JP7780757B2 - Compound identification method and compound identification device - Google Patents

Description

Article 30, Paragraph 2 of the Patent Act will be applied to the 7th Next Generation Initiative Decommissioning Technology Conference (NDEC-7) 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.

Decommissioning nuclear power plants requires measures to deal with contaminated water, remove fuel from spent fuel pools, remove fuel debris, and dispose of waste. In reactors with melted cores, molten fuel, control rods, and other materials are thought to have solidified into fuel debris. In reactors that use boron carbide ( _B4C ) as a control material, borides, which are approximately twice as hard as oxides, are produced. Therefore, it is efficient to remove fuel debris while distinguishing between metals, oxides, and borides. It has also been confirmed that there is a correlation between the boron concentration and hardness of borides. To investigate fuel debris removal methods, it is necessary to identify the material and hardness of unknown samples. Furthermore, the creation of material and hardness distributions within unknown samples is also being considered.

Methods for qualitative and quantitative analysis of elements in samples include inductively coupled plasma atomic emission spectrometry (ICP-AES) and laser ablation inductively coupled plasma mass spectrometry (LA-ICP-MS). When a sample is externally energized, the elements contained in the sample emit light with wavelengths specific to the element, and elemental analysis is performed by measuring the wavelength and amount of emitted light. ICP-AES can sometimes be inaccurate due to interference. Patent Document 1 also discloses an invention for a foreign matter analysis device that can accurately and easily analyze foreign matter contained in a sample.

Laser-induced breakdown spectroscopy (LIBS) is another method for elemental analysis, regardless of whether the material is in a solid, liquid, or gaseous state. LIBS is an atomic emission spectroscopy method that uses a high-energy pulsed laser as an excitation source. It enables analysis in a short time without complex pretreatment processes and is capable of simultaneously measuring elements ranging from light elements (boron and oxygen) to heavy elements (metals). Therefore, it is expected to be applied to obtaining elemental information on high-level waste, hardness measurement, and criticality control based on boron distribution information. Patent Document 2 also discloses an invention for a method for calculating the composition of oxygen and boron in metal-based materials and the hardness of the material.

In LIBS, when a high-peak-power laser with a pulse width of less than a nanosecond is irradiated onto the object to be measured, a plasma of tens of thousands of degrees Celsius is instantly formed. At this time, the object to be measured is atomized and excited by the plasma, and as the plasma disappears, the excited atoms relax to their ground state, emitting element-specific fluorescent light. This is measured to identify and quantify the element being measured. Fluctuations in the laser output and the plasma generated by the elements that make up the object to be measured cause minute fluctuations in time and space, which ultimately result in fluctuations in the measured values.

In the case of LIBS, background due to the plasma state (time, space, etc.) and, in the case of energy discrimination processing, scattering components caused by the interaction between radiation and matter have a significant impact on measurement values. In spectroscopic measurements that use radiation such as lasers, X-rays, and neutrons as probes, although there are various factors and their influence varies in magnitude, a function to cancel or correct these is desired to achieve highly accurate measurements.

Patent No. 6507757 Japanese Patent Application Laid-Open No. 2020-067337

However, when analyzing unknown solid samples, such as alloys or melts containing many elements, spectral data with many emission wavelengths is obtained, making it difficult to identify the elements or compounds. For example, the emission lines of iron (Fe) and zirconium (Zr) can overlap so much that quantitative analysis can be hindered.

In addition, when measurements are made using spectrometers that use not only lasers but also X-rays, neutrons, etc., the dark current is measured with the spectrometer closed to eliminate its effects. In order to eliminate the effects of the measurement environment (air environment, inert environment, etc.), some methods use blank data (measurements without a sample) to eliminate the effects. However, these effects cannot be applied in cases where they are not added to the device's functions or where similar samples do not exist.

In addition, with laser-based optical emission spectrometry (LIBS and LA-ICP-MS) and neutrons, which have self-shielding, it is necessary to evaluate sample-specific errors, but with neutrons and plasma emission, it is often impossible to control the irradiation energy.

In particular, with LIBS measurements, the spectral baseline (a continuous line formed by connecting points thought to be the background) can fluctuate depending on the ambient gas. For example, argon (Ar) atmospheres have high luminescence sensitivity, which can result in white noise, which broadens the baseline over a wide wavelength range. Furthermore, multiple detectors are used to detect light with wavelengths from 190 nm to 900 nm, but because the detectable wavelength range varies depending on the characteristics of the optical system, deviations in the baseline occur at wavelengths where the detectors switch.

As such, there is a demand for reducing background noise components (variation errors such as dark current in spectrometers) in spectral analysis such as spectroscopic analysis and radiation analysis, optimizing probe conditions for light and radiation on the object being measured in various analyses (mitigating damage caused by the probe to the object being measured), and identifying various compounds (including metals, ceramics, etc.).

The present invention therefore aims to provide a method for classifying and estimating metal elements and compounds contained in unknown samples. Another aim is to evaluate sample-specific errors and correct them in advance when classifying and estimating metal elements and compounds.

To solve the above problems, the compound identification method of the present invention factorizes the spectral data of a standard sample of a suspected compound species to derive factors and factor scores representing the influence of the factors; learns and sets in advance a score threshold for classifying the compound species using the factor scores in a decision tree and a factor application order for applying the factor scores to the decision tree; calculates factor scores for an unknown sample using the factors of the standard sample for the spectral data of the unknown sample; and identifies the compound species contained in the unknown sample by classifying the factor scores of the unknown sample using the score threshold in the decision tree in accordance with the factor application order.

In the compound identification method, the spectral data is corrected to a baseline approximated by a penalized asymmetric least squares method so that errors are evaluated more severely when the measured value is below the estimated value than when it is above the estimated value.

In the compound identification method, the factorization is characterized in that a matrix Y with K rows and N columns, consisting of the number of wavelength samples (wavelength dimension) K and the number of samples N in the spectral data, is decomposed using M specified non-negative basis vectors, and the influence of the obtained factors is derive as the factor score.

In the compound identification method, the decision tree is characterized in that, when classifying the compound types, factors and score thresholds are derived that minimize the cost function using impurity.

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

The compound identification device of the present invention is characterized by having an irradiation device that irradiates the sample with a laser, a spectrometer that resolves the light emitted from the sample irradiated with the laser, and multiple detectors that acquire spectral data generated by the spectrometer by dividing it into wavelengths within a predetermined range, and identifying compounds contained in the sample using the spectral data input from the detectors.

This invention makes it possible to classify and estimate the metal elements and compounds contained in unknown samples. By decomposing the spectral data of known substances into multiple factors in advance and training the system to classify elements and compounds using a decision tree, this can be used to classify the components of unknown samples.

In addition, by evaluating sample-specific errors and correcting them in advance, metal elements and compounds can be classified and estimated with high accuracy. Corrections can also be made retrospectively, eliminating errors, including white noise.

1 is a schematic view of a measuring device in a compound identification method according to the present invention. 1 is a flowchart showing the flow of the preprocessing step and the learning step of the compound identification method of the present invention. FIG. 1 is a diagram illustrating the correction of spectral data in the compound identification method of the present invention. FIG. 1 is a diagram illustrating the correction of spectral data in the compound identification method of the present invention. FIG. 1 is a diagram illustrating factorization of spectral data in the compound identification method of the present invention. FIG. 1 is a diagram illustrating factorization of spectral data in the compound identification method of the present invention. FIG. 1 is a diagram illustrating classification by a decision tree in the compound identification method of the present invention. FIG. 1 is a diagram illustrating classification by a decision tree in the compound identification method of the present invention. FIG. 1 is a diagram illustrating classification by a decision tree in the compound identification method of the present invention. FIG. 1 is a flowchart showing the flow of operational steps of a compound identification method according to the present invention. FIG. 1 shows the results of classification by 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 with the same functions will be assigned the same reference numerals, and repeated explanations may be omitted.

This invention relates to an algorithm for two processes in spectroscopic measurement systems that use radiation such as visible light, infrared light, X-rays, gamma rays, and neutrons: reducing background fluctuations caused by the stability of the measurement system, and identifying the spectrum of the measurement target. This algorithm consists of a machine learning method that uses measurement data (including predicted data) as training data, and a data processing system that results from this method. Its implementation is expected to improve measurement accuracy and robustness. Furthermore, because it can be implemented as software processing, it is highly adaptable to analysis programs installed in existing spectroscopic analysis equipment.

To improve the analytical accuracy of element identification and quantification using LIBS (Laser-Induced Breakdown Spectroscopy), parameters that characterize the plasma state (time, space, etc.) (reference parameters) are measured simultaneously in addition to the elemental emission spectrum, and the correlation between the two is used in the correction process. Specifically, reference parameters that are strongly correlated with the intensity and shape of the atomic emission spectrum being measured are set. Traditionally, these were set by highly skilled and experienced technicians who understood the fluctuating components and characteristics. To achieve high analytical accuracy, evaluation that takes into account the spectral fluctuation mechanisms is necessary.

Machine learning is used to set the reference parameters. While conventional methods search for peaks with a certain width in the spectrum of the object to be measured, this invention applies a machine learning method (NMF: non-negative matrix factorization) to big data that is difficult for humans to handle (all spectral components consisting of sets of wavelength-count values (light intensity values) obtained from various analytical devices, etc.) to construct the reference parameters. By using all spectral components rather than signal peaks, it is possible to evaluate spectral intensity, shape, etc. when identifying unknown samples.

The present invention is inspired by methods for extracting specific sounds, such as violins and violas, from an orchestra, or methods for extracting a specific individual from a crowd, and has been applied to a measurement device. The algorithm of the present invention involves processing the spectral data of a standard sample in the preprocessing phase, training it using NMF and decision trees, and then applying it to unknown samples in the operation phase.

Figure 1 is an overview of the measurement device used in the compound identification method of the present invention. The measurement device 100 is equipped with a chamber 200, an irradiation device 300, a spectrometer 400, a detector 410, and a control device 220, and performs LIBS measurements.

The chamber 200 is a vacuum vessel or the like, and the sample 110 is placed inside after reducing the pressure using a vacuum pump 210. During the learning phase, known samples such as the standard sample 110a are primarily used as the sample 110, while during the operational phase, unknown samples 110b to be measured and analyzed are used. Standard samples 110a are those with known compositions, such as metal elements or compounds, while unknown samples 110b are those with unknown compositions, 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. The irradiation light 310 is an Nd:YAG laser or the like with a wavelength of 1064 nm, which uses a crystal of YAG (yttrium, aluminum, garnet) doped with Nd (neodymium) as a medium.

By irradiating the surface of the sample 110 with irradiation light 310, the surface is converted into plasma, and the emitted light 320 is detected as it returns to its ground state. The emitted light 320 is collimated using a collimator or similar device before entering the spectrometer 400. The spectrometer 400 is a prism or diffraction grating, which separates the light into wavelengths and acquires spectral data 420 representing the intensity distribution by arranging multiple detectors 410 side by side, dividing the data into wavelengths within a predetermined range, and sends the data to the control device 220.

The control device 220 is a computer or the like, which controls each device and classifies and analyzes the sample 110 based on the acquired spectral data 420. If necessary, the system may also be equipped with a sequencer 230, an electrically operated stage 240, a cooling water circulation system, etc.

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

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

In the preprocessing and learning stages, the expected compound species are set and a standard sample 110a is prepared, the acquired spectral data 420 of the standard sample 110a is factorized to derive the factors and factor scores of the standard sample 110a, and the score threshold for classifying the compound species by factor scores in the decision tree and the order in which the factors are applied to the decision tree are learned and set in advance as decision tree rules.

The pre-processing stage consists of standard sample preparation L1 and spectral data acquisition L2 in Figure 2. In the standard sample preparation L1 step, compound types such as metal elements and compounds that are expected to be components of the unknown sample 110b to be measured are set in advance, and these metal elements and compounds are prepared as standard sample 110a. Standard sample 110a is data prepared in advance to determine the influence and extent of each factor when identifying the compound types contained in unknown sample 110b.

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

Next, the steps involved in the pretreatment stage described above will be explained with reference to Figures 3 and 4, which show examples obtained by performing LIBS measurements. Figures 3 and 4 explain the correction of spectral data in the compound identification method.

Figure 3(a) shows a spectrum obtained by dividing the LIBS measurement system into multiple measurement wavelength ranges and using a highly sensitive spectrometer for each range to broaden the bandwidth (190 nm to 900 nm). As shown in Figure 3(a), it can be seen that there is a shift in the baseline 430 at the boundary of the wavelength range of the detector 410. This shift causes discontinuities in the baseline 430 due to the sensitivity and dark current of the detector 410.

In wavelength ranges where discontinuities occur, the measurement results contain many uncertain factors, making the results unreliable. Furthermore, if the target has a spectrum in the discontinuous area, it is not possible to calculate the accurate area, and in some cases it may not be possible to distinguish the spectrum.

Figure 3(b) is an enlarged spectrum of a portion of the spectral data 420 in Figure 3(a). As shown in Figure 3(b), unless optimal smoothing is applied for the detector resolution, the baseline will be adjusted more than necessary.

Therefore, the inventors devised a method for correcting the deviation (discontinuity) of the baseline 430 at the boundary of the wavelength ranges of the multiple detectors 410 by asymmetric weighted penalty least squares smoothing (asymmetric least squares method with a penalty term) shown in Fig. 3( _c ). The asymmetric least squares method with a penalty term derives a preferable baseline 430 by minimizing an evaluation function that not only minimizes the difference between the measured value and the estimated value but also minimizes the smoothness of the baseline 430, determining the desirability (goodness) of the estimated value z _i of the baseline 430 relative to the measured value y i.

From the viewpoint of evaluation, first, from the viewpoint of the squared error Σ _i (y _i - _{z i} ) ² , an asymmetric evaluation is performed in which the error is very strict when y _i ≦ z _i and very lenient when y _i > z _i . Furthermore, from the viewpoint of the fluctuation of z _i , a strict evaluation is performed against changes in the fluctuation. w _i is an asymmetric weight, and if y _i > z _i , it becomes p, and if y _i ≦ z _i, it becomes (1-p). p (p ∈ [0, 1]: p is a very small value) plays a role in changing w _i depending on whether the measured value exceeds the estimated value. μ is a regularization parameter, and the fluctuation (non-smoothness) of the estimated value Σ _i (Δ ² z _i ) ² is the penalty term. Note that Δ ² represents a second-order differential operator, and w _i and μ are hyperparameters empirically determined by the designer. The optimal solution can be found by changing the parameters.

As a result, the spectral data 420 is corrected to a baseline 430 approximated using a penalized asymmetric least squares method so that errors are assessed more severely when the measured value is below the estimated value than when the measured value is above the estimated value.

Figure 4 shows the spectrum obtained by correcting the spectrum data 420 in Figure 3 using the correction shown in Figure 3(c). Note that Figure 4(a) shows the spectrum obtained by correcting the spectrum data 420 in Figure 3(a), and Figure 4(b) shows an enlarged view of a portion of the spectrum data 420 in Figure 4(a).

As shown in Figure 4(a), the smoothed baseline 430 after correction eliminates the discontinuity in the spectrum in Figure 3(a), and background removal makes it easier to identify the shape and intensity of the spectral data 420. This improves analytical accuracy in terms of element identification and quantification, making spectral evaluation, which was previously difficult, possible. Note that this method is not limited to the LIBS measurement in this example, but can be applied to all measurement devices that use the spectrometer 400.

In addition to asymmetric weighted penalty least-squares smoothing, the system also uses pattern recognition to remove baseline 430 in the absence of a sample, improving the performance specific to various analytical devices and reducing sample dependency. This system is applicable not only to analytical methods using laser light sources such as Raman spectroscopy and LA-ICP-MS, but also to all measurement devices 100 that use a spectrometer 400.

An example of removal using pattern recognition is to perform pattern recognition on the spectrum obtained when measuring only the measurement environment under conditions where no sample is present (blank shot), and then remove it at the baseline. This is generally done by simply subtracting the conditions under which no sample is present, but by making full use of pattern recognition, it is possible to make corrections to match the baseline, which fluctuates with each measurement.

Figures 5 and 6 are diagrams explaining the factorization of spectral data in the compound identification method. In the NMF execution L3 step, the control device 220 performs nonnegative matrix factorization (NMF) on the corrected spectral data 420.

NMF uses machine learning to decompose all overlapping emission spectrum data into a specified number of factors, and then uses an algorithm to determine which factors to focus on and what thresholds to use in a decision tree to correctly classify substances for a group of sample data represented by pairs of factor scores and substance names. NMF then classifies elements and compounds from the unknown sample 110b by dividing them into related factors.

The spectral data 420 of known substances is converted into a low-dimensional factor score vector using a dimension reduction method in advance, and an algorithm that classifies elements, compounds, etc. using a decision tree is trained to estimate the components in the unknown sample 110b.

In the learning stage, the spectral data 420 (a group of K-dimensional vectors) shown in Figure 5(a) is grouped together and divided into M factors 440 using NMF, reducing the dimension to make it easier to handle with a decision tree. NMF can be applied to any type of data as long as the spectral 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 = 11,562 types of wavelengths) and is represented by a K × N (K rows and N columns) matrix. The factor H is a representation of some factor common to the K-dimensional vector data as M basis vectors (for example, M = 3 factors) and is represented by a K × M (K rows and M columns) matrix. The factor score U is a coupling coefficient for the N pieces of measurement data Y and the M factors H and is represented by an M × N (M rows and N columns) matrix. M is a hyperparameter.

For the standard sample 110a, a matrix consisting of the number of wavelength samples (wavelength dimension) and the number of samples in the spectral data 420 is factorized using non-negative values, and the resulting factors (e.g., factors 1 to 3) are each set as factors 440 for the standard sample 110a, and coefficients representing the degree of influence of each factor 440 on the standard sample 110a are derived as factor scores 450. The factor scores 450 indicate which factors 440 have the greatest influence on the measurement data.

As an example using LIBS, four types of compounds, primarily borides, were selected, and 37 pieces (N = 37 x 4) of spectral data 420 were extracted from each of the standard samples 110a, resulting in high-dimensional samples (number of wavelength dimensions for each spectrum, k = 11,562). The results of reducing the dimensions using NMF (using three factors 440 with different feature values and factor scores 450 as reference parameters) are shown below. As shown in Figure 6, after taking into account the influence of the factor scores 450 on each factor 440, the measured spectral data (Figure 5(a)) can be expressed as a linear sum of the three factors.

Factor 440 may standardize the spectral data 420 using the mean and standard deviation to align the intensity scale for each wavelength. If standardization is not required, this step may be omitted. Note that sample damage due to laser irradiation and sample-specific self-shielding during radiation measurement can be alleviated by standardization.

Figures 7 to 9 are diagrams explaining 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 sets a score threshold 460 for classification using the factor score 450 when applying the factors 440 of the standard sample 110a, as well as the factor application order, which determines which factor 440 of the standard sample 110a to apply first to classify 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 is a search algorithm that correctly classifies substances using the training data, using the focused factor 440 (reference parameter) and the score threshold 460 of the factor score 450 used for that determination.

The combination of NMF and decision tree 500 in the learning stage can be applied to analytical methods that use spectral data 420, such as dynamic light scattering and inductively coupled plasma mass spectrometry (ICP-MS), as well as any type of data, as long as it can be expressed in matrix form with "0" and "positive values." It is also expected to be applicable to image data from TEM (transmission electron microscope), SEM (scanning electron microscope), and associated EDX (energy dispersive X-ray analysis) data.

7A, in order to perform conditional branching in the decision tree 500, a cost function J that calculates the impurity of the factor score 450 of the standard sample 110a is used, and a score threshold _tk (1≦k≦m) that minimizes the cost function J is derived. Note that the cost function J is calculated using Gini impurity or the like, where k is the factor number, m is the number of factors 440, _mleft is the number of factors to be divided into one group (left subset), _mright is the number of factors to be divided into the other group (right subset), _Gleft is the impurity of the left subset, and _Gright is the impurity of the right subset.

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

As shown in FIG. 7B , when the compound species (atmosphere) are Ar and _N2 , a score threshold 460 (e.g., 114.155) of the factor score 450 is derived for the factor 440 on the horizontal axis side, and in the decision tree 500, if the factor score 450 is smaller 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).

Figure 8(a) shows an example in which boride spectra (four compound types: CrB, FeB, NiB, and ZrB) measured with LIBS are decomposed into factors 440, feature quantities (factor scores 450) are extracted, and a decision tree 500 is created. The order of application of factors 440 is assumed to be the second factor, then the third factor, and then the first factor.

At node N1 of the decision tree 500, the score threshold 460 for the factor score 450 in the second factor is derived as 108.615, and as shown in Figure 8(b), compounds with a score greater than the score threshold 460 are classified as compounds containing Cr and B, and the rest are classified at node N2.

At node N2 of the decision tree 500, the score threshold 460 for the factor score 450 in the third factor is derived as 65.837, and as shown in Figure 9(a), compounds with scores greater than the score threshold 460 are classified as compounds containing Ni and B, and the rest are classified at node N3.

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

Chromium (Cr) is classified at the beginning of decision tree 500, rather than zirconium (Zr) or iron (Fe), which have relatively large spectra, and decision tree 500 shows a more accurate classification than human predictions. This is thought to be due to the fact that Zr and Fe contain many peaks that are affected by interference (classification based simply on signal peaks without interference, as in conventional methods, results in a lower accuracy rate).

The control device 220 stores the factors 440 and factor scores 450 obtained from the standard sample 110a, as well as the score threshold for the factor score 450 and the order in which the factors 440 are applied as rules for the decision tree 500, in a storage device such as a database, and creates a compound identification index. During the learning phase, it is determined what type of standard sample 110a the data is from and how the factor scores 450 are for each factor 440. During the operation phase, this is used as a compound identification index when classifying unknown samples 110b.

Figure 10 is a flowchart showing the flow of the operational stages of the compound identification method. The operational stages include unknown sample measurement S1, spectral data correction S2, factor decomposition 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 compound identification indices, and the unknown sample 110b is classified by sorting it using the factor score 450 for each factor 440 of the unknown sample 110b.

In the operational phase, the factor score 450 for the unknown sample 110b is calculated for the spectral data 420 of the unknown sample 110b using the factors 440 of the standard sample 110a, and the factor score 450 for the unknown sample 110b is divided by the score threshold 460 in the decision tree 500 according to the order in which the factors 440 are applied, and the class of the unknown sample 110b is estimated, thereby identifying the compound species contained in the unknown sample 110b.

In the unknown sample measurement S1 step, the measurement device 100 irradiates the unknown sample 110b with a laser, and the spectral data 420 sent from the detector 410 is input into the control device 220.

In the spectral data correction S2 step, the control device 220 also performs baseline correction on the spectral data 420 of the unknown sample 110b, similar to the spectral data acquisition L2 step.

In the factorization process S3 step, the control device 220 performs NMF factorization on the corrected spectral data 420 of the unknown sample 110b using the factors 440 of the standard sample 110a, and calculates the factor score 450 of the unknown sample 110b. Note that standardization processing may also be performed if necessary.

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

By applying the decision tree 500, the factor scores 450 are associated with compound types. Common factors 440 (factors 1 to m) 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. The compound types are then classified through the decision tree 500.

In fact, when a eutectic compound of stainless steel (SUS) and boron carbide (B ₄ C) was measured as the unknown sample 110b, the accuracy rate was 100% for four types of test data x eight pieces.

FIG. 11 shows the results of classification using the compound identification method. The factor scores 450 may be mapped to estimate the components and composition ratios of the unknown sample 110b. As shown in FIG. 11(a), it is also possible to estimate whether a 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 sample contains a large amount of _CrB2 .

By controlling the coordinates on two axes (X and Y) and measuring the distance between the stage and the sample using a displacement meter, height Z-axis information can be obtained and the sample shape determined. Coordinate control is easy using an electric stage. LIBS measurement results can be obtained by pinpointing the position of the sample with the X-Y-Z axes known and performing LIBS measurements. From the LIBS measurement results, the range of the emission spectrum of a characteristic element can be selected and the area value and height calculated. To display element mapping information in two dimensions, the X- and Y-coordinate information can be set on two axes, and the selected element can be mapped by displaying the intensity of the selected element's emission spectrum in color.

This invention makes it possible to classify and estimate the metal elements and compounds contained in unknown samples. By decomposing the spectral data of known substances into multiple factors in advance and training the system to classify elements and compounds using a decision tree, this can be used to classify the components of unknown samples.

In addition, by evaluating sample-specific errors and correcting them in advance, metal elements and compounds can be classified and estimated with high accuracy. Corrections can also be made retrospectively, eliminating errors, including white noise.

In the present invention, reference parameters that characterize the plasma state (time, space, etc.) are used to determine whether to accept or reject the atomic emission spectrum to be measured. Machine learning is also used to set the reference parameters. While conventional methods search for the spectrum of the measurement target as a peak with a certain width, the present invention uses NMF to machine-learn big data (all spectral components obtained from various analytical devices, etc.), which is difficult for humans to handle, to construct the reference parameters. By processing all spectral components rather than processing signal peaks as in conventional methods, the amount of data used for evaluation increases to ^103. As a result, it is now possible to evaluate spectral intensity, shape, and other parameters, which were previously difficult to assess, and there is a prospect of achieving high analytical accuracy.

Depending on the type, baseline correction used in preprocessing can be applied to each spectrum using functions such as asymmetric weighted penalty least-squares smoothing or machine learning such as pattern recognition. If you have an environment where you can measure known samples or use databases from public research institutions as training data, you can retroactively improve the performance of various analytical instruments you already own (not only analytical methods using laser light sources such as Raman scattering and LA-ICP-MS, but all measuring instruments that use spectroscopy) and reduce sample dependency.

By simply using databases (spectroscopy and radiation spectrum information) from public research institutions and measurements of standard samples with known material compositions as training data, it is possible to easily improve the performance of various analytical instruments and reduce sample dependency without having to modify the hardware, such as by implementing it in the instrument's analysis software. A wide variety of applications are expected, including not only spectroscopic measurements using light sources such as lasers, X-rays, and neutrons as probes, but also electron microscopes (transmission electron microscopes: TEM-EDX, scanning electron microscopes: SEM-EDX) equipped with energy dispersive X-ray analysis (EDX) as a spectroscopic function, and electron probe microanalyzers (EPMA) that visualize spectroscopic data as surface information.

The above describes examples of the present invention, but the present invention is not limited to these. For example, NMF is a method used in machine learning to separate orchestral sounds into factors such as violins and violas, and can create and classify characteristic sounds (reference parameters) for violins and violas. In this way, the present invention can be applied to a variety of big data.

It can also be used to create material distributions and hardness distributions within unknown samples, or to apply them 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: Emission light 400: Spectrometer 410: Detector 420: Spectral data 430: Baseline 440: Factor 450: Factor score 460: Score threshold 500: Decision tree

Claims (6)

The control device
factorizing spectral data of a standard sample of a predicted compound species to derive factors and factor scores representing the influence of the factors; learning and setting in advance a score threshold for classifying the compound species using the factor scores in a decision tree and a factor application order for applying the factor scores to the decision tree;
calculating factor scores for the unknown sample using the factors of the standard sample for the spectral data of the unknown sample, and classifying the factor scores of the unknown sample using the decision tree in accordance with the factor application order and the score threshold, thereby identifying the compound species contained in the unknown sample;
A compound identification method characterized by: The spectral data is corrected to a baseline fitted using a penalized asymmetric least squares method so that errors are assessed more severely 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 . The factorization involves decomposing a K-row, N-column matrix consisting of K wavelength dimensions and N sample numbers of the spectral data using M specified non-negative basis vectors, and deriving the influences of the factors obtained as the factor scores.
3. The compound identification method according to claim 1 or 2. The decision tree derives factors and score thresholds that minimize a cost function using impurity when classifying the compound species.
4. The compound identification method according to claim 1, wherein the compound is a nucleotide sequence. Executing 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 unknown sample with a laser;
a spectrometer for resolving light emitted from the unknown sample irradiated by the laser;
a detector for acquiring spectral data generated by the spectrometer;
Identifying compounds contained in the unknown sample using the spectral data input from the detector;
6. The compound identification device according to claim 5.