Spectral decomposition device and spectral decomposition method

Abstract

The present invention provides a spectral decomposition device that makes it possible to achieve accuracy equal to or greater than that of decomposition by optimization, that does not require the same number of supervised learning processes as the number of decompositions, and that can be constructed with a single machine learning model. A spectral decomposition device (10) according to the present invention comprises: an input unit (101) for inputting an optical spectrum; a decomposition result candidate prediction unit (102) for obtaining decomposition result candidates, which are candidates for results of decomposing one optical spectrum into basis spectra, and degrees of certainty, which indicate the prediction reliabilities for the decomposition result candidates; a decomposition result candidate selection unit (103) for selecting, on the basis of the degrees of certainty, a decomposition result, which is a result of decomposing one optical spectrum into basis spectra, from among the decomposition result candidates; and an output unit (104) for outputting one or both of the decomposition result candidates and the decomposition result. The decomposition result candidate prediction unit (102) inputs one optical spectrum and obtains, using a machine learning model, not less than one set of a decomposition result candidate, which includes basis spectra and weights indicating the proportions of the basis spectra, and a degree of certainty for the decomposition result candidate.

Description

Spectral decomposition apparatus and spectral decomposition method

[0001] This invention relates to an apparatus and method for decomposing a spectrum obtained from a sample.

[0002] The physical property information provided by a sample can be quantified as a spectral representation. For example, a three-dimensional fluorescence spectrum (fluorescence fingerprint) 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. The fluorescence characteristics of the sample are quantified as a two-dimensional array.

[0003] The shape and intensity of the spectral spectrum obtained by measuring a sample can be considered as the result of superimposing spectral spectra, each with a unique shape for each component contained in the sample. In the following, the spectral spectrum for each component will be referred to as the base spectrum, and the index representing the mixing ratio of each base spectrum that constitutes the spectral spectrum will be referred to as the weight.

[0004] Spectral decomposition is a technique that predicts which ground spectra are superimposed on a given spectral spectrum obtained from a sample, and how weighted these superimpositions are. By using spectral decomposition to predict the ground spectra and their weights (i.e., decomposing the spectral spectrum), it is possible to observe the composition of a sample.

[0005] Conventional spectral decomposition methods involve preparing multiple spectral lines, having the observer define a model equation based on an assumed number of decompositions (the number of basis spectra to constitute the line), and then predicting the decomposed basis spectra and weights through optimization that minimizes the error between the multiple spectral lines and the model equation. Such methods require the observer to set the number of decompositions, necessitating specialized knowledge. Furthermore, optimal decomposition cannot be achieved without collecting a sufficient number of spectral lines for optimization, resulting in time-consuming and costly sample collection and spectral measurement. Additionally, because the above-mentioned error is used as the objective function in optimization, spectral lines with characteristics such as closely spaced base peaks or relatively small weights compared to other weights may not achieve proper decomposition when using error-based optimization.

[0006] An example of a conventional spectral decomposition device is described in Patent Document 1. In the spectral data analysis device described in Patent Document 1, by adding a regularization term to the model equation, it is possible to obtain decomposition results that are closer to the true ground spectrum compared to a device that uses a model equation without a regularization term.

[0007] Conventional spectral decomposition techniques involve collecting multiple pairs of decomposition results obtained by optimizing multiple spectral spectra, and then using a training dataset obtained by preparing multiple pairs of input data (a single spectral spectrum) and training data (its decomposition result) to supervise learning a machine learning model that infers the number of decompositions, and a machine learning model that infers the decomposition results prepared according to the number of decompositions.

[0008] Japanese Patent Publication No. 2019-87042

[0009] The spectral data analysis device described in Patent Document 1 has the problem that it requires time and cost for sample collection and spectral measurement because optimal resolution cannot be achieved unless a sufficient number of spectral data are collected for optimization.

[0010] Conventional spectral decomposition techniques involve learning decomposition patterns obtained through optimization. However, such techniques cannot learn decomposition patterns that are difficult to decompose through optimization, and cannot achieve decomposition accuracy higher than that of optimization. Furthermore, since it is necessary to train the same number of machine learning models as the number of decompositions, if the same number of datasets as the number of decompositions cannot be prepared, it is not possible to perform decomposition according to the number of decompositions. For this reason, such techniques require the preparation of the same number of datasets as the number of decompositions, which increases the time and cost involved in sample collection and spectral measurement.

[0011] The objective of the present invention is to provide a spectral decomposition device and spectral decomposition method that can achieve higher accuracy than decomposition by optimization, do not require the same number of supervised learning sessions as the number of decompositions, and can be constructed with a single machine learning model.

[0012] The spectral decomposition apparatus according to the present invention comprises an input unit that inputs a spectral spectrum obtained by measuring a sample, with the spectrum showing the components of the sample as the base spectrum, an input unit that inputs the spectral spectrum, an input unit that predicts a candidate decomposition result which is a candidate result of decomposing one spectral spectrum into the base spectrum, and an accuracy indicating the predictive reliability of the candidate decomposition result, an input unit that selects a candidate decomposition result which is a result of decomposing one spectral spectrum into the base spectrum from the candidate decomposition result based on the accuracy, and an output unit that outputs one or both of the candidate decomposition result and the decomposition result. The input unit that predicts a candidate decomposition result inputs one spectral spectrum and uses a machine learning model to obtain one or more pairs of candidate decomposition result which include the base spectrum and a weight indicating the composition ratio of the base spectrum, and the accuracy for the candidate decomposition result.

[0013] The spectral decomposition method according to the present invention is performed on a spectral decomposition device and comprises a spectral spectrum obtained by measuring a sample, wherein the spectrum showing the components of the sample is used as the base spectrum, and includes an input step in which the spectral decomposition device receives the spectral spectrum, a decomposition result candidate prediction step in which the spectral decomposition device obtains a decomposition result candidate, which is a candidate result of decomposing one of the spectral spectra into the base spectrum, and an accuracy indicating the predictive reliability of the decomposition result candidate, a decomposition result candidate selection step in which the spectral decomposition device selects a decomposition result, which is a result of decomposing one of the spectral spectra into the base spectrum, from among the decomposition result candidates based on the accuracy, and an output step in which the spectral decomposition device outputs one or both of the decomposition result candidates and the decomposition result. In the decomposition result candidate prediction step, the spectral decomposition device receives one of the spectral spectra as input and obtains one or more pairs of the decomposition result candidate, which includes the base spectrum and a weight indicating the composition ratio of the base spectrum, and the accuracy for the decomposition result candidate, using a machine learning model.

[0014] According to the present invention, it is possible to provide a spectral decomposition device and spectral decomposition method that can achieve higher accuracy than decomposition by optimization, do not require the same number of supervised learning sessions as the number of decompositions, and can be constructed with a single machine learning model.

[0015] This figure shows an example of the hardware configuration of a spectral decomposition device. This figure shows an example of the functional blocks of a spectral decomposition device. This figure shows an example of fluorescence fingerprint data, which is an example of a spectral spectrum. This figure shows an example of a visualized fluorescence fingerprint. This figure shows an example of fluorescence fingerprint data including missing values. This figure shows an example of a candidate decomposition result when N in equation (1) is 4. This flowchart shows an example of a method for creating composite data. This figure shows an example of a window that the output device displays on the monitor display. This is an example of a flowchart showing the processes performed by the spectral decomposition device.

[0016] The spectral decomposition apparatus and spectral decomposition method according to the present invention take a spectral spectrum as input and uses a machine learning model that calculates the accuracy of one or more sets of base spectra and weights, and the sets with high accuracy as the decomposition result, thereby automatically determining the decomposition result, including the number of decompositions. Synthetic data obtained from pseudo-base spectra and weights is used as training data for the machine learning model, and the decomposition result is learned using the pre-synthesis information as input to the synthesized data. In this invention, by comprehensively setting the proximity of peaks between pseudo-base spectra and the magnitude of weights, a training dataset is constructed from a large amount of synthesized data, including data that is difficult to decompose by optimization, making it possible to perform decomposition with a higher accuracy than decomposition by optimization. Furthermore, by using the accuracy metric, it is possible to train a single machine learning model on decomposition patterns with various numbers of decompositions, so the training of the machine learning model does not require the same number of training datasets as the number of decompositions.

[0017] In this invention, a single spectral spectrum can be decomposed into multiple sets of basis spectra and weights with a precision exceeding that of decomposition by optimization, without having to prepare the same number of training datasets as the number of decompositions in advance and train a machine learning model with these training datasets.

[0018] Hereinafter, a spectral decomposition apparatus and spectral decomposition method according to embodiments of the present invention will be described with reference to the drawings. The spectral decomposition apparatus according to this embodiment can perform the spectral decomposition method according to this embodiment. In the drawings referenced herein, the same or corresponding components are denoted by the same reference numerals, and repeated descriptions of these components may be omitted.

[0019] The spectral decomposition device according to this embodiment performs spectral decomposition by inferring and obtaining one or more decomposition result candidates, each consisting of a base spectrum and a weight set, from a single spectral spectrum obtained by measuring a sample, and by inferring and obtaining an accuracy indicating the prediction reliability for each decomposition result candidate, and selecting a decomposition result candidate according to the accuracy. The base spectrum is a spectrum that constitutes the spectral spectrum and is a spectrum that shows the components of the sample. In this embodiment, one or more decomposition result candidates are output as the decomposition result of a single spectral spectrum.

[0020] The spectral decomposition apparatus according to this embodiment comprises, for example, a computing unit and a storage device that stores a computer program executed by the computing unit. The computing unit receives a single spectral spectrum as input and outputs a decomposition result corresponding to the input spectral spectrum.

[0021] In this embodiment of spectral decomposition, a machine learning model that does not require the same number of supervised learning sessions as the number of decompositions can be used to infer decomposition results, including the number of decompositions, from a single spectral spectrum, consisting of a set of base spectra and weights, with an accuracy exceeding that of optimization.

[0022] In this embodiment, one or more candidate resolution results and their accuracy are obtained from a single spectral spectrum. One or more candidate resolution results are selected from among the candidates according to their accuracy, and the selected one or more candidate resolution results are output as the spectral spectrum resolution results. Each candidate resolution result consists of a pair of a basis spectrum and a weight. The accuracy indicates the predictive confidence of the candidate resolution result.

[0023] The spectral decomposition apparatus according to this embodiment will be explained using Figures 1 to 6.

[0024] <Hardware Configuration of Spectral Decomposition Device> Figure 1 shows an example of the hardware configuration of the spectral decomposition device 10.

[0025] The spectral decomposition device 10 comprises an arithmetic unit 12, a storage device 11, an interface 13, and a bus 14. The spectral decomposition device 10 can be connected to external devices, such as a spectral measuring device 21, a control device 20, and an output device 22. The spectral decomposition device 10 may also be connected to an input device (not shown) for the user to input instructions or data.

[0026] The arithmetic unit 12, the storage device 11, and the interface 13 send and receive information from each other via the bus 14.

[0027] The arithmetic unit 12 is a device that performs various processes within the spectral decomposition unit 10, and can be composed of, for example, a CPU (Central Processing Unit) or an FPGA (Field-Programmable Gate Array). The functions realized by the arithmetic unit 12 will be described later.

[0028] The storage device 11 is a storage device that stores programs executed by the arithmetic unit 12, parameters of machine learning models, a database of base spectra and substance names, and processing results of the spectral decomposition device. The storage device 11 can be composed of, for example, an HDD (Hard Disk Drive), flash memory, an SSD (Solid State Drive), RAM (Random Access Memory), and ROM (Read Only Memory).

[0029] Interface 13 is a communication device that transmits and receives signals to and from devices located outside the spectral decomposition device 10, such as a spectral measuring device 21, a control device 20, and an output device 22.

[0030] The spectral measuring device 21 measures the spectrum of the sample and outputs the measurement result.

[0031] The control device 20 controls the spectrum measuring device 21.

[0032] The output device 22 outputs the processing results from the spectral decomposition device 10. The output device 22 is composed of, for example, a monitor display or a printer. The output device 22 may also be composed of a head-mounted display.

[0033] The means by which the spectral decomposition device 10 transmits and receives signals to and from external devices (for example, a spectral measuring device 21, a control device 20, and an output device 22) can be any means, for example, a wired connection using a cable such as an optical fiber, or a wireless connection using wireless communication technology such as Bluetooth®.

[0034] The spectral decomposition device 10 does not necessarily need to be installed in the same facility as the control device 20, spectral measuring device 21, and output device 22; it may be located on a separate computer or in a so-called cloud. In such cases, a LAN (Local Area Network) or WAN (Wide Area Network) can be used to transmit and receive signals between the interface 13 and external devices (for example, the spectral measuring device 21, the control device 20, and the output device 22).

[0035] <Functional Configuration of Spectrum Decomposition Device 10> Figure 2 shows an example of the functional blocks of the spectrum decomposition device 10. The spectrum decomposition device 10 comprises an input unit 101, a decomposition result candidate prediction unit 102, a decomposition result candidate selection unit 103, and an output unit 104 as functional units.

[0036] The functional units, namely the input unit 101, the decomposition result candidate prediction unit 102, the decomposition result candidate selection unit 103, and the output unit 104, may be implemented by an arithmetic unit 12 that executes a predetermined computer program, or they may be implemented by dedicated hardware. These functional units will be described below.

[0037] The spectral decomposition device 10 is designed to handle a single type of spectral spectrum (e.g., fluorescence fingerprints, Raman spectra, and fluorescence spectra). The spectral decomposition device 10 according to this embodiment is capable of performing spectral decomposition on various spectral spectra.

[0038] The input unit 101 receives a single spectral spectrum input from the interface 13. Furthermore, the input unit 101 receives setting values ​​for the spectral decomposition device 10 from an external source. These setting values ​​include the machine learning model used by the decomposition result candidate prediction unit 102 (described later), hyperparameters related to the selection of decomposition result candidates used by the decomposition result candidate selection unit 103, and the output format of the output unit 104. The data format of the spectral spectrum input to the input unit 101 is a one-dimensional or more array data composed of one or more axes, with each element storing the corresponding spectral intensity.

[0039] The decomposition result candidate prediction unit 102 determines candidate decomposition results, which are the results of decomposing a single spectral spectrum input by the input unit 101 into a base spectrum, and the accuracy of each candidate decomposition result. Specifically, for a single spectral spectrum, the decomposition result candidate prediction unit 102 determines candidate decomposition results consisting of one or more sets of base spectra and weights, and the accuracy indicating the prediction reliability of each candidate decomposition result. For example, the decomposition result candidate prediction unit 102 takes a single spectral spectrum as input and uses a machine learning model to determine one or more sets of candidate decomposition results including base spectra and weights, and their accuracy. The decomposition result candidate prediction unit 102 can train the machine learning model through supervised learning.

[0040] The decomposition result candidate selection unit 103 selects a decomposition result (final decomposition result), which is the result of decomposing one spectral spectrum into a base spectrum, from among the decomposition result candidates based on their accuracy. Specifically, the decomposition result candidate selection unit 103 selects one or more decomposition result candidates from among the decomposition result candidates based on the decomposition result candidates and their accuracy obtained by the decomposition result candidate prediction unit 102, and uses the selected decomposition result candidates as the final decomposition result. For example, the decomposition result candidate selection unit 103 selects a decomposition result candidate whose accuracy is greater than a predetermined threshold as the final decomposition result. Alternatively, for example, the decomposition result candidate selection unit 103 selects a decomposition result candidate whose accuracy is close to a predetermined value (for example, close to 1) as the final decomposition result. In this way, the decomposition result candidate selection unit 103 predicts the decomposition result of a spectral spectrum.

[0041] The output unit 104 outputs either a candidate for the disassembly result or the actual disassembly result, or both. The output unit 104 may output the accuracy along with the candidate for the disassembly result, or it may output the accuracy along with the actual disassembly result.

[0042] It should be noted that the functions of the spectral decomposition device 10 do not necessarily have to be configured as shown in Figure 2 (input unit 101, decomposition result candidate prediction unit 102, decomposition result candidate selection unit 103, and output unit 104). The spectral decomposition device 10 can be equipped with any functional units as long as the processing corresponding to the operation of each functional unit can be realized.

[0043] Figure 9 is an example of a flowchart showing the processes performed by the spectral decomposition device 10. Steps S901 to S904 shown in Figure 9 correspond to the input unit 101, the decomposition result candidate prediction unit 102, the decomposition result candidate selection unit 103, and the output unit 104, which are functional units shown in Figure 2.

[0044] The input step S901 is performed by the input unit 101. In the input step S901, the input unit 101 receives a single spectral signal from the interface 13. Furthermore, in the input step S901, the input unit 101 receives a set value from an external source. This set value includes the machine learning model used in the decomposition result candidate prediction step S902 (described later), hyperparameters related to the selection of decomposition result candidates used in the decomposition result candidate selection step S903, and the output format in the output step S904.

[0045] The decomposition result candidate prediction step S902 is performed by the decomposition result candidate prediction unit 102. In the decomposition result candidate prediction step S902, the decomposition result candidate prediction unit 102 determines, for one spectral spectrum input in the input step S901, a decomposition result candidate consisting of one or more sets of base spectra and weights, and an accuracy indicating the prediction reliability of each decomposition result candidate.

[0046] The decomposition result candidate selection step S903 is performed by the decomposition result candidate selection unit 103. In the decomposition result candidate selection step S903, the decomposition result candidate selection unit 103 selects one or more decomposition result candidates from among the decomposition result candidates based on the decomposition result candidates and their accuracy obtained in the decomposition result candidate prediction step S902, and uses the selected decomposition result candidates as the final decomposition result.

[0047] The output step S904 is performed by the output unit 104. In the output step S904, the output unit 104 outputs either a candidate for disassembly result or the actual disassembly result, or both. The output unit 104 may output the accuracy along with the candidate for disassembly result, or it may output the accuracy along with the actual disassembly result.

[0048] <Configuration and Operation of Each Part> Below, we will explain in detail the operation of the input unit 101, the decomposition result candidate prediction unit 102, the decomposition result candidate selection unit 103, and the output unit 104, as shown in Figure 2.

[0049] The input unit 101 receives a single spectral data and setting values ​​for the spectral decomposition device 10 via the interface 13. The data format of the spectral data is a one-dimensional or more-dimensional array data composed of one or more axes. For example, if the spectral data is a fluorescence fingerprint, the data format of the spectral data is a two-dimensional array composed of two axes, excitation wavelength and emission wavelength, with emission intensity stored in each element.

[0050] The input unit 101 outputs the input spectral spectrum to the decomposition result candidate prediction unit 102. The input unit 101 also outputs the input setting value to the function units that use the setting value (decomposition result candidate prediction unit 102, decomposition result candidate selection unit 103, and output unit 104).

[0051] Figure 3 shows an example of fluorescence fingerprint data 30, which is an example of a spectral spectrum. The spectral spectrum is assumed to have a measurement band and emission intensity within the measurement band recorded with an arbitrary quantization width. That is, the fluorescence fingerprint data 30 has the excitation wavelength and emission wavelength, which are the measurement bands of the spectral spectrum, and the emission intensity corresponding to the combination of excitation wavelength and emission wavelength, recorded with an arbitrary quantization width.

[0052] As shown in Figure 3, in the fluorescence fingerprint data 30, the excitation wavelength, emission wavelength, and emission intensity are recorded with an arbitrary quantization width. Figure 3 shows an example of a fluorescence fingerprint measured with a quantization width of 5 nm from 200 nm to 700 nm for both the excitation and emission light. For example, in the fluorescence fingerprint data 30, the first row, first column records the emission intensity of the emitted light at 200 nm in relation to the excitation light at 200 nm, the second row, second column records the emission intensity of the emitted light at 205 nm in relation to the excitation light at 200 nm, and the first row, first column records the emission intensity of the emitted light at 200 nm in relation to the excitation light at 205 nm.

[0053] The fluorescence fingerprint data 30 shown in Figure 3 is a 101 × 10¹ matrix data because the fluorescence fingerprint was recorded with a quantization width of 5 nm from 200 nm to 700 nm relative to the excitation and emission light.

[0054] In the following, the spectrum consisting of all emission wavelengths at a particular excitation wavelength will be called the emission spectrum, and the spectrum consisting of all excitation wavelengths at a particular emission wavelength will be called the excitation spectrum.

[0055] Figure 4 shows an example of a visualized fluorescent fingerprint 31. The visualized fluorescent fingerprint 31 is shown as contour data with the excitation wavelength on the vertical axis and the emission wavelength on the horizontal axis, displaying the emission intensity for each combination of excitation wavelength and emission wavelength.

[0056] The visualized fluorescent fingerprint 31 shown in Figure 4 is an example of the fluorescent fingerprint data 30 shown in Figure 3 being visualized as a two-dimensional contour map. When humans observe fluorescent fingerprints, they are often displayed as the visualized fluorescent fingerprint 31 shown in Figure 4. However, the visualized fluorescent fingerprint 31 is not limited to contour data; it may also be displayed in other formats, such as grayscale images or heatmap images where emission intensity is assigned to brightness values ​​or colors, or as a bird's-eye view. The shape of the fluorescent fingerprint will vary depending on the properties of the fluorescence characteristics being studied.

[0057] Let's return to the explanation of the input unit 101 (Figure 2).

[0058] The setting values ​​input to the input unit 101 are parameters that the observer (user) uses to adjust the settings of the decomposition result candidate prediction unit 102, the decomposition result candidate selection unit 103, and the output unit 104. The spectral decomposition device 10 changes various settings according to the setting values. The setting values ​​are input to the input unit 101 by the user using an input device such as a mouse or keyboard to connect from the control device 20 to the interface 13.

[0059] The measurement bandwidth and quantization width of the spectral spectrum input to the input unit 101 may be limited to fixed values ​​or may be variable within a certain range. When the measurement bandwidth and quantization width are fixed, the decomposition result candidate prediction unit 102, the decomposition result candidate selection unit 103, and the output unit 104 shall be designed assuming a fixed measurement bandwidth and quantization width. When the measurement bandwidth and quantization width are variable within a certain range, the output to the decomposition result candidate prediction unit 102 is provided in a form that stores the emission intensity of the actual spectrum at the minimum value of the corresponding measurement bandwidth and quantization width.

[0060] Figure 5 shows an example of fluorescence fingerprint data 30 that includes missing values. The fluorescence fingerprint data 32 shown in Figure 5 is an example in which the emission intensity of the fluorescence fingerprint data 30 shown in Figure 3 is stored in a space of 1 nm quantization width within the applicable measurement band, where both the excitation wavelength and emission wavelength are from 200 nm to 700 nm, and the minimum quantization width is 1 nm. Since the quantization width of the fluorescence fingerprint data 30 shown in Figure 3 is 5 nm, the emission intensity is stored at 5 nm intervals in the fluorescence fingerprint data 32 shown in Figure 5.

[0061] In the fluorescence fingerprint data 32 shown in Figure 5, x represents a missing value. The input unit 101 may interpolate the missing value using polynomial fitting or Gaussian process regression with the values ​​stored in the fluorescence fingerprint data 32. The input unit 101 outputs the fluorescence fingerprint data 32 with the interpolated missing values ​​as a spectral spectrum to the decomposition result candidate prediction unit 102. Alternatively, the input unit 101 may output the fluorescence fingerprint data 32 with fixed values ​​such as 0 or -1 stored for the missing values ​​to the decomposition result candidate prediction unit 102. Alternatively, the input unit 101 may output the fluorescence fingerprint data 32 containing the missing values ​​as they are to the decomposition result candidate prediction unit 102. A method for performing spectral decomposition on spectral spectrum data containing missing values ​​will be explained later.

[0062] The input unit 101 may perform any preprocessing on the spectral spectrum output to the candidate resolution prediction unit 102. This preprocessing is intended, for example, to suppress the effects of noise such as Rayleigh scattering.

[0063] In the visualized fluorescence fingerprint 31 shown in Figure 4, the fluorescence fingerprint contains signals that act as noise. For example, wavelength bands containing Rayleigh scattering are generally considered noise components when analyzing fluorescence properties from fluorescence fingerprints. In Rayleigh scattering, it is not the fluorescence emitted by the sample that is scattered, but the light reflected by the sample. Therefore, when Rayleigh scattering is detected, features unrelated to fluorescence properties are reflected in the fluorescence fingerprint.

[0064] Rayleigh scattering is characterized by its regional detection in wavelength bands where the excitation wavelength is a natural number multiple or reciprocal multiple of the emission wavelength, due to the influence of the diffraction grating. For this reason, for example, a pre-processing step can be performed by arbitrarily setting a wavelength that matches the above condition and a wavelength band in the vicinity of this wavelength, and converting the emission intensity in this wavelength band to an arbitrary fixed value such as zero. Alternatively, a pre-processing step can be performed by measuring the fluorescence fingerprint of a standard sample and subtracting the fluorescence fingerprint of the standard sample from each spectrum.

[0065] The input unit 101 can suppress the influence of noise by performing such preprocessing, for example. The input unit 101 may also perform preprocessing in any manner on noise caused by Raman scattering, artifacts, or the like to suppress the influence of the noise.

[0066] The description returns to the decomposition result candidate prediction unit 102, the decomposition result candidate selection unit 103, and the output unit 104 shown in FIG. 2. Next, the decomposition result candidate prediction unit 102 will be described.

[0067] The decomposition result candidate prediction unit 102 receives one spectral spectrum from the input unit 101, and uses a machine learning model trained by supervised learning for the input one spectral spectrum to obtain one or more decomposition result candidates and the accuracy of each decomposition result candidate. The decomposition result candidate includes a set of a base spectrum constituting the spectral spectrum and a weight indicating the composition ratio of the base spectrum in the sample. The accuracy is the prediction reliability of each decomposition result candidate. The higher the accuracy, the higher the possibility that the base spectrum of the decomposition result candidate corresponding to this accuracy is included in the spectral spectrum (that is, the prediction reliability is higher).

[0068] In the following, taking a fluorescence fingerprint as an example of a spectral spectrum, an example of a method for obtaining decomposition result candidates and accuracy using a machine learning model trained by supervised learning will be described.

[0069] Let F be the spectral spectrum (fluorescence fingerprint) input by the input unit 101, and let D(·) be the machine learning model used by the decomposition result candidate prediction unit 102. The output D(F) represents N decomposition result candidates â shown in Equation (1) １ , â ２ , ..., â Ｎ and the accuracy p １ , p ２ , ..., p Ｎ . Further, the n-th decomposition result candidate â ｎ is composed of the base spectrum Êx related to the excitation wavelength shown in Equation (2) ｎ , the base spectrum Êm related to the emission wavelength ｎ , and the weight ŵ ｎ .

[0070]

[0071]

[0072] Each ground spectrum corresponds to the excitation wavelength axis and emission wavelength axis of the fluorescence fingerprint, and the excitation wavelength of the fluorescence fingerprint F is S Ｅｘ Let the emission wavelength be S Ｅｍ Therefore, ^Ex ｎ The number of elements is S Ｅｘ It is a one-dimensional vector, and ^Em ｎ The number of elements is S Ｅｍ It is a one-dimensional vector. And ^Ex ｎ and ^Em ｎ Cartesian product ^Ex ｎ (×)^Em ｎ is a matrix of the same form as F, and its Cartesian product ^Ex ｎ (×)^Em ｎ ni^w ｎ Multiplied by ^w ｎ (^Ex ｎ (×)^Em ｎ ) is ^a in F ｎ The constituent elements are shown.

[0073] The number of types of ground spectra depends on the type of spectral spectrum. If the spectral spectrum is a fluorescence fingerprint, there are two types of ground spectra: a ground spectrum related to the excitation wavelength and a ground spectrum related to the emission wavelength.

[0074] Figure 6 shows an example of a candidate decomposition result 60 when N in equation (1) is set to 4. As an example, Figure 6 shows four candidate decomposition results 60, i.e., four sets of base spectrum 62 and weight 63, along with the accuracy 64 and input spectral spectrum 61 corresponding to each candidate decomposition result.

[0075] The spectral spectrum 61 is input data for the machine learning model D, which seeks candidate resolution results. In the example shown in Figure 6, the spectral spectrum 61 is a spectrum composed of two axes, excitation wavelength and emission wavelength, and is a fluorescence fingerprint illustrated using contour lines.

[0076] The base spectrum 62 is represented by the base spectrum 62a with respect to the excitation wavelength, the base spectrum 62b with respect to the emission wavelength, and the direct product 62c. The base spectrum 62a with respect to the excitation wavelength is given by ^Ex in equation (2). ｎ This is shown in a graph where the vertical axis is the excitation wavelength axis (from bottom to top) and the horizontal axis is the emission intensity axis (from right to left). The ground spectrum 62b with respect to the emission wavelength is given by ^Em in equation (2). ｎ This is shown in a graph where the vertical axis is the emission intensity axis (from top to bottom) and the horizontal axis is the emission wavelength axis (from left to right). The direct product 62c shows the direct product of the ground spectrum 62a with respect to the excitation wavelength and the ground spectrum 62b with respect to the emission wavelength, and is illustrated in the same manner as the spectral spectrum 61.

[0077] The weight 63 is a scalar value representing the weight of each basis spectrum 62.

[0078] The accuracy 64 is a scalar value representing the predictive confidence of the corresponding decomposition result candidate 60. The greater the accuracy 64 is above a predetermined threshold, for example, the closer it is to 1, the more likely it is that the corresponding decomposition result candidate 60 is included in the spectral spectrum 61.

[0079] The machine learning model D shown in equation (1) can be any machine learning model as long as it has a structure that takes a spectral spectrum F as input and produces N sets of outputs. As shown in equation (1), the machine learning model D has a spectral spectrum F as input and outputs the decomposition result candidate a １ , ^a ２ ,...,^a Ｎ and accuracy p １ , p ２ , ..., p Ｎ That is the case.

[0080] The machine learning model D is designed, for example, by fixing the wavelength band (measurement band) and quantization width of the spectral spectrum, as described in the explanation of the input unit 101.

[0081] Alternatively, when interpolating missing values, the machine learning model D uses an arbitrary multilayer perceptron that, for example, takes data obtained by smoothing a two-dimensional matrix spectral spectrum F into a one-dimensional vector as input, outputs a one-dimensional vector concatenating all N decomposition result candidates and N accuracy values, and divides the output into the form shown on the right side of equation (1) to produce N sets of outputs. In this example, the output of the multilayer perceptron is N × (S Ｅｘ +S Ｅｍ This results in a one-dimensional vector of form (+1 + 1). However, each basis spectrum is normalized to a specific value such as 1, and the weights represent the weights corresponding to the normalized basis spectrum. Regarding accuracy, since the range of values ​​must be kept within a specific range, a sigmoid function or similar is used to keep the range of values ​​within a specific range such as 0 to 1. In addition to multilayer perceptrons, any machine learning model such as CNN (Convolutional Neural Network) may be used.

[0082] For example, if the machine learning model D receives a spectral spectrum F as input in a way that includes missing values, it can treat each wavelength band where emission intensity exists as sequential data with elements (excitation wavelength, emission wavelength, emission intensity), and output N decomposition result candidates using machine learning models that handle sequential data, such as RNN (Recurrent Neural Network) or Transformer.

[0083] As an example of a learning algorithm for realizing the machine learning model D shown in equation (1), we will describe a supervised learning algorithm that uses pre-prepared decomposition results of spectral data. The method for creating the pre-prepared decomposition results will be described later.

[0084] Equations (3) and (4) show examples of the structure of training datasets used for supervised learning.

[0085]

[0086]

[0087] In equation (3), T represents the training dataset for the machine learning model D, and S represents the spectral spectrum F.Ｔ １ , F Ｔ ２ , ..., F Ｔ Ｓ And, the pre-prepared disassembly result A Ｔ １ , A Ｔ ２ , ..., A Ｔ Ｓ It is composed of the following. Each spectral data point is input data for the machine learning model D, and the pre-prepared decomposition results serve as training information used to optimize the parameters of the machine learning model D.

[0088] Equation (4) represents a certain teacher information A Ｔ Ｓ M ｓ This is a breakdown of the decomposition results of the set, and the ground spectrum Ex with respect to the excitation wavelength. Ｔ，ｓ １ , Ex Ｔ，ｓ ２ , ..., Ex Ｔ，ｓ Ｍｓ And the basis spectrum Em with respect to emission wavelength Ｔ，ｓ １ , Em Ｔ，ｓ ２ , , , Em Ｔ，ｓ Ｍｓ And the weight lol Ｔ，ｓ １ ,w Ｔ，ｓ ２ , ..., lol Ｔ，ｓ Ｍｓ It is composed of the following. However, each base spectrum is a spectrum normalized so that the maximum intensity is a constant such as 1, and each weight is a weight for the normalized spectrum.

[0089] The number of candidate decomposition results N output by the machine learning model D is M １ , M ２ , ..., M Ｓ The machine learning model D is designed so that the value is greater than or equal to the maximum value. In this way, a spectral decomposition device 10 can be designed in which the decomposition result candidate selection unit 103 can spectrally decompose within a range of 1 to N decompositions.

[0090] In supervised learning, a learning algorithm based on the output obtained when each piece of input data in the learning dataset T is input to the machine learning model D and each piece of supervision information optimizes the parameter group constituting the machine learning model D. Specifically, among N sets of decomposition result candidates, the spectral spectrum of certain learning data is defined as F ｓ Ｔ , the supervision information A for the corresponding decomposition result ｓ Ｔ In the decomposition number M ｓ sets of decomposition result candidates and the accuracy (the output D(F ｓ Ｔ ) of the machine learning model D), the parameters of the machine learning model D are optimized so that the basis spectrum and the weight are made similar to A and the accuracy becomes a value close to 1, while for (N-M ｓ Ｔ ) sets of decomposition result candidates and accuracy, the parameters of the machine learning model D are optimized so that the accuracy becomes a value close to 0 (zero). ｓ

[0091] In the following, the procedure of supervised learning for the spectral spectrum F of certain learning data is illustrated. In practice, the procedure illustrated below is applied to each spectral spectrum of the learning data to optimize the parameters of the machine learning model D. At the time of optimization, optimization using a statistical value of loss values for a plurality of samples such as batch processing may be performed. ｓ Ｔ

[0092] In order to realize the above processing, first, a cost function is set for the output D(F ｓ Ｔ ) which is composed of N sets of decomposition result candidates (sets of basis spectrum and weight) and accuracy, and the supervision information A composed of M sets of basis spectra and weights ｓ ｓ Ｔ , and bipartite matching is performed to search for M sets of outputs similar to A from the N sets of outputs. ｓ Ｔ ｓ

[0093] An example of the cost function is shown in Formula (5). In Formula (5), C ｎｍ is the learning data F ｓ ​​​​​Ｔ the n-th output among N sets of outputs of the machine learning model D for, and the teacher information A ｓ Ｔ the m-th teacher information among M sets of teacher information, represents the cost of the combination thereof.

[0094]

[0095] SIM(·) is an arbitrary function for calculating the similarity between basis spectra, and calculates the similarity between spectra obtained by concatenating the basis spectrum related to the excitation wavelength and the basis spectrum related to the emission wavelength. Without concatenating these basis spectra, the similarity may be calculated between each pair of basis spectra, and a value obtained by summing the respective similarities or an average value of the respective similarities may be taken as the similarity between the basis spectra. As the similarity, for example, cosine similarity or the like can be used. λ １ , λ ２ are coefficients for adjusting the value of each term in Equation (5). In order to perform matching with emphasis on the similarity of the basis spectra, λ １ may be set to a value larger than λ ２ .

[0096] Using Equation (5), for each of the N sets of outputs and M ｓ sets of teacher information, by calculating the cost of all combinations, an N×M matrix having each cost as an element ｓ is defined as a cost matrix, and according to this cost matrix, the linear assignment problem is optimized to perform bipartite matching. As a result of the bipartite matching, among the N sets of outputs, M sets of outputs similar to A ｓ Ｔ ｓ sets of outputs and M ｓ sets of teacher information are matched one-to-one, and (N-M ｓ ) sets of outputs become outputs that are not matched with any teacher information.

[0097] ​Next, for outputs that match the training information, the parameters of machine learning model D are optimized so that the base spectrum and weights are similar to those of the training information, while the accuracy is greater than a predetermined threshold (for example, a value close to 1). For outputs that do not match the training information, the parameters of machine learning model D are optimized so that the accuracy, or the weights and accuracy, are less than or equal to a predetermined threshold (for example, a value close to 0).

[0098] An example of a loss function for performing the above optimization is explained. A loss function is calculated for each of the N sets of outputs. If the nth output matches the mth training information, the loss function of equation (6) is calculated. If the nth output does not match the mth training information, the loss function of equation (7) is calculated. The sum of the loss function values ​​for all N outputs is given by F. Ｓ Ｔ This is defined as the loss, and the system is optimized using an arbitrary optimization function to minimize this loss.

[0099]

[0100]

[0101] The right-hand side of equation (6) consists of three terms.

[0102] Function L of the first term Ｂ This represents the loss function related to the base spectrum. This function can be set to, for example, the reciprocal of the cosine similarity or the value obtained by subtracting the cosine similarity from 1.

[0103] Function L of the second term ｗ This represents the loss function related to the weights. This function can be set to, for example, the mean squared error or the mean absolute error.

[0104] Function L of the third term Ｐ (1-p ｎ This represents the loss function related to accuracy. Since the accuracy of the matched output needs to be as close to 1 as possible, this function is set to a value obtained by subtracting the accuracy from 1 as the loss. Any loss function can be set as long as the loss is low when the accuracy is high.

[0105] The coefficients γ1, γ2, and γ3 in each term are coefficients that determine the balance of each loss.

[0106] The right-hand side of equation (7) consists of two terms.

[0107] The first term is the same function L as the second term on the right-hand side of equation (6). Ｗ Therefore, the weights of the unmatched outputs are made to approach 0 (zero). However, in the decomposition result candidate selection unit 103, it is not always necessary to use the unmatched weights, so the coefficient η2 may be set to 0 (zero) to make the loss 0 (zero).

[0108] The second term represents the loss function related to accuracy. Since it is necessary to bring the accuracy of unmatched outputs as close to zero as possible, accuracy itself is set as the loss. However, any loss function that results in a lower loss when the accuracy is low may be set.

[0109] The machine learning model D is trained by calculating the loss for all N sets of outputs using equation (6) or equation (7) and optimizing the model so that the sum of the obtained losses is minimized. In this training, the machine learning model D learns the base spectrum and weights of outputs that match the training information to be similar to the training information and to have an accuracy close to 1, and learns the weights and accuracy of outputs that do not match to be close to 0 (zero).

[0110] The learning algorithm using equations (5), (6), and (7) is merely an example. The learning algorithm involves selecting the spectral spectrum of a certain training data from among N pairs of candidate decomposition results and dividing it by F. ｓ Ｔ In this case, the corresponding decomposition result is the training information A ｓ Ｔ Number of decompositions in M ｓ Candidate decomposition results for the number of pairs and their accuracy (output D(F) of machine learning model D) ｓ Ｔ In )), the basis spectrum and weights are A ｓ Ｔ While making it similar to the original model, and optimizing the parameters of the machine learning model D so that the accuracy is close to 1, (N-M ｓAny learning algorithm can be used as long as it can optimize the parameters of the machine learning model D so that the probability of the decomposition result of the number of pairs is close to 0 (zero).

[0111] In this embodiment, by using a machine learning model trained with the illustrated learning algorithm, candidate decomposition results and their accuracy can be obtained simultaneously. The candidate decomposition result selection unit 103 automatically selects candidate decomposition results based on the accuracy, thereby enabling spectral decomposition, including the number of decompositions.

[0112] Conventional techniques sometimes separate the machine learning model that infers the number of decompositions from the machine learning model that performs the decomposition. In such techniques, it is necessary to prepare as many machine learning models for spectral decomposition as there are decompositions, and therefore it is necessary to collect and train sufficient training data according to the number of decompositions.

[0113] On the other hand, in the learning algorithm of this embodiment, a single machine learning model can infer the number of decompositions and learn patterns of spectral decomposition for multiple decompositions by matching with training information using an indicator called accuracy. This results in low data collection costs for learning, high learning efficiency, and the output of an accuracy indicator, which is a predictive confidence index for each candidate decomposition result. Therefore, in this embodiment, as will be described later, the candidate decomposition result selection unit 103 can perform flexible spectral decomposition according to the accuracy setting value set by the user.

[0114] In this embodiment, the machine learning model D is constructed as a model that infers candidate decomposition results and their probability by learning pre-prepared decomposition results as shown in equations (3) and (4).

[0115] Conventional techniques include preparing multiple patterns of spectral decomposition results using PARAFAC (PARArell Factor Analysis), a type of tensor decomposition method, and using them as training data for training.

[0116] PARAFAC is a spectral decomposition method that defines the number of resolutions (number of base spectra), the number of base spectra common to all spectral spectra, and the weights of each base spectrum corresponding to each spectral spectrum, and performs spectral decomposition by optimization based on the weighted sum of the base spectra and the error of the spectral spectra. However, as mentioned above, spectral decomposition methods that involve optimization based on errors require the preparation of multiple patterns of spectral spectra to be decomposed, and in the case of spectral spectra with characteristics such as the peaks of the base spectra being close together, or the weight of the base spectrum being relatively small compared to the other weights, proper decomposition may not be achieved.

[0117] The conventional techniques described above allow for the acquisition of decomposition results from a single spectral spectrum without optimization using a pre-trained machine learning model. However, because the decomposition results obtained by PARAFAC are used as training data for the machine learning model, it is not possible to achieve a decomposition accuracy higher than that of PARAFAC, and there is a problem in that it cannot handle cases where PARAFAC is insufficient for proper spectral decomposition. Furthermore, collecting a sufficient quantity of samples for machine learning with an appropriate number of decompositions using PARAFAC presents challenges, including the high level of expertise and arbitrariness required in determining the number of decompositions, as well as the high cost of collecting spectral data.

[0118] In this example, a pre-trained machine learning model can be used to obtain decomposition results from a single spectral spectrum without optimization. The machine learning model does not use decomposition results obtained by collecting multiple actual spectral spectra and optimizing them, as in PARAFAC, as training information. Instead, it uses a set of base spectra and weights that are pseudo-defined as if they were actual decomposition results as training information for the decomposition results, and is trained using a training dataset that uses the synthesized data obtained from these sets as input data.

[0119] In other words, in this embodiment, instead of using data obtained using PARAFAC as training data, data obtained using the above-mentioned synthetic data is used. Therefore, accuracy exceeding that achieved by optimization can be obtained, and a single machine learning model can be used without requiring the same number of supervised training data as the number of decompositions.

[0120] The following is an example of a method for creating synthetic data for fluorescent fingerprints.

[0121] Equation (8) is a model equation that shows the synthesized data of a fluorescence fingerprint using a pseudo-defined set of base spectra and weights. In equation (8), the synthesized data F Ｐ This is the pseudo-ground spectrum Ex of group M with respect to the excitation wavelength. Ｐ １ , Ex Ｐ ２ , ..., Ex Ｐ Ｍ And the basis spectrum Em with respect to emission wavelength Ｐ １ , Em Ｐ ２ , , , Em Ｐ Ｍ And the weight lol Ｐ １ ,w Ｐ ２ , ..., lol Ｐ Ｍ The set and the error term e Ｐ It is composed of.

[0122]

[0123] Equation (8) is a model equation also used in PARAFAC, and it is a model equation that conforms to the physical properties of fluorescent fingerprints. That is, equation (8) conforms to the physical properties of fluorescent fingerprints, in which the fluorescent fingerprint is expressed as a weighted sum of the direct products of the ground spectra for the excitation wavelength and the emission wavelength corresponding to each component (fluorescent molecule) contained in the sample, and the error term includes Rayleigh scattering as described above. By simplifying the components of equation (8), synthetic data can be created.

[0124] Figure 7 is a flowchart showing an example of a method for creating composite data. The decomposition result candidate prediction unit 102 can create composite data, for example, by following the steps in the flowchart shown in Figure 7.

[0125] The following describes an example of creating composite data following the procedure shown in Figure 7. Note that the flowchart in Figure 7 shows one example of a method for creating composite data. Composite data can be created using any procedure, as long as a pseudo-base spectrum, weights, and error terms are defined, and the composite data is defined by equation (8).

[0126] In S71, composite data F Ｐ The number M of pseudo-base spectra constituting the model is randomly defined within the range of 1 to the number N of outputs of the machine learning model D. The number M can be defined in any way; for example, it may be a value sampled from 1, 2, ..., N according to a uniform distribution, or it may be defined in any other arbitrary way.

[0127] In S72, a pseudo-base spectrum and weights are defined.

[0128] There are several methods for defining a pseudo-ground spectrum. For example, a standard sample consisting of a single fluorescent molecule may be prepared, and the excitation and emission spectra of the standard sample may be measured. The measured excitation and emission spectra may then be normalized to a maximum value of 1 and defined as their respective pseudo-ground spectra. Alternatively, if the pair of excitation and emission spectra is known, such as being stored in a database, these known excitation and emission spectra may be used to define their respective pseudo-ground spectra. If the spectra stored in the database deviate from the measurement bandwidth of the fluorescence fingerprint, the intensity of the missing wavelength band may be interpolated using polynomial fitting or Gaussian process regression based on the intensity of the measured wavelength band. Alternatively, the ground spectrum obtained by PARAFAC may be defined as the pseudo-ground spectrum.

[0129] Weights are defined randomly within a range where the minimum and maximum values ​​are greater than 0. These minimum and maximum values ​​serve as indicators of the acceptable spectrum intensity for the base spectrum. For example, a machine learning model trained on a training dataset with weights defined within the range of a minimum of 0.1 and a maximum of 1.0 will differ from a machine learning model trained on a training dataset with weights defined within the range of a minimum of 0.3 and a maximum of 1.0 in spectral decomposition, in which signals with relatively low emission intensity will be considered as the base spectrum. In other words, by changing the minimum and maximum values ​​of the weights, diversity can be introduced into the decomposition results.

[0130] In this embodiment, flexible spectral decomposition is possible by selecting multiple such machine learning models via setting values. For example, even the ground spectrum with small weights, which is difficult to decompose with PARAFAC, can be decomposed in this embodiment.

[0131] In S73, the spectrum obtained by applying a variation processing to the pseudo-base spectrum defined in S72 is redefined as the pseudo-base spectrum.

[0132] As a variation processing method, for example, the wavelength of the pseudo-base spectrum may be shifted in an arbitrary direction. In the shifted pseudo-base spectrum, missing values ​​will occur at the end in the direction opposite to the shift direction. These missing values ​​can be extrapolated using an arbitrary model, such as Gaussian process regression. Alternatively, as a variation processing method, for example, a monotonically increasing function may be applied to each element of the pseudo-base spectrum. Alternatively, for example, a variation processing method may be applied to each emission intensity of the pseudo-base spectrum by applying a power function of an appropriate constant and normalizing the maximum value to 1.

[0133] By randomly applying various variation processes to the pseudo-base spectrum defined in S72, a wide variety of pseudo-base spectrum patterns can be defined. Note that the process in S73 is optional. If the process in S73 is executed, a wide variety of pseudo-base spectra can be obtained.

[0134] S74 is a conditional branch to prevent the creation of pseudo-base spectra with similar peak positions (wavelength bands). When defining the m-th base spectrum, if the peak of the m-th base spectrum overlaps with one of the (m-1) already defined pseudo-base spectrum peaks, the process returns to S72 and redefines the m-th pseudo-base spectrum. For example, if the peak of the m-th base spectrum perfectly matches one of the already defined pseudo-base spectrum peaks, or if the distance between peaks (closeness of wavelength bands) is within a predetermined range, the process returns to S72.

[0135] If the peaks of the baseline spectra perfectly match, spectral decomposition is impossible unless constraints are placed on the shape of the spectra. The S74 process is performed to avoid defining such synthesized data for fluorescence fingerprints.

[0136] The S74 process also determines how far down the peak distance the machine learning model will decompose. For example, a machine learning model trained on a training dataset where the peak distance between ground spectra is 5 nm or more will produce different decomposition results for ground spectra with close peak distances compared to a model trained on a training dataset where the peak distance is 10 nm or more. For instance, the decomposition results can be made more diverse by deciding whether to decompose ground spectra with close peak distances into one ground spectrum or into two ground spectra.

[0137] In this embodiment, flexible spectral decomposition is possible by allowing the user to select from multiple machine learning models via settings, taking into account the performance of the instrument used to measure the spectral spectrum. For example, even ground spectra with close inter-peak distances, which are difficult to decompose using PARAFAC, can be decomposed in this embodiment.

[0138] S75 is a conditional branch that determines whether or not M sets of pseudo-base spectrum and weights have been defined by repeating the processes of S72, S73, and S74 M times. If M sets of pseudo-base spectrum and weights have been defined, the process proceeds to S76; otherwise, the process returns to S72.

[0139] In S76, the error term used in the synthesized data is defined. The error term may be represented, for example, by a matrix composed of Gaussian noise of the same form as the fluorescence fingerprint. Alternatively, the error term may be a signal that cannot be expressed by the direct product of the ground spectrum with respect to the excitation wavelength and the ground spectrum with respect to the emission wavelength, such that the peak position of the emission wavelength changes with the change in excitation wavelength, and the peak position of the excitation wavelength changes with the change in emission wavelength. Alternatively, the error term may be a matrix of the same form as the fluorescence fingerprint, which has a random constant value when the angle between the excitation wavelength and the emission wavelength falls within a certain range (the ratio of the excitation wavelength to the emission wavelength is in a specific wavelength band), and a value of zero when it does not fall within this range. Alternatively, the error term may be a matrix obtained by applying a Gaussian blur to this matrix, or a matrix obtained by summing multiple such matrices together.

[0140] The S76 process adds a signal (error term) that cannot be represented by the base spectrum of the fluorescence fingerprint, which encourages the machine learning model to learn the difference between this signal and the base spectrum.

[0141] In S77, using the M-group pseudo-base spectrum and weight set and the error term, a composite data of one fluorescence fingerprint is obtained from equation (8), and the decomposition result (pseudo-base spectrum and weight set), which is explicit training information, is also obtained. In S75, it is determined that the M-group pseudo-base spectrum and weight set has been defined, so the decomposition result, which is training information, has been obtained.

[0142] The decomposition result candidate prediction unit 102 performs the process shown in Figure 7 multiple times to randomly create multiple composite data sets, thereby constructing a training dataset as shown in equations (3) and (4). The number of composite data sets included in the training dataset is S in equation (3).

[0143] As mentioned above, conventional techniques use the decomposition results obtained by PARAFAC as training data for machine learning models. Therefore, conventional techniques cannot handle cases where PARAFAC has difficulty performing appropriate spectral decomposition, or where the peaks of the ground spectra are close together, or where the weight of the ground spectrum is relatively small compared to the other weights.

[0144] Furthermore, collecting a sufficient quantity of samples for machine learning using PARAFAC with an appropriate number of resolutions presents challenges, including the high level of expertise and arbitrariness required in determining the number of resolutions, as well as the high cost of collecting spectral data.

[0145] In this embodiment, by creating a training dataset using the process shown in Figure 7, it is possible to comprehensively train a machine learning model at low cost, including cases where decomposition is difficult with PARAFAC. Furthermore, in this embodiment, spectral decomposition is possible even in cases where decomposition is difficult with PARAFAC, and the decomposition results can be made flexible according to the user's wishes.

[0146] Let's return to the explanation of the decomposition result candidate selection unit 103 and output unit 104 shown in Figure 2. Next, we will explain the decomposition result candidate selection unit 103.

[0147] The decomposition result candidate selection unit 103 receives multiple decomposition result candidates from the decomposition result candidate prediction unit 102, and selects one or more decomposition result candidates from the multiple candidates based on indicators such as accuracy, and uses the selected decomposition result candidates as the final decomposition result. In this way, the decomposition result candidate selection unit 103 predicts the decomposition result of the spectral spectrum.

[0148] When the spectral decomposition result candidate prediction unit 102 receives the spectral decomposition device 10 from the spectral decomposition device, it obtains N sets of output D(F). D(F) includes N sets of decomposition result candidates and their accuracy. The decomposition result candidate selection unit 103 can select from the N sets of decomposition result candidates a decomposition result candidate with high accuracy as the spectral decomposition result of the spectral spectrum F.

[0149] In this embodiment, since the accuracy training information is provided as either 1 or 0 (zero), the accuracy output of the machine learning model is distributed to values ​​close to 1 or 0. For this reason, the decomposition result candidate selection unit 103 may, for example, use 0.5, which is an intermediate value between 1 and 0, as the accuracy threshold, and select decomposition result candidates with an accuracy of 0.5 or higher as the spectral decomposition result of the spectral spectrum F. The accuracy threshold may also be set by the user from the input unit 101.

[0150] Alternatively, the decomposition result candidate selection unit 103 may select the base spectrum of a decomposition result candidate that is similar to that of a known substance as the spectral decomposition result. By selecting spectral decomposition results in this way, the decomposition result candidate selection unit 103 can perform spectral decomposition limited to only substances with known base spectra.

[0151] For example, if data for substances with known base spectra is pre-stored in the memory device 11, the decomposition result candidate selection unit 103 may calculate the similarity between the base spectra of the decomposition result candidates and the base spectra of known substances stored in the memory device 11, and then select the base spectra of the decomposition result candidates that have a high corresponding accuracy (i.e., the corresponding accuracy is greater than the threshold) and a high similarity (i.e., this similarity is greater than the threshold) as the spectral decomposition result. The similarity may be cosine similarity or the reciprocal of the inter-peak distance. The threshold for determining high accuracy and the threshold for determining high similarity can be arbitrarily determined in advance.

[0152] Accuracy and the similarity mentioned above do not necessarily have to be used as indicators for selecting spectral decomposition results. For example, the decomposition result candidate selection unit 103 associates the corresponding substance name and ground spectrum with the decomposition result candidate for which the accuracy and similarity are high. The information obtained in this way is called "similar substance information". If there is no ground spectrum for a decomposition result candidate with high similarity, information that there is no similar substance may be included in the similar substance information. The similar substance information obtained in this way can be used as an indicator for selecting spectral decomposition results. The decomposition result candidate selection unit 103 outputs the similar substance information to the output unit 104.

[0153] As described above, the decomposition result candidate selection unit 103 can select a decomposition result candidate using at least one of the following: accuracy, similarity to the ground spectrum of a known substance, and similar substance information, and obtain a final spectral decomposition result. For example, the decomposition result candidate selection unit 103 can use the ground spectrum of a decomposition result candidate with high similarity, linked to the similar substance information, as the final spectral decomposition result.

[0154] The decomposition result candidate selection unit 103 outputs the selected decomposition result candidate to the output unit 104 as the final spectral decomposition result. The decomposition result candidate selection unit 103 outputs the spectral decomposition result, the accuracy and similar substance information of the spectral decomposition result, and the decomposition result candidate to the output unit 104.

[0155] The output unit 104 receives the information output by the decomposition result candidate selection unit 103 and outputs it to the output device 22 via the interface 13. For example, the output unit 104 may output either or both the spectral decomposition result and the decomposition result candidates, and along with this output, it may also output either or both the accuracy of the spectral decomposition result and the similar substance information. The output device 22 displays the information output by the decomposition result candidate selection unit 103 on a monitor display. For example, the output device 22 displays a window on the monitor display and displays the information output by the decomposition result candidate selection unit 103 in this window.

[0156] Figure 8 shows an example of a window that the output device 22 displays on the monitor display.

[0157] Window 80 contains a spectral display unit 81, a setting value operation interface 82, an execution button 83, and a spectral decomposition result display unit 84.

[0158] The spectral display unit 81 displays the spectral spectrum measured by the spectral measuring device 21. For example, the spectral display unit 81 illustrates the fluorescence fingerprint using contour lines, similar to the spectral spectrum 61 (Figure 6).

[0159] The setting value manipulation interface 82 is an interface for the user to manipulate default settings or user-defined settings. For example, the setting value manipulation interface 82 is an interface that allows the user to change settings using an input device such as a mouse or keyboard, starting from a state where default settings are preset. In the example shown in Figure 8, the setting value manipulation interface 82 displays "Display Options," "Spectral Decomposition Model," "Selection Criteria," and "Reset."

[0160] The "Display Options" are elements for setting whether or not to display the accuracy, similar substance information, and decomposition result candidates not selected in the decomposition result candidate selection unit 103 (unselected decomposition result candidates) on the spectral decomposition result display unit 84. Whether or not to display each item can be set, for example, using checkboxes. For unselected decomposition result candidates, the accuracy threshold for unselected decomposition result candidates can be set using an input form.

[0161] The "spectral decomposition model" is an element for selecting one or more machine learning models to be used in the decomposition result candidate prediction unit 102. In the example shown in Figure 8, of the two machine learning models (Model A and Model B) with different peak distances, Model A is selected using a checkbox. If multiple machine learning models are selected, the decomposition results from each machine learning model are displayed in parallel on the spectral decomposition result display unit 84. This display allows the user to compare the respective spectral decomposition results.

[0162] The "selection criteria" is an element used by the decomposition result candidate selection unit 103 to set criteria for selecting a spectral decomposition result from multiple decomposition result candidates. In the example shown in Figure 8, the selection criteria can be specified using checkboxes and sliders. In Figure 8, as an example, the selection criteria "select decomposition result candidates with an accuracy of 0.5 or higher, and do not include similar substance information in the selection criteria" is set.

[0163] The "Reset" button is used to return the above settings to their default values.

[0164] The execution button 83 is a button that allows the spectral decomposition device 10 to perform spectral decomposition in response to user operation. When the execution button 83 is pressed, the spectral decomposition device 10 inputs the spectral spectrum shown in the spectral spectrum display unit 81 and the settings shown in the setting value operation interface 82 to the input unit 101 and performs spectral decomposition.

[0165] The spectral decomposition result display unit 84 displays the spectral decomposition result, which is the output of the output unit 104 obtained when the execution button 83 is operated. If the execution button 83 is not operated, nothing is displayed.

[0166] The spectral decomposition result display unit 84 changes the display and hiding of accuracy, similar substance information, and unselected decomposition result candidates according to the settings in the "Display Options" of the setting value operation interface 82. In the example shown in Figure 8, the similar substance information displays "Base 1: Similar substance name A", "Base 2: Similar substance name B", and "Base 3: No similar substances found". Each base spectrum is shown with peak wavelengths using contour lines and graphs. The weights and accuracy are shown using bar graphs and numerical values. These three similar substance information correspond to the decomposition result candidate 60 selected from the four decomposition result candidates 60 shown in Figure 6 because its accuracy 64 is close to 1.

[0167] The output device 22 does not have to display the information output by the decomposition result candidate selection unit 103 in the format of window 80 shown in Figure 8. The output device 22 may display the spectral decomposition results in any format as long as the setting value can be changed and the spectral decomposition results can be visualized and displayed.

[0168] It should be noted that the present invention is not limited to the embodiments described above, and various modifications are possible. For example, the embodiments described above are explained in detail to make the present invention easier to understand, and the present invention is not necessarily limited to embodiments having all of the described configurations. Furthermore, it is possible to replace parts of the configuration of one embodiment with the configuration of another embodiment. It is also possible to add configurations from other embodiments to the configuration of one embodiment. In addition, it is possible to delete parts of the configuration of each embodiment, or to add or replace other configurations.

[0169] 10...Spectrum decomposition device, 11...Storage device, 12...Calculation unit, 13...Interface, 14...Bus, 20...Control device, 21...Spectrum measurement device, 22...Output device, 30...Fluorescence fingerprint data, 31...Visualized fluorescence fingerprint, 32...Fluorescence fingerprint data, 60...Candidate decomposition result, 61...Spectral spectrum, 62...Base spectrum, 62a...Base spectrum related to excitation wavelength, 62b...Base spectrum related to emission wavelength, 62c...Direct product, 63...Weight, 64...Accuracy, 80...Window, 81...Spectral spectrum display unit, 82...Setting value operation interface, 83...Execute button, 84...Spectrum decomposition result display unit, 101...Input unit, 102...Candidate decomposition result prediction unit, 103...Candidate decomposition result selection unit, 104...Output unit.

Claims

1. A spectral decomposition device comprising: an input unit that inputs a spectral spectrum obtained by measuring a sample, wherein the spectrum showing the components of the sample is used as the base spectrum; a decomposition result candidate prediction unit that determines a decomposition result candidate which is a candidate result of decomposing one spectral spectrum into the base spectrum, and an accuracy indicating the prediction reliability of the decomposition result candidate; a decomposition result candidate selection unit that selects a decomposition result which is a result of decomposing one spectral spectrum into the base spectrum from the decomposition result candidates based on the accuracy; and an output unit that outputs one or both of the decomposition result candidates and the decomposition result, wherein the decomposition result candidate prediction unit inputs one spectral spectrum and determines one or more pairs of the decomposition result candidate, which includes the base spectrum and a weight indicating the composition ratio of the base spectrum, and the accuracy for the decomposition result candidate, using a machine learning model.

2. The spectral decomposition apparatus according to claim 1, wherein the training dataset is synthetic data including the base spectrum and weights that are pseudo-defined assuming the actual decomposition results, the machine learning model is trained by supervised learning using the training dataset, and the decomposition result candidate prediction unit trains the machine learning model by making the accuracy of the decomposition result candidates that match the input synthetic data among the decomposition result candidates obtained by the machine learning model by inputting the synthetic data greater than a predetermined threshold, and bringing the base spectrum and weights closer to the synthetic data.

3. A spectral decomposition apparatus according to claim 1, comprising a storage device that pre-stores data of substances whose base spectra are known, wherein the decomposition result candidate selection unit calculates the similarity between the base spectrum of the decomposition result candidate and the base spectrum stored in the storage device, the decomposition result candidate selection unit obtains similar substance information, which is information linking the corresponding substance name and the base spectrum to the decomposition result candidate, for the base spectrum of the decomposition result candidate whose accuracy is greater than a predetermined threshold and whose similarity is greater than a predetermined threshold, and the output unit outputs the similar substance information together with the decomposition result candidate.

4. The spectral decomposition apparatus according to claim 1, wherein the spectral spectrum is recorded with a measurement band and the emission intensity in the measurement band at an arbitrary quantization width.

5. The spectral decomposition apparatus according to claim 3, wherein the decomposition result candidate selection unit selects the decomposition result from the decomposition result candidates using at least one of the accuracy, similarity, and similar substance information.

6. The spectral decomposition apparatus according to claim 2, wherein the candidate decomposition result prediction unit creates the composite data by defining the pseudo-defined base spectrum and the weights, and when defining the base spectrum, the candidate decomposition result prediction unit ensures that the peaks of the base spectrum to be defined do not overlap with the peaks of the pseudo-base spectrum already defined.

7. A spectral decomposition method comprising: an input step in which the spectral decomposition device is operated to measure a sample and constitutes a spectral spectrum obtained therefrom, wherein the spectrum showing the components of the sample is used as the base spectrum; the spectral decomposition device inputs the spectral spectrum; the spectral decomposition device predicts a candidate decomposition result, which is a candidate for the result of decomposing one of the spectral spectra into the base spectrum, and an accuracy indicating the predictive reliability of the candidate decomposition result; the spectral decomposition device selects a candidate decomposition result from the candidate decomposition result based on the accuracy, which is a result of decomposing one of the spectral spectra into the base spectrum; and the spectral decomposition device outputs one or both of the candidate decomposition result and the decomposition result, wherein in the candidate decomposition result prediction step, the spectral decomposition device inputs one of the spectral spectra and uses a machine learning model to obtain one or more pairs of candidate decomposition result, which include the base spectrum and a weight indicating the composition ratio of the base spectrum, and the accuracy for the candidate decomposition result.

8. The spectral decomposition method according to claim 7, wherein the training dataset is synthetic data including the base spectrum and weights that are pseudo-defined assuming the actual decomposition results, the machine learning model is trained by supervised learning using the training dataset, and in the decomposition result candidate prediction step, the spectral decomposition device sets the accuracy of the decomposition result candidates that match the input synthetic data among the decomposition result candidates obtained by the machine learning model by inputting the synthetic data to a value greater than a predetermined threshold, and brings the base spectrum and weights closer to the synthetic data to train the machine learning model.

9. A spectral decomposition method according to claim 7, wherein a storage device has data of substances whose base spectra are known, in which case the spectral decomposition device calculates the similarity between the base spectra of the decomposition result candidates and the base spectra stored in the storage device, in which case the spectral decomposition device obtains similar substance information, which is information linking the corresponding substance name and the base spectrum to the decomposition result candidate, for the base spectra of the decomposition result candidates whose accuracy is greater than a predetermined threshold and whose similarity is greater than a predetermined threshold, and in the output step, the spectral decomposition device outputs the similar substance information together with the decomposition result candidates.

10. The spectral decomposition method according to claim 7, wherein the spectral spectrum is recorded with a measurement band and the emission intensity in the measurement band at an arbitrary quantization width.

11. The spectral decomposition method according to claim 9, wherein in the step of selecting candidate decomposition results, the spectral decomposition apparatus selects the decomposition result from the candidate decomposition results using at least one of the accuracy, the similarity, and the similar substance information.

12. The spectral decomposition method according to claim 8, wherein in the decomposition result candidate prediction step, the spectral decomposition device defines the pseudo-defined base spectrum and the weights to create the composite data, and in the decomposition result candidate prediction step, when the spectral decomposition device defines the base spectrum, it ensures that the peaks of the base spectrum to be defined do not overlap with the peaks of the pseudo-base spectrum already defined.

Images