Data generation method and device, and recognizer generation method and device

Abstract

One aspect of the data generation device of the present application is a data generation device that analogically generates data used when creating an identifier for detecting a peak observed in a signal waveform by machine learning, including: a parameter frequency information acquisition section (31, 32) that obtains frequency information of a predetermined shape parameter that characterizes the shape of a plurality of signal waveforms collected by reducing a target field for creating the machine learning of the identifier; and a simulated waveform generation section that generates a simulated signal waveform that can include overlap of a plurality of peaks and noise using the frequency information of the shape parameter, and provides the simulated signal waveform as data for training or evaluation of the machine learning. Thus, high-precision teacher data can be analogically generated.

Description

Technical Field

[0001] The present invention relates to a data generation method and apparatus for machine learning, and a method and apparatus for generating a recognizer using the data generation method and apparatus. Background Technology

[0002] In liquid chromatography (LC) or gas chromatography (GC), chromatographic analysis of a sample yields a chromatogram showing peaks corresponding to the components (usually compounds) contained in the sample. Since the position (retention time) of the peaks observed in the chromatogram corresponds to the type of component in the sample, the component can be identified based on its position, i.e., qualitative analysis is possible. Furthermore, since the area or height value of the peak observed in the chromatogram corresponds to the content or concentration of the component corresponding to that peak, the component can be quantified based on its area or height value. Therefore, to improve the accuracy of qualitative or quantitative analysis based on chromatograms, it is crucial to accurately determine the position of the peak apex, the area value, or the height value of the peaks appearing in the chromatogram.

[0003] Generally, the peak waveforms in a chromatogram ideally follow a Gaussian distribution. However, in actual chromatograms, peaks may exhibit leading-out or tailing due to various factors. Furthermore, baseline drift and overlapping of multiple peaks originating from different components due to inadequate separation are also common. Under these various conditions, accurately detecting peaks and determining their area or height values ​​with high precision is not easy. Various methods and algorithms for peak detection have been proposed for practical applications.

[0004] For example, Non-Patent Document 1 discloses a waveform processing method that sets an appropriate baseline correction line for a provided chromatogram, appropriately separates overlapping peaks based on the baseline correction line, and calculates the cumulative area of ​​each separated peak.

[0005] Traditional peak detection methods rely on conditions such as low noise and stable signal waveforms. Therefore, in many cases where these conditions are not met, accurate peak detection is impossible.

[0006] In contrast, in recent years, with the rapid advancement of machine learning technologies such as deep learning, attempts have been made to utilize machine learning in the automatic detection of peaks on chromatograms. In peak detection methods using machine learning, a peak detection recognizer is pre-generated by learning from a large amount of chromatogram waveforms and the correct positions or feature values ​​such as area and height of peaks observed within those waveforms as teacher data (also known as training data or learning data). Then, if the input is a chromatogram waveform, the recognition process using the aforementioned recognizer is performed to detect peaks or calculate the feature values ​​of the detected peaks.

[0007] Existing technical documents

[0008] Non-patent literature

[0009] Non-Patent Document 1: "Confirmation of Peak Waveform Processing", [Online], Shimadzu Corporation, [Searched December 2, 2019], URL Link<URL:http: / / www.an.sh imadzu.co.jp / hp l c / support / l i b / l cta lk / 23 / 23l ab.htm>

[0010] Non-Patent Literature 2: "LabSolvations Inlight™ Optional Software Peak Intelligence™ LC / MS / MS Method Package for Primary Metabolite and Cell Culture Analysis", [Online], Shimadzu Corporation, [Searched December 2, 2019], URL Link<URL:https: / / www.an.sh imadzu.co.jp / l cms / m_package / peak i nte l l i gence.htm>

[0011] Non-patent literature 3: Alec Radford et al., “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks”, [online], [retrieved December 2, 2019], URL link<URL:https: / / arx i v.org / abs / 1511.06434>

[0012] Non-patent literature 4: Ian J. Goodfell et al., 8, “Generative Adversarial Networks”, [online], [retrieved December 2, 2019], URL link<URL:https: / / arx iv.org / pdf / 1406.2661.pdf> Summary of the Invention

[0013] The technical problem that the invention aims to solve

[0014] In peak detection methods utilizing machine learning, improving the performance of the learned recognizer is crucial for enhancing detection accuracy, including the precision of area or height values. While this requires utilizing as much appropriate teacher data as possible for learning, collecting such data through actual sample analysis demands significant effort and time. This is especially true when samples originate from biological sources, where sample collection often requires substantial labor and time, making it difficult to prepare sufficient teacher data.

[0015] In general machine learning, there are known methods that use data that is not actually measured but rather artificially created by simulators or similar means as teacher data. However, it is difficult to create artificial data that appropriately reflects biases in measured data, especially since there have been no methods or devices that can properly simulate chromatogram or spectral waveforms obtained through equipment analysis.

[0016] The present invention was made to solve the above-mentioned technical problems. Its main purpose is to provide a data generation method and apparatus that can be used to generate or evaluate a peak detection identifier, and can simulate and generate data that appropriately reflects the deviation of measured data.

[0017] Furthermore, another objective of the present invention is to provide a method and apparatus for generating an identifier, which can be used to obtain an identifier capable of detecting peaks with high precision and calculating characteristic values ​​such as peak area or height with high precision.

[0018] Solution to the above technical problems

[0019] One aspect of the data generation method of the present invention, which addresses the aforementioned technical problems, is a data generation method that simulates the generation of data used when creating a recognizer for detecting peaks observed in a signal waveform through machine learning, wherein the following steps are performed by a computer:

[0020] The parameter frequency information acquisition step involves determining the frequency information of a predetermined shape parameter characterizing the shape of a signal waveform based on multiple signal waveforms collected by narrowing down to the target domain for machine learning used to create the recognizer.

[0021] The simulated waveform generation step utilizes the frequency information of the shape parameters to generate a simulated signal waveform that includes the overlap of multiple peaks and noise.

[0022] The simulated signal waveform is provided as data for training or evaluation in machine learning.

[0023] Furthermore, one solution to the data generation apparatus of the present invention, which addresses the aforementioned technical problems, is an apparatus for implementing the data generation method described above. This apparatus simulates the generation of data used when creating a recognizer for detecting peaks observed in a signal waveform through machine learning, and comprises:

[0024] The parameter frequency information acquisition unit calculates frequency information of a predetermined shape parameter characterizing the shape of a signal waveform based on multiple signal waveforms collected by narrowing down to the target domain for machine learning used to create the recognizer.

[0025] The analog waveform generation unit uses the frequency information of the shape parameters to generate an analog signal waveform that includes the overlap of multiple peaks and noise.

[0026] The simulated signal waveform is provided as data for training or evaluation in machine learning.

[0027] Furthermore, in order to solve the above-mentioned technical problems, one aspect of the method for generating a recognizer according to the present invention is a method for generating a recognizer using the data generation method of the above-mentioned aspect of the present invention, wherein a learning step is performed by a computer, and a recognizer for detecting peaks observed in the signal waveform is created by machine learning using the simulated signal waveform generated in the simulated waveform generation step as training data.

[0028] Furthermore, in order to solve the above-mentioned technical problems, one solution of the recognizer generation apparatus of the present invention is a recognizer generation apparatus that uses the data generation apparatus of the above-mentioned solution of the present invention. It includes a learning unit that creates a recognizer for detecting peaks observed in a signal waveform by machine learning through the analog signal waveform generated by the analog waveform generation unit as training data.

[0029] In this invention, a “signal waveform” is a waveform obtained by analyzing or measuring a sample or a metric object. Typically, it is a chromatogram waveform or spectral waveform obtained by using a chromatographic device, a mass analysis device, a nuclear magnetic resonance device, an optical analysis device, an X-ray analysis device, etc., which shows the peaks that correspond to the components (compounds or elements, etc.) in the sample.

[0030] Invention Effects

[0031] In the data generation method and apparatus of the above-described scheme of the present invention, for example, a Generative Adversarial Network (hereinafter referred to as "GAN"), disclosed in Non-Patent Documents 3 and 4 as a method of machine learning, is used to learn a distribution (probability distribution) related to the variation (deviation) of the shape parameter of the peak observed on the signal waveform. During GAN-based learning, for example, the shape parameter extracted from the chromatogram waveform obtained by actually analyzing the sample in a chromatographic apparatus is provided as true data. During GAN-based learning, a generative model (generator) reflecting the frequency (distribution) of the shape parameter can be obtained.

[0032] The shape parameters mentioned here can include, for example, peak height, peak width, tailing degree, and leading-out degree. Furthermore, shape parameters can also include the number of peaks, the distance between adjacent peaks, and the position of peaks near the center of the signal waveform. Generally, in peak detection using fitting, the EMG (Exponential Modified Gaussian) function is often used as the peak model function. The reason is that, ideally, the peak waveform in a chromatogram approximates a Gaussian waveform, but in reality, due to leading-out or tailing, it mostly becomes an asymmetric waveform. Therefore, for example, the EMG function can be used as the peak model function, and its parameters (e.g., tailing parameters) can be used as one of the aforementioned shape parameters.

[0033] Furthermore, adversarial learning does not necessarily require the use of neural networks (i.e., it does not need to be a GAN). Instead, it can use algorithms based on the data output of the learning object that can approximate the distribution of that data.

[0034] Furthermore, even with the same analysis and measurement methods, the peak shapes can vary significantly depending on the type of sample being analyzed or the types (categories) of components contained within the sample. Even if multiple signal waveforms with vastly different peak shapes are provided as real data to a GAN-based learning algorithm, proper learning is not possible, making it difficult to obtain a high-precision generative model. Therefore, in this invention, instead of using a general analysis that does not limit the type of sample, the type of component, the purpose of the analysis, the application field, etc., a certain degree of limitation is imposed to perform GAN learning that uses signal waveforms with similar peak shapes, i.e., a certain degree of consistency, as real data.

[0035] According to the data generation method and apparatus of the present invention, the distribution of shape parameters in the chromatogram waveform actually obtained by the chromatographic apparatus can be acquired with high precision. Using a generation model reflecting such a distribution, signal waveforms exhibiting peaks with reasonable parameter deviations, signal waveforms consisting of multiple overlapping peaks, or signal waveforms with noise superimposed on peaks can be generated in a simulated manner. Peak information such as the start and end points of each peak are correctly obtained from such simulated signal waveforms. Therefore, by using such simulated signal waveforms as teacher data or evaluation data when generating a peak detection recognizer through machine learning, labor and time can be saved, and the performance of the peak detection recognizer can be improved.

[0036] Furthermore, the method and apparatus for generating the identifier according to the above-described scheme of the present invention can detect peaks with high precision, or calculate characteristic values ​​such as peak area or height with high precision. Therefore, in addition to being able to detect peaks that cannot be adequately detected in conventional peak detection methods, it is also possible to improve the quantitative accuracy of components utilizing peak area values, etc. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of one embodiment of an LC device utilizing the data generation method of the present invention.

[0038] Figure 2 This is a functional module configuration diagram of one embodiment of the data generation device of the present invention.

[0039] Figure 3 yes Figure 2 The diagram shows the functional module structure of the adversarial learning execution unit in the data generation device.

[0040] Figure 4 It is generated in Figure 1 The diagram shows the functional module configuration of the learning device of the recognizer used in the LC device.

[0041] Figure 5 This is a diagram showing an example of the input data (real data) for a GAN.

[0042] Figure 6 This is a diagram showing an example of GAN output data (dummy data).

[0043] Figure 7 This is a diagram showing an example of the peak waveform generation result.

[0044] Figure 8 It is a waveform diagram used to illustrate the actions of a waveform simulator. Detailed Implementation

[0045] Hereinafter, an example of the data generation apparatus of the present invention and the identifier generation apparatus utilizing the apparatus will be described with reference to the accompanying drawings.

[0046] In the following example, the signal waveform of the object being analyzed is a chromatogram waveform. The peaks observed on the chromatogram are detected through data analysis, and the position (retention time) of the peak apex and the peak area (or height) are determined.

[0047] [Overall Structure and General Operation of an LC Device According to One Embodiment]

[0048] Figure 1 This is an overall configuration diagram of an embodiment of the LC device utilizing the data generation apparatus of the present invention.

[0049] exist Figure 1 In this device, the measuring unit 10 includes: a mobile phase container 11; a pump 12 for drawing and delivering the mobile phase at a substantially constant flow rate (or flow rate); a syringe 13 for injecting the sample solution into the mobile phase; a chromatographic column 14 for separating components contained in the sample solution based on retention time; and a detector 15 for detecting components in the eluent eluted from the chromatographic column 14. The detector 15 can be, for example, an optical detector such as a photodiode array (PDA) detector or a mass analysis device.

[0050] The data analysis unit 20 has the function of digitizing and processing the detection signal output from the detector 15, and includes functional modules such as a data collection unit 21, a peak detection processing unit 22, and a qualitative and quantitative analysis unit 23. The peak detection processing unit 22 includes functional modules such as a recognition model storage unit 221 and a peak determination unit 222. The recognition model storage unit 221 stores recognition models for peak detection that are pre-generated through machine learning. An input unit 24 and a display unit 25, serving as a user interface, are connected to the data analysis unit 20.

[0051] Generally, the data parsing unit 20 is a general-purpose computer. By executing dedicated software (computer programs) installed on that computer, it is possible to achieve... Figure 1 The functions of each module are shown.

[0052] In this LC apparatus, pump 12 draws mobile phase stored in mobile phase container 11 and delivers it to column 14 at a substantially constant flow rate. Injector 13 injects a pre-prepared sample solution into the mobile phase at a predetermined time, according to instructions from a control unit (not shown). The injected sample solution is introduced into column 14 along with the mobile phase flow, and various components contained in the sample solution are separated and eluted over time during their passage through column 14. Detector 15 detects the components contained in the eluent and outputs a detection signal of intensity corresponding to the amount of that component at constant intervals.

[0053] In the data analysis unit 20, the data collection unit 21 samples the detection signal at predetermined time intervals, converts it into digital data, and stores it in a storage device. This stored data constitutes the chromatogram waveform (chromatogram data), and the components detected in the detector 15 are observed as peaks on the chromatogram. The peak detection processing unit 22 receives the chromatogram data and detects significant peaks, determining the position (time) of the peak apex and the peak area (or height). The qualitative and quantitative analysis unit 23 determines the components based on the position information of each peak on the chromatogram and calculates the content of each component based on the peak area (or height) using a pre-created standard curve. In other words, the qualitative and quantitative analysis unit 23 performs qualitative and quantitative analysis of each component contained in the sample and outputs the results to the display unit 25.

[0054] [Peak detection methods and peak characteristic value calculation methods]

[0055] In the LC apparatus of this embodiment, in order to detect peaks appearing in the chromatogram that are to be analyzed and to obtain peak area values ​​as characteristic values, a recognition model generated by machine learning is used.

[0056] Figure 4 This is a diagram showing the modular structure of a learning device used to create a recognition model.

[0057] The learning device includes a teacher data generation unit 50, a learning execution unit 51, and a recognition model construction unit 52. The teacher data generation unit 50 further includes a random noise generation unit 501, a generator 502, and a noise addition unit 503 as lower-level functional modules. As described later, the generator 502 uses a generator 41 created through GAN-based learning. Furthermore, the physical entity of the learning device is a computer, which executes a prescribed program installed on the computer to specifically implement the following... Figure 4 The functional modules shown.

[0058] Generally, when creating a recognition model for peak detection using machine learning, either supervised learning or semi-supervised learning is employed, using teacher data pre-labeled (information on the correct solution). Labels refer to the correct peak location (time), peak area, height, etc. However, adding manual labeling is not only time-consuming but also prone to errors, reducing learning effectiveness. To avoid this, methods using pre-known simulated data with correct solutions as teacher data are known for the learning process. Figure 4 The teacher data generation unit 50 in the learning device shown is a waveform simulator used to simulate the generation of such teacher data.

[0059] [Methods for generating recognition models for peak detection]

[0060] Next, the method for creating generator 502 will be explained.

[0061] In a chromatogram, the peak waveform corresponding to a component in the sample can be represented by a function f(t, Z) with shape parameter Z and retention time t as independent variables. Here, shape parameter Z is, for example, the peak width or the degree of tailing. Generally, an EMG function or similar function can be used as a model function for the peak waveform observed in a chromatogram. Therefore, it is also possible to use an EMG function as a model function to learn the distribution of its shape parameter. However, here, instead of using such an explicit model function, we will use a method that simultaneously learns the function itself representing the peak waveform and the distribution of the shape that the function can have (the distribution of its shape parameter) to determine a model function and its shape parameter distribution that match the characteristics of the chromatographic apparatus or its analytical application.

[0062] Figure 2 This is a schematic diagram of one embodiment of a waveform information estimation device for estimating peak model functions and shape parameter distributions. The device includes functional modules such as a measured data input unit 30, a peak waveform extraction unit 31, an adversarial learning execution unit 32, and a generative model determination unit 33. The device is implemented by executing a pre-defined program installed on the computer. Figure 2 The functional modules shown. Figure 3 Is Figure 2 The diagram shows the functional module structure of the GAN used in the adversarial learning execution unit 32. Additionally, these devices can be included in... Figure 1 It is included in the data parsing unit 20 of the LC device shown, but it can also exist independently.

[0063] exist Figure 2 and Figure 3 In the waveform information estimation device shown, the chromatogram waveform obtained from actual measurements is used as the real data for GAN-based adversarial learning. That is, in Figure 1 In the measurement unit 10 of the LC apparatus shown, or in an equivalent LC apparatus, the chromatogram waveform obtained by actually measuring the sample is used as true data. However, although it is a peak waveform on a chromatogram, the shape of the peak waveform can vary greatly depending on the type of sample (i.e., the type of component), the application area or purpose of the analysis, and the LC separation conditions. Specifically, if the analysis is attempted to address cases where the tailing tends to be larger or cases where the tailing is almost non-existent, problems such as an excessively wide distribution of shape parameters and reduced distribution accuracy can occur. Therefore, such problems can be avoided by narrowing the application area of ​​the analysis.

[0064] Specifically, this study limits the scope of analysis to the analysis of metabolites derived from organisms. That is, the sample is a biological sample, and the components in the sample are metabolites. Of course, this is just one example, and the method described here can be applied to various samples, components, and analysis purposes.

[0065] exist Figure 2 In the illustrated apparatus, the measured data input unit 30 reads a large amount of chromatogram data obtained through measurement. The peak waveform extraction unit 31 removes peaks with extremely low SN ratios or poorly separated peaks from each input chromatogram data, extracting peaks with good waveform shapes. Then, at least one of the peak-related shape parameters, such as the degree of tailing (or leading-out) of the extracted peak, peak width, and SN ratio, is determined. Furthermore, as characteristic shape parameters of the chromatogram waveform that can contain multiple peaks rather than single peaks, at least one of the following is determined: the number of peaks in the chromatogram waveform, the position of peaks near the center of the waveform's time axis, and the distance between adjacent peaks. The latter shape parameter can be used when determining the position of peaks in the chromatogram waveform.

[0066] Furthermore, for example, if the correlation between parameters such as retention time and peak width can be predicted, in order to facilitate the learning described later, a standardization process that reduces the correlation between parameters, such as standardizing the peak width with retention time, can be performed in advance.

[0067] As described above, the peak waveform extraction unit 31 acquires a large number of shape parameters for single peak waveforms and shape parameters characterizing chromatogram waveforms. The adversarial learning execution unit 32 performs GAN-based learning, using single peaks with good waveform shapes as real data. Figure 3 As shown, the adversarial learning execution unit 32 includes a random noise generation unit 40, a generator 41, a data selection unit 42, a recognizer 43, a determination unit 44, and an update processing unit 45. Here, both the generator 41 and the recognizer 43 use neural networks described in Non-Patent Documents 3 and 4.

[0068] During learning, generator 41 uses a neural network to create a time-dependent function for generating fake data, i.e., fake peak waveforms, based on noise input from random noise generation unit 40. Then, by providing an input showing the time elapsed to the created function, fake peak waveform data of the same length as the observed data (real data) is generated. Data selection unit 42 alternately switches between the fake peak waveform data output from generator 41 and the aforementioned real data, inputting them to recognizer 43. Recognizer 43 identifies whether the input data is real, and determination unit 44 determines whether its recognition result is correct; that is, it determines whether real data is recognized as real when input to recognizer 43, and whether fake data is recognized as fake when input to recognizer 43.

[0069] The update processing unit 45 updates the coefficients of the neural network based on the decision result determined by the decision unit 44, thereby improving the performance of the generator 41 and the recognizer 43 respectively. As is well known, in GANs, the generator 41 and the recognizer 43 compete while simultaneously learning to improve their respective performance. Specifically, the neural network in the generator 41 learns to generate a function that can generate fake data that is as close as possible to real data, while the neural network in the recognizer 43 learns to correctly distinguish between real and fake data.

[0070] Through learning in the GAN as described above, generator 41 generates a peak waveform that closely resembles the real peak waveform. However, during this learning process, a generative model containing a model function that approximates the peak waveform and the distribution of the function's parameters is obtained. The generative model determination unit 33, based on the learning results in the GAN implemented by the adversarial learning execution unit 32, determines the generative model used to simulate the generation of the peak waveform. This generative model is generator 41 itself, and also includes waveform information such as the model function representing the shape of the peak waveform and the distribution of its shape parameters.

[0071] Figure 5 This is a graph showing a portion of the real data used in GAN-based learning. On the other hand, Figure 6 This is a diagram showing a portion of the fake data generated by generator 41. By comparing these, it can be seen that by providing an appropriate amount and quality of real data for GAN-based learning, it is possible to obtain a generative model that can output fake data that is extremely close to (virtually indistinguishable from) the real data.

[0072] Furthermore, while GANs typically produce outputs from inputs that are vectors of fixed length, scaling is required to maintain a consistent vector length when the input is a fixed-length vector, especially when peak widths vary significantly. Additionally, the shape of peak waveforms is generally strongly constrained by the monotonicity of their differential values. However, since this constraint cannot be applied to learning, the initial values ​​of the waveform generated by generator 41 differ significantly from the original peak shape. Consequently, a long learning time is likely to occur before the output of a peak waveform with an appropriate shape can be produced. Therefore, given the time required for the generator 41's output to become the shape of the data to be learned, GAN-based learning is empirically known to fail.

[0073] Therefore, to avoid this situation, a function s(t, Z) that returns the distortion time can be created in the neural network included in generator 41, and a Gaussian function (s(t, Z)) using this function can be generated in generator 41. Furthermore, for the independent variable t of these functions, the input t has the same size as the vector size of the real data, so as to generate data with the same size as the waveform vector being learned. Of course, a general Gaussian function or an EMG function can also be used in the neural network included in generator 41.

[0074] Therefore, it is possible to directly learn the distribution of waveform vectors while simultaneously learning the distribution of the time distortion parameter function s(t, Z). That is, in a typical GAN, only "vector" to "vector" learning can be performed, while in the method of this invention, the GAN is extended to be able to learn "time series function" to "vector".

[0075] Furthermore, more preferably, in order to achieve the distribution shown by s(t, Z) ≈ t, skip connections can be used in the network of the function s(t, Z), or as s(t, Z) = t + s'(t, Z), the initial distribution of the parameters of the network weights can be adjusted so that s'(t, Z) takes values ​​near 0 with an appropriate width.

[0076] Furthermore, we also consider cases where the bell-shaped function exhibits end sluggishness, resulting in distortion that applies the intensity approximately uniformly. In such cases, it is useful to use the intensity correction neural network function i(y), with i(Gauss(s(t, Z)) as the generator.

[0077] On the other hand, a general neural network can also be used as the recognizer 43, but here, in order to better capture the bell-shaped waveform like the Gaussian function, a network with one-dimensional convolutional layers and linear combination layers is used.

[0078] In the above description, the essential purpose of the adversarial learning execution unit 32 is to learn a distribution consistent with the distribution of the data point group. In the above example, GAN is used as the learning method. As shown in Non-Patent Documents 3 and 4, neural networks are generally used in GANs. However, it is clear from Non-Patent Document 3 that an appropriate algorithm can be used to replace the neural network. This appropriate algorithm can optimize the function using differentiation (gradient).

[0079] Furthermore, here, the KL (Kul back-Leibler) distance is used as the similarity between the observed waveform D and the distribution of the function f generated by the shape parameter Z. However, in addition to the KL distance, various indicators such as the JS (Jensen-Shannon) distance or the Wasserstein distance can be used to evaluate the consistency of the distributions.

[0080] As described above, it is possible to obtain a generation model that simulates the peak waveform observed on the measured chromatogram, that is, it is possible to obtain the distribution information of the model function and shape parameters.

[0081] Figure 4 The generator 502 shown uses the above-described generation model to output a pseudo waveform based on the input noise, which closely resembles the shape of the peak waveform on the chromatogram. The model function for this pseudo waveform is f(t, Z), and the shape parameter of this model function is deformed according to a predetermined probability distribution. In the learning device, noise generated by the random noise generation unit 501 is input to the generator 502, and the peak waveform with a shape according to the distribution of the shape parameter is obtained as described above.

[0082] Figure 7 (A) and (B) are examples of the shapes of a single peak waveform generated by generator 502. The shape of this peak waveform is obtained as a function of time, thus allowing for free variation of the peak width and peak height.

[0083] The noise addition unit 503 adds an appropriate background noise component to the above peak waveform to obtain teacher data. On the other hand, since the peak before noise addition is the peak waveform of the positive solution, it becomes a label corresponding to the teacher data.

[0084] Because generator 502 possesses a probability distribution of shape parameters that accurately reflects the shape changes of the measured peak waveform, it can output a waveform that correctly simulates the changes in peak shape corresponding to the target analytical region. Of course, there is no limit to the number of waveforms generated in the teacher data generation unit 50. Therefore, in Figure 4 In the learning device shown, high-performance recognition models can be efficiently constructed through learning using a large amount of good teacher data. The recognition models created in this way are stored in the recognition model storage unit 221 of the LC device in this embodiment. Therefore, in the LC device of this embodiment, peaks observed in the chromatogram obtained through measurement can be accurately detected, and characteristic values ​​such as peak area or height can be calculated with high precision.

[0085] exist Figure 4In the teacher data generation unit 50 shown, when generating simulated chromatogram waveforms of observed multiple peaks as teacher data, multiple peaks are generated in generator 502. Figure 8 (A) shows a peak distribution based on a prescribed peak shape. Similarly, based on the distribution of shape parameters such as the number of peaks derived from the provided data, the position of peaks near the center of the waveform's time axis, and the distance between adjacent peaks, the number of peaks or the position of each peak is determined, thereby forming a mixed-peak waveform combining multiple peaks. Background noise based on the distribution of shape parameters such as the SN ratio (see reference) is then added to this mixed-peak waveform. Figure 8 (B)) generates such as Figure 8 The simulated chromatogram waveform shown in (C) is used as teacher data. By using such chromatogram waveforms as teacher data, it is possible to construct an identification model that can detect multiple overlapping peaks and output feature values ​​such as the area or height of these peaks.

[0086] Furthermore, the teacher data generation unit 50 can be used not only to generate teacher data during learning, but also to generate evaluation data used when evaluating (or testing) the recognition model created through learning.

[0087] Furthermore, a sufficient amount of teacher data is generally required when learning peak waveform shapes. Therefore, sometimes data obtained from samples targeting a fairly broad field (target) is used to perform peak waveform shape learning. For example, in the case of LC devices, there are cases where the analysis is not of a specific substance targeting a specific biological sample such as blood or urine, but rather data obtained from general analysis of biological samples is used for learning. That is, there are cases where the teacher data includes a dataset covering a broader field than the actual field where the peak detection of the present invention is to be applied. In such cases, it is of course also considered to augment the data by simulating only the area near the parameter points corresponding to the sample data of the actual application field. That is, by pre-learning the distribution of peak shape parameters targeting a broad field, investigating the range within the peak shape parameter distribution where the peak clusters of samples targeting a narrow field are concentrated, it is possible to generate simulated data based on the shape parameter distribution limited to that range.

[0088] Furthermore, as mentioned above, in practical applications, there are many situations where it is necessary to handle mixed peaks with overlapping ends of adjacent peaks. However, in many analytical devices such as LC and GC devices, linearity in the analysis can be considered to be ensured. Therefore, when the object of analysis is the chromatogram waveform, the signal value of a mixed peak can also be regarded as the value obtained by simply adding the signal values ​​corresponding to multiple individual peaks.

[0089] The above embodiments are examples of applying the present invention to an LC device, but it is obvious that the present invention can be applied when processing or analyzing signal waveforms obtained from various analysis devices other than LC devices or GC devices.

[0090] Specifically, the present invention can also be used for peak detection of mass spectra obtained by mass analysis devices, nuclear magnetic resonance spectra obtained by nuclear magnetic resonance devices, optical spectra such as absorption spectra or reflection spectra obtained by optical analysis devices, and X-ray energy spectra obtained by X-ray analysis devices.

[0091] [Various options]

[0092] Those skilled in the art will understand that the above exemplary embodiments are specific examples of the following schemes.

[0093] (Item 1) One aspect of the data generation method of the present invention is a data generation method that simulates the generation of data used when creating a recognizer for detecting peaks observed in a signal waveform through machine learning, wherein the following steps are performed by a computer:

[0094] The parameter frequency information acquisition step involves determining the frequency information of a predetermined shape parameter characterizing the shape of a signal waveform based on multiple signal waveforms collected by narrowing down to the target domain for machine learning used to create the recognizer.

[0095] The simulated waveform generation step utilizes the frequency information of the shape parameters to generate a simulated signal waveform that includes the overlap of multiple peaks and noise.

[0096] The simulated signal waveform is provided as data for training or evaluation in machine learning.

[0097] (Item 2) In the data generation method described in Item 1, it can be configured such that, in the parameter frequency information acquisition step, the shape parameters obtained from the multiple signal waveforms are used as input, and adversarial learning is performed using two opposing models, a generation model and a recognition model, to obtain a generation model containing the frequency information of the shape parameters. In the simulated waveform generation step, the generation model is used to generate a simulated signal waveform.

[0098] (Item 9) One aspect of the data generation apparatus of the present invention is a data generation apparatus that simulates the generation of data used when creating a recognizer for detecting peaks observed in a signal waveform through machine learning, comprising:

[0099] The parameter frequency information acquisition unit calculates frequency information of a predetermined shape parameter characterizing the shape of a signal waveform based on multiple signal waveforms collected by narrowing down to the target domain for machine learning used to create the recognizer.

[0100] The analog waveform generation unit uses the frequency information of the shape parameters to generate an analog signal waveform that includes the overlap of multiple peaks and noise.

[0101] The simulated signal waveform is provided as data for training or evaluation in machine learning.

[0102] (Item 10) In the data generation apparatus described in Item 9, it can be configured such that, in the parameter frequency information acquisition unit, shape parameters obtained from the plurality of signal waveforms are used as input, and adversarial learning is performed using two opposing models, a generation model and a recognition model, to obtain a generation model containing frequency information of the shape parameters, and the analog waveform generation unit uses the generation model to generate an analog signal waveform.

[0103] In the data generation methods described in items 1 and 2, and the data generation apparatus described in items 9 and 10, for example, it is possible to acquire with high accuracy information on the distribution of model functions and shape parameters such as peak tailing degree and peak width for simulating chromatogram waveforms obtained by a chromatographic apparatus with high precision. Therefore, it is possible to freely generate simulated signal waveforms that accurately reflect deviations from the actually obtained signal waveforms and accurately obtain peak information such as the start and end points of each peak. Thus, it is possible to save labor and time used for creating teacher data, etc., while improving the performance of the identifier used to detect peaks observed in the signal waveform.

[0104] (Item 3) In the data generation method described in Item 1 or Item 2, it can be configured such that, in the parameter frequency information acquisition step, isolated peaks with a SN ratio greater than or equal to a reference value are extracted from the signal waveform and the shape parameters associated with the peaks are determined.

[0105] (Item 11) Similarly, in the data generation apparatus described in Item 9 or Item 10, it can be configured such that the parameter frequency information acquisition unit extracts isolated peaks from the signal waveform and peaks with an SN ratio of more than a reference value and calculates the shape parameters associated with the peaks.

[0106] According to the data generation method described in item 3 and the data generation apparatus described in item 11, the accuracy of the generative model obtained through adversarial learning such as GAN can be improved.

[0107] (Item 4) In the data generation method described in Item 3, it can be set such that the shape parameters associated with the peak include any one of peak height, peak width, and peak tailing degree.

[0108] (Item 12) Furthermore, in the data generation apparatus described in Item 11, it is possible to configure the shape parameters associated with the peak to include any one of peak height, peak width, and peak tailing degree.

[0109] According to the data generation method described in item 4 and the data generation apparatus described in item 12, the accuracy of the generated model can be improved by learning shape parameters related to appropriate peaks.

[0110] (Item 5) In any of the data generation methods described in items 1 to 4, the shape parameters can be configured to include the number of peaks, the distance between adjacent peaks, and the position of the peaks located near the center of the signal waveform. In the simulated waveform generation step, the peak positions in the simulated signal waveform are determined by the frequency information of these shape parameters.

[0111] (Item 13) In any of the data generation apparatuses described in items 9 to 12, the shape parameters can be configured such that the number of peaks, the distance between adjacent peaks, and the position of the peaks located near the center of the signal waveform are included, and the analog waveform generation unit determines the peak positions in the analog signal waveform based on the frequency information of these shape parameters.

[0112] According to the data generation method described in item 5 and the data generation apparatus described in item 13, it is possible to generate signal waveforms that simulate the overlap of multiple peaks with high precision.

[0113] (Item 6) In any of the data generation methods described in items 1 to 5, it can be configured such that, in the simulated waveform generation step, noise is added to a waveform containing one or more peaks generated using the generation model to generate a simulated signal waveform.

[0114] (Item 14) Furthermore, in any of the data generation apparatuses described in items 9 to 13, the analog waveform generation unit can be configured to generate an analog signal waveform by adding noise to a waveform containing one or more peaks generated using the generation model.

[0115] According to the data generation method described in item 6 and the data generation apparatus described in item 14, teacher data that accurately simulates signal waveforms with significant noise can be obtained. Therefore, a generation model with high accuracy in peak detection or peak area calculation can be obtained.

[0116] (Item 7) In any of the data generation methods described in items 1 to 6, it can be set such that the multiple signal waveforms collected in the target field are obtained by analyzing chromatogram waveforms or spectral waveforms of a specific type of sample.

[0117] (Item 15) Furthermore, in any of the data generation apparatuses described in items 9 to 14, it is possible to configure the multiple signal waveforms collected in the target area to be chromatogram waveforms or spectral waveforms obtained by analyzing a specific type of sample.

[0118] According to the data generation method described in item 7 and the data generation apparatus described in item 15, since the deviation of the shape parameters of the chromatogram waveform or the spectrum waveform is reduced, a high-precision generation model can be created, and the accuracy of the simulated signal waveform is also improved.

[0119] (Item 8) One aspect of the invention of item 8 is a method for generating a recognizer using the data generation method described in any one of items 1 to 7, which can be configured to perform a learning step by using machine learning to generate a simulated signal waveform generated in the simulated waveform generation step as training data to create a recognizer for detecting peaks observed in the signal waveform.

[0120] (Item 16) Furthermore, one aspect of the invention in Item 16 is a recognizer generation device that uses the data generation apparatus described in any one of Items 9 to 15, which can be configured to include a learning unit that creates a recognizer for detecting peaks observed in a signal waveform by machine learning through machine learning using a simulated signal waveform generated by the simulated waveform generation unit as training data.

[0121] According to the method for generating the identifier described in item 8 and the apparatus for generating the identifier described in item 16, an identifier capable of detecting peaks observed in chromatogram waveforms or the like with high precision and calculating the area or height value of the peak with high precision can be obtained. This reduces the rate of missed or false peak detections and improves the accuracy of quantification based on peak area or height values.

[0122] Explanation of reference numerals in the attached figures

[0123] 10 Measurement Department

[0124] 11. Mobile Phase Container

[0125] 12 pumps

[0126] 13 syringes

[0127] 14 chromatographic column

[0128] 15 detectors

[0129] Peak Detection and Processing Department

[0130] 20 Data Analysis Department

[0131] 21 Data Collection Department

[0132] 22 Peak Detection and Processing Department

[0133] 221 Recognition Model Storage Department

[0134] 222 Summit Decision Department

[0135] 23 Qualitative and Quantitative Analysis Department

[0136] 24 Input Section

[0137] 25 Display Unit

[0138] 30 Measured Data Input Section

[0139] 31 Peak Waveform Extraction Section

[0140] 32 Confrontational Learning Executive Department

[0141] 33 Generative Model Decision Department

[0142] 40 Random Noise Generator

[0143] 41 Generator

[0144] 42 Data Selection Department

[0145] 43 Recognizer

[0146] 44 Judgment Department

[0147] 45 Update Processing Department

[0148] 50 Teacher Data Generation Department

[0149] 501 Random Noise Generation Unit

[0150] 502 generator

[0151] 503 Noise Addition Section

[0152] 51 Learning Implementation Department

[0153] 52. Recognition Model Construction Department.

Claims

1. A method for generating a recognizer, characterized in that, include: The steps of obtaining multiple signal waveforms by measuring multiple samples using an analytical device; The step of extracting the single peak contained in each of the multiple signal waveforms and determining the predetermined shape parameters that characterize the shape of the single peak; The steps involve taking the shape parameters as input and performing adversarial learning using two opposing models, a generative model and a recognition model, to obtain a generative model containing a model function with an approximate peak waveform and the distribution of the parameters of that function. The step of using the aforementioned generation model to generate a simulated signal waveform containing overlapping peaks and / or noise; The step of creating an identifier for detecting peaks observed in signal waveforms obtained from measuring samples by using machine learning with the simulated signal waveforms as training or evaluation data.

2. The method for generating the identifier as described in claim 1, characterized in that, In the step of determining the shape parameters, isolated peaks with a single peak whose SN ratio is above a reference value are extracted from the signal waveform, and the shape parameters associated with that peak are determined.

3. The method for generating the identifier as described in claim 2, characterized in that, The shape parameters include any one of peak height, peak width, and peak tailing degree.

4. The method for generating the identifier as described in claim 1, characterized in that, The shape parameters include the number of peaks, the distance between adjacent peaks, and the position of the peaks located near the center of the signal waveform. In the step of generating the simulated signal waveform, the distribution of these shape parameters determines the peak positions in the simulated signal waveform.

5. The method for generating the identifier as described in claim 1, characterized in that, In the step of generating a simulated signal waveform, noise is added to a waveform containing one or more peaks generated using the generation model to generate a simulated signal waveform.

6. A recognition generator generating device, characterized in that, have: The parameter information acquisition unit extracts the single peak contained in each of the multiple signal waveforms obtained by measuring multiple samples using an analysis device, and determines the predetermined shape parameter characterizing the shape of the single peak. The generative model acquisition unit takes the shape parameters as input and performs adversarial learning using two opposing models, a generative model and a recognition model, to obtain a generative model containing a model function with an approximate peak waveform and the distribution of the parameters of the function. The analog waveform generation unit uses the generation model to generate an analog signal waveform that includes the overlap of multiple peaks and / or noise. The identifier acquisition unit creates an identifier for detecting peaks observed in the signal waveform obtained from measuring the sample by using machine learning with the simulated signal waveform as training or evaluation data.

7. The identifier generation apparatus as described in claim 6, characterized in that, The parameter information acquisition unit extracts isolated peaks from the signal waveform and single peaks with a SN ratio greater than or equal to a reference value, and calculates the shape parameters associated with the peak.

8. The identifier generation apparatus as described in claim 7, characterized in that, The shape parameters include any one of peak height, peak width, and peak tailing degree.

9. The identifier generation apparatus as described in claim 6, characterized in that, The shape parameters include the number of peaks, the distance between adjacent peaks, and the position of the peaks located near the center of the signal waveform. The analog waveform generation unit determines the peak positions in the simulated signal waveform by the distribution of these shape parameters.

10. The identifier generation apparatus as described in claim 6, characterized in that, The analog waveform generation unit generates an analog signal waveform by adding noise to a waveform containing one or more peaks generated using the generation model.

Images