here, the multiple second trained enzyme potency prediction models may be configured to input common koji-making conditions and multiple koji grain breakage data, calculate an average value from the output enzyme titers, and determine the average value as the final prediction result.

[0176] The sixth embodiment of the koji-making support system 100 uses a second trained enzyme potency prediction model constructed by performing machine learning on a learning model, which is a machine learning technique, to predict enzyme potency from arbitrary koji-making conditions and data on the breaking down of multiple grains of koji corresponding to those arbitrary koji-making conditions. This allows the user to predict the enzyme potency immediately after koji-making is completed, and to make fine adjustments to mash management based on the prediction results.

[0177] In the sixth embodiment, the multiple-grain-breaking data for koji corresponding to arbitrary koji-making conditions is data obtained by processing image data acquired from multiple grains of koji produced. In the present invention, the multiple-grain-breaking data for koji corresponding to arbitrary koji-making conditions is not limited to the sixth embodiment, but may be predicted values ​​of multiple-grain-breaking data output by inputting arbitrary koji-making conditions into a trained multiple-grain-breaking data prediction model. However, when using predicted values ​​of multiple-grain-breaking data output from a trained multiple-grain-breaking data prediction model that predicts multiple-grain-breaking data from koji-making conditions, the accuracy of predicting the enzyme titer of koji may be inferior to that when using actual measured values ​​of the multiple-grain-breaking data for koji in the sixth embodiment, because it depends on the prediction accuracy of the trained multiple-grain-breaking data prediction model.

[0178] Instead of the enzyme titer as the response variable, it is also possible to predict other physical properties obtained from koji, such as the amount of bacterial cells, as the response variable.

[0179] 7. Seventh embodiment A koji-making support system that extracts candidate combinations of koji-making conditions and breaking distributions for search conditions according to the seventh embodiment will be described.

[0180] In the seventh embodiment, the koji-making conditions are composed of a total of 12 types of condition data: seed amount, moisture, product temperatures at 1, 5, 15, and 22 hours in the bed process, and product temperatures at 1, 5, 10, 15, 20, and 24 hours in the shelf process.

[0181] In the seventh embodiment, three levels are set for each condition data that constitutes one standard koji-making condition (hereinafter referred to as "standard koji-making condition"), and all koji-making conditions that are constituted by combinations of the set condition data are saved, thereby creating koji-making conditions for multiple database configurations.

[0182] The seventh embodiment is configured to use the koji-making conditions recorded in past koji-making experiments as the reference koji-making conditions. The values ​​of the condition data constituting the reference koji-making conditions will be explained as reference values. The three levels are: a level in which a predetermined value is subtracted from the reference value, a level in which the reference value is used as is, and a level in which a predetermined value is added to the reference value.

[0183] In the seventh embodiment, three levels are set for the seeding amount, which is condition data constituting the standard koji-making conditions: a reference value of -0.01 g / kg, the reference value, and a reference value of +0.01 g / kg. Furthermore, three levels are set for the moisture, which is condition data constituting the standard koji-making conditions: a reference value of -1%, the reference value, and a reference value of +1%. Furthermore, three levels are set for each product temperature, which is condition data constituting the standard koji-making conditions: a reference value of -1°C, the reference value, and a reference value of +1°C.

[0184] In the seventh embodiment, three levels are set for each of the 12 types of condition data that make up the standard koji-making conditions, and all koji-making conditions that are made up of combinations of the set condition data are saved. This results in a total of approximately 530,000 (3 12 Koji-making conditions for the database configuration (items) are created.

[0185] The present invention is not limited to the configuration of the seventh embodiment, and the number of levels may be changed. For example, five levels may be set for each product temperature, which is the condition data constituting the reference koji-making conditions: a reference value of −2°C, a reference value of −1°C, the reference value, a reference value +1°C, and a reference value +2°C. Furthermore, the reference koji-making conditions may be configured to use the average value of multiple koji-making conditions recorded in past actual koji-making operations, or to use koji-making conditions that the user deems appropriate. Furthermore, without setting reference koji-making conditions, the koji-making conditions for the database may be configured to use a large number of koji-making conditions accumulated in past actual koji-making operations, or to use a large number of koji-making conditions that the user deems appropriate.

[0186] The seventh embodiment is configured to use the breaking semen distribution output by inputting the koji-making conditions for database construction into the learned breaking semen distribution prediction model of the third embodiment as the breaking semen distribution for database construction.

[0187] In the seventh embodiment, one record constituting a database for extracting candidate combinations of koji-making conditions and breaking-semite distributions is created from one koji-making condition for database construction and one breaking-semite distribution for database construction output by inputting the koji-making condition for database construction into a trained breaking-semite distribution prediction model. By repeating the creation of records according to the above procedure, a database for extracting candidate combinations of koji-making conditions and breaking-semite distributions consisting of approximately 530,000 records is created.

[0188] In the seventh embodiment, the ratio data of koji classes is data expressed as the ratio of each class in the distribution of broken koji.

[0189] In the seventh embodiment, the target data refers to the ratio data of each koji class that constitutes the koji-breaking distribution, and the target data may include one or more data. In the seventh embodiment, it is preferable to increase the number of target data to be used as search conditions. By increasing the number of target data, it is possible to improve the accuracy of extracting optimal combination candidates of koji-making conditions and koji-breaking distributions from the records that constitute the database for extracting combination candidates of koji-making conditions and koji-breaking distributions.

[0190] In the seventh embodiment, all of the ratio data for each koji class can be set as target data. For example, the user can set the ratio data for class 0 as "10," the ratio data for class 1 as "40," the ratio data for class 2 as "30," and the ratio data for class 3 as "20." These ratio data can then be used as target data when searching for data from the database for extracting candidate combinations of koji-making conditions and breaking distributions. Furthermore, instead of setting ratio data for all of the koji classes that make up the breaking distribution, the user can set the ratio data for class 1 as "40," for example, and use this ratio data as target data when searching for data from the database for extracting candidate combinations of koji-making conditions and breaking distributions.

[0191] As shown in Figure 1, the koji-making assistance system 100 of the seventh embodiment, like the first embodiment, comprises a storage means 101, an input means 102, a control means 103, and an output means 104. A database for extracting candidate combinations of koji-making conditions and breaking distributions is stored in the storage means 101. The storage means 101 also stores a koji-making assistance program 400. The input means 102 of the seventh embodiment is a keyboard, and the koji-making assistance system 100 can acquire koji class ratio data input as numerical values ​​via the input means 102.

[0192] As shown in FIG. 12, the koji-making assistance program 400 of the seventh embodiment includes an input step 401, a search step 402, and an output step 403. The input step 401 transfers the koji class ratio data acquired by the input unit 102 to the search step 402. The search step 402 calculates the root mean squared error (RMSE) between the koji class ratio data acquired from the input step 401 and the koji class ratio data included in each record constituting the database for extracting candidate combinations of koji-making conditions and breakage distributions. The control unit 103 then executes a process to extract the record with the smallest RMSE as a candidate combination of koji-making conditions and breakage distributions that satisfies the search criteria. The output step 403 displays the acquired candidate combinations of koji-making conditions and breakage distributions on the output unit 104. By including these steps, the koji-making assistance program 400 can cause the koji-making assistance system 100 to execute a process to extract candidate combinations of koji-making conditions and breakage distributions based on the search criteria.

[0193] In the seventh embodiment, if target data is set containing ratio data for four koji classes, for example, with the ratio data for class 0 set to "10," the ratio data for class 1 set to "40," the ratio data for class 2 set to "30," and the ratio data for class 3 set to "20," search step 402 calculates the RMSE between the ratio data for these koji classes passed from input step 401 and the ratio data for the four koji classes consisting of classes 0 to 3 contained in each record constituting the database for extracting candidate combinations of koji-making conditions and breaking distributions, and obtains the record with the smallest RMSE as a candidate combination of koji-making conditions and breaking distributions that satisfies the search conditions. Furthermore, for example, if the ratio data for class 1 is set to "40" and target data including ratio data for one koji class is set, search step 402 calculates the RMSE between the ratio data for that koji class passed from input step 401 and the ratio data for one koji class consisting of class 1 included in each record that makes up the database for extracting candidate combinations of koji-making conditions and breaking distributions, and obtains the record with the smallest RMSE as a candidate combination of koji-making conditions and breaking distributions that meets the search conditions.

[0194] The present invention is not limited to the configuration of the seventh embodiment, and the search step may be configured to acquire all records whose calculated RMSE is smaller than a certain value, or may be configured to sort the records in ascending order of calculated RMSE and acquire a certain number of records from the top. Furthermore, the search step may be configured to acquire records using other evaluation indexes, such as mean absolute error (MAE), instead of RMSE.

[0195] In the seventh embodiment, the koji-making support system 100 evaluates the accuracy of the extraction results by comparing the target value of the breaking distribution set by the user as a search condition with the actual measured value of the breaking distribution obtained by actually making koji using the extracted koji-making conditions.

[0196] In the seventh embodiment, the target values ​​for the breakage distribution to evaluate the accuracy of the extraction results were set as [Class 0: 20%, Class 1: 40%, Class 2: 30%, Class 3: 10%]. Based on the ratio data of the koji classes that make up the target values ​​for the breakage distribution, the koji-making support system 100 extracted the koji-making conditions consisting of [seeding amount: 0.01 g / kg, moisture: 33.0%, bed process 1 h product temperature: 31.0 ° C, bed process 5 h product temperature: 31.0 ° C, bed process 15 h product temperature: 31.5 ° C, bed process 22 h product temperature: 32.0 ° C, shelf process 1 h product temperature: 33.0 ° C, shelf process 5 h product temperature: 34.0 ° C, shelf process 10 h product temperature: 38.0 ° C, shelf process 15 h product temperature: 42.0 ° C, shelf process 20 h product temperature: 43.0 ° C, shelf process 24 h product temperature: 43.0 ° C]. When koji was actually made using the extracted koji-making conditions, the measured value of the broken-sperm distribution was obtained, consisting of [Class 0: 19%, Class 1: 48%, Class 2: 29%, Class 3: 5%]. The RMSE between the target value of the broken-sperm distribution and the measured value of the broken-sperm distribution is approximately 4.8. This RMSE value is evaluated to be within a range that does not cause any problems in practical use.

[0197] As shown in the above test results, the koji-making assistance system 100 is able to extract candidate combinations of koji-making conditions and breaking distributions with high accuracy in response to search conditions.

[0198] In the seventh embodiment, the breaking distribution for the database is output by inputting the koji-making conditions for the database into the learned breaking distribution prediction model. This configuration makes it possible to create a large number of breaking distributions for the database based on the koji-making conditions for the database.

[0199] The seventh embodiment of the koji-making support system 100 can extract candidate combinations of koji-making conditions and breaking distributions based on one or more of the ratio data for each koji class. Therefore, the user can input a breaking distribution consisting of the ratio data for all the target koji classes, or the ratio data for a koji class that is a part of it, and analyze the koji-making conditions required to obtain the target breaking distribution or ratio data for the koji class.

[0200] The present invention is not limited to the configuration of the seventh embodiment, and may be configured such that records consisting of actual measured values ​​of koji-making conditions and breaking distribution recorded in actual koji-making are added to the database for extracting candidate combinations of koji-making conditions and breaking distribution created by the above-mentioned method, and the koji-making support system 100 may extract candidate combinations of koji-making conditions and breaking distribution using the actual measured value-added database for extracting candidate combinations of koji-making conditions and breaking distribution to which these records have been added.

[0201] 8. Eighth embodiment The following describes a koji-making support system that extracts candidate combinations of koji-making conditions and data on breaking down multiple koji grains in response to search conditions in the eighth embodiment.

[0202] In the eighth embodiment, the koji-making conditions for constructing the database are created in the same way as in the seventh embodiment. The eighth embodiment is configured to use the koji multiple grain breakage data output by inputting the koji-making conditions for constructing the database into the trained koji multiple grain breakage data prediction model of the fourth embodiment as the koji multiple grain breakage data for constructing the database.

[0203] In the eighth embodiment, one record constituting a database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data of koji is created from one set of koji-making conditions for constructing a database and one set of multiple-grain-breaking data of koji output by inputting the koji-making conditions for constructing a database into a trained multiple-grain-breaking data prediction model. By repeating the creation of records according to the above procedure, a database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data of koji consisting of approximately 530,000 records is created.

[0204] In the eighth embodiment, the koji multiple grain breaking data is, as in the fourth embodiment, four types of data: the average value of the broken area of ​​the koji multiple grains, the average value of the black percentage of the multiple grains, the average value of the circumference of the multiple grains, and the average value of the number of colonies of the multiple grains. In the eighth embodiment, it is the average value of each breaking data for 100 koji grains.

[0205] In the eighth embodiment, the target data refers to each piece of data constituting the koji multiple grain breakage data, and the target data may include one or more pieces of data. In the eighth embodiment, it is preferable to increase the number of pieces of target data to be used as search conditions. By increasing the number of pieces of target data, it is possible to improve the accuracy of extracting optimal combination candidates of koji-making conditions and multiple grain breakage data from the records constituting the database for extracting combination candidates of koji-making conditions and multiple grain breakage data.

[0206] In the eighth embodiment, all of the breaking data can be set as target data. For example, the user can set the breaking data for the breaking area to "200," the black rate to "0.20," the circumference to "85," and the colony count to "3." These breaking data can be used as target data when searching for data from the database for extracting combination candidates for koji-making conditions and multiple-grain breaking data. Furthermore, instead of setting all of the breaking data that make up the multiple-grain breaking data, the user can set the breaking data for the breaking area to "200," for example, and use this breaking data as target data when searching for data from the database for extracting combination candidates for koji-making conditions and multiple-grain breaking data.

[0207] As shown in Figure 1, the koji-making assistance system 100 of the eighth embodiment, like the first embodiment, comprises a storage means 101, an input means 102, a control means 103, and an output means 104. A database for extracting combination candidates for koji-making conditions and data on the breaking of multiple grains of koji is stored in the storage means 101. The storage means 101 also stores a koji-making assistance program 400. The input means 102 of the eighth embodiment is a keyboard, and the koji-making assistance system 100 can acquire breaking data input as numerical values ​​via the input means 102.

[0208] As shown in FIG. 12, the koji-making assistance program 400 of the eighth embodiment includes an input step 401, a search step 402, and an output step 403. The input step 401 transfers the koji-breaking data acquired by the input means 102 to the search step 402. The search step 402 calculates the RMSE between the koji-breaking data acquired from the input step 401 and the koji-breaking data included in each record constituting the database for extracting candidate combinations of koji-making conditions and multiple-grain-breaking data, and causes the control means 103 to execute a process of acquiring the record with the smallest RMSE as a candidate combination of koji-making conditions and multiple-grain-breaking data that satisfies the search criteria. The output step 403 causes the output means 104 to display the acquired candidate combinations of koji-making conditions and multiple-grain-breaking data. By including these steps, the koji-making assistance program 400 allows the koji-making assistance system 100 to execute a process of extracting candidate combinations of koji-making conditions and multiple-grain-breaking data based on the search criteria.

[0209] In the eighth embodiment, the RMSE is not calculated using the squared difference between the target koji-breaking data and the koji-breaking data included in each record constituting the database for extracting combination candidates for koji-making conditions and multiple-grain koji-breaking data. Instead, the difference between the target koji-breaking data and the koji-breaking data included in each record constituting the database for extracting combination candidates for koji-making conditions and multiple-grain koji-breaking data is divided by the numerical value of the target koji-breaking data to obtain a relative numerical value relative to the target koji-breaking data. The relative numerical values ​​are each squared, added, and divided by the number of data to obtain the square root of the result. Because the range of values ​​that each koji-breaking data can take varies greatly depending on the type of koji-breaking data, conventional methods calculate RMSE that is strongly influenced by koji-breaking data with large values. On the other hand, the eighth embodiment calculates RMSE using the above-mentioned method, allowing different types of koji-breaking data to be treated and evaluated equally. Details will be described later using specific numerical values.

[0210] In the eighth embodiment, when target data including four pieces of breaking data are set, for example, the breaking data for the breaking area is set to "200", the breaking data for the black rate is set to "0.20", the breaking data for the circumference is set to "85", and the breaking data for the number of colonies is set to "3", the search step 402 converts the difference between these breaking data passed from the input step 401 and the four breaking data consisting of the breaking area, black rate, circumference, and number of colonies contained in each record constituting the database for extracting combination candidates for koji making conditions and multiple grain koji breaking data (for example, one of the records constituting the database for extracting combination candidates for koji making conditions and multiple grain koji breaking data: breaking data for the breaking area "190", breaking data for the black rate "0.18", breaking data for circumference "80", breaking data for the number of colonies "3") into a relative numerical value for the target breaking data. Specifically, the deviation of "10" between the target breaking data (200) for the breaking area and the breaking data for the breaking area of ​​the records constituting the database for extracting candidate combinations of koji-making conditions and multiple-grain koji breaking data was calculated. The deviation of "10" was then divided by the target breaking data (200) to obtain "0.05 (5%)," which was then converted into a relative value of "5." Relative values ​​for the black ratio, circumference, and number of colonies were calculated using the same method, resulting in relative values ​​for each target breaking data (10), "6," and "0," respectively. The four calculated relative values ​​were then squared and added together. The sum was then divided by the number of data (n = 4) and square rooted to obtain an RMSE of "6.3." The record with the smallest calculated RMSE for each record was selected as a candidate combination of koji-making conditions and multiple-grain koji breaking data that met the search criteria.Furthermore, in the eighth embodiment, for example, when the breaking data for the breaking portion area is set to "300" and target data including one type of breaking data is set, the search step 402 calculates the RMSE between the breaking data passed from the input step 401 and one type of breaking data consisting of the breaking portion area included in each record that constitutes the database for extracting candidate combinations of koji-making conditions and multiple koji grain breaking data, and obtains the record with the smallest RMSE as a candidate combination of koji-making conditions and multiple koji grain breaking data that satisfies the search conditions.

[0211] The present invention is not limited to the configuration of the eighth embodiment, and the search step may be configured to acquire all records whose calculated RMSE is smaller than a certain value, or may be configured to sort the records in ascending order of calculated RMSE and acquire a certain number of records from the top. Furthermore, the search step may be configured to acquire records using other evaluation indexes, such as mean absolute error (MAE), instead of RMSE.

[0212] In the eighth embodiment, the koji-making support system 100 evaluates the accuracy of the extraction results by comparing the target value of the data for breaking down multiple grains of koji set by the user as a search condition with the actual measured value of the data for breaking down multiple grains of koji obtained by actually making koji using the extracted koji-making conditions.

[0213] The eighth embodiment is configured to use the multiple-grain-breaking data of koji for database construction, which is output by inputting the koji-making conditions for database construction into the learned multiple-grain-breaking data prediction model. This configuration makes it possible to create a large amount of multiple-grain-breaking data of koji for database construction based on the large amount of koji-making conditions for database construction.

[0214] The eighth embodiment of the koji-making support system 100 can extract candidate combinations of koji-making conditions and multiple-grain koji breaking data based on one or more of the breaking data. Therefore, the user can input the target multiple-grain koji breaking data consisting of all the breaking data or breaking data that is a part of it, and analyze the koji-making conditions required to obtain the target multiple-grain koji breaking data or breaking data.

[0215] The present invention is not limited to the configuration of the eighth embodiment, and may be configured such that records consisting of actual measured values ​​of koji-making conditions and data on multiple koji grains breaking recorded in actual koji making are added to the database for extracting candidate combinations of koji-making conditions and data on multiple koji grains breaking created by the above-mentioned method, and the koji-making support system 100 may extract candidate combinations of koji-making conditions and data on multiple koji grains breaking using the database for extracting candidate combinations of koji-making conditions and data on multiple koji grains breaking of an actual measured value added type to which these records have been added.

[0216] 9. Ninth embodiment A koji-making support system that extracts candidate combinations of koji-making conditions and enzyme activity values ​​for search conditions according to the ninth embodiment will be described.

[0217] In the ninth embodiment, the koji-making conditions for constructing the database are created using the same method as in the seventh embodiment. The breakage distribution is created using the koji-making conditions for constructing the database as input using the same method as in the third embodiment. Furthermore, the enzyme activity values ​​for constructing the database are created using the same method as in the fifth embodiment.

[0218] In the ninth embodiment, one record constituting a database for extracting candidate combinations of a first koji-making condition and an enzyme titer is created from one koji-making condition for constructing a database and one enzyme titer for constructing a database output by inputting the koji-making condition for constructing a database and one breakage distribution output by inputting the koji-making condition for constructing a database into a trained breakage distribution prediction model for a first trained enzyme titer prediction model. By repeating the creation of records according to the above procedure, a database for extracting candidate combinations of a first koji-making condition and an enzyme titer containing multiple accumulated records is created.

[0219] In the ninth embodiment, the enzyme activity data constituting the enzyme activity are α-amylase and glucoamylase, and if α-amylase has a value of 700 (units / g·koji) and glucoamylase has a value of 200 (units / g·koji), the enzyme activity data of α-amylase is "700" and the enzyme activity data of glucoamylase is "200".

[0220] In the ninth embodiment, the target data refers to the enzyme activity data that make up the enzyme activity, and the target data may include one or more data. It is preferable to increase the number of target data to be used as search criteria. By increasing the number of target data, the accuracy of extracting optimal candidate combinations of koji-making conditions and enzyme activity values ​​can be improved from the records that make up the database for extracting candidate combinations of koji-making conditions and enzyme activity values.

[0221] In the ninth embodiment, all of the enzyme potency data can be set as target data. For example, the user can set the enzyme potency data for α-amylase to "700" and the enzyme potency data for glucoamylase to "200," and use these enzyme potency data as target data when searching for data from the database for extracting candidate combinations of the first koji-making conditions and enzyme potencies. Alternatively, instead of setting enzyme potency data for all of the data constituting the enzyme potencies, the user can set the enzyme potency data for α-amylase to "700," and use this enzyme potency data as target data when searching for data from the database for extracting candidate combinations of the first koji-making conditions and enzyme potencies.

[0222] As shown in Figure 1, the koji-making assistance system 100 of the ninth embodiment, like the first embodiment, comprises a storage means 101, an input means 102, a control means 103, and an output means 104. A database for extracting candidate combinations of the first koji-making conditions and enzyme activity values ​​is stored in the storage means 101. The storage means 101 also stores a koji-making assistance program 400. The input means 102 is a keyboard, and the koji-making assistance system 100 can acquire each enzyme activity data, which is target data input as a numerical value via the input means 102.

[0223] As shown in FIG. 12, the koji-making assistance program 400 of the ninth embodiment includes an input step 401, a search step 402, and an output step 403. The input step 401 transfers each enzyme titer data acquired by the input means 102 to the search step 402. The search step 402 calculates the RMSE between each enzyme titer data transferred from the input step 401 and the enzyme titer data included in each record constituting the first database for extracting candidate combinations of koji-making conditions and enzyme titers, and causes the control means 103 to execute a process of acquiring the record with the smallest RMSE as a candidate combination of koji-making conditions and enzyme titers that satisfies the search criteria. The output step 403 causes the output means 104 to display the acquired candidate combinations of koji-making conditions and enzyme titers. By including these steps, the koji-making assistance program 400 can cause the koji-making assistance system 100 to execute a process of extracting candidate combinations of koji-making conditions and enzyme titers based on the search criteria.

[0224] In the ninth embodiment, as in the eighth embodiment, the RMSE is calculated by dividing the difference between each target enzyme potency data and each enzyme potency data included in each record constituting the database for extracting candidate combinations of first koji-making conditions and enzyme potencies by each target enzyme potency data, thereby obtaining a relative value for each target enzyme potency data. These relative values ​​are each squared, added, and divided by the number of data points, and the square root of the result is used. Because the range of values ​​that each enzyme potency data can take varies greatly depending on the type of enzyme potency data, typical methods calculate an RMSE that is strongly influenced by enzyme potency data with large values. Calculating the RMSE using the above method allows different types of enzyme potency data to be treated and evaluated equally.

[0225] In the ninth embodiment, the user sets a target value for the enzyme activity, and the koji-making support system 100 extracts candidate combinations of koji-making conditions and enzyme activity based on the enzyme activity data that constitutes the target value for the enzyme activity. Here, the koji-making support system 100 evaluates the accuracy of the extraction results by comparing the target value for the enzyme activity set by the user as a search condition with the actual measured value of the enzyme activity obtained by actually making koji using the extracted koji-making conditions.

[0226] In the ninth embodiment, α-amylase

[0700] and glucoamylase

[0200] were set as target data for enzyme activity to evaluate the accuracy of the extraction results. Based on the enzyme activity data constituting the target values ​​for the enzyme activity, the koji-making support system 100 extracted the following koji-making conditions: seed amount: 0.02 g / kg, moisture: 34.0%, bed process 1-h product temperature: 31.0°C, bed process 5-h product temperature: 31.0°C, bed process 15-h product temperature: 31.5°C, bed process 22-h product temperature: 32.0°C, shelf process 1-h product temperature: 33.0°C, shelf process 5-h product temperature: 35.0°C, shelf process 10-h product temperature: 40.0°C, shelf process 15-h product temperature: 43.0°C, shelf process 20-h product temperature: 44.0°C, shelf process 24-h product temperature: 44.0°C. When koji was actually made using the extracted koji-making conditions, the actual enzyme activity values ​​for α-amylase

[0720] and glucoamylase

[0210] were obtained. The RMSE between the target enzyme activity value and the actual enzyme activity value was approximately 4.1. This RMSE value is considered to be within a range that is acceptable for practical use.

[0227] As shown in the above test results, the koji-making assistance system 100 is able to extract candidate combinations of koji-making conditions and enzyme activity values ​​with high accuracy for the search conditions.

[0228] The ninth embodiment is configured to use the enzyme titers output by inputting the koji-making conditions and the breaking distribution for database construction into the first trained enzyme potency prediction model as the enzyme titers for database construction. This configuration makes it possible to create a large number of enzyme titers for database construction based on the koji-making conditions for database construction that have been created in large quantities.

[0229] The koji-making support system 100 of the ninth embodiment can extract candidate combinations of koji-making conditions and enzyme potencies based on one or more of the enzyme potency data. Therefore, by inputting all of the target enzyme potency data or a portion of the target enzyme potency data, the user can analyze the koji-making conditions required to obtain the target enzyme potency or enzyme potency data.

[0230] The present invention is not limited to the configuration of the ninth embodiment, and may be configured such that records consisting of actual measured values ​​of koji-making conditions and enzyme titers recorded in actual koji-making are added to the database for extracting candidate combinations of first koji-making conditions and enzyme titers created by the above-mentioned method, and the koji-making support system 100 may extract candidate combinations of koji-making conditions and enzyme titers using the actual measured value-added first database for extracting candidate combinations of koji-making conditions and enzyme titers to which these records have been added.

[0231] The present invention is not limited to the ninth embodiment, and may be a database for extracting candidate combinations of koji-making conditions, a breaking distribution, and an enzyme titer, which is composed of koji-making conditions, a breaking distribution, and an enzyme titer. In this case, the breaking distribution for the database configuration output by inputting the koji-making conditions for the database configuration into a trained breaking distribution prediction model can be used.

[0232] In the present invention, even when using a database for extracting candidate combinations of koji-making conditions, breakage distribution, and enzyme titer, one or more target data from any of the enzyme titer data constituting the enzyme titer can be used as search criteria. Furthermore, when using a database for extracting candidate combinations of koji-making conditions, breakage distribution, and enzyme titer, the number or type of target data can be increased to search criteria by adding one or more target data from any of the condition data constituting the koji-making conditions, or one or more target data from any of the ratio data from any of the koji class constituting the breakage distribution, in addition to one or more target data from any of the enzyme titer data constituting the enzyme titer. By increasing the number or type of target data, optimal candidate combinations of koji-making conditions, breakage distribution, and enzyme titer can be extracted from the records constituting the database for extracting candidate combinations of koji-making conditions, breakage distribution, and enzyme titer.

[0233] In the present invention, when using a database for extracting candidate combinations of koji-making conditions, breaking-sem distribution, and enzyme titer, instead of using one or more target data from any of the enzyme titer data that make up the enzyme titer as the search criteria, one or more target data from any of the condition data that make up the koji-making conditions, or one or more target data from any of the ratio data from any of the koji classes that make up the breaking-sem distribution, may be used as the search criteria.

[0234] 10. Tenth embodiment A koji-making support system that extracts candidate combinations of koji-making conditions and enzyme activity values ​​for search conditions according to the tenth embodiment will be described.

[0235] In the tenth embodiment, the koji-making conditions for constructing the database are created using the same method as in the seventh embodiment. The koji multiple grain breaking data is created using the koji-making conditions for constructing the database as input using the same method as in the fourth embodiment. The enzyme activity values ​​for constructing the database are created using the same method as in the sixth embodiment.

[0236] In the tenth embodiment, one record constituting a database for extracting candidate combinations of second koji-making conditions and enzyme titers is created from one set of koji-making conditions for constructing a database and one set of enzyme titers for constructing a database output by inputting the one set of koji-making conditions for constructing a database and one set of koji multiple grain breakage data output by the trained multiple grain breakage data prediction model using the one set of koji-making conditions for constructing a database into a second trained enzyme titer prediction model. By repeating the creation of records according to the above procedure, a database for extracting candidate combinations of second koji-making conditions and enzyme titers containing multiple accumulated records is created.

[0237] In the tenth embodiment, the enzyme activity data is the activity data of each enzyme that constitutes the enzyme activity, as in the ninth embodiment.

[0238] In the tenth embodiment, the target data refers to the enzyme activity data that make up the enzyme activity, and the target data may include one or more data. It is preferable to increase the number of target data to be used as search conditions. By increasing the number of target data, the accuracy of extracting optimal candidate combinations of koji-making conditions and enzyme activity values ​​can be improved from the records that make up the database for extracting candidate combinations of koji-making conditions and enzyme activity values.

[0239] In the tenth embodiment, all of the enzyme titer data can be set as target data, similarly to the ninth embodiment.

[0240] As shown in Figure 1, the koji-making assistance system 100 of the tenth embodiment, like the first embodiment, comprises a storage means 101, an input means 102, a control means 103, and an output means 104. A database for extracting candidate combinations of second koji-making conditions and enzyme activity values ​​is stored in the storage means 101. The storage means 101 also stores a koji-making assistance program 400. The input means 102 is a keyboard, and the koji-making assistance system 100 can acquire each enzyme activity data, which is target data input as a numerical value via the input means 102.

[0241] As shown in FIG. 12, the koji-making assistance program 400 of the tenth embodiment includes an input step 401, a search step 402, and an output step 403. The input step 401 transfers each enzyme titer data acquired by the input means 102 to the search step 402. The search step 402 calculates the RMSE between each enzyme titer data transferred from the input step 401 and the enzyme titer data included in each record constituting the second database for extracting candidate combinations of koji-making conditions and enzyme titers, and causes the control means 103 to execute a process of acquiring the record with the smallest RMSE as a candidate combination of koji-making conditions and enzyme titers that satisfies the search criteria. The output step 403 causes the output means 104 to display the acquired candidate combinations of koji-making conditions and enzyme titers. By including these steps, the koji-making assistance program 400 can cause the koji-making assistance system 100 to execute a process of extracting candidate combinations of koji-making conditions and enzyme titers based on the search criteria.

[0242] In the tenth embodiment, as in the ninth embodiment, the RMSE is calculated by dividing the difference between each target enzyme titer data and each enzyme titer data included in each record constituting the database for extracting candidate combinations of second koji-making conditions and enzyme titers by each target enzyme titer data, thereby obtaining a relative value for each target enzyme titer data. The relative values ​​are each squared, added, and divided by the number of data, and the square root of the result is used. Because the range of values ​​that each enzyme titer data can take varies greatly depending on the type of enzyme titer data, typical methods calculate an RMSE that is strongly influenced by enzyme titer data with large values. Calculating the RMSE using the above method allows different types of enzyme titer data to be treated and evaluated equally.

[0243] In the tenth embodiment, the user sets a target value for the enzyme activity, and the koji-making support system 100 extracts candidate combinations of koji-making conditions and enzyme activity based on the enzyme activity data that constitutes the target value for the enzyme activity. Here, the koji-making support system 100 evaluates the accuracy of the extraction results by comparing the target value for the enzyme activity set by the user as a search condition with the actual measured value of the enzyme activity obtained by actually making koji using the extracted koji-making conditions.

[0244] The tenth embodiment is configured to use the enzyme titers output by inputting the koji-making conditions for database construction and data on multiple koji grains being broken into the second trained enzyme potency prediction model as the enzyme titers for database construction. This configuration makes it possible to create a large number of enzyme titers for database construction based on the koji-making conditions for database construction that have been created in large quantities.

[0245] The tenth embodiment of the koji-making support system 100 can extract candidate combinations of koji-making conditions and enzyme potencies based on one or more of the enzyme potency data. Therefore, by inputting all of the target enzyme potency data or a portion of the target enzyme potency data, the user can analyze the koji-making conditions required to obtain the target enzyme potency or enzyme potency data.

[0246] The present invention is not limited to the configuration of the 10th embodiment, and may be configured such that records consisting of actual measured values ​​of koji-making conditions and enzyme titers recorded in actual koji-making are added to the database for extracting candidate combinations of second koji-making conditions and enzyme titers created by the above-mentioned method, and the koji-making support system 100 may extract candidate combinations of koji-making conditions and enzyme titers using the actual measured value-added type database for extracting candidate combinations of second koji-making conditions and enzyme titers to which these records have been added.

[0247] The present invention is not limited to the tenth embodiment, and may be a database for extracting candidate combinations of koji-making conditions, multiple-grain-breaking data, and enzyme titers, which is composed of koji-making conditions, multiple-grain-breaking data, and enzyme titers. In this case, the multiple-grain-breaking data for koji can be used as the database-configuration multiple-grain-breaking data, which is output by inputting the koji-making conditions for database configuration into a trained multiple-grain-breaking data prediction model.

[0248] In the present invention, even when using a database for extracting candidate combinations of koji-making conditions, multiple-grain koji breaking data, and enzyme titers, one or more target data from any of the enzyme titer data constituting the enzyme titers can be used as search conditions. Furthermore, when using a database for extracting candidate combinations of koji-making conditions, multiple-grain koji breaking data, and enzyme titers, the number or type of target data can be increased to search conditions by adding one or more target data from any of the condition data constituting the koji-making conditions, or one or more target data from any of the breakage data constituting the multiple-grain koji breaking data, in addition to one or more target data from any of the enzyme titer data constituting the enzyme titers. By increasing the number or type of target data, optimal candidate combinations of koji-making conditions, multiple-grain koji breaking data, and enzyme titers can be extracted from the records constituting the database for extracting candidate combinations of koji-making conditions, multiple-grain koji breaking data, and enzyme titers.

[0249] When using a database for extracting candidate combinations of koji-making conditions, data on the breaking down of multiple koji grains, and enzyme activity, instead of using one or more target data from any of the enzyme activity data that make up the enzyme activity as the search criteria, one or more target data from any of the condition data that make up the koji-making conditions, or one or more target data from any of the breaking down data that make up the data on the breaking down of multiple koji grains may be used as the search criteria.

[0250] 11. Eleventh embodiment We will now explain the 11th embodiment of a koji-making support system that extracts condition data constituting future koji-making conditions that have been modified based on search conditions consisting of the target data's broken-seed distribution and the condition data constituting the koji-making conditions up to now during koji-making.

[0251] In the eleventh embodiment, the condition data acquired in each step of the actual koji-making process will be described as the actual measured value of the condition data. Also, the koji-making conditions extracted from the database for extracting candidate combinations of koji-making conditions and breaking distributions will be described as the predicted value of the koji-making conditions, and the condition data constituting the predicted value of the koji-making conditions will be described as the predicted value of the condition data.

[0252] In the eleventh embodiment, the user sets a target value for the breakage distribution. Furthermore, the koji-making support system 100 of the eleventh embodiment extracts candidate combinations of koji-making conditions and breakage distributions from a database for extracting candidate combinations of koji-making conditions and breakage distributions based on the target value for breakage distribution set by the user, using a method similar to that of the seventh embodiment. The user then actually produces koji according to the predictions of the extracted condition data. If the difference between the predicted value and the actual value of each condition data becomes large and there is a risk of deviation from the target breakage distribution, a new database for extracting candidate combinations of koji-making conditions and breakage distributions is created during koji-making, taking into account the actual measured values ​​of the condition data up to that point, in order to approach the target breakage distribution. The initially set target value for the breakage distribution is set in the created database for extracting candidate combinations of koji-making conditions and breakage distributions, and future condition data is acquired, allowing the koji-making conditions to be revised during koji-making.

[0253] In actual koji production, it is difficult to completely control all of the condition data, such as product temperature, so the predicted values ​​of each condition data may not be the same as the actual measured values ​​of each condition data. In such cases, if koji production is continued according to the predicted values ​​of the koji production conditions, the distribution of the broken grains in the final koji obtained by the koji production may deviate from the target distribution of broken grains. Therefore, in order to approach the target distribution of broken grains, it is necessary to modify the koji production conditions during the koji production.

[0254] The koji-making support system 100 of the eleventh embodiment creates koji-making conditions from the actual measured values ​​of the condition data and the predicted values ​​of the condition data during koji-making. Specifically, the koji-making support system 100 first uses the actual measured values ​​of the condition data for parts of the condition data that make up the koji-making conditions for which actual measured values ​​have already been acquired. Next, the koji-making support system 100 uses the predicted values ​​of the condition data that have been set so far for parts of the condition data that make up the koji-making conditions for which actual measured values ​​have not yet been acquired. In this way, reference koji-making conditions are created that are composed of the actual measured values ​​of each set of condition data and the predicted values ​​of each set of condition data.

[0255] The koji-making support system 100 of the 11th embodiment creates koji-making conditions for multiple database configurations based on the standard koji-making conditions, using the same method as in the 7th embodiment. Note that in the 11th embodiment, the portion of the condition data that constitutes the standard koji-making conditions, which consists of actual measured values, is a fixed value, so no level is set.

[0256] For example, if there are 12 types of condition data that make up the standard koji-making conditions, and among the condition data that make up the standard koji-making conditions, there are 7 types of condition data that are fixed, actual measured values, and 5 types of condition data that are predicted values, then three levels are set for each condition data that is the predicted value of the condition data, and all koji-making conditions that are made up of combinations of the set condition data are saved, resulting in 243 (3 5 Koji-making conditions for the database configuration (items) are created.

[0257] The koji-making support system 100 of the 11th embodiment uses the koji-making conditions for the database configuration to create a database for extracting candidate combinations of koji-making conditions and breaking-seed distributions, which is composed of koji-making conditions for the database configuration and breaking-seed distributions for the database configuration, in a manner similar to that of the 7th embodiment.

[0258] The koji-making support system 100 of the 11th embodiment uses the ratio data of koji classes that constitute the target value of the breaking distribution as target data, and searches for data from the created database for extracting candidate combinations of koji-making conditions and breaking distribution in a manner similar to that of the 7th embodiment, and extracts candidate combinations of koji-making conditions and breaking distribution.

[0259] The koji-making assistance system 100 of the eleventh embodiment acquires the extracted koji-making conditions as koji-making conditions that have been corrected during the koji-making process (hereinafter referred to as "corrected values ​​of the koji-making conditions").

[0260] 1, the koji-making support system 100 of the 11th embodiment, like the first embodiment, includes a storage means 101, an input means 102, a control means 103, and an output means 104. In the 11th embodiment, a database for extracting candidate combinations of koji-making conditions and breaking-seed distributions, which is composed of koji-making conditions for database configuration and breaking-seed distributions for database configuration, is stored in the storage means 101.

[0261] In the eleventh embodiment, the input means 102 is a keyboard, and the koji-making assistance system 100 can acquire the target values ​​of the breakage distribution and the actual measured values ​​of the condition data input as numerical values ​​by the input means 102.

[0262] In the koji-making assistance system 100 of the eleventh embodiment, the storage means 101 stores a koji-making assistance program 400. Furthermore, the control means 103 executes commands according to the koji-making assistance program 400.

[0263] As shown in Figure 12, the koji-making support program 400 of the eleventh embodiment includes an input step 401, a search step 402, and an output step 403. The input step 401 transfers the target value of the breakage distribution, the predicted value of the condition data, and the actual value of the condition data acquired by the input means 102 to the search step 402. The search step 402 creates a database for extracting candidate combinations of koji-making conditions and breakage distributions based on the predicted values ​​of the condition data and the actual values ​​of the condition data constituting the koji-making conditions. The search step 402 calculates the RMSE between the target value of the breakage distribution and the ratio data included in each record constituting the database for extracting candidate combinations of koji-making conditions and breakage distributions. The control means 103 then executes a process to acquire the record with the smallest RMSE as the corrected value of the koji-making conditions that meets the search criteria. The output step 403 displays the acquired corrected value of the koji-making conditions on the output means 104. By providing these steps, the koji-making assistance program 400 can cause the koji-making assistance system 100 to execute the process of correcting the koji-making conditions.

[0264] The koji-making support system 100 of the 11th embodiment evaluates the accuracy of the correction results by comparing the RMSE between the target value of the broken-sperm distribution and the broken-sperm distribution of the output koji when koji-making is continued according to koji-making conditions created from the actual measured values ​​of the condition data up to now and the predicted values ​​of the condition data that have been set up up to now, and the RMSE between the target value of the broken-sperm distribution and the broken-sperm distribution of the output koji when koji-making is continued according to koji-making conditions created from the actual measured values ​​of the condition data up to now and the corrected values ​​of the condition data from now.

[0265] In the 11th embodiment, as shown in Table 6, the RMSE between the initially set target value of the breaking distribution and the breaking distribution extracted from the database for extracting candidate combinations of koji making conditions and breaking distribution was 0.5. As shown in Table 6, when the accuracy of the correction results was evaluated, the RMSE between the initially set target value of the breaking distribution and the breaking distribution of the output koji when koji making was continued according to the koji making conditions created from the actual measured values ​​of the condition data so far and the predicted values ​​of the condition data set so far was 8.7. In addition, the RMSE between the initially set target value of the breaking distribution and the breaking distribution of the output koji when koji making was continued according to the koji making conditions created from the fixed values ​​of the condition data so far and the corrected values ​​of the condition data from now on was 2.6.

[0266] [Table 6]

[0267] As described above, the koji-making support system 100 of the 11th embodiment is capable of correcting the koji-making conditions and supporting koji-making, even if a difference occurs between the predicted values ​​of the koji-making conditions and the actual values ​​of the koji-making conditions during koji-making, so that a koji-breaking distribution closer to the target value of the koji-breaking distribution is obtained.

[0268] In the present invention, the timing for correcting the koji-making conditions is not limited to that of the eleventh embodiment, and may be set appropriately by comparing the predicted values ​​of the koji-making conditions with the actual measured values ​​of the koji-making conditions. For example, the koji-making conditions may be corrected when the difference between the predicted values ​​and the actual measured values ​​of the koji-making conditions exceeds a predetermined value.

[0269] In the present invention, the timing for inputting target data into the created database and extracting target combination candidates is not limited to the seventh, eighth, ninth, tenth, and eleventh embodiments, and may be set as appropriate. For example, in the seventh embodiment, the timing for inputting target data and extracting target combination candidates can be the end of the bed process. In this case, by inputting all of the ratio data of koji classes that make up the target breakage distribution and the actual measured values ​​of the condition data up to the end of the bed process as the target data, it is possible to obtain predicted values ​​for the remaining condition data that make up the koji-making conditions.

[0270] The koji-making support system and program of the present invention can be installed in a koji-making device. For example, if a target koji quality is set in advance, it is possible to select candidate koji-making conditions for obtaining koji of the target quality, automatically transfer the selected candidate koji-making conditions to the koji-making device, and automatically produce koji under the optimal koji-making conditions. [Explanation of symbols]

[0271] 100 Koji making support system 101 Memory means 102 Input Method 103 Control Means 104 Output Method 200 Trained Model Construction Methods 300 Koji Production Support Program 301 Input Step 302 Control Steps 303 Output Step 400 Koji Production Support Program 401 Input Step 402 Search Step 403 Output Step

Claims

1. A koji-making support system that predicts the distribution of broken spermatozoa from koji-making conditions, Teacher's koji making conditions and The training data is a training distribution generated by converting the koji grain classes corresponding to the plurality of koji output by the trained koji grain class estimation model into a ratio of each class, using as training data a training distribution generated by converting the koji grain class corresponding to the plurality of koji output by the trained koji grain class estimation model into a ratio of each class. The trained sperm distribution prediction model was constructed by applying machine learning to a learning model, which is a machine learning method. The koji-making conditions are input to predict the distribution of broken sperm. The trained koji grain class estimation model is Data on the breaking down of a single grain of koji for instructors in a format that can be processed by a computer, obtained by photographing a single grain of koji; The teacher koji one-grain class labels are used as teacher data, and the teacher koji one-grain class labels are classified into multiple classes based on the breakage circumference and breakage inclusion for one grain of koji corresponding to the teacher koji one-grain breakage data. It is built by running machine learning on a learning model, which is a machine learning technique. This koji-making support system is characterized by estimating the koji-grain class corresponding to the koji-grain breaking data for a single koji grain, which data is acquired by photographing the single koji grain and is in a format that can be processed by a computer.

2. A koji-making support system that predicts the enzyme activity of koji based on koji-making conditions and the distribution of broken spermatozoa, Teacher's koji making conditions and A teacher's koji-breaking distribution is generated by converting the koji-grain classes corresponding to the plurality of koji output by the trained koji-grain class estimation model according to claim 1 into a ratio of each class, using as input single-grain koji breakage data in a format that can be processed by a computer, obtained by photographing a single grain of koji obtained from koji made under the teacher's koji-making conditions; The teacher enzyme activity analysis value obtained by analyzing the koji made under the teacher koji-making conditions is used as teacher data, A first trained enzyme titer prediction model constructed by performing machine learning on a learning model, which is a machine learning method, A koji-making support system that predicts enzyme activity by inputting koji-making conditions and a distribution of broken spermatozoa corresponding to the koji-making conditions.

3. A koji-making support system that extracts candidate combinations of koji-making conditions and breaking distributions for search conditions, Koji making conditions for database configuration, A database for extracting combination candidates of koji-making conditions and breaking distributions is created by inputting the koji-making conditions for database configuration into the trained breaking distribution prediction model described in claim 1 and outputting the breaking distribution for database configuration, A koji-making support system characterized by searching a database for extracting candidate combinations of koji-making conditions and breaking distributions using one or more target data from the ratio data of one koji grain class that constitutes the breaking distribution as search conditions, and extracting candidate combinations of koji-making conditions and breaking distributions that satisfy the search conditions.

4. A koji-making support system that extracts candidate combinations of koji-making conditions and enzyme activity values ​​based on search criteria, Koji making conditions for database configuration, For the first trained enzyme activity prediction model described in claim 2, a database for extracting candidate combinations of first koji-making conditions and enzyme activity values ​​is created, which is composed of the koji-making conditions for database construction and the enzyme activity values ​​for database construction output by inputting the koji-making conditions for database construction and the breakage distribution output by the trained breakage distribution prediction model described in claim 1, A koji-making support system characterized by searching a database for extracting candidate combinations of the first koji-making conditions and enzyme titers using one or more target data from the enzyme titer data that constitute the enzyme titers as search conditions, and extracting candidate combinations of koji-making conditions and enzyme titers that satisfy the search conditions.

5. A koji-making support program for causing a computer to execute a process for predicting the distribution of broken grains from koji-making conditions, Teacher's koji making conditions and The training data is a training distribution generated by converting the koji grain classes corresponding to the plurality of koji output by the trained koji grain class estimation model into a ratio of each class, using as training data a training distribution generated by converting the koji grain class corresponding to the plurality of koji output by the trained koji grain class estimation model into a ratio of each class. The trained sperm distribution prediction model was constructed by applying machine learning to a learning model, which is a machine learning method. The koji-making conditions are input and the computer executes a process to predict the distribution of broken fermentation. The trained koji grain class estimation model is The data for the teacher on the breaking down of a single grain of koji is obtained by photographing a single grain of koji and is in a format that can be processed by a computer. The teacher koji one-grain class labels are used as teacher data, and the teacher koji one-grain class labels are classified into multiple classes based on the breakage circumference and breakage inclusion for one grain of koji corresponding to the teacher koji one-grain breakage data. It is built by running machine learning on a learning model, which is a machine learning technique. This koji-making support program is characterized by causing a computer to execute a process of estimating the koji grain class corresponding to the koji grain breaking data for a single koji grain that is obtained by photographing the single koji grain and is in a format that can be processed by a computer.

6. A koji-making support program for causing a computer to execute a process for predicting the enzyme activity of koji from koji-making conditions and the distribution of broken spermatozoa, Teacher's koji making conditions and A teacher's koji-breaking distribution is generated by inputting a plurality of single-grain koji-breaking data in a format that can be processed by a computer and that is acquired by photographing a single grain of koji obtained from koji made under the teacher's koji-making conditions, and converting the single-grain koji classes corresponding to the plurality of koji output by the trained single-grain koji class estimation model according to claim 1 into a ratio of each class; The teacher enzyme activity analysis value obtained by analyzing the koji made under the teacher koji-making conditions is used as teacher data, A first trained enzyme titer prediction model constructed by performing machine learning on a learning model, which is a machine learning method, A koji-making support program that inputs koji-making conditions and a distribution of broken spermatozoa corresponding to the koji-making conditions and causes a computer to execute a process of predicting enzyme activity.

7. A koji-making support program for causing a computer to execute a process of extracting combination candidates of koji-making conditions and breaking distributions for search conditions, Koji making conditions for database configuration, A database for extracting combination candidates of koji-making conditions and breaking distributions is created by inputting the koji-making conditions for database configuration into the trained breaking distribution prediction model described in claim 1 and outputting the breaking distribution for database configuration, A koji-making support program that causes a computer to execute a process of searching a database for extracting candidate combinations of koji-making conditions and koji-breaking distributions using one or more target data from the ratio data of one koji grain class that constitutes the koji-breaking distribution as search conditions, and extracting candidate combinations of koji-making conditions and koji-breaking distributions that satisfy the search conditions.

8. A koji-making support program for causing a computer to execute a process of extracting candidate combinations of koji-making conditions and enzyme activity values ​​based on search criteria, Koji making conditions for database configuration, For the first trained enzyme activity prediction model described in claim 2, a database for extracting candidate combinations of first koji-making conditions and enzyme activity values ​​is created, which is composed of the koji-making conditions for database construction and the enzyme activity values ​​for database construction output by inputting the koji-making conditions for database construction and the breakage distribution output by the trained breakage distribution prediction model described in claim 1, A koji-making support program that causes a computer to execute a process of searching a database for extracting candidate combinations of the first koji-making conditions and enzyme titers using one or more target data from among the enzyme titer data that constitute the enzyme titers as search conditions, and extracting candidate combinations of koji-making conditions and enzyme titers that satisfy the search conditions.

Images