JP2023149303A - Compound identification method and compound identification device - Google Patents

Abstract

To provide a method for classifying and estimating metal elements and compounds included in an unknown sample.SOLUTION: A compound identification method of the present invention performs factorization of spectral data for a standard sample of a compound species to be assumed to derive a factor and a factor score representing the degree of influence of the factor, and learning and setting in advance a score threshold for classifying the compound species according to the factor score in a decision tree and a factor application order to apply the factor score to the decision tree, calculates a factor score in an unknown sample by using a factor of the standard sample to spectral data of the unknown sample, and classifying the factor score of the unknown sample in the decision tree by using the score threshold according to the factor application order to identify the compound species included in the unknown sample.SELECTED DRAWING: Figure 1

Description

Patent Law Article 30, Paragraph 2 application filed 7th Next Generation Initiative Decommissioning Technology Conference (NDEC-7) to be held on March 11, 2022

The present invention relates to a method for classifying and estimating metal elements and compounds contained in an unknown sample.

In preparation for the decommissioning of nuclear power plants, countermeasures for contaminated water, fuel removal from spent fuel pools, fuel debris removal, and waste management are necessary. In a nuclear reactor with a melted core, melted fuel, control rods, etc. are thought to be mixed together and solidified as fuel debris. In nuclear reactors that use boron carbide (B ₄ C) as a control material, borides are produced that are approximately twice as hard as oxides, so fuel debris can be separated while distinguishing between metals, oxides, and borides. It is efficient to take out the It has also been confirmed that there is a correlation between the concentration of boron contained in boride and the hardness. In order to study the fuel debris retrieval method, it is necessary to identify the material and material hardness of the unknown sample. In addition, the creation of material distribution and hardness distribution within unknown samples is also being considered.

Methods for qualitative and quantitative analysis of elements in a sample include inductively coupled plasma optical emission spectrometry (ICP-AES) and laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS). When some energy is applied to a sample from the outside, the elements contained in the sample emit light with a wavelength unique to the element, so elemental analysis is performed by measuring the wavelength and amount of emitted light. With ICP-AES, accurate analysis may not be possible due to interference. Note that, as described in Patent Document 1, an invention of a foreign matter analysis device that can accurately and easily analyze foreign matter contained in a sample is also disclosed.

Further, there is laser-induced breakdown spectroscopy (LIBS) as a method for elemental analysis regardless of the state of solid, liquid, gas, etc. LIBS is an atomic emission spectrometry method that uses a high-energy pulsed laser as an excitation source. It can be analyzed in a short time without complicated pretreatment steps, and can analyze light elements (boron, oxygen) to heavy elements (metals). Since it is possible to simultaneously measure up to 100% of waste, it is expected to be applied to obtaining elemental information on high-level waste, measuring hardness, and criticality control based on boron distribution information. Note that, as described in Patent Document 2, an invention of a method for calculating the composition of oxygen and boron in a material containing metal as a main component and the hardness of the material is also disclosed.

In LIBS, when a measurement target is irradiated with a high peak power laser with a pulse width of nanoseconds or less, plasma at tens of thousands of degrees Celsius is instantaneously formed. At this time, the object to be measured is atomized and excited by plasma formation, and as the plasma disappears, the excited atoms are relaxed to the ground state, accompanied by fluorescence emission unique to the element. This is measured to identify and quantify the element to be measured. Fluctuations in laser output and plasma generated by the elements constituting the measurement target cause minute fluctuations in time and space, resulting in fluctuations in measured values.

In the case of LIBS, the background based on the plasma state (time, space, etc.), and in the case of energy discrimination processing, the scattered components caused by the interaction between radiation and matter greatly influence the measured values. In spectroscopic measurements that use radiation such as lasers, X-rays, and neutrons as probes, although there are various factors and effects, a function to cancel or correct them is desired in order to perform highly accurate measurements. .

Patent No. 6507757 JP2020-067337A

However, when analyzing an unknown solid sample, for alloys or melts containing many elements, spectral data with many emission wavelengths are obtained, and it may be difficult to distinguish between elements and compounds. For example, if a large number of emission lines of iron (Fe) or zirconium (Zr) overlap, quantitative analysis may be hindered.

Furthermore, when measuring not only lasers but also X-rays, neutrons, etc. using a spectrometer, dark current is measured with the spectrometer closed to eliminate its influence. In order to eliminate the measurement environment (air environment, inert environment, etc.), some methods use blank data (measurement without a sample) to remove the influence. However, these effects cannot be applied if they are not added to the device functions or if similar samples do not exist.

In addition, for optical emission spectroscopy using lasers (LIBS and LA-ICP-MS) and for neutrons with self-shielding, it is necessary to evaluate the error inherent in the sample, but for neutrons and plasma emission, the irradiation energy is Often cannot be controlled.

In particular, in LIBS measurements, the baseline of the spectrum (connecting points thought to be background and showing it as a continuous line) may fluctuate depending on the atmospheric gas. For example, in an argon (Ar) atmosphere, the luminescence sensitivity is high and white noise may occur where the baseline is broad over a wide range of wavelengths. In addition, multiple detectors are used to detect light with wavelengths from 190 nm to 900 nm, but since the detectable wavelength range changes depending on the characteristics of the optical system, the baseline may shift at the wavelength at which the detector switches. It will happen.

In this way, it is possible to reduce background noise components (fluctuation errors such as spectrometer dark current) in spectral analyzes such as spectroscopic analysis and radiation analysis, and to optimize the probe conditions using light and radiation for the object to be measured in various analyses. There is a need for the mitigation of damage caused by probes to objects to be measured) and the identification of various compounds (including metals, ceramics, etc.).

Therefore, an object of the present invention is to provide a method for classifying and estimating metal elements and compounds contained in an unknown sample. Another purpose is to evaluate and correct errors inherent in samples when classifying and estimating metal elements and compounds.

In order to solve the above problems, the compound identification method of the present invention decomposes spectral data of standard samples of assumed compound types into factors to derive factors and factor scores representing the degree of influence of the factors. The score threshold for classifying compound types using the factor scores in the decision tree and the order of application of the factors to apply the factor scores to the decision tree are learned and set in advance, and the spectral data of the unknown sample is applied to the standard sample. By calculating the factor score of the unknown sample using a factor, and classifying the factor score of the unknown sample using the decision tree according to the order of application of the factors using the score threshold, the compound species contained in the unknown sample can be determined. It is characterized by identifying.

In the compound identification method, the spectral data is corrected to a baseline approximated by an asymmetric least squares method with a penalty term so that the error is evaluated more strictly when the measured value is below the estimated value than when it is above the estimated value. , is characterized by.

In the compound identification method, the factorization is performed by converting a K-by-N matrix Y consisting of the number of wavelength samples (wavelength dimension) K of the spectral data and the number of samples N using specified M non-negative basis vectors. The method is characterized in that the degree of influence of the factor obtained by decomposition is derived as the factor score.

In the compound identification method, the decision tree is characterized in that factors and score thresholds are derived such that a cost function using impurity has a minimum value when classifying the compound type.

The compound identification method is characterized in that the factor scores are mapped to estimate components and composition ratios of the unknown sample.

Further, the compound identification device of the present invention includes an irradiation device that irradiates the sample with a laser, a spectrometer that decomposes light emitted from the sample irradiated with the laser, and a spectrum generated by the spectrometer. A plurality of detectors that acquire data separately for each wavelength in a predetermined range, and the compound contained in the sample is identified using spectrum data input from the detectors. .

According to the present invention, metal elements and compounds contained in an unknown sample can be classified and estimated. By decomposing the spectrum data of known substances into multiple factors in advance and learning how to classify elements and compounds using a decision tree, the system can be used to classify the components of unknown samples.

Furthermore, by evaluating errors inherent in the sample and correcting them in advance, metal elements and compounds can be classified and estimated with high accuracy. Note that this can be corrected afterward, and errors including white noise can be removed.

FIG. 1 is a schematic diagram of a measuring device in the compound identification method of the present invention. 1 is a flowchart showing the flow of a preprocessing stage and a learning stage of the compound identification method of the present invention. FIG. 3 is a diagram illustrating correction of spectral data in the compound identification method of the present invention. FIG. 3 is a diagram illustrating correction of spectral data in the compound identification method of the present invention. FIG. 2 is a diagram illustrating factorization of spectral data in the compound identification method of the present invention. FIG. 2 is a diagram illustrating factorization of spectral data in the compound identification method of the present invention. FIG. 2 is a diagram illustrating classification using a decision tree in the compound identification method of the present invention. FIG. 2 is a diagram illustrating classification using a decision tree in the compound identification method of the present invention. FIG. 2 is a diagram illustrating classification using a decision tree in the compound identification method of the present invention. FIG. 2 is a flowchart showing the flow of operational steps of the compound identification method of the present invention. FIG. 2 is a diagram showing the results of classification using the compound identification method of the present invention.

Embodiments of the present invention will be described in detail below with reference to the drawings. Components having the same functions are designated by the same reference numerals, and repeated explanations may be omitted.

The present invention relates to a measurement system for spectroscopic analysis that uses radiation such as visible light-infrared light, This relates to an algorithm that targets two types of processing. This algorithm consists of a machine learning method that uses measured data (including predicted data) as learning data, and the configuration of a data processing system resulting from this method, and is expected to improve measurement accuracy and robustness as it is implemented. Furthermore, since it can be implemented as software processing, it is highly easy to implement into analysis programs included in existing spectroscopic analyzers.

In order to improve the analytical accuracy of element identification and quantification using LIBS (laser-induced breakdown spectroscopy), we simultaneously measure parameters (reference parameters) that characterize the plasma state (time, space, etc.) in addition to elemental emission spectra. Then, the correlation between the two is used for correction processing. Specifically, reference parameters that are strongly correlated with the intensity and shape of the atomic emission spectrum to be measured are set. In the past, measurements were made by an experienced measurer who understood the fluctuation components and characteristics. In order to obtain high analytical accuracy, it is necessary to conduct evaluations that understand the spectral variation mechanism.

Machine learning is used to set reference parameters. In the conventional method, the spectrum of the measurement target is searched as a peak with a certain width, but in the present invention, it is difficult to handle big data (wavelength-count value (light intensity) obtained from various analytical devices etc.). A machine learning method (NMF: non-negative matrix factorization) is applied to a set of all spectral components (values) to construct reference parameters. By using all spectral components rather than signal peaks, it is possible to evaluate the spectrum including its intensity and shape when identifying an unknown sample.

The present invention was inspired by a method for extracting a specific sound such as a violin or viola from an orchestra, or a method for extracting a specific individual from a crowd, and was developed into a measuring device. In the algorithm of the present invention, the spectral data of the standard sample processed in the preprocessing phase is trained using NMF and decision trees in the learning phase, and unknown data is learned in the operation phase. Applied to samples.

FIG. 1 is an overview diagram of a measuring device in the compound identification method of the present invention. The measurement device 100 includes a chamber 200, an irradiation device 300, a spectrometer 400, a detector 410, a control device 220, and the like, and performs LIBS measurement.

The chamber 200 is a vacuum container or the like, and the sample 110 is placed inside the chamber 200, which is depressurized using a vacuum pump 210. As the sample 110, a known sample such as a standard sample 110a is mainly used in the learning stage, and an unknown sample 110b to be measured and analyzed is used in the operational stage. The standard sample 110a is a sample whose composition is known in advance, such as a metal element or a compound, and the unknown sample 110b is a sample whose composition is unknown, such as fuel debris brought in from a nuclear power plant.

The irradiation device 300 is a laser head or the like equipped with a laser oscillator, and irradiates the sample 110 in the chamber 200 with irradiation light 310 such as a pulsed laser. As the irradiation light 310, a Nd:YAG laser with a wavelength of 1064 nm is used, which uses a crystal in which Nd (neodymium) is added to YAG (yttrium, aluminum, garnet) as a medium.

By irradiating the irradiation light 310, the surface of the sample 110 is turned into plasma, and when the surface returns to the ground state, the emitted light 320 is detected. The emitted light 320 is made parallel using a collimator or the like, and then enters the spectroscope 400. The spectrometer 400 is a prism, a diffraction grating, or the like, and separates light into wavelengths, and obtains spectrum data 420 representing the intensity distribution by arranging a plurality of detectors 410 in parallel for each wavelength in a predetermined range. , to the control device 220.

The control device 220 is a computer or the like, and controls each device, as well as classifying and analyzing the sample 110 from the acquired spectrum data 420. In addition, a sequencer 230, an electric stage 240, cooling water circulation equipment, etc. may be provided as necessary.

For example, the stage 240 on which the sample 110 is placed is moved under the control of the sequencer 230 to scan multiple points. The laser irradiation position is obtained from the position information (XYZ axis coordinate information) of the stage 240, and the shape information (height information) of the sample 110 is obtained using the displacement meter, and the coordinates are mapped by the control device 220. to visualize shape and intensity information in three dimensions.

FIG. 2 is a flowchart showing the flow of the preprocessing stage and learning stage of the compound identification method. The preprocessing stage includes standard sample preparation (setting the compound type of the standard sample) L1, spectrum data acquisition (baseline estimation) L2, and the learning stage includes NMF execution (obtaining the factors and factor scores of the standard sample) L3, and determination. Tree rule generation (setting of score threshold and factor application order) has step L4.

In the preprocessing stage and the learning stage, a standard sample 110a is prepared by setting the assumed compound type, and the acquired spectrum data 420 of the standard sample 110a is decomposed into factors to derive the factors and factor scores of the standard sample 110a. , a score threshold for classifying compound species by factor score in a decision tree, and the order in which factors are applied to the decision tree are learned and set in advance as rules of the decision tree.

The steps related to the pre-processing stage consist of standard sample preparation L1 and spectrum data acquisition L2 shown in FIG. In the standard sample preparation L1 step, compound types such as metal elements and compounds assumed to be components of the unknown sample 110b to be measured are set, and the metal elements and compounds are prepared as the standard sample 110a. The standard sample 110a serves as data prepared in advance in order to see which factor has what degree of influence when identifying the compound species contained in the unknown sample 110b.

In the spectrum data acquisition L2 step, spectrum data 420 sent from the detector 410 (including data stored in advance in a storage device, etc.) is input to the control device 220. Since the spectrum data 420 includes noise and the like and the baseline 430 (baseline) is disturbed, the control device 220 performs baseline estimation. In the baseline estimation, the spectral data 420 is corrected so that the baseline 430 becomes smooth.

Next, an example of the process related to the above-mentioned pretreatment stage obtained by performing LIBS measurement will be described with reference to FIGS. 3 to 4. 3 and 4 are diagrams illustrating correction of spectral data in the compound identification method.

FIG. 3(a) is a spectrum obtained by dividing a plurality of measurement wavelength ranges in the LIBS measurement system and using a highly sensitive spectrometer for each to achieve a wide band (from 190 nm to 900 nm). As shown in FIG. 3A, it can be seen that the baseline 430 deviates at the boundary of the wavelength range of the detector 410. The deviation causes discontinuity in the baseline 430 due to the sensitivity of the detector 410 and dark current.

In the wavelength range where discontinuity occurs, the measurement results include many uncertain factors, making the results unreliable. Furthermore, in the case of an object with a spectrum in a discontinuous portion, it is not possible to accurately calculate the area, and in some cases, the spectrum cannot be distinguished.

Moreover, FIG. 3(b) is a spectrum obtained by enlarging a part of the spectrum data 420 of FIG. 3(a). As shown in FIG. 3(b), unless optimal smoothing is applied to the resolution of the detector, the baseline will be adjusted more than necessary.

Therefore, the inventors used asymmetric weight penalty least squares smoothing (asymmetric least squares method with penalty term) as shown in FIG. Continuity). The asymmetric least squares method with penalty term determines the desirability (goodness) of the estimated value z _i of the baseline 430 with respect to the measured value y _i so that the difference between the measured value and the estimated value is not only minimized but also smoothed. A preferable baseline 430 is derived by minimizing the evaluation function.

From the viewpoint of evaluation, first, from the viewpoint of the squared error Σ _i (y _i −z _i ) ² , when y _i ≦z _i , the error is made very strict, and when y _i >z _i , the error is An asymmetrical evaluation is performed in which the value of z _i is made very lenient, and furthermore, from the perspective of the fluctuation of z i , a strict evaluation is performed on changes in the fluctuation. w _i is an asymmetric weight, which is p if y _i >z _i and (1-p) if y _i ≦z _i . p (p∈[0,1]: p is a very small value) has the role of changing w _i depending on whether the measured value exceeds the estimated value. μ is a regularization parameter, and the variation (non-smoothness) of the estimated value Σ _i (Δ ² z _i ) ² is a penalty term. Note that Δ ² represents a second-order differential operator, and w _i and μ are hyperparameters determined empirically by the designer. All you have to do is change the parameters and find the optimal solution.

As a result, the spectral data 420 is based on a base approximated by the asymmetric least squares method with a penalty term so that the error is evaluated more strictly when the measured value is less than the estimated value than when the measured value exceeds the estimated value. Corrected to line 430.

FIG. 4 shows a spectrum obtained by correcting the spectrum data 420 in FIG. 3 by the correction shown in FIG. 3(c). Note that FIG. 4(a) is a spectrum obtained by correcting the spectral data 420 in FIG. 3(a), and FIG. 4(b) is a spectrum obtained by enlarging a part of the spectral data 420 in FIG. 4(a). .

As shown in FIG. 4(a), the smoothed baseline 430 after correction eliminates the spectral discontinuity in FIG. 3(a), and background removal identifies the shape and intensity of the spectral data 420. You will find it easier to do so. This improves analytical accuracy in terms of element identification and quantitative performance, and enables spectrum evaluation, which has been difficult until now. Note that this embodiment is not limited to the LIBS measurement of this embodiment, and is applicable to all measurement devices that use the spectrometer 400.

Furthermore, by applying not only asymmetric weight penalty least squares smoothing but also pattern recognition based removal from the baseline 430 under the condition of no sample, it is possible to improve the performance specific to various analytical devices and reduce sample dependence. It is applicable not only to analysis methods using a laser light source such as Raman spectroscopy and LA-ICP-MS, but also to all measurement devices 100 that use the spectrometer 400.

As an example of removal by pattern recognition, there is a method in which a spectrum obtained when only the measurement environment is measured without a sample (dry shot) is pattern recognized and removed as a baseline. Generally, this is done by simply subtracting the condition where there is no sample, but by making full use of pattern recognition, it is possible to make corrections that match the baseline that moves up and down with each measurement.

FIGS. 5 and 6 are diagrams illustrating factor decomposition of spectral data in the compound identification method. In the NMF execution L3 step, the controller 220 performs non-negative matrix factorization (NMF) on the corrected spectral data 420.

NMF uses machine learning to decompose the entire overlapped emission spectrum data into a specified number of factors, and uses a decision tree to determine which factors to focus on for the sample data group expressed as a pair of factor score and substance name. An algorithm is used to find out what threshold value can be used to accurately classify substances, and by classifying them into related factors, elements and compounds are classified from the unknown sample 110b.

By converting the spectrum data 420 of the known substance into a low-dimensional factor score vector using a dimension reduction method in advance, and training an algorithm that classifies elements and compounds using a decision tree, it is possible to Estimate the components of.

In the learning stage, the spectral data 420 (K-dimensional vector group) shown in FIG. 5(a) is grouped together, divided into M factors 440 using NMF, and reduced in dimension so that it can be easily handled by a decision tree. NMF is applicable regardless of the type of data as long as the spectrum data can be expressed as a matrix of 0 and positive values.

As shown in FIG. 5(b), N pieces of measurement data Y (y ₁ to y _N ) are approximately decomposed into factors H and factor scores U. The measurement data y _i is vector data (intensity) of K dimensions (for example, K=11562 types of wavelengths), and is represented by a K×N (K rows and N columns) matrix. Factor H is a factor common to K-dimensional vector data expressed as M (for example, M = 3 factors) basis vectors, and is a K×M (K rows and M columns) matrix. expressed. The factor score U is a coupling coefficient regarding N measurement data Y and M factors H, and is represented by an M×N (M rows and N columns) matrix. Note that M is a hyperparameter.

In the case of the standard sample 110a, a matrix consisting of the number of wavelength samples (wavelength dimension) of the spectrum data 420 and the number of samples is factorized using non-negative values, and the obtained factors (for example, the first to third factors) are each used as the standard. The factor 440 of the sample 110a is assumed to be a factor 440, and a coefficient representing the degree of influence of each factor 440 in the standard sample 110a is derived as a factor score 450. The factor score 450 indicates which factor 440 greatly influences the measurement data.

As an example using LIBS, four types of compounds, mainly borides, are set, and 37 (N = 37 x 4) spectral data 420 are extracted from each as the standard sample 110a. The number of dimensions was k=11562). The results of reducing the dimension using NMF (three factors 440 with different feature amounts and a factor score 450 as reference parameters) are shown. As shown in FIG. 6, after considering the influence of the factor score 450 on each factor 440, the measured spectrum data (FIG. 5(a)) can be expressed by a linear sum of three factors.

Note that the factor 440 may standardize the spectrum data 420 using an average value and a standard deviation in order to equalize the intensity scale for each wavelength. If standardization is not desired, this step may be omitted. Note that there are fragility of samples due to laser irradiation and self-shielding peculiar to samples during radiation measurement, so standardization can alleviate these problems.

7 to 9 are diagrams illustrating classification using a decision tree in the compound identification method. In the decision tree rule generation L4 step, the control device 220 creates a decision tree 500 using the factor score 450 of the standard sample 110a as a feature, and creates a decision tree 500 for classification based on the factor score 450 when applying the factor 440 of the standard sample 110a. The score threshold 460 and the order in which factors are applied are set to determine which factor 440 of the standard sample 110a is applied for classification using the decision tree 500.

The decision tree 500 is a classification method that determines "True" or "False" based on reference parameters obtained by NMF, and for learning data, the focused factor 440 (reference parameter) and the factor score used for the determination are This is a search algorithm that uses a score threshold of 460 of 450 to correctly classify substances.

The combination of NMF and decision tree 500 in the learning stage can be used not only for analysis methods that have spectral data 420 such as dynamic light scattering and inductively coupled plasma mass spectrometry (ICP-MS), but also for "0" and "positive values". If a matrix representation can be written, it can be applied regardless of the type of data. It can also be expected to be applied to image data such as TEM (transmission electron microscope), SEM (scanning electron microscope), and accompanying EDX (energy dispersive X-ray analysis).

As shown in FIG. 7(a), in order to make a conditional branch in the decision tree 500, a cost function J that calculates impurity for the factor score 450 of the standard sample 110a is used, and a score that minimizes the cost function J is used. A threshold value t _k (1≦k≦m) is derived. The cost function J is calculated using Gini impurity, etc., where k is the factor number, m is the number of factors 440, m _left is the number of factors (left subset) divided into one, and m _right is divided into the other. The number of factors (right subset), G _left is the impurity of the left subset, and G _right is the impurity of the right subset.

Note that if only one type of compound exists in the subset, it is a pure subset, and if multiple types of compounds are mixed, it is an impure subset. At this time, the impurity of the subset is expressed by Gini impurity or cross entropy based on the ratio of each compound contained in the subset.

As shown in FIG. 7(b), when the compound type (atmosphere) is Ar and _N2 , a score threshold 460 (for example, 114.155) of the factor score 450 is derived and determined for the factor 440 on the horizontal axis side. In the tree 500, if the factor score 450 is less than the score threshold 460, the factor 440 is classified into the left subset (estimated as _N2 ), and if the factor score 450 is greater than the score threshold 460, the factor 440 is classified into the right subset (estimated as Ar). being classified.

Figure 8(a) shows that the boride spectrum (four compound types: CrB, FeB, NiB, and ZrB) measured by LIBS is decomposed into factors of 440, and the feature values (factor score of 450) are calculated. An example of extracting and creating a decision tree 500 will be shown. It is assumed that the order of application of the factor 440 is the second factor, the third factor, and the first factor.

In the node N1 of the decision tree 500, the score threshold 460 of the factor score 450 in the second factor is derived as 108.615, and as shown in FIG. 8(b), those larger than the score threshold 460 are compounds containing Cr and B. , and the others are classified as node N2.

In the node N2 of the decision tree 500, the score threshold 460 of the factor score 450 in the third factor is derived as 65.837, and as shown in FIG. 9(a), those larger than the score threshold 460 are compounds containing Ni and B. , and the others are classified as node N3.

At node N3 of the decision tree 500, the score threshold 460 of the factor score 450 in the first factor is derived as 117.886, and as shown in FIG. 9(b), the score threshold 460 is greater than the compound containing Zr and B. Others are classified as compounds containing Fe and B.

Chromium (Cr) is classified earlier in decision tree 500 than zirconium (Zr) and iron (Fe), which have a relatively rich spectrum, and is classified more accurately in decision tree 500, contrary to human expectations. . This is thought to be due to the fact that Zr and Fe contain many peaks that are affected by interference (if the classification is simply based on signal peaks without interference, as in the conventional method, the correct answer rate will decrease) .

In the control device 220, the factor 440 and the factor score 450 obtained from the standard sample 110a, as well as the score threshold of the factor score 450 and the factor application order of the factor 440 as rules for the decision tree 500 are stored in a storage device such as a database, and the compound Create an identification index. In the learning stage, we understand what kind of standard sample 110a the data is from and how the factor score 450 is for each factor 440, and in the operational stage, we use it as a compound identification index when classifying the unknown sample 110b. Make use of it.

FIG. 10 is a flowchart showing the flow of the operational steps of the compound identification method. The operation stage includes steps of unknown sample measurement S1, spectral data correction S2, factorization processing S3, and decision tree classification processing S4. Each factor 440 and its factor score 450 of the standard sample 110a are prepared in advance as a compound identification index, and the unknown sample 110b is divided into classes by sorting by the factor score 450 of each factor 440 of the unknown sample 110b.

In the operation stage, a factor score 450 for the unknown sample 110b is calculated using the factor 440 of the standard sample 110a for the spectrum data 420 of the unknown sample 110b, and a decision tree 500 is used to calculate the unknown sample 110b according to the order of application of the factors 440. The class of the unknown sample 110b is estimated by dividing the factor score 450 by the score threshold 460, thereby identifying the compound species contained in the unknown sample 110b.

In the unknown sample measurement S1 step, the unknown sample 110b is irradiated with a laser in the measurement device 100, and the spectrum data 420 sent from the detector 410 is input to the control device 220.

In the spectral data correction step S2, the control device 220 performs baseline correction on the spectral data 420 of the unknown sample 110b as well as in the spectral data acquisition L2 step.

In step S3 of the factorization process, the controller 220 performs factorization using NMF on the corrected spectrum data 420 of the unknown sample 110b using the factor 440 of the standard sample 110a, and calculates the factor score 450 of the unknown sample 110b. calculate. Note that standardization processing may be performed as necessary.

In step S4 of the decision tree classification process, the control device 220 classifies the factor scores 450 of the unknown sample 110b into the factors 440 based on the rules (learned algorithm) of the decision tree 500. In the decision tree 500, the factor score 450 of the unknown sample 110b is divided into classes according to the score threshold 460 according to the order of application of the factors 440, and the class to which the unknown sample belongs is estimated, thereby determining the compound species contained in the unknown sample 110b. identify

By multiplying the decision tree 500, the factor scores 450 and compound types are associated. Common factors 440 (first to mth factors) are derived from all the spectral data 420 of the standard sample group, and each spectral data 420 is converted into a factor score 450 through each factor. Then, the compounds are classified into compound types through the decision tree 500.

In fact, when we measured a eutectic compound of stainless steel (SUS) and boron carbide (B ₄ C) as unknown sample 110b, we obtained a 100% correct answer rate with 4 types of test data x 8 pieces. Ta.

FIG. 11 is a diagram showing the results of classification by the compound identification method. The factor score 450 may be mapped to estimate the components and composition ratio of the unknown sample 110b. As shown in FIG. 11(a), it is also possible to estimate whether or not the simulated sample contains elements contained in a certain compound. Furthermore, as shown in FIG. 11(b), it is also possible to estimate the composition ratio to determine whether a large amount of CrB ₂ is contained.

By controlling the two-axis coordinates of the X-axis and Y-axis and measuring the distance of the sample placed between the stage and the displacement meter using a displacement meter, information on the height Z-axis is obtained and the shape of the sample is grasped. Coordinate control is easy using an electric stage. LIBS measurement results are obtained by pinpoint LIBS measurement of the position of the sample with the XYZ axes known. A range of emission spectra of characteristic elements is selected from the LIBS measurement results, and area values and heights are calculated. In order to display the mapping information of two-dimensional elements, the selected element may be mapped by setting the X-coordinate and Y-coordinate information on two axes and showing the intensity of the selected element's emission spectrum in color.

According to the present invention, metal elements and compounds contained in an unknown sample can be classified and estimated. By decomposing the spectrum data of known substances into multiple factors in advance and learning how to classify elements and compounds using a decision tree, the system can be used to classify the components of unknown samples.

Furthermore, by evaluating errors inherent in the sample and correcting them in advance, metal elements and compounds can be classified and estimated with high accuracy. Note that this can be corrected afterward, and errors including white noise can be removed.

In the present invention, adoption or rejection of an atomic emission spectrum to be measured is determined based on reference parameters that characterize the plasma state (time, space, etc.). Machine learning is also used to set reference parameters. In the conventional method, the spectrum of the measurement target is searched as a peak with a certain width, whereas in the present invention, big data (all spectral components obtained from various analytical devices) that is difficult for humans to handle is searched using NMF. Build reference parameters using machine learning. As a result, the amount of data used for evaluation increases to 103 by processing all spectral components instead of processing based on signal peaks, which was done in the conventional method.As a result, the amount of data used for evaluation increases to 10 ³ It has become possible to perform evaluations that include shapes, etc., and there is also the prospect of establishing high analytical accuracy.

Depending on the type of baseline correction used in preprocessing, it can be applied to each spectrum by using a function such as the asymmetric weight penalty least squares smoothing method or machine learning such as pattern recognition. If you have a database from a public research institution or an environment where you can measure known samples as learning data, you can improve the performance of various analyzers you already own by retrofitting them (laser light sources such as Raman scattering and LA-ICP-MS). It is possible to reduce dependence not only on the analysis method used, but also on all measuring instruments that use a spectrometer) and on the sample.

All you need to do as learning data is to measure databases from public research institutes (spectral and radiation spectrum information) and standard samples whose material compositions are known, and you can easily use it without modifying the hardware by implementing it in the analysis software of the device. It is possible to improve the performance specific to various analytical devices and reduce sample dependence. In addition to spectroscopic measurements using light sources such as lasers, X-rays, and neutrons as probes, electron microscopes (transmission electron microscopes: TEM-EDX, scanning electron microscopes: A wide variety of applications are expected, including SEM-EDX), electron probe microanalyzer (EPMA), etc., to convert spectral data into images as surface information.

Although the embodiments of the present invention have been described above, the present invention is not limited thereto. For example, NMF is a method used in machine learning that divides orchestral sounds into factors such as violin and viola, and can create and classify characteristic sounds (reference parameters) of violin and viola. In this way, the present invention can be applied to various types of big data.

Furthermore, material distribution and hardness distribution within an unknown sample may be created, or may be applied to multivariate analysis.

100: Measuring device 110: Sample 110a: Standard sample 110b: Unknown sample 200: Chamber 210: Vacuum pump 220: Control device 230: Sequencer 240: Stage 300: Irradiation device 310: Irradiation light 320: Emitted light 400: Spectrometer 410: Detector 420: Spectral data 430: Baseline 440: Factor 450: Factor score 460: Score threshold 500: Decision tree

Claims (6)

A score threshold for classifying the compound type by the factor score in a decision tree by decomposing the spectral data of a standard sample of the assumed compound type into factors to derive a factor and a factor score representing the degree of influence of the factor. The order in which factors are applied to apply the factor scores to the decision tree is learned and set in advance,
A factor score of the unknown sample is calculated using the factors of the standard sample for the spectral data of the unknown sample, and the factor scores of the unknown sample are classified by the score threshold using the decision tree in accordance with the order of application of the factors. By identifying the compound species contained in the unknown sample,
A compound identification method characterized by: The spectral data is corrected to a baseline approximated by an asymmetric least squares method with a penalty term so that the error is evaluated more strictly when the measured value is below the estimated value than when it is above the estimated value.
The compound identification method according to claim 1, characterized in that: The factor decomposition is performed by decomposing the K-by-N matrix consisting of the wavelength dimension K and the number of samples N of the spectral data using specified M non-negative basis vectors, and calculating the degree of influence of the factor as described above. derived as a factor score,
The compound identification method according to claim 1 or 2, characterized in that: The decision tree derives factors and score thresholds such that a cost function using impurity has a minimum value when classifying the compound type.
The compound identification method according to any one of claims 1 to 3, characterized in that: carrying out the compound identification method according to any one of claims 1 to 4,
A compound identification device characterized by: an irradiation device that irradiates the sample with a laser;
a spectrometer that decomposes light emitted from the sample irradiated with the laser;
a detector for acquiring spectral data generated by the spectrometer;
identifying compounds contained in the sample using spectral data input from the detector;
The compound identification device according to claim 5, characterized in that: