Waveform analyzing method and waveform analyzing device

Abstract

To correctly detect peaks having different widths included in an analysis target waveform.SOLUTION: A waveform analyzing device includes: a first trained model storage section (44) that stores a first trained model, which is constructed by machine learning in which a first window is applied to first reference waveform data and outputs a first index of a peak portion or a non-peak portion for first partial data; a second trained model storage section (44) that stores a second trained model, which is constructed by machine learning in which a second window different from the first window is applied to second reference waveform data and outputs an index of a peak portion or a non-peak portion for second partial data; a first index output processing section (55) that inputs first analysis target partial data into the first trained model to output the first index; a second index output processing section (56) that inputs second analysis target partial data into the second trained model to output the second index; and a peak portion estimation section (58) that estimates a peak portion from the first index and the second index.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method and apparatus for analyzing waveforms obtained by measuring a sample using an analytical device. [Background technology]

[0002] Liquid chromatographs and gas chromatographs are used to identify and quantify components in a sample. In a chromatograph, the components in a sample are separated using a column, and the components that flow out of the column are detected in sequence. A chromatogram is then created, with the horizontal axis representing time and the vertical axis representing detection intensity, to detect peaks, and the concentration or amount of the compound corresponding to the peak is determined from its area and height. Another widely used technique is to obtain spectral waveforms from liquid chromatographs and gas chromatographs. Spectral waveforms, with the horizontal axis representing wavelength or mass-to-charge ratio and the vertical axis representing detection intensity, are often used for substance identification.

[0003] Various methods have been put into practical use to detect peaks in chromatograms. In recent years, new peak detection methods using machine learning have been proposed and put into practical use (e.g., Patent Document 1, Non-Patent Documents 1 and 2).

[0004] Patent Document 1 describes a waveform analysis technology that constructs a trained model by performing machine learning using data from multiple reference waveforms whose peak positions are known as training data, and then uses the trained model to estimate peaks contained in data from a waveform to be analyzed. As an example, the technology describes a trained model constructed by performing machine learning on a training model using semantic segmentation, which is used in the field of image analysis, using data from multiple extracted ion chromatograms (EICs) whose peak positions are known and obtained by selected ion monitoring (SIM) or multiple reaction monitoring (MRM) measurements as training data. The trained model is input with a predetermined number of measurement data points extracted from the extracted ion chromatograms of the analysis target obtained by SIM or MRM measurements, and is then caused to output an index (label) indicating whether the measurement data points belong to a peak or a non-peak portion. The frame (extraction range) used to extract the predetermined number of measurement data points to be input to the trained model from the one-dimensional data constituting the waveform to be analyzed is called a window. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2021 / 064924 [Non-patent literature]

[0006] [Non-Patent Document 1] "Peakintelligence for GCMS LabSolutions Insight waveform processing software," [online], [searched March 14, 2024], Shimadzu Corporation, Internet <URL:https: / / www.an.shimadzu.co.jp / products / gas-chromatograph-mass-spectrometry / gc-ms-software / peakintelligence-for-gcms / index.html> [Non-patent document 2] "Peakintelligence for LCMS: Optional peak processing software for LabSolutions LCMS and LabSolutions Insight," [online], [Retrieved March 14, 2024], Shimadzu Corporation, Internet <URL: https: / / www.an.shimadzu.co.jp / products / liquid-chromatograph-mass-spectrometry / lc-ms-software / peakintelligence / index.html> [Non-patent document 3] "Reducing pesticide data analysis time with Peakintelligence for GCMS," [online], [Retrieved March 14, 2024], Shimadzu Corporation, Internet <URL: https: / / www.an.shimadzu.co.jp / sites / an.shimadzu.co.jp / files / pim / pim_document_file / an_jp / applications / application_note / 21749 / an_01-00585-jp.pdf> [Non-patent document 4] "High-speed pharmaceutical impurity analysis in accordance with the European Pharmacopoeia using Nexera-i MT," [online], [searched March 14, 2024], Shimadzu Corporation, Internet <URL: https: / / www.an.shimadzu.co.jp / sites / an.shimadzu.co.jp / files / pim / pim_document_file / an_jp / applications / application_note / 17727 / an_l518.pdf> Summary of the Invention [Problem to be solved by the invention]

[0007] When analyzing a known target component contained in a sample (target analysis), for example, a mass spectrometer is used as a detector, and SIM or MRM measurements are performed using ions generated from the target component as target ions to create an extracted ion chromatogram. In targeted analysis, peaks can be detected from a waveform within a limited time period corresponding to the retention time of any target component (e.g., a waveform including a 1.5-minute peak portion) within the entire measurement time of the chromatograph (see, for example, Non-Patent Document 3). Furthermore, SIM and MRM measurements are highly selective for the target component, resulting in narrow, sharp peaks. When detecting peaks from such waveforms, the waveform analysis technique described in Patent Document 1 can be suitably used.

[0008] On the other hand, when comprehensively analyzing unknown components contained in a sample (non-target analysis), the positions (retention times) at which peaks appear are unknown, so peaks must be detected from the waveform over the entire chromatographic measurement time (e.g., over 60 minutes). Furthermore, when a PDA detector or UV detector is used as the chromatographic detector, peaks with various widths can appear, ranging from short peaks (e.g., peak widths of about 0.5 minutes) to long peaks (e.g., peak widths of over 5 minutes) from the peak start point to the peak end point (see, for example, Non-Patent Document 4). When the present inventors applied the waveform analysis technique described in Patent Document 1 to such chromatograms, they found that peaks could not be correctly detected in some cases.

[0009] Although the example described here is one in which peaks are detected from a chromatogram obtained using a chromatograph, the same problems as those described above also exist when peaks are detected from other types of waveforms.

[0010] The problem to be solved by the present invention is to provide a technique capable of correctly detecting peaks of different widths contained in a waveform obtained by measuring a sample using an analytical device. [Means for solving the problem]

[0011] One aspect of the present invention, which has been made to solve the above-mentioned problems, is a waveform analysis method for analyzing a waveform that is composed of analysis target data acquired by measuring a sample using an analytical device, and has a first parameter on the horizontal axis and a second parameter on the vertical axis, comprising: a first trained model construction step of constructing a first trained model using, as training data, first reference waveform data constituting a first reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, where a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, is known, and applying machine learning to a first window that extracts data within a predetermined range in the horizontal axis direction from the first reference waveform data, and when first partial data corresponding to the first window is input, outputting a first index representing a peak portion or a non-peak portion for each of a plurality of first partial data elements constituting the first partial data; a second trained model construction step of constructing a second trained model using second reference waveform data as training data, the second reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, with a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, known, by machine learning that applies a second window that extracts data of a predetermined range having a width different from that of the first window in the horizontal axis direction from the second reference waveform data, and that outputs a second index representing a peak portion or a non-peak portion for each of a plurality of second partial data elements that constitute the second partial data when second partial data corresponding to the second window is input; a first index output step of extracting first analysis target partial data corresponding to the first window from the analysis target data, inputting the extracted first analysis target partial data to the first trained model, and outputting the first index for each of a plurality of first analysis target data elements constituting the first analysis target partial data; a second index output step of extracting second analysis target partial data corresponding to the second window from the analysis target data and inputting the extracted second analysis target partial data into the second trained model, and outputting the second index for each of a plurality of second analysis target data elements constituting the second analysis target partial data; a peak portion estimating step of estimating a peak portion from the analysis target data based on the first index outputted in the first index outputting step and the second index outputted in the second index outputting step; Equipped with.

[0012] Another aspect of the present invention, which has been made to solve the above-mentioned problems, is a waveform analyzer used to analyze a waveform that is composed of analysis target data acquired by measuring a sample using an analytical device, and has a first parameter on the horizontal axis and a second parameter on the vertical axis, a first trained model storage unit in which a first trained model is stored, the first trained model being constructed by machine learning applying a first window that extracts data within a predetermined range in the horizontal axis direction from the first reference waveform data, the first reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, with a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, as training data, and the first trained model being configured to output a first index representing a peak portion or a non-peak portion for each of a plurality of first partial data elements that constitute the first partial data when first partial data corresponding to the first window is input; a second trained model storage unit in which a second trained model is stored, the second trained model being constructed by machine learning that applies a second window that extracts data in a predetermined range of width different from the first window in the horizontal axis direction from the second reference waveform data, the second reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, and in which a pair of the value of the first parameter and the value of the second parameter that is the position of a peak portion is known, when second partial data corresponding to the second window is input, and outputs a second index that represents a peak portion or a non-peak portion for each of a plurality of second partial data elements that constitute the second partial data; a first index output processing unit that extracts first analysis target partial data corresponding to the first window from the analysis target data, inputs the extracted first analysis target partial data to the first trained model, and outputs the first index for each of a plurality of first analysis target data elements that constitute the first analysis target partial data; a second index output processing unit that extracts second analysis target partial data corresponding to the second window from the analysis target data, inputs the extracted second analysis target partial data to the second trained model, and outputs the second index for each of a plurality of second analysis target data constituting the second analysis target partial data; a peak portion estimation unit that estimates a peak portion from the analysis target data based on the first index output by the first index output unit and the second index output by the second index output unit; Equipped with. [Effects of the Invention]

[0013] The inventor discovered that in order to correctly detect peak portions contained in waveform data, it is necessary to analyze the entire peak for each peak and to estimate the peak portion by analyzing a sufficient number of measurement data points, and thus came up with the present invention.

[0014] Applying a first window (or a second window) to reference waveform data (or analysis target data) refers to applying a first window (or a second window) having a predetermined width in the horizontal direction to the reference waveform data (or analysis target data) and extracting partial data located within the first window (or the second window). Typically, this process is performed multiple times by moving adjacent first and second windows (sliding windows) from the start position of the reference waveform data or analysis target data toward the end position while overlapping them in the horizontal direction. Estimating peak portions based on first and second indices refers to estimating peak portions based on the fact that indices representing peak portions are output in first and second analysis target data elements positioned next to each other in the horizontal direction. By performing these estimations across the entire analysis target data constituting the waveform to be analyzed, peak and non-peak portions contained in the waveform data can be estimated.

[0015] In the present invention, when machine learning is performed using reference waveform data in which the positions of peak portions are known as training data, not only is machine learning performed using a first window that extracts data in a predetermined range in the horizontal direction to construct a first trained model, but also machine learning is performed using a second window that extracts data in a predetermined range with a width different from that of the first window to construct a second trained model. Data to be analyzed is then input to both the first trained model and the second trained model, and indicators (first indicator and second indicator) representing peak portions or non-peak portions are output from each. In this way, the present invention uses a first trained model that uses the first window and a second window with a width different from that of the first window. In this way, by using a second trained model that is suitable for detecting peaks having a width different from that of the first trained model, peaks that cannot be detected in the second window can be detected in the first window, and peaks that cannot be detected in the first window can be detected in the second window, thereby estimating the positions of peaks with different widths and correctly detecting them. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a diagram showing the configuration of a main part of a liquid chromatograph system including an embodiment of a waveform analyzer according to the present invention. [Figure 2] This figure shows the relationship between the sampling rate in a PDA detector and the half-width of the peak that is suitable for using that sampling rate as a measurement condition. [Figure 3] A diagram showing the relationship between the sampling rate and the time constant in a PDA detector. [Figure 4] 1 is a flowchart showing the procedure for creating a trained model in one embodiment of the waveform analysis method according to the present invention. [Figure 5] FIG. 10 is a diagram showing the relationship between the sampling rate, the peak half-width, the range that can be regarded as the peak portion when the peak is approximated by a Gaussian function, and the window width. [Figure 6] FIG. 10 is a diagram illustrating the state in which the first window, second window, and third window are applied to training data. [Figure 7]1 is a flowchart showing a procedure for estimating peak portions contained in unanalyzed chromatogram data in one embodiment of the waveform analysis method according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A waveform analysis method and a waveform analysis device according to embodiments of the present invention will be described below with reference to the accompanying drawings.

[0018] 1 shows the essential configuration of a liquid chromatograph system 1 including the waveform analyzer of this embodiment. The liquid chromatograph system 1 includes a liquid chromatograph unit 10 and a control and processing unit 40. A part of the control and processing unit 40 corresponds to the waveform analyzer of the present invention. Generally, a chromatographic waveform (chromatogram) obtained from a liquid chromatograph system or a gas chromatograph system is composed of data with the horizontal axis representing time and the vertical axis representing detection intensity, but the waveform to be analyzed in the present invention is not limited to a chromatogram as in this embodiment and may be a spectral waveform or the like.

[0019] The liquid chromatograph unit 10 includes a mobile phase container 11 containing a mobile phase, a liquid delivery pump 12 that delivers the mobile phase contained in the mobile phase container 11, an injector 13 that injects a liquid sample, a column 14 that separates components contained in the liquid sample, and a detector 15 that detects the components that flow out of the column 14 in sequence. The liquid chromatograph unit 10 also includes an autosampler 16 that is set with sample containers containing multiple liquid samples and introduces the multiple liquid samples into the injector 13 in the order specified in the measurement conditions. The detector 15 can be a type appropriate for the components to be detected, such as a mass spectrometer, ultraviolet absorbance detector (UV detector), photodiode array detector (PDA detector), refractive index detector (RID), or electrical conductivity detector.

[0020] The control and processing unit 40 includes a memory unit 41. The memory unit 41 includes a reference waveform data memory unit 42, a measurement data memory unit 43, and a trained model memory unit 44. The reference waveform data memory unit 42 stores measurement data (reference waveform data) that has been acquired by measurement using a detector 15 such as a mass spectrometer, an ultraviolet absorbance detector (UV detector), a photodiode array detector (PDA detector), a refractive index detector (RID), or an electrical conductivity detector, and that has undergone peak detection and the like, together with information such as the measurement conditions (including the sampling rate) and the type of detector.

[0021] Figure 2 shows an example of the relationship between the sampling rate of a PDA detector (Shimadzu SPD-M10Avp / M20A / M30A / 30AM / M40) and the half-width of the peak that can be correctly detected using that sampling rate. Correctly detecting a peak means constructing the peak using a sufficient number of measurement points to accurately represent its shape. For example, a sampling rate of 5 msec can accurately represent the shape of a peak with a half-width of 0.06 sec or more. Figure 3 also shows the relationship between the sampling rate and the time constant for some models (SPD-M30A, SPD-M40).

[0022] Analysis results are often provided to users as waveforms displayed on a screen. The waveforms provided to users are two-dimensional figures, and the data that constitute them is numerical information obtained by digitally converting detector signals. For example, reference waveform data used in machine learning is typically two-dimensional data in which the output signal values ​​from a detector are arranged in time series. Alternatively, the time series information, i.e., the sampling interval, is known information, and the reference waveform can be reproduced even if the time series information is removed. Therefore, reference waveform data may be one-dimensional data. As long as the sampling time interval is known, it is possible to restore time series data by arranging the data by the sampling time interval. In many cases, the sampling time interval is included in the measurement conditions and is therefore known information. Reference waveform data is used in machine learning and is data in which the data elements that constitute the peak portions contained therein are identified. Reference waveform data includes multiple chromatogram data acquired under the same measurement conditions and using the same type of analytical instrument and detector. This reference waveform data may be data previously measured using the liquid chromatograph system 1 of this embodiment, or it may be data acquired from a database that stores data measured using an analytical instrument other than the liquid chromatograph system 1 of this embodiment. It is preferable that the reference waveform data used in machine learning and the analysis target data to be analyzed have the same data structure. Furthermore, the measurement data storage unit 43 may store measurement conditions used to measure various compounds. Furthermore, the measurement data storage unit 43 sequentially stores chromatogram data acquired by the liquid chromatograph unit 10. The trained model storage unit 44 stores a first trained model, a second trained model, and a third trained model created by the trained model creation unit 51 (described later).

[0023] The control and processing unit 40 includes, as functional blocks, a trained model creation unit 51, a measurement condition setting unit 52, a measurement execution unit 53, a window setting unit 54, a first index output processing unit 55, a second index output processing unit 56, a third index output processing unit 57, a peak portion estimation unit 58, and an analysis result output unit 59. Note that the trained model creation unit 51 may be removed from the liquid chromatograph system 1 to create a package for a customer. In this case, before delivery to the customer, the developer creates a trained model using the trained model creation unit 51, and the trained model is stored in the trained model storage 44. In this case, the trained model creation unit 51 may be removed from the liquid chromatograph system 1. The control and processing unit 40 is actually a general-purpose personal computer, and the above-mentioned functional blocks are realized by executing a pre-installed waveform analysis program on the computer's processor. The control and processing unit 40 is also connected to an input unit 6 consisting of a keyboard, mouse, etc., and a display unit 7 consisting of an LCD display, etc.

[0024] Next, we will explain a method for analyzing a chromatogram using the chromatography mass spectrometry system of this embodiment. In the chromatography mass spectrometry system of this embodiment, when the waveform analysis program is executed, a screen is displayed on the display unit 7, which allows the user to select either creating a trained model or analyzing chromatogram data.

[0025] First, the procedure for creating a trained model will be described with reference to the flowchart in Figure 4.

[0026] When creation of a trained model is selected, the trained model creation unit 51 prepares an untrained trained model (step 1). Various trained models capable of performing semantic segmentation can be suitably used for this trained model. Semantic segmentation is generally used to analyze images composed of pixel data distributed two-dimensionally, but in this embodiment, it is applied to the analysis of chromatogram waveform data composed of multiple data elements acquired at a predetermined sampling interval. Examples of trained models capable of performing semantic segmentation include U-Net, SeGNet, and PSPNet (see, for example, Patent Document 1). In this embodiment, U-Net is used.

[0027] Next, the trained model creation unit 51 displays a screen on the display unit 7 that prompts the user to specify the type of training data to be read from the reference waveform data storage unit 42. This screen may, for example, prompt the user to select the type of detector in a pull-down menu. As described above, the reference waveform data storage unit 42 stores reference waveform data that has been acquired by measurements using a mass spectrometer, ultraviolet absorbance detector (UV detector), photodiode array detector (PDA detector), refractive index detector (RID), electrical conductivity detector, etc., and that has undergone peak detection, etc., along with information on the type of detector.

[0028] When the type of detector is selected (step 2), the trained model creation unit 51 reads out a plurality of pieces of reference waveform data acquired using the selected detector from the reference waveform data storage unit 42. When reading out the reference waveform data, information on the sampling rate may also be read out from the reference waveform data storage unit 42.

[0029] The number of data points input to the learning model is arbitrary, but inputting a large number of data points requires time for processing. Therefore, it is desirable to adopt a number of data points appropriate for the hardware power (processing capability). On the other hand, if the number of data points is too small, machine learning will be performed based on information that is unable to reproduce the waveform to be analyzed with sufficient accuracy, as explained by the sampling theorem, and a learned model that estimates peak portions will be constructed. In this embodiment, the number of data points input to U-Net is 1,024. Then, 1,024 measurement data elements extracted at equal intervals from the beginning (the origin side of the measurement time on the horizontal axis; the same applies below) of the reference waveform data are input to U-Net as a set to train the learning model. The frame (range) within which this set of measurement data is extracted from the reference waveform data is called a window. Therefore, if the window width is narrow along the time axis, 1,024 data elements are extracted at short time intervals within a narrow time range. On the other hand, if the window width is wide along the time axis, 1,024 data elements are extracted at long time intervals within a long time range. From the perspective of waveform reproducibility, the shorter the time interval between data elements, the better. However, the inventors have found that it is appropriate to set the width of the first window, which has the narrowest width in the time axis direction, to a width equivalent to 1,024 times the detector's sampling rate. The reason for this is that even if the number of data elements input to the learning model is set to a value greater than the detector's sampling rate (even if the time interval between data elements is shorter than the detector's sampling rate), it is not possible to obtain data from the detector with a number of points greater than the detector's sampling performance. Figure 5 shows the relationship between the sampling rate, peak half-width, the range (±3σ) that can be considered a peak when the peak is approximated by a Gaussian function, and the window length (sampling interval × 1,024). Thus, it is desirable to determine the minimum window size, i.e., the minimum time interval between data elements input to the learning model, based on the detector's sampling rate. Alternatively, a window size and minimum interval between input points appropriate for the detector may be stored in a memory unit.

[0030] As shown in the upper part of Figure 6, the trained model creation unit 51 performs machine learning on the entire range of the reference waveform data by moving the position of the first window while overlapping the window range by a predetermined width in the time axis direction (sliding window). When using sliding windowing, it is recommended to move the window so that 1 / 3 to 1 / 2 of the window width overlaps between adjacent windows. This allows almost all peaks appearing on the chromatogram to be contained entirely within one of the windows. Although it is possible to distinguish peak and non-peak portions even if a portion of a peak is located outside the window, configuring the peak to be located entirely within the window improves the accuracy of peak identification. For ease of understanding, Figure 6 shows a two-dimensional graph with time on the horizontal axis and detection intensity on the vertical axis, in which a window has been applied to the reference waveform. However, the reference waveform data constituting the reference waveform may be intensity information arranged in a time series in the order of sampling, and may be one-dimensional data that does not include the sampling order or the time series data itself (i.e., does not have values ​​in the time axis direction). In this case, the window is also one-dimensional, i.e., it defines only the length (number of data points) without any information about the vertical or horizontal axis. Also, depending on the window width selected, it is possible to include the entire reference waveform data in one window, as in the third window described below.

[0031] In this way, machine learning applying the first window is performed on all reference waveform data (step 3). By performing this type of machine learning, a trained model is created that, when measurement data is input, outputs labels (indicators) representing the attributes of each data element constituting the measurement data. The output labels can be, for example, peak start point, peak end point, single peak, tailing-processed peak, fully separated peak, vertically split peak, and non-peak portion, as described in Patent Document 1. Labels other than non-peak portion are assigned to peak portions. Furthermore, tailing-processed peak, fully separated peak, and vertically split peak are assigned to portions where multiple peaks overlap (superimposed peak portion), and a method suitable for separating these multiple peaks is output as a label. Note that these labels are examples of preferred embodiments, and in the present invention, it is sufficient to output labels (indicators) representing at least peak portions and non-peak portions. Note that the procedure for performing machine learning and the content of the labels themselves are the same as those in Patent Document 1, and therefore detailed description will be omitted here.

[0032] The trained model creation unit 51 then performs the same machine learning as above, applying a second window with a width different from that of the first window, to all of the reference waveform data (step 4). Note that in this embodiment, since the minimum window width in the time axis direction is set for the first window, it is sufficient to specify a width for the second window that is wider than that of the first window.

[0033] In a liquid chromatograph, various types of detectors are used depending on the components to be detected, such as a mass spectrometer, an ultraviolet absorbance detector (UV detector), a photodiode array detector (PDA detector), a refractive index detector (RID), and an electrical conductivity detector, as described above. The shape and width of the peaks that appear in a chromatogram vary depending on the type of detector. Taking advantage of the fact that there is a tendency in the shape of the peaks detected by each detector, in this embodiment, the width of the second window is determined for each type of detector so that the widest peak among the peaks expected for each detector is completely contained in one window.

[0034] For example, when the detector is a mass spectrometer, the expected peak width is approximately 1.5 minutes at most, whereas when a PDA detector or UV detector is used, peaks with widths of 5 minutes or 10 minutes may appear. Therefore, as described above, the width of the second window is predetermined depending on the type of detector. For example, when the detector is a mass spectrometer, the width of the second window is predetermined to 3 minutes, and when the detector is a PDA detector or UV detector, the width of the second window is predetermined to 15 minutes. The width of the second window is, for example, 1.5 to 2 times the maximum expected peak width. When performing sliding window analysis, the window is moved so that 1 / 3 to 1 / 2 of the window width overlaps between adjacent windows, as shown schematically in the middle of Figure 6.

[0035] When the second window is used, the measurement data present within the second window is divided into 1,024 points at equal intervals and input to the learning model. Note that this is just one example, and the present invention is not limited to this. When the second window is used, the number of measurement data points present within the second window may be adjusted to 1,024 points by summing, averaging, or thinning out multiple pieces of measurement data, before being input to U-Net and machine learning is performed.

[0036] The trained model creation unit 51 further performs the same machine learning as above, applying a third window having a width equivalent to the entire measurement time period, to all the reference waveform data (step 5).

[0037] Liquid chromatographs sometimes perform gradient analysis, in which the mixing ratio of multiple mobile phases is gradually changed during measurement. Gradient analysis often results in a gradual increase (or decrease) in the baseline over the entire measurement time period, known as drift. Machine learning of only a portion of the reference waveform data, such as the first or second window, makes it difficult to obtain a trained model that can correctly identify such drifts and peaks. Therefore, in this embodiment, as shown schematically in the lower part of Figure 6, machine learning is performed using a single third window that corresponds to the entire measurement time period. Even when using the third window, the number of measurement data points within the third window exceeds the number of data points input to U-Net. Therefore, when using the third window, the number of measurement data points within the third window is adjusted to 1,024 by, for example, increasing the interval between input points, averaging multiple measurement data, or thinning out the measurement data, before inputting the data into U-Net and performing machine learning.

[0038] By performing the above processing, the trained model creation unit 51 stores a first trained model using a first window determined according to the sampling interval and a second trained model using a second window determined according to the type of detector in the trained model storage unit 44. Furthermore, a third trained model using a third window having a width equivalent to the entire measurement time is constructed and stored in the trained model storage unit 44 together with information on the corresponding detector type. Note that if the second window has a width equivalent to the entire measurement time period, the second window and the third window will be the same, and therefore it is not necessary to construct and store a third trained model.

[0039] Next, the procedure for analyzing the waveform of an unanalyzed chromatogram will be described with reference to the flowchart of FIG.

[0040] When the user places a sample in the autosampler 16 and instructs the start of analysis, the measurement condition setting unit 52 reads out the measurement conditions stored in the measurement data storage unit 43 and displays them on the screen of the display unit 7. These measurement conditions include the type of detector to be used for the measurement and information on its sampling rate. When the user selects the measurement conditions to be used from the displayed measurement conditions (or makes appropriate changes) and instructs the start of measurement, the measurement condition setting unit 52 creates a batch file that executes the measurement under the selected conditions and saves it in the measurement data storage unit 43.

[0041] When the user instructs the execution of a measurement, the measurement execution unit 53 executes a batch file stored in the measurement data storage unit 43 to perform chromatographic analysis of the sample, acquires measurement data constituting a chromatogram, and stores the acquired measurement data in the measurement data storage unit 43. Like the reference waveform data, this measurement data is data in which output signals from a detector are arranged in time series, and corresponds to the analysis target data in the present invention. Here, an example has been described in which a chromatogram is acquired by measuring a sample using the measurement execution unit 53, but chromatogram data may also be acquired by, for example, reading in chromatogram data acquired in advance.

[0042] After acquiring chromatogram data by measuring a sample or reading acquired data (step 11), when the user issues an instruction to analyze the chromatogram data, the window setting unit 54 creates a chromatogram from the read data and displays it on the screen of the display unit 7 (step 12). The window setting unit 54 also determines the widths of the first, second, and third windows based on the sampling rate, detector type, and total measurement time described in the measurement conditions, and displays these values ​​on the display unit 7. The width of the first window is the sampling rate x 1,024, the width of the second window is a value associated with the detector type, and the width of the third window is the total measurement time. The user checks the values ​​of each window displayed on the display unit 7 and determines these values ​​by performing a predetermined input operation (step 13). It is desirable that the window size used for estimation be the same as the window size used for machine learning, but the present invention is not limited thereto. For example, even if the window size used for reference waveform data during machine learning differs from the window size used for analysis waveform data during estimation, a slight difference in size has little effect on estimation accuracy, and this effect can be further reduced by performing normalization processing in advance.

[0043] Once the width of each window is determined, the first index output processor 55 reads 1,024 measurement data points from the beginning of the chromatogram data and inputs them into the first trained model. Again, the windows are moved while overlapping adjacent windows by 1 / 3 to 1 / 2 of their width. The first trained model outputs a label for each input chromatogram data element: peak start point, peak end point, single peak, tailing peak, fully separated peak, vertically split peak, or non-peak portion (step 14). In this embodiment, a label is output for each data element arranged along the time axis. Note that, in this embodiment, one-dimensional data in which only detection intensities are arranged in chronological order is used as input data, but time-series information corresponding to each detection intensity (information on the time when data for each detection intensity was acquired) is restored and assigned a label. The output label corresponds to the first index representing a peak portion or a non-peak portion in this invention. More specifically, only the label for the non-peak portion corresponds to an index representing a non-peak portion, and all other labels correspond to indexes representing peak portions. The position of the first window is shifted so that the ranges of adjacent windows overlap, 1,024 data elements are input, and a label is output for each data element. This process is carried out over the entire measurement range. In this way, one or more labels are output for all measurement data elements (multiple labels are output for measurement data located in the overlapping window area).

[0044] Next, the second index output processor 56 performs a process to reduce the number of data points contained within the chromatogram when the second window is applied to the chromatogram to 1,024 points. Specifically, similar to when the second window is applied to the training data, the processor 56 widens the time interval between data points extracted from multiple measurement data, sums them, averages them, or thins out the measurement data. The second index output processor 56 then reads 1,024 measurement data points from the beginning of the chromatogram data and inputs them into the second trained model. The second trained model outputs one of the following labels for each input measurement data element: peak start point, peak end point, single peak, tailing peak, fully separated peak, vertically split peak, or non-peak portion. The processor 56 shifts the position of the second window so that the ranges of adjacent windows overlap, inputs 1,024 data elements, and outputs a label for each data element. This process is performed across the entire measurement range. In this way, one or more labels are output for all measurement data elements (step 15). Here too, multiple labels are output for the measurement data located in the overlapping portion of the windows. In this embodiment, the width of the first window is narrower than the width of the second window, so the first window (i.e., the first trained model) can detect narrow peaks that would be missed by the second window (i.e., the second trained model), and conversely, the second window can accurately detect wide peaks that do not fit into the first window and for which the first window has low detection accuracy.

[0045] Furthermore, the third index output processing unit 57 performs a process to reduce the total number of measurement points to 1,024 points. Specifically, similar to when the third window is applied to the training data, the third index output processing unit 57 widens the time interval between data points extracted from multiple measurement data, sums them, averages them, or thins out the measurement data. The third index output processing unit 57 then inputs the 1,024 measurement data points into the third trained model. The third trained model outputs one of the following labels for each input measurement data element: peak start point, peak end point, single peak, tailing-processed peak, fully separated peak, vertically split peak, or non-peak portion. In this way, one label is output for each measurement data element (step 16).

[0046] When processing is complete, applying all windows to the chromatogram data to be analyzed, the peak portion estimation unit 58 determines a label for each measurement data element. If multiple labels are output for the same measurement data element (measurement point), they are integrated. Then, a peak portion is estimated based on the label for each measurement data element (step 17). If different labels are output for the same measurement data element, a label for the measurement data element (measurement point) is determined based on a predetermined priority. Specifically, for example, if labels for peak portions and non-peak portions are output, the peak portion is prioritized. Furthermore, regarding single peaks and overlapping peaks (tailing-processed peaks, fully separated peaks, and vertically separated peaks), overlapping peaks are prioritized. This makes it possible to avoid overlooking the existence of a peak or erroneously estimating that an overlapping peak that requires peak separation is a single peak.

[0047] Finally, the analysis result output unit 59 displays the analysis results (labels for each measurement data element) together with the chromatogram of the analysis target on the display unit 7 (step 18), thereby enabling the user to confirm the peaks estimated to be present on the chromatogram of the analysis target.

[0048] When analyzing a known target component contained in a sample (target analysis), for example, a mass spectrometer is used as a detector, and SIM or MRM measurements are performed using ions generated from the target component as target ions to create an extracted ion chromatogram. In targeted analysis, peaks are detected from the waveform during a limited time period (e.g., 1.5 minutes) corresponding to the retention time of the target component out of the total measurement time of the chromatograph (see, for example, Non-Patent Document 3). Furthermore, SIM and MRM measurements are highly selective for the target component, resulting in narrow, sharp peaks. The waveform analysis technology described in Patent Document 1 was developed with the detection of peaks from such waveforms in mind.

[0049] On the other hand, when comprehensively analyzing unknown components contained in a sample (non-target analysis), the positions (retention times) at which peaks appear are unknown, so peaks must be detected from waveform data for the entire chromatographic measurement time (e.g., over 60 minutes). Furthermore, when a PDA detector or UV detector is used as the chromatographic detector, peaks with various widths from the peak start point to the peak end point can appear, ranging from short ones (e.g., peak widths of about 0.5 minutes) to long ones (e.g., peak widths of over 5 minutes) (see, for example, Non-Patent Document 4).

[0050] In this case, if a window is set with a sufficient number of measurement points allocated to a narrow peak, the narrow peak can be correctly detected, but the entire wide peak cannot fit within a single window, making it difficult to correctly detect the peak portion using a trained model. On the other hand, if a window is set with a sufficient number of measurement points allocated to a wide peak, the wide peak can be correctly detected, but the input data corresponding to the narrow peak will be small (for example, only one or two points), making it impossible to detect the narrow peak. As mentioned above, the number of sampling points for the window is fixed (1024 points in this embodiment), and a wide window width in the axial direction is equivalent to a wide time interval between the extracted data elements. Therefore, a narrow peak may fall between two adjacent data points, and in the worst case, not even one point may be detected. Furthermore, because none of these trained models covered the entire measurement range in a single window, they were unable to correctly capture trends in baseline fluctuations across the entire measurement range, resulting in baseline fluctuations being mistaken for peaks and peaks being mistaken for the baseline (non-peak portion).

[0051] One possible solution to the above problem is to expand tiny peaks in the time direction (and intensity direction) to create measurement data that represents pseudo-wide peaks, and then train a learning model on this data. However, in this case, the machine learning would involve training data in which not only the peaks but also the noise level are stretched in the time direction. Because such time-stretched noise is not detected in actual measurements, this type of machine learning would result in a trained model that has learned waveforms that do not appear in actual measurement data. As a result, such a trained model would be unable to correctly distinguish between non-peak portions (noise portions) and peak portions contained in the data to be analyzed obtained in actual measurements.

[0052] In contrast, in this embodiment, as described above, a first trained model, a second trained model, and a third trained model are created by machine learning using three windows of different widths. This makes it possible to use the first trained model, which can correctly detect narrow peaks, and the second trained model, which can correctly detect wide peaks. Furthermore, the third trained model, which uses the entire measurement range as one window, can correctly detect a baseline that fluctuates over the entire measurement time range and correctly distinguish between baseline fluctuations and peaks.

[0053] The above embodiment is merely an example and can be modified as appropriate in accordance with the spirit of the present invention.

[0054] In the above embodiment, three trained models were created by machine learning using three windows of different widths, but the number of window types and trained models may be two, four, or more.

[0055] In the above embodiment, the data to be analyzed is chromatogram data obtained by measurement using a liquid chromatograph. However, the waveform analysis method and waveform analysis device according to the present invention can be used to analyze various types of data. For example, chromatogram data obtained by measurement using a gas chromatograph and measurement data other than chromatogram data can also be analyzed in a similar manner. Furthermore, for example, an optical spectrum (a waveform representing the change in detected intensity on the wavelength or wavenumber axis) obtained by measurement using a spectrophotometer and a mass spectrum obtained by measurement using a mass analyzer can also be analyzed in a similar manner.

[0056] In the above embodiment, the second trained model was constructed by applying a second window with a width previously associated with the type of detector to training data and performing machine learning. However, multiple trained models may be created by applying windows of different widths to training data for a single detector, and information regarding the window width (and detector type) may be associated and stored in the trained model storage unit 44. In this case, the user can change the width of the second window by checking the shape of the chromatogram created from the chromatogram data to be analyzed, displayed on the display unit 7. When the user changes the width of the second window, the second indicator output processing unit 56 reads out the second trained model corresponding to the changed width of the second window from the trained model storage unit 44 and outputs a label for each measurement data element in the same manner as described above. Furthermore, if there is no fluctuation, such as drift, in the baseline of the chromatogram to be analyzed over the entire measurement time, the peak portion may be estimated without using the third window.

[0057] Alternatively, in the above configuration, instead of changing the window width, the user may input the value of the expected peak width. In that case, the value obtained by multiplying the input peak width by a predetermined constant (e.g., 1.5 or 2) is set as the width of the second window, and processing is performed in the same manner as above.

[0058] Furthermore, in the above embodiment, U-Net was used for the first trained model, the second trained model, and the third trained model, but different trained models may be used for each trained model. Neural networks can be suitably used for these trained models, including various architectures such as those that perform semantic segmentation, those that perform object detection (SSD), those that use a stochastic model, recurrent neural networks (RNN), and transformers. Since each of these multiple types of architectures has its own strengths and weaknesses, the accuracy of peak detection can be further improved by constructing trained models by appropriately using them.

[0059] In the above embodiment, one label is output for each piece of measurement data, and the analysis results are displayed on the display unit 7. However, depending on the architecture constituting the trained model, it may be possible to output multiple labels and the accuracy of each label as an inference result for one measurement data element. The U-Net described in the above embodiment is one such architecture. In the above embodiment, only the most accurate label is output. However, when using such a trained model, other labels and their accuracy may be displayed on the display unit 7 in addition to the most accurate label. This allows the user to more accurately estimate the peak, even if the label determined to be most accurate is incorrect, by changing it to the label with the next highest accuracy.

[0060] [Aspect] It will be apparent to those skilled in the art that the above-described exemplary embodiments are examples of the following aspects.

[0061] (Section 1) One aspect of the present invention is a waveform analysis method for analyzing a waveform that is composed of analysis target data acquired by measuring a sample using an analytical device, the waveform having a first parameter on the horizontal axis and a second parameter on the vertical axis, the method comprising: a first trained model construction step of constructing a first trained model using, as training data, first reference waveform data constituting a first reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, where a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, is known, and applying machine learning to a first window that extracts data within a predetermined range in the horizontal axis direction from the first reference waveform data, and when first partial data corresponding to the first window is input, outputting a first index representing a peak portion or a non-peak portion for each of a plurality of first partial data elements constituting the first partial data; a second trained model construction step of constructing a second trained model using second reference waveform data as training data, the second reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, with a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, known, by machine learning that applies a second window that extracts data of a predetermined range having a width different from that of the first window in the horizontal axis direction from the second reference waveform data, and that outputs a second index representing a peak portion or a non-peak portion for each of a plurality of second partial data elements that constitute the second partial data when second partial data corresponding to the second window is input; a first index output step of extracting first analysis target partial data corresponding to the first window from the analysis target data, inputting the extracted first analysis target partial data to the first trained model, and outputting the first index for each of a plurality of first analysis target data elements constituting the first analysis target partial data; a second index output step of extracting second analysis target partial data corresponding to the second window from the analysis target data and inputting the extracted second analysis target partial data into the second trained model, and outputting the second index for each of a plurality of second analysis target data elements constituting the second analysis target partial data; a peak portion estimating step of estimating a peak portion from the analysis target data based on the first index outputted in the first index outputting step and the second index outputted in the second index outputting step; Equipped with.

[0062] (Section 2) Another aspect of the present invention is a waveform analysis device used to analyze a waveform having a first parameter on the horizontal axis and a second parameter on the vertical axis, the waveform being composed of analysis target data acquired by measuring a sample using an analytical device, the waveform analysis device comprising: a first trained model storage unit in which a first trained model is stored, the first trained model being constructed by machine learning applying a first window that extracts data within a predetermined range in the horizontal axis direction from the first reference waveform data, the first reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, with a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, as training data, and the first trained model being configured to output a first index representing a peak portion or a non-peak portion for each of a plurality of first partial data elements that constitute the first partial data when first partial data corresponding to the first window is input; a second trained model storage unit in which a second trained model is stored, the second trained model being constructed by machine learning that applies a second window that extracts data in a predetermined range of width different from the first window in the horizontal axis direction from the second reference waveform data, the second reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, and in which a pair of the value of the first parameter and the value of the second parameter that is the position of a peak portion is known, when second partial data corresponding to the second window is input, and outputs a second index that represents a peak portion or a non-peak portion for each of a plurality of second partial data elements that constitute the second partial data; a first index output processing unit that extracts first analysis target partial data corresponding to the first window from the analysis target data, inputs the extracted first analysis target partial data to the first trained model, and outputs the first index for each of a plurality of first analysis target data elements that constitute the first analysis target partial data; a second index output processing unit that extracts second analysis target partial data corresponding to the second window from the analysis target data, inputs the extracted second analysis target partial data to the second trained model, and outputs the second index for each of a plurality of second analysis target data elements that constitute the second analysis target partial data; a peak portion estimation unit that estimates a peak portion from the analysis target data based on the first index output by the first index output unit and the second index output by the second index output unit; Equipped with.

[0063] The inventors have discovered that in order to correctly detect peak portions contained in waveform data, it is necessary to analyze the entire peak for each peak and to analyze a sufficient number of measurement data points to estimate the peak portion.

[0064] Applying a first window (or a second window) to reference waveform data (or analysis target data) refers to applying a first window (or a second window) having a predetermined width in the horizontal direction to the reference waveform data (or analysis target data) and extracting partial data located within the first window (or the second window). Typically, this process is performed multiple times by moving adjacent first and second windows (sliding windows) from the start position of the reference waveform data or analysis target data toward the end position while overlapping them in the horizontal direction. Estimating peak portions based on first and second indices refers to estimating peak portions based on the fact that indices representing peak portions are output in first and second analysis target data elements positioned next to each other in the horizontal direction. By performing these estimations across the entire analysis target data constituting the waveform to be analyzed, peak and non-peak portions contained in the waveform data can be estimated.

[0065] In the waveform analysis method according to paragraph 1 and the waveform analysis device according to paragraph 2, when performing machine learning using reference waveform data in which the positions of peak portions are known as training data, not only is machine learning performed using a first window that extracts data in a predetermined range in the horizontal direction to construct a first trained model, but also machine learning is performed using a second window that extracts data in a predetermined range with a width different from that of the first window to construct a second trained model. The first reference waveform data and the second reference waveform data may be the same or different, but it is preferable that the first reference waveform data include narrow peaks and the second reference waveform data include wide peaks. Then, data to be analyzed is input to both the first trained model and the second trained model, and indices (first and second indices) representing peak portions or non-peak portions are output from each. In the waveform analysis method according to paragraph 1 and the waveform analysis device according to paragraph 2, by using the first trained model that uses the first window and the second window that has a width different from that of the first window, peaks of different widths can be correctly detected.

[0066] (Section 3) The waveform analyzer according to paragraph 3 is the waveform analyzer according to paragraph 2, further comprising: The analysis target data is associated with information on a sampling rate in a measurement in which the analysis target data was acquired, The width of the first window is determined based on the sampling rate.

[0067] In the waveform analyzer according to the third aspect, the peak portion can be estimated from the analysis target data by directly using each measurement data element that constitutes the measurement data.

[0068] (Section 4) The waveform analyzer according to paragraph 4 is a waveform analyzer according to paragraph 2 or 3, the analysis target data is associated with information on the type of detector used in the measurement to acquire the analysis target data; The width of the second window is predetermined depending on the type of the detector.

[0069] The width of the peaks that appear in the measurement data varies depending on the type of detector used in measuring the sample. In the waveform analyzer according to paragraph 4, the width of the second window is determined depending on the type of detector used when acquiring the data to be analyzed, thereby making it possible to correctly detect the peaks contained in the data to be analyzed.

[0070] (Section 5) The waveform analyzer according to paragraph 5 is a waveform analyzer according to paragraph 2 or 3, The second window extracts the entire data to be analyzed.

[0071] When performing gradient analysis in a liquid chromatograph or temperature-programmed analysis in a gas chromatograph, a gradual increase (or decrease) in the baseline over the entire measurement time period, known as drift, occurs. It is difficult to obtain a trained model that can correctly distinguish between such drifts and peaks by simply performing machine learning on a portion of the reference waveform data. The waveform analysis device according to paragraph 5 constructs a trained model by performing machine learning using a window that corresponds to the entire measurement time period in which the data to be analyzed was acquired, thereby enabling the correct discrimination between drifts and peaks.

[0072] (Section 6) The waveform analyzer according to paragraph 6 is a waveform analyzer according to any one of paragraphs 2 to 5, The first trained model and the second trained model are configured with different architectures.

[0073] In the waveform analysis device according to any one of paragraphs 2 to 5, a neural network can be suitably used to construct a training model, including various architectures such as those that perform semantic segmentation, those that perform object detection (SSD), those that use a stochastic model, recurrent neural networks (RNN), transformers, etc. In the waveform analysis device according to paragraph 6, the accuracy of peak detection can be further improved by constructing a trained model by appropriately using these multiple types of architectures.

[0074] (Section 7) The waveform analyzer according to paragraph 7 is a waveform analyzer according to any one of paragraphs 2 to 6, When the index output from the first index output step and the index output from the second index output step are different for the same measurement data element, the peak portion estimation unit prioritizes the index representing the peak portion and estimates the peak portion from the analysis target data.

[0075] In the waveform analyzer according to paragraph 7, even if an index representing a peak portion and an index representing a non-peak portion are output for the same measurement data element, the peak portion is given priority, so that all peaks contained in the data to be analyzed can be detected without omission. [Explanation of symbols]

[0076] 1...Liquid chromatograph system 10...Liquid chromatograph section 11...Mobile phase container 12...Liquid transfer pump 13...Injector 14...Column 15...Detector 16...Autosampler 40...Control and processing section 41...Storage section 42...Reference waveform data storage section 43...Measurement data storage unit 44…Trained model memory section 51...Trained model creation unit 52...Measurement condition setting section 53...Measurement execution unit 54...Window settings section 55...First index output processing unit 56...Second index output processing unit 57...Third index output processing unit 58...Peak part estimation section 59...Analysis result output section 6...Input section 7...Display section

Claims

1. A waveform analysis method for analyzing a waveform that is composed of analysis target data acquired by measuring a sample using an analytical device, the waveform having a first parameter on the horizontal axis and a second parameter on the vertical axis, comprising: a first trained model construction step of constructing a first trained model using, as training data, first reference waveform data constituting a first reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, where a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, is known, and applying machine learning to a first window that extracts data within a predetermined range in the horizontal axis direction from the first reference waveform data, and when first partial data corresponding to the first window is input, outputting a first index representing a peak portion or a non-peak portion for each of a plurality of first partial data elements constituting the first partial data; a second trained model construction step of constructing a second trained model using second reference waveform data as training data, the second reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, with a pair of the value of the first parameter and the value of the second parameter being known, and applying a second window that extracts data in a predetermined range in the horizontal axis direction from the second reference waveform data, the second window having a width different from that of the first window, when second partial data corresponding to the second window is input, to output a second index representing a peak portion or a non-peak portion for each of a plurality of second partial data elements that constitute the second partial data; a first index output step of extracting first analysis target partial data corresponding to the first window from the analysis target data, inputting the extracted first analysis target partial data to the first trained model, and outputting the first index for each of a plurality of first analysis target data elements constituting the first analysis target partial data; a second index output step of extracting second analysis target partial data corresponding to the second window from the analysis target data, inputting the extracted second analysis target partial data to the second trained model, and outputting the second index for each of a plurality of second analysis target data elements constituting the second analysis target partial data; a peak portion estimating step of estimating a peak portion from the analysis target data based on the first index outputted in the first index outputting step and the second index outputted in the second index outputting step; A waveform analysis method comprising:

2. A waveform analyzer used to analyze a waveform having a first parameter on the horizontal axis and a second parameter on the vertical axis, the waveform being composed of analysis target data acquired by measuring a sample using an analytical device, a first trained model storage unit in which a first trained model is stored, the first trained model being constructed by machine learning applying a first window that extracts data within a predetermined range in the horizontal axis direction from the first reference waveform data, the first reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, with a pair of the value of the first parameter and the value of the second parameter, which is the position of a peak portion, as training data, and the first trained model being configured to output a first index representing a peak portion or a non-peak portion for each of a plurality of first partial data elements that constitute the first partial data when first partial data corresponding to the first window is input; a second trained model storage unit in which a second trained model is stored, the second trained model being constructed by machine learning applying a second window that extracts data in a predetermined range of width different from the first window in the horizontal axis direction from the second reference waveform data, the second reference waveform having the first parameter on the horizontal axis and the second parameter on the vertical axis, and a pair of the value of the first parameter and the value of the second parameter that is the position of a peak portion, as training data, and that outputs a second index representing a peak portion or a non-peak portion for each of a plurality of second partial data elements that constitute the second partial data when second partial data corresponding to the second window is input; a first index output processing unit that extracts first analysis target partial data corresponding to the first window from the analysis target data, inputs the extracted first analysis target partial data to the first trained model, and outputs the first index for each of a plurality of first analysis target data elements that constitute the first analysis target partial data; extracting second analysis target partial data corresponding to the second window from the analysis target data and inputting the second analysis target partial data to the second trained model, and providing the second index output processing unit for each of a plurality of second analysis target data elements constituting the second analysis target partial data; a peak portion estimation unit that estimates a peak portion from the analysis target data based on the first index output by the first index output unit and the second index output by the second index output unit; A waveform analysis device comprising:

3. The analysis target data is associated with information on a sampling rate in a measurement in which the analysis target data was acquired, 3. The waveform analyzer according to claim 2, wherein the width of the first window is determined based on the sampling rate.

4. the analysis target data is associated with information on the type of detector used in the measurement to acquire the analysis target data; 3. The waveform analyzer according to claim 2, wherein the width of the second window is predetermined depending on the type of the detector.

5. The waveform analyzer according to claim 2 , wherein the second window extracts the entire analysis target data.

6. The waveform analysis device according to claim 2 , wherein the first trained model and the second trained model are configured with different architectures.

7. 3. The waveform analysis device according to claim 2, wherein when the index output from the first index output step and the index output from the second index output step are different for the same measurement data element, the peak portion estimation unit prioritizes the index representing the peak portion and estimates the peak portion from the analysis target data.

Images