Spectral analysis device, spectral analysis system, and spectral analysis method

Abstract

Disclosed is a spectroscopic analysis device that quantifies / recognizes characteristics of a sample on the basis of spectroscopic data of the sample generated with one or more excitation wavelengths. The spectroscopic analysis device includes a computing device and a memory storing a program executed by the computing device. The arithmetic device receives as input spectral data obtained from a sample, acquires wavelength selection information indicating the degree of contribution to quantification / recognition and / or the necessity of the contribution to quantification / recognition for each wavelength band of the spectral data, selects a wavelength band used for quantification / recognition of characteristics of the sample from the wavelength bands of the spectral data on the basis of the wavelength selection information, and performs quantification / recognition of the characteristics of the sample on the basis of the selected wavelength band. The characteristics of the sample are quantified / identified on the basis of the detection intensity corresponding to the selection result of the wavelength band and output.

Description

[0001] References

[0002] This application claims the benefit of priority based on Japanese patent application No. 2023-055642, filed on March 30, 2023, the contents of which are hereby incorporated by reference into this application. Technical Field

[0003] The present invention relates to spectral analysis. Background Art

[0004] In a spectroscopic analysis device, a spectrum is obtained by irradiating a sample with excitation light after spectroscopy and measuring the light detected from the sample. For example, a three-dimensional fluorescence spectrum called a fluorescence fingerprint is measured using a spectrofluorophotometer. When a sample is irradiated with excitation light of a specific wavelength, the excited sample emits fluorescence of various wavelengths. This is called a fluorescence spectrum, and the intensity of the fluorescence at each fluorescence wavelength is called fluorescence intensity.

[0005] The sample emits different fluorescence spectra by changing the wavelength of the excitation light. Therefore, by sequentially changing the wavelength of the excitation light and measuring the fluorescence spectrum, a three-dimensional spectrum with three axes, namely the excitation wavelength, the fluorescence wavelength, and the fluorescence intensity, is obtained. The above-mentioned detection wavelength is equivalent to the fluorescence wavelength, and the detection intensity is equivalent to the fluorescence intensity. By analyzing the shape of the obtained three-dimensional spectrum, it is possible to quantify the components contained in the sample and identify the types.

[0006] As a conventional method for analyzing spectral data, for example, there is a method of performing multivariate analysis using each detection intensity in a spectrum as an explanatory variable.

[0007] Patent Document 1 discloses a method of estimating the content of a target substance while limiting the excitation wavelength used in the estimation by performing sparse estimation on spectral data using light absorption characteristic values ​​(detection intensities) at a plurality of excitation wavelengths as explanatory variables.

[0008] Patent Document 2 discloses a method of setting a rectangular measurement window for a fluorescent fingerprint and performing multivariate analysis using the integrated value of the fluorescence intensity within the measurement window.

[0009] Prior art literature

[0010] Patent Literature

[0011] Patent Document 1: Japanese Patent Application Publication No. 2018-013418

[0012] Patent Document 2: Japanese Patent Application Publication No. 2015-180895 Summary of the invention

[0013] Problems to be solved by the invention

[0014] The technology described in Patent Document 1 automatically limits the bands used in quantification / identification by performing sparse extrapolation using all detection intensities in spectral data as explanatory variables. However, as in the technology described in Patent Document 1, when sparse extrapolation and multivariate analysis are performed using the detection intensities of all bands in the spectrum as explanatory variables, depending on the status of the noise components included in the training data, a quantitative / identification model focusing on the noise components that do not originally participate in quantification / identification will be constructed, and the performance of the quantitative / identification model will be greatly reduced. If the bands are to be selected manually by experts in order not to use such bands in quantification / identification, professional knowledge is required, so the increase in analysis time becomes a problem.

[0015] In addition, the technology described in Patent Document 2 sets a rectangular measurement window and uses the integral value of the fluorescence intensity within the measurement window for multivariate analysis. This limits the band used in quantification / identification. However, the above-mentioned measurement window needs to be designed manually, or explored by repeatedly constructing a regression equation based on the fluorescence intensity within the randomly set measurement window and resetting the measurement window based on the quantitative / identification results. It requires professional knowledge to distinguish the effective band for quantification / identification of the analysis object, and in the case of random attempts, there is a possibility that a band containing noise components will be selected depending on the condition of the noise components contained in the training data.

[0016] In view of the above problems, the present invention aims to provide a spectral analysis device that calculates information for selecting a wavelength band effective for quantification / identification, namely wavelength selection information, and has an automatic wavelength selection function based on the wavelength selection information, thereby performing high-precision quantification / identification without the need for professional knowledge and shortening the time spent on wavelength selection.

[0017] Technical solutions to solve problems

[0018] One embodiment of the present invention is a spectral analysis device for quantifying / identifying characteristics of a sample based on spectral data of the sample generated using one or more excitation wavelengths, the device comprising a computing device and a memory for storing a program executed by the computing device, the computing device receiving one or more spectral data obtained from one or more samples as input, obtaining wavelength selection information indicating the contribution to quantification / identification and / or whether it is necessary for each band of the spectral data, selecting a band used in quantification / identification of characteristics of the sample from the bands of the spectral data based on the wavelength selection information, and quantifying / identifying the characteristics of the sample based on the detection intensity corresponding to the selection result of the band and outputting it.

[0019] Effects of the Invention

[0020] According to one embodiment of the present invention, it is possible to realize spectral analysis that can perform high-precision quantification and identification in a shortened time. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a diagram showing an example of the hardware configuration of the spectrum analysis device according to the first embodiment.

[0022] Figure 2 This is a diagram showing an example of a functional block diagram of the spectrum analysis device according to the first embodiment.

[0023] Figure 3A This is a diagram showing an example of spectrum data.

[0024] Figure 3B This is a diagram showing an example of spectrum data.

[0025] Figure 4A This is a diagram showing an example of wavelength selection information.

[0026] Figure 4B This is a diagram showing an example of wavelength selection information.

[0027] Figure 5 This is a diagram showing an example of the wavelength selection result.

[0028] Figure 6 This is a diagram showing an example of the processing flow of the spectrum analysis method of Example 1.

[0029] Figure 7 This is a diagram showing an example of a functional block diagram of the spectrum analysis device according to the second embodiment.

[0030] Figure 8 This is a diagram showing an example of a graph illustrating the relationship between the number of wavelength bands selected by automatic wavelength selection and the quantification / identification accuracy.

[0031] Fig. 9 This is a diagram showing an example of the processing flow of the spectrum analysis method of Example 2.

[0032] Fig.10 This is a diagram showing an example of the hardware configuration of the spectrum analysis system of Example 3. DETAILED DESCRIPTION

[0033] Hereinafter, embodiments of the spectrum analysis method, device, and system of the present invention will be described with reference to the accompanying drawings. In the following description and the accompanying drawings, components having the same functional structure are denoted by the same reference numerals to omit repeated description.

[0034] Example 1

[0035] <Hardware structure of spectrum analyzer>

[0036] In the first embodiment, a spectrum analyzer is described which estimates information for selecting a wavelength band effective for quantification / identification (quantification or identification), that is, wavelength selection information, and has an automatic wavelength selection function for automatically selecting a wavelength band used for quantification / identification based on the wavelength selection information, thereby shortening the analysis time and achieving high-precision quantification / identification without requiring professional knowledge. Quantification is the estimated amount, and identification is the estimated type.

[0037] use Figure 1 The hardware structure of the spectroscopic analysis device of Example 1 is described. The spectroscopic analysis device 100 includes an interface 110 , a computing device 111 , a memory 112 , and a bus 113 . The interface 110 , the computing device 111 , and the memory 112 transmit and receive information via the bus 113 .

[0038] The components of the spectroscopic analysis device 100 are described below. The interface 110 is a communication device that transmits and receives signals to and from a device located outside the spectroscopic analysis device 100 via a network. Devices that transmit and receive signals to and from the interface 110 include a control device 120 that controls a spectrometric device 121, a spectrometric device 121 that measures three-dimensional spectrum data, and a display device 122 such as a monitor and a printer that displays the processing results of the spectroscopic analysis device 100.

[0039] The method of sending and receiving signals with the external device may be a wired connection method using a cable such as an optical fiber, or a wireless connection method using a wireless communication technology such as Bluetooth. In addition, the spectrum analysis device 100 does not need to be set in the same facility as the control device 120 or the spectrum measurement device 121, and the display device 122, and may be set and mounted in a separate computer or cloud. In this case, the signal transmission and reception between the interface 110 and the external device is connected, for example, via a LAN (Local Area Network) or a WAN (Wide Area Network).

[0040] The computing device 111 is a device that executes various processes in the spectroscopic analysis device 100, and is, for example, a CPU (Central Processing Unit) or an FPGA (Field-Programmable Gate Array). The functions executed by the computing device 111 will be described later.

[0041] The memory 112 is a device for storing programs executed by the operation device 111 , parameters of quantitative / identification models, processing results, and the like, and is a HDD, SSD, RAM, ROM, flash memory, and the like.

[0042] <Functional Configuration of Spectral Analysis Device 100 >

[0043] Figure 2 This is an example of a functional block diagram of an embodiment of the spectrum analysis device 100. Each functional unit may be realized by the operation device 111 executing a predetermined program, or may be realized by dedicated hardware.

[0044] The spectrum analysis device 100 includes an input unit 201, a wavelength selection information calculation unit 202, an automatic wavelength selection unit 203, an estimation unit 204, and an output unit 205 as functional units. Each functional unit will be described below.

[0045] The input unit 201 receives one or more discretized pieces of spectrum data input from the interface 110 .

[0046] The wavelength selection information calculation unit 202 reads the wavelength selection information calculation model stored in the memory 112, calculates the wavelength selection information based on the input spectrum data, or obtains the wavelength selection information stored in the memory 112. Alternatively, the wavelength selection information may be calculated by combining the wavelength selection information calculated based on the input spectrum data with the wavelength selection information stored in the memory 112.

[0047] The automatic wavelength selection unit 203 automatically selects a wavelength band used for quantification / identification based on the wavelength selection information, and outputs the selection result as a wavelength selection result.

[0048] The estimation unit 204 reads one of the quantitative / identification models stored in the memory 112 , and performs quantitative / identification based on the spectrum data based on the wavelength selection result obtained by the automatic wavelength selection unit 203 .

[0049] The output unit 205 outputs at least one of the wavelength selection information, the wavelength selection result, and the quantification / identification result.

[0050] In addition, the above functions do not need to be Figure 2 The functional parts of the same structure, as long as it can achieve the same processing as the actions of each functional module. Figure 6 An example of a processing flow chart of Example 1 is shown in FIG. Figure 2 The elements of the functional block diagram shown correspond to each other.

[0051] In the input step 601 , one or more discretized spectral data input from the interface 110 are received.

[0052] In the wavelength selection information calculation step 602, the wavelength selection information calculation model stored in the memory 112 is read, and the wavelength selection information is calculated based on the above-mentioned input spectrum data, or the wavelength selection information stored in the memory 112 is obtained. Alternatively, the wavelength selection information can be calculated by combining the wavelength selection information calculated based on the above-mentioned input spectrum data with the wavelength selection information stored in the memory 112.

[0053] In the automatic wavelength selection step 603, a wavelength band used for quantification / identification is automatically selected based on the wavelength selection information, and the selection result is output as a wavelength selection result.

[0054] In the calculation step 604, one of the quantitative / identification models stored in the memory 112 is read, and based on the wavelength selection result obtained in the above-mentioned automatic wavelength selection step 603, quantitative / identification is performed according to the spectral data.

[0055] The output step 605 outputs at least one of the wavelength selection information, the selection result, and the quantification / identification result.

[0056] Hereinafter, the detailed operation will be described with each functional unit as the subject, but each step corresponding to each functional unit may be replaced as the subject.

[0057] <Structure and operation of each part>

[0058] The operations of the input unit 201 , the wavelength selection information calculation unit 202 , the automatic wavelength selection unit 203 , and the estimation unit 204 among the functional units will be described in detail below.

[0059] The input unit 201 receives one or more discretized spectral data via the interface 110. The spectral data refers to one-dimensional data with one wavelength band and detection intensity as axes, or multidimensional data with two or more wavelength bands and detection intensity as axes. The spectral data received by the input unit is output to the wavelength selection information calculation unit 202, or the spectral data is stored in the memory 112 and output to the wavelength selection information calculation unit 202.

[0060] As an example of spectral data, Figure 3A The fluorescence fingerprint data is shown in . The discretized fluorescence fingerprint data 300 is an example of fluorescence fingerprint data stored in a discretized manner. Figure 3B The fluorescence fingerprint data visualization example 301 is shown. The fluorescence fingerprint data visualization example 301 is an example in which the vertical axis is the excitation light wavelength, the horizontal axis is the detection wavelength, and the detection intensity at each combination of the excitation and detection wavelengths is displayed as contour line data.

[0061] As shown in the discretized fluorescence fingerprint data 300, the fluorescence fingerprint data is measured and recorded in a discrete manner. The discretized fluorescence fingerprint data 300 shows an example in which 200nm to 700nm are measured at intervals of 5nm for the excitation light / detection light. For example, the detection intensity of the detection light 200nm relative to the excitation light 200nm is recorded in the first row and first column of the discretized fluorescence fingerprint data 300, the detection intensity of the detection light 205nm relative to the excitation light 200nm is recorded in the first row and second column, and the detection intensity of the detection light 200nm relative to the excitation light 205nm is recorded in the second row and first column. In addition, since 500nm is recorded at intervals of 5nm, the number of columns and rows of the discretized fluorescence fingerprint data 300 is 100×100.

[0062] The fluorescent fingerprint data visualization example 301 is an example of two-dimensionally imaging the discretized fluorescent fingerprint data 300 as contour data. When people observe the fluorescent fingerprint data, they often use the fluorescent fingerprint data visualization example 301 to confirm. In addition to contour data, there are also cases where the detection intensity is assigned to the brightness value or color. Grayscale images, heat map images, bird's-eye views, etc. are used to represent. The shape of the fluorescent fingerprint is different according to the properties of the sample to be measured, and the identification of the sample and the quantification of specific components are performed based on the difference in the shape.

[0063] In the fluorescence fingerprint data visualization example 301, the fluorescence fingerprint data includes noise components that do not participate in quantification / identification. For example, a band including Rayleigh scattering is generally regarded as a noise component in quantification / identification using fluorescence fingerprints. Rayleigh scattering is not fluorescence emitted by a sample, but light reflected from a sample. Because Rayleigh scattering is detected, features unrelated to the characteristics of the sample are reflected in the fluorescence fingerprint.

[0064] For example, a method can be considered to automatically limit the wavelength band used for quantification / identification by performing sparse estimation using all detection intensities in spectral data as explanatory variables. However, when analyzing spectral data containing Rayleigh scattering, there is a risk that the accuracy of quantification / identification will be reduced.

[0065] For example, when the pattern of Rayleigh scattering has an accidental correlation with the target value of quantification / identification of training data due to the influence of the measurement environment, etc., when constructing a quantification / identification model with the detection intensity of Rayleigh scattering included in the explanatory variable, there is a case where a quantification / identification model is constructed that focuses on the pattern of Rayleigh scattering, although it is a feature that is not originally related to the characteristics of the sample. In order to avoid this, experts select wavelengths in advance so that the band containing Rayleigh scattering is not used in quantification / identification.

[0066] Rayleigh scattering has the characteristic of being detected in intervals near the wavelength band where the excitation wavelength is a natural multiple or a reciprocal multiple of the detection wavelength due to the influence of the diffraction grating, and the width of the interval varies depending on the measurement conditions, environment, and sample of the fluorescence fingerprint. Therefore, for example, an expert needs to set the width according to the obtained spectral data to determine the interval of Rayleigh scattering.

[0067] In addition, noise components such as Raman scattering and noise during measurement are also causes of reduced quantification / identification performance similar to Rayleigh scattering, so in order to reduce the influence of noise components on quantification / identification, experts make a pre-selection.

[0068] In the method disclosed in this embodiment, based on the input spectral data, the contribution to the effectiveness of quantification / identification is calculated by digitizing, and / or the necessary / unnecessary bands for quantification / identification are extracted, which are used as wavelength selection information. Here, the contribution can take values ​​of more than 3 levels. Then, the band used in quantification / identification is automatically selected in advance, and quantification / identification is performed using only the detection intensity corresponding to the selected band as the explanatory variable. By automatically performing the pre-selection, the cost of manual selection by experts can be reduced, and a high-precision quantification / identification function can be provided.

[0069] The wavelength selection information calculation unit 202 receives one or more spectrum data from the input unit 201 and / or the memory 112. The wavelength selection information calculation unit 202 reads the wavelength selection information calculation model stored in the memory 112, and calculates the wavelength selection information based on the received spectrum data, or obtains the wavelength selection information stored in the memory 112.

[0070] The wavelength selection information refers to the contribution to the effectiveness of quantification / identification expressed numerically for each wavelength band constituting the received spectral data, and / or the necessary / unnecessary wavelength bands expressing the effectiveness of quantification / identification with necessary / unnecessary true / false values. The wavelength selection information calculation model for calculating the wavelength selection information is classified into two types: a contribution calculation model for calculating the contribution based on the input spectral data, and a necessary / unnecessary wavelength band calculation model for calculating necessary / unnecessary wavelength bands. The contribution calculation model and the necessary / unnecessary wavelength band calculation model can be constructed using a machine learning method or a manually designed algorithm.

[0071] As a method of calculating the contribution from the spectral data received by the input unit 201, for example, there is a method of reading a contribution calculation model that has been supervised learned using a plurality of quantification / identification tasks from the memory 112, and calculating the contribution to quantification / identification corresponding to each band in the spectral data. The contribution calculation model is a model that receives spectral data of the same format as one or more samples as input, calculates all the contributions to quantification / identification of each band constituting the spectral data, and converts all the contributions into tensors to output as a contribution map isomorphic to the spectral data.

[0072] Here is an example of a contribution calculation model using Multi-Layer Perceptron (MLP). MLP is a model that repeatedly performs linear mapping and nonlinear mapping on input data more than once, and performs regression or recognition. The model is constructed by optimizing the parameters of the linear mapping. First, for one or more spectral data received, the feature quantities such as the average of each band in equation 1 and the variance of each band in equation 2 are calculated. Using the feature quantities of multiple samples, a more appropriate contribution can be calculated. Here x s n are the detection intensities of one or more pieces of spectrum data of the same format received by the input unit 201 , s represents the index of a sample, S represents the number of samples, n represents the index of a band constituting the spectrum data, and N represents the number of bands.

[0073] [Mathematical formula 1]

[0074]

[0075] [Mathematical formula 2]

[0076]

[0077] Next, as in Formula 3, the MLP using the above-mentioned feature quantity as input is used to calculate a vector representing the contribution of each band. C () represents a function that performs MLP operation processing with the above-mentioned feature quantity as input and outputs the contribution of each band to quantification / identification as an N-dimensional vector, and each element of the output N-dimensional vector represents the contribution of each band. For example, the mathematical formula c n Indicates the contribution of the band with index n. The contribution calculated by equation 3 is used as a tensor, and the contribution map is isomorphic to the spectral data, which is the output of the contribution calculation model.

[0078] [Mathematical formula 3]

[0079]

[0080] The above MLP CThis is a model that pre-optimizes the parameters of the linear mapping through supervised learning using the target value of the contribution (target contribution map) as training information. A method for generating a target contribution map of training information includes, for example, preparing a training data set and a verification data set consisting of a plurality of spectral data and a group of quantitative / identified training information, and performing a linear regression using the detection intensity of only one band as an explanatory variable for each band constituting the spectral data (in the case of multidimensional data, each combination of bands is represented), using the training data set, and using the quantitative / identification accuracy relative to the verification data set as a target contribution map.

[0081] In addition, the accuracy of quantification / identification when a combination of multiple bands is used as an explanatory variable can be fully calculated, and the result of averaging each band can be used as a target contribution map. In addition, in order to prevent the contribution from being specifically targeted at the above-mentioned verification data set, a plurality of training data sets and verification data set segmentation modes can be generated for a data set consisting of a plurality of spectral data and quantitative / identification training information groups, and the target contribution maps obtained in each mode can be summarized by statistics such as average or maximum value as the target contribution map, thereby improving the generalization performance.

[0082] The loss function in supervised learning is, for example, Mean Squared Error in quantitative cases, and Cross Entropy in recognition cases. The optimization method uses, for example, the steepest descent method or the stochastic gradient descent method. The above loss function and optimization method are examples, and other loss functions and optimization methods can also be used. The calculation of the feature quantity can be a general statistic such as Mathematical Formula 1 and Mathematical Formula 2, or a machine learning method such as Convolutional Neural Network or Transformer, or a manually designed feature quantity such as HOG.

[0083] In addition, the wavelength selection information calculation unit 202 may store the contribution obtained by inputting the training data set to the contribution calculation model in the memory 112, and read it out when the spectrum data is newly input and use it as the contribution. In addition, the average value or maximum value of the contribution stored in the memory 112 and the contribution obtained by inputting the newly input spectrum data to the contribution calculation model may be taken, and the result of the aggregation may be used as the contribution.

[0084] Next, an example of a method for extracting necessary / unnecessary bands is described. An example of a method for extracting necessary / unnecessary bands uses a rule-based model different from machine learning. For example, as a method for extracting necessary / unnecessary bands for quantification / identification, there is a method of using a necessary / unnecessary band calculation model that automatically extracts a detection band corresponding to Rayleigh scattering as an unnecessary interval for each excitation band.

[0085] The necessary / unnecessary band calculation model is described in Example 1. First, in each excitation band, a band whose excitation wavelength is a natural number multiple and a reciprocal number multiple of the detection wavelength among the detection bands constituting the spectrum data is extracted as a central band.

[0086] For each central band, in the detection band smaller than the central band, the local minimum of the detection intensity is detected, and the detection band corresponding to the local minimum and closest to the central band is used as the left band. Similarly, in the detection band larger than the central band, the local minimum of the detection intensity is detected, and the detection band corresponding to the local minimum and closest to the central band is used as the right band. For the group of the left band and the right band, the interval above the left band and below the right band is regarded as Rayleigh scattering, thereby automatically extracting unnecessary bands.

[0087] The above method is also the same when the excitation band corresponding to Rayleigh scattering is automatically extracted as an unnecessary interval for each detection band. The extracted unnecessary bands can be regarded as unnecessary bands, and the unextracted intervals can be regarded as necessary bands. Similar to the contribution map, all necessary / unnecessary bands can be transformed into a necessary / unnecessary band map that is a tensor isomorphic to the spectral data.

[0088] The method for extracting necessary / unnecessary bands for quantification / identification is not limited to this, and may also be a method of calculating the statistics (average, variance, covariance, etc.) of the detection intensity of each band and treating bands below or above a certain threshold as necessary or unnecessary bands. In addition, when the range of occurrence of the spectrum to be focused on is limited to a certain extent, only the interval may be treated as a necessary band, and the others may be treated as unnecessary bands.

[0089] In addition, the wavelength selection information calculation unit 202 may store the necessary / unnecessary bands obtained by inputting the training data set into the necessary / unnecessary band calculation model in the memory 112, and read them out when the spectrum data is newly input and use them as necessary / unnecessary bands. In addition, the necessary / unnecessary bands stored in the memory 112 and the necessary / unnecessary bands obtained by inputting the newly input spectrum into the necessary / unnecessary band calculation model may be logically obtained by taking the logical sum or logical AND, and the result of the collection may be used as the necessary / unnecessary band.

[0090] As an example of wavelength selection information, Figure 4A An example of contribution map visualization 401 is shown in FIG. Figure 4B 4 shows a necessary / unnecessary band map visualization example 402. Each wavelength selection information corresponds to the fluorescence fingerprint data visualization example 301. The contribution map visualization example 401 visualizes the contribution map indicating the effectiveness of each band in the fluorescence fingerprint data visualization example 301 for quantification / identification as contour line data.

[0091] In addition to the contour data, it is also possible to use a grayscale image or a heat map image, a bird's-eye view, etc. that assigns the detection intensity to a brightness value or color for visualization. The necessary / unnecessary band mapping visualization example 402 uses true and false values ​​to indicate whether each band of the fluorescent fingerprint data visualization example 301 is necessary or unnecessary for quantification / identification, and the necessary bands are visualized as white within the dotted line, and the unnecessary bands are visualized as black within the dotted line. In addition, the color assigned to the true and false values ​​can be any color as long as it is different for true and false.

[0092] The wavelength selection information calculation unit 202 outputs the information of the contribution and the necessary / unnecessary wavelength band as wavelength selection information to the automatic wavelength selection unit 203. In addition, the wavelength selection information calculation unit 202 may calculate only one of the contribution or the necessary / unnecessary wavelength band and output it as wavelength selection information. In addition, the result of summarizing the information of the contribution and the necessary / unnecessary wavelength band may be output as wavelength selection information. For example, the summarization method includes summarizing the false ( Figure 4B The black band in the figure is considered as contribution 0, and the true ( Figure 4B The wavelength band (white in the image) is regarded as contribution 1, and the method of multiplying the contribution of multiple values ​​is used. By summarizing, more suitable wavelength selection can be performed.

[0093] The automatic wavelength selection unit 203 automatically selects a wavelength band used in quantification / identification based on the wavelength selection information, and outputs the selection result as a wavelength selection result. The automatic selection method may, for example, select a wavelength band with a contribution degree above a fixed threshold as a wavelength band used in quantification / identification, or may select a wavelength band excluding an unnecessary wavelength band or a necessary wavelength band as a wavelength band used in quantification / identification. In addition, the wavelength selection information and / or the wavelength selection result may be output to the output unit without passing through the estimation unit, or may be output to the estimation unit.

[0094] As an example of the wavelength selection result, Figure 5501 is a visualization example of the wavelength selection result. The visualization example 501 of the wavelength selection result corresponds to the visualization example 301 of the fluorescence fingerprint data, the visualization example 401 of the contribution degree, and the visualization example 402 of the necessary / unnecessary bands. In the visualization example 501 of the wavelength selection result, for each band of the fluorescence fingerprint data visualization example 301, a true / false value is used to indicate whether the band is pre-selected as a band to be used in quantification / identification, and the pre-selected band is visualized as white within the dotted line, and the unpre-selected band is visualized as black within the dotted line. In addition, the color assigned to the true / false value can be any color as long as it is different for true and false.

[0095] In the wavelength selection result visualization example 501, the false ( Figure 4B Bands (black in ) and bands whose contribution is less than a preset threshold in the contribution visualization example 401 are selected as bands not used in quantification / identification, and bands above the threshold are selected as bands used in quantification / identification.

[0096] The estimation unit 204 reads any quantitative / identification model stored in the memory 112, and performs quantitative / identification using the detection intensity corresponding to the wavelength selection result obtained in the automatic wavelength selection unit 203 as an explanatory variable. In the case of the wavelength selection result visualization example 501, the fluorescence intensity corresponding to the white band within the dotted line is used as an explanatory variable. The quantitative / identification model can use a linear regression model such as Logistic regression or Lasso regression using the above-mentioned detection intensity as an explanatory variable, or a machine learning model such as Neural Network, Random Forest, Support Vector Machine, etc.

[0097] Example 2

[0098] In Example 2, a spectral analysis device described in Example 1 is described in which a function of updating and building a wavelength selection information calculation model and a function of building a quantification / identification model are added to the spectral analysis device described in Example 1, thereby enabling automatic wavelength selection and high-precision quantification / identification for quantification / identification problems set by the user.

[0099] <Hardware structure of spectrum analyzer>

[0100] The hardware structure of the spectrum analysis device of Example 2 is as follows Figure 1 The hardware configuration of the spectrum analyzer of the first embodiment shown is the same as that of the embodiment, so the description thereof will be omitted.

[0101] <Functional structure of spectrum analyzer>

[0102] Figure 7This is an example of a functional block diagram of a spectrum analyzer according to Embodiment 2. The spectrum analyzer 700 includes an input unit 701, a wavelength selection information calculation unit 702, an automatic wavelength selection unit 703, an estimation unit 704, and an output unit 705 as functional units. Each functional unit will be described below.

[0103] The input unit 701 receives one or more pieces of spectrum data or one or more sets of spectrum data and training information via the interface 110. The spectrum data or the set of spectrum data and training information received by the input unit 701 is output to the wavelength selection information calculation unit 702, or is stored in the memory 112 and output to the wavelength selection information calculation unit 702.

[0104] The wavelength selection information calculation unit 702 calculates the wavelength selection information based on one or more spectral data or one or more sets of spectral data and training information, or obtains the wavelength selection information stored in the memory 112. In addition, when receiving the set of spectral data and training information, the wavelength selection information calculation model stored in the memory 112 may be read, and the wavelength selection information calculation model may be updated using the set of spectral data and training information, or a new wavelength selection information calculation model may be constructed and stored in the memory 112. With respect to one or more spectral data or one or more sets of spectral data and training information, the information input from the input unit 701 may be directly received, or received via the memory 112, or both may be used simultaneously.

[0105] The automatic wavelength selection unit 703 automatically selects a wavelength band used for quantification / identification based on the wavelength selection information or the wavelength selection information and the quantification / identification result output by the estimation unit 704, and outputs the selected result as a wavelength selection result.

[0106] The estimation unit 704 constructs a quantitative / identification model using one or more sets of spectral data and training information, or performs quantitative / identification of one or more spectral data, based on the wavelength selection result obtained by the automatic wavelength selection unit 703. The information input from the input unit 701 is received via the automatic wavelength selection unit 703 or the memory 112 regarding one or more spectral data or one or more sets of spectral data and training information. Alternatively, both may be used simultaneously.

[0107] The output unit 705 outputs at least one of the wavelength selection information, the selection result, and the quantitative / identification result, and outputs information to assist the user in wavelength selection and / or determination of a quantitative / identification model as necessary.

[0108] In addition, the above functions do not need to be Figure 7 The functional part is generally composed of only those processes that can achieve the same operation as each functional module. Fig. 9An example of a processing flow chart of Example 2 is shown in FIG. Figure 7 The elements of the functional block diagram shown correspond to each other.

[0109] In the input step 901, one or more pieces of spectral data or one or more sets of spectral data and training information are received via the interface 110. The spectral data or the set of spectral data and training information received in the input step 901 are used in the wavelength selection information calculation step, or are stored in the memory 112 and used in the wavelength selection information calculation step.

[0110] In the wavelength selection information calculation step 902, the wavelength selection information is calculated based on one or more spectral data or one or more sets of spectral data and training information, or the wavelength selection information stored in the memory 112 is obtained. In addition, when the set of spectral data and training information is received, the wavelength selection information calculation model stored in the memory 112 can be read, and the wavelength selection information calculation model can be updated using the set of spectral data and training information, or a new wavelength selection information calculation model can be constructed and stored in the memory 112.

[0111] Regarding one or more pieces of spectral data or one or more sets of spectral data and training information, the information input in step 901 may be directly received, or received via the memory 112 , or both may be used simultaneously.

[0112] In the automatic wavelength selection step 903, based on the wavelength selection information or the wavelength selection information and the quantification / identification result outputted in the estimation step 904, a wavelength band used in the quantification / identification is automatically selected, and the selected result is outputted as a wavelength selection result.

[0113] In the calculation step 904, based on the wavelength selection result obtained in the automatic wavelength selection step 903, a quantitative / identification model is constructed using one or more sets of spectral data and training information, or quantitative / identification of one or more spectral data is performed. Regarding one or more spectral data or one or more sets of spectral data and training information, the information input from the input step 901 is received via the automatic wavelength selection step 903 or the memory 112. Alternatively, both may be used at the same time.

[0114] The wavelength selection information and / or the selection result and / or the quantitative / identification result are output in the output step 905. In addition, information for assisting the user in wavelength selection and / or determining a quantitative / identification model is output as necessary.

[0115] Hereinafter, the detailed operation will be described with each functional unit as the subject, but each step corresponding to each functional unit may be replaced as the subject.

[0116] <Structure and operation of each part>

[0117] The operations of the input unit 701 , the wavelength selection information calculation unit 702 , the automatic wavelength selection unit 703 , and the estimation unit 704 in the functional unit will be described in detail below.

[0118] The input unit 701 receives one or more spectral data or one or more sets of spectral data and training information via the interface 110. The spectral data or the set of spectral data and training information received by the input unit may be stored in the memory 112 or directly output to the wavelength selection information calculation unit 702.

[0119] When receiving only one or more pieces of spectrum data, the wavelength selection information calculation unit 702 performs the same processing as the wavelength selection information calculation unit 202 described in Example 1. When receiving one or more sets of spectrum data and training information, the wavelength selection information calculation model stored in the memory 112 is read to calculate the wavelength selection information based on the set of spectrum data and training information, or to update and construct the wavelength selection information calculation model and calculate the wavelength selection information. Each processing is described below.

[0120] As a method for calculating contribution based on a combination of spectral data and training information, there are, for example, a method of pre-generating a contribution calculation model corresponding to spectral data by performing supervised learning using multiple quantitative / identification tasks, and a method of calculating contribution based on one or more combinations of spectral data and training information.

[0121] In Example 1, a method for calculating the contribution using only one or more spectral data is described. However, by estimating the contribution using a combination of spectral data and training information, the contribution to quantification / identification can be estimated more accurately. In Example 2, an example of calculating the contribution using an already constructed contribution calculation model or updating / constructing a contribution calculation model using newly input spectral data and training information is described.

[0122] Here, an example of a contribution calculation model using a Multi-Layer Perceptron (MLP) is described. First, a feature quantity is calculated for a set of one or more spectral data and training information received by an input unit. The MLP described in Formula 4 Z It is an MLP that takes a set of spectral data and training information as input and outputs features that are isomorphic to the spectral data. s n Indicates that MLP Z The index of the input sample is the nth element of the feature quantity obtained by combining the spectral data of s with the training information. Here, y sis the training information of the sample index s in the set of one or more spectral data and training information received by the input unit 701. In Mathematical Formula 4, the training information y s This is explained as a scalar value, but if the training information is a vector, then if the MLP is designed to receive each element of the vector as input Z , then the feature quantity can also be output.

[0123] [Formula 4]

[0124]

[0125] Similar to Math. 1 and Math. 2 of Example 1, Math. 5 and Math. 6 are used to calculate the average of Math. 5 and the variance of Math. 6 for the feature quantity obtained by Math. 4.

[0126] [Formula 5]

[0127]

[0128] [Mathematical formula 6]

[0129]

[0130] Thereafter, by comparing the MLP used in Example 1 C The same model calculates the contribution by inputting the results of equations 5 and 6, and outputs a contribution map. As in equations 4, 5, and 6, by using the feature quantity that combines the spectral data and the training information, the contribution can be calculated taking the training information into consideration.

[0131] The above MLP Z and MLP C The parameters of the linear mapping are optimized by supervised learning using the target contribution map as training information. For the target contribution map, training data and verification data of the spectral data are prepared, and linear regression is performed on only one band of each band that constitutes the spectral data. The quantitative / identification accuracy corresponding to the spectral data not used in the linear regression is used as training information.

[0132] Alternatively, the quantification / identification accuracy when the combination of multiple bands is used as the explanatory variable can be fully calculated, and the average result of each band can be used as training information. In addition, in order to prevent the contribution from being specifically targeted at the above-mentioned verification data, multiple training data and verification data patterns can be generated for the spectral data, and the target contribution map obtained under each pattern can be summarized by statistics such as average or maximum value, thereby improving the generalization performance.

[0133] As feature quantities, general statistics such as Equation 1, Equation 2, Equation 4, and Equation 5 may be used, or machine learning methods such as Convolutional Neural Network, Transformer, and Neural Process, or manually designed feature quantities such as HOG may be used.

[0134] In the case of a method for calculating contribution based on a combination of spectral data and training information, it is possible to construct a contribution calculation model by performing supervised learning using quantitative / identified training information instead of using a target contribution map as training information. For example, instead of outputting contribution using equation 3, the parameters of a linear regression using the detection intensity of each band constituting the spectral data as an explanatory variable can be estimated as in equation 7. MLP W () Calculate the weights and biases multiplied by each detection intensity as the parameters of linear regression. For example, w n The index of the band constituting the spectral data is the nth weight, and b represents the bias.

[0135] [Mathematical formula 7]

[0136]

[0137] The weighted sum of the estimated parameters and spectral data is calculated as shown in Formula 8, and supervised learning is performed on the calculation results and the training information for quantification / identification. ^s The index of the sample is the estimated value of s for quantification / identification. The spectral data used in parameter estimation and the spectral data used in the calculation of the weighted sum of the set of quantitative / identification training information can be the same, or different data can be used to improve generalization performance.

[0138] [Mathematical formula 8]

[0139]

[0140] After estimating the parameters using the model of the estimated linear regression parameters learned as described above, the contribution can be calculated by the absolute value of the weight obtained, for example, as in Mathematical Formula 9. The method for calculating the contribution is not limited to this. In the case of quantitative calculation, the change in weight when the positive and negative values ​​of the training information are reversed can be used as the contribution. In the case of binary classification, the change in weight when the recognition value of the training information is reversed in two classes can be used as the contribution. In the case of multi-classification recognition of more than three classes, the contribution obtained by considering each class as the positive class and the other class as the negative class can be calculated, and the contribution of the multi-classification recognition can be obtained by taking the average of the contribution degrees.

[0141] [Mathematical formula 9]

[0142] c n =|w n | (9)

[0143] The automatic wavelength selection unit 703 reads the wavelength band stored in the memory 112, automatically selects it as the wavelength band used in the quantification / identification, or automatically selects the wavelength band used in the quantification / identification based on the wavelength selection information and / or the quantification / identification result obtained in the estimation unit 704, and outputs the selected result as the wavelength selection result. The automatic selection method can select a wavelength band with a contribution degree above a fixed threshold value, or a wavelength band excluding unnecessary wavelength bands or a necessary wavelength band, as in the first embodiment. The threshold value is a hyperparameter that affects the quantification / identification accuracy.

[0144] In addition, the threshold corresponding to the wavelength selection information may be automatically determined by cooperating with the construction of the quantitative / identification model and the quantitative / identification result in the estimation unit 704 described later. The threshold automatic determination method is described after the detailed description of the estimation unit 704. In addition, the wavelength selection information and / or the wavelength selection result may be output to the output unit without passing through the estimation unit, or may be output to the estimation unit.

[0145] The estimation unit 704 constructs a quantitative / identification model with the detection intensity corresponding to the wavelength selection result obtained in the automatic wavelength selection unit 703 as an explanatory variable and the training information as a response variable, and stores it in the memory 112, or reads the quantitative / identification model stored in the memory 112 to perform quantitative / identification of the spectral data. In addition, the quantitative / identification result can also be output to the automatic wavelength selection unit 703. The quantitative / identification model can use a linear regression model such as Logistic regression or Lasso regression with the above-mentioned detection intensity as an explanatory variable, or a machine learning model such as Neural Network, Random Forest, Support Vector Machine, etc.

[0146] The method of preselecting a wavelength in the automatic wavelength selection unit 703 based on the wavelength selection information and / or the quantitative / identification result obtained in the estimation unit 704 is described. As an example, there is a method of automatically determining a contribution threshold value according to the quantitative / identification accuracy. First, the set of spectrum data and training information received by the input unit 701 is divided into training data and verification data, and the wavelength selection information calculation unit 702 calculates the wavelength selection information using the training data.

[0147] In Example 1, for the wavelength selection information obtained, an arbitrary initial value is used as a threshold, and the training data is used to construct a quantitative / identification model with the detection intensity corresponding to the band with a contribution degree above the threshold as an explanatory variable. For the obtained quantitative / identification model, the quantitative / identification accuracy is calculated using the verification data.

[0148] For example, a target accuracy is set in advance. If the calculated quantitative / identification accuracy is less than the target accuracy, the accuracy is improved by lowering the threshold and increasing the number of explanatory variables (bands) used in quantitative / identification. If the calculated quantitative / identification accuracy is above the target accuracy, the threshold is increased to re-construct a quantitative / identification model that achieves the target accuracy and reduces the number of explanatory variables (bands) used in quantitative / identification.

[0149] By repeatedly adjusting the threshold value corresponding to the wavelength selection information and constructing the quantitative / identification model, a band with a minimum number of bands used in the quantitative / identification and a target accuracy can be automatically selected. Alternatively, a method can be adopted in which the maximum number of bands used in the quantitative / identification is pre-set, and the threshold value is repeatedly adjusted and the quantitative / identification model is constructed in the same manner as above to automatically select a band with a maximum number of bands or less and a highest quantitative / identification accuracy.

[0150] Regarding the method of automatically determining the contribution threshold based on the relationship between the threshold and the quantitative / identification accuracy, instead of automatically determining the threshold, the output unit 705 may be used to present the relationship between the quantitative / identification accuracy and the number of bands selected in advance to the user. Figure 8 An example is a graph showing the relationship between the quantification / identification accuracy and the number of bands selected in advance, and it is assumed that the user selects a quantification / identification model from the graph using a mouse or the like via the interface 110 .

[0151] The quantification / identification accuracy and the pre-selected number of wavelengths in the selected quantification / identification model are displayed in "Accuracy" and "Number of Wavelengths." Thus, the user can select a quantification / identification model in consideration of the required quantification / identification accuracy and the pre-selected number of wavelengths.

[0152] The output unit 705 outputs at least one of the wavelength selection information, the selection result, and the quantitative / identification result, and outputs information to assist the user in wavelength selection and / or determination of a quantitative / identification model as necessary.

[0153] In the case of quantification, in addition to the estimated quantitative results of each sample, evaluation values ​​of the quantitative model such as the determination coefficient and RMSE (Root Mean Squared Error) can also be output, and graph data such as a calibration curve can also be output. In the case of recognition, in addition to the estimated recognition results of each sample, evaluation values ​​of the recognition model such as the recognition rate, precision rate, and recall rate can also be output. These outputs are displayed on the display device 122 via the interface 110.

[0154] Example 3

[0155] In Example 3, a spectrum analysis system is described which shortens the analysis time and realizes high-precision quantification / identification by preselecting a wavelength band effective for quantification / identification of spectral data using the spectrum analysis device 700 described in Example 2. The spectrum analysis system described in Example 3 has two stages: a training stage in which the preselected wavelength band is determined using a training data set provided by a user and a quantification / identification model is constructed, and an estimation stage in which only the detection intensity of the wavelength band corresponding to the wavelength selection result is measured for the evaluation sample and quantification / identification is performed using the quantification / identification model constructed in the training stage. Below, after explaining the hardware structure, the details are explained in each stage.

[0156] <Hardware structure of spectrum analysis system>

[0157] exist Fig.10 2 shows a hardware configuration diagram of the third embodiment. The spectrum analysis system 1000 of the third embodiment is composed of a control device 120 , a spectrum measuring device 121 , a spectrum analysis device 700 , and a display device 122 .

[0158] The control device 120 sets the wavelength band of the excitation light irradiated to the sample and / or the detection wavelength band of the light reflected, transmitted, absorbed, or emitted by the sample in the spectrum measuring device 121 described later.

[0159] The spectrum measuring device 121 measures the spectrum corresponding to the wavelength set in the control device 120 for the training sample and the evaluation sample, and inputs the discretized spectrum data to the spectrum analyzing device 700 .

[0160] The spectrum analyzer 700 is the spectrum analyzer described in Example 2. In the training phase, a training data set is constructed based on the spectrum data of the training sample output by the spectrum measuring device 121 and the training information input by the user, and the wavelength band used for quantification / identification is pre-selected and a quantification / identification model is constructed using the training data set. In addition, in the evaluation phase, the wavelength selection result of the wavelength band used for quantification / identification calculated and constructed in the training phase and the quantification / identification model are used to perform quantification / identification using the spectrum data (evaluation data) of the evaluation sample output by the spectrum measuring device 121 as input.

[0161] The display device 122 presents the wavelength selection information and / or the wavelength selection result and / or the quantification / identification result output by the spectrum analysis device 700 to the user.

[0162] Each device operates differently in the training phase and the estimation phase. Therefore, the operation of each device will be described separately for the training phase and the estimation phase.

[0163] <Operation of each device (training phase)>

[0164] The control device 120 sets the wavelength band of the excitation light irradiated on the training sample and / or the detection wavelength band of the light reflected, transmitted, absorbed, or emitted by the sample in the spectrum analyzer 700. In the training phase, all the wavelength bands that may become candidates for the wavelength band used in quantification / identification are set as the measurement wavelength band.

[0165] The spectrometer 121 measures a spectrum for a training sample for obtaining training data, obtains discretized spectrum data, and inputs the discretized spectrum data to the spectrum analyzer 700 .

[0166] The spectrum analyzer 700 receives the spectrum data of the training sample measured by the spectrum measuring device 121 and the training information corresponding to the spectrum data, calculates the wavelength selection information by the method described in Example 2, selects the wavelength band used in the quantification / identification in advance, and constructs the quantification / identification model with the detection intensity corresponding to the wavelength selection result as the explanatory variable. The wavelength selection result is sent to the control device 120 and set as the measurement wavelength band in the estimation stage.

[0167] The display device 122 displays the quantitative / recognition accuracy corresponding to the training data set output by the spectrum analysis device 700, the wavelength selection information shown in FIG. 4, and the Figure 5 The wavelength selection results shown are as follows: Figure 8 Information is shown to assist the user in making wavelength selection and / or quantitative / identification model decisions.

[0168] <Operation of each device (Evaluation phase)>

[0169] The control device 120 sets the wavelength selection result output by the spectrum analyzer 700 in the training phase as the wavelength band of the excitation light irradiated to the evaluation sample and the wavelength bands of the reflected light, transmitted light, absorbed light, etc. detected from the sample.

[0170] The spectrum measuring device 121 measures the spectrum corresponding to the wavelength set by the control device 120 for the evaluation sample, obtains discretized spectrum data by the same method as in the training stage, and inputs it to the spectrum analyzing device 700 .

[0171] The spectrum analysis device 700 receives the spectrum data of the evaluation sample measured by the spectrum measuring device 121, and performs quantification / identification by the method described in Example 1 using the quantification / identification model constructed in the training phase.

[0172] The display device 122 presents to the user the quantitative determination / identification results of the evaluation sample output by the spectrum analysis device 700. In addition, the acquired spectrum and the like may be visually displayed as necessary.

[0173] <Modification>

[0174] The calculation method of wavelength selection information is described in Examples 1 and 2, and the construction method of quantitative / identification model is described in Example 2. However, any data expansion method may be used for the training data set provided by the user to increase the spectral data used in the calculation of wavelength selection information or in the construction of quantitative / identification model. Examples of data expansion methods include adding random noise, weighted linear sum of multiple spectral data, and range change.

[0175] The automatic wavelength selection unit 203 (703) can pre-select the bands based on the wavelength selection information, limiting it to a specific type of band. For example, in the discretized fluorescence fingerprint data 300, the bands can be pre-selected limited to the excitation wavelength, or the bands can be selected limited to the detection wavelength. In addition, in embodiments 1 and 2, the contribution and the calculation of the necessary / unnecessary bands are described in parallel, but one of the processes can be performed first and then the other. For example, the necessary / unnecessary bands can be calculated, and the spectral data after excluding the unnecessary bands from the spectral data can be input to calculate the contribution, etc.

[0176] In the third embodiment, a method of using the spectral data measured by the spectrometer 121 as a training data set is described, but the spectral data prepared in advance may be stored in the memory 112 and used as a part of the training data. In addition, the evaluation data may be stored in the memory 112 in advance, and the accuracy corresponding to the evaluation data of the model newly constructed in the training stage and the model stored in the memory 112, or the number and accuracy of the pre-selected bands, may be compared, and a model with higher accuracy or a model with higher accuracy and a smaller number of pre-selected bands may be determined based on a certain index and stored in the memory 112.

[0177] In each embodiment, the spectrum analyzer 100 (700) can store input signals and output signals, models, etc. of each functional unit of the spectrum analyzer 100 (700) in an external device. For example, the spectrum information, training information, wavelength selection information, wavelength selection results, quantitative / identification results, quantitative / identification models, wavelength selection information calculation models, etc. stored in the memory 112 can be stored in NAS (Network Attached Storage) or cloud storage via the interface 110, and the above-mentioned signals and models stored in NAS or cloud storage can also be received by the interface 110 and stored in the memory 112.

[0178] In each embodiment, after receiving the discretized spectral data, the input unit 201 (701) can change the resolution by performing a zoom-in or zoom-out process using the nearest neighbor interpolation method or the linear interpolation method. For example, in Embodiment 3, after constructing a quantitative / identification model using the discretized fluorescence fingerprint data 300 measured at 1 nm intervals as a training data set, if it is necessary to perform a quantitative / identification on the evaluation data measured at 5 nm intervals, the input unit 201 can perform a zoom-in process on the discretized fluorescence fingerprint data 300 of the evaluation data in a manner that the height (number of excitation bands) and the width (number of detection bands) become 5 times, thereby applying the quantitative / identification model generated using the training data set. The same is true when performing a zoom-out process.

[0179] In each embodiment, the resolution of the spectral data used in the calculation of the contribution map or the necessary / unnecessary band map can be changed by performing an enlargement / reduction process using the nearest neighbor interpolation method or the linear interpolation method on the contribution map or the necessary / unnecessary band map. For example, in Example 3, after calculating the contribution map using the discretized fluorescence fingerprint data 300 measured at 1 nm intervals, in order to make it a contribution map corresponding to the discretized fluorescence fingerprint data measured at 5 nm intervals, the height (number of excitation bands) and the width (number of detection bands) are reduced to 1 / 5, and the band can be pre-selected for the data measured at 5 nm intervals.

[0180] In addition, the present invention is not limited to the above-mentioned embodiments, and includes various modified examples. For example, the above-mentioned embodiments are described in detail in order to explain the present invention in an easy-to-understand manner, and are not limited to all structures that must be described. In addition, a part of the structure of a certain embodiment can be replaced with the structure of other embodiments, and the structure of other embodiments can be added to the structure of a certain embodiment. In addition, for a part of the structure of each embodiment, other structures can be added, deleted, or replaced.

[0181] In addition, for each of the above structures, functions, processing units, etc., part or all of them can be implemented by hardware, for example, by designing in an integrated circuit. In addition, each of the above structures, functions, etc. can also be implemented by software by interpreting and executing programs that implement each function through a processor. Information such as programs, tables, files, etc. that implement each function can be stored in a recording device such as a memory, a hard disk, an SSD (Solid State Drive), or a recording medium such as an IC card or an SD card.

[0182] In addition, the control lines and information lines are shown as necessary for explanation, and not all the control lines and information lines on the product are necessarily shown. In fact, it can be considered that almost all structures are connected to each other.

Claims

1. A spectral analysis device for quantifying / identifying characteristics of a sample based on spectral data of the sample generated using one or more excitation wavelengths, characterized in that: comprising a computing device and a memory for storing a program executed by the computing device, The computing device, Acquire wavelength selection information that indicates the contribution to quantification / identification and / or whether it is necessary for each band of spectral data. receives as input the spectral data obtained from a sample, selecting a wavelength band for quantification / identification of characteristics of the sample from the wavelength bands of the spectral data based on the wavelength selection information, Based on the detection intensity corresponding to the selected result of the wavelength band, the characteristics of the sample are quantified / identified and output.

2. The spectrum analysis device according to claim 1, characterized in that: The computing device outputs at least one of the wavelength selection information and the selection result of the wavelength band.

3. The spectrum analysis device according to claim 1, characterized in that: The computing device calculates the contribution of each wavelength of the spectral data required for quantification / identification based on the received one or more spectral data, and generates the wavelength selection information based on the calculated contribution.

4. The spectrum analysis device according to claim 1, characterized in that: The computing device specifies a section including a noise component in the spectrum data, and includes the specified section in the wavelength selection information as a section unnecessary for quantification / identification.

5. The spectrum analysis device according to claim 1, characterized in that: The computing device uses the accuracy of quantification / identification when quantifying / identifying each band of the spectral data using the detection intensity of the band as an explanatory variable as training information of contribution, and uses multiple groups of training information of contribution corresponding to the spectral data to construct a model for calculating the wavelength selection information.

6. The spectrum analysis device according to claim 1, characterized in that: The arithmetic device constructs a model for calculating the wavelength selection information using a plurality of sets of quantitative / identification training information corresponding to spectral data.

7. The spectrum analysis device according to claim 6, characterized in that: The computing device determines the value of the hyperparameter based on the relationship between the wavelength selection result and the quantitative / identification accuracy when the hyperparameter affecting the quantitative / identification accuracy is changed, or prompts the user with information on the relationship to assist the user in determining the value of the hyperparameter.

8. A spectrum analysis system, characterized in that: include: A control device for setting the wavelength band of the excitation light and the wavelength band of the detection light irradiated on the sample; a spectrometer for measuring spectral data corresponding to the wavelength band set in the control device for the sample; and The spectrum analysis device according to claim 1, The control device controls the spectrometer to measure light in a wavelength band corresponding to the result of wavelength selection obtained by the spectrometer.

9. The spectrum analysis system according to claim 8, characterized in that: Also includes a display device, The spectrometer sends the spectrum data to the spectrum analyzer. The display device receives and displays the analysis result of the spectrum analysis device.

10. A method of spectral analysis, wherein a device is used to quantify / identify the characteristics of a sample based on the spectral data of the sample generated using one or more excitation wavelengths, characterized in that: The device obtains wavelength selection information indicating the contribution to quantification / identification and / or whether it is necessary for each wavelength band of the spectral data, The device receives as input spectral data obtained from a sample, The device selects a wavelength band for quantification / identification of characteristics of the sample from the wavelength bands of the spectral data based on the wavelength selection information, The device quantifies / identifies the characteristics of the sample based on the detection intensity corresponding to the selection result of the wavelength band and outputs it.

11. The spectrum analysis method according to claim 10, characterized in that: The device calculates the contribution of each wavelength of the spectral data required for quantification / identification based on the received one or more spectral data, and generates the wavelength selection information based on the calculated contribution.

12. The spectrum analysis method according to claim 10, characterized in that: The device specifies a section including noise components in the spectrum data, and includes the specified section in the wavelength selection information as a section unnecessary for quantification / identification.

13. The spectrum analysis method according to claim 10, characterized in that: The device uses the accuracy of quantification / identification when quantifying / identifying each band of spectral data using the detection intensity of the band as an explanatory variable as training information of contribution, and uses multiple groups of training information of contribution corresponding to the spectral data to construct a model for calculating the wavelength selection information.

14. The spectrum analysis method according to claim 10, characterized in that: The apparatus constructs a model for calculating the wavelength selection information using multiple sets of quantitative / identification training information corresponding to spectral data.

15. The spectrum analysis method according to claim 14, characterized in that: The device determines the value of the hyperparameter based on the relationship between the wavelength selection result and the quantitative / identification accuracy when the hyperparameter affecting the quantitative / identification accuracy is changed, or prompts the user with information about the relationship to assist the user in determining the value of the hyperparameter.

Images