Learning methods and learning programs

Abstract

To perform learning reflecting noise that can be generated when an analysis device analyzes a sample.SOLUTION: A learning method includes: acquiring a pseudo noise waveform indicating an assumed noise (step S2); acquiring a pseudo peak waveform not including pseudo noise (step S4); generating a pseudo signal waveform by adding the pseudo noise waveform and the pseudo peak waveform (step S6); and updating an estimation model on the basis of the pseudo signal waveform (step S8). The acquisition of the pseudo noise waveform has: causing an analysis device to execute noise measurement generating the pseudo noise waveform a plurality of times; calculating a similarity degree of a plurality of pseudo noise waveforms generated by the noise measurement performed the plurality of times; and executing predetermined processing according to the similarity degree.SELECTED DRAWING: Figure 9

Description

Technical Field

[0001] The present disclosure relates to a learning method and a learning program.

Background Art

[0002] International Publication No. 2020 / 070786 (Patent Document 1) discloses a chromatograph system. This chromatograph system separates and detects peaks of unresolved peaks in a chromatogram by AI (Artificial Intelligence) using an estimation model. Then, the chromatograph system performs qualitative analysis or quantitative analysis of a sample based on the peaks.

[0003] In Patent Document 1, it is disclosed that a computer performs learning for updating an estimation model. This computer acquires a plurality of chromatograms each having a peak, and creates a chromatogram of unresolved peaks by adding the plurality of chromatograms together. The computer updates the estimation model using the plurality of chromatograms as learning data and the created chromatogram as teacher data.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Generally, in an analytical apparatus such as a chromatograph, a signal waveform (for example, a chromatogram) including a noise waveform of noise that may occur when analyzing a sample may be generated. Patent Document 1 does not disclose learning in which the noise waveform is reflected.

[0006] This disclosure was made to solve these problems, and its purpose is to perform learning that reflects the noise waveforms that may be generated when the analytical instrument analyzes a sample. [Means for solving the problem]

[0007] The learning method of this disclosure is a learning method for learning an estimation model used to detect peaks in signal waveforms created by an analytical instrument that analyzes a sample. The learning method comprises acquiring noise waveforms that may occur when the analytical instrument performs analytical processing, and learning an estimation model based on the noise waveforms. Furthermore, acquiring noise waveforms involves acquiring multiple noise waveforms generated by noise measurements performed multiple times by the analytical instrument, calculating the degree of similarity of the multiple noise waveforms, and performing predetermined processing according to the degree of similarity.

[0008] The learning program of this disclosure is a learning program that causes a computer to update an estimation model used to detect peaks in signal waveforms created by an analytical instrument that analyzes a sample. The learning program causes the computer to acquire noise waveforms that may occur when the analytical instrument performs analytical processing, and to learn the estimation model based on the noise waveforms. Acquiring noise waveforms involves acquiring multiple noise waveforms generated by noise measurements performed multiple times by the analytical instrument, calculating the degree of similarity of the multiple noise waveforms, and performing predetermined processing according to the degree of similarity. [Effects of the Invention]

[0009] According to this disclosure, the analytical instrument can perform learning that reflects the noise that may occur when analyzing a sample. [Brief explanation of the drawing]

[0010] [Figure 1] This is a diagram showing an example of the configuration of an analysis system. [Figure 2] This is a block diagram showing the hardware configuration of the learning device 30. [Figure 3] This is a diagram to explain pseudochromatograms. [Figure 4] This is an example of a blank chromatogram that does not contain unexpected noise. [Figure 5] This is an example of a blank chromatogram containing unexpected noise M. [Figure 6] This is a functional block diagram of the learning device. [Figure 7] This is a functional block diagram of the GAN execution unit and the update unit. [Figure 8] This is a functional block diagram of the noise generation unit. [Figure 9] This is a flowchart showing the processing of the analysis system. [Figure 10] This is a flowchart showing the process in step S2. [Figure 11] This figure shows an example of the results obtained from a predetermined measurement. [Figure 12] This is a flowchart of the process in step S2 of the second embodiment. [Figure 13] This is an example of a specified image. [Modes for carrying out the invention]

[0011] Embodiments of the present invention will be described in detail below with reference to the drawings. The same or corresponding parts in the drawings are denoted by the same reference numerals, and their descriptions will not be repeated.

[0012] <First Embodiment> [Analysis System] This disclosure relates to a learning technique for updating an estimation model used to detect peaks in signal waveforms created by an analytical instrument. Analytical instruments include, for example, gas chromatographs (GC), liquid chromatographs (LC), mass spectrometers, spectrophotometers, and X-ray analyzers.

[0013] For example, the signal waveform may be a chromatogram waveform or a mass spectrum waveform. When the analyzer is a spectrophotometer, the signal waveform becomes an absorption spectrum waveform. When the analyzer is an X-ray analyzer, the signal waveform becomes an X-ray spectrum waveform.

[0014] In addition, the learning (learning process) of the estimation model (estimation model 121 described later) includes a process of newly generating (constructing) an unconstructed estimation model and a process of updating an already constructed estimation model. "Updating the estimation model" includes a process of updating the parameters of the estimation model. Also, the estimation model updated (optimized) by the learning process is also referred to as a "trained model". Also, the estimation model before learning and the trained estimation model are collectively referred to as the "estimation model". Hereinafter, an example where "the learning of the estimation model is an update of the estimation model" will be mainly described.

[0015] In the present embodiment, an analysis system employing a liquid chromatograph will be described. FIG. 1 is a diagram showing a configuration example of the analysis system 100. The analysis system 100 includes an analyzer 35, an input device 61, a display device 65, and a learning device 30. The analyzer 35 includes a measurement unit 10 and a data analysis device 25. The data analysis device 25 and the learning device 30 are configured by, for example, an information processing device (for example, a PC (Personal Computer)). In the example of FIG. 1, the data analysis device 25 and the learning device 30 are shown separately, but they may be integrated. The learning device 30 corresponds to the "computer" of the present disclosure.

[0016] The input device 61 is a pointing device such as a keyboard or mouse, and receives commands from the user. The display device 65 is composed of, for example, a liquid crystal display (LCD) panel. The display device 65 displays various images. When a touch panel is used as the user interface, the input device 61 and the display device 65 are integrally formed. The input device 61 is connected to the data analysis device 25 and the learning device 30. The display device 65 is also connected to the data analysis device 25.

[0017] The data analysis device 25 has a control unit 20. The control unit 20 controls the measurement unit 10. The measurement unit 10 includes a mobile phase container 11, a pump 12, an injector 13, a column 14, a detector 15, and a placement unit 18. The sample S to be analyzed is placed in the placement unit 18. The sample S becomes a sample solution by being dissolved in a predetermined solvent. The mobile phase is contained in the mobile phase container 11. The pump 12 sucks the mobile phase contained in the mobile phase container 11 and delivers it to the column 14 at a substantially constant flow rate (or flow rate).

[0018] The injector 13 injects a predetermined amount of sample solution into the mobile phase at a predetermined timing in accordance with instructions from the control unit 20. The injected sample solution is introduced into the column 14 via the flow of the mobile phase. Various components contained in the sample solution are separated and eluted over time as they pass through the column 14. In other words, the column 14 separates the components contained in the sample solution according to their retention time.

[0019] The detector 15 detects components in the eluate eluted from the column 14. The detector 15 outputs a detection signal with an intensity corresponding to the amount of the component to the data analysis device 25. The detector 15 may be an optical detector, such as one employing a photodiode array (PDA) detector.

[0020] In addition to the control unit 20 described above, the data analysis device 25 includes a data acquisition unit 110, a peak detection processing unit 111, and an analysis unit 117.

[0021] The data acquisition unit 110 samples the detection signal output from the detector 15 at predetermined time intervals and converts it into digital data. The data acquisition unit 110 stores this digital data in a predetermined storage area (not shown). This digital data is data representing a chromatogram waveform (hereinafter also referred to as "chromatogram data").

[0022] The peak detection processing unit 111 estimates (derives) the peaks of the chromatogram from the chromatogram data collected by the data acquisition unit 110 using AI (Artificial Intelligence).

[0023] In this embodiment, the peak detection processing unit 111 includes a model storage unit 114 and a peak determination unit 116. The model storage unit 114 stores, for example, an estimated model 121 (neural network) generated by machine learning. This estimated model 121 is represented by, for example, a predetermined function. The predetermined function is, for example, an EMG (Exponetially Modified Gaussian) function.

[0024] Furthermore, the peak determination unit 116 inputs the chromatogram obtained from the chromatogram data collected by the data acquisition unit 110 into the estimation model 121. The estimation model 121 outputs the peaks of the chromatogram. In this way, the peak detection processing unit 111 estimates the peaks of the chromatogram obtained from the chromatogram data collected by the data acquisition unit 110 and outputs them to the analysis unit 117.

[0025] The duration (retention time) at which a peak is observed corresponds to the type of component. The chromatogram is transmitted to a data analysis device. The data analysis device identifies the components based on the retention times of the peaks in the chromatogram. This identification is also referred to as "qualitative analysis."

[0026] Furthermore, the peak height and peak area of ​​a chromatogram correspond to the concentration or content of the sample's components. A data analysis device identifies the concentration or content of the sample's components from the peak height or area values ​​in the chromatogram. This identification is also referred to as "quantitative analysis."

[0027] The analysis unit 117 determines the position (time) of the peak top and the area value (or height) of the peak output from the peak determination unit 116. The analysis unit 117 identifies the components from the position information of each peak on the chromatogram. The analysis unit 117 also calculates the content of each component from the peak area value (or height value) using a pre-created calibration curve. In this way, the analysis unit 117 performs qualitative and quantitative analysis of each component contained in the sample. The analysis unit 117 displays the qualitative and quantitative analysis results on the display device 65.

[0028] Furthermore, the learning device 30 updates the estimated model 121 (performs training on the estimated model 121), as will be described later. By performing a predetermined operation on the input device 61, the mode of the analysis system 100 is switched to the learning mode. The learning device 30 updates the estimated model 121 during this learning mode.

[0029] [Hardware configuration of the learning device] Figure 2 is a block diagram showing the hardware configuration of the learning device 30 according to this embodiment. As shown in Figure 2, the learning device 30 comprises a control device 21, a storage device 19, a media reader 17, and a communication interface 23 as its main hardware elements.

[0030] The control unit 21 updates the estimated model 121, as described later. The control unit 21 is composed of, for example, a CPU (Central Processing Unit), an FPGA (Field Programmable Gate Array), and a GPU (Graphics Processing Unit). The control unit 21 may consist of at least one of the CPU, FPGA, and GPU, or it may consist of a CPU and FPGA, an FPGA and a GPU, a CPU and a GPU, or all of the CPU, FPGA, and GPU. The control unit 21 may also consist of processing circuitry.

[0031] The storage device 19 includes a volatile storage area (e.g., a working area) for temporarily storing program code, work memory, etc., when the control device 21 executes an arbitrary program. For example, the storage device 19 is composed of a volatile memory device such as DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory). Furthermore, the storage device 19 includes a non-volatile storage area. For example, the storage device 19 is composed of a non-volatile memory device such as a hard disk or SSD (Solid State Drive).

[0032] In this embodiment, an example is shown in which volatile and non-volatile storage areas are included in the same storage device 19. However, volatile and non-volatile storage areas may be included in different storage devices. For example, the control device 21 may include volatile storage areas, and the storage device 19 may include non-volatile storage areas. The learning device 30 may include a microcomputer that includes the control device 21 and the storage device 19.

[0033] The storage device 19 stores the estimated model 121 and the control program 122. The estimated model 121 includes a neural network and parameters used in the processing of the neural network.

[0034] The estimation model 121 includes at least a program capable of machine learning, and its parameters are optimized (adjusted, updated) by performing machine learning based on training data (teaching data). The learning device 30 transmits the optimized estimation model (estimated model 121A in Figure 7) to the data analysis device 25. The data analysis device 25 updates the estimation model 121 stored in the model storage unit 114 with the transmitted estimation model 121A (optimized estimation model). In this way, the peak detection processing unit 111 can improve the accuracy of peak estimation by estimating peaks using the updated estimation model 121.

[0035] The process of updating the parameters of the estimated model 121 is also called the "learning process." The estimated model 121 optimized by the learning process is also called the "trained model." In this embodiment, the estimated model 121 before learning and the estimated model 121 after learning are collectively referred to as the "estimated model." In particular, the trained estimated model 121 is also called the "trained model." The control program 122 is a program executed by the control device 21.

[0036] The media reader 17 accepts a recording medium 130, such as a removable disk, and acquires data stored on the recording medium 130. This data is, for example, a control program. The control program 122 may also be stored on the recording medium 130 (for example, a removable disk) and distributed as a program product. Alternatively, the control program 122 may be provided by an information provider as a program product that can be downloaded via the so-called Internet. The control device 21 reads the program from the recording medium 130 or the program provided via the Internet. The control device 21 stores the read program in a predetermined storage area (storage area of ​​the storage device 19). The control device 21 executes the learning process described later by executing the stored control program 122.

[0037] The recording medium 130 is not limited to DVD-ROM (Digital Versatile Disk Read Only Memory), CD-ROM (compact disc read-only memory), FD (Flexible Disk), or hard disk, but may also be a medium that permanently stores a program, such as magnetic tape, cassette tape, optical disc (MO (Magnetic Optical Disc) / MD (Mini Disc) / DVD (Digital Versatile Disc)), optical card, mask ROM, EPROM (Electronically Programmable Read-Only Memory), EEPROM (Electronically Erasable Programmable Read-Only Memory), or semiconductor memory such as flash ROM. Furthermore, the recording medium 130 is a non-temporary medium from which the control program 122 and the like can be read by a computer.

[0038] The communication interface 23 is an interface for connecting the data analysis device 25 and enables data input and output between the learning device 30 and the data analysis device 25. In this embodiment, the learning device 30 acquires measured chromatogram data and blank chromatogram data transmitted from the data analysis device 25 via the communication interface 23. Hereinafter, measured chromatogram data and blank chromatogram data will be collectively referred to as chromatogram data. The learning device 30 also transmits the updated estimation model 121 to the data analysis device 25 via the communication interface 23.

[0039] [Pseudochromatogram] Next, we will explain pseudochromatograms. Pseudochromatograms are information that the learning device 30 uses to update the estimated model 121. Figure 3 is a diagram illustrating pseudochromatograms. In the chromatograms shown in Figures 3(A) to 3(C), the vertical axis represents signal intensity, and the horizontal axis represents time (retention time).

[0040] Figure 3(A) shows an example of pseudo-noise. Pseudo-noise is noise that simulates expected noise. Expected noise is noise that may occur (is expected to occur) when the analyzer 35 analyzes the sample S. Causes of expected noise include contamination of the column 14 with impurities or baseline drift. There is one or more types of expected noise. In this embodiment, pseudo-noise is generated when the analyzer 35 performs a blank measurement.

[0041] Here, blank measurement is a process performed by the analyzer 35 executing an analytical process while no sample S is placed in the placement section 18. Blank measurement includes, for example, the following first blank measurement, second blank measurement, third blank measurement, and fourth blank measurement.

[0042] The first blank measurement is, for example, when the analyzer 35 is configured as a gas chromatograph, a process in which only the carrier gas is analyzed without the sample S being placed in the placement unit 18.

[0043] The second blank measurement is a process in which a first pseudo-sample is generated consisting only of a solvent without adding a reagent (e.g., a conductive reagent) and the sample, and the first pseudo-sample is analyzed. The third blank measurement is a process in which a second pseudo-sample is generated consisting only of a reagent and a solvent without adding the sample, and the pseudo-sample is analyzed.

[0044] Furthermore, the fourth blank measurement utilizes the same pretreatment performed when analyzing a sample. The fourth blank measurement involves performing the same pretreatment as described above on either the carrier gas, the first pseudo-sample, or the second pseudo-sample, and then analyzing the substance after the pretreatment.

[0045] Furthermore, among the first to fourth blank measurements, contamination occurs most frequently and noise intensity increases in the order of fourth blank measurement, third blank measurement, second blank measurement, and first blank measurement. In other words, the noise intensity that can occur is highest in the fourth blank measurement and lowest in the first blank measurement.

[0046] In this embodiment, the measurement unit 10 generates a blank chromatogram by performing N blank measurements (where N is an integer of 2 or more). This blank chromatogram corresponds to pseudo-noise. Furthermore, this pseudo-noise corresponds to the "noise waveform" in this disclosure. Thus, since blank measurement is a measurement for generating pseudo-noise, it is also referred to as "noise measurement."

[0047] Figure 3(B) shows an example of a false peak. A false peak is a waveform that simulates a peak in a chromatogram. False peaks do not contain noise and include both peaks and baselines. As will be described later, false peaks are information generated when the analyzer 35 analyzes the sample S.

[0048] Figure 3(C) shows an example of a pseudochromatogram. The learning device 30 generates a pseudochromatogram by adding pseudo-noise and pseudo-peaks. The learning device 30 then updates the estimation model 121 based on the generated pseudochromatogram. Specifically, the learning device 30 updates the estimation model 121 using the pseudochromatogram as training data and the pseudo-peaks as training data.

[0049] Furthermore, the measurement unit 10 generates a chromatogram by analyzing a newly identified unknown sample. In this analysis, the detection signal from the detector 15 may contain the aforementioned assumed noise. In this case, a chromatogram with noise in the baseline is generated.

[0050] The data analysis device 25 detects peaks by inputting chromatogram data containing the assumed noise into the updated estimation model 121. Here, as described above, the estimation model 121 is trained using pseudo-chromatograms as training data and pseudo-peaks as training data. Therefore, by using the estimation model 121, the data analysis device 25 can appropriately detect peaks, i.e., chromatograms that exclude the assumed noise. In this way, even when a chromatogram containing assumed noise is generated when analyzing a new unknown sample, the accuracy of peak detection can be improved by using the estimation model 121. Furthermore, the learning device 30 can learn the estimation model under the same conditions as those generated when analyzing an unknown sample (a situation in which a signal waveform is generated).

[0051] Furthermore, the learning device 30 may update the estimation model 121 using pseudo-noise as training data without including pseudo-peaks. In other words, the training data (training dataset) includes at least one of a first pseudo-chromatogram that includes pseudo-peaks and a second pseudo-chromatogram that does not include pseudo-peaks. The first pseudo-chromatogram is, for example, a pseudo-chromatogram in which pseudo-peaks and pseudo-noise are added together. The second pseudo-chromatogram is, for example, a pseudo-chromatogram that does not include pseudo-peaks and is composed of pseudo-noise.

[0052] [Regarding unexpected noise] Next, we will explain unintended noise. As mentioned above, pseudo-noise is generated by performing blank measurements. However, pseudo-noise may contain unintended noise. Unintended noise is noise that cannot be included when the measurement unit 10 analyzes the data. Therefore, if the learning device 30 updates the estimation model 121 using pseudo-noise that contains unintended noise, the accuracy of peak detection using the estimation model 121 will decrease (the quality of the estimation model will decrease). Unintended noise is also called "unintended noise."

[0053] Here, we will explain why unexpected noise may be included. For example, during blank measurement, an unexpected foreign substance may be introduced into one of the components or the solvent of the measurement unit 10. Such a component is, for example, the pump 12. If a blank measurement is performed with such foreign substance present, a peak (unexpected noise) caused by the foreign substance will be included in the false noise.

[0054] Furthermore, for example, when the measurement unit 10 is composed of a gas chromatograph, an unknown sample is analyzed, and the unknown sample is derivatized using a derivatization reagent. When the measurement unit 10 is composed of a gas chromatograph, the sample S is placed in the placement unit 18, while a blank measurement is performed using a derivatization reagent. Consequently, the blank measurement includes peaks (unintended noise) originating from the derivatization reagent as false noise.

[0055] Therefore, in this embodiment, the learning device 30 determines whether or not unintended noise is included in the pseudo-noise based on the correlation coefficient of each of the N blank chromatograms. If it is determined that unintended noise is included, the pseudo-noise containing the unintended noise is excluded and the estimation model 121 is updated. This improves the accuracy of peak detection using the estimation model 121.

[0056] Figures 4 and 5 are examples of blank chromatograms generated by performing 60 (=N) blank measurements. Figure 4 is an example of a blank chromatogram without unintended noise (pseudo-noise). Figure 5 is an example of a blank chromatogram with unintended noise M (pseudo-noise).

[0057] As shown in the upper part of Figure 4, if no unexpected noise is present, a phenomenon may occur where different expected noise is included in the blank chromatogram each time N blank measurements are performed. Considering this phenomenon, the correlation coefficient between one chromatogram Bn (n=1,...N) among the N blank chromatograms and another chromatogram Bm (m=1,...N) that is different from Bn will be low. Note that m and n are different (m≠n).

[0058] Therefore, the learning device 30 calculates L correlation coefficients between chromatogram Bn and other chromatograms Bm. Note that L = N There are two C2s. The learning device 30 then calculates the average or sum of the L correlation coefficients. In this embodiment, the learning device 30 calculates the average of the L correlation coefficients. If the average is less than a predetermined threshold, it is determined that the blank chromatogram does not contain unexpected noise.

[0059] The lower part of Figure 4 shows an example of a heatmap of L correlation coefficients. The vertical axis of this heatmap represents chromatogram Bn, and the horizontal axis represents the other chromatograms Bm. As shown in the lower part of Figure 4, if unintended noise is not included in the pseudo-noise, the correlation coefficients will be low overall. In the example in the lower part of Figure 4, the average value of the L correlation coefficients is 0.03. The correlation coefficient corresponds to the "similarity" in this disclosure. The similarity is a value that indicates the degree of similarity between chromatogram Bn, one of N chromatograms, and the other chromatograms Bm. In this embodiment, the more similar chromatogram Bn and the other chromatograms Bm are, the larger the similarity (correlation coefficient). As a modification, a similarity value that decreases as chromatogram Bn and the other chromatograms Bm become more similar may be adopted. Note that the degree of similarity may also be expressed using other parameters, for example, the distance between a chromatogram Bn with a similarity of 1 and another chromatogram Bm (e.g., the Fréchet distance).

[0060] On the other hand, as shown in the upper part of Figure 5, if unexpected noise is present, a peak may occur in the same holding time in any of the N blank measurements. In the upper part of Figure 5, a peak occurs as unexpected noise M. Therefore, the average value of the L correlation coefficients becomes large. Thus, the learning device 30 determines that unexpected noise is present in the blank chromatogram if the average value of the L correlation coefficients is greater than or equal to the above threshold.

[0061] The lower part of Figure 5 shows an example of a heatmap of L correlation coefficients. As shown in the lower part of Figure 5, when unintended noise is included in the pseudo-noise, the correlation coefficients are generally higher. In the example in the lower part of Figure 5, the average value of the L correlation coefficients is assumed to be 0.97.

[0062] [Functional blocks for learning devices] Figure 6 is a functional block diagram of the learning device 30. The learning device 30 includes an input unit 32, an extraction unit 34, a GAN execution unit 36, a noise generation unit 38, and an update unit 40. The learning device 30 is a device that updates the estimated model 121 in learning mode. When the user performs a predetermined operation on the input device 61, the mode of the analysis system 100 is switched to learning mode.

[0063] The input unit 32 receives measured chromatogram data and blank chromatogram data. Measured chromatogram data is obtained by the measurement unit 10 measuring a sample set in the placement unit 18. The sample may be an unknown sample with unknown components, or a known sample with known components. Blank chromatogram data is obtained by the measurement unit 10 measuring a sample when it is not set in the placement unit 18. As a modification, at least one of the measured chromatogram data and blank chromatogram data may be generated by an apparatus equivalent to the measurement unit 10.

[0064] The measured chromatogram data is input to the extraction unit 34, and the blank chromatogram data is input to the noise generation unit 38. The chromatogram data is assigned a flag. This flag is information used to determine whether the data is measured chromatogram data or blank chromatogram data. Measured chromatogram data is assigned a measured flag to indicate that it is measured chromatogram data. Blank chromatogram data is assigned a blank flag to indicate that it is blank chromatogram data. The input unit 32 determines whether the chromatogram data is measured chromatogram data or blank chromatogram data by determining the flag attached to the chromatogram data. The input unit 32 transmits the measured chromatogram data to the extraction unit 34 and the blank chromatogram data to the noise generation unit 38.

[0065] The extraction unit 34 extracts peaks with good waveform shapes from the measured chromatogram. As an example of a specific method, the extraction unit 34 removes predetermined peaks from the measured chromatogram data. These predetermined peaks include, for example, peaks with extremely low signal-to-noise ratios and peaks with insufficient separation. The extraction unit 34 then calculates shape parameters (hereinafter also referred to as "peak parameters") related to the extracted peaks. The peak parameters include, for example, tailing degree, leading degree, peak width, and signal-to-noise ratio (SN ratio), at least one of these.

[0066] Furthermore, the extraction unit 34 calculates not only peak parameters but also chromatogram parameters. Chromatogram parameters include the number of peaks in the chromatogram, the position of peaks near the center of the waveform's time axis, and the distance between adjacent peaks. Chromatogram parameters are used to determine the position of peaks in the chromatogram waveform.

[0067] Thus, the extraction unit 34 outputs the extracted peak parameters and chromatogram parameters as real peak waveforms to the GAN execution unit 36. Alternatively, the extraction unit 34 may extract peaks from the measured chromatogram data based on the current estimation model 121A. The GAN execution unit 36 ​​performs processing using a Generative Adversarial Network (GAN). The technology for GANs is disclosed, for example, in International Publication No. 2021 / 261202. Specifically, the GAN execution unit 36 ​​updates the generator 41 for generating pseudo-peak waveforms. The GAN execution unit 36 ​​transmits information about the generator 41 (for example, updated parameters). Therefore, the generator 41 of the GAN execution unit 36 ​​and the generator 41 of the update unit 40 are the same. Details of the processing of the GAN execution unit 36 ​​will be described later. The pseudo-peak waveforms are used for training the estimation model 121.

[0068] Furthermore, the noise generation unit 38 generates and acquires a pseudo-noise waveform (see Figures 4 and 5 above) based on the blank chromatogram data from the input unit 32. Here, the pseudo-noise waveform is used for training the estimation model 121. The processing of the noise generation unit 38 will be described later.

[0069] The update unit 40 creates a pseudo-peak (see Figure 3(B)). The update unit 40 then generates a pseudo-chromatogram by adding the pseudo-peak to the pseudo-noise generated by the noise generation unit 38. The learning device 30 then updates the estimated model 121A using the pseudo-chromatogram as training data and the pseudo-peak as training data. The estimated model 121A is stored in the storage unit 52 (see Figure 7). The update unit 40 transmits the updated estimated model 121A to the data analysis device 25. The data analysis device 25 updates the estimated model 121 stored in the model storage unit 114 with the estimated model 121A transmitted from the learning device 30. Therefore, the estimated model 121 and the estimated model 121A are synchronized. If the data analysis device 25 and the learning device 30 are integrated, the update unit 40 updates the estimated model 121.

[0070] Figure 7 is a functional block diagram of the GAN execution unit 36 ​​and the update unit 40. The GAN execution unit 36 ​​includes a random noise generation unit 46, a generator 41, a data selection unit 42, a discriminator 43, a determination unit 44, and an update processing unit 45. A predetermined neural network is used as the generator 41 and the discriminator 43. This predetermined neural network is disclosed, for example, in the following references 1 and 2.

[0071] Reference 1 is "Alec Radford, et al., 'Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks,' [online], accessed January 26, 2022], Internet<URL:https: / / arxiv.org / abs / 1511.06434> "

[0072] Reference 2 is "Ian J. Goodfellow, et al., 'Generative Adversarial Nets,' [online], [accessed January 26, 2022], Internet."<URL:https: / / arxiv.org / pdf / 1406.2661.pdf> "

[0073] The random noise generation unit 46 generates noise randomly and outputs the noise to the generator 41. The data of the real peak waveform transmitted from the extraction unit 34 is input to the data selection unit 42.

[0074] The generator 41 creates a function based on the noise from the random noise generation unit 46 and the neural network. This function takes time (elapsed time) as an argument and outputs false peak waveform data. The generator 41 generates false peak waveform data by inputting time (elapsed time) into the created function. Furthermore, the number of data points in the false peak waveform is the same as the number of data points in the real peak waveform output from the extraction unit 34.

[0075] The data selection unit 42 alternately switches between false peak waveform data and genuine peak waveform data and inputs them to the classifier 43. The classifier 43 identifies whether the input waveform data is genuine or not. The classifier 43 outputs the identification result to the determination unit 44. The determination unit 44 determines whether the identification result is correct or not. In other words, this determination confirms whether the classifier 43 identified genuine peak waveform data as genuine when it was input to the classifier 43, and whether it identified false peak waveform data as false when it was input to the classifier 43.

[0076] The update processing unit 45 updates the neural network parameters (coefficients) based on the determination result from the determination unit 44 so that the performance of the generator 41 and the discriminator 43 improves. In this way, GANs perform learning by having the generator 41 and the discriminator 43 compete with each other to improve their respective performance. Specifically, the update processing unit 45 updates the neural network parameters in the generator 41 so that the generator 41 generates a high-quality function. Here, a high-quality function is a function that can generate false peak waveform data that is as close as possible to real peak waveform data.

[0077] Next, the update unit 40 will be explained using Figure 7. The update unit 40 includes a random noise generation unit 46, a generator 41, an adder 53, a learning execution unit 51, and a memory unit 52. As described above, the generator 41 of the GAN execution unit 36 ​​and the generator 41 of the update unit 40 are the same. In other words, the generator 41 updated by the update processing unit 45 of the GAN execution unit 36 ​​is also used in the update unit 40. Furthermore, the random noise generation unit 46 of the GAN execution unit 36 ​​is also used in the update unit 40.

[0078] As described above, the random noise generation unit 46 generates noise randomly and outputs the random noise to the generator 41. The generator 41 generates a pseudo-peak waveform. This pseudo-peak waveform corresponds to the "peak waveform" in this disclosure. As described above, the performance of the generator 41 is improved by the update processing unit 45. Therefore, the generator 41 can generate a pseudo-peak waveform that is close to the real peak waveform. The generated pseudo-peak waveform is output to the adder 53 and the learning execution unit 51.

[0079] The summing unit 53 generates a pseudo-chromatogram (see Figure 3(C)) by adding the pseudo-peak waveform from the generator 41 (see Figure 3(B)) and the pseudo-noise waveform from the noise generation unit 38 (see Figure 3(A)). This pseudo-chromatogram corresponds to the "signal waveform" in this disclosure. The learning execution unit 51 updates the estimation model 121A using the pseudo-chromatogram as training data and the pseudo-peak as training data.

[0080] The learning execution unit 51 then transmits the updated estimation model 121A to the data analysis device 25. The data analysis device 25 updates the estimation model 121 to estimation model 121A.

[0081] Figure 8 is a functional block diagram of the noise generation unit 38. The noise generation unit 38 includes a calculation unit 62 and a determination unit 64.

[0082] As described above, in learning mode, the user performs N blank measurements. From these N blank measurements, the data analysis device 25 creates N blank chromatograms. The calculation unit 62 obtains these N blank chromatograms from the data analysis device 25. The calculation unit 62 calculates the L correlation coefficients mentioned above using the N blank chromatograms. Furthermore, the calculation unit 62 calculates the average value of these L correlation coefficients. The calculated average value is output to the judgment unit 64.

[0083] The determination unit 64 determines whether the average value is greater than or equal to the threshold value. If the average value is less than the threshold value, it means that no unexpected noise waveforms are included in the blank chromatogram (Figure 4). In this case, the determination unit 64 acquires N blank chromatograms as pseudo-noise waveforms. The determination unit 64 then transmits the pseudo-noise waveforms to the summation unit 53.

[0084] On the other hand, the case where the average value is above the threshold is when an unexpected noise waveform is included in the blank chromatogram (Figure 5). Therefore, in this case, the determination unit 64 discards N blank chromatograms.

[0085] As described above, when analyzing the sample S in the analyzer 35, a chromatogram containing noise (assumed noise) that may occur may be generated. Therefore, as shown in Figure 3, in learning mode, the learning device 30 generates a pseudo-chromatogram by adding pseudo-noise that simulates the assumed noise and pseudo-peaks that simulate peaks. Then, the learning device 30 updates the estimation model 121 based on the pseudo-peaks and pseudo-chromatogram. This improves the accuracy of peak detection compared to "peak detection using an estimation model that does not reflect assumed noise".

[0086] Furthermore, in this embodiment, by performing N blank measurements, the learning device 30 generates N blank chromatograms as pseudo-peaks. However, if foreign matter is introduced into the measurement unit 10 during the blank measurement, the blank chromatograms may contain waveforms of unintended noise. Therefore, the learning device 30 calculates the average value of L correlation coefficients for the N blank chromatograms. If the average value is below a threshold, the learning device 30 uses the N blank chromatograms as pseudo-noise waveforms. On the other hand, if the average value is below a threshold, the learning device 30 discards the N blank chromatograms. Thus, the learning device 30 can determine if unintended noise waveforms are included in the pseudo-noise waveforms and can perform processing according to the unintended noise waveforms. In addition, the learning device 30 automatically excludes pseudo-noise waveforms that contain unintended noise waveforms. Therefore, the deterioration of the quality of the estimation model 121 can be suppressed without burdening the user.

[0087] Furthermore, the assumed noise waveform is generated by performing a blank measurement without using the sample S. Therefore, the learning device 30 can generate the assumed noise waveform without consuming the sample S.

[0088] Furthermore, the degree of similarity is the average value of the correlation coefficients (L correlation coefficients) of each of the N pseudo-noise waveforms. Therefore, the learning device 30 can use the known parameter of correlation coefficient to determine whether or not an unexpected noise waveform is included in the pseudo-noise waveform.

[0089] Furthermore, the learning device 30 acquires pseudo-peak waveforms using a GAN. Therefore, since it can acquire pseudo-peak waveforms that are close to real peak waveforms, the quality of the pseudo-peak waveforms can be improved.

[0090] [flowchart] Figure 9 is a flowchart showing the processing of the analysis system 100 including the learning device 30. In step S2, the learning device 30 generates and acquires a pseudo-noise waveform. Next, in step S4, the learning device 30 generates and acquires a pseudo-peak waveform. The part in parentheses in step S4 will be explained in the second embodiment. Next, in step S6, the learning device 30 generates and acquires a pseudo-chromatogram by adding the pseudo-noise waveform and the pseudo-peak waveform. Next, in step S8, the learning device 30 updates the estimated model 121A (estimated model 121) based on the pseudo-peak waveform and the pseudo-chromatogram.

[0091] Figure 10 is a flowchart showing the process in step S2. In step S20, the measurement unit 10 performs N blank measurements. Next, in step S22, the learning device 30 calculates the correlation coefficient (L correlation coefficients) for each of the N blank chromatograms obtained from the N blank measurements. Next, in step S24, the learning device 30 calculates the average value of the L correlation coefficients.

[0092] Next, in step S26, the learning device 30 determines whether the average value is less than a threshold. If the average value is less than a threshold (i.e., if it is determined to be YES in step S26), that is, if no unexpected noise is included in the blank chromatogram, the learning device 30 acquires N blank chromatograms as pseudo-noise. The learning device 30 then transmits the pseudo-noise (N blank chromatograms) to the summing unit 53.

[0093] On the other hand, if the average value is above the threshold (i.e., if NO is determined in step S26), that is, if unexpected noise is included in the blank chromatogram, the learning device 30 discards N blank chromatograms.

[0094] <Second Embodiment> In the first embodiment, the learning device 30 was described as having a configuration that acquires pseudo-noise by blank measurement. In the second embodiment, the learning device 30 acquires pseudo-noise by placing a sample in the placement unit 18 and performing a predetermined measurement on the sample. The sample may be a known sample with known components, or an unknown sample with unknown components.

[0095] Here, when the signal waveform generated by the analyzer 35 is represented in two dimensions, the axis in the first direction (X-axis) is represented by intervals, and the axis in the second direction (Y-axis) is represented by intensity. In this embodiment, the interval represented by the axis in the first direction includes a first interval and a second interval. The predetermined measurement is a measurement in which a peak based on the sample (hereinafter also referred to as "actual peak") occurs in the first interval of the signal waveform, while no peak is detected in the second interval, which is different from the first interval. In the second interval, assumed noise (hereinafter also referred to as "actual noise") is detected.

[0096] If the analytical instrument is a chromatograph, the signal waveform is a chromatogram, the first interval is the first time zone, and the second interval is the second time zone. If the analytical instrument is another instrument (for example, an X-ray analyzer), the signal waveform is an X-ray spectral waveform, the first interval is the first energy band, and the second interval is the second energy band. The first time zone (first interval) is also called the peak interval, and the second time zone (second interval) is also called the non-peak interval. In this embodiment, the learning device 30 updates the estimation model 121 by treating the actual peak waveform as a pseudo-peak waveform and the actual noise waveform as a pseudo-noise waveform. As a modification, the learning device 30 may update the estimation model 121 based only on the actual peak waveform. As another modification, the learning device 30 may update the estimation model 121 based only on the actual noise waveform.

[0097] The specified measurement is, for example, a measurement in MRM (Multiple Reaction Monitoring) mode (hereinafter also referred to as "MRM measurement"). MRM is a method in which various ions ionized by an ionization probe are selected in the first stage mass spectrometer, the precursor ions are dissociated in a collision cell, and then a specific ion is detected from the broken-down ions (product ions) in the second stage mass spectrometer. The component is defined by its compound number. The specified measurement may also be other measurements, such as a scan measurement. The specified measurement corresponds to the "noise measurement" in this disclosure.

[0098] Figure 11 shows an example of the results obtained by a predetermined measurement. In the example in Figure 11, compound number X1, compound number X2, and compound number X3 are shown as the numbers of compounds that can be identified by MRM measurement.

[0099] In the example shown in Figure 11, the component corresponding to compound number X1 is detected by a peak in the time period 5.9–6.5. The component corresponding to compound number X2 is detected by a peak in the time period 11.75–12.05. The component corresponding to compound number X3 is detected by a peak in the time period 15.7–16.1.

[0100] Furthermore, since the sample measured in the predetermined measurement is a known sample, the first time period (the time period in which the actual peak occurs) and the second time period (the time period in which the actual noise occurs without the actual peak) are predetermined. In the example in Figure 11, the first interval corresponds to the time period from 5.9 to 6.5 (the time period corresponding to compound number X1). The second interval corresponds to the time period from 15.7 to 16.1 (the time period corresponding to compound number X3). In the first interval, the actual peak P is detected. In the second interval, the actual noise Q is detected.

[0101] Furthermore, the learning device 30 calculates the average value of the correlation coefficients of N chromatograms obtained by N predetermined measurements during the second time period. The second time period is a time period in which no peaks caused by known samples occur. If the average value is below a threshold, the learning device 30 determines that no unexpected noise has occurred. Therefore, the learning device 30 acquires the chromatogram from the second time period as pseudo-waveform noise. On the other hand, if the average value is above a threshold, the learning device 30 determines that unexpected noise has occurred. Therefore, the learning device 30 discards the pseudo-waveform noise from the chromatogram from the second time period.

[0102] Figure 12 is a flowchart of the process in step S2 of the second embodiment. The learning method in the second embodiment is represented by the flowchart in Figure 9. Furthermore, in the second embodiment, the process enclosed in parentheses in step S4 is executed.

[0103] In step S42 of Figure 12, the measurement unit 10 performs N sample measurements (measurements of known samples). Next, in step S44, the learning device 30 calculates the correlation coefficient of N chromatograms in the non-peak interval. Then, the learning device 30 performs the processing from step S24 onward.

[0104] Furthermore, once the processing in step S2 is completed, the learning device 30 executes the processing in step S4. In step S4, the learning device 30 acquires the waveform of the actual peak P in the first section as the real peak waveform (see Figure 7).

[0105] According to the second embodiment, the learning device 30 acquires the actual noise waveform Q generated in the second interval as a pseudo-noise waveform. The learning device 30 then calculates the average value of the correlation coefficients of the N actual noise waveforms Q (step S24 in Figure 12). In step S26, the learning device 30 determines whether the average value is below a threshold. If it is determined to be YES in step S26, in step S28A, the N chromatograms acquired in step S42 are acquired as pseudo-noise waveforms. On the other hand, if it is determined to be NO in step S26, the N chromatograms are discarded in step S30A. In this way, the learning device 30 determines whether or not an unexpected noise waveform is included in the pseudo-noise waveform based on the average value. Therefore, the learning device 30 can generate pseudo-noise waveforms under the same conditions as when analyzing a sample. Thus, the quality of the pseudo-noise waveforms can be improved.

[0106] Furthermore, the learning device 30 acquires the actual peak waveform P as a pseudo-peak waveform. Therefore, the learning device 30 can effectively utilize the actual peak waveform that may be acquired along with the actual noise waveform.

[0107] <Third Embodiment> In the first or second embodiment, a configuration was described in which, in step S26 of Figure 10 or Figure 12, if the average value is above a threshold (NO in step S26), the learning device 30 automatically discards N blank chromatograms. However, if the average value is above a threshold (NO in step S26), the learning device 30 may also notify the user that the degree of similarity (average value of correlation coefficients) is above a threshold. This notification can be achieved, for example, by displaying a predetermined image on the display device 65.

[0108] Figure 13 is an example of a predetermined image. Figure 13 shows an example where a predetermined image is displayed in the display area 62A of the display device 65. In the example in Figure 13, the predetermined image displayed is the phrase, "The average value of the correlation coefficient is above the threshold." By displaying this predetermined image, the learning device 30 can make the user aware that the average value of the correlation coefficient is above the threshold, that is, that the pseudo-noise waveform contains an unintended noise waveform. Therefore, the user can exclude the pseudo-noise waveform by inputting a predetermined command to the input device 61 (manually by the user). In addition, the user can visually identify the pseudo-noise waveform that contains the unintended noise.

[0109] [Pattern] Those skilled in the art will understand that the above-described exemplary embodiments are specific examples of the following embodiments.

[0110] (Section 1) A learning method relating to one aspect of the present disclosure is a learning method for learning an estimation model used to detect peaks in signal waveforms created by an analytical device that analyzes a sample. The learning method comprises acquiring noise waveforms that may occur when the analytical device performs analytical processing, and learning an estimation model based on the noise waveforms. Furthermore, acquiring noise waveforms involves acquiring multiple noise waveforms generated by noise measurements performed multiple times by the analytical device, calculating the degree of similarity of the multiple noise waveforms, and performing predetermined processing according to the degree of similarity.

[0111] With this configuration, it is possible to determine if unintended noise waveforms are included in the noise waveform and to perform processing according to the noise.

[0112] (Clause 2) The learning method described in paragraph 1, which includes performing a predetermined process, includes acquiring multiple noise waveforms whose similarity is less than a threshold. The predetermined process includes discarding multiple noise waveforms whose similarity is greater than or equal to a threshold.

[0113] This configuration allows for the automatic exclusion of unintended noise waveforms. Therefore, it is possible to suppress the degradation of the estimation model quality without burdening the user.

[0114] (Clause 3) The learning method described in paragraph 1 or 2, wherein the prescribed processing includes processing to notify the user that the degree of similarity is equal to or greater than a threshold.

[0115] This configuration allows users to recognize when the degree of similarity exceeds a certain threshold.

[0116] (Article 4) The learning method described in any one of Articles 1 to 3, wherein the noise measurement is a blank measurement performed by the analyzer when no sample is placed in the analyzer.

[0117] With this configuration, noise waveforms can be generated without consuming the sample. (Article 5) The learning method described in any one of Articles 1 to 3, wherein noise measurement is a measurement in which an analytical instrument analyzes a known sample whose components are known, thereby generating a real peak waveform in the first section of the signal waveform. Acquiring a noise waveform includes acquiring a real noise waveform generated in the second section in which no real peak is detected as a noise waveform.

[0118] With this configuration, noise waveforms can be generated under the same conditions as when analyzing a sample. Therefore, the quality of the noise waveforms can be improved.

[0119] (Clause 6) The learning method described in Clause 5, which involves learning an estimation model, includes learning an estimation model based on a real peak waveform and a noise waveform.

[0120] With this configuration, the actual peak waveform, which can be acquired along with the actual noise waveform, can be effectively utilized.

[0121] (Clause 7) The learning method described in any one of paragraphs 1 to 6, wherein the degree of similarity is the average value of the correlation coefficients of the multiple noise waveforms.

[0122] With this configuration, it is possible to determine whether or not unintended noise waveforms are included in the acquired noise waveform using a known parameter called the correlation coefficient.

[0123] (Clause 8) A learning method described in any one of Clauses 1 to 7, comprising acquiring a peak waveform that does not contain noise waveforms, and generating a signal waveform by adding a noise waveform and a peak waveform, wherein learning an estimation model includes learning an estimation model based on the signal waveform.

[0124] With this configuration, it is possible to train an estimation model under the same conditions as those generated when analyzing an unknown sample (a situation in which a signal waveform is generated).

[0125] (Section 9) The learning method described in Section 8, wherein the peak waveform is obtained by a generative adversarial network.

[0126] This configuration allows for an improvement in the quality of the peak waveform. (Clause 10) A learning program according to one embodiment is a learning program that causes a computer to update an estimation model used to detect peaks in signal waveforms created by an analytical device that analyzes a sample. The learning program causes the computer to acquire noise waveforms that may occur when the analytical device performs analytical processing, and to learn an estimation model based on the noise waveforms. Acquiring noise waveforms involves acquiring multiple noise waveforms generated by noise measurements performed multiple times by the analytical device, calculating the degree of similarity of the multiple noise waveforms, and performing predetermined processing according to the degree of similarity.

[0127] Furthermore, regarding the embodiments and modifications described above, it was intended from the outset that the configurations described in the embodiments could be appropriately combined, including combinations not mentioned in the specification, to the extent that no inconvenience or inconsistency arises.

[0128] The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims rather than by the foregoing description, and all modifications within the meaning and scope equivalent to the claims are intended to be included. [Explanation of Symbols]

[0129] 10 Measurement unit, 11 Mobile phase container, 12 Pump, 13 Injector, 14 Column, 15 Detector, 18 Arrangement unit, 19 Storage device, 20 Control unit, 21 Control device, 23 Communication interface, 25 Data analysis device, 30 Learning device, 32 Input unit, 34 Extraction unit, 35 Analysis device, 36 Execution unit, 38 Noise generation unit, 40 Update unit, 41 Generator, 42 Data selection unit, 43 Classifier, 44 Judgment unit, 45 Update processing unit, 46 Random noise generation unit, 51 Learning execution unit, 52 Storage unit, 53 Addition unit, 61 Input device, 62 Calculation unit, 62A Display area, 64 Judgment unit, 65 Display device, 100 Analysis system, 110 Data acquisition unit, 111 Peak detection processing unit, 114 Model storage unit, 116 Peak determination unit, 117 Analysis unit, 121, 121A Estimation model, 122 Control program, 130 Recording medium.

Claims

1. A learning method for training an estimation model used to detect peaks in signal waveforms created by an analytical device that analyzes a sample, The aforementioned learning method is The acquisition of noise waveforms that may occur when the aforementioned analytical device performs analytical processing, The process includes learning the estimation model based on the noise waveform, Acquiring the aforementioned noise waveform means The analysis device acquires multiple noise waveforms generated by noise measurements performed multiple times, To calculate the degree of similarity of the aforementioned multiple noise waveforms, If the degree of similarity is greater than the threshold, the first process is executed. A learning method comprising: if the degree of similarity is less than the threshold, performing a second process different from the first process.

2. The second process includes a process for acquiring the plurality of noise waveforms whose similarity is less than the threshold, The learning method according to claim 1, wherein the first process includes a process of discarding the plurality of noise waveforms whose similarity is greater than the threshold.

3. The learning method according to claim 1 or 2, wherein the first process includes a process of notifying the user that the degree of similarity is greater than the threshold.

4. The learning method according to any one of claims 1 to 3, wherein the noise measurement is a blank measurement performed by the analyzer when no sample is placed in the analyzer.

5. The noise measurement described above is a measurement in which a real peak waveform is generated in the first section of the signal waveform by analyzing a known sample whose components are known using the analytical device. The learning method according to any one of claims 1 to 3, wherein acquiring the noise waveform includes acquiring the actual noise waveform that occurs in a second interval in which the actual peak waveform is not detected as the noise waveform.

6. The learning method according to claim 5, wherein learning the estimation model includes learning the estimation model based on the actual peak waveform and the noise waveform.

7. The learning method according to any one of claims 1 to 6, wherein the degree of similarity is the average value of the correlation coefficients of each of the plurality of noise waveforms.

8. The aforementioned learning method further, To obtain a peak waveform that does not contain the aforementioned noise waveform, The system includes generating a signal waveform by adding the noise waveform and the peak waveform, The learning method according to any one of claims 1 to 7, wherein learning the estimation model includes learning the estimation model based on the signal waveform.

9. The learning method according to claim 8, wherein the peak waveform is obtained by a generative adversarial network.

10. A learning program that causes a computer to train an estimation model used to detect peaks in signal waveforms created by an analytical device that analyzes a sample, The learning program is provided to the computer, The acquisition of noise waveforms that may occur when the aforementioned analytical device performs analytical processing, Based on the noise waveform, the estimation model is trained. Acquiring the aforementioned noise waveform means The analysis device acquires multiple noise waveforms generated by noise measurements performed multiple times, To calculate the degree of similarity of the aforementioned multiple noise waveforms, If the degree of similarity is greater than the threshold, the first process is executed. A learning program comprising: if the degree of similarity is less than the threshold, executing a second process different from the first process.

Images