Spectra generation apparatus and method, machine learning model construction apparatus

Abstract

The present application provides a kind of spectral generation device, for the machine learning model with spectrum as input, generates the data set of multiple pseudo-spectra and attribute information for learning the machine learning model corresponding to purpose.The spectral generation device for generating the group of pseudo-spectrum and attribute information for constructing machine learning model has: input unit, which is input by the group of spectrum data set and generation request of spectrum or spectrum and sample information;Foundation acquisition unit, which acquires the foundation spectrum of generated pseudo-spectrum;Generation unit, which generates pseudo-spectrum according to foundation spectrum;And output unit, which outputs the group of generated pseudo-spectrum and attribute information.

Description

Technical Field

[0001] This invention relates to a spectrometer and method for generating spectra, and a machine learning model building apparatus. Background Technology

[0002] It can quantify the physical properties of a sample into a spectrum. For example, a three-dimensional fluorescence spectrum obtained using a spectrophotometer is obtained by irradiating the sample with excitation light and measuring the intensity of the fluorescence emitted by the sample at this time. The information of the fluorescence characteristics of the sample is quantified into a two-dimensional array.

[0003] By quantifying spectra, it is possible to construct models that use machine learning to infer the properties of a sample. For example, supervised learning can be used to construct models that take spectra as input and infer quantitative values ​​such as the physical properties or compound concentrations of a sample corresponding to the input spectra, taking the spectra as input. Models that infer categories such as the type of sample can also be constructed in the same way.

[0004] Spectra sometimes contain information beyond the desired properties of the sample, acting as noise that does not show correlation with the target quantitative value or category. Furthermore, spectra tend to have high correlations (collinearity) between adjacent elements, so when learning machine learning models that use spectral elements as explanatory variables, the optimization of parameters related to highly correlated elements is sometimes unstable. Therefore, feature quantities are sometimes extracted from spectra depending on the objective.

[0005] Regarding the model for extracting this feature, machine learning models are sometimes used. The feature can take the form corresponding to the objective, such as the result of removing noise from the spectrum, the peak coordinates and peak information of the spectrum, or a feature vector with a lower dimension than the spectrum. For example, in the case of a machine learning model for estimating the peak coordinates and peak information of a spectrum, it can be constructed using supervised learning with a training dataset consisting of a pre-prepared spectrum and a set of peak coordinates and peaks contained within the spectrum.

[0006] To build a machine learning model based on the aforementioned spectra, a training dataset is needed to optimize the parameters of the machine learning model. However, in the case of spectra, there are many procedures from sample preparation to measurement, making it difficult to collect a sufficient amount of training data.

[0007] Patent Document 1 discloses a method that separates the spectrum (original spectrum) obtained from a sample into an analytical spectrum and a non-analytical spectrum, performs various modifications on the non-analytical spectrum, and generates a hypothetical spectrum by adding the modified non-analytical spectrum to the analytical spectrum, thereby rapidly increasing the amount of training data without actually preparing samples.

[0008] Existing technical documents

[0009] Patent documents

[0010] Patent Document 1: Japanese Patent Application Publication No. 2023-74746 Summary of the Invention

[0011] The problem that the invention aims to solve

[0012] The technique described in Patent Document 1 uses frequency separation technology to decompose the spectrum (original spectrum) obtained from the sample into a spectrum (analytical spectrum) that forms the basis for the analysis of the sample and a non-analytical spectrum that is different from the analytical spectrum. Furthermore, in Patent Document 1, various modifications are applied to the non-analytical spectrum, and it is added to the analytical spectrum to generate a hypothetical spectrum. Moreover, in Patent Document 1, by establishing a correlation between pre-prepared quantitative values ​​and categories representing the properties of the sample, the amount of training data for the machine learning model used to learn and estimate the quantitative values ​​and categories is increased.

[0013] However, since the number of modes in the spectrum used for analysis does not increase, overlearning may occur in machine learning models that estimate the quantitative values ​​and categories due to insufficient number of modes.

[0014] Furthermore, in addition to the aforementioned machine learning models that estimate quantitative values ​​and categories, machine learning models that focus on spectra also include models that estimate the result after removing noise signals from the spectrum, models that estimate the peak coordinates and peak values ​​of the spectrum, and models that compress the dimensions into feature vectors with lower dimensions than the spectrum, etc., which are models that estimate quantitative values ​​and categories associated with the sample.

[0015] However, Patent Document 1 does not mention a method for generating or learning training data for these models. For example, in the case of these models, it is not only related to the construction of a generalized model that the patterns of the analytical spectra obtained from the original spectra are comprehensively learned, but also to the patterns of the shapes that can be obtained from the spectra. However, in the method described in Patent Document 1, the number of patterns in the training data is insufficient.

[0016] The present invention was made in view of the above-mentioned problems, and its object is to provide a spectrum generation apparatus and method, and a machine learning model building apparatus for generating a dataset of multiple pseudo-spectral and attribute information for learning a machine learning model corresponding to a target, for a machine learning model that takes a spectrum as input.

[0017] Methods for solving problems

[0018] To address the aforementioned issues, one aspect of the present invention provides a spectral generation apparatus that generates a set of pseudo-spectrums and attribute information for constructing a machine learning model that uses spectra as explanatory variables. The spectral generation apparatus comprises: an input unit that receives a spectral dataset and a generation request, the spectral dataset consisting of one or more spectra or a set of one or more spectra and sample information, and the generation request adjusting the generation result; a base acquisition unit that acquires one or more base spectra that constitute the generated pseudo-spectrums; a generation unit that assigns weights to the one or more base spectra obtained by the base acquisition unit and generates pseudo-spectrums different from the spectra input to the input unit based on the weighted base spectra; and an output unit that outputs a set of one or more generated pseudo-spectrums and attribute information, wherein the base spectra acquired by the base acquisition unit include at least one factor spectrum that constitutes the spectra input to the input unit or a spectrum obtained by performing a predetermined modification process on the factor spectra, and the attribute information output from the output unit corresponding to the pseudo-spectrums includes information on at least one of the base spectra used to generate the pseudo-spectrums, the predetermined modification process, and the weights.

[0019] Invention Effects

[0020] According to the present invention, it is possible to generate a set of pseudo-spectral and attribute information for constructing a machine learning model that uses the spectrum as an explanatory variable. Attached Figure Description

[0021] Figure 1 This is a hardware structure diagram of the spectrum generation device.

[0022] Figure 2 This is a functional block diagram of the spectrum generation device.

[0023] Figure 3 This is a table representing an example of a spectrum.

[0024] Figure 4 This is a diagram representing an example of a spectrum.

[0025] Figure 5 This is an illustrative diagram showing an example of a group representing pseudo-spectral and attribute information.

[0026] Figure 6 This is the processing flow of the spectrum generation device.

[0027] Figure 7 This is a functional block diagram of the spectrum generation device in Example 2.

[0028] Figure 8 This is a chart representing an example of the verification results.

[0029] Figure 9 These are other charts that represent examples of verification results. Detailed Implementation

[0030] Hereinafter, embodiments of the present invention will be described based on the accompanying drawings. The spectral generation apparatus of this embodiment extracts a basic spectrum characteristic of a limited number of actual spectra, expands and reconstructs it, thereby generating more pseudo-spectrums. Actual spectra refer to the spectra obtained from actual measurements of samples.

[0031] The spectral generation apparatus of this embodiment, for example, obtains one or more basic spectra from a spectral dataset consisting of one or more spectra or a group of one or more spectra and sample information, assigns weights to each basic spectrum, calculates a weighted sum based on the basic spectrum and the weights, and thereby outputs a group of one or more pseudo-spectrums with different shapes than the spectra contained in the aforementioned spectral dataset, which can be used for learning by various machine learning models, and a group of attribute information associated with the pseudo-spectrums.

[0032] The spectrum generation apparatus of this embodiment includes, for example, a computing unit and a memory storing the computer program executed by the computing unit. The computing unit receives the spectral dataset as input and outputs one or more sets of pseudo-spectral data and attribute information.

[0033] The spectral generation apparatus according to this embodiment can generate pseudo-spectrums with a large number of patterns required for constructing various machine learning models corresponding to the purpose.

[0034] In this embodiment, the spectrum generation apparatus can construct training pseudo-spectral datasets and validation pseudo-spectral datasets using one or more sets of pseudo-spectral and attribute information for learning a target machine learning model. The training pseudo-spectral dataset is then used to learn the target machine learning model. The spectrum generation apparatus validates the machine learning model learned using the validation pseudo-spectral dataset, and recursively generates additional training pseudo-spectral datasets based on the validation results. Using this spectrum generation apparatus, a machine learning model construction system can be obtained that performs additional learning on a machine learning model and outputs a target machine learning model and validation results.

[0035] Example 1

[0036] use Figures 1-6 The first embodiment is described below.

[0037] <Hardware Structure of the Spectrum Generation Device>

[0038] In Example 1, a set of pseudo-spectral and attribute information that can be used as learning data for various machine learning models is generated based on a spectral dataset containing a small number of spectra.

[0039] use Figure 1The hardware structure of the spectrum generation device 10 will be described. The spectrum generation device 10 includes, for example, a memory 11, a processing unit 12, an interface 13, and a bus 14. The memory 11, the processing unit 12, and the interface 13 transmit and receive information via the bus 14.

[0040] The various parts of the spectrum generating apparatus 10 will be described. Interface 13 is a communication device for transmitting and receiving signals with devices 20, 21, and 22 located outside the spectrum generating apparatus 10. Devices that transmit and receive signals with 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 apparatus 10. Output device 22 can be, for example, a monitor display, a printer, etc. Output device 22 can also be configured as a head-mounted display. The spectrum generating apparatus 10 can also be connected to an input device (not shown) for user input of instructions or data.

[0041] These external devices 20, 21, and 22 can transmit and receive signals from the spectrum generating device 10 via wired connections using cables such as optical fibers, or via wireless connections using wireless communication technologies such as Bluetooth (registered trademark).

[0042] The spectrum generation device 10 does not necessarily need to be located in the same facility as the control device 20, the spectrum measuring device 21, and the output device 22; it can also be located in another computer or in the so-called cloud. In such cases, the transmission and reception of signals between the interface 13 and the external devices 20, 21, and 22 can be achieved using, for example, a LAN (Local Area Network) or a WAN (Wide Area Network).

[0043] The arithmetic unit 12 is a device that performs various processes within the spectrum generation device 10, such as a CPU (Central Processing Unit) or an FPGA (Field-Programmable Gate Array). The functions performed by the arithmetic unit 12 will be described later.

[0044] The memory 11 is a device that stores the program executed by the computing device 12, the parameters of the quantitative / identification model, the processing results, etc., such as HDD (Hard Disk Drive), flash memory, SSD (Solid State Drive), RAM (Random Access Memory), ROM (Read Only Memory), etc.

[0045] <Functional Structure of Spectrum Generating Device 10>

[0046] Figure 2 This is a functional block diagram of the spectrum generation apparatus 10. The functional units 101, 102, 103, and 104 shown can be implemented by the arithmetic unit 12 executing a predetermined computer program, or by dedicated hardware. The spectrum generation apparatus 10 includes, for example, an input unit 101, a basic acquisition unit 102, a generation unit 103, and an output unit 104 as functional units. Each functional unit will be described below.

[0047] The input unit 101 receives one or more spectra or a group of more than one set of spectra and tags associated with sample information input from the interface 13. Furthermore, the input unit 101 also receives a request to adjust the generated pseudo-spectrum. The spectra are in the form of a one-dimensional or higher array of data consisting of one or more axes, with each element storing the corresponding spectral intensity.

[0048] The basic acquisition unit 102 acquires one or more basic spectra as constituent elements of the spectrum based on one or more spectra or a group of one or more spectra and tags input to the input unit 101. The basic spectrum may also include at least one or more basic vectors based on the spectra input to the input unit 101 or vectors obtained by arbitrarily modifying the basic vectors.

[0049] The generation unit 103 assigns weights to the base spectrum acquired by the base acquisition unit 102, calculates a weighted sum based on the base spectrum and the weights, thereby generating a pseudo spectrum, and acquires attribute information associated with the pseudo spectrum. The attribute information includes at least one of the following: the base spectrum used when generating the pseudo spectrum, the content of the variation processing, and the assigned weights.

[0050] The output unit 104 outputs one or more sets of pseudo-spectral and attribute information. The destination of the output information is the output device 22.

[0051] Furthermore, the aforementioned functions do not require, as Figure 2 The functional units 101 to 104 shown are configured as shown, as long as they can perform processing equivalent to the operation of each functional unit.

[0052] Figure 6 This is an example of a processing flowchart. Each step S1~S4 is related to... Figure 2 The functional units 101 to 104 shown correspond to this.

[0053] In input step S1, one or more spectra or a group of more than one set of spectra and a group of tags associated with the spectra are received from interface 13, along with a generation request for adjusting the generated pseudo-spectrum. The spectrum is in the form of a one-dimensional or higher array of data consisting of one or more axes, with the corresponding spectral intensity stored in each element.

[0054] In the basic acquisition step S2, based on one or more spectra received in the input step S1, one or more basic spectra are acquired as constituent elements of the spectra. The basic spectrum includes at least one basic vector based on the spectrum input to the input unit 101 or a vector obtained by arbitrarily modifying the basic vector.

[0055] In generation step S3, weights are assigned to the base spectrum obtained in the base acquisition step S2, and a weighted sum is calculated based on the base spectrum and the weights to generate a pseudo spectrum. Attribute information associated with the pseudo spectrum is also obtained. The attribute information includes at least one of the following: the base spectrum used in generating the pseudo spectrum, the applied modification process, and the assigned weights.

[0056] In step S4, output one or more sets of pseudo-spectral and attribute information.

[0057] <Structure and Movement of Each Part>

[0058] The operation of the input unit 101, the basic acquisition unit 102, the generation unit 103 and the output unit 104 described above will be described in detail below.

[0059] Input unit 101 receives one or more spectra or a group of one or more spectra and tags associated with the spectra, as well as a generation request, via 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 for excitation wavelength and emission wavelength, with emission intensity stored in each element. The spectrum or the group of spectra and tags received by input unit 101 is output to basic acquisition unit 102, or stored in memory 11 and output to basic acquisition unit 102. A generation request refers to structured data containing parameters for adjusting the generated pseudo-spectrum.

[0060] Figure 3 This represents 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. Figure 4 Example 31 of visualization of three-dimensional fluorescence spectrum. In Example 31 of visualization of three-dimensional fluorescence spectrum, the vertical axis is the excitation wavelength and the horizontal axis is the emission wavelength. The emission intensity of each combination of excitation and emission wavelengths is displayed as contour lines.

[0061] As shown in the three-dimensional fluorescence spectrum 30, the excitation wavelength, emission wavelength, and emission intensity are recorded with arbitrary quantization widths. The three-dimensional fluorescence spectrum 30 represents an example of measuring wavelengths from 200 nm to 700 nm relative to the excitation / emission light with a quantization width of 5 nm. For example, the emission intensity of the detection light at 200 nm corresponding to the excitation light at 200 nm is recorded in the first row and first column of the three-dimensional fluorescence spectrum 30; the emission intensity of the detection light at 205 nm corresponding to the excitation light at 200 nm is recorded in the second row and second column of the first row; and the emission intensity of the detection light at 200 nm corresponding to the excitation light at 205 nm is recorded in the first column of the second row. Since 500 nm wavelengths are recorded at 5 nm intervals, the three-dimensional fluorescence spectrum 30 has 100 columns and 100 rows.

[0062] Example 31 of the three-dimensional fluorescence spectrum visualization is an example of visualizing the three-dimensional fluorescence spectrum 30 as contour lines in a two-dimensional image. When observing a three-dimensional fluorescence spectrum, it is often confirmed through a display like that in Example 31. Besides contour lines, it can also be displayed using grayscale images, thermal images, bird's-eye views, etc., where emission intensity is assigned to brightness values ​​and colors. The shape of the three-dimensional fluorescence spectrum varies depending on the properties of the fluorescence characteristics of the object.

[0063] The generation request is a parameter used by an external (user) entity to adjust the settings of the input unit 101, the basic acquisition unit 102, and the generation unit 103. The spectrum generation device 10 changes various settings according to the generation request. The user sets the generation request from the control device 20 via the interface 13 using a mouse, keyboard, etc.

[0064] When the input unit 101 outputs to the basic acquisition unit 102, or stores the data in the memory 11 and outputs it to the basic acquisition unit 102, it can also perform arbitrary preprocessing on the spectrum.

[0065] In Example 31 of three-dimensional fluorescence spectroscopy visualization, the three-dimensional fluorescence spectrum includes signals that can be considered noise. For example, in cases where fluorescence properties are analyzed based on the three-dimensional fluorescence spectrum, bands containing Rayleigh scattering typically become noise components. Rayleigh scattering is not fluorescence emitted by the sample, but rather light reflected from the sample. By detecting Rayleigh scattering, features unrelated to fluorescence properties are reflected in the three-dimensional fluorescence spectrum.

[0066] Rayleigh scattering is characterized by the following: due to the influence of the diffraction grating, the excitation wavelength is regionally detected in the vicinity of a band that is a natural multiple or a reciprocal multiple of the emission wavelength. Therefore, for example, an arbitrary band can be set, and the emission intensity of the wavelength and its surrounding bandwidth region, consistent with the above conditions, can be converted to any fixed value such as 0. Alternatively, the three-dimensional fluorescence spectrum of a standard sample can be measured, and preprocessing can be performed by subtracting the three-dimensional fluorescence spectrum of the standard sample from each spectrum, thereby suppressing the influence of noise. Raman scattering and other generated artifacts can also be preprocessed in any manner.

[0067] In the method disclosed in this embodiment, the spectrum generation device 10 acquires one or more basic spectra as constituent elements of the spectrum based on the input spectrum, and assigns various weights to the basic spectra. Then, the spectrum generation device 10 outputs one or more sets of pseudo-spectrums represented by the weighted sum of multiple basic spectra and a set of attribute information associated with the pseudo-spectrums.

[0068] The spectrum generation apparatus 10 of this embodiment can generate pseudo-spectrums of various modes by acquiring a base spectrum and changing the weights involved in the base spectrum, and output attribute information associated with the pseudo-spectrums. Therefore, the spectrum generation apparatus 10 can flexibly create datasets for learning various machine learning models according to their purposes.

[0069] The basic acquisition unit 102 receives one or more spectra from the input unit 101 or the memory 11 and acquires basic spectra. Here, basic spectra refer to the constituent elements of the generated pseudo-spectrum, equivalent to the basic vectors when the data to be generated is represented in a vector space. The spectrum generation apparatus 10 acquires various basic spectra based on the spectral dataset and changes the weights involved in the basic spectra, thereby enabling the generation of various pseudo-spectrums.

[0070] The following describes the methods for obtaining the fundamental spectra. Multiple methods for obtaining the fundamental spectra can also be used. The resulting list of fundamental spectra is referred to as the fundamental spectrum list.

[0071] As an example of a method for acquiring a base spectrum, a method for acquiring a factor spectrum based on the spectrum received by the input unit 101 and using that factor spectrum as the base spectrum will be described. Here, the factor spectrum is equivalent to the base vector when the spectrum received by the input unit 101 is used as a vector, and represents the spectrum that constitutes the components of the spectrum received by the input unit 101.

[0072] As an example of a method for obtaining factor spectra, a method for obtaining factor spectra for three-dimensional fluorescence spectra using PARAFAC (PARArell Factor Analysis), a tensor decomposition method, is illustrated. In PARAFAC, the linear relationship between the concentration of fluorescent molecules in a sample and the fluorescence intensity is utilized, and multiple three-dimensional fluorescence spectra are represented using the trilinear model described in Equation 1.

[0073] [Formula 1]

[0074]

[0075] In equation 1, F' k Let E represent a tensor containing the number of excitation bands. x and the number of transmission bands E m The K three-dimensional fluorescence spectra F1, F2, ..., F2 constituted K The k-th sample F k Represented by a trilinear model. N is a hyperparameter representing the number of fluorescent molecule species contained in the sample. n It is the inherent emission spectrum of the nth fluorescent molecule from 1 to N, which is E x A vector dimension. An emission spectrum represents the distribution of light emitted at various excitation wavelengths when fluorescent molecules are irradiated with excitation light, at a fixed emission wavelength. n It is the inherent excitation spectrum of the nth fluorescent molecule, and it is E m A dimensional vector. The excitation spectrum represents the distribution of light emitted by fluorescent molecules excited by excitation light of a fixed wavelength at various emission wavelengths. kn It refers to the relative concentration of the nth fluorescent molecule contained in the sample corresponding to the kth sample.

[0076] The circled × (hereinafter (×)) refers to the operation of the direct product of vectors. That is, a n (×)b n E with the same shape as the three-dimensional fluorescence spectrum x ×E m Matrix. The c obtained by multiplying this matrix by the relative concentration. kn (a) n (×)b n () represents the signal generated by the nth fluorescent molecule in the kth sample. Equation 1 is the expression for the three-dimensional fluorescence spectrum F. k The sum of the first to the Nth signals, F' k The numerical formula obtained through modeling.

[0077] In PARAFAC, the emission spectra a1, ..., a N The various elements, excitation spectrum b1, ..., b N The various elements, and the relative concentrations c of each fluorescent molecule in each sample. 11 c 1N c KN As variables, each variable is optimized based on the objective function shown in Equation 2. Equation 2 represents the optimization of the three-dimensional fluorescence spectra F1, F2, ..., F... K and tensors F'1, F'2, ..., F' represented by the trilinear model K The average of the squared errors of each element is minimized. F ijk This represents the emission intensity of the k-th sample at the i-th excitation wavelength and the j-th emission wavelength. in The emission spectrum a of the nth fluorescent molecule n The value of the i-th excitation wavelength in b. jn b represents the excitation spectrum of the nth fluorescent molecule. n The value of the j-th emission wavelength in the equation.

[0078] The objective function shown in Equation 2 can be optimized using any optimization method. For example, it can be optimized using ALS (Alternating Least Squares) or the steepest descent method.

[0079] [Formula 2]

[0080]

[0081] The optimized results of emission spectrum, excitation spectrum, and relative concentration obtained by Equation 2 can also be standardized in a way that unifies the maximum values ​​of each emission spectrum and excitation spectrum. With standardization, the scale of relative concentration is changed.

[0082] Arbitrary regularization terms can also be added to the objective function shown in Equation 2. For example, a feature extraction function can be defined that takes each three-dimensional fluorescence spectrum or a tensor represented by a trilinear model as input. For each pair of tensors represented by a trilinear model corresponding to each three-dimensional fluorescence spectrum, the squared error of the output of the feature extraction function is added as a regularization term to Equation 2. The feature extraction function can read from memory 112 any optimized model used to estimate the compound concentration and compound species of the three-dimensional fluorescence spectrum. By adding such a regularization term to Equation 2, when optimization is performed to make the tensor represented by the trilinear model similar to the three-dimensional fluorescence spectrum, regularization can be performed to make elements such as compound concentration and compound species that can be inferred from the three-dimensional fluorescence spectrum similar as well.

[0083] The number N of fluorescent molecules in Equations 1 and 2 is initially unknown. Therefore, PARAFAC results for N from multiple modes can be obtained, scores can be calculated, and the PARAFAC result corresponding to the N with the highest score can be used as the final PARAFAC result. Alternatively, the PARAFAC result corresponding to N from a single mode can be used as the final PARAFAC result. Core consistency, widely used in PARAFAC, can be used as the scoring function.

[0084] The optimized product of the emission and excitation spectra is a1 (×)b1, ..., a N (×)b N It is a collection of three-dimensional fluorescence spectra representing the inherent fluorescence properties of N fluorescent molecules, and is a factor spectrum constituting the three-dimensional fluorescence spectrum received by the input unit 101. This factor spectrum is added to the basic spectrum list as a basic spectrum.

[0085] The optimization results of the relative concentration obtained by PARAFAC can also be stored in memory 11 in association with the factor spectrum. When a set of spectrum and tag is received through input unit 101, the tag information can also be stored in memory 11 together with the optimization results of the relative concentration.

[0086] Furthermore, as a method for obtaining factor spectra, a method using spectra obtained from a standard sample composed of a single fluorescent molecule is described. Since each fluorescent molecule possesses inherent fluorescence properties, in the case of a sample composed of a single fluorescent molecule, the obtained three-dimensional fluorescence spectrum becomes a two-dimensional matrix represented by a factor spectrum, with the scale linearly varying according to the concentration of the fluorescent molecules. Therefore, when the three-dimensional fluorescence spectrum input to the input unit is a three-dimensional fluorescence spectrum obtained from a sample composed of a single fluorescent molecule, the three-dimensional fluorescence spectrum itself can be considered as a factor spectrum and added to the basic spectrum list. When the input unit 101 receives a spectrum of a sample composed of a single fluorescent molecule and a set of tags containing sample information such as the fluorescent molecule name and structural formula, the tag information can also be stored in the memory 11.

[0087] Alternatively, the result of applying arbitrary modification processing to the base spectrum stored in the base spectrum list can be appended to the base spectrum list. The factor spectrum described above is the result of obtaining the constituent elements of the spectrum received by the input unit 101. In the subsequent generation unit 103, if only the above factor spectrum is used, a pseudo spectrum with a similar shape to the above spectrum but different weights for each factor spectrum is generated. When modification processing is applied to the above base spectrum, a pseudo spectrum with constituent elements different from the above spectrum can be generated. As a result, the machine learning model can learn pseudo spectra with more diverse patterns than previous spectrum generation methods, realizing a general machine learning model. Hereinafter, examples of modifications to the base spectrum will be explained.

[0088] For example, the base spectrum can be shifted in any direction. Depending on the measurement environment and sample condition, a portion of the factor spectrum constituting the three-dimensional fluorescence spectrum may shift towards a shorter emission wavelength. Thus, the entire factor spectrum may shift under various conditions. If the machine learning model has not learned to produce such a shifted spectrum, and this shifted spectrum is input during inference, accurate inference may not be possible. Therefore, by shifting the base spectrum in any direction, the shift in the factor spectrum observed in the actual spectrum can be virtually reproduced in the subsequent generation unit 103. The shifted base spectrum produces missing values ​​at the end in the direction opposite to the shifted direction, but these can be interpolated using, for example, "0", or extrapolated using an arbitrary model such as a Gaussian process regression model.

[0089] For example, it is also possible to apply a monotonically increasing function to each element of the basic spectrum. When a monotonically increasing function is applied to each element of the basic spectrum, the shape of the basic spectrum can be changed within a range that does not alter the magnitude relationships between the elements. For example, in three-dimensional fluorescence spectroscopy, let the emission intensity at the i-th excitation wavelength and the j-th emission wavelength of a certain basic spectrum B be set as B. ij When the monotonically increasing function is set as f(), it can be modified in the form of expression 3, B' ij This is the result of the change processing. Equation 3 applies a monotonically increasing function f() to each element of the basic spectrum, standardizing it so that the maximum value of the applied result is the same as the maximum value before application. f() can be an exponential function exp(B ij Alternatively, it can be obtained by adding 1 to the input value in a way that the output does not take a negative value; this is the logarithmic function ln(B). ij +1).

[0090] [Formula 3]

[0091]

[0092] When performing the aforementioned modification process, if the base spectrum used in the modification process is a factor spectrum, and relative concentration information or label information is associated with it, the aforementioned relative concentration and label information can also be associated with the base spectrum that has undergone modification process and stored in memory 11.

[0093] Alternatively, the results of recursively applying the variation processing described above to the basic spectrum can also be added to the basic spectrum list.

[0094] Log information about the creation process of each acquired basic spectrum is stored in memory 11. For example, it includes information on whether the basic spectrum is a factor spectrum; if it is a factor spectrum, it includes information on the spectrum used when acquiring the factor spectrum; if it is not a factor spectrum, it includes information on the modification process.

[0095] The above-mentioned method for obtaining the basic spectrum can be preset with fixed conditions, or it can be arbitrarily adjusted by an external (user) using a generation request.

[0096] The generation unit 103 assigns weights to the basic spectra acquired by the basic acquisition unit 102, calculates a weighted sum of the basic spectra and the weights, and thereby generates pseudo-spectrums. By using various basic spectra acquired by the basic acquisition unit 102, pseudo-spectrums of various modes can be generated.

[0097] The weighted sum is expressed by equation 4. In equation 4, F' represents the pseudospectrum, and B... m w represents the m-th fundamental spectrum in the fundamental spectrum list.m This represents the weight of the m-th fundamental spectrum. E represents noise, which can be any high-frequency noise, low-frequency noise, or no noise at all (E = 0). The following provides an example of how to set the weights.

[0098] [Formula 4]

[0099]

[0100] For example, fixed weight candidates can be assigned to each basic spectrum, and combinations of these weight candidates can be defined. For instance, with M basic spectra, if 11 weight candidates are assigned in units of 0.1 within the range of 0 to 1 ([0.0, 0.1, ..., 1.0]), the number of combinations of these weight candidates is 11 to the power of M. Pseudo-spectrums are generated in each combination according to Equation 4. Furthermore, to avoid using basic spectra associated with the same factor spectra, the weight of any basic spectrum other than the factor spectrum itself or those obtained by applying modifications to the factor spectrum can be set to 0.

[0101] For example, weights can be set according to any probability distribution. For instance, as shown in Equation 5, a multidimensional normal distribution can also be used to set the weights. In Equation 5, N() represents an M-dimensional normal distribution that takes 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, where the m-th element is related to w in Equation 4. m Correspondingly, the input for the multivariate normal distribution in Equation 5 can be set randomly for the mean vector, variance, and covariance matrix, or it can be set to fixed values ​​arbitrarily.

[0102] Given that the relative concentration information is correlated with each basic spectrum, the mean, variance, and covariance of the relative concentration can also be calculated and used as the mean vector and variance-covariance matrix of Equation 5. By employing this method, since the weighted combination is generated based on the trend of the relative concentration information correlated with the actual spectrum obtained during tensor decomposition, the generated result of Equation 4 can be set as both an unobserved spectrum and a pseudo-spectrum that naturally reflects the actual fluorescence characteristics.

[0103] [Formula 5]

[0104]

[0105] The weights mentioned above can be set as fixed conditions in advance, or they can be arbitrarily adjusted by an external party (user) using a generated request.

[0106] The attribute information used in the generation of the pseudo spectrum is associated with the pseudo spectrum generated as described above. The attribute information includes information associated with the basic spectrum, such as the basic spectrum acquired by the basic acquisition unit 102, relative concentration information, and log information, stored in the memory 11, as well as the weights set by the generation unit 103 and the weight setting method such as the probability distribution.

[0107] In this way, the spectrum generation device 10 can easily prepare datasets for use in the learning of desired machine learning models by managing and associating various information arising from the generation of pseudo-spectrums. Furthermore, the spectrum generation device 10 can also use the attribute information itself for the training and validation of machine learning models. Examples of applying attribute information to the construction and validation of machine learning models will be described later.

[0108] The output unit 104 outputs a set of one or more pseudo-spectral and attribute information generated by the generation unit 103. The set of pseudo-spectral and attribute information is output as structured data.

[0109] The output unit 104 can also display the output results on the output device 22 via the interface 13.

[0110] Figure 5 Example 40 shows the generated results displayed to the user. Example 40 includes, for example, a pseudo-spectral visualization example 41, a basic spectral visualization example 42, and an attribute information display example 43. Pseudo-spectral visualization example 41 and basic spectral visualization example 42 are visualized using contour lines, similar to the three-dimensional fluorescence spectral visualization example 30. Attribute information display example 43 displays attribute information in text form. Specific names and values ​​are recorded in the "..." section within attribute display example 43. For example, if the "acquisition method" is specified, information such as "PARAFAC" or "single fluorescent molecule sample" is displayed. Figure 5 The generated results shown in Example 40 are an example, demonstrating that any display format can be used to observe pseudo-spectrum, fundamental spectrum, and attribute information.

[0111] Example 2

[0112] use Figures 7-9Example 2 will be described below. This example focuses on the differences from Example 1. In this example, the spectrum generation apparatus 10 described in Example 1 is effectively used to generate the training pseudo-spectral dataset and the validation pseudo-spectral dataset required for learning the desired machine learning model. The machine learning model building apparatus 50 of this example uses the training pseudo-spectral dataset to learn the machine learning model and uses the validation pseudo-spectral dataset to validate the learned machine learning model. The machine learning model building apparatus 50 generates the training pseudo-spectral dataset again based on the validation results and performs additional learning. Alternatively, the machine learning model building apparatus 50 efficiently provides the desired machine learning model by outputting the learned machine learning model and the validation results.

[0113] <Hardware Configuration of the Spectral Analysis System>

[0114] The machine learning model building device 50 in this embodiment and Figure 1 Similarly, the hardware structure described can be configured, for example, as a computer system with memory, a computing device, and interfaces (none of which are shown). Figure 1 The machine learning model building device 50 can be connected to the control device 20, the spectral measuring device 21, and the output device 22.

[0115] Figure 7 This is an example of a functional block diagram of a machine learning model building device 50. The functional units 51, 52, 53, and 54 shown in the diagram can be implemented by a computing device that executes a predetermined computer program, or by dedicated hardware.

[0116] The machine learning model building apparatus 50 includes, for example, a design unit 51, a learning unit 52, a verification unit 53, and a model output unit 54 as functional units, and has a built-in spectrum generation device 10. The functional units will be described below.

[0117] Design unit 51 receives a function request from an external source. Based on the received function request, design unit 51 inputs a generation request to the spectrum generation device 10, which has a pre-input spectrum dataset. Design unit 51 uses the pseudo-spectrum and attribute information output from the spectrum generation device 10 to construct a training pseudo-spectrum dataset and a validation pseudo-spectrum dataset. Furthermore, design unit 51 uses the training pseudo-spectrum dataset and the validation pseudo-spectrum dataset to design a machine learning model and objective function.

[0118] Learning Department 52 uses a pseudo-spectral dataset for training to learn machine learning models.

[0119] The verification unit 53 uses a pseudo-spectral dataset for verification to verify the performance of the machine learning model learned by the learning unit 52. Based on the verification results of the machine learning model's performance, the verification unit 53 outputs a function request to the design unit 51 to reconstruct the pseudo-spectral dataset for training and perform additional learning, or to determine the final machine learning model and the final machine learning model's verification results.

[0120] The model output section 54 outputs the final machine learning model and validation results.

[0121] The above functions do not need to be as follows Figure 7 The functions can be structured in the same way as the individual functional blocks, as long as they can perform the same processing as the functions.

[0122] <Structure and Movement of Each Part>

[0123] The following details the operations of the design department 51, learning department 52, verification department 53, and model output department 54.

[0124] Design unit 51, for example, receives a function request input through an interface not shown. Based on the received function request, design unit 51 inputs a generation request to the spectrum generation device 10, which has a pre-input spectrum dataset. Design unit 51 uses a combination of pseudo-spectral data and attribute information, which are outputs of the spectrum generation device 10, to construct a training pseudo-spectral dataset and a validation pseudo-spectral dataset. Design unit 51 uses these constructed datasets to design a machine learning model and an objective function.

[0125] Feature requests can also include the functionalities of the machine learning model. For example, dimensionality compression, peak coordinate detection, and inference of noise-removed spectra. Design machine learning models corresponding to these objectives.

[0126] Depending on the purpose, machine learning models can use either neural networks such as Convolutional Neural Networks (CNN) and Transformers, or Support Vector Machines (SVM).

[0127] For the machine learning model described above, an objective function that effectively utilizes attribute information is designed. In actual spectra, information such as the base spectrum and weights is unknown and difficult to determine. On the other hand, in the case of pseudo-spectrums, attribute information representing the properties of the spectrum can be used for machine learning, thus enabling machine learning that better reflects the properties of the spectrum compared to actual spectra.

[0128] For example, in the case of a machine learning model for dimensionality compression of three-dimensional fluorescence spectra, an objective function such as Equation 6 can be designed using the weight information of the base spectra as components of the pseudo-spectrum. In Equation 6, F' and F” represent pairs of pseudo-spectrums, σθ represents a machine learning model aimed at dimensionality compression with parameters set to θ, inputs set to pseudo-spectrums, and outputs set to vectors, w' and w” represent the weights of the base spectra constituting F' and F” as vectors of each component, and d() represents a distance function. The distance function can be any function including Euclidean distance.

[0129] Equation 6 is a function for pairs of pseudospectrals, aimed at ensuring that the distance between the vectors resulting from dimensionality compression matches the distance between the vectors representing the weights. That is, pairs with similar weight patterns in the pseudospectrals are learned in a similar manner to achieve similar dimensionality compression results. The pattern of the weights related to the base spectrum represents the proportion of the base spectrum, which is a component of the spectrum, contained within it. This is not the pattern of the weights themselves, and compared to comparing the difference between the distance between the pairs of pseudospectrals themselves and the distance in the dimensionality compression result, it is unaffected by factors other than the base spectrum, such as noise. Therefore, it is possible to learn a dimensionality compression method that better reflects the characteristics of the spectrum.

[0130] [Formula 6]

[0131]

[0132] For example, in the case of a machine learning model that detects peak coordinates, the objective function can be set to calculate the peak coordinates of each fundamental spectrum in the attribute information and detect them. By designing the objective function to detect peak coordinates derived from the fundamental spectra, learning can be performed without detecting peak coordinates derived from noise.

[0133] For example, in the case of a machine learning model that estimates a spectrum with noise removed, an objective function can also be designed to estimate the result of subtracting noise from the attribute information from the pseudo-spectrum.

[0134] Additionally, feature requests can also include attribute-based performance information for machine learning models.

[0135] For example, in the case of a machine learning model that performs dimensionality compression, the squared error between the value of the distance function of the pattern pair in Equation 6 and the value of the distance function of the compressed result pair can be set to be less than a fixed threshold in each sample of the validation pseudospectral dataset.

[0136] For example, in the case of a machine learning model that detects peak coordinates, the allowable estimation error of the estimated peak coordinates can be set as the performance requirement.

[0137] For example, in the case of a machine learning model that assumes a noise-removed spectrum, the noise removal performance, such as the signal-to-noise ratio, can be set as the desired performance.

[0138] Based on the aforementioned functional requests, the spectrum generation device 10 is used to construct training pseudo-spectral datasets and validation pseudo-spectral datasets. For example, when learning a machine learning model for dimensionality compression using pairs of pseudo-spectrums and their associated weights, the spectrum generation device 10 uses a set of two pseudo-spectrums and attribute information as elements in both the training and validation pseudo-spectral datasets to construct datasets with an arbitrary number of elements.

[0139] For example, in the case of a machine learning model that detects peak coordinates, the spectral generation device 10 is used to comprehensively acquire the basic spectra after shifting the factor spectra, and pseudo-spectrals are generated based on these basic spectra, thereby enabling the construction of a training pseudo-spectral dataset and a validation pseudo-spectral dataset for patterns with a large number of peak coordinates.

[0140] By using the spectrum generation device 10, the pseudo-spectral dataset for verification can be adjusted into a dataset for verifying the requested performance. For example, in a machine learning model that performs dimensionality compression, when the requested performance is set in relation to the distance function contained in Equation 6, the distance patterns related to the weights are set at fixed intervals such as [0.0, 0.1, ..., 1.0]. The dataset obtained by randomly generating pairs of pseudo-spectral data corresponding to pairs of patterns with weights for each distance and pairs of basic information associated with them is used as the pseudo-spectral dataset for verification, thereby enabling the verification of the performance of each distance pattern.

[0141] Figure 8 Example 60 is a verification result display where the model output unit 54 outputs the verification result obtained by the verification unit 53 and prompts the user. The horizontal axis represents the distance between pairs based on attribute information (weighted patterns), the vertical axis represents the distance between pairs based on the output of the machine learning model (dimensionality compression result), the black circle represents the average value of the verification results obtained from each sample of the pseudo-spectral dataset used for verification, the consistency line 601 represents the line where these distances are consistent, and the area enclosed by the threshold line 602 represents the allowable range of the requested performance.

[0142] As described above, by using a pseudo-spectral dataset that causes the values ​​of attribute information to change comprehensively as a validation pseudo-spectral dataset, it is possible to verify the relationship between the changed attribute information and the output of the machine learning model.

[0143] Figure 9Example 61 shows the verification result displayed by the model output unit 54, which outputs the verification result obtained by the verification unit 53 and prompts the user. In Example 61, the horizontal axis represents a certain attribute information (numerical value), the vertical axis represents the value of the objective function, and the black circle represents the average value of the verification result obtained from each sample in the pseudo-spectral dataset used for verification. If all the black circles are below the threshold line 610, it indicates that the requested performance is met.

[0144] Learning Department 52 uses the machine learning model designed by Design Department 51, the training pseudo-spectral dataset, and the objective function to learn the machine learning model. The optimization method for the objective function can be any algorithm. The algorithm can be either the steepest descent method or the stochastic gradient descent method.

[0145] The verification unit 53 uses the verification pseudo-spectral dataset constructed by the design unit 51 to verify the performance of the machine learning model learned by the learning unit 52. Based on the verification results, it outputs a function request to the design unit 51 again, thereby expanding the training pseudo-spectral dataset, and the learning unit 52 learns the machine learning model again, or decides on the final machine learning model.

[0146] For example, if a threshold is set as a function request, such as the objective function value for the verification pseudo-spectral dataset being below (or above) a specific value, and the function request is met, the machine learning model learned by the learning unit 52 will be determined as the final machine learning model, and will be stored in the memory 11 along with the verification results. If the function request is not met, the design unit 51 will expand the training pseudo-spectral dataset, the learning unit 52 will learn again, and the verification unit 53 will verify again.

[0147] In this way, by repeatedly using the spectrum generator 10 to generate training pseudo-spectral datasets and validation pseudo-spectral datasets that meet the purpose, and recursively expanding the training pseudo-spectral datasets based on the validation results and performing relearning, it is possible to efficiently learn machine learning models that meet functional requirements.

[0148] The model output unit 54 outputs the final machine learning model and verification results determined by the verification unit 53. The model output unit 54 can also, as described in Embodiment 1, display the output results using an output device (not shown). The model output unit 54 can also, as in Verification Result Display Example 60 and Verification Result Display Example 61, display, in addition to the verification results based on request performance, the relationship between arbitrary attribute information and the output in a graphical form.

[0149] <Variation Example>

[0150] This invention is not limited to the embodiments described above, but includes various modifications. For example, the embodiments described above are examples given in detail to facilitate understanding of the invention, and are not necessarily limited to having all the described structures. Furthermore, with respect to a part of the structure of each embodiment, other structures can be added, deleted, or replaced.

[0151] Furthermore, the aforementioned structures, functions, and processing units can be partially or entirely implemented in hardware, for example, through integrated circuit design. Alternatively, the aforementioned structures and functions can be implemented in software by a processor interpreting and executing programs that implement each function. The programs, tables, files, and other information implementing each function can be stored in recording devices such as memory, hard disks, SSDs (Solid State Drives), or recording media such as IC cards or SD cards.

[0152] Furthermore, control lines and information lines refer to the lines deemed necessary in the specifications, but do not necessarily represent all control lines and information lines on the product. In fact, almost all structures can be considered interconnected.

[0153] The spectrum generation device 10 or the machine learning model building device 50 can also be connected to the storage medium MM. The storage medium MM is configured as, for example, a memory device, a hard disk device, an optical disk device, a magnetic tape device, etc., and stores computer programs and data non-transiently.

[0154] The storage medium MM can forward computer programs used to implement the main functions of the spectrum generation device 10 or the machine learning model building device 50 to the spectrum generation device 10 or the machine learning model building device 50. Alternatively, it can also transfer and store part or all of the computer programs used to implement the main functions of the spectrum generation device 10 or the machine learning model building device 50 to the storage medium MM. By connecting the storage medium MM to another computer (not shown) and installing the computer programs stored in the storage medium MM on the other computer, the other computer can function as the spectrum generation device 10 or the machine learning model building device 50.

[0155] In the above embodiments, the invention described below is obviously practicable.

[0156] (Statement 1) A spectrum generation apparatus that generates a set of pseudo-spectrums and attribute information for constructing a machine learning model that uses spectra as explanatory variables, wherein the spectrum generation apparatus comprises: an input unit that receives a spectrum dataset and a generation request, the spectrum dataset consisting of one or more spectra or a set of one or more spectra and sample information, the generation request adjusting the generation result; a base acquisition unit that acquires one or more base spectra that become components of the generated pseudo-spectrums; a generation unit that assigns weights to one or more base spectra obtained by the base acquisition unit and generates pseudo-spectrums different from the spectra input to the input unit based on the weighted base spectra; and an output unit that outputs a set of one or more sets of the generated pseudo-spectrums and attribute information, wherein the base spectra acquired by the base 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 performing a predetermined modification process on the factor spectrum, and the attribute information output from the output unit corresponding to the pseudo-spectrums includes information on at least one of the base spectra used to generate the pseudo-spectrums, the predetermined modification process, and the weights.

[0157] (Statement 2) According to the spectrum generation apparatus of Statement 1, the generation unit sets weights on one or more basic spectra obtained by the basic acquisition unit, calculates a weighted sum based on the basic spectra and the weights, thereby generating a pseudo-spectrum with a shape different from the spectrum input to the input unit.

[0158] (Statement 3) According to the spectrum generation apparatus described in Statement 1 or 2, the basic acquisition unit processes one or more spectra input to the input unit as a tensor consisting of an axis of the sample direction and an axis of one or more spectral directions, and decomposes the spectrum into a group of one or more factor spectra and relative concentration vectors using a tensor decomposition method, wherein the basic spectrum contains the factor spectra, and the relative concentration vector is a vector of the same size as the number of spectra.

[0159] (Statement 4) The spectral generation apparatus according to any one of Statements 1 to 3, wherein the basic acquisition unit determines a feature extraction function based on the generation request, the optimization function of the tensor decomposition method has a regularization term, the purpose of which is to make the output of the feature extraction function when the spectrum input to the input unit is set as input similar to the output of the feature extraction function when the reconstruction result of the group of factor spectra and concentration vectors as tensor decomposition results is set as input, the generation unit acquires the output of the feature extraction function with the pseudo-spectrum as input, and includes the output in the attribute information.

[0160] (Statement 5) The spectral generation apparatus according to any one of Statements 1 to 4, wherein the input unit receives one or more spectra associated with a standard sample, and the base acquisition unit includes the spectrum in the base spectrum.

[0161] (Statement 6) The spectrum generation apparatus according to any one of Statements 1 to 5, wherein the generation unit sets the weights according to an arbitrary probability distribution and includes the probability distribution information in the attribute information.

[0162] (Statement 7) The spectral generation apparatus according to any one of Statements 1 to 6, wherein the generation unit sets the weights according to a probability distribution of a statistic in which the concentration vector is included in the parameters, and the attribute information includes information on the probability distribution.

[0163] (Statement 8) A machine learning model building apparatus, comprising a spectrum generation device as described in any one of Statements 1 to 7, for building a machine learning model with spectrum as input using a combination of pseudo-spectrums and basic information, the machine learning model building apparatus comprising: a design unit that receives arbitrary function requests for adjusting the machine learning model as the object of construction, uses the spectrum generation device to build a training pseudo-spectral dataset and a validation pseudo-spectral dataset, and defines the specifications of the machine learning model; a learning unit that uses the training pseudo-spectral dataset to learn the machine learning model; a validation unit that uses the validation pseudo-spectral dataset to validate the machine learning model; and a model output unit that outputs the machine learning model and the validation results of the validation device, the validation unit validating the performance of the machine learning model learned by the learning unit, outputting a function request to the design unit based on the validation results, thereby expanding the training pseudo-spectral dataset, and relearning and validating the machine learning model, the design unit using the pseudo-spectrum as input to design an objective function based on the attribute information.

[0164] (Statement 9) The machine learning model building apparatus according to Statement 8, wherein the design unit includes a constraint term in the objective function for learning the machine learning model, the purpose of which is to make the latent variables or outputs of the machine learning model similar when the pseudo-spectrum with similar attribute information is used as input.

[0165] (Statement 10) The machine learning model construction apparatus according to Statement 8 or 9, wherein the verification unit includes in the verification result the output of the machine learning model when each pseudospectrum contained in the verification pseudospectral dataset is taken as input and data representing the relational nature of attribute information.

[0166] (Statement 11) A spectral generation method generates a set of pseudo-spectrums and attribute information for constructing a machine learning model with spectra as explanatory variables, wherein the spectral generation method performs the following steps: an input step, receiving a spectral dataset and an arbitrary generation request, wherein the spectral dataset consists of one or more spectra or a set of one or more spectra and sample information, and the generation request adjusts the generation result; a base acquisition step, acquiring one or more base spectra that become constituent elements of the generated pseudo-spectrum; a generation step, assigning weights to one or more base spectra obtained in the base acquisition step, calculating a weighted sum based on the base spectra and the weights, thereby generating a pseudo-spectrum with a shape different from the spectra received in the input step; and an output step, outputting a set of one or more sets of the pseudo-spectrums and attribute information, wherein the base spectra acquired in the base acquisition step include at least one or more factor spectra that are constituent elements of the spectra input in the input step or spectra obtained by arbitrarily modifying the factor spectra, and the attribute information includes information on at least one of the generated base spectra, the modification, and the weights.

[0167] (Statement 12) According to the spectral generation method described in Statement 11, in the basic acquisition step, one or more spectra input in the input step are processed as tensors consisting of axes of the sample direction and one or more spectral directions. Using a tensor decomposition method, the spectra are decomposed into a group of one or more factor spectra and relative concentration vectors. The factor spectra are included in the basic spectrum, and the relative concentration vectors are vectors of the same size as the number of spectra.

[0168] (Statement 13) According to the spectrum generation method described in Statement 11 or 12, in the basic acquisition step, a similarity function is determined based on the generation request, and the optimization function of the tensor decomposition method has a regularization term. The purpose of the regularization term is to make the spectrum received in the input step similar to the reconstruction result based on the similarity function with the spectrum input in the input step and the reconstruction result of the group of factor spectra and concentration vectors as the result of tensor decomposition. In the generation step, the output of the feature extraction function with the pseudo-spectrum as input is obtained, and the output is included in the attribute information.

[0169] (Representation 14) The spectral generation method according to any one of Representations 11 to 13, wherein, in the input step, one or more spectra associated with a standard sample are received, and in the base acquisition step, the spectra are included in the base spectrum.

[0170] (Statement 15) The spectral generation method according to any one of Statements 11 to 14, wherein, in the generation step, the weights are set according to an arbitrary probability distribution, and the information of the probability distribution is included in the attribute information.

[0171] Explanation of reference numerals in the attached figures

[0172] 10: Spectrum generation device

[0173] 20: Control device

[0174] 21: Spectroscopic measuring device

[0175] 22: Output device

[0176] 50: Machine learning model building device

[0177] 51: Design Department

[0178] 52: Study Department

[0179] 53: Verification Department

[0180] 54: Model output section

[0181] 101: Input Department

[0182] 102: Basic Acquisition Department

[0183] 103: Production Department

[0184] 104: Output section.

Claims

1. A spectrum generation apparatus for generating a set of pseudo-spectral and attribute information for constructing a machine learning model that uses spectra as explanatory variables, characterized in that, The spectrum generating device includes: The input unit receives a spectral dataset and a generation request, wherein the spectral dataset consists of one or more spectra or a group of more than one set of spectra and sample information, and the generation request adjusts the generation result. The basic acquisition unit acquires one or more basic spectra that become the constituent elements of the generated pseudo-spectrum; The generation unit assigns weights to one or more basic spectra obtained by the basic acquisition unit, and generates pseudo-spectrums that are different from the spectra input to the input unit based on the weighted basic spectra. as well as The output unit outputs one or more sets of the generated pseudo-spectral and attribute information. The basic spectrum acquired by the basic acquisition unit includes at least one factor spectrum that is a component of the spectrum input to the input unit, or a spectrum obtained by performing a predetermined modification process on the factor spectrum. The attribute information output from the output unit corresponding to the pseudo spectrum includes information on at least one of the base spectrum used to generate the pseudo spectrum, the predetermined variation process, and the weight.

2. The spectrum generation device according to claim 1, characterized in that, The generation unit assigns weights to one or more basic spectra obtained by the basic acquisition unit, and calculates a weighted sum based on the basic spectra and the weights, thereby generating a pseudo-spectrum with a shape different from the spectrum input to the input unit.

3. The spectrum generation device according to claim 1, characterized in that, The basic acquisition unit processes one or more spectra input to the input unit as a tensor consisting of an axis in the sample direction and an axis in one or more spectral directions. Using tensor decomposition, the spectrum is decomposed into a set of more than one group of factor spectra and relative concentration vectors. The factor spectrum is included in the basic spectrum. The relative concentration vector is a vector of the same size as the number of spectra.

4. The spectrum generation device according to claim 3, characterized in that, The basic acquisition unit determines the feature extraction function based on the generated request. The optimization function of the tensor decomposition method has a regularization term. The purpose of this regularization term is to make the output of the feature extraction function when the spectrum input to the input part is taken as the input, similar to the output of the feature extraction function when the reconstruction result of the group of factor spectra and concentration vectors as the tensor decomposition result is taken as the input. The generation unit acquires the output of the feature extraction function and includes the output in the attribute information.

5. The spectrum generation apparatus according to claim 1, characterized in that, The input unit receives one or more spectra associated with a standard sample, and the base acquisition unit includes the spectrum in the base spectrum.

6. The spectrum generation apparatus according to claim 2, characterized in that, The generation unit sets the weights according to an arbitrary probability distribution, and includes the probability distribution information in the attribute information.

7. The spectrum generation apparatus according to claim 3 or 4, characterized in that, The generation unit sets the weights according to the probability distribution of the statistics of the concentration vector included in the parameters, and includes the probability distribution information in the attribute information.

8. A machine learning model building apparatus, characterized in that, The machine learning model building apparatus includes the spectral generation apparatus according to any one of claims 1 to 7, which uses a combination of pseudo-spectrum and basic information to construct a machine learning model with spectrum as input. The machine learning model building device includes: The design department receives arbitrary function requests for adjusting the machine learning model as the object of construction, uses the spectral generation device to construct training pseudo-spectral datasets and validation pseudo-spectral datasets, and also defines the specifications of the machine learning model. The learning department uses the training pseudospectral dataset to learn the machine learning model; The verification department uses the pseudo-spectral dataset for verification to verify the machine learning model; as well as The model output unit outputs the verification results of the machine learning model and the verification device. The verification unit verifies the performance of the machine learning model learned by the learning unit, and outputs a function request to the design unit based on the verification results, thereby expanding the training pseudo-spectral dataset for further learning and verification of the machine learning model. The design department takes the pseudo-spectrum as input and designs an objective function based on the attribute information.

9. The machine learning model building apparatus according to claim 8, characterized in that, The design department includes a constraint term in the objective function used for learning the machine learning model. The purpose of the constraint term is to make the latent variables or outputs of the machine learning model similar when pseudo-spectral data with similar attribute information is used as input.

10. The machine learning model building apparatus according to claim 8, characterized in that, The verification unit includes in the verification result the output of the machine learning model when each pseudospectrum in the pseudospectral dataset used for verification is taken as input, and data representing the relational nature of attribute information.

11. A method for generating spectra, comprising generating a set of pseudo-spectral and attribute information for constructing a machine learning model with spectra as explanatory variables, characterized in that, The spectrum generation method performs the following steps: The input step receives a spectral dataset and an arbitrary generation request, wherein the spectral dataset consists of one or more spectra or a group of more than one set of spectra and sample information, and the generation request adjusts the generation result; The basic acquisition steps involve acquiring one or more fundamental spectra that become the constituent elements of the generated pseudo-spectrum. The generation step assigns weights to one or more basic spectra obtained in the basic acquisition step, calculates a weighted sum based on the basic spectra and the weights, thereby generating a pseudo-spectrum with a different shape than the spectrum received in the input step; and The output step outputs one or more sets of the pseudo-spectral and attribute information. The basic spectrum obtained in the basic acquisition step includes at least one factor spectrum that is a component of the spectrum input in the input step, or a spectrum obtained by subjecting the factor spectrum to arbitrary modification. The attribute information includes information on at least one of the base spectrum used for generation, the variation processing, and the weights.

12. The spectral generation method according to claim 11, characterized in that, In the basic acquisition step, one or more spectra input in the input step are processed as a tensor consisting of the axis of the sample direction and the axes of one or more spectral directions. Using tensor decomposition, the spectrum is decomposed into a set of more than one group of factor spectra and relative concentration vectors. The factor spectrum is included in the base spectrum, and the relative concentration vector is a vector of the same size as the number of spectra.

13. The spectral generation method according to claim 12, characterized in that, In the basic acquisition step, The feature extraction function is determined based on the generated request. The optimization function of the tensor decomposition method has a regularization term. The purpose of this regularization term is 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 the reconstruction result of the group using the factor spectrum and concentration vector as the tensor decomposition result is used as input. In the generation step, the output of the feature extraction function that takes the pseudo-spectrum as input is obtained, and the output is included in the attribute information.

14. The spectral generation method according to claim 11, characterized in that, In the input step, one or more spectra associated with a standard sample are received, and in the base acquisition step, the spectra are included in the base spectrum.

15. The spectral generation method according to claim 11, characterized in that, In the generation step, the weights are set according to an arbitrary probability distribution, and the probability distribution information is included in the attribute information.

Images