Non-contact component estimation system

Abstract

This invention provides a non-contact component estimation system using near-infrared spectroscopy that can estimate various components of an object, such as its water content, with high accuracy, regardless of various state changes in the object. [Solution] The non-contact component estimation system comprises absorption spectrum acquisition devices 11 and 12 that acquire and output near-infrared absorption spectra of a substance, a feature data generation unit 22 that generates feature data from spectral data from the absorption spectrum acquisition devices 11 and 12, a machine learning model that takes input data based on the feature data as input and outputs component data of the substance, and a learning model management unit 6 that manages the machine learning model. The learning model management unit 6 generates a machine learning model by performing supervised machine learning on mixed feature data, which is a mixture of feature data from at least two or more of the aforementioned substances, each with different absorption spectral characteristics, as learning data.

Description

Technical Field

[0001] The present invention relates to a non-contact component estimation system for estimating various components such as moisture contained in a substance.

Background Art

[0002] Patent Document 1 discloses a technique for measuring the water content ratio of a soil material that is a ground material or an aggregate for concrete using a near-infrared moisture meter. In this technique, a part of the soil material is placed as a sample in a groove-shaped accommodating portion and relatively moved with respect to the near-infrared moisture meter. During this relative movement, near-infrared rays are irradiated from the near-infrared moisture meter onto the surface of the soil material in the accommodating portion of the container. Since moisture has the property of absorbing and attenuating near-infrared rays, the returned near-infrared rays are attenuated by the moisture in the soil material. The near-infrared moisture meter measures the moisture content of the soil material using the relationship (calibration curve) between the absorption amount of near-infrared rays in the soil material and the water content ratio of the soil material.

[0003] Patent Document 2 discloses a non-destructive moisture measurement device for measuring the moisture state in a living plant non-destructively by near-infrared spectroscopy. This non-destructive moisture measurement device includes a near-infrared spectroscopy sensor that irradiates a cultivated plant, which is the measurement object, with near-infrared light and non-destructively measures the absorbance (reflectance) of the near-infrared light of this cultivated plant, and a calibration curve showing the correspondence between the moisture content (water content rate) of the cultivated plant and the absorbance. The non-destructive moisture measurement device calculates the moisture content (water content rate) of the living cultivated plant using the absorbance measured by the near-infrared spectroscopy sensor and the data of the calibration curve.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] In the moisture measurement techniques described in Patent Documents 1 and 2, it is necessary to create a calibration curve in advance that shows the correspondence between the moisture content and absorbance of the object being measured, and the accuracy of this calibration curve affects the measurement accuracy. Furthermore, if conditions other than moisture content of the object being measured (e.g., surface condition) change, the measurement accuracy may deteriorate.

[0006] In view of the above circumstances, the object of the present invention is to provide a non-contact component estimation system using near-infrared spectroscopy that can estimate various components of an object, such as its water content, with high accuracy, despite various changes in the object's state. [Means for solving the problem]

[0007] The present invention provides a non-contact component estimation system for estimating various components such as water contained in a substance, comprising: an absorption spectrum acquisition device that acquires and outputs a near-infrared absorption spectrum of the substance; a feature data generation unit that generates feature data from the spectral data from the absorption spectrum acquisition device; a machine learning model that takes input data based on the feature data as input and outputs component data of the substance; and a learning model management unit that manages the machine learning model. The learning model management unit generates the machine learning model by performing supervised machine learning on mixed feature data, which is a mixture of the feature data of at least two or more substances, each with different absorption spectral characteristics, as training data.

[0008] In this configuration, a machine learning model is constructed that takes feature data generated from spectral data (near-infrared absorption spectral data) from an absorption spectrum acquisition device as input data and outputs component data of a substance (e.g., water content data). By providing spectral data, which is the near-infrared absorption spectrum of a substance whose components are to be estimated, to this machine learning model, the component data of that substance is output. To accurately estimate the component data of a substance, the construction of a machine learning model, that is, supervised machine learning for the machine learning model, is important. In this invention, mixed feature data, in which the feature data of at least two or more substances with different absorption spectral characteristics are mixed, is used as training data in machine learning. In machine learning, there is a tendency for estimation accuracy to improve by increasing the number of training data, but this also increases the risk of overfitting, which can suddenly cause a decrease in estimation accuracy. In this invention, instead of simply increasing the number of training data, by using mixed feature data, in which the feature data of at least two or more substances with different absorption spectral characteristics are mixed, as training data, overfitting is suppressed, and furthermore, it has become possible to generate a machine learning model with good estimation accuracy even with a relatively small number of training data. While mixed feature data can be used simultaneously, it is also possible to construct a final machine learning model by first building a machine learning model using feature data obtained from the substance to be estimated, and then later training it with feature data obtained from substances other than the substance to be estimated.

[0009] In one embodiment of the present invention, one of food, pharmaceuticals, resins, minerals, soil, wood, and paper materials is selected as the substance to be estimated, and at least one other than the selected substance to be estimated is selected as a reference substance. The mixed feature data includes the feature data obtained from the substance to be estimated and the feature data obtained from the reference substance. In this configuration, when constructing a machine learning model, not only the feature data obtained from the substance to be estimated, but also the feature data obtained from substances other than the substance to be estimated is used as training data. For example, if the substance to be estimated is food, not only the feature data of food but also the feature data of soil is used as training data. By using feature data obtained from substances other than the substance to be estimated that have the same components (component ratios) as the substance to be estimated, the absorption spectral characteristics of substances with different tissue structures from the substance to be estimated are also taken into consideration, overfitting is suppressed, and accurate component estimation becomes possible with a smaller amount of training data than a machine learning model trained using only the feature data of the substance to be estimated.

[0010] In one embodiment of the present invention, one of rice for eating, rice for sake, sweet potato, soybean, and wheat is selected as the grain to be estimated for component estimation, and at least one other grain besides the selected grain to be estimated for component estimation is selected as a reference grain, and the mixed feature data includes the feature data obtained from the grain to be estimated for component estimation and the feature data obtained from the reference grain. In this configuration, not only the feature data obtained from the grain to be estimated for component estimation but also the feature data obtained from different types of grains are used as training data. Even in such a case, the accuracy of component estimation is improved even with a smaller amount of training data than a training model generated using only the feature data obtained from the grain to be estimated for component estimation.

[0011] Furthermore, in one of the other embodiments of the present invention, one of rice for eating, rice for sake, sweet potato, soybean, and wheat is selected as the grain to be estimated for component estimation, and at least one other than the selected grain to be estimated for component estimation is selected as a reference grain, and the mixed feature data includes the feature data obtained from the grain to be estimated for component estimation, the feature data obtained from the reference grain, and the feature data obtained from substances other than grains. In this configuration as well, the accuracy of component estimation is improved even with a smaller amount of training data than a learning model generated using only the feature data obtained from the grain to be estimated for component estimation as training data.

[0012] While spectral data output from an absorption spectrum acquisition device may be processed in its original form and used as feature data, the absorption spectral intensity for a substance varies depending on the specific frequency band. To normalize such intensity fluctuations that differ for each specific frequency band, it is preferable to differentiate the absorption spectrum and use the resulting differential data. In particular, such normalization is advantageous in the present invention, where feature data for different substances are used. For this reason, the present invention is proposed to include a differential processing unit that differentiates the spectral data from the absorption spectrum acquisition device and outputs differential data, and a feature data generation unit that generates feature data from the differential data.

[0013] Since the spectral data output from the absorption spectrum acquisition device exhibits complex curves, higher-order differentiation is more suitable than first-order differentiation for its differential processing. However, considering the computational load, second-order differentiation is appropriate. Therefore, in this invention, it is proposed that the differential processing unit obtains the differential data by performing second-order differentiation on the spectral data.

[0014] When the substance whose components are to be estimated is a substance used in the manufacturing process at a manufacturing site, the absorption spectrum acquisition device is placed at each manufacturing process, so it is advantageous for the absorption spectrum acquisition device to have a compact configuration. For this reason, in the present invention, the absorption spectrum acquisition device is configured to have an infrared light sensor unit consisting of a light irradiation unit that irradiates the substance with infrared light and a light receiving unit that receives reflected light from the substance, and a spectroscopic unit that generates the near-infrared absorption spectrum from the reflected light. In this case, if the infrared light sensor unit and the spectroscopic unit are connected by a branch signal line, the spectroscopic unit can be used in common by multiple infrared light sensor units.

[0015] The learning model of the present invention is machine-learned using feature data from spectral data of different substances as training data. The locations where spectral data for each substance is generated, the locations where feature data from each spectral data is generated, and the locations where the learning model is generated using each feature data are likely to be far apart from each other. To overcome such geographical problems, the present invention proposes that the absorption spectrum acquisition device be located at the manufacturing site where products are manufactured from the substances, the feature data generation unit be located at a first service provider capable of data communication with the manufacturing site, the learning model management unit be located at a second service provider capable of data communication with the first service provider, and the machine learning model used in practice, built on the trained learning model data generated by the learning model management unit, be located at the first service provider.

[0016] Other features, functions, and effects of the present invention will be revealed by the following description of the invention with reference to the drawings. [Brief explanation of the drawing]

[0017] [Figure 1] This is a schematic diagram of a non-contact component estimation system. [Figure 2] This is a schematic diagram of an infrared light sensor unit. [Figure 3] This is a schematic spectral diagram showing an example of the near-infrared absorption spectrum of an object being examined. [Figure 4] It is a graph schematically showing the second derivative data of the near-infrared absorption spectrum. [Figure 5] It is a schematic diagram showing a graph showing the change in moisture during the sake brewing process, which is an example of component information. [Figure 6] It is a schematic diagram explaining the process of generating learning data for constructing a learning model. [Figure 7] It is a table showing the evaluation results of a learning model learned using a combination of learning data obtained from each of Koshihikari, sake rice koji rice in the dried state, and sake rice steamed rice. [Figure 8] It is a table showing the evaluation results of a learning model learned using a combination of learning data obtained from each of cooked rice, sake rice koji rice, sake rice steamed rice, and cooked rice. [Figure 9] It is a table showing the evaluation results of a learning model learned using a combination of learning data obtained from each of cooked rice, sake rice koji rice, sake rice steamed rice, and sweet potato. [Figure 10] It is a schematic diagram showing a distributed non-contact component estimation system using a data communication network.

Embodiments for Carrying Out the Invention

[0018] In the following embodiments, the non-contact component estimation system according to the present invention is configured as a non-contact moisture estimation system. In a system for estimating various components other than moisture (such as lipids and proteins), in the following description, "moisture" is replaced with "component to be estimated (such as lipids and proteins)". The schematic configuration of this non-contact moisture estimation system will be described using FIG. 1. This non-contact moisture estimation system estimates the moisture contained in substances such as food, grains, drugs, resins, minerals, soil, wood, and paper materials as inspection objects. Representative names are listed as the substance names here, but each substance contains many types of names. For example, grains include cooked rice, sake rice, sweet potato, soybeans, wheat, and the like.

[0019] The non-contact moisture estimation system consists of an absorption spectrum acquisition device 1, a data preprocessing unit 2, a learning model unit 4, a moisture information notification unit 5 as a component information notification unit, and a learning model management unit 6.

[0020] The absorption spectrum acquisition device 1, which generates and outputs near-infrared absorption spectra of a material, comprises an infrared light sensor unit 11 and a spectroscopic unit 12. As shown in Figure 2, the infrared light sensor unit 11 has a light irradiation unit 11a that irradiates the material with infrared light (near-infrared light or infrared light including near-infrared light) and a light receiving unit 11b that receives the reflected light that returns after the infrared light irradiated onto the material is reflected by the material. The spectroscopic unit 12 spectrally analyzes the reflected light sent from the light receiving unit 11b via an optical fiber to generate the near-infrared absorption spectrum of the material being inspected. Figure 3 shows the near-infrared absorption spectra of rice in three different states over time in the sake brewing process (soaked rice, steamed rice, and cooled rice) as the material being inspected. In this near-infrared absorption spectrum, the horizontal axis is wavelength and the vertical axis is absorbance (corresponding to moisture content). Note that in Figure 2, the absorption spectrum acquisition device 1 is schematically illustrated for explanatory purposes, but it can be configured in various forms depending on its application. For example, if the absorption spectrum acquisition device 1 is incorporated into a rice cooker or the like, it will be used in a form housed in a heat-resistant box. Furthermore, if an infrared light transmission window is formed, a configuration will be adopted in which infrared light is passed through the infrared light transmission window to perform infrared light measurements.

[0021] The data preprocessing unit 2 generates feature data from the spectral data obtained from the absorption spectrum acquisition device 1. In this embodiment, the data preprocessing unit 2 is equipped with a feature data generation unit 22 that generates feature data that represents the characteristics of the spectral data. The absorption spectrum represented by the spectral data is a wave curve in a coordinate system where the vertical axis is absorbance and the horizontal axis is wavelength (or wave number), as shown in Figure 3, and the feature data generation unit 22 converts the shape features of such a wave curve into feature data. A simple and effective conversion technique for characterizing the shape of a wave curve is to differentiate the wave curve. Particularly preferred is to perform a second derivative on the wave curve. Since the absorption spectrum has extrema in multiple specific frequency bands, it is particularly preferable to calculate feature data from the second derivative values ​​in multiple specific frequency bands that show extrema. For this reason, the data preprocessing unit 2 is equipped with a differentiation processing unit 21 that performs a second derivative on the spectral data for multiple predetermined wavelength region units, and a feature data generation unit 22 that converts the second derivative data from the differentiation processing unit 21 into feature data. In this invention, the differential processing is not limited to second derivatives, so a configuration may be adopted in which the differential processing unit 21 performs first derivatives, and the feature data generation unit 22 converts the first derivative data from the differential processing unit 21 into feature data. In the generation of feature data from differential data by the feature data generation unit 22, the second derivative values ​​in each specific frequency band may be used directly as feature data, but it is also possible to use vector data as feature data, in which the second derivative values ​​in each specific frequency band and preferably the group of second derivative values ​​over time are also vector elements. Figure 4 shows a schematic graph of second derivative data, which is the result of second derivative processing of the near-infrared absorption spectra of rice in its state over time (soaked rice, steamed rice, cooled rice) as shown in Figure 3. In this graph, the horizontal axis is wavelength, and the vertical axis is second derivative absorbance.

[0022] The learning model unit 4 shown in Figure 1 can be constructed as a linear regression model using partial least squares regression (PLS regression), but in this embodiment, it is constructed as a nonlinear regression model such as a neural network or a random forest model. In any case, the learning model unit 4 comprises a machine learning model 40 (hereinafter simply referred to as the learning model 40) and an input data creation unit 41 that creates input data to be input to the learning model 40 from feature data. Furthermore, the learning model unit 4 includes a moisture content determination unit 42 that determines the moisture content of the inspected object from the output data as an estimation result output from the learning model 40. The moisture content determination unit 42 can also take feature data of spectral data obtained from measurements of the inspected object at each manufacturing process or from time-series measurements performed at a specific manufacturing process (for example, measurements of steamed rice continuously discharged in the steaming process, or measurements at each unit steaming time in the steaming process) as input data, determine the time-series moisture content of the inspected object at each manufacturing process from the output time-series output data, and output the moisture content as time-series data.

[0023] The moisture content output from the moisture content determination unit 42 is converted into notification data by the moisture information notification unit 5 and transmitted as moisture notification information to the department that needs it, for example, a computer installed at the manufacturing site (an example of a site management computer 7) or a portable terminal carried by the manufacturing manager at the manufacturing site (an example of a site management computer 7). In order for the manufacturing manager to grasp the material moisture content over time in the manufacturing process, moisture data as component data estimated at multiple manufacturing times in the manufacturing process is necessary. Therefore, the manufacturing time is attached to the moisture data as attribute information along with data related to the manufacturing process. An example of moisture notification information is shown in Figure 5. The moisture notification information exemplified here is a graph showing the moisture content (moisture ratio) estimated using near-infrared absorption spectra in each process of sake production, which consists of brown rice, polishing, soaking, steaming, aging, koji making, and fermentation.

[0024] The learning model unit 4, particularly the learning model 40, is substantially composed of a computer program. By installing the trained learning model data, which is a computer program, into the computer, or by providing various coefficients to the learning model program, the computer functions as the learning model 40, and consequently as the learning model unit 4. Of course, at least a part of the learning model unit 4 may be constructed from hardware. The management of such a learning model unit 4, particularly the learning model 40, is performed by the learning model management unit 6.

[0025] The learning model management unit 6 generates a learning model 40 by performing supervised machine learning on mixed feature data, which is a mixture of feature data from at least two or more substances (tested objects) with different absorption spectral characteristics. Here, the at least two or more substances with different absorption spectral characteristics are selected from heterogeneous materials such as food, grains, pharmaceuticals, resins, minerals, soil, wood, and paper.

[0026] The specific method for generating the learning model 40 by the learning model management unit 6 will be explained below using Figure 6. In this embodiment, the learning model 40 outputs moisture data for a test object (substance to be estimated for moisture content) which is one of the substances consisting of food, grains, medicines, resins, minerals, soil, wood, and paper materials. The learning data used to construct this learning model 40 is mixed feature data which includes feature data obtained from the substance to be estimated (test object) and feature data obtained from at least one other substance (reference substance) among the substances to be estimated.

[0027] Figure 6 schematically illustrates the process of generating mixed training data for constructing the learning model 40. Here, the mixed training data used includes feature data obtained by performing a second differential operation on spectral data, which is the near-infrared absorption spectrum of the first substance (the test object) (referred to here as principal feature data to distinguish it from the feature data of other substances), and feature data obtained by performing a differential operation on spectral data, which is the near-infrared absorption spectrum of the first and second substances (reference substances) other than the test object (referred to here as first and second minor feature data to distinguish them from the feature data of other substances). Of course, third minor feature data, fourth minor feature data, ... may also be used.

[0028] The training data creation unit 61 creates primary training data from primary feature data, primary secondary training data from primary secondary feature data, and secondary secondary training data from secondary secondary feature data. The total number of training data points can be around one thousand. Although not shown in the diagram, this machine learning is supervised machine learning, so the measured water content of each substance from which the feature data was obtained is added to the training data as training data.

[0029] The learning model 40a is trained using mixed learning data, which contains a mixture of primary learning data, first secondary learning data, and second secondary feature data. Once training is complete, the learning model management unit 6 outputs the trained learning model data and provides it to the learning model 40 of the learning model unit 4 (see Figure 1).

[0030] Figure 7 shows the model evaluation of moisture estimation in a learning model 40 trained using only training data obtained from rice (Koshihikari), a learning model 40 trained using only training data obtained from sake rice koji rice, a learning model 40 trained using only training data obtained from sake rice steamed rice, and a learning model 40 trained using mixed training data in which the above training data are mixed as primary training data, first secondary training data, and second secondary training data. This model evaluation is expressed as the mean absolute error (MAE). In the example in Figure 7, the number of data points used is the number of actual data points and the number of data points obtained by copying the actual data and amplifying it to more than 1000, with the number of actual data points shown in parentheses. The model evaluation for each number of data points is also shown, with the model evaluation for the number of actual data points shown in parentheses. Figure 7 shows that the model evaluation with the number of copied and amplified data points is the same as or slightly better than the model evaluation with the number of actual data points. Furthermore, compared to the learning model 40 trained using single-type learning data consisting only of rice, which is the grain targeted for moisture estimation (grain targeted for component estimation, substance targeted for component estimation), the learning model 40 trained using mixed learning data, which includes learning data for sake rice koji and sake rice steamed rice mixed with the learning data for rice, shows a better or nearly identical model evaluation.

[0031] Figure 8 shows the model evaluation of moisture estimation in the learning model 40 trained using only single-type training data obtained from rice (Koshihikari), the model evaluation of moisture estimation in the learning model 40 trained using only single-type training data obtained from sake rice koji rice, the model evaluation of moisture estimation in the learning model 40 trained using only single-type training data obtained from sake rice steamed rice, the model evaluation of each moisture estimation in the learning model 40 trained using integrated mixed training data which combines training data obtained from rice, sake rice koji rice, and sake rice steamed rice, the model evaluation of moisture estimation in the learning model 40 trained using only training data obtained from cooked rice, and the model evaluation of moisture estimation in the learning model 40 trained using integrated mixed training data which adds training data obtained from cooked rice instead of rice training data. Here again, the model evaluation is the mean absolute error (MAE). The model evaluation of moisture estimation in the learning model 40 trained using only single-type training data is shown in parentheses. Figure 8 shows that the model evaluation of the learning model 40 trained using integrated mixed learning data, which includes learning data obtained from cooked rice instead of learning data from raw rice, is superior to the model evaluation of the other learning models 40.

[0032] Figure 9 shows the evaluation results of the learning model 40, which was trained using combinations of training data obtained from table rice (Koshihikari), sake rice koji, sake rice steamed rice, and sweet potato. The model evaluation of moisture estimation in the learning model 40 trained using only single-type training data obtained from table rice (Koshihikari), the model evaluation of moisture estimation in the learning model 40 trained using only single-type training data obtained from sake rice koji, the model evaluation of moisture estimation in the learning model 40 trained using only single-type training data obtained from sake rice steamed rice, and the model evaluation of moisture estimation in the learning model 40 trained using integrated mixed training data that combines training data obtained from table rice, sake rice koji, and sake rice steamed rice are the same as in Figure 8. The model evaluation of moisture estimation for sweet potato in the learning model 40 trained using only training data obtained from sweet potato is not good. However, in the integrated mixed training data, the model evaluation of sweet potato moisture estimation in the training model 40 trained using integrated mixed training data, which includes training data obtained from sweet potatoes instead of training data obtained from sake rice koji rice, is improved compared to the model evaluation of the training model 40 trained using only training data obtained from sweet potatoes.

[0033] Furthermore, one of the following grains—rice for eating, rice for sake, sweet potato, soybean, or wheat—is selected as the grain for moisture estimation, at least one of the other grains is selected as the reference grain, and a substance other than a grain (such as clay) may also be selected as the reference substance. In this case, the mixed feature data includes feature data obtained from the grain for moisture estimation (primary feature data), feature data obtained from the reference grain (first secondary feature data), and feature data obtained from the substance other than a grain (such as clay) (second secondary feature data).

[0034] Next, we will describe some specific embodiments. (Generation of general-purpose learning models) For rice with varying moisture content, the second derivative data (relationship between wavelength and second derivative absorbance) of the near-infrared absorption spectrum (showing the relationship between wavelength and absorbance) is obtained. Feature data generated from each second derivative data set is used to create training data, and this training data is used to generate a general-purpose learning model for moisture estimation.

[0035] (Application to the sake brewing process) A general-purpose learning model is further trained using training data created from feature data generated from the second derivative data (relationship between wavelength and second derivative absorbance) of near-infrared absorption spectra (showing the relationship between wavelength and absorbance) in the raw rice processing process, which consists of milling, soaking, steaming, aging, koji making, and fermentation, thereby generating a learning model for sake brewing. The generated learning model for sake brewing can estimate the moisture content based on the near-infrared absorption spectra from absorption spectrum acquisition devices 1 placed at each process.

[0036] (Application to the rice manufacturing process) A general-purpose learning model is further trained using training data created from feature data generated from the second derivative data (relationship between wavelength and second derivative absorbance) of near-infrared absorption spectra (showing the relationship between wavelength and absorbance) in the rice manufacturing process, which consists of milling, soaking, steaming (cooking), and molding, thereby generating a learning model for rice manufacturing.

[0037] (Application to the soy sauce manufacturing process) A general-purpose learning model is further trained using training data created from feature data generated from the second derivative data (relationship between wavelength and second derivative absorbance) of near-infrared absorption spectra (showing the relationship between wavelength and absorbance) in the soy sauce manufacturing process, which consists of processes such as milling, soaking, steaming, roasting, koji making, and fermentation, using soybeans alone or soybeans and wheat as raw materials, to generate a learning model for soy sauce manufacturing.

[0038] (Application to the shochu manufacturing process) A general-purpose learning model is further trained using feature data generated from the second derivative data (relationship between wavelength and second derivative absorbance) of the near-infrared absorption spectrum (showing the relationship between wavelength and absorbance) at each stage of the shochu manufacturing process, which consists of steaming, drying, and roasting, using sweet potatoes, potatoes, or rice as raw materials. This training data is then used to generate a learning model for shochu manufacturing.

[0039] (Application of heterogeneous substances in moisture estimation) For example, in generating a learning model 40 that uses grains as the target for moisture estimation, not only learning data obtained from grains, but also learning data obtained from other materials (food, medicine, resin, mineral, soil, wood, paper, etc.) may be used. Conversely, in generating a learning model 40 that uses the above-mentioned other materials as the target for moisture estimation, learning data obtained from grains, etc., may be used in addition to learning data obtained from other materials.

[0040] As described above, when estimating the water content of a new substance, it is possible to estimate the water content of the new substance by further training a general-purpose learning model using training data created from the characteristic data of the new substance. Since this only involves further training of a general-purpose learning model, it has the advantage of requiring less training data compared to generating training model data for estimating the water content of a new substance from scratch. Furthermore, if a general-purpose learning model is not available, the learning model 40a can be trained using training data created from the characteristic data of multiple substances, and even in this case, satisfactory training model data can be generated with a small amount of training data.

[0041] Next, a distributed non-contact moisture estimation system will be explained using Figure 10. In this non-contact moisture estimation system, multiple manufacturing sites or processes, a first service provider SP1, and a second service provider SP2 are interconnected by data communication lines such as the Internet, public data communication lines, or private data communication lines. Each manufacturing site or process is equipped with a site management computer 7, an absorption spectrum acquisition device 1, and a data preprocessing unit 2. The first service provider SP1 is equipped with a learning model unit 4 and a moisture information notification unit 5. The second service provider SP2 is equipped with a learning model management unit 6. In this distributed non-contact moisture estimation system, the learning model management unit 6 of the second service provider SP2 generates trained learning model data for the learning model 40 based on the feature data (first feature data, second feature data, third feature data, ...) sent from each manufacturing site or process and their respective training data (actually measured moisture content). The machine learning model 40 used in practice is provided in the learning model unit 4 of the first service provider SP1 and is built based on trained learning model data sent from the learning model management unit 6. The learning model unit 4 built in the first service provider SP1 estimates the moisture content of the target material based on feature data (first feature data, second feature data, third feature data, ...) sent from each manufacturing site or each manufacturing process.

[0042] [Another embodiment] (1) In the embodiments described above, it was assumed that the same learning model 40 would be used for various test objects, but different learning models 40 may be used depending on the test object. For example, depending on the test object, an appropriate one can be selected from linear regression, decision trees, random forests, neural networks, support vector machines, etc. In this case, appropriate training data will be created from the same feature data.

[0043] (2) In the embodiment described above, the data preprocessing unit 2, the learning model unit 4, the moisture information notification unit 5, and the learning model management unit 6 were each constructed as separate units (computers). However, in small-scale manufacturing facilities, all of these may be constructed as a single computer. In this case, the reception of spectral data from the absorption spectrum acquisition device 1 may be done online or by batch processing using a memory device.

[0044] (3) In the embodiments described above, the non-contact component estimation system was configured as a non-contact moisture estimation system. However, it can be configured as a non-contact component estimation system that measures lipids, proteins, and various other components that have absorption in the near-infrared region, and in that case, the components and measurement methods described in the embodiments described above can be applied.

[0045] Furthermore, the configurations disclosed in the above embodiments (including other embodiments, the same applies hereinafter) can be applied in combination with configurations disclosed in other embodiments, as long as no inconsistencies arise. Moreover, the embodiments disclosed herein are illustrative, and the embodiments of the present invention are not limited thereto, and can be modified as appropriate without departing from the object of the present invention. [Industrial applicability]

[0046] This invention is applied to a non-contact component estimation system that estimates various components of a substance, such as water content, using near-infrared absorption spectroscopy. [Explanation of symbols]

[0047] 1: Absorption spectrum acquisition device 2: Data preprocessing unit 4: Learning Model Unit 5: Moisture Information Notification Unit (Component Information Notification Unit) 6: Learning Model Management Department 7: Site Management Computer 11: Infrared light sensor unit 11a: Light irradiation part 11b: Light receiving part 12: Spectroscopic Unit 21: Differential Calculus Section 22: Feature Data Generation Unit 40: Learning Model (Machine Learning Model) 40a: Learning model for training 41: Input Data Creation Section 42: Moisture content determination unit 61: Training Data Creation Department SP1: First Service Provider SP2: Second Service Provider

Claims

1. A non-contact component estimation system that estimates various components such as water contained in a substance, An absorption spectrum acquisition device that acquires and outputs the near-infrared absorption spectrum of the aforementioned substance, A feature data generation unit that generates feature data from spectral data from the absorption spectrum acquisition device, A machine learning model that takes input data based on the aforementioned feature data as input and outputs component data of the substance, A learning model management unit that manages the aforementioned machine learning model, Equipped with, The learning model management unit is a non-contact component estimation system that generates a machine learning model by performing supervised machine learning on mixed feature data, which is a mixture of the feature data of at least two or more substances, each with different absorption spectral characteristics, as learning data.

2. A non-contact component estimation system according to claim 1, wherein one of food, medicine, resin, mineral, soil, wood, and paper material is selected as a component estimation target substance, at least one other than the selected component estimation target substance is selected as a reference substance, and the mixed feature data includes the feature data obtained from the component estimation target substance and the feature data obtained from the reference substance.

3. A non-contact component estimation system according to claim 1, wherein one of rice for eating, rice for making sake, sweet potato, soybean, and wheat is selected as a component estimation target grain, and at least one other than the selected component estimation target grain is selected as a reference grain, and the mixed feature data includes the feature data obtained from the component estimation target grain and the feature data obtained from the reference grain.

4. A non-contact component estimation system according to claim 1, wherein one of rice for eating, rice for making sake, sweet potato, soybean, and wheat is selected as a grain to be estimated for component estimation, and at least one other than the selected grain to be estimated for component estimation is selected as a reference grain, and the mixed feature data includes the feature data obtained from the grain to be estimated for component estimation, the feature data obtained from the reference grain, and the feature data obtained from the substance other than the grain.

5. The non-contact component estimation system according to claim 1, further comprising a differential processing unit that performs differential processing on the spectral data from the absorption spectrum acquisition device and outputs differential data, and a feature data generation unit that generates the feature data from the differential data.

6. The non-contact component estimation system according to claim 5, wherein the differential processing unit obtains the differential data by performing a second derivative on the spectral data.

7. The non-contact component estimation system according to claim 1, wherein the absorption spectrum acquisition device comprises an infrared light sensor unit consisting of a light irradiation unit that irradiates the substance with infrared light and a light receiving unit that receives reflected light from the substance, and a spectroscopic unit that generates the near-infrared absorption spectrum from the reflected light.

8. The absorption spectrum acquisition device is located in a manufacturing process for producing a product from the substance, the feature data generation unit is located in a first service provider capable of data communication with the manufacturing process, and the learning model management unit is located in a second service provider capable of data communication with the first service provider. The non-contact component estimation system according to any one of claims 1 to 7, wherein the machine learning model used in practice is constructed based on the trained learning model data generated by the learning model management unit and is deployed to the first service provider.

Images