Data generation method, learning model generation method, computer program, information processing device, and analysis device

Abstract

To provide a data generation method for generating training data including a number of spectra by generating a number of spectra from a small number of spectra, a computer program, an information processor, and an analysis device.SOLUTION: In a data generation method for generating training data for allowing a learning model for outputting an analysis result of a sample to be learned when spectra obtained from the sample are inputted, original spectra as the spectra obtained from the sample are acquired, the original spectra and non-analysis spectra differing from analysis spectra as a basis of analysis of the sample included in the original spectra are used to generate multiple virtual spectra that include same analysis spectra as the analysis spectra included in the original spectra and have a different shape from that of the original spectra, and training data including multiple datasets for relating the virtual spectra to the analysis result of the sample, is generated.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a data generation method for generating training data including spectra for machine learning, a learning model generation method, a computer program, an information processing device, and an analysis device. [Background technology]

[0002] The spectrum of a signal from a sample is acquired and the sample is analyzed based on the acquired spectrum. For example, components contained in the sample are identified based on a fluorescent X-ray spectrum. For example, Raman scattered light from a sample such as a cell is measured and the type of sample is determined based on the Raman spectrum. One method for obtaining sample analysis results based on spectra is to use a learning model trained to output an analysis result when a spectrum is input. To obtain a trained learning model, machine learning must be performed using training data containing spectra and correct analysis results. Machine learning requires training data containing many combinations of spectra and analysis results. However, it is difficult to obtain a large number of spectra through actual measurements. Patent Document 1 describes a method for generating multiple spectra contained in training data from a small number of spectra by superimposing random noise generated using random numbers on the spectrum. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-101524 Summary of the Invention [Problem to be solved by the invention]

[0004] To generate a learning model that outputs reliable analysis results, training data containing as many combinations of spectra and analysis results as possible is required. However, simply superimposing random noise onto the spectrum does not generate an adequate number of spectra. Therefore, a technology that can generate a larger number of spectra from a smaller number of spectra is needed to generate training data.

[0005] The present invention has been made in consideration of the above circumstances, and its purpose is to provide a data generation method, a learning model generation method, a computer program, an information processing device, and an analysis device that generate training data containing a large number of spectra by generating a large number of spectra from a small number of spectra. [Means for solving the problem]

[0006] A data generation method according to one embodiment of the present invention is a data generation method for generating training data for training a learning model that outputs an analysis result of a sample when a spectrum obtained from the sample is input, the data generation method comprising the steps of: acquiring an original spectrum, which is a spectrum obtained from the sample; using the original spectrum and a non-analytical spectrum that is different from the analytical spectrum contained in the original spectrum and that serves as the basis for analysis of the sample, to generate a plurality of virtual spectra that contain analytical spectra that are identical to the analytical spectrum contained in the original spectrum and have a different shape from the original spectrum; and generating training data that includes a plurality of data sets, each of which associates the virtual spectrum with the analysis result of the sample.

[0007] In one aspect of the present invention, a virtual spectrum is generated using an original spectrum and a non-analytical spectrum different from the analytical spectrum that serves as the basis for sample analysis, and training data including the virtual spectrum is generated. The virtual spectrum includes an analytical spectrum that is identical to the analytical spectrum included in the original spectrum, but has a different shape from the original spectrum. A large number of virtual spectra can be generated from a small number of original spectra by processing using multiple different non-analytical spectra, such as by adding the non-analytical spectrum to the original spectrum. Because the non-analytical spectrum has almost no effect on the analysis results, the virtual spectrum can be associated with the analysis results of the sample, and training data including a large number of spectra can be generated.

[0008] A data generation method according to one aspect of the present invention is characterized in that the virtual spectrum is generated by separating a non-analysis spectrum from the original spectrum, and adding a non-analysis spectrum different from the non-analysis spectrum separated from the original spectrum to the spectrum obtained after the non-analysis spectrum has been separated from the original spectrum.

[0009] In one aspect of the present invention, a virtual spectrum is generated by separating a non-analytical spectrum from an original spectrum and adding other non-analytical spectra to the original spectrum. The non-analytical spectra included in the virtual spectrum are different from the non-analytical spectra included in the original spectrum, so the virtual spectrum is different from the original spectrum.

[0010] A data generation method according to one aspect of the present invention is characterized in that the non-analysis spectrum separated from the original spectrum is processed by adding a predetermined value, shifting a wavenumber, or inverting a value centered on a specific value contained in the non-analysis spectrum, and the virtual spectrum is generated by adding the processed non-analysis spectrum to the spectrum obtained after the non-analysis spectrum has been separated from the original spectrum.

[0011] In one embodiment of the present invention, a virtual spectrum is generated by processing a non-analytical spectrum and adding the processed non-analytical spectrum to the spectrum obtained after separating the non-analytical spectrum from the original spectrum. By performing processing such as adding a predetermined value, shifting a wavenumber, or inverting a value, a large number of different processed non-analytical spectra can be generated from a small number of non-analytical spectra. Therefore, a large number of virtual spectra different from the original spectrum can be generated using a large number of processed non-analytical spectra.

[0012] A data generation method according to one aspect of the present invention is characterized in that it acquires a plurality of original spectra, and generates the virtual spectrum by separating non-analysis spectra from one original spectrum and adding non-analysis spectra separated from other original spectra to the spectrum obtained after the separation.

[0013] In one aspect of the present invention, a virtual spectrum is generated by adding a non-analytical spectrum separated from another original spectrum to a spectrum obtained after separating a non-analytical spectrum from one original spectrum. By using a non-analytical spectrum obtained by actual measurement, a virtual spectrum close to a real spectrum can be obtained. Training data close to training data containing real spectra can be obtained, and a learning model that provides accurate output according to the real spectrum can be generated by machine learning.

[0014] A data generation method according to one aspect of the present invention is characterized in that it acquires a plurality of original spectra, processes the non-analysis spectra by adding together a plurality of non-analysis spectra separated from the plurality of original spectra, and generates the virtual spectrum by adding the processed non-analysis spectra to the spectrum obtained after the non-analysis spectra have been separated from the original spectra.

[0015] In one embodiment of the present invention, a non-analytical spectrum is processed by adding multiple non-analytical spectra together. By using multiple non-analytical spectra obtained by actual measurements, a virtual spectrum that is close to the actual spectrum can be obtained. Furthermore, by changing the method for adding multiple non-analytical spectra, a large number of processed non-analytical spectra can be generated, thereby generating a large number of virtual spectra.

[0016] A data generation method according to one embodiment of the present invention is characterized in that it acquires a plurality of original spectra, divides a plurality of non-analytical spectra separated from the plurality of original spectra into a plurality of split spectra falling within a plurality of wavenumber ranges, processes the non-analytical spectra by combining the split spectra with each other among the plurality of non-analytical spectra, and generates the virtual spectrum by adding the processed non-analytical spectra to the spectrum obtained after separating the non-analytical spectra from the original spectrum.

[0017] In one embodiment of the present invention, a non-analytical spectrum is processed by dividing the non-analytical spectrum into a plurality of split spectra included in a plurality of wavenumber ranges, and combining the split spectra among the plurality of non-analytical spectra. By combining the plurality of split spectra in various ways, a large number of processed non-analytical spectra can be generated, and a large number of virtual spectra can be generated.

[0018] A data generating method according to one aspect of the present invention is characterized in that a non-analysis spectrum is separated from the original spectrum by separating a component having a specific frequency band from the original spectrum.

[0019] In one embodiment of the present invention, non-analytical spectra are separated from the original spectra by frequency separation, which makes it possible to easily separate non-analytical spectra that have different frequency bands from the analytical spectra that form the basis of the sample analysis results.

[0020] A data generation method according to one aspect of the present invention is characterized in that the original spectrum is a Raman spectrum, the analytical spectrum is a component caused by Raman scattered light contained in the Raman spectrum, and the non-analytical spectrum is a fluorescence spectrum contained in the Raman spectrum.

[0021] In one aspect of the present invention, the original spectrum is a Raman spectrum, the analytical spectrum is a component in the Raman spectrum that is due to Raman scattered light, and the non-analytical spectrum is a fluorescence spectrum. By using a fluorescence spectrum obtained by actual measurement as the non-analytical spectrum, a virtual spectrum that is close to the actual spectrum can be obtained. Using a learning model, it becomes possible to easily obtain sample analysis results based on the Raman spectrum.

[0022] A data generating method according to one aspect of the present invention is characterized in that the virtual spectrum is generated by adding a noise spectrum to a spectrum obtained by adding the non-analysis spectrum to the original spectrum.

[0023] In one embodiment of the present invention, a virtual spectrum is generated by adding a noise spectrum to a spectrum obtained by adding a non-analytical spectrum to an original spectrum. By adding a noise spectrum, a larger number of virtual spectra can be generated.

[0024] A data generation method according to one aspect of the present invention is characterized in that it generates the virtual spectrum by separating a noise spectrum having a specific frequency band from the original spectrum, and adding a noise spectrum having the specific frequency band that is different from the noise spectrum separated from the original spectrum to a spectrum obtained by separating the noise spectrum from the original spectrum and adding the non-analysis spectrum.

[0025] In one embodiment of the present invention, a noise spectrum having a specific frequency band is separated from an original spectrum, and a virtual spectrum is generated by adding noise spectra having the same frequency band. Because the frequency band of the noise spectrum remains unchanged, a large number of virtual spectra can be generated without affecting the analysis results.

[0026] A learning model generation method according to one aspect of the present invention is characterized in that it acquires training data generated by the data generation method according to the present invention, and generates a learning model that, when a spectrum obtained from a sample is input, outputs an analysis result of the sample by learning using the training data.

[0027] In one aspect of the present invention, a learning model is generated by learning using training data, which outputs a sample analysis result when a spectrum is input. Since the training data includes many combinations of spectra and sample analysis results, machine learning can be used to generate a learning model that outputs a highly accurate analysis result according to the spectrum.

[0028] A computer program according to one embodiment of the present invention is a computer program that causes a computer to execute a process of generating training data for training a learning model that outputs an analysis result of a sample when a spectrum obtained from the sample is input, the computer program acquiring an original spectrum that is a spectrum obtained from the sample, and using the original spectrum and a non-analytical spectrum that is different from the analytical spectrum contained in the original spectrum and that serves as the basis for analysis of the sample, to generate a plurality of virtual spectra that contain analytical spectra that are identical to the analytical spectrum contained in the original spectrum and have a different shape from the original spectrum, and generating training data including a plurality of data sets, each of which associates the virtual spectrum with an analysis result of the sample.

[0029] An information processing device according to one embodiment of the present invention is an information processing device that performs processing to generate training data for training a learning model that outputs an analysis result of a sample when a spectrum obtained from the sample is input, and is characterized by comprising: a spectrum acquisition unit that acquires an original spectrum, which is a spectrum obtained from the sample; a virtual spectrum generation unit that uses the original spectrum and a non-analytical spectrum that is different from the analytical spectrum contained in the original spectrum and that serves as the basis for analysis of the sample, to generate a plurality of virtual spectra that include analytical spectra that are identical to the analytical spectrum contained in the original spectrum and have shapes different from the original spectrum; and a data generation unit that generates training data including a plurality of data sets, each of which associates the virtual spectrum with an analysis result of the sample.

[0030] In one embodiment of the present invention, an information processing device (computer) generates a virtual spectrum using an original spectrum and a non-analytical spectrum, and generates training data including the virtual spectrum. A large number of virtual spectra can be generated by processing using a plurality of non-analytical spectra, such as by adding a non-analytical spectrum to an original spectrum. The virtual spectrum can be associated with the analysis result of a sample, and training data including a large number of spectra can be generated.

[0031] An analytical apparatus according to one aspect of the present invention is an analytical apparatus for analyzing a sample, comprising: a measurement device that measures a physical phenomenon occurring in the sample and generates a spectrum representing characteristics of the measured physical phenomenon; and an information processing device, wherein the information processing device comprises: a virtual spectrum generation unit that uses an original spectrum generated by the measurement device and a non-analytical spectrum that is different from the analytical spectrum contained in the original spectrum and that serves as the basis for analysis of the sample, to generate a plurality of virtual spectra that include an analytical spectrum that is identical to the analytical spectrum contained in the original spectrum and have a different shape from the original spectrum; and a data generation unit that generates training data including a plurality of data sets, each of which associates the virtual spectrum with an analysis result of the sample.

[0032] In one aspect of the present invention, an analysis apparatus includes a measurement device and an analysis apparatus. The measurement device generates an original spectrum, and an information processing device generates a virtual spectrum from the original spectrum and generates training data. A large number of virtual spectra can be generated from the original spectrum generated by the measurement device. Training data containing a large number of spectra can be generated from the limited number of original spectra that can be generated by the measurement device. [Effects of the Invention]

[0033] The present invention can generate training data containing a large number of spectra, and therefore has excellent effects such as being able to generate a learning model that has been trained to output an analysis result when a spectrum is input through machine learning using the training data. [Brief explanation of the drawings]

[0034] [Figure 1] FIG. 2 is a block diagram showing an example of the configuration of an analysis device and an example of the configuration of a measurement device. [Figure 2] FIG. 2 is a block diagram illustrating an example of a functional configuration of an information processing device. [Figure 3] FIG. 1 is a conceptual diagram illustrating the function of a learning model. [Figure 4] 10 is a flowchart illustrating an example of a processing procedure for generating training data. [Figure 5] FIG. 1 is a conceptual diagram showing an example of a method for separating an analytical spectrum, a non-analytical spectrum, and a noise spectrum from an original spectrum. [Figure 6] FIG. 10 is a conceptual diagram showing an example in which a plurality of non-analysis spectra are generated to be added to an analysis spectrum. [Figure 7] FIG. 10 is a conceptual diagram showing an example in which a plurality of virtual spectra are generated from an original spectrum. [Figure 8] 10 is a flowchart showing the steps of a process for generating a learning model. [Figure 9] 10 is a flowchart showing the procedure of a process for analyzing a sample 3 using a learning model. DETAILED DESCRIPTION OF THE INVENTION

[0035] The present invention will now be described in detail with reference to the drawings showing embodiments thereof. In this embodiment, training data is generated to generate a learning model by machine learning that outputs analysis results corresponding to spectra obtained from a sample. FIG. 1 shows an example of the configuration of an analytical device 100 and is a block diagram showing an example of the configuration of a measurement device 2. The analytical device 100 is a device that analyzes samples and performs processing to generate training data. The analytical device 100 includes an information processing device 1 and a measurement device 2. The measurement device 2 applies some kind of stimulus, such as light irradiation, to a sample, measures a physical phenomenon occurring in the sample in response to the stimulus, and generates a spectrum representing the characteristics of the measured physical phenomenon. For example, the measurement device 2 measures fluorescent X-rays or Raman scattered light emitted from the sample 3 and generates a fluorescent X-ray spectrum or Raman spectrum. FIG. 1 shows an example in which the measurement device 2 is a Raman scattered light measurement device. The information processing device 1 executes a data generation method for generating training data. The information processing device 1 is connected to the measurement device 2.

[0036] In this embodiment, an example is mainly shown in which the measurement device 2 is a Raman scattered light measurement device. The measurement device 2 includes a sample holder 25 that holds the sample 3, an irradiation unit 21 that irradiates laser light, a beam splitter 242, and a lens 243. For example, the sample 3 is a biological sample such as a cell. The sample holder 25 is, for example, a sample stage on which the sample 3 is placed. The sample holder 25 may be in a form other than a sample stage. The irradiation unit 21 includes a laser light source. The laser light irradiated by the irradiation unit 21 is reflected by the beam splitter 242, passes through the lens 243, and is irradiated onto the sample 3.

[0037] The measuring device 2 further includes a mask 241, a spectroscope 23, a detector 22 for detecting light, and a controller 26. The mask 241 has a slit. Raman scattered light is generated in the portion of the sample 3 irradiated with the laser light. The generated Raman scattered light is collected by a lens 243, passes through a beam splitter 242, and passes through the slit in the mask 241 to be narrowed into a thin beam and enters the spectroscope 23. In FIG. 1, the laser light and Raman scattered light are indicated by solid arrows. The spectroscope 23 separates the incident Raman scattered light. The detector 22 detects the light of each wavelength separated by the spectroscope 23. The measuring device 2 includes an optical system consisting of numerous optical components such as mirrors, lenses, and filters for guiding, collecting, and separating the laser light and Raman scattered light. In FIG. 1, the optical system other than the mask 241, the beam splitter 242, and the lens 243 is omitted. The measuring device 2 may be configured such that the laser light from the irradiation unit 21 passes through the beam splitter 242 and the Raman scattered light is reflected by the beam splitter 242. The optical system does not have to include the mask 241, the beam splitter 242, or the lens 243.

[0038] The irradiation unit 21, the detection unit 22, and the spectroscope 23 are connected to a control unit 26. The control unit 26 is connected to the information processing device 1. The control unit 26 includes a calculation unit that executes calculations to control each unit of the measurement device 2, a memory, and a communication unit that transmits and receives data to and from the information processing device 1.

[0039] The control unit 26 controls each part of the measurement device 2. The irradiation unit 21 is controlled to be turned on or off by the control unit 26. The spectrometer 23 is controlled by the control unit 26 to determine the wavelength of light that is separated and detected by the detection unit 22. The detection unit 22 outputs a signal corresponding to the intensity of light of each detected wavelength to the control unit 26. The control unit 26 receives the signal output by the detection unit 22 and generates a Raman spectrum based on the wavenumber of the light separated by the spectrometer 23 and the light intensity indicated by the input signal. The Raman spectrum is data in which the wavenumber of light is associated with the intensity of the detected light. The intensity of light may be associated with the wavelength or energy of the light instead of the wavenumber.

[0040] The control unit 26 can adjust the measurement conditions for measuring Raman scattered light. The measurement conditions include, for example, the intensity of the laser light irradiated onto the sample 3, the irradiation time, the number of times the intensity of the Raman scattered light is integrated, the width of the slit through which the laser light passes, the wavelength of the laser light, or the focal length. For example, the control unit 26 adjusts the intensity of the laser light by changing the output of the laser light source included in the irradiation unit 21. For example, the control unit 26 changes the output of the laser light source by adjusting the energy injected into the laser light source. The intensity of the laser light may also be adjusted using a filter that attenuates the laser light. For example, the control unit 26 measures the intensity of the Raman scattered light of each wavelength detected by the detection unit 22 while changing the wavelength of the light to be dispersed by the spectrometer 23, repeats detection for each wavelength, and adjusts the number of times the measured intensity of the Raman scattered light is integrated.

[0041] For example, the mask 241 may have multiple slits with different widths, and the control unit 26 may adjust the width of the slits by changing the slits through which the laser light passes. For example, the irradiation unit 21 may include multiple laser light sources with different wavelengths, and the control unit 26 may adjust the wavelength of the laser light by changing the laser light source to be used. For example, the measurement device 2 may include multiple lenses 243 with different focal lengths, and the control unit 26 may adjust the focal length by changing the lens 243 to be used. The measurement device 2 may be configured to change the positions of the components constituting the optical system or the position of the sample holder 25 when adjusting the focal length.

[0042] FIG. 2 is a block diagram showing an example of the functional configuration of the information processing device 1. The information processing device 1 is configured using a computer such as a personal computer or a server device. The information processing device 1 includes a calculation unit 11, a memory 12, a drive unit 13, a storage unit 14, an operation unit 15, a display unit 16, and a communication unit 17. The calculation unit 11 is configured using, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a multi-core CPU. The calculation unit 11 may also be configured using a quantum computer. The memory 12 stores temporary data generated in conjunction with calculations. The memory 12 is, for example, a RAM (Random Access Memory). The drive unit 13 reads information from a recording medium 10 such as an optical disc or a portable memory. The storage unit 14 is non-volatile, for example, a hard disk or a non-volatile semiconductor memory.

[0043] The calculation unit 11 causes the drive unit 13 to read the computer program 141 recorded on the recording medium 10, and stores the read computer program 141 in the storage unit 14. The calculation unit 11 executes processing required for the information processing device 1 in accordance with the computer program 141. The computer program 141 may be a computer program product. The computer program 141 may be downloaded from outside the information processing device 1. Alternatively, the computer program 141 may be pre-stored in the information processing device 1. In these cases, the information processing device 1 does not need to be equipped with the drive unit 13.

[0044] The operation unit 15 receives input of information such as text by receiving operations from the user. The operation unit 15 is, for example, a touch panel, a pen tablet, a keyboard, or a pointing device. The display unit 16 displays images. The display unit 16 is, for example, a liquid crystal display or an EL display (Electroluminescent Display). The operation unit 15 and the display unit 16 may be integrated. The communication unit 17 is connected to the measurement device 2. The communication unit 17 may be connected to the measurement device 2 via a communication network such as the Internet or a LAN. The communication unit 17 sends and receives data to and from the measurement device 2.

[0045] The information processing device 1 may be configured with multiple computers, and data may be stored in a distributed manner across the multiple computers, and processing may be executed in a distributed manner across the multiple computers. The information processing device 1 may be realized using cloud computing, or may be realized by multiple virtual machines provided within a single computer.

[0046] The information processing device 1 includes a learning model 142. The learning model 142 is realized by the calculation unit 11 executing information processing in accordance with a computer program 141. For example, the memory unit 14 stores data recording parameters of the learning model 142, and the calculation unit 11 uses the parameters to execute information processing in accordance with the computer program 141, thereby realizing the learning model 142. The learning model 142 may be configured with hardware. The learning model 142 may be realized using a quantum computer. Alternatively, the learning model 142 may be provided outside the information processing device 1, and the information processing device 1 may execute processing using the external learning model 142.

[0047] FIG. 3 is a conceptual diagram showing the function of the learning model 142. The learning model 142 is trained so as to output an analysis result of the sample 3 when a spectrum obtained from the sample 3 is input. In the example of this embodiment, the spectrum input to the learning model 142 is a Raman spectrum. The analysis result of the sample 3 is, for example, the type of cells contained in the sample 3, which is a biological sample. For example, the analysis result of the sample 3 is the type of elements or compounds contained in the sample 3. For example, the learning model 142 is configured using a neural network or a support vector machine.

[0048] The measuring device 2 detects Raman scattered light generated from the sample 3 and generates a Raman spectrum. The control unit 26 inputs the generated Raman spectrum to the information processing device 1. The information processing device 1 accepts the Raman spectrum input via the communication unit 17, and the calculation unit 11 stores the accepted Raman spectrum in the memory unit 14. The measuring device 2 changes the measurement conditions and generates a Raman spectrum under each measurement condition. For example, the measuring device 2 changes the measurement conditions, such as the laser light intensity, irradiation time, number of integrations, slit width, laser light wavelength, or focal length. The measuring device 2 inputs the Raman spectrum generated under each measurement condition to the information processing device 1, and the information processing device 1 stores each Raman spectrum in the memory unit 14. Furthermore, when the sample 3 is changed, the measuring device 2 generates multiple Raman spectra for multiple samples 3, and the information processing device 1 stores the multiple Raman spectra in the memory unit 14.

[0049] The information processing device 1 analyzes the sample 3, such as determining the type of cells contained in the sample 3, based on the Raman spectrum. The information processing device 1 stores the analysis results of the sample 3 in the storage unit 14 in association with the Raman spectrum. Alternatively, the information processing device 1 receives the analysis results of the sample 3 based on the Raman spectrum when the user operates the operation unit 15. The information processing device 1 stores the received analysis results of the sample 3 in the storage unit 14 in association with the Raman spectrum.

[0050] The information processing device 1 performs processing to generate training data necessary for learning the learning model 142. FIG. 4 is a flowchart showing an example of the processing procedure for generating training data. Hereinafter, step is abbreviated as S. The calculation unit 11 of the information processing device 1 executes processing in accordance with the computer program 141.

[0051] The information processing device 1 acquires an original spectrum, which is a spectrum generated by the measurement device 2 (S11). As described above, the measurement device 2 measures Raman scattered light and inputs the Raman spectrum to the information processing device 1. In S11, the information processing device 1 receives the Raman spectrum via the communication unit 17, and the calculation unit 11 stores the received Raman spectrum in the memory unit 14, thereby acquiring the original spectrum. In this embodiment, the original spectrum is the Raman spectrum generated by the measurement device 2. Furthermore, the analysis results of the sample 3 are stored in the memory unit 14 in association with the original spectrum.

[0052] A plurality of measurement devices 2 may be connected to the information processing device 1. The information processing device 1 may acquire original spectra from a plurality of measurement devices 2. The information processing device 1 may acquire original spectra other than the original spectrum generated by the connected measurement device 2. For example, an original spectrum may be generated by a measurement device other than the connected measurement device 2 and input to the information processing device 1. In S11, the information processing device 1 acquires a plurality of original spectra and stores them in the memory unit 14. The processing of S11 corresponds to the spectrum acquisition unit.

[0053] The information processing device 1 then normalizes the multiple original spectra (S12). In S12, the calculation unit 11 performs normalization so that the light intensities at specific wavenumbers are the same. For example, the calculation unit 11 performs normalization by dividing the light intensities at the specific wavenumbers by the light intensities at other wavenumbers.

[0054] The information processing device 1 separates the analytical spectrum, non-analytical spectrum, and noise spectrum from the normalized raw spectrum (S13). The light measured by the measurement device 2 also contains light other than Raman scattered light, and the raw spectrum contains components caused by light other than Raman scattered light. The analytical spectrum is a spectrum that forms the basis of the analysis results of the sample 3, and is the component contained in the raw spectrum that is caused by Raman scattered light. The non-analytical spectrum is a spectrum different from the analytical spectrum, and is the component contained in the raw spectrum that is caused by light other than Raman scattered light. For example, the light measured by the measurement device 2 contains fluorescence, the Raman spectrum contains a fluorescence spectrum, and the non-analytical spectrum is a fluorescence spectrum. The non-analytical spectrum has almost no effect on the analysis results of the sample 3.

[0055] FIG. 5 is a conceptual diagram showing an example of a method for separating an analytical spectrum, a non-analytical spectrum, and a noise spectrum from an original spectrum. In S13, the information processing device 1 performs frequency separation to separate components having a specific frequency band from the original spectrum. The horizontal axis of the spectrum shown in FIG. 5 represents wavenumber, and the vertical axis represents signal intensity. Frequency separation makes it possible to easily separate a non-analytical spectrum, which has a different frequency band from the analytical spectrum.

[0056] The calculation unit 11 separates a non-analysis spectrum consisting of frequency components equal to or less than a first threshold from the original spectrum by performing a first frequency separation that separates frequency components equal to or less than a first threshold from the original spectrum. The first frequency separation may separate frequency components below the first threshold. The first threshold is predetermined to a frequency value lower than the frequency band of the analysis spectrum. The first threshold is pre-stored in the storage unit 14 or is included in the computer program 141. The fluorescence spectrum is a component having a lower frequency band than the component caused by Raman scattered light, and can be separated from the original spectrum by the first frequency separation.

[0057] Furthermore, the calculation unit 11 separates the analysis spectrum and the noise spectrum by performing a second frequency separation to separate frequency components equal to or less than a second threshold from the spectrum obtained after separation of the non-analysis spectrum. The second threshold is predetermined to a frequency value that exceeds the first threshold and is higher than the frequency band of the analysis spectrum. The second threshold is pre-stored in the storage unit 14 or is included in the computer program 141. The separated frequency components equal to or less than the second threshold are the analysis spectrum. That is, the analysis spectrum consists of frequency components ranging from the first threshold to the second threshold. The frequency components exceeding the second threshold are the noise spectrum. In the second frequency separation, frequency components less than the second threshold may also be separated.

[0058] For example, the calculation unit 11 performs an FFT (fast Fourier transform) and an inverse FFT to perform first frequency separation and second frequency separation. The calculation unit 11 stores the analysis spectrum, non-analysis spectrum, and noise spectrum separated from the original spectrum in the storage unit 14. The information processing device 1 may include a first LPF (low-pass filter) having a cutoff frequency as a first threshold value and a second LPF having a cutoff frequency as a second threshold value. In this embodiment, the information processing device 1 performs the first frequency separation using the first LPF and the second frequency separation using the second LPF. In the process of generating training data, the spectrum included in the training data is generated by adding the non-analysis spectrum to the analysis spectrum.

[0059] The information processing device 1 then generates a plurality of non-analysis spectra to be added to the analysis spectra (S14). In S14, for example, the calculation unit 11 uses a non-analysis spectrum separated from one original spectrum as a non-analysis spectrum to be added to the analysis spectrum separated from another original spectrum.

[0060] Alternatively, the calculation unit 11 processes a non-analytical spectrum and uses the processed non-analytical spectrum as a non-analytical spectrum to be added to an analytical spectrum. For example, the calculation unit 11 processes a non-analytical spectrum by adding multiple non-analytical spectra separated from multiple original spectra. In this case, the calculation unit 11 multiplies one or both of the non-analytical spectra to be added by a specific positive or negative value before adding them. In this way, adding also includes multiplying by a positive or negative value before adding. The calculation unit 11 may process a non-analytical spectrum by adding three or more non-analytical spectra. The calculation unit 11 may also add a non-analytical spectrum obtained by another method to a non-analytical spectrum separated from an original spectrum. For example, a fluorescence spectrum generated by a measurement device that measures fluorescence may be added. For example, a theoretically generated fluorescence spectrum may be added. By changing the method of adding multiple non-analytical spectra, a large number of processed non-analytical spectra can be generated.

[0061] For example, the calculation unit 11 processes the non-analysis spectrum by adding a predetermined positive or negative value to the non-analysis spectrum. For example, the calculation unit 11 processes the non-analysis spectrum by shifting the wavenumber. In the wavenumber shift, the wavenumber associated with the light intensity included in the non-analysis spectrum is changed to another wavenumber shifted by a predetermined amount from the original wavenumber. For example, the calculation unit 11 processes the non-analysis spectrum by inverting the spectrum along the horizontal or vertical axis around a specific value included in the non-analysis spectrum. In the horizontal axis inversion, the light intensity values ​​associated with two wavenumbers whose absolute difference from the specific wavenumber is equal are swapped. In the vertical axis inversion, a value twice the value obtained by subtracting the specific intensity value from the light intensity associated with each wavenumber is added to the light intensity associated with each wavenumber. A large number of mutually distinct spectra can be generated from a small number of non-analysis spectra.

[0062] For example, the calculation unit 11 processes the non-analysis spectrum by dividing the non-analysis spectrum into multiple split spectra along the horizontal axis and combining the split spectra among the multiple non-analysis spectrums. That is, the calculation unit 11 divides the wavenumber of the entire non-analysis spectrum into multiple wavenumber ranges and divides the non-analysis spectrum into split spectra included in each wavenumber range. For example, the calculation unit 11 exchanges some split spectra included in the non-analysis spectrum with split spectra included in other non-analysis spectra. For example, the calculation unit 11 randomly combines multiple split spectra extracted from multiple non-analysis spectra. By combining multiple split spectra in various ways, a large number of processed non-analysis spectra can be generated. Even when the horizontal axis of the original spectrum is represented by a physical quantity other than wavenumber, such as wavelength or energy, the calculation unit 11 processes the non-analysis spectrum in a similar manner.

[0063] In S14, the calculation unit 11 may select a non-analysis spectrum obtained by another method as the non-analysis spectrum to be added to the analytical spectrum. For example, the calculation unit 11 may select a fluorescence spectrum generated by a measurement device that measures fluorescence as the non-analysis spectrum to be added to the analytical spectrum. For example, the calculation unit 11 may select a theoretically generated fluorescence spectrum as the non-analysis spectrum to be added to the analytical spectrum.

[0064] 6 is a conceptual diagram showing an example of generating multiple non-analysis spectra to be added to an analysis spectrum. For a non-analysis spectrum 41 separated from an original spectrum, multiple non-analysis spectra 42 to be added to the analysis spectrum are generated. A large number of different non-analysis spectra 41 are generated, including a non-analysis spectrum 42 separated from another original spectrum, a non-analysis spectrum 42 obtained by adding a predetermined value to a non-analysis spectrum 41, a non-analysis spectrum 42 obtained by adding multiple non-analysis spectra 41, a non-analysis spectrum 42 obtained by shifting a non-analysis spectrum 41, and a non-analysis spectrum 42 obtained by inverting a non-analysis spectrum 41.

[0065] The information processing device 1 then generates a plurality of noise spectra to be added to the analysis spectrum (S15). In S15, for example, the calculation unit 11 uses a noise spectrum separated from one original spectrum as a noise spectrum to be added to the analysis spectrum separated from another original spectrum. As a result, the noise spectrum to be added to the analysis spectrum has the same frequency band as the noise spectrum separated from the original spectrum. Alternatively, the calculation unit 11 generates a noise spectrum having the same frequency band as the noise spectrum separated from the original spectrum, and uses the generated noise spectrum as the noise spectrum to be added to the analysis spectrum.

[0066] The information processing device 1 then generates multiple virtual spectra by adding the generated non-analysis spectrum and noise spectrum to the analysis spectrum (S16). The virtual spectrum contains an analysis spectrum identical to the analysis spectrum contained in the original spectrum, but has a different shape from the original spectrum. In S16, the calculation unit 11 generates a virtual spectrum by adding one of the generated non-analysis spectrums and one of the generated noise spectrums to the analysis spectrum. The calculation unit 11 generates multiple virtual spectra by repeating the addition while changing the combination of the analysis spectrum, non-analysis spectrum, and noise spectrum. Multiple non-analysis spectrums or multiple noise spectra may be added to the analysis spectrum. The calculation unit 11 may generate a virtual spectrum without separating or adding noise spectra.

[0067] 7 is a conceptual diagram showing an example in which a plurality of virtual spectra are generated from an original spectrum. Through the processes of S12 to S16, a large number of virtual spectra 44 are generated from an initial small number of original spectra 43. A large number of non-analysis spectra or noise spectra are added to the original spectra 43, thereby generating a large number of virtual spectra 44.

[0068] In the processes of S12 to S16, a non-analysis spectrum is separated from the original spectrum, and a different non-analysis spectrum is added to generate a virtual spectrum. Separating a non-analysis spectrum from the original spectrum means adding a non-analysis spectrum multiplied by a negative value to the original spectrum. That is, in the processes of S12 to S16, a virtual spectrum is generated by adding a non-analysis spectrum to the original spectrum.

[0069] Furthermore, by adding noise spectra, a larger number of virtual spectra can be generated. By separating a noise spectrum from an original spectrum and adding another noise spectrum, the noise contained in the spectrum is changed, and a larger number of virtual spectra can be generated. Alternatively, by adding a noise spectrum having the same frequency band as the noise spectrum separated from the original spectrum, a larger number of virtual spectra can be generated without changing the frequency band of the noise spectrum. Because the frequency band of the noise spectrum to be added is the same as the frequency band of the noise spectrum contained in the original spectrum, the noise spectrum to be added has almost no effect on the analysis results.

[0070] A virtual spectrum is a virtual Raman spectrum generated from a Raman spectrum, which is an original spectrum. The non-analytical or noise spectrum contained in the virtual spectrum is different from the non-analytical or noise spectrum contained in the original spectrum, so the virtual spectrum and the original spectrum have different shapes, and the virtual spectrum is a different spectrum from the original spectrum. Multiple virtual spectra contain slightly different non-analytical or noise spectra. Even if multiple virtual spectra contain the same analytical spectrum, they contain different non-analytical or noise spectra, making them different spectra. A virtual spectrum contains an analytical spectrum that is the same as the analytical spectrum contained in the original spectrum, and the non-analytical and noise spectra have almost no effect on the analysis results. Therefore, the analysis results corresponding to the virtual spectrum are the same as the analysis results corresponding to the original spectrum. If the analytical spectra contained in multiple virtual spectra are the same, the analysis results corresponding to the multiple virtual spectra are the same. Therefore, the virtual spectrum can be associated with the analysis results of sample 3.

[0071] The processes of S12 to S16 correspond to the virtual spectrum generation unit. In S12 to S16, the information processing device 1 may generate a virtual spectrum by a method other than separating a non-analysis spectrum from an original spectrum and adding other non-analysis spectra or noise spectra. For example, the calculation unit 11 may generate a virtual spectrum by adding other non-analysis spectra or noise spectra to the original spectrum without separating the non-analysis spectra from the original spectrum. For example, the calculation unit 11 may generate a virtual spectrum by connecting non-analysis spectra along the horizontal axis of the analysis spectrum or original spectrum. For example, the calculation unit 11 may generate a virtual spectrum by arranging the original spectrum or analysis spectrum and non-analysis spectra in a two-dimensional array.

[0072] The information processing device 1 then generates training data including a plurality of data sets, each of which associates a virtual spectrum with an analysis result of the sample 3 (S17). In S17, the calculation unit 11 generates a data set that associates a virtual spectrum generated from the original spectrum with an analysis result of the sample 3 associated with the original spectrum. The calculation unit 11 generates a plurality of data sets, generates training data including the plurality of data sets, and stores the training data in the storage unit 14. Since a large number of virtual spectra are obtained, training data including a large number of data sets is obtained. The processing of S17 corresponds to the data generation unit. After S17 is completed, the information processing device 1 ends the processing of generating training data.

[0073] In the above description of the processes of S11 to S17, an example has been shown in which a large number of virtual spectra are generated from a plurality of original spectra, but in the processes of S11 to S17, a large number of virtual spectra may be generated from a single original spectrum. Note that the processes of S11 to S17 may be executed by an information processing device that is not connected to the measurement device 2. In this embodiment, the original spectrum and the analysis results of the sample 3 are input to the information processing device, and the information processing device executes the processes of S11 to S17 using the input data.

[0074] Through the processes of S11 to S17, training data including a large number of virtual spectra is generated from the initial small number of raw spectra. The generated training data includes a large number of virtual spectra, and it becomes possible to train the learning model 142 using the training data.

[0075] In the processes of S11 to S17, a non-analytical spectrum obtained by actual measurement and separated from the original spectrum is used as a non-analytical spectrum to be added to the analytical spectrum to generate a virtual spectrum, thereby obtaining a virtual spectrum that is close to the actual spectrum. A spectrum obtained by adding together multiple non-analytical spectra obtained by actual measurement is used as a non-analytical spectrum to be added to the analytical spectrum, thereby obtaining a virtual spectrum that is close to the actual spectrum. In this embodiment, the non-analytical spectrum is a fluorescence spectrum included in the Raman spectrum. A virtual spectrum that is close to the actual Raman spectrum is obtained by using a fluorescence spectrum obtained by actual measurement as the non-analytical spectrum.

[0076] By generating training data containing virtual spectra that are close to the real spectrum, training data that is close to the training data containing the real spectrum is generated. By using training data that is close to the training data containing the real spectrum, it becomes possible to generate a learning model by machine learning that produces accurate output according to the real spectrum.

[0077] The information processing device 1 executes a learning model generation method. That is, the information processing device 1 performs processing to generate a learning model 142 that has been trained to output an analysis result of the sample 3 when a Raman spectrum is input, by learning using training data. FIG. 8 is a flowchart showing the processing steps for generating the learning model 142. The information processing device 1 acquires training data (S21). In S21, the calculation unit 11 acquires the training data by reading out the training data stored in the memory unit 14. Alternatively, the training data is input to the information processing device 1, and the information processing device 1 acquires the training data.

[0078] The information processing device 1 then uses the training data to generate a learning model 142 that has been trained to output an analysis result of the sample 3 when a Raman spectrum is input (S22). In S22, the calculation unit 11 inputs the virtual spectrum included in the training data to the learning model 142, and trains the learning model 142.

[0079] The learning model 142 performs calculations in response to the input of the virtual spectrum and outputs the analysis results of the sample 3. The analysis results of the sample 3 are, for example, the types of cells contained in the sample 3. For example, the analysis results output are the probabilities that the types of cells contained in the sample 3 are each of multiple types. The calculation unit 11 adjusts the calculation parameters of the learning model 142 so as to reduce the error between the analysis results output by the learning model 142 and the analysis results associated with the virtual spectrum input in the training data. In other words, the parameters are adjusted so that analysis results that are substantially identical to the analysis results associated with the virtual spectrum are output. The calculation unit 11 performs machine learning of the learning model 142 by repeating processing using multiple data sets included in the training data and adjusting the parameters of the learning model 142.

[0080] The calculation unit 11 stores the learned data recording the adjusted final parameters in the storage unit 14. In this way, a learning model 142 is generated that has been trained to output an analysis result of the sample 3 when a Raman spectrum is input. After S22 is completed, the information processing device 1 ends the process of generating the learning model 142. Note that the processes of S21 to S22 may be executed by an information processing device different from the information processing device 1 that executed the processes of S11 to S17. In this embodiment, the information processing device receives training data as input and executes the processes of S21 to S22 using the input training data.

[0081] The analytical device 100 analyzes the sample 3 using the trained learning model 142. FIG. 9 is a flowchart showing the steps of a process for analyzing the sample 3 using the learning model 142. The measuring device 2 measures Raman scattered light from the sample 3, generates a Raman spectrum, and inputs the Raman scattered light to the information processing device 1. The information processing device 1 acquires the Raman spectrum (S31). In S31, the information processing device 1 acquires the Raman spectrum by receiving the Raman spectrum input from the measuring device 2 via the communication unit 17, and the calculation unit 11 stores the Raman spectrum in the memory unit 14.

[0082] The information processing device 1 inputs the Raman spectrum to the learning model 142 (S32). In S32, the calculation unit 11 inputs the Raman spectrum to the learning model 142 and causes the learning model 142 to execute processing. In response to the input of the Raman spectrum, the learning model 142 outputs the analysis result of the sample 3. The information processing device 1 acquires the analysis result of the sample 3 output by the learning model 142 (S33).

[0083] The information processing device 1 then outputs the analysis results of the sample 3 (S34). For example, in S34, the calculation unit 11 displays the Raman spectrum and the analysis results of the sample 3 on the display unit 16. After S34 is completed, the information processing device 1 ends the process of analyzing the sample 3.

[0084] The analysis results of the sample 3 are obtained by the processes of S31 to S34. By using the learning model 142, the analysis results of the sample 3 can be easily obtained. For example, it is possible to easily determine the type of cells contained in the sample 3. Note that the processes of S31 to S34 may be executed by an information processing device different from the information processing device 1 that executed the processes of S11 to S17 or S21 to S22.

[0085] As described above in detail, in this embodiment, a virtual spectrum is generated by processing the original spectrum and the non-analytical spectrum, such as by adding the non-analytical spectrum to the original spectrum obtained from sample 3, and training data including the virtual spectrum is generated. By using multiple different non-analytical spectra, a large number of virtual spectra can be generated from a small number of original spectra. Because the non-analytical spectra have almost no effect on the analysis results, the analysis results based on the virtual spectrum are identical to the analysis results based on the original spectrum. Therefore, the virtual spectrum can be associated with the analysis results of sample 3, and training data including a large number of data sets in which the virtual spectrum and the analysis results of sample 3 are associated can be generated.

[0086] The generated training data includes many combinations of Raman spectra and analysis results of sample 3, so machine learning using the training data can be used to generate a learning model 142 that outputs accurate analysis results according to the Raman spectra. By using the generated learning model 142, analysis results of sample 3 based on the Raman spectra can be easily obtained. For example, the type of cells contained in sample 3 can be easily determined.

[0087] In this embodiment, the spectrum is mainly a Raman spectrum, but the spectrum may be a spectrum other than a Raman spectrum. For example, the spectrum may be a fluorescent X-ray spectrum. The spectrum may be a spectrum other than a spectrum of light or radiation.

[0088] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. In other words, embodiments obtained by combining technical means modified appropriately within the scope of the claims are also included in the technical scope of the present invention. [Explanation of symbols]

[0089] 100 Analyzer 1. Information processing equipment 10 Recording media 141 Computer Programs 142 Learning Model 2. Measuring equipment 3. Sample 41, 42 Non-analytical spectra 43 Raw Spectrum 44 Virtual Spectrum

Claims

1. 1. A data generation method for generating training data for training a learning model that outputs an analysis result of a sample when a spectrum obtained from the sample is input, the method comprising: A raw spectrum is obtained from the sample, generating a plurality of virtual spectra, each of which includes an analytical spectrum identical to an analytical spectrum included in the original spectrum and a non-analytical spectrum different from an analytical spectrum that serves as a basis for analysis of the sample, and each of which has a different shape from the original spectrum; generating training data including a plurality of data sets each of which associates the virtual spectrum with an analysis result of the sample; A data generation method comprising:

2. separating the non-analytical spectrum from the original spectrum; The virtual spectrum is generated by adding a non-analytical spectrum different from the non-analytical spectrum separated from the original spectrum to the spectrum obtained after the non-analytical spectrum has been separated from the original spectrum.

2. The data generation method according to claim 1.

3. processing the non-analytical spectrum separated from the original spectrum by adding a predetermined value, shifting a wavenumber, or inverting a value centered on a specific value contained in the non-analytical spectrum; The virtual spectrum is generated by adding the processed non-analytical spectrum to the spectrum obtained after separating the non-analytical spectrum from the original spectrum.

3. The data generating method according to claim 2.

4. Acquire multiple raw spectra, The virtual spectrum is generated by adding a non-analytical spectrum separated from another original spectrum to the spectrum obtained by separating the non-analytical spectrum from one original spectrum.

3. The data generating method according to claim 2.

5. Acquire multiple raw spectra, processing the non-analytical spectra by adding together the non-analytical spectra separated from the plurality of original spectra; The virtual spectrum is generated by adding the processed non-analytical spectrum to the spectrum obtained after separating the non-analytical spectrum from the original spectrum.

3. The data generating method according to claim 2.

6. Acquire multiple raw spectra, Dividing the non-analytical spectra separated from the plurality of original spectra into a plurality of split spectra included in a plurality of wavenumber ranges, and combining the split spectra with each other among the plurality of non-analytical spectra, thereby processing the non-analytical spectra; The virtual spectrum is generated by adding the processed non-analytical spectrum to the spectrum obtained after separating the non-analytical spectrum from the original spectrum.

3. The data generating method according to claim 2.

7. Separating a non-analytical spectrum from the original spectrum by separating components having a specific frequency band from the original spectrum.

7. The data generating method according to claim 2, wherein the data generating method is a data generating method for generating a data set.

8. The original spectrum is a Raman spectrum, the analytical spectrum is a component caused by Raman scattered light contained in the Raman spectrum, and the non-analytical spectrum is a fluorescence spectrum contained in the Raman spectrum.

8. The data generating method according to claim 1, wherein the data generating method is a data generating method for generating a data set.

9. The virtual spectrum is generated by adding a noise spectrum to the spectrum obtained by adding the non-analysis spectrum to the original spectrum.

9. The data generating method according to claim 1, wherein the data generating method is a data generating method for generating a data set.

10. Separating a noise spectrum having a specific frequency band from the original spectrum; The virtual spectrum is generated by adding a noise spectrum having the specific frequency band and different from the noise spectrum separated from the original spectrum to the spectrum obtained by separating the noise spectrum from the original spectrum and adding the non-analysis spectrum.

9. The data generating method according to claim 8.

11. Acquiring training data generated by the data generation method according to any one of claims 1 to 10, By learning using the training data, a learning model is generated that outputs an analysis result of a sample when a spectrum obtained from the sample is input. A learning model generation method characterized by:

12. A computer program that causes a computer to execute a process of generating training data for training a learning model that outputs an analysis result of a sample when a spectrum obtained from the sample is input, the computer program comprising: A raw spectrum is obtained from the sample, generating a plurality of virtual spectra, each of which includes an analytical spectrum identical to an analytical spectrum included in the original spectrum and a non-analytical spectrum different from an analytical spectrum that serves as a basis for analysis of the sample, and each of which has a different shape from the original spectrum; generating training data including a plurality of data sets each of which associates the virtual spectrum with an analysis result of the sample; A computer program that causes a computer to execute a process.

13. An information processing device that performs processing to generate training data for training a learning model that outputs an analysis result of a sample when a spectrum obtained from the sample is input, a spectrum acquisition unit that acquires a raw spectrum, which is a spectrum obtained from a sample; a virtual spectrum generating unit that uses the original spectrum and a non-analytical spectrum that is different from the analytical spectrum included in the original spectrum and serves as the basis for analysis of the sample, and generates a plurality of virtual spectra that include analytical spectra identical to the analytical spectra included in the original spectrum and have shapes different from those of the original spectrum; a data generating unit that generates training data including a plurality of data sets each of which associates the virtual spectrum with the analysis result of the sample.

1. An information processing device comprising:

14. In an analytical device for analyzing a sample, a measurement device that measures a physical phenomenon occurring in a sample and generates a spectrum that represents a characteristic of the measured physical phenomenon; an information processing device, The information processing device includes: a virtual spectrum generating unit that uses an original spectrum generated by the measurement device and a non-analytical spectrum that is different from the analytical spectrum included in the original spectrum and that serves as the basis for analysis of the sample, to generate a plurality of virtual spectra that include analytical spectra identical to the analytical spectra included in the original spectrum and have shapes different from those of the original spectrum; a data generation unit that generates training data including a plurality of data sets each of which associates the virtual spectrum with an analysis result of the sample; An analytical device comprising:

Images