Information processing device, information processing method, computer program, and learning model

Abstract

The properties of diesel fuel are estimated using the analysis results of diesel fuel, which are obtained by an analytical device that measures the content of substances in the sample. [Solution] The information processing device comprises a memory for storing a learning model, which is a computer program for estimating the properties of diesel fuel, and a processor for executing the learning model. The processor performs an operation that includes the steps of: pre-processing the analysis results from an analyzer that measures the content of substances in a diesel fuel sample to generate input data for input to the learning model; and executing processing by the learning model based on the generated input data.

Description

[Technical Field]

[0001] The present disclosure relates to an information processing apparatus, an information processing method, a computer program, and a learning model . [Background Art]

[0002] Patent Document 1 discloses a method in which a reference sample is introduced into each of a gas chromatograph and a mass spectrometer, a specific process is performed on the obtained analysis result to obtain a series of coefficients, crude oil or a boiling point fraction thereof is introduced into each of the gas chromatograph and the mass spectrometer, the same specific process as described above is performed on the obtained analysis result, and the resulting data is multiplied by the series of coefficients to obtain a predicted value of a property or properties of the crude oil or the boiling point fraction thereof. [Prior Art Literature] [Patent Literature]

[0003] [Patent Document 1] Japanese National Publication of International Patent Application No. 11-508363 [Summary of the Invention] [Problem to be Solved by the Invention]

[0004] However, in the method disclosed in Patent Document 1, it is necessary to analyze the reference sample and the crude oil or the boiling point fraction thereof not only by gas chromatography but also by a mass spectrometer. [Means for Solving the Problem]

[0005] An information processing device according to one aspect of the present disclosure is an information processing device comprising: a memory for storing a learning model which is a computer program for estimating the properties of diesel fuel; and a processor for executing the learning model, wherein the processor performs an operation that includes the steps of: pre-processing the analysis results from an analyzer for measuring the content of substances in a diesel fuel sample to generate input data for input to the learning model; and executing processing by the learning model based on the generated input data. [Effects of the Invention]

[0006] According to this disclosure, the properties of diesel fuel can be estimated using the analysis results of diesel fuel performed by an analytical instrument that measures the content of substances in a sample, such as a gas chromatograph. [Brief explanation of the drawing]

[0007] [Figure 1] Figure 1 shows an example of the configuration of a machine learning system according to an embodiment. [Figure 2] Figure 2 is a block diagram showing an example of the hardware configuration of a learning device according to this embodiment. [Figure 3] Figure 3 is a graph showing an example of analysis results obtained by gas chromatography. [Figure 4] Figure 4 shows an example of the properties of the diesel fuel to be estimated according to the embodiment. [Figure 5] Figure 5 illustrates an example of data integration and splitting. [Figure 6] Figure 6 is a graph showing an example of the relationship between the number of principal components and the cumulative contribution in the analysis results data. [Figure 7] Figure 7 shows an example of the number of principal components in the analysis results data for each property of diesel fuel. [Figure 8] Figure 8 is a diagram illustrating an example of how training data is generated. [Figure 9] Figure 9 is a diagram illustrating an example of the learning function of a learning device according to an embodiment. [Figure 10]Figure 10 shows an example of the configuration of a regression model, which is a neural network. [Figure 11] Figure 11 is a diagram illustrating an example of input data generation. [Figure 12] Figure 12 is a flowchart showing an example of the operation of the learning device during the learning phase according to the embodiment. [Figure 13] Figure 13 is a flowchart showing an example of the operation of the learning device in the estimation phase according to the embodiment. [Modes for carrying out the invention]

[0008] <Summary of the embodiments of this disclosure> The embodiments of this disclosure are outlined below.

[0009] (1) The information processing device according to this embodiment includes a memory for storing a learning model, which is a computer program for estimating the properties of diesel fuel, and a processor for executing the learning model, wherein the processor performs an operation that includes the steps of: pre-processing the analysis results from an analytical device that measures the content of substances in a diesel fuel sample to generate input data for input to the learning model; and executing processing by the learning model based on the generated input data. This makes it possible to estimate the properties of diesel fuel using the analysis results of diesel fuel from an analytical device that measures the content of substances in a sample.

[0010] (2) In (1) above, the preprocessing may include a dimensionality reduction process to reduce the number of dimensions of the analysis results. This reduces the amount of input data and reduces the processing load on the learning model.

[0011] (3) In the above (2), the analysis result is data indicating a relationship between a first element and a second element that are different from each other, and the dimension reduction process may be a process of reducing the number of dimensions of the analysis result while retaining characteristics of the relationship between the first element and the second element. This enables reduction of the data amount of input data while retaining characteristics of the multidimensional input data.

[0012] (4) In the above (3), the dimension reduction process may include a process of reducing the number of data points of the first element by principal component analysis. Principal component analysis enables reduction of the data amount of input data while retaining characteristics of the input data.

[0013] (5) In the above (3) or (4), the dimension reduction process may include a process of lowering the resolution of the first element in the analysis result. This enables reduction of the data amount of input data through a simple process.

[0014] (6) In any one of the above (1) to (5), the input data is teacher data for supervised learning, and is data in which property data indicating properties of light oil and analysis result data indicating the analysis result are combined, and the processing by the learning model may be processing for performing machine learning with the learning model, which is a model before learning, to generate a trained model. This enables generation of a trained model for estimating properties of light oil through supervised learning.

[0015] (7) In the above (6), the step of generating the input data includes generating first teacher data corresponding to a first property of the light oil and second teacher data corresponding to a second property of the light oil, and the process of generating the trained model may be a process of generating a first trained model for estimating the first property using the first teacher data, and generating a second trained model for estimating the second property using the second teacher data. This enables generation of a plurality of trained models respectively corresponding to a plurality of properties of light oil.

[0016] (8) In (6) or (7) above, the operation further includes the step of determining whether the error in the property values ​​of the diesel fuel as output data obtained when evaluation data showing the analysis results of the diesel fuel by the analyzer is input to the generated trained model is smaller than a specified value, and if the error in the property values ​​of the diesel fuel as output data exceeds a specified value, the processor may discard the trained model and perform the steps of generating the input data and performing the processing by the trained model again. This allows the estimation performance of the generated trained model to be evaluated and retraining to be performed if the estimation performance of the trained model is low.

[0017] (9) In any one of (1) to (5) above, the input data is data to be input to the trained model, which is a trained model, and includes analysis result data showing the analysis result, and the processing by the trained model may be a process in which the trained model processes the input data to estimate the properties of the diesel fuel. This makes it possible to estimate the properties of the diesel fuel from the analysis results of the diesel fuel by the analyzer using the trained model.

[0018] (10) In (9) above, the properties of the diesel fuel include a first property and a second property that are different from each other, and the trained model may include a first trained model for estimating the first property and a second trained model for estimating the second property. This allows for the use of different trained models depending on the properties of the diesel fuel to be estimated, making it possible to prepare trained models specialized for each property and improving the estimation accuracy for each property.

[0019] (11) The information processing method according to this embodiment is an information processing method using a learning model, which is a computer program for estimating the properties of diesel fuel, and includes the steps of: pre-processing the analysis results of an analytical device that measures the content of substances in a diesel fuel sample to generate input data for input to the learning model; and executing processing by the learning model based on the generated input data. This makes it possible to estimate the properties of diesel fuel using the analysis results of diesel fuel from an analytical device that measures the content of substances in a sample.

[0020] (12) The computer program according to this embodiment is a computer program for generating input data to be input to a learning model, which is a computer program for estimating the properties of diesel fuel, and causes the computer to perform the step of preprocessing the analysis results from an analytical device that measures the content of substances in a diesel fuel sample, and generating input data to be input to the learning model. As a result, input data for a learning model for estimating the properties of diesel fuel can be generated using the analysis results of diesel fuel from an analytical device that measures the content of substances in a sample.

[0021] (13) The learning model according to this embodiment is a computer program for estimating the properties of diesel fuel, and causes the computer to perform the step of estimating the properties of diesel fuel based on input data generated by preprocessing the analysis results of an analytical device that measures the content of substances in a diesel fuel sample. This makes it possible to estimate the properties of diesel fuel using the analysis results of diesel fuel from an analytical device that measures the content of substances in a sample.

[0022] This disclosure can be realized not only as an information processing device having the characteristic configuration described above, an information processing method using the characteristic processing as a step, a computer program for causing the information processing device to execute the characteristic processing, and a learning model for causing the information processing device to execute the characteristic processing, but also as an information processing system including the information processing device, or as part or all of the information processing device being a semiconductor integrated circuit.

[0023] <Details of the embodiments of this disclosure> The embodiments of this disclosure will be described in detail below with reference to the drawings. At least some of the embodiments described below may be combined in any way.

[0024] [1. Machine Learning Systems] Figure 1 shows an example of the configuration of a machine learning system according to an embodiment. The machine learning system 1 according to the embodiment includes a learning device 10 and a terminal device 20. The learning device 10 is an example of an "information processing device".

[0025] The learning device 10 and the terminal device 20 are connected, for example, by a communication line 30, and can communicate data with each other. Figure 1 shows an example where the learning device 10 and the terminal device 20 are connected one-to-one, but the learning device 10 may be connected to multiple terminal devices 20 via a network.

[0026] The learning device 10 performs supervised learning, which is a type of machine learning. The terminal device 20 includes, for example, an input device and a display device. The user can, for example, operate the terminal device 20 to instruct the start of learning.

[0027] In machine learning, a learning model is generated to estimate the properties of diesel fuel from the analysis results of diesel fuel performed by an analytical instrument. Hereafter, the learning model before machine learning will also be called the "pre-training model," and the learning model after machine learning will also be called the "trained model."

[0028] In this embodiment, the analysis results of diesel fuel performed by an analytical instrument are used as input data to the trained model. Here, the diesel fuel subjected to analysis by the analytical instrument is sometimes referred to as the "diesel fuel sample." The output data of the trained model is an estimated value of the properties of the diesel fuel. That is, in the trained model, the analysis results of the diesel fuel are the explanatory variables, and the properties of the diesel fuel are the dependent variable. The analytical instrument used is a device that measures the content of substances in a sample, such as a gas chromatograph. In this embodiment, an example in which analysis results data from a gas chromatograph are used as input data to the trained model will be described.

[0029] The user can input analysis results data from a gas chromatograph into the terminal device 20. The terminal device 20 transmits the analysis results data to the learning device 10. The learning device 10 uses the analysis results data to perform machine learning or inputs the analysis results data into a trained model to estimate the properties of diesel fuel. For example, when the learning device 10 finishes machine learning, it transmits the machine learning results to the terminal device 20, and the terminal device 20 can display the learning results on the display device. Furthermore, when the learning device 10 estimates the properties of diesel fuel using the trained model, it transmits the estimation results to the terminal device 20, and the terminal device 20 can display the estimation results on the display device.

[0030] For example, the learning device 10 is comprised of a computer. In a specific example, the learning device 10 is a computer with large-scale and high-speed computing capabilities, such as a supercomputer. The learning device 10 may be a computer dedicated to machine learning, or it may be a general-purpose computer. The learning device 10 may be a server, and the terminal device 20 may be a client.

[0031] For example, the learning device 10 may have the functions of a terminal device 20. In this case, the machine learning system 1 is composed of one learning device 10.

[0032] [2. Hardware configuration of the learning device] Figure 2 is a block diagram showing an example of the hardware configuration of a learning device according to this embodiment.

[0033] The learning device 10 includes a processor 101, a non-volatile memory 102, a volatile memory 103, and an interface (hereinafter also referred to as "IF") 104. The processor 101, the non-volatile memory 102, the volatile memory 103, and the IF 104 are each connected to one another by a bus (data bus). The processor 101, the non-volatile memory 102, the volatile memory 103, and the IF 104 can each transmit data to one another via the bus.

[0034] The processor 101 is, for example, a CPU (Central Processing Unit). However, the processor 101 is not limited to a CPU. The processor 101 may be, for example, a GPU (Graphics Processing Unit) or an NPU (Neural Processing Unit). In a specific example, the processor 101 is a multi-core processor. The processor 101 may be a single-core processor. The processor 101 may include multiple processors or cores and be capable of performing parallel processing. The processor 101 is configured to execute computer programs. The processor 101 may include, for example, an ASIC (Application Specific Integrated Circuit) as part, or programmable hardware such as an FPGA (Field Programmable Gate Array) or a CPLD (Complex Programmable Logic Device) as part.

[0035] The volatile memory 103 is a semiconductor memory such as SRAM (Static Random Access Memory) or DRAM (Dynamic Random Access Memory). The non-volatile memory 102 is a rewritable storage device such as flash memory or a hard disk. The non-volatile memory 102 stores the training data generation program 110, the learning program 120, and the input data generation program 210.

[0036] The training data generation program 110 is a computer program for generating training data. The learning program 120 is a computer program for performing machine learning using the training data. The learning function of the learning device 10 is realized when the learning program 120 is executed by the processor 101. The processor 101 generates trained models 230_1, 230_2, 230_3, ... by executing machine learning using the learning program 120.

[0037] The generated trained models 230_1, 230_2, 230_3, ... are stored in the non-volatile memory 102. Each of the trained models 230_1, 230_2, 230_3, ... is a computer program for estimating the properties of diesel fuel based on input data. The input data generation program 210 is a computer program for generating input data for the trained models 230_1, 230_2, 230_3, ...

[0038] IF104 is a communication interface for communication with the terminal device 20. For example, IF104 is an Ethernet interface ("Ethernet" is a registered trademark).

[0039] [3. Functions of the learning device] The learning device 10 has a training data creation function, a learning function, an input data creation function, and an inference function. The functions of the learning device 10 are described below.

[0040] [3-1. Training Data Creation Function] The training data creation function is realized by the processor 101 executing the training data generation program 110. The training data creation function is a function that preprocesses the analysis results of diesel fuel by gas chromatography and creates (generates) training data. The training data is input data to be input into the pre-trained model 130.

[0041] The learning device 10 learns the relationship between the properties of diesel fuel and the results of diesel fuel analysis by gas chromatography. The training data used for supervised learning is a dataset that includes the results of diesel fuel analysis by gas chromatography, which are the input data for the trained model, and the diesel fuel property values, which are the correct output values ​​of the trained model.

[0042] [3-1-1. Analysis Results Data] Figure 3 is a graph (chromatogram) showing an example of analysis results by gas chromatography. In Figure 3, the vertical axis represents signal intensity, and the horizontal axis represents time (retention time). A gas chromatograph (hereinafter also referred to as "GC") is a device for separating a sample into its constituent components (substances) and quantifying each component. The analytical principle of GC is explained below. When a sample is introduced into the GC, the sample, including the solvent components, is heated and vaporized in the sample vaporization chamber. The vaporized sample moves through the column with the mobile phase (carrier gas), but the movement speed differs for each substance (e.g., compound). Therefore, there is a difference in the time it takes for each compound to reach the outlet of the column, and the compounds are separated. In GC, the amount of compound that reaches the outlet of the column is continuously measured from the time of sample introduction and output as an electrical signal. Therefore, in the chromatogram, a peak in signal intensity appears at the time when each compound is detected. The time from sample introduction to the appearance of the peak is called the "retention time".

[0043] The pattern of the chromatogram differs depending on the composition of the sample. Therefore, when diesel fuel is analyzed by GC, a chromatogram (analysis result) corresponding to the composition of the diesel fuel is output. Since the properties of diesel fuel differ depending on its composition, there is a correlation between the properties of the diesel fuel and the chromatogram.

[0044] [3-1-2. Property Data] Figure 4 shows an example of the properties of the diesel fuel to be estimated according to the embodiment. The properties of the diesel fuel to be estimated by the learning device 10 include density (15℃) [g / cm³]. 3 ], kinematic viscosity (30℃) [mm 2 This includes 18 items: [s], flash point [°C], 10% distillation temperature [°C], 50% distillation temperature [°C], 90% distillation temperature [°C], cetane index [-], sulfur content [mass ppm], HFRR [μm], cloud point [°C], clogging point [°C], pour point [°C], saturation content [volume fraction %], olefin content [volume fraction %], aromatic content [volume fraction %], mono-ring aromatic content [volume fraction %], bi-ring aromatic content [volume fraction %], and tri-ring or more aromatic content [volume fraction %]. Hereinafter, density (15°C) is also referred to as "Property 1," kinematic viscosity (30°C) as "Property 2," flash point as "Property 3," 10% distillation temperature as "Property 4," 50% distillation temperature as "Property 5," 90% distillation temperature as "Property 6," cetane index as "Property 7," sulfur content as "Property 8," HFRR as "Property 9," cloud point as "Property 10," clogging point as "Property 11," pour point as "Property 12," saturation content as "Property 13," olefin content as "Property 14," aromatic content as "Property 15," mono-ring aromatic content as "Property 16," bi-ring aromatic content as "Property 17," and three or more-ring aromatic content as "Property 18."

[0045] Here, "density" is the mass per unit volume of the sample. "Density (15°C)" is the density of the sample at 15°C. Density can be measured, for example, in accordance with JIS K 2249 (2011) "Petroleum products - Method for determining density". "Kinematic viscosity" is the value obtained by dividing viscosity by the density of the liquid under the same conditions (temperature, pressure), and represents the degree of resistance when the liquid flows due to the action of gravity. "Kinematic viscosity (30°C)" is the kinematic viscosity of the sample at 30°C. Kinematic viscosity can be measured, for example, in accordance with JIS K 2283 (2000) "Crude oil and petroleum products - Test method for kinematic viscosity and method for calculating viscosity index". "Flash point" is the lowest temperature of the sample at which, when an ignition source is brought close to the sample vapor, the sample vapor emits a flash of light and burns instantaneously, and the flame propagates across the liquid surface. The flash point can be measured, for example, in accordance with JIS K 2265-3 (2007) "Method for determining flash point - Part 3". The "distillation temperature" is the temperature corresponding to the amount of distillate at a certain point during distillation. The distillation temperature can be measured, for example, in accordance with JIS K 2254 (2018) "Petroleum products - Method for determining distillation properties". The "cetane index" is a value indicating the autoignition properties of diesel fuel. The cetane index can be measured, for example, in accordance with JIS K 2280-5 (2013) "Petroleum products - Method for determining octane number, cetane number and cetane index - Part 5: Cetane index". The "sulfur content" is the proportion of sulfur and sulfur compounds contained in the sample. The sulfur content can be measured, for example, in accordance with JIS K 2541 (2003) "Crude oil and petroleum products - Test method for sulfur content". "HFRR" is an index that expresses the lubricity of diesel fuel, determined in micrometers by the diameter of the wear marks produced when a test disc immersed in diesel fuel and a test steel ball are brought into contact while vibrating. HFRR can be measured, for example, by conforming to JPI-5S-50-98 "Diesel Fuel - Lubricity Test Method".

[0046] The "cloud point" is the temperature at which, when the sample is cooled without stirring, the sample at the bottom of the test tube becomes hazy or begins to cloud due to the precipitation of paraffin wax. The cloud point can be measured, for example, in accordance with JIS K 2269 (1987) "Test Method for Pour Point and Cloud Point of Crude Oil and Petroleum Products". The "clogging point" is the temperature at which, when the sample is cooled and drawn through a wire mesh with a mesh opening of 45 μm, the filtration time of 20 mL of the sample exceeds 60 seconds. The clogging point can be measured, for example, in accordance with JIS K 2288 (2000) "Petroleum Products - Diesel Fuel - Test Method for Clogging Point". The "pour point" is the lowest temperature at which the sample flows. The pour point can be measured, for example, in accordance with JIS K 2269 (1987) "Test Method for Pour Point and Cloud Point of Crude Oil and Petroleum Products". The "saturation content" is the proportion of paraffinic hydrocarbons and naphthenic hydrocarbons contained in the sample. The saturated content can be measured, for example, by following JPI-5S-49-07 "Petroleum Products - Hydrocarbon Type Test Method - High-Performance Liquid Chromatography Method". The "olefin content" is the proportion of unsaturated hydrocarbons contained in the sample. The olefin content can be measured, for example, by following JPI-5S-49-07 "Petroleum Products - Hydrocarbon Type Test Method - High-Performance Liquid Chromatography Method". The "aromatic content" is the proportion of aromatic hydrocarbons contained in the sample. The aromatic content can be measured, for example, by following JPI-5S-49-07 "Petroleum Products - Hydrocarbon Type Test Method - High-Performance Liquid Chromatography Method". The "mono-ring aromatic content" is the proportion of mono-ring aromatic hydrocarbons contained in the sample. The "mono-ring aromatic content" can be measured, for example, by following JPI-5S-49-07 "Petroleum Products - Hydrocarbon Type Test Method - High-Performance Liquid Chromatography Method". The "bi-ring aromatic content" is the proportion of bi-ring aromatic hydrocarbons contained in the sample. "Two-ring aromatic content" can be measured, for example, by following JPI-5S-49-07 "Petroleum Products - Hydrocarbon Type Test Method - High-Performance Liquid Chromatography Method". "Three-ring or more aromatic content" is the proportion of three-ring or more aromatic hydrocarbons contained in the sample. "Three-ring or more aromatic content" can be measured, for example, by following JPI-5S-49-07 "Petroleum Products - Hydrocarbon Type Test Method - High-Performance Liquid Chromatography Method".

[0047] [3-1-3. Data Integration and Splitting] For example, the user inputs diesel fuel property data and GC analysis result data into the terminal device 20. For each diesel fuel, one property data and one analysis result data are input into the terminal device 20. The terminal device 20 receives a large amount of diesel fuel property data and analysis result data. The terminal device 20 transmits the input property data and analysis result data to the learning device 10. The learning device 10 stores the received property data and analysis result data, for example, in a non-volatile memory 102.

[0048] The learning device 10 integrates the characteristic data and the analysis result data using the training data generation program 110 to generate integrated data. Furthermore, the learning device 10 divides the integrated data to generate divided data. For example, the learning device 10 can reduce the dimensionality of the analysis result data by applying downsampling, as described later, to the analysis result data, and then integrate the downsampled analysis result data with the characteristic data.

[0049] Figure 5 illustrates an example of data integration and splitting. The learning device 10 correlates the property data 301 and the analysis result data 302 with each other to generate integrated data 310. That is, the learning device 10 integrates the property data 301 and analysis result data 302 for the same diesel fuel to generate integrated data for that diesel fuel. The learning device 10 similarly generates integrated data for multiple diesel fuels.

[0050] Furthermore, the learning device 10 divides the integrated data 310 according to the properties (items) of diesel fuel, and generates divided data 320_1, 320_2, 320_3, ... That is, divided data 320_1 is data related to property 1, divided data 320_2 is data related to property 2, divided data 320_3 is data related to property 3, and so on for divided data 320_4 and so on. The divided data group 320A is formed by the divided data 320_1, 320_2, 320_3, ... for all properties.

[0051] Each segmented data set 320_1, 320_2, 320_3, ... contains all of the analysis result data 302 (all of the analysis result data 302 after downsampling). That is, segmented data set 320_1 is a combination of the data for property 1 from the property data 301 and the analysis result data 302. Segmented data set 320_2 is a combination of the data for property 2 from the property data 301 and the analysis result data 302. Segmented data set 320_3 is a combination of the data for property 3 from the property data 301 and the analysis result data 302. The same applies to segmented data sets 320_4 and below.

[0052] [3-1-4. Dimensionality Reduction] The analysis results data obtained by GC are multidimensional data with thousands to tens of thousands of dimensions. Therefore, the learning device 10 according to this embodiment performs dimensionality reduction processing on the analysis results data. In other words, preprocessing includes dimensionality reduction processing. This is expected to improve the accuracy of learning and inference, as well as reduce the processing load, compared to using the analysis results data directly for machine learning or inference. However, if the number of dimensions in the analysis results data is within an acceptable range for processing, dimensionality reduction processing can be omitted.

[0053] In this embodiment, the dimensionality reduction process includes downsampling and principal component analysis.

[0054] As described above, a chromatogram is data that shows the relationship between retention time (first element) and signal intensity (second element). GC samples signal intensity at each unit time. Therefore, chromatogram data contains signal intensity data for the number of samples taken. Downsampling is a process that reduces the number of samples in the chromatogram data. Downsampling is performed, for example, to convert the analysis result data into a format that can be processed by the learning device 10 or a format that is easy to process. Specifically, downsampling is a process that decimates the signal intensity sampled at each unit time at equal intervals in the time domain. In other words, downsampling is a process that reduces the temporal resolution of the chromatogram data, which is time-series data. That is, downsampling is a process that reduces the resolution of retention time, which is the first element in the analysis result. In this embodiment, downsampling does not reduce the resolution of signal intensity, which is the second element. However, downsampling may be a process that reduces the resolution of signal intensity along with the resolution of retention time. Downsampling can reduce, for example, tens of thousands of data points to several thousand to about 10,000 data points (dimensions).

[0055] Principal component analysis (PCA) is a process that reduces the number of dimensions of data by converting multivariate data into fewer variables (principal components). In this embodiment, the number of principal components in the analysis result data is reduced to tens or less in the PCA.

[0056] Figure 6 is a graph showing an example of the relationship between the number of principal components and the cumulative contribution in the analysis result data. In Figure 6, the vertical axis represents the cumulative contribution, and the horizontal axis represents the number of principal components. As the number of principal components increases from 0, the cumulative contribution increases rapidly, reaching approximately 90% when the number of principal components is 10. When the number of principal components exceeds 10, the cumulative contribution increases further gradually and saturates at a value close to 100% (approximately 95%). Therefore, it can be seen that the characteristics of the original data are sufficiently preserved even when the number of principal components is tens or less. Since the number of dimensions (explanatory variables) of the analysis result data after downsampling is several thousand to about 10,000, the number of explanatory variables can be greatly reduced by principal component analysis. In other words, principal component analysis is an example of a process that reduces the number of dimensions of the analysis result while preserving the characteristics of the relationship between the first element, retention time, and the second element, signal intensity.

[0057] The number of principal components is determined to a value appropriate for each property of diesel fuel. For example, by changing the number of principal components in the analysis result data, a score for the number of principal components can be calculated for each target variable (property), and the number of principal components (number of explanatory variables) can be determined based on the score corresponding to each target variable. The score is, for example, a value (evaluation value) used to evaluate the estimation performance of each of the trained models 230_1, 230_2, 230_3, ... which are generated (constructed) for each property. The score is, for example, a value calculated based on the estimated values ​​output from each of the trained models 230_1, 230_2, 230_3, ... The score is, for example, the mean absolute error (MAE) of the estimated values. For example, if MEA is adopted as the score, the number of principal components for which the score corresponding to a certain property (target variable) is smaller than a predetermined value (threshold) can be determined as the number of principal components corresponding to that property.

[0058] Figure 7 shows an example of the number of principal components in the analysis results data for each property of diesel fuel. In the example in Figure 7, the number of principal components is set to 15 or 25. Specifically, the number of principal components for kinematic viscosity (30°C), flash point, 10% distillation temperature, 50% distillation temperature, 90% distillation temperature, cetane index, HFRR, cloud point, clogging point, and pour point is "15". The number of principal components for density (15°C), sulfur content, saturated content, olefin content, aromatic content, mono-ring aromatic content, bi-ring aromatic content, and tri-ring or more aromatic content is "25".

[0059] Hereafter, the analysis results data after principal component analysis will also be referred to as "dimensionality reduction data." As mentioned above, the number of dimensions in the dimensionality reduction data is assigned to each characteristic. Hereafter, the dimensionality reduction data corresponding to characteristic 1 will be referred to as "dimensionality reduction data 1," the dimensionality reduction data corresponding to characteristic 2 will be referred to as "dimensionality reduction data 2," and the dimensionality reduction data corresponding to characteristic 3 will be referred to as "dimensionality reduction data 3." The same applies to dimensionality reduction data 4 and below.

[0060] [3-1-5. Generating training data] Figure 8 illustrates an example of the generation of training data. Each segmented data 320_1, 320_2, 320_3, ... included in segmented data group 320A is subjected to principal component analysis to generate training data group 330A, which consists of training data 330_1, 330_2, 330_3, ... for each characteristic. Training data 330_1 is a combination of characteristic 1 data from characteristic data 301 and dimensionality reduction data 1. Training data 330_2 is a combination of characteristic 2 data from characteristic data 301 and dimensionality reduction data 2. Training data 330_3 is a combination of characteristic 3 data from characteristic data 301 and dimensionality reduction data 3. The same applies to training data 330_4 and subsequent data.

[0061] [3-2. Learning Function] The learning function is implemented by the processor 101 executing the learning program 120. The learning function is a function that trains the pre-training model 130 using training data 330_1, 330_2, 330_3, ... so that when analysis result data (dimensionality reduction data) is input, it outputs values ​​that approximate the correct property values.

[0062] Figure 9 is a diagram illustrating an example of the learning function of the learning device according to the embodiment. Training data 330_1, 330_2, 330_3, ... are input to the learning device 10, and machine learning is performed. In the machine learning performed by the learning device 10, trained models 230_1, 230_2, 230_3, ... are generated (constructed) for each characteristic. Trained model 230_1 is an estimation model for estimating characteristic 1, trained model 230_2 is an estimation model for estimating characteristic 2, and trained model 230_3 is an estimation model for estimating characteristic 3. The same applies to trained models 230_4 and below. The trained model group 230A is composed of trained models 230_1, 230_2, 230_3, ...

[0063] Each of the trained models 230_1, 230_2, 230_3, ... is a regression model. A regression model is a model that expresses, for example, the relationship between an explanatory variable and a dependent variable using mathematical formulas. Examples of regression models include linear regression models, multiple regression analysis models, polynomial regression models, logistic regression, random forests, gradient boosting decision trees, Support vector machines, neural networks, etc., can be used.

[0064] Figure 10 shows an example of the configuration of a regression model, which is a neural network. A neural network includes an input layer, a hidden layer, and an output layer.

[0065] The input layer contains multiple nodes (shown as circles in the diagram). Each node in the input layer is input to the principal components s1, s2, s3, ... from the dimensionality reduction data.

[0066] The intermediate layer consists of one or more processing layers. In the example shown in Figure 10, the intermediate layer has a multiphase structure. The intermediate layer contains multiple nodes.

[0067] The output layer, for example, contains one node.

[0068] In the input, hidden, and output layers, nodes in adjacent layers are connected by edges (synaptic connections, shown as line segments in the diagram). Each edge is assigned a weight (connection weight). Machine learning tunes the weight of each edge.

[0069] The nodes in the output layer correspond to the properties of the target being estimated. The output layer nodes output the estimated property value A.

[0070] [3-3. Input Data Creation Function] The input data creation function is realized by the processor 101 executing the input data generation program 210. The input data creation function is a function that creates (generates) input data by preprocessing the analysis results of diesel fuel by gas chromatography. The input data is data to be input into the trained models 230_1, 230_2, 230_3, ...

[0071] Figure 11 illustrates an example of input data generation. The input data for the trained models 230_1, 230_2, 230_3, ... is generated based on the analysis results of diesel fuel obtained by gas chromatography. Specifically, the input data creation function generates input data 400 by reducing the dimensionality of the analysis result data 303. In other words, input data 400 is dimensionality-reduced data.

[0072] Specifically, the dimensionality reduction process in the input data creation function includes downsampling and principal component analysis. In other words, the dimensionality reduction process in the input data creation function is the same as the dimensionality reduction process in the training data creation process. That is, the data format of input data 400 is the same as the data format of dimensionality reduction data 1, 2, 3, ... included in training data 330_1, 330_2, 330_3, ...

[0073] The input data creation function generates input data with a number of dimensions corresponding to the properties to be estimated. Specifically, if the properties to be estimated are kinematic viscosity (30°C), flash point, 10% distillation temperature, 50% distillation temperature, 90% distillation temperature, cetane index, HFRR, cloud point, clogging point, or pour point, the number of dimensions of the generated input data will be "15". If the properties to be estimated are density (15°C), sulfur content, saturated content, olefin content, aromatic content, mono-ring aromatic content, bi-ring aromatic content, or three or more-ring aromatic content, the number of dimensions of the generated input data will be "25".

[0074] [3-4. Inference Function] The inference function is implemented by the processor 101 executing the pre-trained models 230_1, 230_2, 230_3, ... The inference function is the function of estimating the properties of diesel fuel using the pre-trained models 230_1, 230_2, 230_3, ...

[0075] When input data is provided to the trained models 230_1, 230_2, 230_3, ..., the trained models 230_1, 230_2, 230_3, ... estimate the properties of diesel fuel based on the input data and output the estimation results.

[0076] [4. Operation of the learning device] [4-1. Learning Phase] The operation of the learning device 10 during the learning phase is described below. The learning phase is the phase in which the learning device 10 performs machine learning. Figure 12 is a flowchart showing an example of the operation of the learning device during the learning phase according to the embodiment. The processor 101 performs the following operations by executing the training data generation program 110 and the learning program 120.

[0077] For example, the user pre-measures the properties 1, 2, 3, ... of several diesel fuel samples. Furthermore, each diesel fuel sample is subjected to GC analysis to obtain analysis result data 302. The user operates the terminal device 20 to input the property data 301 for the multiple diesel fuel samples into the terminal device 20. The input property data 301 is transmitted from the terminal device 20 to the learning device 10. The processor 101 accepts the received property data 301 (step S101).

[0078] The user operates the terminal device 20 to input analysis result data 302 for multiple diesel fuel samples into the terminal device 20. The input analysis result data 302 is transmitted from the terminal device 20 to the learning device 10. The processor 101 receives the received analysis result data 302 (step S102).

[0079] The processor 101 performs downsampling of the analysis result data 302 (step S103).

[0080] The processor 101 integrates the analysis result data 302 after downsampling with the property data 301 to generate integrated data 310 (step S104). The processor 101 then divides the generated integrated data 310 according to property to generate divided data 320_1, 320_2, 320_3, ... (step S105).

[0081] The processor 101 performs principal component analysis on each of the analysis result data contained in the partitioned data 320_1, 320_2, 320_3, ... to generate training data 330_1, 330_2, 330_3, ... (step S106). In principal component analysis, each analysis result data is reduced in dimensionality to the number of dimensions corresponding to each of the partitioned data 320_1, 320_2, 320_3, ...

[0082] The processor 101 inputs the generated training data 330_1, 330_2, 330_3, ... into the pre-trained model 130 and performs machine learning (step S107). This generates trained models 230_1, 230_2, 230_3, ... for each characteristic.

[0083] Processor 101 evaluates the estimation performance of each of the trained models 230_1, 230_2, 230_3, ... Specifically, the estimation performance of each of the trained models 230_1, 230_2, 230_3, ... is evaluated using evaluation data generated in the same way as the training data 330_1, 330_2, 330_3, .... The analysis result data (dimensionality reduction data) included in the evaluation data is input to the trained models 230_1, 230_2, 230_3, ... Processor 101 inputs the analysis result data of multiple evaluation data sets into each of the trained models 230_1, 230_2, 230_3, ... Multiple estimates are output from each of the trained models 230_1, 230_2, 230_3. The processor 101 calculates the mean absolute error (MAE) of the estimated values ​​as an evaluation value for each of the trained models 230_1, 230_2, and 230_3. The processor 101 evaluates the estimation performance by comparing the calculated evaluation value with a threshold. The threshold is set for each trained model 230_1, 230_2, 230_3, ... Note that the evaluation value is not limited to the mean absolute error (MAE) of the estimated values, but may also be, for example, the root mean square error (RMSE) of the estimated values.

[0084] The processor 101 determines whether or not to perform retraining based on the evaluation result of the estimated performance (step S108). The determination of whether or not to perform retraining is made for each trained model 230_1, 230_2, 230_3, ... That is, if the MAE calculated for each trained model 230_1, 230_2, 230_3, ... is smaller than the threshold, the processor 101 determines that the estimated performance of the trained models 230_1, 230_2, 230_3, ... is high and does not perform retraining (NO in step S108). If the MAE is above the threshold, the processor 101 determines that the estimated performance of the trained models 230_1, 230_2, 230_3, ... is low and decides to perform retraining (YES in step S108).

[0085] Suppose it is determined that the trained model 230_p corresponding to characteristic p should not be retrained, but the trained model 230_q corresponding to characteristic q should be retrained. The processor 101 discards (deletes) the trained model 230_q and returns to step S106. At this time, the trained model 230_p that was determined not to be retrained is not discarded. In retraining, for example, the number of dimensions of the analysis result data is changed from the first machine learning run. The processor 101 performs machine learning using training data that includes the analysis result data with the new number of dimensions, and generates a new trained model 230_q.

[0086] If it is determined that retraining is not necessary for all trained models 230_1, 230_2, 230_3, ... (NO in step S108), the operation of the learning device 10 in the learning phase ends.

[0087] [4-2. Estimation Phase] The operation of the learning device 10 during the estimation phase is described below. The estimation phase is the phase in which the learning device 10 estimates the properties of diesel fuel using the trained models 230_1, 230_2, 230_3, ... Figure 13 is a flowchart showing an example of the operation of the learning device during the estimation phase according to the embodiment. The processor 101 executes the following operations by running the input data generation program 210 and the trained models 230_1, 230_2, 230_3, ...

[0088] For example, the terminal device 20 displays an input screen for specifying the properties of the diesel fuel to be estimated. The user can operate the terminal device 20 to specify the properties to be estimated from among several diesel fuel properties. The terminal device 20 transmits the specification of the properties of the diesel fuel to be estimated to the learning device 10. The processor 101 accepts the received specification of the properties of the diesel fuel to be estimated (step S201).

[0089] As described above, each property is assigned a number of dimensions. The processor 101 determines the number of dimensions corresponding to the specified diesel fuel property to be estimated (step S202).

[0090] For example, the user pre-analyzes a diesel fuel sample to be analyzed by GC and obtains analysis result data 303. The user operates the terminal device 20 to input the analysis result data 303 for the diesel fuel sample to be analyzed. The terminal device 20 transmits the input analysis result data 303 to the learning device 10. The processor 101 accepts the received analysis result data 303 (step S203).

[0091] The processor 101 performs downsampling of the analysis result data 303 (step S204).

[0092] The processor 101 performs principal component analysis on the downsampled analysis result data 303 to generate input data 400 (step S205). In principal component analysis, the analysis result data 303 is reduced in dimensionality to the number of dimensions determined in step S202.

[0093] The processor 101 selects a pre-trained model from among several pre-trained models 230_1, 230_2, 230_3, ... that corresponds to the specified property of the target to be estimated (step S206).

[0094] The processor 101 inputs the input data 400 to the selected trained model and performs the estimation of the properties of the diesel fuel (step S207). The trained model outputs the estimation result of the properties of the diesel fuel.

[0095] The processor 101 transmits (outputs) the estimated results of the diesel fuel properties output from the trained model to the terminal device 20 (step S208). The terminal device 20 displays the received estimated results of the diesel fuel properties on its display device. This completes the operation of the training device 10 in the estimation phase.

[0096] [5. Variant] In the embodiments described above, the input data for the trained models 230_1, 230_2, 230_3, ... were used as the results of diesel fuel analysis by GC. However, the embodiments are not limited to this, and the results of diesel fuel analysis by analytical instruments other than GC may be used, as long as the analytical instrument measures the content of substances in the sample. For example, the results of diesel fuel analysis by liquid chromatography, Fourier transform infrared spectrophotometer (FT-IR), or nuclear magnetic resonance (NMR) can be used.

[0097] A liquid chromatograph is a device that pressurizes a liquid mobile phase using a pump or the like and passes it through a column, separating and detecting analytes (substances in a sample) by utilizing the difference in interactions (adsorption, partitioning, ion exchange, size exclusion, etc.) between the stationary phase and the mobile phase. The analytical result data of a liquid chromatogram, like that of a GC, is a chromatogram that shows the relationship between time (retention time) and signal intensity. Therefore, when using the analytical results of a liquid chromatograph, the number of dimensions of the analytical result data can be reduced by the same dimensionality reduction process as in the embodiment described above.

[0098] FT-IR is a device that performs structural analysis or quantification of a sample by irradiating it with infrared light and measuring the transmitted or reflected light. Since the infrared light irradiated onto the sample is absorbed based on the vibrational and rotational motion of molecules in the substance, the content of the substance is expressed in the transmittance or absorbance at the corresponding wavenumber or wavelength. The analysis result from FT-IR is a graph (IR spectrum) showing the relationship between wavenumber or wavelength and transmittance or absorbance. In the FT-IR analysis result, the first element is wavenumber or wavelength, and the second element is transmittance or absorbance. When using the analysis result from FT-IR, the dimensionality of the analysis result can be reduced by downsampling to lower the resolution at wavenumber or wavelength and performing principal component analysis.

[0099] NMR is a device used to analyze the molecular structure or physical properties of compounds in a sample. Atomic nuclei with a nuclear spin quantum number of 1 that is not 0 possess a magnetic dipole moment, and when placed in a magnetic field, the nucleus resonates with electromagnetic waves (radio waves) of a specific frequency due to Zeeman fission. The resonance frequency differs depending on the nuclide, but even within the same nuclide, differences arise depending on the chemical or magnetic environment in which each atomic nucleus is placed. This is called the chemical shift, and the type of functional group contained in a compound can be estimated from the chemical shift of the NMR signal. The analysis result by NMR is a graph (NMR spectrum) that shows the relationship between the resonance frequency (chemical shift) and the relative signal intensity. In the analysis result by NMR, the first element is the resonance frequency, and the second element is the relative signal intensity. When using the analysis result by NMR, the dimensionality of the analysis result can be reduced by downsampling to lower the resolution at the resonance frequency and performing principal component analysis.

[0100] In the embodiments described above, each of the trained models 230_1, 230_2, 230_3, ... estimated one corresponding property, but this is not limited to this. A single trained model that estimates multiple properties may be generated by machine learning. In this case, the training data can be a dataset that combines the correct values ​​for each of the multiple properties to be estimated with the analysis result data.

[0101] In the embodiments described above, the learning device 10 is configured to perform both machine learning and inference of diesel fuel properties, but the invention is not limited to this configuration. For example, the learning device 10 may be configured to perform machine learning but not inference of diesel fuel properties. For example, the trained models 230_1, 230_2, 230_3, ... generated by machine learning may be stored in the non-volatile memory of an estimation device, which is a different computer from the learning device 10, and the estimation device may perform inference of diesel fuel properties.

[0102] In the embodiment described above, the learning device 10 performs downsampling of the analysis data in the learning phase, then integrates the property data and the analysis result data, and then divides the integrated data. However, the invention is not limited to this configuration. The learning device 10 may integrate the analysis result data before downsampling with the property data and then downsample the integrated data, or it may divide the integrated data before downsampling and then downsample each divided data.

[0103] [6. Supplementary Notes] The embodiments disclosed herein are illustrative and not restrictive in all respects. The scope of rights in this disclosure is defined by the claims rather than by the embodiments described above, and includes the meaning of equivalents to the claims and all modifications within that scope. [Explanation of symbols]

[0104] 1. Machine Learning Systems 10. Learning device (information processing device) 20 Terminal devices 30 communication lines 101 Processors 102 Non-volatile memory 103 Volatile memory 104 Interface (IF) 110 Training Data Generation Program 120 Learning Programs 130 Pre-trained models 210 Input Data Generation Program 230_1, 230_2, 230_3, ... Pre-trained models 230A Pre-trained model group 301 Property Data 302,303 Analysis Results Data 310 Integrated Data 320_1,320_2,320_3,... Split data 320A Divided Data Set 330_1, 330_2, 330_3, ... Training data 330A Training Data Set 400 Input Data

Claims

1. A memory that stores a learning model, which is a computer program for estimating the properties of diesel fuel, An information processing device comprising a processor that executes the aforementioned learning model, The aforementioned processor, The process involves preprocessing the analysis results obtained by an analytical device that measures the content of substances in a diesel fuel sample, and generating input data for input into the learning model. The steps include: performing processing by the learning model based on the generated input data; Perform an action that includes Information processing device.

2. The aforementioned preprocessing includes a dimensionality reduction process to reduce the number of dimensions of the analysis results. The information processing apparatus according to claim 1.

3. The aforementioned analysis results are data showing the relationship between the first and second elements, which are different from each other. The dimensionality reduction process is a process that reduces the number of dimensions of the analysis result while preserving the characteristics of the relationship between the first element and the second element. The information processing apparatus according to claim 2.

4. The dimensionality reduction process includes a process of reducing the number of dimensions of the analysis result by principal component analysis. The information processing apparatus according to claim 3.

5. The dimensionality reduction process includes a process to reduce the resolution of the first element in the analysis result. The information processing apparatus according to claim 3.

6. The aforementioned input data is training data for supervised learning, and is data that combines property data showing the properties of the diesel fuel and analysis result data showing the analysis results. The processing performed by the aforementioned learning model involves performing machine learning using the learning model, which is the model before training, to generate a trained model. The information processing apparatus according to any one of claims 1 to 5.

7. The step of generating the input data includes generating first training data corresponding to the first properties of the diesel fuel and second training data corresponding to the second properties of the diesel fuel. The process for generating the pre-trained model involves generating a first pre-trained model for estimating the first characteristic using the first training data, and generating a second pre-trained model for estimating the second characteristic using the second training data. The information processing apparatus according to claim 6.

8. The operation further includes the step of determining whether the error in the property value of the diesel fuel, which is output data obtained when evaluation data showing the analysis results of the diesel fuel by the analyzer is input to the generated trained model, is smaller than a specified value. If the error in the property value of the diesel fuel as output data exceeds a specified value, the processor discards the trained model and repeats the steps of generating the input data and performing the processing by the trained model. The information processing apparatus according to claim 6.

9. The aforementioned input data is data to be input to the trained model, which is a trained model, and includes analysis result data showing the analysis results. The processing by the aforementioned learning model is a process in which the trained model processes the input data to estimate the properties of the diesel fuel. The information processing apparatus according to any one of claims 1 to 5.

10. The properties of the diesel fuel include a first property and a second property that are different from each other. The trained model includes a first trained model for estimating the first property and a second trained model for estimating the second property. The information processing apparatus according to claim 9.

11. An information processing method using a learning model, which is a computer program for estimating the properties of diesel fuel, The process involves preprocessing the analysis results obtained by an analytical device that measures the content of substances in a diesel fuel sample, and generating input data for input into the learning model. The steps include: performing processing by the learning model based on the generated input data; including, Information processing methods.

12. A computer program for generating input data to be input into a learning model, which is a computer program for estimating the properties of diesel fuel, On the computer, The step involves preprocessing the analysis results obtained by an analytical device that measures the content of substances in a diesel fuel sample, and generating input data for input into the learning model. To execute Computer program.

13. A learning model, which is a computer program for estimating the properties of diesel fuel, On the computer, A step of estimating the properties of the diesel fuel based on input data generated by pre-processing the analysis results from an analytical device that measures the content of substances in a diesel fuel sample. To execute A learning model.

Images