JP7842442B2 - AI-based analysis method and AI analysis device - Google Patents

Description

This invention relates to an AI-based analysis method , an AI analyzer, and a program for operating an AI analyzer, particularly characterized by the inability to evaluate using a single function, and instead requiring the classification of events and data, followed by the use of functions prepared for each classification. For example, regarding soil analysis, it relates to a method for analyzing soil samples to evaluate soil fertility characteristics, including component analysis and chemical properties such as cation exchange capacity (CEC) and anion exchange capacity (AEC). It is also a technology that can be applied to evaluating the properties of materials when their composition is unknown, such as in welding.

As an analysis method using AI, Patent Document 1 discloses a method for learning a predictive model, a device for learning a predictive model, and a plant control system. In the learning of the predictive model described in Patent Document 1, the predictive model for predicting the target variable from at least one explanatory variable is updated based on a first evaluation value and a second evaluation value. The first evaluation value is calculated as an index indicating the prediction error of the predictive model, and the second evaluation value is calculated as an index indicating the degree of agreement between the feature quantities related to the sensitivity direction of at least some of the explanatory variables to the target variable and the tolerance range set for the feature quantities based on known information.

In this case, the relationship between the explanatory variables and the dependent variable is classified, and the explanatory variables themselves are also expressed as sensitivity directional errors. That is, the correlation between the explanatory and dependent variables is classified in advance, and the error is predicted based on this classification of correlations. Therefore, error estimation also requires the use of different calculation formulas for each pre-classified error. In other words, the explanatory function used to determine the explanatory variables is always unique, and the learning efficiency is improved by pre-classifying the relationships included in the errors.

Conventionally, for soil analysis, a soil analysis apparatus and method, such as that described in Patent Document 2, are known. According to Patent Document 2, an alternating magnetic field is applied to the soil to be analyzed, the detection signal output from the detection coil is measured and processed by the measurement unit, and the correlation data between the quantitative values of soil fertility traits stored in the memory unit and the estimated values of soil fertility traits obtained from the detection signal is used by the estimation unit to estimate the soil fertility traits, including the CEC, of the soil to be analyzed based on the detection signal. Here, the estimation unit can accurately estimate the values of soil fertility traits, including the CEC, by creating the correlation data using so-called PLS regression analysis during the estimation of soil fertility traits.

Furthermore, according to the third embodiment described in Patent Document 2, the estimation unit can perform CEC estimation in a simple manner by using the detection signal output from the detection coil in response to the excitation signal as a parameter, and can also measure soil fertility traits other than CEC.

Patent No. 6831030 Patent No. 6562490

National Agriculture and Food Research Organization, Japan Soil Inventory, https://soil-inventory.dc.affrc.go.jp/

The prediction model learning method described in Patent Document 1 uses a pre-constructed prediction model. Although it can be updated using machine learning, it cannot make predictions using any model other than the one newly trained with data prepared separately for training.

Regarding errors in measurement data, there are variations due to the measurement object itself, external factors, and human factors. However, since so-called systematic errors can be easily classified using other methods, this invention focuses only on so-called random errors.

Incidentally, in soil analysis according to Patent Document 2, the accuracy of the analysis sometimes decreased depending on the type and condition of the soil, making it difficult to perform accurate soil analysis. As with the soil analysis described above, measurement targets where the accuracy of the analysis decreases depending on the type and condition are often found in phenomena that depend on the natural environment or the skills of the worker.

Classification is typically performed by academic experts based on a comprehensive assessment of analysis results using diverse analytical methods. However, this approach is currently far from being feasible in the industrial sector due to its lack of immediacy and ease of use. Therefore, the inventors of this invention have discovered an AI-based analysis method that allows for the selection of various predictive models and the construction of new models based on the state of explanatory variables, thereby completing this invention.

In view of the above, the present invention aims to provide an AI-based analysis method, an AI analysis device , and a program for operating this AI analysis device, which enable easy and highly accurate automation of analysis and evaluation without being affected by measurement conditions including accidental errors such as external factors and human factors.

The above objective is that, according to the present invention, in order to analyze a first objective variable relating to the features contained in the object of analysis , measurement data is obtained from the object of analysis, and a first estimation function is obtained using the measurement data as explanatory variables , thereby estimating the first objective variable ,
The object of analysis is any of the following: soil, concrete structure, ceramic structure, metal including alloys, and metal structure including alloys that has undergone processing involving alteration, wherein the material, material properties, crystalline state, material morphology, and material structure exhibit multiple states.
If the correlation between the estimated value of the first estimation function and the quantitative value of the first dependent variable is 0.7 or less and cannot be uniquely determined ,
The system is configured to classify the aforementioned measurement data into multiple target variables using a second estimation function, and to obtain multiple first estimation functions for each classification.
In the first step,
In order to select the first estimation function to be used from among the multiple first estimation functions created for each classification, the second objective variable is estimated by classifying the second estimation function using the measurement data, and the first estimation function to be used is selected.
In the second step,
This is achieved by an AI analysis method that uses measurement data related to the first estimation function selected in the first step as explanatory variables, and estimates the first objective variable using the first estimation function .

In the above configuration, preferably, evaluation data that serves as explanatory variables are used as training data, and the first or second objective variable is obtained as quantitative values of the training data and used as comparison data, and AI analysis is performed.
Preferably, classification is performed using a second estimation function with evaluation data, and based on the obtained classification results, a first estimation function for the first dependent variable defined for each classification in the classification data is selected, the first dependent variable is estimated using the first estimation function with evaluation data, and the classification of the subject of analysis and the estimated value of the first dependent variable are determined.
Preferably, in order to create multiple classified first estimation functions, when the entire measurement data is analyzed and estimated using the same analytical method, multiple classifications are created by classifying the degree of deviation of the estimated values from the training data. Preferably, multiple first estimation functions are created by classification using the similarity of the measurement data.

According to the above configuration, regarding the measurement data, first, multiple classifications are defined by comparing the quantitative value of the target variable to be analyzed with corresponding comparison data, and an estimation function is defined for each classification. Then, a corresponding classification is selected for each measurement data, and the target variable to be analyzed is analyzed and evaluated from the measurement data using the estimation function corresponding to the selected classification. Therefore, since classifications are defined using the measurement data, it is not necessary to acquire data for defining classifications separately from the measurement data, which can simplify the process. Furthermore, since the classification to which the measurement data should be assigned is selected before analyzing and evaluating the measurement data, and the analysis and evaluation of the measurement data is performed using the estimation function corresponding to the selected classification, it becomes possible to perform the analysis and evaluation of the measurement data more accurately.

In the above configuration, an estimation function corresponding to the selection of classification, preferably a second estimation function corresponding to the selection of classification, is obtained by one of the following analysis methods: linear analysis, PLS regression analysis, SVM (Support Vector Machine), neural network, or supervised AI analysis method.
Multiple first estimation functions corresponding to the target variable are preferably obtained by one of the following analytical methods: clustering, principal component analysis, unsupervised AI analysis, linear analysis, PLS regression analysis, SVM (Support Vector Machine), neural network, or supervised AI analysis. Measurement data are preferably obtained by one of the following methods using multiple frequency bands: electromagnetic measurement, acoustic measurement, or optical measurement.
Preferably, the primary objective variable is to obtain estimated values of soil fertility characteristics, including agricultural indicators of soil such as cation exchange capacity (CEC), base saturation, silicic acid, iron oxide, and nutrients such as potassium, magnesium, calcium, iron, manganese, zinc, copper, molybdenum, boron, chlorine, nickel, nitrogen, carbon, and phosphorus. Preferably, the measurement data is obtained by generating an alternating magnetic flux by applying multiple excitation signals of different frequencies to an excitation coil, and detecting the disturbed magnetic flux depending on the state of the object being analyzed using a detection coil. Preferably, the excitation signals have a frequency band of about 5 to 100 times from the lowest frequency to the highest frequency. Preferably, the excitation signals have five or more different frequencies, and the number of excitation signals is one or more greater than the sum of the number of analyses and the number of classifications.
Preferably, when the total number of each item is evaluated using the same evaluation function for one or more items to be analyzed , the standard deviation is used to classify the items based on their degree of deviation from the correlation, such as -3σ or less, -3σ to -1σ, ±1σ, 1σ to 3σ, and 3σ or more.
Furthermore, the present invention relates to an AI analysis device in which a material, material composition, crystal state, material form, or material structure exhibits multiple states, and which analyzes any of the following: soil, concrete structures, ceramic structures, metals including alloys, and metal structures including alloys that have undergone processing involving alteration; and in order to analyze a first objective variable relating to the features contained in the analysis target, measurement data is obtained from the analysis target, and the measurement data is used as explanatory variables to obtain a first estimation function for the first objective variable, thereby estimating the first objective variable .
A sensor having a coil,
A measurement unit that generates excitation signals for each frequency to be input to a coil in order to apply an alternating magnetic field to the object to be analyzed, and processes the detection signal output from the coil based on the generated excitation signals,
A storage unit that stores data relating to the correlation between the quantitative value of the target variable of each analysis target and the sensor and measurement unit, for multiple analysis targets,
An estimation unit estimates the target variable to be analyzed using the data stored in the memory unit, based on the detection signal processed by the measurement unit.
It is equipped with,
The estimation unit obtains the complex amplitude ratio of the detection signal to the excitation signal as measurement data from the detection signal, defines multiple classifications from the quantitative values of multiple different target variables stored in the memory unit and the corresponding complex amplitude ratios of the detection signals as comparison data, and defines a first estimation function for each classification.
This is achieved by an AI analysis device in which, for each measurement data, the estimation unit selects a classification corresponding to that measurement data using a second estimation function , and then estimates the first target variable to be analyzed from the measurement data using a first estimation function corresponding to the selected classification.
The above-mentioned AI analysis device program is achieved by a program that enables a computer to function as an AI analysis device.

According to the present invention, by adding classifications of the objects to be analyzed, it is possible to provide an AI analysis method , an AI analysis device, and a program for operating this AI analysis device that can accurately determine soil fertility characteristics such as soil CEC.

This is a flowchart illustrating an example of the AI analysis method of the present invention. This is a block diagram showing the overall configuration of one embodiment of the analytical apparatus according to the present invention. This is an explanatory diagram illustrating the analytical method of the present invention. This figure shows the relationship between the estimated and quantitative values of CEC in Comparative Example 1. This figure shows the relationship between the estimated and quantitative values of CEC in Example 1. This figure shows the relationship between the estimated and quantitative values of CEC in Example 2. This figure shows the results of grouping using the k-means method in Comparative Example 2.

Embodiments of the present invention will be described below with reference to the drawings.
The AI analysis method of the present invention has the following features.
In other words, in an AI analysis method that obtains an estimation function for a first objective variable using measurement data that serves as explanatory variables acquired to analyze a first objective variable related to features included in the object of analysis, if the correlation between the estimation function and the quantitative value of the first objective variable is 0.7 or less and cannot be uniquely determined, multiple estimation functions are used, the number of such estimation functions is defined as the number of classifications, and the measurement data is classified into multiple objective variables using the multiple estimation functions for estimation. The reason for setting the correlation value to 0.7 or less when classification is unnecessary is that statistically, a correlation value of 0.7 or less is the criterion for whether the estimation is correct. If the accuracy of the estimation is required, the correlation value may be further set to 0.8 or less or 0.9 or less. The explanatory variables acquired to analyze the first objective variable related to features included in the object of analysis are, for example, the detection signals from the detection coil 12 described later, and are called measurement data. An estimation function for the first objective variable is obtained from the measurement data. Note that AI (Artificial Intelligence) analysis refers to the calculation methods and algorithms used when estimating the objective variable from explanatory variables.

Classification is necessary when, for example, analyzing one or more analysis targets using an AI analysis method without classification results in an insufficient correlation between the estimation function and the quantitative value of the first dependent variable, such that the correlation is 0.7 or less. When a single evaluation function is insufficient for estimation, it becomes necessary to use multiple estimation functions that classify the targets.

In this invention, measurement data acquired for analysis is used to classify the data for selecting the estimation function for the analysis.
Figure 1 is a flowchart illustrating an example of the preparation and actual measurement for AI analysis requiring classification. As shown in Figure 1, the preparation involves measuring measurement data, considering classification methods, and preparing an estimation function for the analysis from the measurement data. The estimation function is also called an evaluation formula.
Specifically, this method includes a first step of using multiple estimation functions as classification items and estimating a second dependent variable using measurement data in order to analyze the selection of such classification items as a second dependent variable, and a second step of selecting a classification related to the first estimation function from the results of the first step and estimating the first dependent variable using measurement data for that classification.
For the estimation function of the first dependent variable, classification is performed to obtain a high correlation by creating multiple estimation functions. Preferably, measurement data is used for estimating the first and second dependent variables. Preferably, in order to create the multiple classified estimation functions, when the entire measurement data is analyzed and estimated using the same analytical method, multiple classifications are created by classifying the degree of deviation of the estimated values from the training data. Multiple estimation functions may also be created by classification using the similarity of the measurement data.

According to the above configuration, by utilizing these analytical methods, the estimation function can be easily and accurately defined based on measurement data without having to separately acquire measurement data for defining the estimation function.

Examples of analytical objects for which the estimation function cannot be uniquely determined include materials, material properties, crystalline states, material morphologies, and material structures that can exist in multiple states, such as soil, concrete structures, ceramic structures, and metals containing alloys or metal structures containing alloys that have undergone alteration-related processing. In the case of soil, this includes differences due to regional formation variations as well as anthropogenic changes such as agricultural work; in concrete structures, it includes differences due to the origin of the concrete, moisture content, and type of reinforcing steel; in ceramic structures, it includes differences due to the origin of the raw materials for sintered ceramic structures; and in metals, it includes differences due to changes in the state of alloying elements, crystalline states, transformations, alloying, and grain refinement or enlargement resulting from nitriding, carbonization, plasma, ion, or heat treatment and processing. These are all examples of objects that include items that cannot be definitively classified.

For example, when the object of analysis is soil, it is possible to utilize known data such as soil inventories (non-patent literature) or classify the data using unsupervised learning analyses such as clustering and principal component analysis, but this does not always guarantee a high correlation. The classification should be performed in a way that allows for accurate estimation of the primary target variable. For example, when analyzing the weld strength of spot welds made of the same metal, if the metal material is a mixture of iron and aluminum, classifying the measurement data into iron and aluminum, and then estimating the weld strength of the spot welds between irons using the primary target variable, makes correlation easier, and it becomes possible to estimate a correlation of 0.9 or higher.

For example, if we perform an AI analysis on the entire training data used to estimate the first target variable, the estimation will not be accurate, resulting in an estimated function with a low correlation function. The results obtained here are then classified based on their degree of deviation from the estimated function. For example, using the standard deviation σ, classification can be done using values of 1x and 3x the standard deviation σ, allowing for classifications such as: a group with results more than -3x (below -3σ), a group between -3x and -1x (-3σ to -σ), a group between -1x and +1x (-σ to +σ), a group between +1x and +3x (+σ to +3σ), and a group more than +3x (+3σ). However, care must be taken to ensure that the number of measurement data points for each group is not too small compared to the training data. Furthermore, based on the results of analysis using these group classification methods, the 1x and 3x threshold values may be reviewed and adjusted to ensure appropriate correlation between the data in each group.

The above example is just one instance; while clear classification is possible in learning, in actual measurements, there are cases where classification is not easily determined. In such cases, using known classifications can lead to highly accurate analysis.

The primary objective of classification is to prepare multiple estimation functions for the first dependent variable and select the appropriate one from the same data. Therefore, classification only needs to enable grouping that allows for learning of classification methods, each producing multiple estimation formulas that yield high correlations. Furthermore, it is desirable that the estimation of the second dependent variable—which category each group belongs to—can be analyzed using the same measurement data.

(Analytical device for implementing AI analysis methods)
Figure 2 shows the configuration of one embodiment of an analytical apparatus for implementing the AI analysis method according to the present invention. As shown in Figure 2, the analytical apparatus 1 consists of a sensor 10, a measurement unit 20, and a data processing unit 30.

The sensor 10 comprises an excitation coil 11 and a detection coil 12, and a magnetic path forming section 13. The sensor 10 is housed in a sensor holder 14, for example, made of metal, to block external magnetic fields. The sample 15 to be measured is placed in the measurement section of the sensor holder 14. The sample 15 may be housed in a container and placed in the measurement section of the sensor holder 14. Alternatively, measurement is performed by directly placing the sample 15 in the measurement section. The sensor 10 is supported, for example, within the sensor holder 14 by a non-magnetic gap filler (not shown).

The magnetic path forming section 13 consists of, for example, a bottom section 13a, a cylindrical section 13b, and a shaft section 13c, with the bottom section 13a supporting the cylindrical section 13b and the shaft section 13c. An excitation coil 11 and a detection coil 12 are mounted on the shaft section 13c. Since the sensor 10 has the detection coil 12 and the sample 15 positioned in the magnetic path formed by the excitation coil 11 and the magnetic path forming section 13, the signal detected by the detection coil 12 can be influenced according to the permeability of the sample 15 to be analyzed. The sensor 10 described above is just one example configuration; other configurations with similar functions may also exist. For example, the excitation coil 11 and the detection coil 12 may be composed of a single coil.

The measurement unit 20 comprises an oscillation unit 21, a signal processing unit 22, and a control unit 23. The oscillation unit 21 repeatedly generates a signal of a certain frequency and gradually increases or decreases the frequency of that signal. The signal oscillated from the oscillation unit 21 is split into an excitation signal and a reference signal. The excitation signal is transmitted to the excitation coil 11 and output to the signal processing unit 22 as a reference signal. The signal processing unit 22 calculates the temporal change of the detection signal from the detection coil 12 relative to the excitation signal using the reference signal from the oscillation unit 21. At this time, the signal processing unit 22 has a Fourier transform function and converts the time-axis signal into a frequency-axis signal. The signal processing unit 22 also digitizes the detection signal from the sensor 10 and outputs it to the data processing unit 30. The control unit 23 inputs and outputs data and various control signals to and from the data processing unit 30, and controls the oscillation unit 21 and the signal processing unit 22.

In this case, the data processing unit 30 estimates the target variable for analysis from the digital data of the detection signal processed by the signal processing unit 22. In the following explanation, the target of analysis will be soil, the target variable will be soil fertility traits, and the estimation of CEC will be used as an example of soil fertility traits. The data processing unit 30 is composed of a computer comprising an input/output interface unit 31 that interfaces with the control unit 23, a storage device 32 equipped with a main memory and an auxiliary storage device, an arithmetic unit that performs calculations such as arithmetic operations, and a control device that controls the storage device and the arithmetic unit. The data processing program is stored in the auxiliary storage device, and when the data processing program is deployed to the arithmetic unit and executed, the data processing unit 30 functionally provides the storage unit 33 and estimation unit 34 shown in Figure 2.

The memory unit 33 stores data relating to the correlation between the quantitative value of the target variable and the estimated value of the target variable obtained from the processed detection signal measured using the sensor 10 and the measurement unit 20, for the object of analysis. The estimation unit 34 applies an alternating magnetic field to the object of analysis using the sensor 10 and estimates the CEC of the object of analysis using the data stored in the memory unit 33, based on the detection signal processed by the measurement unit 20.

The estimation unit 34 can analyze the target variable according to the following analytical method, using the quantitative values of the target variable that have been determined in advance. Figure 3 is an explanatory diagram illustrating an example of the analytical method of the present invention, using soil CEC analysis as an example.

As shown in Figure 3, the analysis is performed as follows.
(Step ST1: Preparation of learning samples)
In step ST1, a learning sample 41 is prepared that corresponds to the final sample 47 to be analyzed. The learning sample 41 may be part or all of the sample 47 to be analyzed. One soil or multiple soils with different compositions are prepared. If necessary, the soil to be analyzed may be air-dried and then crushed using a mortar and pestle, and the resulting air-dried crushed soil may be used as the soil sample.

In step ST1, the learning sample 41 is an object measured by the sensor 10, and examples include soil, metal members joined by spot welding, and metals such as reinforcing bars in concrete. If the learning sample 41 is soil, the target variable may be soil fertility characteristics such as CEC (cation exchange capacity).

Furthermore, the learning sample 41 can be used not only to estimate CEC but also to estimate other soil fertility traits. Here, other soil fertility traits that can be estimated include AEC, total carbon (T-C), total nitrogen (T-N), available phosphate (Av-P), total phosphorus (P), iron (Fe), aluminum (Al), and elements important for plant nutrition such as K (potassium), Ca (calcium), and Mg (magnesium). These other soil fertility traits can also be estimated using the analytical method described in Example 4 of Patent Document 2 (Patent Document 2, paragraphs 0083-0086 and 0087-0094). However, if the soil falls outside a specific soil category, or if soil classification is required, the classification estimation method of this invention must be added.

(Step ST2: Training data and Step ST3: Comparison data)
In step ST2, the learning sample 41 to be analyzed is measured with the sensor 10 to obtain learning data 42, i.e., learning measurement data. At this time, the sensor 10 may be placed in close proximity to the soil that will be the learning sample 41 for measurement, or the sensor 10 may be placed in close proximity to the air-dried pulverized soil described above for measurement. Then, in step ST3, for each soil sample 41, the quantitative value of CEC, which will be comparative data 43, is determined using, for example, the known Schorenberger method.
Thus, in step ST2, training data 42 and comparison data 43 necessary for AI learning and estimation are obtained in order to obtain the estimation function 45 for obtaining the first target variable in step ST4. The training data 42 and comparison data 43 are linked to characteristic information such as field location and academic soil classification results related to soil classification and stored in the database of the storage unit 33.

Here, the actual classification of soil depends on the soil condition at the time of measurement, making its evaluation difficult. As shown in Non-Patent Document 1, it is possible to use the results of several academically established classifications, but this has the drawback of not covering all fields and land, and not necessarily matching the soil condition at the time of measurement. In particular, soil components, structure, and composition such as CEC can be used as a guideline in agricultural land because the soil condition changes during farming, but they may be insufficient for classification in AI analysis estimation. Therefore, it is necessary to understand the classification of each soil type from its current state, and AI analysis using measurement data is performed.

(Steps ST4 and ST6 (AI pre-training))
An AI analysis is performed using the comparison data 43, which consists of quantitative values of soil samples obtained in steps ST2 and ST3 and stored in the memory unit 33, to create a first estimation function 45 for obtaining the first target variable. A second estimation function 46 for obtaining the second classification target variable is created for each item of the classification data 44. Step ST3 is also called pre-training. The results are stored in the memory unit 33.
For the classification data 44, which is the second dependent variable, if the analysis results from the first estimation function 45 do not yield a good agreement, such as a correlation coefficient of 0.7 or less, the data that is particularly difficult to agree with will be separated and classified using the method described later in order to improve the overall results.
The AI analysis in steps ST4 and ST5, which obtains the first and second objective variables, is performed using one of the following analysis methods: unsupervised AI analysis methods such as clustering and principal component analysis, or supervised AI analysis methods such as linear analysis, SVM (Support Vector Machine), PLS regression analysis, and neural networks.
The methods used in step ST4 to obtain the first estimation function 45 and step ST5 to obtain the second estimation function 46 may be the same or different. After obtaining the first estimation function 45 through AI learning in step ST4, the process may proceed to step ST7 to determine whether the verification result of the first estimation function 45 is appropriate or not. If the verification result is not appropriate (NO), the process returns to step ST5 to prepare for classification. If the verification result is appropriate (YES), the process proceeds to the actual evaluation (step ST8) as described later.
Here, as a method for obtaining the second estimation function 46, clustering is a commonly used classification method. However, depending on the classification data 44, clustering may be inappropriate if it contains strong features, and in such cases, deep learning methods such as neural networks, which are supervised AI analysis methods, may be effective.
The classification in Step ST6 may also be based on academic classification or human judgment, without using AI analysis methods.

If a second estimation function 46 and the classifications based thereon are defined, a first estimation function 45 for analysis and evaluation is defined for each classification defined as classification data 44 in step ST5.

According to the above configuration, by utilizing these analytical methods, a classification evaluation function can be easily and accurately created based on the measurement data.

(Actual evaluation from Step ST8)
Next, we move on to the actual evaluation process. An actual sample 47 to be analyzed, whose CEC and soil classification-related characteristics are unknown, is prepared in the same way as the soil sample in step ST1, and evaluation data 48 consisting of complex amplitude and phase within a predetermined frequency range is acquired by the sensor 10. Specifically, the evaluation data 48 is obtained by generating an alternating magnetic flux by applying an alternating current to the excitation coil 11, and detecting the magnetic flux that is disturbed by the state of the analysis target using the detection coil 12. At this time, multiple excitation signals of different frequencies may be applied to the excitation coil 11. The excitation signals may have a frequency band of about 5 to 100 times from the lowest frequency to the highest frequency. Furthermore, the excitation signals may have five or more different frequencies and be one or more greater than the sum of the number of analyses and the number of classifications.

After defining the classification data 44 and the first estimation function 45 for each classification in this manner, as described later, from step ST8, evaluation data 48 is measured for the sample to be analyzed 47, a classification determination is made using the second estimation function 46, the corresponding classification is selected in step ST11, and the analysis result 50 is obtained from the evaluation data 48 using the first estimation function 45 corresponding to the selected classification in step ST14, and the evaluation result is finalized in step ST16.

(Step ST10: Examining the necessity of classification)
In this case, if it is clear that the sample to be analyzed 47 belongs to a specific classification, for example, if the first estimation function 45 can be identified for the evaluation data 48, then classification by calculation of the second estimation function 46 is unnecessary, and the process can proceed from step ST12 to step ST13, and in step ST14, the CEC of the soil that will result in the desired analysis result 50 may be estimated.

(Step ST11: Implementation of classification)
In step ST11, estimation is performed using a second estimation function 46 to select the estimation function to be used from among the first estimation functions 45 created for each classification of the evaluation data 48. Based on the calculation results using the second estimation function 46, it is determined which classification of the classification data 44 the evaluation data 48 corresponds to, and the first estimation function 45 is selected.

(Step ST13: Consider the necessity of analysis)
After the classification in step ST11, you choose whether or not to estimate the soil's CEC. If you only want to perform classification without analysis, for example, to determine which classification the soil belongs to for crop selection in the target soil, you can proceed to step ST16 via step ST15, confirm the classification result 49 obtained in step ST11, and finish.

(Step ST14 Analysis)
In step ST14, when obtaining the analysis result 50 of the sample 47 to be analyzed, that is, when estimating the CEC of the soil, in step ST11, the soil is classified using the evaluation data 48 and a second estimation function 46 for the objective variable for the second classification. Based on the obtained classification result 49, a first estimation function 45 for the first objective variable, created in step ST4, is selected for each classification defined in the classification data 44. The CEC of the soil, which will be the analysis result 50, is estimated using the evaluation data 48, and in step ST16, the soil classification and the estimated CEC value are confirmed and the process is completed.

(Step ST16)
This completes the soil classification and CEC analysis estimation. However, if the results indicate that the estimated values are inappropriate, the process will be restarted from step ST3 for retraining, the classification will be redone in steps ST5 and ST6, and the analysis will be redone in steps ST11 and ST14. If necessary, the method for acquiring training data in step ST2 will also be reviewed.

(Measurement method using electromagnetic measurement)
The following describes in detail the method for acquiring training data 42 and measurement data 48, as well as the classification and analysis estimation of soil, when the target of analysis is soil and measurements are taken using the electromagnetic measurement method shown in Figure 2, with CEC and other quantitative values of soil fertility traits as the objective variables.

In the soil analysis apparatus 1, a learning sample 41, which will become the sample 15, is placed in the measurement section of the sensor holding unit 14. Under the control of the control unit 23, the oscillation unit 21 oscillates signals at each frequency, increasing the frequency stepwise at arbitrary intervals (e.g., several kHz to several hundred kHz) within a specified frequency range (e.g., several kHz to several hundred kHz), and outputs them to the excitation coil 11. The signal detected by the detection coil 12 is processed by the signal processing unit 22, converted into a digital signal, and output to the data processing unit 30. In the estimation unit 34, the processed detection signal output from the signal processing unit 22 is stored in the storage unit 33. This value becomes the learning data 42.

The frequency range of the AC generated by the oscillator 21 can be, for example, 1 Hz to 80 Hz, 50 Hz to 5 kHz, 500 Hz to 2.5 kHz, or approximately 1 kHz to 100 kHz. That is, the frequency band of the AC may be approximately 5 to 100 times the minimum to maximum frequency. Furthermore, the AC frequencies may be set to be five or more distinct frequencies, and to be at least one greater than the sum of the number of analyses and the number of classifications.

A portion of the soil sample, which is the learning sample 41, is taken and analyzed using a known quantitative analysis method such as the Schorenberger method as shown below, and the resulting value is used as comparison data 43.
In addition to CEC, an example of a method for measuring quantitative values of various soil fertility traits is:
CEC, potassium, magnesium, calcium: Schörenberger method, etc. Total carbon (T-C): Combustion method using a C/N coder, etc. Total nitrogen (T-N): Combustion method using a C/N coder, etc. Available phosphorus (Av-P): Truog method, etc. Potassium, magnesium, calcium, iron, manganese, zinc, copper, molybdenum, boron, chlorine, nickel: Elemental analyzers such as X-ray fluorescence analysis are available.
In this case, if a quantitative analysis method is available for the analysis target, a supervised learning method can be used as the AI analysis method. If quantitative analysis cannot be performed on the analysis target, an unsupervised learning AI analysis method can be used without using the comparison data 43. It is also possible to perform supervised learning using qualitative analysis that includes ambiguity or images as the comparison data 43.

In the estimation unit 34, pre-training (step ST4) is performed using AI analysis, with the training data 42 as the explanatory function and the comparison data 43 as the target variable. AI analysis refers to the use of one of the following analysis methods: supervised AI analysis methods such as linear analysis, SVM, PLS regression analysis, and neural networks; or, if comparison data 43 is unavailable, unsupervised AI analysis methods such as clustering analysis and principal component analysis. This analysis is performed by the estimation unit 34 using a program stored in the storage device 32, within the data processing unit 30. When the comparison data 43 consists of quantitative values of various soil fertility traits, including CEC, the first estimation function 45 is obtained through AI analysis.

In electromagnetic measurement methods and other measurement techniques, the measurement results are affected by the material and structure of the object being measured, as well as the ratio and distribution of materials if multiple materials are mixed or combined. Therefore, it is necessary to change the measurement conditions and the estimation function depending on the object being measured. In the case of soil, the first estimation function 45 cannot be uniquely determined due to influences such as soil type and soil improvement by agricultural work. Therefore, it is necessary to create a first estimation function 45 for each group of objects being measured by dividing the learning process according to a classification. This grouped classification becomes the classification data 44.

In this case, as classification data 44, there is the "Soil Classification" (see Non-Patent Document 1) from the National Agriculture and Food Research Organization's Japan Soil Inventory, which is an academic classification and can be used as a reference. However, since farms managing individual fields may have implemented soil improvement measures, it is difficult to classify appropriately using only the above classification. Furthermore, soil classification requires tasks such as deep soil sampling and analysis of its cross-sectional condition, making soil analysis not easy.

Therefore, to obtain a second estimation function 46 with soil classification as the second objective variable, AI learning (step ST6) is performed on soil classification estimation using the aforementioned training data 42. Both analysis and classification can be performed using the same training data 42.

For soil classification, referencing the aforementioned academic soil classifications, a first estimation function 45 is created for each classification defined in the classification data 44, and the analysis is performed. If the results are insufficient, particularly for mismatched data, reclassification is performed, referring to the value of the standard deviation σ, and the estimations created for each classification are selected to yield desirable results. The classification here may deviate from academic classifications, but it is used as a classification that reflects the actual situation. Furthermore, if academic classifications are not available, unsupervised learning methods such as clustering analysis and principal component analysis can be used. In this case, as will be discussed later, it is important to note that appropriate classification may not be possible due to the influence of other factors.
The determined classification is used as the classification data 44, which is the target variable, and the training data 42 is used as the explanatory variable to perform AI learning. The resulting estimation function becomes the second estimation function 46 for the second target variable, which is the classification.

Next, in the actual measurement, the soil sample 47, which is the sample to be analyzed, is prepared in step ST9. Specifically, similar to step ST2, the sample 15 is placed on the measurement unit of the sensor holder 14, or the sensor 10 is brought into direct contact with the soil sample 15 in a field or similar location. Under the same conditions as during learning, and under the control of the control unit 23, the oscillation unit 21 oscillates signals at each frequency, gradually increasing the frequency within a specified frequency range (e.g., several kHz to several hundred kHz) and at interval frequencies (e.g., several kHz), and outputs these signals to the excitation coil 11. For each frequency signal, the signal detected by the detection coil 12 is processed by the signal processing unit 22, converted into a digital signal, and output to the data processing unit 30. In the estimation unit 34, the processed detection signal output from the signal processing unit 22 is stored in the storage unit 33. This value becomes the evaluation data 48.

The obtained evaluation data 48 is used to select a classification using the second estimation function 46 obtained during pre-training, and the analysis is estimated using the first estimation function 45 based on the classification result 49. This results in obtaining the classification result 49 and analysis result 50 of the analyte sample 47 in step ST16.

Furthermore, as the measurement data is increased, the data used for AI training can be increased through sampling and re-evaluation, thereby improving the accuracy of analysis and evaluation.

This section describes an example of soil analysis. Here, for verification purposes, the analysis target sample 47 is used as the training sample 41, and the training data 42, which includes comparison data 43, is used as the evaluation data 48. The correlation obtained from recalculations using the first estimation function 45 and the second estimation function 46, acquired after training, is examined.

In this analysis example, soil samples were collected from five fields: A, G, S1, S2, N1, and N2. According to the academic soil classification in Patent Document 1, the soil samples used fall into four categories: gravelly ordinary brown lowland soil (A), humus-rich sublayer lowland allophanic black soil (G), typical ordinary humid black soil (S1, S2), and gravelly ordinary allophanic black soil (N1, N2). The fields indicated in parentheses correspond to each of these categories.

Each soil sample (training sample 41) is analyzed and evaluated using the soil analyzer 1. First, the soil sample is irradiated with a magnetic field, for example, varying the frequency from 10 kHz to 200 kHz in 1.5 kHz increments. The transmitted magnetic field is then detected by the detection coil 12 of the soil analyzer 1 and measured by the signal processing unit 22. The real and imaginary parts of the detection signal are used as training data 42. The same training sample 41 is also subjected to CEC analysis using the Schorenberger method, and the results are used as comparison data 43. Using this training data 42 and comparison data 43, a neural network is implemented as an AI analysis method.

(Comparative Example 1)
For reference, we performed CEC estimation using a neural network from all measurement data without classification.
Figure 4 shows the relationship between the estimated and quantitative values of CEC for Comparative Example 1. The horizontal axis of Figure 4 represents the estimated value (meq/100g) obtained from the first estimation function 45 of CEC, and the vertical axis of Figure 4 represents the quantitative value (meq/100g) of CEC for the comparison data 44. The squares (■), black circles (●), and triangles (△) in the figure represent the classification categories AZ, SN, and S, respectively. As shown in Figure 4, the correlation between the estimated and quantitative values of CEC for 455 subjects was 0.71. At this time, it was 0.66 for the lower 95% and 0.75 for the upper 95%, and the p-value was 0.0001 or less. Although there is a certain degree of correlation, improvement will be made by classification.

In the results, the estimated results were examined based on their deviation from the comparison data 44, and selections were made at the field level. For this data, it was appropriate to classify the fields into three categories: Fields A and G, and Fields S1, S2, N1, and Field N2.

(Example 1)
In Example 1, AI learning was performed using a neural network with CEC estimation as the first objective variable for each of the three classifications, and a first estimation function 45 corresponding to each classification was obtained. Figure 5 shows the relationship between the estimated value and the quantitative value of CEC in Example 1. The horizontal axis of Figure 5 is the estimated value (meq/100g) obtained from the first estimation function 45 of CEC, and the vertical axis is the quantitative value of CEC (meq/100g) which is the comparison data 44. The square marks (■), black circles (●), and triangle marks (△) in the figure represent the classification categories AZ, SN, and S, respectively. As shown in Figure 5, the correlation between the estimated value and the quantitative value of CEC was 0.92 for 455 subjects. At this time, it was 0.90 for the lower 95%, 0.93 for the upper 95%, and the p-value was 0.0001 or less. In Example 1, the correlation improved from 0.71 in Comparative Example 1 to 0.92 by applying the classification.

(Example 2)
In Example 2, AI learning is performed using a neural network with the three groups as the objective variables and the training data 42 as the explanatory variables for these three classifications, and a second estimation function 46, which is a second objective variable for classification, is created. Using this second estimation function 46, evaluation estimation of the training data 42 is performed in place of the actual evaluation data 48, and the training data 42 is reclassified. This reclassification reclassifies the data into closer groups. Based on this reclassification result, AI learning is performed for each classification to obtain a first estimation function 45 for the first CEC analysis. CEC estimation is performed using the reclassification result and the respective first estimation functions 45. Figure 6 shows the relationship between the estimated value and the quantitative value of CEC in Example 2. The horizontal axis of Figure 6 is the estimated value of CEC (meq/100g), and the vertical axis is the quantitative value of CEC (meq/100g). The squares (■), black circles (●), and triangles (△) in the figure represent the classification categories AZ, SN, and S, respectively. As shown in Figure 6, the correlation between the estimated and quantitative values of CEC for 455 subjects was 0.98. At this time, the correlation was 0.97 for the lower 95% and 0.98 for the upper 95%, with a p-value of 0.0001 or less.

Thus, the correlation in Comparative Example 1, where analysis and evaluation were performed without classification, was 0.72. In Example 1, where classification was performed, the correlation improved to 0.92, and in Example 2, where reclassification using AI-based classification estimation was performed, the correlation improved to 0.98.

In other words, by using AI-based reclassification estimation, we were able to perform CEC estimation with an even higher correlation.

(Comparative Example 2)
In AI learning and analysis, clustering is sometimes used for analysis, and this was compared as Comparative Example 2. Clustering analysis groups similar data together. Common methods include hierarchical clustering and k-means.
Figure 7 shows an example where the results of grouping using the k-means method were used as the second estimation function 46 for subsequent analysis. The horizontal axis shows the distribution of data in an exaggerated state due to the real and imaginary parts, which are parameters of the measured data. The k-means method allows for grouping similar data based on the characteristics of the data (1, 2, and 3 in Figure 7). Even when the grouping shown in Figure 7 is used as the classification result, and the first estimation function 45 from step ST3 onwards is created for each classification using the same training data 42, and the classification analysis from step ST8 onwards is performed using the same training data 42 as evaluation data 48, the estimated values of the CEC in the analysis result 50 were ^R² values of 0.36, 0.70, and 0.48 for the verification values of groups 1, 2, and 3, respectively. These values indicate that the accuracy of the estimation is significantly worse compared to the correlation coefficients of Examples 1 and 2. This differs from the results of the classification based on academic soil classification used in Examples 1 and 2, and is therefore due to some feature quantity included in the measurement content, which does not match the purpose of this study. This is because the classification results differed from those obtained based on the actual objects being analyzed, and as shown in Comparative Example 2, this is an example where caution is needed in unsupervised learning.

The present invention can be implemented in various forms without departing from its spirit. For example, in the embodiments described above, the acquisition of the second estimation function 46 for defining the classification and the acquisition of the first estimation function 45 for inferring the CEC are both performed using neural analysis. However, it is clear that the invention is not limited to this, and other analytical methods, such as linear analysis, SVM (Support Vector Machine), PLS regression analysis, supervised AI analysis methods, and unsupervised AI analysis methods, may also be used for the definitions.

Furthermore, while the above-described embodiment explained the estimation of CEC (Centralized Electricity Consumption) among soil fertility characteristics, it is clear that the method is not limited to this and can also be applied to the estimation of other soil fertility characteristics, or to AI analysis of welding strength in spot welds, concrete structures with reinforcing bars, etc.

Furthermore, it may be possible to learn the classification system of soil taxonomy as a classification criterion for the soil to be analyzed, and then perform classification using the re-estimation results based on that learning.

In the embodiment described above, the measurement data was acquired solely by electromagnetic measurement using a detection coil 12 utilizing multiple frequency bands. However, it may also be acquired by acoustic measurement such as ultrasound, or optical measurement using light.

1. Soil analysis device 10. Sensor 11. Excitation coil 12. Detection coil 13. Magnetic path forming unit 14. Sensor holding unit 15. Sample 20. Measurement unit 21. Oscillation unit 22. Signal processing unit 23. Control unit 30. Data processing unit 31. Input/output interface unit 32. Memory device 33. Storage unit 34. Estimation unit 41. Training sample 42. Training data 43. Comparison data 44. Classification data 45. First estimation function (analysis)
46. Second estimation function (classification)
47. Sample to be analyzed 48. Data for evaluation 49. Classification results 50. Analysis results

Claims (15)

In an AI analysis method that estimates a first objective variable related to features included in the object of analysis by obtaining measurement data from the object of analysis and obtaining a first estimation function using the measurement data as explanatory variables,
The object of analysis is any of the following: soil, concrete structure, ceramic structure, metal including alloys, and metal structure including alloys that has undergone processing involving alteration, wherein the material, material properties, crystalline state, material morphology, and material structure exhibit multiple states.
If the correlation between the estimated value of the first estimation function and the quantitative value of the first dependent variable is 0.7 or less and cannot be uniquely determined ,
The system is configured to classify the aforementioned measurement data into multiple target variables using a second estimation function, and to obtain multiple first estimation functions for each classification.
In the first step,
In order to select the first estimation function to be used from among the multiple first estimation functions created for each classification, the second objective variable is estimated by classifying the second estimation function using the measurement data, and the first estimation function to be used is selected.
In the second step,
An AI analysis method that uses measurement data related to the first estimation function selected in the first step as explanatory variables, and estimates the first target variable using the first estimation function . The AI analysis method according to claim 1, wherein the evaluation data that serves as the explanatory variables is used as training data, the first objective variable or the second objective variable is obtained as quantitative values of the training data and used as comparison data, and AI analysis is performed . Using the aforementioned evaluation data, classification is performed using a second estimation function, and based on the obtained classification results, a first estimation function for the first dependent variable defined for each classification in the classification data is selected.
Using the evaluation data, the first target variable is estimated using the first estimation function.
The AI analysis method according to claim 2 , which determines the classification of the object to be analyzed and the estimated value of the first objective variable . The AI analysis method according to claim 1, wherein, in order to create a plurality of classified first estimation functions, the entire measurement data is analyzed and estimated using the same analysis method, and a plurality of classifications are created by classifying the estimated values according to the degree of deviation from the training data. The AI analysis method according to claim 1, wherein a plurality of first estimation functions are created by classification using the similarity of the measurement data. The AI analysis method according to claim 1 , wherein a second estimation function corresponding to the selection of the above classification is obtained by any of the following analysis methods: linear analysis, PLS regression analysis, SVM (Support Vector Machine), neural network, or supervised AI analysis method. The AI analysis method according to claim 1, wherein a plurality of first estimation functions corresponding to the objective variable are obtained by any of the following analysis methods: clustering, principal component analysis, unsupervised AI analysis method, linear analysis, PLS regression analysis, SVM (Support Vector Machine), neural network, or supervised AI analysis method. The AI analysis method according to any one of claims 1 to 7, wherein the measurement data is acquired by electromagnetic measurement, acoustic measurement, or optical measurement using multiple frequency bands. An AI analysis method according to any one of claims 1 to 8, wherein the first objective variable is to obtain estimated values of soil fertility traits, which include soil agricultural indicators such as cation exchange capacity (CEC), base saturation, silicic acid, iron oxide, and nutrients such as potassium, magnesium, calcium, iron, manganese, zinc, copper, molybdenum, boron, chlorine, nickel, nitrogen, carbon, and phosphorus . The AI analysis method according to any one of claims 1 to 9 , wherein the measurement data is used to generate an alternating magnetic flux by applying a plurality of excitation signals of different frequencies to an excitation coil, and the magnetic flux that is disturbed by the state of the object to be analyzed is detected using a detection coil. The AI analysis method according to claim 10 , wherein the excitation signal has a frequency bandwidth of approximately 5 to 100 times from the lowest frequency to the highest frequency. The AI analysis method according to claim 10 , wherein the excitation signals have five or more distinct frequencies and are one or more greater than the sum of the number of analyses to be performed and the number of classifications. An AI analysis method according to any one of claims 1 to 12, wherein, when the total number of each of the one or more items to be analyzed is evaluated using the same evaluation function , the standard deviation is used to classify the items according to the degree of deviation from the correlation, such as -3σ or less, -3σ to -1σ, ±1σ, 1σ to 3σ, and 3σ or more. An AI analysis device that analyzes one of the following materials, including soil, concrete structures, ceramic structures, metals containing alloys, and metal structures containing alloys that have undergone processing involving alteration, in order to analyze a first objective variable related to the features contained in the analysis target, by acquiring measurement data from the analysis target and obtaining a first estimation function of the first objective variable using the measurement data as explanatory variables, thereby estimating the first objective variable ,
A sensor having a coil,
A measurement unit that generates excitation signals for each frequency to be input to the coil in order to apply an alternating magnetic field to the object to be analyzed, and processes the detection signal output from the coil using the generated excitation signals,
A storage unit that stores data relating to the correlation between the quantitative value of the target variable of each analysis target and the sensor and the measurement unit, for multiple analysis targets,
An estimation unit estimates the target variable of the analysis target using the data stored in the storage unit based on the detection signal processed by the measurement unit,
It is equipped with,
The estimation unit obtains the complex amplitude ratio of the detection signal to the excitation signal as measurement data from the detection signal, defines multiple classifications from the quantitative values of multiple different target variables stored in the storage unit and the corresponding complex amplitude ratios of the detection signals as comparison data, and defines a first estimation function for each classification.
An AI analysis device in which the estimation unit selects a classification corresponding to each measurement data using a second estimation function, and estimates the first target variable of the analysis target from the measurement data using a first estimation function corresponding to the selected classification. A program that causes a computer to function as the AI analysis device described in claim 14 .