Analysis device, analysis method, program for analysis device, analytical learning device, analytical learning method, and program for analytical learning device

Abstract

Provided is an analysis device for analyzing a measurement sample on the basis of spectral data obtained from the measurement sample. The analysis device comprises: a correlation data storage unit for storing correlation data indicating the correlation between spectral data regarding reference samples for which a total analysis value of a predetermined plurality of components is already known and the total analysis values of the reference samples; and a calculation main unit that applies the correlation data to spectral data regarding a measurement sample, and calculates the total analysis value of a predetermined plurality of components included in the measurement sample. The reference samples include a first reference sample including the predetermined plurality of components, and a second reference sample composed of one or more of the components included in the first reference sample. The correlation data indicates a machine learning model calculated using teacher data, namely, first reference sample data including the spectral data regarding the first reference sample and the total analysis value of the predetermined plurality of components included in the first reference sample, and second reference sample data including the spectral data regarding the second reference sample and the total analysis value of the predetermined plurality of components included in the second reference sample.

Description

[Technical field]

[0001] The present invention relates to an analytical device that analyzes a measurement sample based on spectral data obtained from the measurement sample. [Background technology]

[0002] Conventionally, for example, FID analyzers and spectroscopic analyzers such as FTIR have been used to measure the concentration and amount of total hydrocarbons (hereinafter, also referred to as THC) contained in automobile exhaust gases and the like.

[0003] Although FID analyzers excel in analytical precision, they require the supply of hydrogen gas (H2) as a combustion support gas and helium gas (He) as a carrier gas, which makes them difficult to handle and increases running costs.

[0004] On the other hand, FTIR analyzers have the advantage of being easier to use and having lower running costs, but they suffer from poor analytical accuracy. In other words, FTIR analyzers perform a two-step calculation in which the concentration of each hydrocarbon (HC) is calculated individually from the optical spectrum, and then these are weighted and added together, and errors that may occur in measuring the concentration of each HC are superimposed on errors that may occur in setting the weighting coefficients, making it difficult to improve measurement accuracy.

[0005] In order to improve the measurement accuracy of an FTIR analyzer, Patent Document 1 describes an FTIR analyzer that calculates the correlation between the THC concentration and the optical spectrum obtained by shining light on a reference sample by machine learning, and estimates the THC concentration by applying the optical spectrum of the measurement sample to a machine learning model that indicates the calculated correlation. Patent Document 1 describes the use of a gas (e.g., automobile exhaust gas) that contains multiple types of hydrocarbons, which is the same type as the measurement sample, as the reference sample. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] WO2019 / 031331 publication Summary of the Invention [Problem to be solved by the invention]

[0007] However, since many hydrocarbons highly correlated with each other are mixed in exhaust gas, there is a problem that it is difficult to separate and learn the contribution of each hydrocarbon to the THC concentration when machine learning is performed using only the optical spectrum of the exhaust gas as training data, as in the above-mentioned analysis device of Patent Document 1. Therefore, for example, if the composition of the hydrocarbons in the measurement sample deviates from the hydrocarbon components in the learned data, it may be difficult to obtain sufficient analysis accuracy, which is called overlearning.

[0008] Therefore, the main objective of the present invention is to improve the measurement accuracy in an analytical device such as an FTIR spectroscopic analyzer that estimates the total analytical value of a specified number of components such as THC from the spectral data of a measurement sample.

[0009] That is, the analytical apparatus of the present invention analyzes a measurement sample based on spectral data obtained from the measurement sample, and comprises a correlation data storage unit that stores correlation data indicating a correlation between the spectral data of a reference sample in which a total analytical value of a predetermined number of components is known and the total analytical value of the reference sample, and a calculation main unit that applies the correlation data stored in the correlation data storage unit to the spectral data of the measurement sample to calculate the total analytical value of the predetermined number of components contained in the measurement sample, wherein the reference sample includes a first reference sample containing the predetermined number of components and a second reference sample consisting of one or more components contained in the first reference sample, and the correlation data is data indicating a machine learning model calculated by machine learning using as teacher data first reference sample data including the spectral data of the first reference sample and the total analytical value of the first reference sample, and second reference sample data including the spectral data of the second reference sample and the total analytical value of the second reference sample.

[0010] The spectral data includes not only the spectral data of light transmitted through, reflected from, or scattered by the measurement sample (or reference sample), but also the spectral data of light absorbed by the measurement sample (or reference sample) (absorption spectral data), corrected absorption spectral data in which the effects of interfering components contained in the measurement sample (or reference sample) are reduced or removed, and even the mass spectrum obtained by ionizing the measurement sample (or reference sample). The total analytical value is the total of the physical quantities of a plurality of components, such as the total concentration of the components or the total mass of the components.

[0011] In this case, the machine learning model used to calculate the total analysis value is trained using not only the measurement data of a first reference sample containing multiple components (e.g., exhaust gas containing multiple hydrocarbons, etc.) but also the measurement data of a second reference sample (e.g., hydrocarbons) consisting of one or multiple components contained in the first reference sample as training data, so that the accuracy is higher than that of a model that learns the contribution of each component to the total analysis value of multiple components, for example. By using a machine learning model with improved accuracy and improved robustness against changes in the components of the measurement sample, the analysis accuracy can be improved in an analysis device that estimates the total analysis value of a specified component from the spectral data of the measurement sample. Note that the second reference sample only needs to contain at least the components contained in the first reference sample, and may contain components other than the components contained in the first reference sample.

[0012] It is preferable to use, as the second reference sample, one or more components constituting the predetermined multiple components. In this way, by individually learning the contribution of each component (e.g., hydrocarbons) to the total analytical value (e.g., THC concentration), it is possible to avoid overfitting the machine learning model and further improve the analytical accuracy of the analyzer.

[0013] As the second reference sample, it is preferable to use one or more components whose total analysis value is zero. In this case, the components constituting the second reference sample are preferably the components contained in the measurement sample. With this, it is possible to learn the spectra of components that do not contribute to the total analytical value of the specified multiple components, and if such components are contained in the measured sample, it is possible to prevent them from being erroneously added to the total analytical value of the specified multiple components.

[0014] Incidentally, when the measurement sample is exhaust gas, for example, as shown in Figure 8, the THC concentration and H2O concentration in the exhaust gas appear to have a linear relationship at first glance, but the relationship may change depending on the operating conditions (dotted line area). In other words, it can be said that there is a spurious correlation between the THC concentration and H2O concentration in the exhaust gas. When learning such measurement data, the data in the dotted line area, which occurs less frequently, is not given importance, and a learning model is created that uses the wavenumber representing absorption by H2O to estimate THC, resulting in a deterioration in analytical accuracy in the dotted line area.

[0015] Therefore, it is preferable to use a second reference sample that is composed of one or more components that have a spurious correlation with the total analytical value. In this way, by having the system learn that components that have a spurious correlation with the total analysis value of a specified number of components do not contribute to the total analysis value, it is possible to avoid learning the above-mentioned spurious correlation, and the analytical accuracy of the analysis device can be further improved.

[0016] In addition, for example, in the case of an automobile or the like, when the engine misfires or the temperature is extremely low, fuel combustion deteriorates significantly, and the exhaust gas contains a large amount of unburned fuel vapor and contains a larger amount of heavier hydrocarbon components than normal exhaust gas components. Therefore, it is preferable to use the fuel that generates the exhaust gas as the second reference sample. By learning the fuel spectrum, such a system can perform highly accurate analysis under a wide range of conditions, including the above-mentioned engine misfires and extremely low temperatures.

[0017] In order to utilize a machine learning model that is more accurate and robust to changes in the components of the measurement sample, it is preferable that the second reference sample is composed of the components contained in the first reference sample.

[0018] In addition, it is preferable that in the analytical apparatus, the correlation data storage unit stores multiple correlation data calculated for each fuel type, and the calculation main unit switches the correlation data to be applied to the spectral data obtained from the measurement sample depending on the fuel type generating the measurement sample. In this way, the correlation data calculated for each combustion species can be used separately, thereby further improving the analytical accuracy.

[0019] In a specific embodiment in which the effects of the present invention are significantly exhibited, the measurement sample or the first reference sample is an automobile exhaust gas, and the component to be analyzed is a hydrocarbon (HC), and the total analytical value of a plurality of predetermined components is a THC concentration. In addition, the present analytical device is preferably an FTIR type, and when analyzing THC, it is preferable to measure the total analytical value of the reference sample using an FID analytical device.

[0020] The present invention also includes an analytical learning device that is specialized for the function of calculating only correlation using only a reference sample. In this case, the system preferably includes a receiving unit that receives spectral data obtained from a reference sample for which the total analytical value of a predetermined plurality of components is known, a reference sample data storage unit that stores reference sample data including the total analytical value and spectral data of a plurality of different reference samples, and a correlation calculation unit that uses the reference sample data as teacher data to calculate a common correlation between the spectral data and the total analytical value of each reference sample by machine learning, wherein the reference samples include a first reference sample containing the predetermined plurality of components and a second reference sample consisting of one or more components contained in the first reference sample, and the reference sample data includes first reference sample data including the spectral data of the first reference sample and the total analytical value of the predetermined plurality of components contained in the first reference sample, and second reference sample data including the spectral data of the second reference sample and the total analytical value of the predetermined plurality of components contained in the second reference sample. Effect of the Invention

[0021] According to the present invention thus configured, it is possible to improve the measurement accuracy in an analysis device such as an FTIR spectroscopic analysis device. [Brief description of the drawings]

[0022] [Figure 1] 1 is an overall view of an exhaust gas measurement system including an analysis device according to an embodiment of the present invention; [Diagram 2] FIG. 2 is a schematic diagram showing the entire analysis device according to the embodiment. [Diagram 3] FIG. 2 is a functional block diagram of a processor in the embodiment. [Figure 4] 4 is a flowchart showing the operation of the analysis device in the embodiment. [Diagram 5] 6 is a graph showing experimental results using the analysis device of the embodiment. [Figure 6] 6 is a graph showing experimental results using the analysis device of the embodiment. [Figure 7] FIG. 13 is a functional block diagram of a processor according to another embodiment. [Figure 8] A diagram explaining the spurious correlation between THC and H2O. [Explanation of symbols]

[0023] 100...Analyzer 51...Main analysis section 52 Total analysis value calculation section 521 Correlation calculation unit 522... Calculation body 53 Reception DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] An analysis device 100 according to one embodiment of the present invention will be described below with reference to the drawings.

[0025] The analysis device 100 of this embodiment constitutes a part of an exhaust gas measurement system 200. This exhaust gas measurement system 200 includes a chassis dynamo 300, an FID analysis device 400, and the analysis device 100, as shown in FIG. This analytical device 100 is a Fourier transform infrared spectroscopic analyzer, commonly referred to as FTIR, and is used to simultaneously calculate the concentration of one or more components, such as inorganic compounds, hydrocarbons, nitrogen compounds, etc., contained in a measurement target. Specifically, this analytical device 100 (hereinafter, for the sake of distinction, it may also be referred to as FTIR analytical device 100) includes a light source 1, an interferometer (spectroscopic section) 2, a sample cell 3, a photodetector 4, a calculation processing device 5, etc., as shown in FIG. 2.

[0026] The light source 1 emits light having a broad spectrum (continuous light including light of many wave numbers), and may be, for example, a tungsten-iodine lamp or a high-intensity ceramic light source.

[0027] As shown in the figure, the interferometer 2 utilizes a so-called Michelson interferometer, which is equipped with one half mirror (beam splitter) 21, a fixed mirror 22, and a movable mirror 23. The light from the light source 1 that enters this interferometer 2 is split into reflected light and transmitted light by the half mirror 21. One light is reflected by the fixed mirror 22, and the other is reflected by the movable mirror 23, returns to the half mirror 21, and is combined before being emitted from the interferometer 2.

[0028] The sample cell 3 is a transparent cell into which exhaust gas, which is a measurement sample, is introduced, and the light emitted from the interferometer 2 passes through the measurement sample in the sample cell 3 and is guided to the photodetector 4 . The photodetector 4 is what is referred to here as a MCT photodetector 4 .

[0029] The arithmetic processing device 5 includes analog electric circuits having a buffer, an amplifier, etc., digital electric circuits having a CPU, a memory, a DSP, etc., and an A / D converter and the like interposed therebetween.

[0030] The arithmetic processing device 5, by cooperation of a CPU and its peripheral devices in accordance with a predetermined program stored in the memory, performs the function of a main analysis unit 51 which calculates transmitted light spectrum data indicating the spectrum of light transmitted through the sample from the output value of the photodetector 4, calculates absorption spectrum data from the transmitted light spectrum data, identifies various components contained in the measurement sample, and calculates the concentration (or amount) of each component, as shown in FIG. 3.

[0031] The main analysis section 51 includes a spectral data generation section 511 and an individual component analysis section 512 .

[0032] When the movable mirror 23 is moved back and forth and the light intensity transmitted through the sample is observed with the position of the movable mirror 23 on the horizontal axis, in the case of light with a single wavenumber, the light intensity draws a sine curve due to interference. On the other hand, the actual light transmitted through the sample is continuous light and the sine curve differs for each wavenumber, so the actual light intensity is a superposition of the sine curves drawn by each wavenumber, and the interference pattern (interferogram) takes the form of a wave packet.

[0033] The spectrum data generating unit 511 obtains the position of the movable mirror 23 by a range finder such as a HeNe laser (not shown), obtains the light intensity at each position of the movable mirror 23 by the photodetector 4, and converts the interference pattern obtained from these into transmitted light spectrum data with each wave number component on the horizontal axis by performing a fast Fourier transform (FFT). Then, based on the transmitted light spectrum data measured in advance with the sample cell empty, for example, the transmitted light spectrum data of the measurement sample is further converted into absorption spectrum data.

[0034] The individual component analysis section 512 identifies various components contained in the measurement sample from, for example, the positions (wave numbers) and heights of the peaks in the absorption spectrum data, and calculates the concentration (or amount) of each component.

[0035] Thus, the analyzer 100 of this embodiment is used as an exhaust gas analyzer for measuring the THC concentration (or amount) in exhaust gas, which is a measurement sample. In this embodiment, as shown in Fig. 3, the arithmetic processing device 5 is further provided with functions as a reception unit 53, a total analysis value calculation unit 52, etc., so that the THC concentration (or amount) of the measurement sample can be measured with high accuracy.

[0036] The receiving unit 53 receives the THC concentration of a gas (exhaust gas in this case) containing multiple types of hydrocarbons measured by the FID analyzer 400. The exhaust gas whose THC concentration is known as measured by the FID analyzer 400 is hereinafter referred to as a first reference sample.

[0037] Incidentally, this first reference sample is also introduced into the FTIR analysis device 100 together with the FID analysis device 400, and its absorption spectrum data is also acquired by the main analysis unit 51. The reception unit 53 also receives the absorption spectrum data of the first reference sample, which is intermediate information calculated by the main analysis unit 51, and links this to the THC concentration of the first reference sample measured by the FID analysis device 400 to obtain first reference sample data, which is then stored in the reference sample data storage unit D1 set in a predetermined area of ​​memory.

[0038] In the analysis device 100 of this embodiment, the receiving unit 53 is configured to receive not only the exhaust gas but also the THC concentrations of a plurality of single component gases. The THC concentrations of the single component gases are determined in advance, and each single component gas with a known THC concentration is hereinafter referred to as a second reference sample.

[0039] Specifically, the second reference sample is composed of one or more components contained in the first reference sample, more specifically, (1) Methane (CH4), toluene (C7H8) and octane (C8H 18 ) and other hydrocarbon gases (components constituting the specified multiple components in the claims), (2) Gases containing carbonyl carbon (carbon having a C=O double bond) (formaldehyde, formic acid, etc.), etc., that are insensitive to FID analysis equipment (components whose total analysis value of a specified number of components as claimed is zero), (3) Inorganic gases (H2O, CO2, CO, NO, NO2, N2O, NH3, etc.), etc., that have a pseudo-correlation with the THC concentration in exhaust gas (components that have a pseudo-correlation with the total analytical value of a specified number of components, as referred to in the claims) etc.

[0040] The THC concentration of the second reference sample may be measured in advance by the FID analyzer 400, or may be calculated based on the single component gas concentration if the single component gas concentration is known. If the single component gas constituting the second reference sample does not theoretically contribute to the THC concentration, the THC concentration may be set to zero without measurement. In this embodiment, the THC concentration of the "hydrocarbon gas" that is sensitive to the FID analyzer is measured by the FID analyzer 400. On the other hand, the THC concentrations of the "FID insensitive gas" and the "pseudo correlation gas", which are gases that do not theoretically contribute to the THC concentration, are set to zero without being measured by the FID analyzer 400.

[0041] This second reference sample is introduced into the FTIR analysis device 100 in the same manner as the first reference sample, and its absorption spectrum data is acquired by the main analysis unit 51. The reception unit 53 also receives the absorption spectrum data of the second reference sample, which is intermediate information calculated by the main analysis unit 51, and links this to the THC concentration of the second reference sample to obtain second reference sample data, which is stored in the reference sample data storage unit D1.

[0042] Furthermore, in this embodiment, the surrounding condition data including at least the temperature and pressure of the first and second reference samples can be acquired by a sensor (not shown) provided in the system or an input by an operator. The receiving unit 53 is configured to acquire the surrounding condition data of the first and second reference samples, attach the data to the first and second reference sample data, and store the data in the reference sample data storage unit D1.

[0043] The total analysis value calculation unit 52 calculates the concentration of THC in a measurement sample (exhaust gas) from the absorption spectrum data of the measurement sample using multiple reference sample data stored in the reference sample data storage unit D1 as teacher data, and more specifically, includes a correlation calculation unit 521 and a calculation main unit 522. Note that this THC corresponds to the multiple components referred to in the claims, and the THC concentration corresponds to the total analysis value referred to in the claims.

[0044] The correlation calculation unit 521 refers to a plurality of first reference sample data stored in the reference sample data storage unit D1 as teacher data, performs machine learning on the correlation between the absorption spectrum data and the THC concentration that is common to the first reference sample data, and calculates a machine learning model.

[0045] Here, the correlation calculation unit 521 also refers to a plurality of second reference sample data as training data in order to calculate a more accurate machine learning model.

[0046] Specifically, the correlation calculation unit 521 uses the second reference sample data including the absorption spectrum data of the hydrocarbon gas and the THC concentration as teacher data, and thereby learns the contribution of each hydrocarbon component in the absorption spectrum data of the first reference sample individually. In addition, the correlation calculation unit 521 uses the absorption spectrum data of the FID insensitive gas and the second reference sample data containing the THC concentration (=0) as training data, so that the correlation calculation unit 521 learns that components containing carbonyl carbon (i.e., components for which the FID analyzer has no sensitivity) do not contribute to the THC concentration. In addition, the correlation calculation unit 521 learns that the pseudo-correlated gas does not contribute to the THC concentration (i.e., avoids learning the pseudo-correlation) by using the second reference sample data including the absorption spectrum data of the pseudo-correlated gas and the THC concentration (=0) as teacher data.

[0047] The correlation data indicating the machine learning model calculated by the correlation calculation unit 521 in this manner is stored in a correlation data storage unit D2 set in a predetermined area of ​​the memory.

[0048] In addition, in the correlation calculation section 521, learning is repeated every time each reference sample data is added, and the correlation is updated, so that the more each reference sample data is, the more the correlation accuracy improves.

[0049] Furthermore, the correlation calculation unit 521 in this embodiment is configured to calculate the correlation using the surrounding condition data of each reference sample as a parameter, i.e., the correlation changes depending on the temperature, pressure, etc. of each reference sample, but it is not necessary to refer to the surrounding condition data when calculating the correlation.

[0050] The calculation main unit 522 calculates the THC concentration of the measurement sample by fitting the correlation calculated by the correlation calculation unit 521 to the spectral data of the measurement sample. At this time, since the reception unit 53 is configured to acquire peripheral situation data of the measurement sample, the calculation main unit 522 applies the correlation corresponding to the peripheral situation data of the measurement sample when calculating the THC concentration.

[0051] Next, the operation of the exhaust gas measurement system 200 having such a configuration will be described with reference to FIG.

[0052] The learning operation will be described. First, the absorption spectrum data and THC concentration of the first reference sample are obtained (step S1). Specifically, an automobile is driven on the chassis dynamo 300, and the exhaust gas, which is the first reference sample, is guided to the FID analyzer 400 and the FTIR analyzer 100. In addition to driving the automobile on the chassis dynamo 300, an engine connected to the engine dynamo may be driven, or a drive system component such as a transmission may be driven on a drive system dynamo. Then, the FID analyzer 400 measures the THC concentration, while the FTIR analyzer 100 measures the absorption spectrum data of the exhaust gas by the main analysis unit 51. In this embodiment, the FID analyzer 400 and the FTIR analyzer 100 perform exhaust gas measurement (analysis) in synchronization with each other at a certain timing (several msec to several sec), for example. For each measurement, the receiving unit 53 acquires the exhaust gas absorption spectrum data calculated by the main analysis unit 51 and the THC concentration of the exhaust gas analyzed by the FID analysis device 400, and stores them one after another in the reference sample data storage unit D1 as first reference sample data. At this time, the receiving unit 53 acquires the temperature and pressure of the exhaust gas, attaches them to the reference sample data, and stores them in the reference sample data storage unit D1.

[0053] However, the condition of a vehicle's engine changes over time as the engine speed fluctuates, and the condition of the exhaust gas (components, pressure, temperature) also changes accordingly. Therefore, by the sequential measurements described above, data can be obtained from multiple reference samples, each of which has at least different THC concentrations.

[0054] Next, the absorption spectrum data and the THC concentration of the single component gas, which is the second reference sample, are obtained (step S2).

[0055] Specifically, first, a hydrocarbon gas (single component) as the second reference sample is introduced from a gas cylinder or the like to the FID analyzer 400 and the FTIR analyzer 100. Then, as with the first reference sample described above, the THC concentration is measured by the FID analyzer 400, the absorption spectrum data of the hydrocarbon gas is measured by the FTIR analyzer 100, and the measured THC concentration and absorbance spectrum data are stored as the second reference sample data in the reference sample data storage unit D1. Here, multiple different types of hydrocarbon gases (methane, toluene, octane, etc.) are introduced in sequence to the FID analyzer 400 and the FTIR analyzer 100, and the THC concentration and absorbance spectrum data for each hydrocarbon gas is acquired and stored in sequence.

[0056] Next, an FID insensitive gas (single component) as a second reference sample is introduced from a gas cylinder or the like to the FTIR analyzer 100, and the absorption spectrum data of the FID insensitive gas is measured by the FTIR analyzer 100. Here, the THC concentration of the FID insensitive gas (i.e., zero) is input in advance by an operator to the reception unit 53. Each time the absorption spectrum data of the FID insensitive gas is measured, the reception unit 53 associates it with the previously input THC concentration and stores it in the reference sample data storage unit D1 as the second reference sample data. As with the hydrocarbon gas, a plurality of different types of FID insensitive gases (formaldehyde, formic acid, etc.) are introduced in sequence to the FTIR analyzer 100, and the THC concentration and absorbance spectrum data of each FID insensitive gas are stored in sequence.

[0057] Next, the second reference sample, a pseudo-correlated gas (single component gas), is introduced from a gas cylinder or the like into the FTIR analysis device 100, and similarly to the FID insensitive gas, the measured absorption spectrum data is linked to the THC concentration input in advance (i.e., zero) and stored in the reference sample data storage unit D1 as the second reference sample data.

[0058] Then, the correlation calculation unit 521 refers to many pieces of first reference sample data and second reference sample data stored in the reference sample data storage unit D1, calculates the correlation between the absorption spectrum data and the THC concentration by machine learning (step S3), and stores the resulting correlation in the correlation data storage unit D2 (step S4). This ends the learning.

[0059] After the learning is completed in this manner, it becomes possible to perform actual THC concentration measurement using only the FTIR analyzer 100, without using the FID analyzer 400. In measuring the THC concentration, another vehicle to be tested is placed on the chassis dynamometer 300 and driven, and its exhaust gas is introduced into the FTIR analyzer 100.

[0060] In the FTIR analyzer 100, the main analysis unit 51 acquires the absorption spectrum data of the exhaust gas (step S5). Then, the total analysis value calculation unit 52 (calculation main unit 522) applies the correlation stored in the correlation data storage unit D2 to the absorption spectrum data to calculate the THC concentration (step S6).

[0061] Thus, the present analytical device 100 configured as described above learns not only measurement data of exhaust gas (first reference sample) containing multiple hydrocarbon components, but also measurement data of single-component gas (second reference sample) such as hydrocarbon gas, FID insensitive gas, and pseudo-correlated gas as teacher data, so that it is possible to calculate a highly accurate machine learning model that has learned information on the contribution of each hydrocarbon component to the THC concentration, components that have no contribution to the THC concentration, and components that have a pseudo-correlation.The use of such a highly accurate machine learning model makes it possible to estimate the THC concentration with high accuracy from the spectral data of the measurement sample.

[0062] FIG. 5 shows a comparison between the analytical accuracy of the THC concentration by the analytical device 100 (analytical device A) of one embodiment of the present invention and the analytical accuracy of the THC concentration by a conventional analytical device (analytical device B). The analytical device A calculates the THC concentration using a machine learning model trained on the above-mentioned exhaust gas (first reference sample), hydrocarbon gas, FID insensitive gas, and pseudo-correlated gas (second reference sample) as teacher data. On the other hand, the analytical device B calculates the THC concentration using a machine learning model trained on only the exhaust gas (first reference sample) as teacher data. Using these analytical devices, exhaust gas from a vehicle test-driven under various conditions was analyzed and the THC concentration was calculated. The exhaust gas was also analyzed by an FID analyzer to calculate the THC concentration, and the error (estimated error) from the THC concentration measured by each analytical device was calculated. The results are shown in FIG. 5. FIG. 6 is a graph showing the state of the exhaust gas (gas concentration / THC ratio) in the dotted line area and the two-dot chain line area in FIG. 5.

[0063] As can be seen from Figure 5, in the case of analyzer B, which learned only from exhaust gas as training data, the estimation error was about 10% on average. In contrast, in the case of analyzer A, which learned from exhaust gas as well as hydrocarbon gas, FID insensitive gas, and pseudo-correlated gas as training data, the estimation error was 5% or less for all measurement targets, indicating that the analytical accuracy was improved. In particular, the effect of improving the analytical accuracy by learning from a second reference sample such as a hydrocarbon gas as training data was more pronounced under conditions where the proportion of methane was high (the two-dot chain line area in Figures 5 and 6) compared to conditions where the proportion of methane was low (the dotted line area in Figures 5 and 6).

[0064] The present invention is not limited to the above-described embodiment. For example, since the absorbance of THC contains moisture and other interfering components in the measurement target, the THC concentration may be calculated based on a corrected absorption spectrum that reduces or eliminates the effects of these components, thereby further improving the accuracy of the analysis.

[0065] Furthermore, an analytical learning device that performs only correlation calculations may be configured by removing the functions of individual component analysis unit 512 and calculation main unit 522 from this analytical device and introducing only a reference sample. The correlations found by this analytical learning device can be used in other FTIR analytical devices.

[0066] Furthermore, in calculating the correlation, not only values ​​related to the physical attributes of the samples, such as the temperature and pressure of the first and second reference samples, but also other surrounding condition data may be used as parameters. For example, engine combustion information (information related to supercharging, EGR, rich / stoichiometric / lean, laminar flow, uniform flow, direct injection, port injection, etc.), engine head shape, ignition timing, catalyst configuration, fuel type, fuel oxygen amount, inorganic gas components, soot concentration, SOF concentration, engine model, engine speed, load information, hot start, cold start, oxygen concentration, catalyst temperature, gear ratio, room temperature, etc. may be taken into account as parameters. In addition, the calculated correlation may be divided into one or more parameters (for example, engine model, combustion system, catalyst, fuel type, etc.). That is, the analysis device 100 may have multiple types of correlation data classified into one or more parameters as the correlation data stored in the correlation data storage unit D2.

[0067] Conversely, it is also possible to configure the system so that all or part of the surrounding conditions data is not used as a parameter for calculating the correlation, but rather the surrounding conditions data that has a strong influence (i.e., is highly related) on the THC concentration (total analysis value) calculated and measured by the analysis device is extracted. In this way, it is possible to grasp surrounding condition data that is highly correlated with THC concentration, i.e., design parameters, making it possible to provide this to automobile manufacturers and catalyst manufacturers as a design and development support system.

[0068] Depending on the surrounding condition data used for the correlation, for example, if the engine model is different, re-learning may be required each time. In that respect, if the amount of surrounding condition data to be taken into account is reduced and a correlation between the spectrum data and the THC concentration (total analysis value) can be obtained, versatility will be increased. Specifically, it is possible to consider a configuration in which no surrounding condition data is taken into account in the correlation calculation, or in which the physical state of the sample itself (e.g., pressure, temperature, refractive index, viscosity, etc.) is taken into account in the correlation calculation, but other external attributes (e.g., engine model, ignition timing, etc.) are not taken into account in the correlation calculation.

[0069] When the temperature of the exhaust gas (first reference sample) is used as a parameter as a surrounding condition in calculating the correlation, it is preferable to measure the outlet temperature of the tailpipe of the automobile on the chassis dynamometer 300 by a sensor or the like and use the measured temperature as the parameter.

[0070] The range of the spectrum data used for learning and analysis may be only the wave number range in which the analysis target component is included, or may be expanded to a predetermined range beyond that. Also, the wave number range of the interference component may be excluded.

[0071] Specifically, the range of the spectral data used for learning and analysis is set to 2800 cm -1 More than 3200cm -1 The following may be used: If the range of the spectrum data is as described above, it will include the wavenumber range of the HC, which is the target component of analysis, as well as the wavenumber range of the interfering component water (approximately 3400 cm -1 Since the above-mentioned substances are eliminated, the influence of water on the calculation of the THC concentration of the measurement sample can be reduced, and the measurement accuracy can be further improved.

[0072] The concentration of each of the HCs having a small molecular weight may be determined in the main analysis section 51, and the total concentration of one or more HCs having a large molecular weight may be subtracted from the THC concentration.

[0073] In another embodiment, the correlation data storage unit D2 may store a plurality of (e.g., two or more) pieces of correlation data calculated for each concentration range of the THC concentration. In this case, the correlation calculation unit 521 may divide the THC concentration into a plurality of (e.g., two or more) concentration ranges, calculate the correlation between the absorption spectrum data and the THC concentration for each concentration range, and store the correlation data in the correlation data storage unit D2. Then, when the calculation main unit 522 receives the absorption spectrum data of the measurement sample, it may select one piece of correlation data suitable for the absorption spectrum data from the plurality of correlation data stored in the correlation data storage unit D2 based on the area of ​​the absorption spectrum data, and fit the selected correlation data to the spectral data of the measurement sample to calculate the THC concentration of the measurement sample.

[0074] In another embodiment, the correlation data storage unit D2 may store a plurality of correlation data calculated for each fuel type (e.g., gasoline, alcohol content, bio-based ether, etc.). In this case, the calculation main unit 522 may select correlation data for the closest fuel type from the plurality of correlation data stored in the correlation data storage unit D2 according to the fuel type that generates the exhaust gas that is the measurement sample, and calculate the THC concentration of the measurement sample using the selected correlation data. In this case, the calculation main unit 522 may be configured to identify the fuel type based on the concentration of each individual component calculated by the individual component analysis unit 512 and the absorption spectrum data generated by the spectrum data generation unit 511.

[0075] In this case, the analysis device 100 may further include a fuel type correlation data storage unit that stores fuel type correlation data calculated by machine learning on the correlation between the absorption spectrum data of the measurement sample and the fuel type. In this case, when the calculation main unit 522 receives the absorption spectrum data of the measurement sample, the calculation main unit 522 may apply the correlation indicated by the fuel type correlation data to the absorption spectrum data to identify the fuel type. Then, the calculation main unit 522 may select correlation data corresponding to the identified fuel type from the correlation data storage unit D2 to calculate the THC concentration.

[0076] The analytical device 100 of the above embodiment calculates the correlation between the spectral data of the reference sample and the THC concentration by itself, but is not limited to this. The analytical device 100 of another embodiment may use a correlation calculated in advance by another analytical learning device that only performs correlation calculation, and calculate the THC concentration directly from the spectral data of the measurement sample based on the correlation. Specifically, the analytical device 100 may be configured such that the arithmetic processing device 5 does not have the functions of the reference sample data storage unit D1 and the correlation calculation unit 521, as shown in Fig. 7. Here, the reception unit 53 may receive correlation data (i.e., learned data) indicating a correlation previously calculated by another analytical learning device via a network or the like, and store the correlation data in advance in the correlation data storage unit D2. The calculation main unit 522 may then fit the correlation data previously stored in the correlation data storage unit D2 to the absorption spectrum data of the measurement sample to calculate the THC concentration of the measurement sample. Note that accepting unit 53 may accept new correlation data from other analytical learning devices at predetermined intervals and periodically update the correlation data stored in correlation data storage unit D2.

[0077] In the above embodiment, the second reference sample is all of the hydrocarbon gas, the FID insensitive gas, and the pseudo-correlated gas, but this is not limited thereto. In other embodiments, only some of these may be used as the second reference sample, or a single component gas other than these may be used. In addition, the second reference sample is not limited to gas, and may be liquid.

[0078] In the above embodiment, the hydrocarbon gas, the FID insensitive gas, and the pseudo-correlated gas used as the second reference sample are each composed of a single component, but this is not limited thereto. In other embodiments, the second reference sample may be composed of multiple components, as long as it is different from the first reference sample and its THC concentration is known.

[0079] In another embodiment, the fuel (which may be in either liquid or gaseous state) that is the combustion source that generates the exhaust gas that is the measurement sample may be used as the second reference.

[0080] In the above embodiment, the first reference sample and the measurement sample are exhaust gases, but they may be air, other gases, or liquids.

[0081] The first reference sample does not need to be the same type as the measurement sample, and may be a standard gas generated by mixing multiple components to be analyzed with a main component such as nitrogen. In this case, since the total analytical value of the multiple components is known, there is no need to use a separate analytical device to analyze the multiple components of the reference sample.

[0082] Furthermore, the components to be analyzed are not limited to hydrocarbons (HC), but may include other components such as non-methane hydrocarbons (NMHC), non-methane, non-ethane hydrocarbons (NMNEHC), petroleum hydrocarbons (PH) in soil, volatile organic compounds (VOC) in the ambient air, the calorific value of petroleum fuels, nitrogen oxides, and dioxins.

[0083] The analytical device of the present invention can be applied to any analytical device that measures a plurality of components and measures the sum of the components. For example, the analytical device can be applied to an analytical device that irradiates a measurement sample with light and analyzes the spectrum, an analytical device that ionizes a measurement sample and analyzes the mass spectrum obtained, an NDIR, or a mass spectrometer, and can also be applied to a scattering type particle size distribution measuring device other than a spectroscopic analyzer. Furthermore, the present invention is not limited to the analysis of exhaust gas from automobiles, but can also be used to analyze exhaust gas from internal combustion engines of ships, aircraft, agricultural machines, machine tools, etc.

[0084] Furthermore, the present invention is not limited to the above-described embodiment, and it goes without saying that various modifications are possible without departing from the spirit and scope of the present invention. [Industrial Applicability]

[0085] According to the present invention, it is possible to improve the measurement accuracy in an analysis device such as an FTIR spectrometer.

Claims

1. A method for analyzing a measurement sample based on spectral data obtained from the measurement sample, a correlation data storage unit storing correlation data indicating a correlation between spectrum data of a reference sample having a known total analysis value of a plurality of predetermined components and the total analysis value of the plurality of predetermined components contained in the reference sample; a calculation body that applies correlation data stored in the correlation data storage unit to spectrum data obtained from the measurement sample to calculate a total analysis value of the predetermined plurality of components contained in the measurement sample, The reference sample includes a first reference sample including the predetermined plurality of components, and a second reference sample including one or more components that are part of the first reference sample; The correlation data is First reference sample data including spectral data of the first reference sample and the total analysis value of the first reference sample; An analytical apparatus showing a machine learning model calculated using, as training data, spectral data of the second reference sample and second reference sample data including the total analysis value of the second reference sample.

2. The analytical device according to claim 1 , wherein the second reference sample is one or more components that make up the predetermined plurality of components.

3. 3. The analytical apparatus according to claim 1, wherein the second reference sample comprises one or more components that do not contribute to a total analytical value of the predetermined plurality of components in a reference sample whose total analytical value is known.

4. 4. The analytical apparatus according to claim 1, wherein the second reference sample comprises one or more components that have a spurious correlation with the total analysis value.

5. 5. The analytical device according to claim 1, wherein the second reference sample is a fuel that produces exhaust gas.

6. 6. The analytical device according to claim 1, wherein the measurement sample or the first reference sample is an exhaust gas, and the predetermined plurality of components are hydrocarbons.

7. 7. The analyzer according to claim 6, wherein the total analysis value of the predetermined plurality of components is a concentration of total hydrocarbons contained in the exhaust gas.

8. The analytical device according to any one of claims 1 to 7, which is of an FTIR type.

9. 9. The analyzer according to claim 1, wherein the total analysis value of the first reference sample and the total analysis value of the second reference sample are measured by an FID analyzer.

10. The correlation data storage unit stores a plurality of correlation data calculated for each fuel type, 10. The analytical device according to claim 1, wherein the calculation main body switches correlation data to be applied to the spectral data obtained from the measurement sample depending on the type of fuel that generates the measurement sample.

11. A method for analyzing a measurement sample based on spectral data obtained from the measurement sample, comprising: a correlation data storage step of storing correlation data indicating a correlation between spectrum data of a reference sample having a known total analysis value of a plurality of predetermined components and the total analysis value of the plurality of predetermined components contained in the reference sample; and a calculation step of calculating a total analysis value of the predetermined plurality of components contained in the measurement sample by applying the stored correlation data to the spectrum data obtained from the measurement sample, The reference sample includes a first reference sample including the predetermined plurality of components, and a second reference sample including one or more components that are part of the first reference sample; The correlation data is First reference sample data including spectral data of the first reference sample and the total analysis value of the first reference sample; An analytical method showing a machine learning model calculated using, as training data, spectral data of the second reference sample and second reference sample data including the total analysis value of the second reference sample.

12. A program installed in an analytical device that analyzes a measurement sample based on spectral data obtained from the measurement sample, a function of serving as a correlation data storage unit that stores correlation data indicating a correlation between spectrum data of a reference sample, the total analysis value of which is known, and the total analysis value of the predetermined plurality of components contained in the reference sample; and causing the analytical device to function as a calculation main unit that applies correlation data stored in the correlation data storage unit to spectrum data obtained from the measurement sample to calculate a total analysis value of the predetermined plurality of components contained in the measurement sample. The reference sample includes a first reference sample including the predetermined plurality of components, and a second reference sample including one or more components that are part of the first reference sample; The correlation data is First reference sample data including spectral data of the first reference sample and the total analysis value of the first reference sample; A program for an analytical device showing a machine learning model calculated using, as training data, spectral data of the second reference sample and second reference sample data including the total analysis value of the second reference sample.

13. a receiving unit for receiving spectrum data obtained from a reference sample having a known total analysis value of a plurality of predetermined components; a reference sample data storage unit storing reference sample data including total analysis values ​​and spectral data of the predetermined plurality of components of a plurality of different reference samples; a correlation calculation unit that uses the reference sample data as teacher data to calculate a common correlation between the spectral data of each reference sample and the total analysis value by machine learning; The reference sample includes a first reference sample including the predetermined plurality of components, and a second reference sample including one or more components that are part of the first reference sample; The reference sample data is First reference sample data including spectral data of the first reference sample and the total analysis value of the first reference sample; and second reference sample data including spectral data of the second reference sample and the total analysis value of the second reference sample.

14. a spectral data receiving step of receiving spectral data obtained from a reference sample having a known total analysis value of a plurality of predetermined components; a reference sample data storage step of storing reference sample data including total analysis values ​​and spectral data of the predetermined plurality of components of the plurality of reference samples different from each other; a correlation calculation step of calculating a common correlation between the spectral data of each reference sample and the total analysis value by machine learning using the reference sample data as teacher data, The reference sample includes a first reference sample including the predetermined plurality of components, and a second reference sample including one or more components that are part of the first reference sample; The reference sample data is First reference sample data including spectral data of the first reference sample and a total analysis value of the predetermined plurality of components contained in the first reference sample; A learning method for analysis, comprising: second reference sample data including spectral data of the second reference sample and a total analysis value of the predetermined multiple components contained in the second reference sample.

15. a function of serving as a reception unit for receiving spectrum data obtained from a reference sample having a known total analysis value of a plurality of predetermined components; a function as a reference sample data storage unit that stores reference sample data including total analysis values ​​and spectral data of the predetermined plurality of components of a plurality of different reference samples; a function of a correlation calculation unit that calculates a common correlation between the spectral data of each reference sample and a total analysis value by machine learning using the reference sample data as teacher data in the analytical learning device; The reference sample includes a first reference sample including the predetermined plurality of components, and a second reference sample including one or more components that are part of the first reference sample; The reference sample data is First reference sample data including spectral data of the first reference sample and the total analysis value of the first reference sample; A program for an analytical learning device, comprising: second reference sample data including spectral data of the second reference sample and the total analysis value of the second reference sample.

Images