Spectrum generation device and method, and machine learning model construction device

Abstract

The purpose of the present invention is to provide a spectrum generation device that generates a dataset of a plurality of pseudo-spectra and attribute information for a machine learning model having a spectrum as an input, in order to cause a machine learning model to learn according to a purpose. This spectrum generation device, which generates a set of a pseudo-spectrum and attribute information for constructing a machine learning model, is provided with: an input unit to which a spectrum or a spectrum dataset comprising a set of a spectrum and sample information, and a generation request are input; a basis acquisition unit that acquires a basis spectrum from which a pseudo-spectrum is to be generated; a generation unit that generates a pseudo-spectrum from the basis spectrum; and an output unit that outputs a set of the generated pseudo-spectrum and attribute information.

Description

Spectrum generation device and method, machine learning model construction device

[0001] The present invention relates to a spectrum generation device and method, and a machine learning model construction device.

[0002] The physical property information of a sample can be quantified as a spectrum. For example, a three-dimensional fluorescence spectrum obtained using a spectrofluorometer is a spectrum obtained by irradiating a sample with excitation light and measuring the intensity of the fluorescence emitted by the sample, and the information on the fluorescence properties of the sample is quantified as a two-dimensional array.

[0003] By quantifying the spectra, a machine learning model can be used to construct a model that estimates the properties of a sample. For example, a model that takes a spectrum as input and estimates quantitative values ​​such as the physical properties or compound concentrations of the sample corresponding to the input spectrum can be constructed by supervised learning using a training dataset consisting of multiple pairs of spectra and quantitative values ​​prepared in advance. Models that estimate categories such as the type of sample can also be constructed in a similar manner.

[0004] Spectra may contain information other than the desired properties of a sample, as well as noise signals that show no correlation with the target quantitative value or category. Furthermore, because spectra tend to have high correlations between adjacent elements (collinearity), when training a machine learning model using each spectral element as an explanatory variable, optimization of parameters related to highly correlated elements may not be stable. Therefore, feature values ​​may be extracted from spectra depending on the purpose.

[0005] A machine learning model may also be used for the model that extracts this feature. The feature may have a form depending on the purpose, such as information on the peak coordinates and peak values ​​of the spectrum obtained by removing noise signals from the spectrum, or a feature vector with a lower dimension than the spectrum. For example, a machine learning model that estimates information on the peak coordinates and peak values ​​of the spectrum can be constructed by supervised learning using a training data set that is composed of a previously prepared spectrum and pairs of peak coordinates and peak values ​​contained in the spectrum.

[0006] To build a machine learning model for the above spectra, a training dataset is required to optimize the parameters of the machine learning model. However, in the case of spectra, there are many steps involved, from sample preparation to measurement, making it often difficult to collect a sufficient amount of training dataset.

[0007] Patent Document 1 discloses a method for rapidly increasing the number of training data sets without actually preparing samples, by separating a spectrum (original spectrum) obtained from a sample into an analytical spectrum and a non-analytical spectrum, subjecting the non-analytical spectrum to various variations, and generating a virtual spectrum by adding the varied non-analytical spectrum to the analytical spectrum.

[0008] Japanese Patent Application Laid-Open No. 2023-74746

[0009] The technology described in Patent Document 1 uses frequency separation technology to decompose a spectrum (raw spectrum) obtained from a sample into a spectrum (analytical spectrum) that serves as the basis for analyzing the sample and a non-analytical spectrum that is different from the analytical spectrum. Patent Document 1 then generates a virtual spectrum by subjecting the non-analytical spectrum to various variations and adding it to the analytical spectrum. Furthermore, Patent Document 1 links the non-analytical spectrum to quantitative values ​​and categories that represent the properties of the sample, which have been prepared in advance, thereby increasing the amount of training data used to train a machine learning model that estimates the quantitative values ​​and categories.

[0010] However, since the number of patterns of the analytical spectrum itself does not increase, there is a possibility that overlearning may occur due to an insufficient number of patterns in the machine learning model that estimates the quantitative values ​​and categories.

[0011] Furthermore, in addition to the machine learning models that estimate the quantitative values ​​and categories mentioned above, there are also other machine learning models for spectra that have purposes other than estimating quantitative values ​​and categories associated with samples, such as models that estimate the results after removing noise signals from a spectrum, models that estimate spectral peak coordinates and peak values, and models that compress a spectrum into a feature vector with lower dimensionality than the spectrum.

[0012] However, Patent Document 1 does not mention a method for generating or learning training data for learning these models. For example, in the case of such models, comprehensive learning of all possible spectral shape patterns, not just the pattern of the analysis spectrum obtained from the original spectrum, leads to the construction of a generalized model, but the method described in Patent Document 1 does not provide an adequate number of training data patterns.

[0013] The present invention has been made in consideration of the above-mentioned problems, and aims to provide a spectrum generation device and method, and a machine learning model construction device, which generate a data set of multiple pseudospectra and attribute information for training a machine learning model according to a purpose for a machine learning model that uses a spectrum as input.

[0014] In order to solve the above-described problems, according to one aspect of the present invention, a spectrum generation device for generating pairs of pseudospectrums and attribute information for constructing a machine learning model using spectra as explanatory variables includes an input unit that receives a spectral dataset consisting of one or more spectra or one or more pairs of spectra and sample information, and a generation request for adjusting the generation result; a basis acquisition unit that acquires one or more basis spectra that are components of the pseudospectrum to be generated; a generation unit that sets weights on the one or more basis spectra obtained by the basis acquisition unit and generates, from the weighted basis spectra, a pseudospectrum that is different from the spectrum input to the input unit; and an output unit that outputs one or more pairs of the generated pseudospectrum and attribute information, wherein the basis spectra acquired by the basis acquisition unit include at least one or more factor spectra that are components of the spectrum input to the input unit, or spectra obtained by applying a predetermined variation process to a factor spectrum, and the attribute information that is associated with the pseudospectrum and output from the output unit includes at least one of information on the basis spectra, the predetermined variation process, and the weight used to generate the pseudospectrum.

[0015] According to the present invention, it is possible to generate a pair of pseudospectrum and attribute information for constructing a machine learning model using a spectrum as an explanatory variable.

[0016] FIG. 1 is a hardware configuration diagram of a spectrum generating device; FIG. 2 is a functional block diagram of a spectrum generating device; FIG. 3 is a table showing an example of a spectrum; FIG. 4 is a diagram showing an example of a spectrum; FIG. 5 is an explanatory diagram showing an example of a set of a pseudospectrum and attribute information; FIG. 6 is a processing flow of a spectrum generating device; FIG. 7 is a functional block diagram of a spectrum generating device according to a second embodiment; FIG. 8 is a graph showing an example of a verification result; and FIG. 9 is another graph showing an example of a verification result.

[0017] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The spectrum generating device according to this embodiment generates many pseudospectrums by extracting, expanding, and reconstructing characteristic basis spectra from a small number of real spectra. A real spectrum is a spectrum obtained by actually measuring a sample.

[0018] The spectrum generation device of this embodiment, for example, acquires one or more basis spectra from a spectral dataset consisting of one or more spectra or one or more pairs of spectra and sample information, sets weights to each basis spectrum, and calculates a weighted sum of the basis spectra and the weights, thereby outputting one or more pairs of pseudospectrums with shapes different from those of the spectra included in the spectral dataset and attribute information associated with the pseudospectrum, which can be used for training various machine learning models.

[0019] The spectrum generating device of this embodiment includes, for example, an arithmetic unit and a memory that stores a computer program executed by the arithmetic unit. The arithmetic unit receives the spectral dataset as input and outputs one or more sets of pseudospectra and attribute information.

[0020] The spectrum generation device of this embodiment can generate pseudospectrums with a large number of patterns necessary for constructing various machine learning models according to the purpose.

[0021] The spectrum generation device of this embodiment can also construct a training pseudospectral dataset and a validation pseudospectral dataset using one or more pairs of pseudospectrum and attribute information to train a target machine learning model, and train the target machine learning model using the training pseudospectral dataset. The spectrum generation device verifies the trained machine learning model using the validation pseudospectral dataset and recursively generates additional training pseudospectral datasets depending on the verification results. Use of this spectrum generation device can provide a machine learning model construction system that additionally trains a machine learning model and outputs the target machine learning model and the verification results.

[0022] The first embodiment will be described with reference to FIGS.

[0023] <Hardware Configuration of Spectrum Generating Device> In the first embodiment, a set of pseudospectrums and attribute information that can be used as training data for various machine learning models is generated based on a spectrum data set including a small number of spectra.

[0024] 1, the hardware configuration of a spectrum generating device 10 will be described. The spectrum generating device 10 includes, for example, a memory 11, a computing device 12, an interface 13, and a bus 14. The memory 11, the computing device 12, and the interface 13 transmit and receive information via the bus 14.

[0025] The components of the spectrum generating device 10 will now be described. The interface 13 is a communication device that transmits and receives signals to and from devices 20, 21, and 22 external to the spectrum generating device 10. Devices that transmit and receive signals to and from the interface 13 include, for example, a spectrum measuring device 21 that measures the spectrum of a sample and outputs the measurement results, a control device 20 that controls the spectrum measuring device 21, and an output device 22 that outputs the processing results from the spectrum generating device 10. The output device 22 is, for example, a monitor display or a printer. The output device 22 may be configured as a head-mounted display. An input device (not shown) that allows a user to input instructions, data, etc. may be connected to the spectrum generating device 10.

[0026] The method for transmitting and receiving signals between these external devices 20, 21, and 22 and the spectrum generating device 10 may be a wired connection using a cable such as optical fiber, or a wireless connection using a wireless communication technology such as Bluetooth (registered trademark).

[0027] The spectrum generating device 10 does not necessarily have to be installed in the same facility as the control device 20, the spectrum measuring device 21, and the output device 22, but may be installed on another computer or in a so-called cloud. In such a case, signals may be transmitted and received between the interface 13 and the external devices 20, 21, and 22 using, for example, a local area network (LAN) or a wide area network (WAN).

[0028] The arithmetic unit 12 is a device that executes various processes within the spectrum generating device 10, and is, for example, a CPU (Central Processing Unit) or an FPGA (Field-Programmable Gate Array). The functions executed by the arithmetic unit 12 will be described later.

[0029] The memory 11 is a device that stores programs executed by the calculation device 12, parameters of quantitative / discrimination models, processing results, etc., and is, for example, a hard disk drive (HDD), flash memory, solid state drive (SSD), random access memory (RAM), read only memory (ROM), etc.

[0030] <Functional Configuration of Spectrum Generating Device 10> Figure 2 is a functional block diagram of the spectrum generating device 10. Each of the illustrated functional units 101, 102, 103, and 104 may be implemented by a calculation device 12 that executes a predetermined computer program, or by dedicated hardware. The spectrum generating device 10 includes, as functional units, an input unit 101, a basis acquisition unit 102, a generation unit 103, and an output unit 104, for example. Each functional unit will be described below.

[0031] The input unit 101 accepts one or more spectra or one or more pairs of a spectrum and a label associated with sample information, input from the interface 13. The input unit 101 also accepts a generation request for adjusting the pseudospectrum to be generated. The spectrum is formatted as one or more dimensional array data consisting of one or more axes, and each element stores the corresponding spectral intensity.

[0032] The basis acquisition unit 102 acquires one or more basis spectra, which are components of a spectrum, based on one or more spectra or one or more pairs of spectra and labels input to the input unit 101. The basis spectra may include at least one or more basis vectors based on the spectra input to the input unit 101 or vectors obtained by performing any perturbation processing on a basis vector.

[0033] The generator 103 generates a pseudospectrum by setting weights on the basis spectra acquired by the basis acquisition unit 102 and calculating a weighted sum of the basis spectra and the weights, and also acquires attribute information associated with the pseudospectrum. The attribute information includes at least one of information on the basis spectra used when generating the pseudospectrum, the details of the variation process, and the set weights.

[0034] The output unit 104 outputs one or more pairs of pseudospectrum and attribute information to the output device 22.

[0035] It should be noted that the above functions do not need to be configured exactly as the functional units 101 to 104 shown in FIG. 2, but rather it is sufficient if the processing corresponding to the operation of each functional unit can be realized.

[0036] 6 shows an example of a processing flow diagram, in which steps S1 to S4 correspond to the functional units 101 to 104 shown in FIG.

[0037] The input step S1 accepts one or more spectra or one or more pairs of a spectrum and a label associated with the spectrum, and a generation request for adjusting the pseudospectrum to be generated, input from the interface 13. The spectrum is formatted as one or more dimensional array data consisting of one or more axes, and each element stores the corresponding spectral intensity.

[0038] In the basis acquisition step S2, one or more basis spectra, which are components of the spectrum, are acquired based on the one or more spectra received in the input step S1. The basis spectra include at least one or more basis vectors based on the spectrum input to the input unit 101 or vectors obtained by performing any perturbation processing on a basis vector.

[0039] In the generating step S3, a weight is assigned to the basis spectrum acquired in the basis acquiring step S2, and a weighted sum is calculated from the basis spectrum and the weight to generate a pseudospectrum and acquire attribute information associated with the pseudospectrum. The attribute information includes at least one of information on the basis spectrum used to generate the pseudospectrum, the applied variation processing, and the set weight.

[0040] In the output step S4, one or more pairs of pseudospectrum and attribute information are output.

[0041] <Configuration and Operation of Each Unit> The operations of the input unit 101, the basis acquisition unit 102, the generation unit 103, and the output unit 104 will be described in detail below.

[0042] The input unit 101 receives one or more spectra or one or more pairs of spectra and labels associated with the spectra, and a generation request via the interface 13. A spectrum refers to a multidimensional array composed of one or more axes. For example, in the case of a three-dimensional fluorescence spectrum, it is a two-dimensional array composed of axes of excitation wavelength and emission wavelength, and each element stores emission intensity. The spectrum or the pair of spectrum and label received by the input unit 101 is output to the basis acquisition unit 102, or is output to the basis acquisition unit 102 while being stored in the memory 11. The generation request refers to structure data including parameters for adjusting the pseudospectrum to be generated.

[0043] Fig. 3 shows a three-dimensional fluorescence spectrum as an example of a spectrum. Three-dimensional fluorescence spectrum 30 is an example of a three-dimensional fluorescence spectrum. Fig. 4 shows a visualized example 31 of a three-dimensional fluorescence spectrum. In the visualized example 31 of a three-dimensional fluorescence spectrum, the vertical axis represents the excitation wavelength and the horizontal axis represents the emission wavelength, and the emission intensity for each combination of excitation and emission wavelength is displayed as contour data.

[0044] As shown in the three-dimensional fluorescence spectrum 30, the excitation wavelength, emission wavelength, and emission intensity are recorded with an arbitrary quantization interval. The three-dimensional fluorescence spectrum 30 shows an example in which excitation light / emission light was measured from 200 nm to 700 nm with a quantization interval of 5 nm. For example, the emission intensity of 200 nm detection light in response to 200 nm excitation light is recorded in the first row and first column of the three-dimensional fluorescence spectrum 30; the emission intensity of 205 nm detection light in response to 200 nm excitation light is recorded in the first row and second column; and the emission intensity of 200 nm detection light in response to 205 nm excitation light is recorded in the second row and first column. The number of columns and rows of the three-dimensional fluorescence spectrum 30 is 100 × 100, since 500 nm wavelengths are recorded at 5 nm intervals.

[0045] A three-dimensional fluorescence spectrum visualization example 31 is an example in which a three-dimensional fluorescence spectrum 30 is visualized as a two-dimensional image in the form of contour line data. When humans observe a three-dimensional fluorescence spectrum, they often confirm it using a display such as the three-dimensional fluorescence spectrum visualization example 31. In addition to contour line data, the spectrum may also be displayed as a gray image or heat map image in which emission intensity is assigned to a brightness value or color, or as a bird's-eye view. The shape of the three-dimensional fluorescence spectrum varies depending on the nature of the target fluorescence characteristic.

[0046] The generation request is a parameter that allows an external device (user) to adjust the settings of the input unit 101, the basis acquisition unit 102, and the generation unit 103. The spectrum generation device 10 changes various settings in accordance with the generation request. The generation request is set by the user using a mouse, keyboard, or the like from the control device 20 via the interface 13.

[0047] The input unit 101 may perform any preprocessing on the spectrum when outputting it to the basis acquisition unit 102 or when saving it in the memory 11 and outputting it to the basis acquisition unit 102 .

[0048] In the three-dimensional fluorescence spectrum visualization example 31, the three-dimensional fluorescence spectrum contains noise signals. For example, a wavelength band including Rayleigh scattering is generally considered to be a noise component when analyzing fluorescence characteristics from a three-dimensional fluorescence spectrum. Rayleigh scattering is not fluorescence emitted by the sample, but light reflected from the sample. By detecting Rayleigh scattering, features unrelated to the fluorescence characteristics are reflected in the three-dimensional fluorescence spectrum.

[0049] Rayleigh scattering is characterized by the effect of the diffraction grating, which results in detection in a region near a wavelength band where the excitation wavelength is a natural number multiple or the reciprocal multiple of the emission wavelength. Therefore, the influence of noise can be suppressed by preprocessing, such as by setting an arbitrary wavelength band and converting the emission intensity of the wavelength matching the above condition and the surrounding bandwidth to an arbitrary fixed value such as 0, or by measuring the 3D fluorescence spectrum of a standard sample and subtracting the 3D fluorescence spectrum of the standard sample from each spectrum. Preprocessing can also be performed in any manner to reduce Raman scattering and other artifacts.

[0050] In the method disclosed in this embodiment, the spectrum generation device 10 acquires one or more basis spectra, which are components of the spectrum, based on an input spectrum, and assigns various weights to the basis spectra.The spectrum generation device 10 then outputs one or more pairs of a pseudospectrum represented by a weighted sum of the multiple basis spectra and attribute information associated with the pseudospectrum.

[0051] The spectrum generation device 10 of this embodiment acquires basis spectra and changes weights associated with the basis spectra to generate pseudospectra of various patterns while outputting attribute information associated with the pseudospectra. This enables the spectrum generation device 10 to flexibly create training datasets for various machine learning models according to their purposes.

[0052] The basis acquisition unit 102 receives one or more spectra from the input unit 101 or the memory 11 and acquires basis spectra. Here, a basis spectrum refers to a component of a pseudospectrum to be generated, and corresponds to a basis vector when data that can be generated is expressed in vector space. The spectrum generation device 10 acquires various basis spectra based on a spectral data set and changes the weights associated with the basis spectra, thereby enabling the generation of a variety of pseudospectrums.

[0053] A method for acquiring a basis spectrum will be described below. A plurality of methods for acquiring a basis spectrum may be used. Hereinafter, a list of acquired basis spectra will be referred to as a basis spectrum list.

[0054] As an example of a method for acquiring a basis spectrum, a method will be described in which a factor spectrum is acquired based on the spectrum received by the input unit 101 and the factor spectrum is used as a basis spectrum. Here, the factor spectrum corresponds to a basis vector when the spectrum received by the input unit 101 is regarded as a vector, and represents a spectrum that is a component of the spectrum received by the input unit 101.

[0055] As an example of a method for obtaining a factor spectrum, a method for obtaining a factor spectrum for a three-dimensional fluorescence spectrum using PARAFAC (PARArell Factor analysis), which is a type of tensor decomposition method, will be described. PARAFAC utilizes the property that the relationship between the concentration of fluorescent molecules contained in a sample and the fluorescence intensity is linear, and expresses multiple three-dimensional fluorescence spectra using a trilinear model described in Equation 1.

[0056] In equation 1, F' ｋ is the number of excitation wavelength bands E ｘ , number of emission wavelength bands E ｍ K three-dimensional fluorescence spectra F １ , F ２ ,... ,F Ｋ The kth sample F ｋ is a tensor expressed in a trilinear model. N is a hyperparameter that represents the number of types of fluorescent molecules contained in the sample. ｎ is the emission spectrum specific to the nth fluorescent molecule from 1 to N, and E ｘ The emission spectrum is a vector of dimension b. The emission spectrum is a spectrum that shows the distribution of light emitted at each excitation wavelength when a fluorescent molecule is irradiated with excitation light at a fixed emission wavelength. ｎ is the excitation spectrum specific to the nth fluorescent molecule, and E ｍ The excitation spectrum is a c-dimensional vector. The excitation spectrum is a spectrum that shows the distribution of light emitted by fluorescent molecules when excited by excitation light of a fixed wavelength at each emission wavelength. ｋｎ denotes the relative concentration of the nth fluorescent molecule contained in the specimen corresponding to the kth sample.

[0057] The circled × (hereafter (×)) indicates the operation of the direct product of vectors. ｎ (×) b ｎ is the E ｘ ×E ｍ The matrix is ​​obtained by multiplying the matrix by the relative concentration, c ｋｎ (a ｎ (×) b ｎ) represents the signal produced by the nth fluorescent molecule in the kth sample. ｋ is the sum F' of the first to Nth signals. ｋ This is the modeled formula.

[0058] In PARAFAC, the emission spectrum a １ ,... ,a Ｎ Each element of the excitation spectrum b １ ,... ,b Ｎ and the relative concentration c of each fluorescent molecule in each sample １１ ,... ,c １Ｎ ,... ,c ＫＮ Using the variables, each variable is optimized based on the objective function shown in Equation 2. Equation 2 expresses the three-dimensional fluorescence spectrum F １ , F ２ ,... ,F Ｋ and the tensor F' expressed in trilinear-model １ , F' ２ ,... ,F' Ｋ It means minimizing the mean square error of each element of F. ｉｊｋ denotes the emission intensity of the kth sample at the ith excitation wavelength and the jth emission wavelength. ｉｎ is the emission spectrum of the nth fluorescent molecule a ｎ b indicates the value of the ith excitation wavelength in ｊｎ is the excitation spectrum of the nth fluorescent molecule b ｎ denotes the value of the j-th emission wavelength in

[0059] The objective function shown in Equation 2 may be optimized by any optimization method, such as ALS (Alternating Least Square) or steepest descent.

[0060] The optimized results of the emission spectrum, excitation spectrum, and relative concentration obtained by Equation 2 may be normalized to unify the maximum values ​​of the emission spectrum and excitation spectrum, and the scale of the relative concentration may be changed in accordance with the normalization.

[0061] An arbitrary regularization term may be added to the objective function shown in Equation 2. For example, a feature extraction function may be defined that takes each three-dimensional fluorescence spectrum or a tensor expressed using a trilinear model as input, and the squared error of the output of the feature extraction function may be added as a regularization term to Equation 2 for pairs of tensors expressed using the trilinear model corresponding to each three-dimensional fluorescence spectrum. The feature extraction function may load from memory 112 an arbitrary optimized model that estimates the compound concentration or compound type of the three-dimensional fluorescence spectrum. Adding such a regularization term to Equation 2 makes it possible to perform regularization that makes elements that can be inferred from the three-dimensional fluorescence spectrum, such as the compound concentration or compound type, similar when optimizing the tensor expressed using the trilinear model to resemble the three-dimensional fluorescence spectrum.

[0062] The number N of fluorescent molecules in Equations 1 and 2 is originally unknown. Therefore, PARAFAC results for multiple patterns of N are obtained, scores are calculated for the results, and the PARAFAC result corresponding to N with the highest score may be used as the final PARAFAC result, or the PARAFAC result corresponding to one pattern of N may be used as the final PARAFAC result. Core consistency, which is widely used in PARAFAC, may be used as the scoring function.

[0063] The direct product of the optimized emission spectrum and excitation spectrum a １ (×) b １ ,... ,a Ｎ (×) b Ｎ is a collection of three-dimensional fluorescence spectra that represent the specific fluorescence characteristics of N types of fluorescent molecules, and is a factor spectrum that constitutes the three-dimensional fluorescence spectrum received by the input unit 101. This factor spectrum is added to the basis spectrum list as a basis spectrum.

[0064] The optimization result of the relative concentration obtained by PARAFAC may be linked to the factor spectrum and stored in the memory 11. When the input unit 101 receives a pair of a spectrum and a label, the label information may also be stored in the memory 11 together with the optimization result of the relative concentration.

[0065] Another method for obtaining factor spectra will be described, which uses a spectrum obtained from a standard sample composed of a single type of fluorescent molecule. Because each fluorescent molecule has its own unique fluorescent properties, the resulting three-dimensional fluorescence spectrum for a sample composed of a single type of fluorescent molecule will be a two-dimensional matrix represented by a single factor spectrum whose scale changes linearly with the concentration of the fluorescent molecule. Therefore, if the three-dimensional fluorescence spectrum input to the input unit is obtained from a sample composed of a single type of fluorescent molecule, the three-dimensional fluorescence spectrum itself may be considered a factor spectrum and added to the basis spectrum list. When the input unit 101 receives a pair of a spectrum of a sample composed of a single type of fluorescent molecule and a label containing sample information such as the fluorescent molecule name and structural formula, the label information may also be stored in the memory 11.

[0066] The results of applying any variation processing to a base spectrum stored in the basis spectrum list may be added to the basis spectrum list. The factor spectrum described above is the result of acquiring the components of the spectrum received by the input unit 101. If the subsequent generation unit 103 uses only the factor spectrum, a pseudospectrum similar in shape to the spectrum but with different weights applied to each factor spectrum is generated. If variation processing is applied to the basis spectrum, a pseudospectrum with components different from those of the spectrum can be generated. This enables the machine learning model to learn pseudospectra with more diverse patterns than conventional spectrum generation methods, leading to the realization of a versatile machine learning model. Examples of variations in the basis spectrum are described below.

[0067] For example, a variation process may be performed to shift the basis spectrum in an arbitrary direction. Depending on the spectrum measurement environment and the state of the sample, some of the factor spectra constituting the three-dimensional fluorescence spectrum may shift in the direction of shortening the emission wavelength. As such, the entire factor spectrum may shift under various conditions. If the machine learning model has not learned a spectrum in which such a shift has occurred, accurate inference may not be performed when the spectrum in which the shift has occurred is input during inference. Therefore, by shifting the basis spectrum in an arbitrary direction, the subsequent generation unit 103 can simulate the shift phenomenon of the factor spectrum observed in a real spectrum. The shifted basis spectrum will have missing values ​​at the end in the direction opposite to the shift direction. However, this may be interpolated with, for example, "0" or extrapolated using an arbitrary model such as a Gaussian process regression model.

[0068] For example, a variation process may be performed by applying a monotonically increasing function to each element of the base spectrum. When a monotonically increasing function is applied to each element of the base spectrum, the shape of the base spectrum can be changed within a range that does not change the magnitude relationship between the elements. For example, in a three-dimensional fluorescence spectrum, the emission intensity at the i-th excitation wavelength and the j-th emission wavelength of a certain base spectrum B is expressed as B ｉｊ When a monotonically increasing function is f(), the variation process may be performed in the form of Equation 3. ｉｊ is the result of the variation processing, and Equation 3 applies a monotonically increasing function f() to each element of the base spectrum, and normalization is performed so that the maximum value of the application result is the same as the maximum value before application. f() is an exponential function exp(B ｉｊ ) or a logarithmic function ln(B ｉｊ +1) is also acceptable.

[0069] When performing the variation processing described above, if the base spectrum used in the variation processing is a factor spectrum and is associated with relative concentration information or label information, the base spectrum that has undergone the variation processing may also be associated with the above relative concentration and label information and stored in memory 11.

[0070] The results of recursively applying the variation process to the basis spectrum by the variation process described above may also be added to the basis spectrum list.

[0071] Log information on the creation process of each acquired base spectrum is stored in memory 11. For example, the information includes information on whether the base spectrum is a factor spectrum, information on the spectrum used to acquire the factor spectrum if it is a factor spectrum, and information on the variation process if it is not a factor spectrum.

[0072] The method for obtaining the base spectrum may be set to fixed conditions in advance, or may be arbitrarily adjusted by an external party (user) using a generation request.

[0073] The generator 103 generates a pseudospectrum by setting weights for the basis spectra acquired by the basis acquisition unit 102 and calculating a weighted sum of the basis spectra and the weights. Using various basis spectra acquired by the basis acquisition unit 102 makes it possible to generate pseudospectrums of various patterns.

[0074] The weighted sum is expressed by Equation 4. In Equation 4, F' is the pseudospectrum, B ｍ is the mth basis spectrum in the basis spectrum list, w ｍ represents the weight applied to the m-th basis spectrum. E represents noise, and any high-frequency noise or low-frequency noise may be used, or no noise may be generated, i.e., E = 0. An example of how to set the weight is described below.

[0075] For example, a fixed weight candidate may be set for each basis spectrum, and a combination of the weight candidates for the basis spectra may be set. For example, when there are M basis spectra, if 11 weight candidates are set for each basis spectrum in increments of 0.1 in the range of 0 to 1, i.e., [0.0, 0.1, . . . , 1.0], the number of combinations of weight candidates is 11 to the Mth power, and a pseudospectrum is generated for each combination according to Equation 4. Furthermore, in order to avoid using basis spectra associated with the same factor spectrum, the weights of all but one of the factor spectra themselves or the basis spectra obtained by applying variation processing to the factor spectra may be set to 0.

[0076] For example, weights may be set according to any probability distribution. For example, weights may be set using a multidimensional normal distribution as shown in Equation 5. In Equation 5, N() represents an M-dimensional normal distribution with an M-dimensional mean vector μ and an M×M variance-covariance matrix Σ as input. w represents an M-dimensional weight vector generated according to the M-dimensional normal distribution, and the m-th element is w in Equation 4. ｍ In the input of the multivariate normal distribution in Equation 5, the mean vector and variance-covariance matrix may be set randomly or may be set to any fixed value.

[0077] When the relative concentration information is associated with each basis spectrum, the mean, variance, and covariance of the relative concentrations may be calculated and used as the mean vector and variance-covariance matrix of Equation 5. By adopting this method, a combination of weights is generated based on the tendency of the relative concentration information associated with the actual spectrum obtained during tensor decomposition, and therefore the generated result of Equation 4 can be a spectrum that has not actually been observed, and at the same time, a pseudospectrum that naturally reflects the actual fluorescence characteristics.

[0078] The weights may be set in advance as fixed conditions, or may be arbitrarily adjusted by an external party (user) using a generation request.

[0079] The pseudospectrum generated as described above is associated with attribute information used to generate the pseudospectrum. The attribute information includes the basis spectra acquired by the basis acquisition unit 102, information associated with the basis spectra stored in the memory 11, such as relative concentration information and log information, weights set by the generation unit 103, and a method for setting the weights, such as a probability distribution.

[0080] In this way, the spectrum generation device 10 can easily prepare a data set to be used for training a desired machine learning model by linking and managing various pieces of information resulting from the generation of pseudospectrums. Furthermore, the spectrum generation device 10 can also use the attribute information itself for training and verifying a machine learning model. Examples of using attribute information for building and verifying a machine learning model will be described later.

[0081] The output unit 104 outputs one or more pairs of pseudospectrum and attribute information generated by the generation unit 103. The pairs of pseudospectrum and attribute information are output as structure data.

[0082] The output unit 104 may display the output result on the output device 22 via the interface 13 .

[0083] FIG. 5 shows a generated result display example 40 presented to the user. The generated result display example 40 includes, for example, a pseudospectrum visualization example 41, a base spectrum visualization example 42, and an attribute information display example 43. The pseudospectrum visualization example 41 and the base spectrum visualization example 42 are visualized using contour lines, similar to the three-dimensional fluorescence spectrum visualization example 30. The attribute information display example 43 displays the attribute information in text format. Specific names and numerical values ​​are entered in the "..." boxes in the attribute display example 43. For example, for "acquisition method," information such as "PARAFAC" or "single fluorescent molecule sample" is displayed. The generated result display example 40 shown in FIG. 5 is just one example, and any display format that allows observation of the pseudospectrum, base spectrum, and attribute information can be used.

[0084] Example 2 will be described using FIGS. 7 to 9 . The differences between Example 2 and Example 1 will be mainly described. In this example, the spectrum generation device 10 described in Example 1 is utilized to generate a training pseudospectral dataset and a validation pseudospectral dataset required for training a desired machine learning model. A machine learning model construction device 50 of this example trains a machine learning model using the training pseudospectral dataset and validates the trained machine learning model using the validation pseudospectral dataset. The machine learning model construction device 50 generates another training pseudospectral dataset based on the validation results, and performs additional training. Alternatively, the machine learning model construction device 50 efficiently provides a desired machine learning model by outputting the trained machine learning model and the validation results.

[0085] <Hardware Configuration of Spectral Analysis System> The machine learning model construction device 50 according to this embodiment is configured as a computer system including, for example, a memory, a calculation device, and an interface (none of which are shown), similar to the hardware configuration described in Fig. 1. The machine learning model construction device 50 can be connected to the control device 20, the spectrum measurement device 21, and the output device 22, as described in Fig. 1.

[0086] 7 is an example of a functional block diagram of a machine learning model construction device 50. Each of the illustrated functional units 51, 52, 53, and 54 may be realized by a computing device that executes a predetermined computer program, or may be realized by dedicated hardware.

[0087] The machine learning model construction device 50 includes, as functional units, for example, a design unit 51, a learning unit 52, a verification unit 53, and a model output unit 54, and incorporates the spectrum generation device 10. Each functional unit will be described below.

[0088] The design unit 51 receives a function request input from an external device. Based on the received function request, the design unit 51 inputs a generation request to the spectrum generation device 10, to which a spectrum dataset has been input in advance. The design unit 51 constructs a training pseudospectrum dataset and a validation pseudospectrum dataset using a pair of pseudospectrum and attribute information output from the spectrum generation device 10. Furthermore, the design unit 51 uses the training pseudospectrum dataset and the validation pseudospectrum dataset to design a machine learning model and an objective function.

[0089] The learning unit 52 uses the training pseudospectral data set to train a machine learning model.

[0090] The verification unit 53 uses the verification pseudospectral dataset to verify the performance of the machine learning model trained by the learning unit 52. Depending on the verification result of the performance of the machine learning model, the verification unit 53 outputs a function requirement to the design unit 51, causing the design unit 51 to reconstruct a training pseudospectral dataset and perform additional learning, or to determine a final machine learning model and the verification result of the final machine learning model.

[0091] The model output unit 54 outputs the final machine learning model and the verification results.

[0092] The above functions do not need to be configured exactly as the functional units in FIG. 7, but only need to be able to realize processing corresponding to the operation of each functional block.

[0093] <Configuration and Operation of Each Unit> The operations of the design unit 51, learning unit 52, verification unit 53, and model output unit 54 will be described in detail below.

[0094] The design unit 51 receives a function request, for example, input from an interface (not shown). Based on the received function request, the design unit 51 inputs a generation request to the spectrum generation device 10, to which a spectrum dataset has been input in advance. The design unit 51 constructs a training pseudospectrum dataset and a validation pseudospectrum dataset using a pair of pseudospectrums and attribute information that are the output of the spectrum generation device 10. The design unit 51 uses these constructed datasets to design a machine learning model and an objective function.

[0095] The functional requirements may include functions that the machine learning model must have, such as performing dimensionality reduction, detecting peak coordinates, estimating noise-removed spectra, etc. A machine learning model is designed to meet these objectives.

[0096] The machine learning model may be a neural network such as a Convolutional Neural Network (CNN) or a Transformer, or a Support Vector Machine (SVM), depending on the purpose.

[0097] For the machine learning model described above, an objective function that utilizes attribute information is designed. With actual spectra, information such as the basis spectrum and weights is unknown and difficult to identify. On the other hand, with pseudospectrums, attribute information that represents the properties of the spectrum can be used in machine learning, making it possible to perform machine learning that is more in line with the properties of the spectrum than with actual spectra.

[0098] For example, in the case of a machine learning model that performs dimensionality reduction on three-dimensional fluorescence spectra, an objective function such as that shown in Equation 6 may be designed that uses weight information applied to the base spectra that are components of the pseudospectrum. In Equation 6, F' and F'' represent a pair of pseudospectrums, σθ represents a parameter θ, and is a machine learning model that aims for dimensionality reduction with the pseudospectrum as the input and the vector as the output. w' and w'' represent vectors whose elements are weights applied to the base spectra that make up F' and F'', and d() represents a distance function. Any function including Euclidean distance may be used as the distance function.

[0099] Equation 6 is a function intended to match the distance between vectors representing the dimension reduction result and the distance between vectors representing weights for pairs of pseudospectra. In other words, pairs with similar pseudospectral weight patterns are trained to have similar dimension reduction results. The weight pattern applied to the basis spectra represents the proportion of the basis spectra, which are components of the spectrum, contained in the spectrum. Compared to comparing the difference between the distance between the pseudospectral pair itself and the distance between the dimension reduction result, rather than the weight pattern, this is not influenced by factors other than the basis spectra, such as noise, and therefore makes it possible to train a dimension reduction method that is more in line with the characteristics of the spectrum.

[0100] For example, in the case of a machine learning model that detects peak coordinates, the peak coordinates for each base spectrum in the attribute information may be calculated and the objective function may be set to detect these. By designing the objective function to detect peak coordinates derived from the base spectrum, it becomes possible to perform learning that does not detect peak coordinates derived from noise.

[0101] For example, in the case of a machine learning model that estimates a spectrum from which noise has been removed, an objective function may be designed that estimates the result of subtracting noise in the attribute information from the pseudospectrum.

[0102] Furthermore, the functional requirements may include required performance based on attribute information for the machine learning model.

[0103] For example, in the case of a machine learning model that performs dimensionality reduction, it may be set that the squared error between the distance function value of the pair of weight patterns in Equation 6 and the distance function value of the pair of compression results is less than a fixed threshold for each sample in the validation pseudospectral dataset.

[0104] For example, in the case of a machine learning model that detects peak coordinates, the allowable estimation error of the estimated peak coordinates may be set as the required performance.

[0105] For example, in the case of a machine learning model that estimates a spectrum from which noise has been removed, noise removal performance such as an S / N ratio may be set as the required performance.

[0106] Based on the above-described functional requirements, a training pseudospectral data set and a validation pseudospectral data set are constructed using the spectrum generation device 10. For example, when a machine learning model for the purpose of dimensionality reduction is trained using a pair of pseudospectrums and a pair of weights associated with the pseudospectrums, the spectrum generation device 10 is used to construct datasets with any number of elements, with a set of two pseudospectrums and attribute information being used as one element of the training pseudospectral data set and the validation pseudospectral data set.

[0107] For example, in the case of a machine learning model that detects peak coordinates, the spectrum generation device 10 can be used to comprehensively obtain basis spectra obtained by shifting factor spectra, and pseudospectrums can be generated based on these basis spectra, thereby constructing a training pseudospectrum data set and a validation pseudospectrum data set that have a large number of peak coordinate patterns.

[0108] The verification pseudospectral dataset can be adjusted to become a dataset for verifying required performance by using the spectrum generation device 10. For example, in a machine learning model that performs dimensionality reduction, when required performance is set for the distance function included in Equation 6, distance patterns related to weight patterns are set at regular intervals such as [0.0, 0.1, . . . , 1.0], and pairs of pseudospectrum corresponding to pairs of weight patterns having each distance and pairs of basis information associated with the pairs are randomly generated to create a verification pseudospectral dataset, thereby enabling performance verification for each distance pattern.

[0109] FIG. 8 shows an example 60 of a verification result display that is presented to the user after the model output unit 54 outputs the verification results obtained by the verification unit 53. The horizontal axis represents the distance based on pairs of attribute information (weight patterns), the vertical axis represents the distance based on pairs of machine learning model outputs (dimensionality reduction results), and the black circles represent the average values ​​of the verification results obtained for each sample of the verification pseudospectral dataset. The coincidence line 601 is the line where these distances coincide, and the area between the threshold lines 602 represents the allowable range of the required performance.

[0110] As described above, by using a pseudospectral dataset obtained by comprehensively varying the values ​​of attribute information as a validation pseudospectral dataset, it is possible to verify the relationship between the varied attribute information and the output of the machine learning model.

[0111] 9 shows a verification result display example 61 that is output by the model output unit 54 from the verification unit 53 and presented to the user. In the verification result display example 61, the horizontal axis represents certain attribute information (numerical values), the vertical axis represents the value of the objective function, and the black circles represent the average value of the verification results obtained for each sample in the verification pseudospectral data set. When all the black circles are below a threshold line 610, this indicates that the required performance is met.

[0112] The learning unit 52 learns the machine learning model using the machine learning model designed by the design unit 51, the training pseudospectral data set, and the objective function. Any algorithm may be used to optimize the objective function. The algorithm may be steepest descent or stochastic gradient descent.

[0113] The verification unit 53 verifies the performance of the machine learning model learned by the learning unit 52 using the verification pseudospectral dataset constructed by the design unit 51, and outputs functional requirements to the design unit 51 again based on the verification results, thereby expanding the training pseudospectral dataset and training the machine learning model again in the learning unit 52, or determining it as the final machine learning model.

[0114] For example, if a threshold is set as a functional requirement, such as the value of the objective function for the verification pseudospectral dataset being equal to or less than a specific value (or equal to or more than a specific value), and if the functional requirement is satisfied, the machine learning model trained by the learning unit 52 is determined to be the final machine learning model, and is stored together with the verification results in the memory 11. If the functional requirement is not satisfied, the design unit 51 expands the training pseudospectral dataset, and the learning unit 52 performs learning again, and the verification unit 53 performs verification again.

[0115] In this way, by using the spectrum generator 10 to generate a training pseudospectral data set and a validation pseudospectral data set that are tailored to the purpose, and then repeating the process of recursively expanding the training pseudospectral data set based on the validation results and the process of relearning, it is possible to efficiently train a machine learning model that satisfies functional requirements.

[0116] The model output unit 54 outputs the final machine learning model determined by the verification unit 53 and the verification result. As described in the first embodiment, the model output unit 54 may display the output result on an output device (not shown). As shown in verification result display examples 60 and 61, the model output unit 54 may display the relationship between any attribute information and the output in a graph-like format in addition to the verification result based on the required performance.

[0117] <Modifications> The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0118] Furthermore, the above-described configurations, functions, processing units, etc. may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. Furthermore, the above-described configurations, functions, etc. may be implemented in software by a processor interpreting and executing a program that implements each function. Information such as the program, table, and file that implements each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card or SD card.

[0119] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and do not necessarily show all the control lines and information lines in the product. In reality, it can be assumed that almost all components are interconnected.

[0120] The spectrum generating device 10 or the machine learning model construction device 50 may be connected to a storage medium MM. The storage medium MM is configured as, for example, a memory device, a hard disk device, an optical disk device, a magneto-optical disk device, a magnetic tape device, or the like, and non-temporarily stores computer programs and data.

[0121] The storage medium MM can transfer a computer program for realizing the main functions of the spectrum generating device 10 or the machine learning model building device 50 to the spectrum generating device 10 or the machine learning model building device 50. Alternatively, a computer program for realizing some or all of the main functions of the spectrum generating device 10 or the machine learning model building device 50 can be transferred to and stored in the storage medium MM. By connecting the storage medium MM to another computer (not shown) and installing the computer program stored in the storage medium MM in the other computer, the other computer can function as the spectrum generating device 10 or the machine learning model building device 50.

[0122] It is clear that the above-described embodiments enable the invention expressed as follows.

[0123] (Representation 1) A spectrum generation device that generates pairs of pseudospectrums and attribute information for constructing a machine learning model using spectra as explanatory variables, the spectrum generation device comprising: an input unit that receives input of a spectral dataset consisting of one or more spectra or one or more pairs of spectra and sample information, and a generation request that adjusts the generation result; a basis acquisition unit that acquires one or more basis spectra that are components of the pseudospectrum to be generated; a generation unit that sets weights on the one or more basis spectra obtained by the basis acquisition unit and generates, from the weighted basis spectra, a pseudospectrum that is different from the spectrum input to the input unit; and an output unit that outputs one or more pairs of the generated pseudospectrum and attribute information, wherein the basis spectra obtained by the basis acquisition unit include at least one or more factor spectra that are components of the spectrum input to the input unit or spectra obtained by performing a predetermined variation process on the factor spectra, and the attribute information that is associated with the pseudospectrum and output from the output unit includes information on at least one of the basis spectra, the predetermined variation process, and the weight used to generate the pseudospectrum.

[0124] (Representation 2) A spectrum generating device according to Representation 1, wherein the generating unit sets weights to one or more basis spectra obtained by the basis acquisition unit, and calculates a weighted sum of the basis spectra and the weights to generate a pseudospectrum having a shape different from that of the spectrum input to the input unit.

[0125] (Representation 3) A spectrum generating device according to claim 1 or 2, wherein the basis acquisition unit treats one or more spectra input to the input unit as a tensor consisting of a sample direction axis and one or more spectrum direction axes, and uses a tensor decomposition technique to decompose the spectrum into one or more pairs of factor spectra and relative concentration vectors, the factor spectra being included in the basis spectra, and the relative concentration vectors being vectors with the same size as the number of spectra.

[0126] (Representation 4) A spectrum generation device according to any one of Representations 1 to 3, wherein the basis acquisition unit determines a feature extraction function based on the generation request, and the optimization function of the tensor decomposition method has a regularization term that aims to make the output of the feature extraction function when the spectrum input to the input unit is used as input similar to the output of the feature extraction function when a reconstruction result using a set of a factor spectrum and a concentration vector, which are the tensor decomposition results, is used as input, and the generation unit obtains the output of the feature extraction function when the pseudospectrum is used as input, and includes the output in the attribute information.

[0127] (Representation 5) A spectrum generating device described in any one of Representations 1 to 4, wherein the input unit receives spectra associated with one or more standard samples, and the basis acquisition unit includes the spectra in the basis spectrum.

[0128] (Representation 6) A spectrum generating device according to any one of Representations 1 to 5, wherein the generating unit sets the weights according to an arbitrary probability distribution, and includes information on the probability distribution in the attribute information.

[0129] (Representation 7) A spectrum generating device according to any one of Representations 1 to 6, wherein the generating unit sets the weights according to a probability distribution whose parameters include statistics of the density vectors, and information on the probability distribution is included in the attribute information.

[0130] (Representation 8) A machine-learning model construction device includes the spectrum generation device according to any one of Representations 1 to 7, and constructs a machine-learning model using a set of pseudospectrum and basis information, the machine-learning model construction device receiving any function request for adjusting the machine-learning model to be constructed, and constructing a training pseudospectrum dataset and a validation pseudospectrum dataset using the spectrum generation device. The machine-learning model construction device further includes a design unit that defines specifications of the machine-learning model, a learning unit that trains the machine-learning model using the training pseudospectrum dataset, a verification unit that verifies the machine-learning model using the validation pseudospectrum dataset, and a model output unit that outputs the machine-learning model and a verification result obtained by the verification device. The verification unit verifies the performance of the machine-learning model trained by the learning unit, and expands the training pseudospectrum dataset by outputting a function request to the design unit in accordance with the verification result, and re-trains and verifies the machine-learning model ... design unit uses the pseudospectrum as input and designs an objective function based on the attribute information.

[0131] (Representation 9) A machine learning model construction device according to Representation 8, wherein the design unit includes a constraint term in the objective function used for training the machine learning model, the constraint term being aimed at ensuring that the latent variables or outputs of the machine learning model are similar when pseudospectra with similar attribute information are input.

[0132] (Representation 10) A machine learning model construction device according to either of Representation 8 or 9, wherein the verification unit includes in the verification result data representing the relationship between the output of the machine learning model when each pseudospectrum included in the verification pseudospectrum dataset is used as input and attribute information.

[0133] (Representation 11) A spectrum generation method for generating a pair of pseudospectrum and attribute information for constructing a machine learning model using a spectrum as an explanatory variable, the method comprising: an input step for receiving a spectral dataset consisting of one or more spectra or one or more pairs of spectra and sample information, and an arbitrary generation request for adjusting the generation result; a basis acquisition step for acquiring one or more basis spectra that will be components of the pseudospectrum to be generated; a generation step for generating a pseudospectrum having a shape different from that of the spectrum received in the input step by assigning weights to the one or more basis spectra obtained in the basis acquisition step and calculating a weighted sum of the basis spectra and the weights; and an output step for outputting one or more pairs of the pseudospectrum and attribute information, wherein the basis spectra obtained in the basis acquisition step include at least one or more factor spectra that are components of the spectrum input in the input step, or spectra obtained by subjecting the factor spectra to arbitrary variation processing, and the attribute information includes at least one of the basis spectra, the variation processing, and the weights used for generation.

[0134] (Representation 12) A spectrum generation method according to Representation 11, wherein the basis acquisition step treats one or more spectra input in the input step as a tensor having a sample direction axis and one or more spectrum direction axes, and decomposes the spectrum into one or more sets of factor spectra and relative concentration vectors using a tensor decomposition technique, the factor spectra being included in the basis spectra, and the relative concentration vectors being vectors with the same size as the number of spectra.

[0135] (Representation 13) A spectrum generation method according to either of Representation 11 or 12, wherein the basis acquisition step determines a similarity function based on the generation request, the optimization function of the tensor decomposition technique has a regularization term aimed at making the spectrum received in the input step similar to the reconstruction result based on the result of the similarity function, which inputs the spectrum input in the input step and a reconstruction result using a set of a factor spectrum and a concentration vector, which are the results of tensor decomposition, and the generation step acquires the output of a feature extraction function to which the pseudospectrum is input, and includes the output in the attribute information.

[0136] (Representation 14) A spectrum generation method described in any one of Representations 11 to 13, wherein the input step receives spectra associated with one or more standard samples, and the basis acquisition step includes the spectra in the basis spectra.

[0137] (Representation 15) A spectrum generation method according to any one of Representations 11 to 14, wherein the generation step sets the weights according to an arbitrary probability distribution, and includes information on the probability distribution in the attribute information.

[0138] 10: Spectrum generating device, 20: Control device, 21: Spectrum measuring device, 22: Output device, 50: Machine learning model building device, 51: Design unit, 52: Learning unit, 53: Verification unit, 54: Model output unit, 101: Input unit, 102: Basis acquisition unit, 103: Generation unit, 104: Output unit

Claims

1. A spectrum generation device that generates pairs of pseudospectrums and attribute information for constructing a machine learning model that uses spectra as explanatory variables, comprising: an input unit that receives input of a spectral dataset consisting of one or more spectra or one or more pairs of spectra and sample information, and a generation request that adjusts the generation result; a basis acquisition unit that acquires one or more basis spectra that are components of the pseudospectrum to be generated; a generation unit that sets weights on the one or more basis spectra obtained by the basis acquisition unit and generates, from the weighted basis spectra, a pseudospectrum that is different from the spectrum input to the input unit; and an output unit that outputs one or more pairs of the generated pseudospectrum and attribute information, wherein the basis spectra obtained by the basis acquisition unit include at least one factor spectrum that is a component of the spectrum input to the input unit or a spectrum obtained by applying a predetermined variation process to the factor spectrum, and the attribute information associated with the pseudospectrum and output from the output unit includes information on at least one of the basis spectra, the predetermined variation process, and the weight used to generate the pseudospectrum.

2. A spectrum generating device according to claim 1, wherein the generating unit sets weights to one or more basis spectra obtained by the basis acquisition unit, and calculates a weighted sum of the basis spectra and the weights to generate a pseudospectrum having a shape different from that of the spectrum input to the input unit.

3. A spectrum generating device according to claim 1, wherein the basis acquisition unit treats one or more spectra input to the input unit as a tensor consisting of a sample direction axis and one or more spectrum direction axes, and decomposes the spectrum into one or more sets of factor spectra and relative concentration vectors using a tensor decomposition technique, and includes the factor spectra in the basis spectra, and the relative concentration vector is a vector of the same size as the number of spectra.

4. A spectrum generation device according to claim 3, wherein the basis acquisition unit determines a feature extraction function based on the generation request, and the optimization function of the tensor decomposition method has a regularization term that aims to make the output of the feature extraction function when the spectrum input to the input unit is used as input similar to the output of the feature extraction function when a reconstruction result using a set of a factor spectrum and a concentration vector, which are the tensor decomposition results, is used as input, and the generation unit obtains the output of the feature extraction function and includes it in the attribute information.

5. A spectrum generating device according to claim 1, wherein the input unit receives spectra associated with one or more standard samples, and the basis acquisition unit includes the spectra in the basis spectra.

6. A spectrum generating device according to claim 2, wherein the generating unit sets the weights in accordance with an arbitrary probability distribution, and the attribute information includes information on the probability distribution.

7. A spectrum generating device according to claim 3 or claim 4, wherein the generating unit sets the weights in accordance with a probability distribution whose parameters include statistics of the density vectors, and the attribute information includes information on the probability distribution.

8. A machine learning model construction device comprising the spectrum generating device according to any one of claims 1 to 7, which constructs a machine learning model that takes a spectrum as input, using a set of pseudospectrum and basis information, the machine learning model construction device comprising: a design unit that receives any function request for adjusting the machine learning model to be constructed, and uses the spectrum generating device to construct a training pseudospectrum dataset and a validation pseudospectrum dataset, and further defines the specifications of the machine learning model; a learning unit that trains the machine learning model using the training pseudospectrum dataset; a verification unit that verifies the machine learning model using the validation pseudospectrum dataset; and a model output unit that outputs the machine learning model and a verification result by the verification device, wherein the verification unit verifies the performance of the machine learning model trained by the learning unit, and expands the training pseudospectrum dataset by outputting a function request to the design unit in accordance with the verification result, and re-trains and verifies the machine learning model; and the design unit uses the pseudospectrum as input and designs an objective function based on the attribute information.

9. A machine learning model construction device according to claim 8, wherein the design unit includes a constraint term in the objective function used for training the machine learning model, the constraint term being aimed at ensuring that the latent variables or outputs of the machine learning model are similar when pseudospectrums with similar attribute information are input.

10. A machine learning model construction device according to claim 8, wherein the verification unit includes in the verification result data representing the relationship between the output of the machine learning model when each pseudospectrum included in the verification pseudospectrum dataset is used as input and attribute information.

11. A spectrum generation method for generating pairs of pseudospectrums and attribute information for constructing a machine learning model using spectra as explanatory variables, the spectrum generation method comprising: an input step for receiving a spectral dataset consisting of one or more spectra or one or more pairs of spectra and sample information, and an arbitrary generation request for adjusting the generation result; a basis acquisition step for acquiring one or more basis spectra that will be components of the pseudospectrum to be generated; a generation step for generating a pseudospectrum having a shape different from that of the spectrum received in the input step by setting weights to the one or more basis spectra obtained in the basis acquisition step and calculating a weighted sum of the basis spectra and the weights; and an output step for outputting one or more pairs of the pseudospectrum and attribute information, wherein the basis spectra obtained in the basis acquisition step include at least one or more factor spectra that are components of the spectrum input in the input step, or spectra obtained by applying arbitrary variation processing to the factor spectra, and the attribute information includes information on at least one of the basis spectra, the variation processing, and the weights used for generation.

12. A spectrum generation method according to claim 11, wherein the basis acquisition step treats one or more spectra input in the input step as a tensor consisting of a sample direction axis and one or more spectrum direction axes, and decomposes the spectrum using a tensor decomposition technique into one or more sets of factor spectra and relative concentration vectors, the factor spectra being included in the basis spectra, and the relative concentration vectors being vectors with the same size as the number of spectra.

13. A spectrum generation method according to claim 12, wherein the basis acquisition step determines a feature extraction function based on the generation request, the optimization function of the tensor decomposition method has a regularization term that aims to make the output of the feature extraction function when the spectrum input in the input step is used as input similar to the output of the feature extraction function when a reconstruction result using a set of a factor spectrum and a concentration vector, which are the tensor decomposition results, is used as input, and the generation step obtains the output of the feature extraction function when the pseudospectrum is used as input, and includes it in the attribute information.

14. A spectrum generation method according to claim 11, wherein the input step receives spectra associated with one or more standard samples, and the basis acquisition step includes the spectra in the basis spectra.

15. A spectrum generation method according to claim 11, wherein said generating step sets said weights in accordance with an arbitrary probability distribution, and said attribute information includes information on the probability distribution.

Images