Patent Application
20210201140
The present invention relates to a sample analysis apparatus and a sample analysis program, and is particularly suitable for use in an apparatus for analyzing a characteristic of a frequency spectrum obtained by spectroscopic measurement of a sample using a terahertz wave to obtain information about the sample.
Conventionally, a spectroscopic apparatus that measures a characteristic of a substance using an electromagnetic wave has been provided. In the spectroscopic apparatus, a sample transmits or reflects an electromagnetic wave, and a physical property or a chemical property of the sample is measured from a change in the electromagnetic wave caused by an interaction between the electromagnetic wave and the sample. A frequency spectrum of the sample observed by this spectroscopic measurement has a spectrum structure unique to the sample. In particular, in spectroscopy using a terahertz wave, which is a type of electromagnetic wave, an intermolecular interaction due to a hydrogen bond, etc. is observed.
However, since the intermolecular interaction occurring in the sample in response to the terahertz wave has a complicated process, spectra observed by spectroscopic measurement overlap each other, and thus it is more difficult to extract a peak at a specific frequency than in infrared spectroscopy or photoelectron spectroscopy. Therefore, it is difficult to determine a location in the frequency spectrum at which the characteristic of the sample appears or a type of waveform in which the characteristic of the sample appears, and there is a problem that it is extremely difficult to find the characteristic.
In recent years, in view of such a problem, attempts have been made to create a learning model from a spectral shape of terahertz spectroscopy data and analyze the spectroscopy data using machine learning (for example, see Non-Patent Document 1). Non-Patent Document 1 describes that spectroscopic measurement of a potassium chloride aqueous solution having a plurality of concentrations is performed by terahertz total reflection spectroscopy, and machine learning is performed using a plurality of pieces of spectroscopy spectrum data obtained in this way as learning data, thereby creating a model for predicting the potassium chloride concentration from the spectroscopy data.
Non-Patent Document 1: The 65th Japan Society of Applied Physics, Spring Academic Lecture, “Analysis of Terahertz Spectroscopic Measurement Data of Ion Hydration State Using Machine Learning” (Mar. 20, 2018, Daiki Kawakami, Hitoshi Tabata)
According to Non-Patent Document 1, when learning data was learned by the least squares method, the Ridge method, and the Lasso method, respectively, it was reported that the learning model by the Ridge method was able to predict the potassium chloride concentration of test data most accurately. As is well known, to improve the accuracy of prediction in machine learning, it is important how to appropriately extract a feature quantity of input data using a learning model. However, Non-Patent Document 1 does not mention a scheme of extracting a feature quantity from spectroscopy spectrum data of a terahertz wave.
The invention has been made to solve the above-described problems, and an object of the invention is to enable more accurate prediction of information about a sample from a frequency spectrum obtained by performing spectroscopic measurement on the sample using a terahertz wave.
To solve the above-mentioned problem, in the invention, a learning model for predicting information about a sample is prepared using a parameter determining a property for each of a plurality of fitting functions corresponding to a generation source of composite waveform fit to a frequency spectrum obtained by spectroscopic measurement of the sample using a terahertz wave, and a parameter obtained for the sample to be predicted is applied to the learning model, thereby predicting the information about the sample to be predicted.
According to the invention configured as described above, it is possible to predict information about a sample to be predicted using a learning model in which a feature quantity of a frequency spectrum obtained by performing spectroscopic measurement on a sample is represented by parameters that determine properties of a plurality of fitting functions. That is, a composite waveform of the plurality of fitting functions is fit to a frequency spectrum of a terahertz wave of the sample. For this reason, the parameters that determine the properties of the fitting functions reflect a property of the sample hidden in the frequency spectrum. Therefore, even when the frequency spectrum obtained by performing spectroscopic measurement on the sample using the terahertz wave is one in which a difference in property of the sample is unlikely to clearly appear as a feature of a waveform such as a peak of a specific frequency, information about the sample to be predicted can be accurately predicted by the learning model.
Hereinafter, a description will be given of a first embodiment of the invention with reference to drawings.
Each functional block of the learning device 10A and the prediction device 20A can be configured by any of hardware, a digital signal processor (DSP), and software. For example, in the case of being configured by software, each of the above functional blocks is actually configured to include a CPU, a RAM, a ROM, etc. of a computer, and implemented by an operation of a program stored in a recording medium such as the RAM, the ROM, a hard disk, or a semiconductor memory. This description is similarly applied to each functional block included in the sample analysis apparatus according to the second and third embodiments described later.
The learning data input unit 11A inputs a parameter obtained for a learning sample and information about the learning sample as learning data for each of a plurality of learning samples. Here, the plurality of learning samples are samples whose information about the sample (for example, a type of substance used as the sample and a property of the substance (physical property, chemical property, etc.)) is known, and have different properties. It is possible to use a solid or gaseous substance as the sample. However, in the present embodiment, it is possible to use, as the sample, a liquid whose characteristic particularly hardly appears on a frequency spectrum obtained by spectroscopic measurement using a terahertz wave.
For example, in a case where a liquid is used as a sample, it is considered that whether or not a foreign substance is mixed in the liquid is analyzed by terahertz spectroscopy measurement and machine learning. In this case, a plurality of solutions in which additives are intentionally mixed with a liquid in which no foreign substance is mixed is prepared and used as a learning sample. Here, when a plurality of learning samples is prepared by mixing different types of additives, it is possible to predict a type of foreign substance mixed in a liquid to be analyzed based on a learning model created by machine learning using the learning samples. In the case of this example, a known property of a substance (liquid) used as the learning sample is a property that “a foreign substance of a type mixed as an additive is mixed in the liquid”.
In addition, when a plurality of learning samples is prepared by changing the mixing amount of the same type of additives, it is possible to predict the amount of a specific type of foreign substance mixed in the liquid to be analyzed based on a learning model created by machine learning using the learning samples. In the case of this example, a known property of the substance (liquid) used as the learning sample is “the amount of the specific foreign substance mixed as an additive”. Furthermore, when a plurality of learning samples is created by changing the mixing amount of each of different types of additives, it is possible to predict a type of foreign substance and the amount of the foreign substance mixed in the liquid to be analyzed based on a learning model created by machine learning using the learning samples.
In addition, by creating a liquid containing no foreign substance and a liquid prepared by mixing the same or different additives as a plurality of learning samples, it is possible to predict whether or not a foreign substance is mixed in the liquid to be analyzed (presence or absence of mixing of the foreign substance) based on a learning model created by machine learning using the learning samples. In the case of this example, a known property of the substance (liquid) used as the learning sample is a property that “mixing of the foreign substance is absent” or “mixing of the foreign substance is present”.
“Information about the learning sample” input as one piece of the learning data is known information about the sample. The information about the learning sample is used as teacher data for machine learning. Hereinafter, the information about the learning sample may be referred to as teacher data. As described above, for example, in the case of creating a plurality of learning samples by mixing different types of additives, the information about the learning samples is information indicating the types of the additives. Since it is possible to detect a type of additive mixed when the learning sample is created, information specifying the type of the additive may be used as teacher data. Note that the teacher data in this case maybe used name information of the additive or identification information uniquely allocated to each type of additive.
In addition, a “parameter obtained for the learning sample” input as another piece of the learning data is a parameter calculated by analyzing a frequency spectrum obtained by spectroscopic measurement using a terahertz wave. In other words, a composite waveform of a plurality of fitting functions is fit to a frequency spectrum obtained from a terahertz wave signal of a learning sample detected by the spectroscopic apparatus, and values that determine properties of the plurality of fitting functions used for the fitting are used as parameters.
In the present embodiment, a plurality of normal distribution functions differing in at least one of a center frequency, the amplitude, and a width is used as the plurality of fitting functions. In addition, in the present embodiment, at least one of a center frequency, the amplitude, a width, and the area of a predetermined region in a function waveform of a normal distribution function is used as a parameter that determines a property of a fitting function. For example, when all four of the center frequency, the amplitude, the width, and the area are used as parameters, and n normal distribution functions are used to generate a composite waveform fit to a frequency spectrum of a certain learning sample, 4×n parameters correspond to a “parameter obtained for the learning sample”. Note that details of calculation of this parameter will be described later.
The learning model creation unit 12A creates a learning model using the learning data (parameters and teacher data) input from the learning data input unit 11A, and stores the created learning model in the learning model storage unit 30A. For example, the learning model creation unit 12A creates the learning model by applying a known machine learning algorithm (for example, a machine learning algorithm using a neural network) using the above-described learning data.
As illustrated in
In addition, the output layer may have one node that outputs a value indicating the amount of foreign substance mixed in the liquid. Alternatively, it is possible to adopt a learning model in which a section from an upper limit to a lower limit, which can be assumed as the amount of foreign substance mixed in the liquid, is divided into a plurality of ranges, the same number of nodes as the number of divided ranges are provided in the output layer, and an index value indicating a probability of the amount of mixed foreign substance is output to each node. Further, it is possible to adopt a learning model in which the same number of nodes as the number of a plurality of types of additives used for creating the learning sample is provided in the output layer, and an index value indicating a probability of the type of the mixed foreign substance is output to each node.
As described above, a configuration of the learning model created by the learning model creation unit 12A is appropriately set depending on the object to be analyzed. That is, the neural network illustrated in
The prediction data input unit 21A inputs a parameter obtained for a sample to be predicted as prediction data. The “parameter obtained for the sample to be predicted” input by the prediction data input unit 21A as the prediction data refers to a parameter calculated by analyzing a frequency spectrum obtained by spectroscopic measurement using a terahertz wave. This analysis method is the same as analysis performed on the frequency spectrum of the learning sample. That is, a composite waveform of a plurality of fitting functions is fit to a frequency spectrum obtained from a terahertz wave signal of a sample to be predicted (hereinafter, referred to as a prediction sample) detected by the spectroscopic apparatus, and values that determine properties of the plurality of fitting functions used for the fitting are used as parameters.
Here, the parameters input by the prediction data input unit 21A are of the same type as the parameters input by the learning data input unit 11A. That is, in case that the learning data input unit 11A inputs 4×n parameters (the center frequency, the amplitude, width, and the area for each of the n normal distribution functions used to generate the composite waveform) for one learning sample, the prediction data input unit 21A also inputs 4×n parameters for the prediction sample.
The sample information prediction unit 22A predicts information about the prediction sample by applying the prediction data (parameter) input by the prediction data input unit 21A to the learning model stored in the learning model storage unit 30A. That is, the sample information prediction unit 22A inputs the parameter input by the prediction data input unit 21A to the input layer of the learning model, thereby acquiring the index value output from the output layer as information predicted for the prediction sample. In the case of mixing of the foreign substance described above as an example, the information predicted for the prediction sample corresponds to the presence or absence of mixing of the foreign substance in a liquid used as the prediction sample, the type of foreign substance mixed in the liquid, the amount of foreign substance mixed in the liquid, etc.
Here, a method of calculating the above-described parameters will be described in detail.
Each of the functional blocks 201 to 204 can be configured by any of hardware, a DSP, and software. For example, in the case of being configured by software, each of the functional blocks 201 to 204 is actually configured to include a CPU, a RAM, a ROM, etc. of a computer, and implemented by an operation of a program stored in a recording medium such as the RAM, the ROM, a hard disk, or a semiconductor memory.
Here, in the present embodiment, terahertz wave signals of a plurality of learning samples are detected by the spectroscopic apparatus 300, and processing of the parameter calculation apparatus 200 is performed on each of the terahertz wave signals. In addition, in the present embodiment, a terahertz wave signal of a prediction sample is detected by the spectroscopic apparatus 300, and processing of the parameter calculation apparatus 200 is performed on the terahertz wave signal. In other words, the parameter calculation apparatus 200 serves as both a learning parameter calculation unit and a prediction parameter calculation unit of the claims.
The frequency spectrum acquisition unit 201 obtains a frequency spectrum representing an absorbance with respect to a frequency based on the terahertz wave signal detected by the spectroscopic apparatus 300. The spectroscopic apparatus 300 causes a sample to be measured disposed on an optical path to transmit or reflect a terahertz wave, and detects the terahertz wave applied to the sample in such a manner. In the present embodiment, various known types can be used as the spectroscopic apparatus 300.
The thinning processing unit 202 thins out an extreme value at a frequency at which absorption of the terahertz wave is increased by water vapor other than the sample from absorbance data for each frequency in the frequency spectrum obtained by the frequency spectrum acquisition unit 201. In addition to the sample, water vapor exists on the optical path of the spectroscopic apparatus 300. Since the terahertz wave is absorbed by water vapor, there is a possibility that a property of the water vapor is included in the acquired frequency spectrum. Therefore, the thinning processing unit 202 performs a process of thinning out an extreme value at a frequency at which absorption of the terahertz wave by the water vapor increases.
Note that the frequency at which the absorption of the terahertz wave by the water vapor increases can be specified using, for example, data provided by NICT (National Institute of Information and Communications Technology). NICT discloses data on a radio wave attenuation rate of air (including water vapor) for terahertz wave communication. By using this data, it is possible to specify the frequency at which the absorption of the terahertz wave by the water vapor increases.
Note that it is possible to construct a vacuum environment or an environment in which there is significantly little water vapor equivalent thereto, and install the spectroscopic apparatus 300 in the environment. In such a case, the thinning processing unit 202 can be omitted.
The fitting processing unit 203 fits a composite waveform of a plurality of normal distribution functions different in at least one of the center frequency, the amplitude, and the width to the frequency spectrum obtained by the frequency spectrum acquisition unit 201. In the present embodiment, the fitting processing unit 203 performs a process of fitting a composite waveform of a plurality of normal distribution functions to a plurality of pieces of absorbance data thinned out by the thinning processing unit 202.
In more detail, under the assumption that the frequency spectrum can be approximated by the overlap of a plurality of normal distribution waveforms, the fitting processing unit 203 calculates a plurality of normal distribution functions that minimizes a residual between absorbance data at each frequency of the frequency spectrum (a plurality of pieces of absorbance data thinned out by the thinning processing unit 202) and a value of the composite waveform at each frequency corresponding thereto by optimization calculation using the center frequency, the amplitude, and the width as variables.
As described above, in the present embodiment, a normal distribution function (Gaussian function) is used as an example of a function used for fitting. In addition, a 1/e width is used as an example of a width of the normal distribution function. In addition, as the area of a predetermined region in a function waveform of the normal distribution function, the area of a waveform region having the amplitude equal to or greater than the amplitude corresponding to the 1/e width is used. In the present embodiment, the number of normal distribution functions to be synthesized can be arbitrarily set.
For each of a plurality of frequency spectra obtained for a plurality of samples collected from the same liquid, the first fitting processing unit 203A performs fitting to a frequency spectrum using a composite waveform of a plurality of normal distribution functions using the center frequency, the amplitude, and the width as parameters and different in at least the center frequency.
The center frequency specification unit 203B groups each center frequency of the normal distribution function used for a plurality of times of fitting to a plurality of frequency spectra (for a plurality of samples collected from the same liquid) by the first fitting processing unit 203A, and specifies a representative center frequency from each group. For example, the center frequency specification unit 203B performs grouping in units of collections of respective center frequencies of a plurality of normal distribution functions obtained by a fitting process, and specifies one or more representative center frequencies from each group.
As illustrated in
Then, the center frequency specification unit 203B specifies one or a plurality of representative center frequencies from each of the groups Gr1 to Gr4. A method of specifying the representative frequency can be arbitrarily set. For example, it is possible to specify, as a representative, one frequency at which a plurality of center frequencies is most concentrated in a group. Alternatively, an average value of a plurality of center frequencies belonging to a group may be calculated, and one center frequency closest to the average value may be specified as a representative. Further, with regard to a group in which a frequency range of the group is wide and a plurality of center frequencies is relatively widely dispersed, such as the groups Gr3 and Gr4, a plurality of center frequencies may be specified at equal intervals as representatives.
Note that when grouping is performed for each collection of center frequencies, only a center frequency whose residual value is smaller than a predetermined value may be used.
The second fitting processing unit 203C fixes n center frequencies specified by the center frequency specification unit 203B for each of a plurality of frequency spectra obtained for a plurality of samples collected from the same liquid, and performs fitting to the frequency spectra again with a composite waveform of n normal distribution functions using the amplitude and the width as parameters. In other words, the second fitting processing unit 203C calculates n normal distribution functions (the center frequency is the one specified by the center frequency specification unit 203B) that minimize a residual between a value of absorbance at each frequency of a frequency spectrum and a value of a composite waveform at each frequency corresponding thereto by optimization calculation using the amplitude and the width as variables.
As illustrated in
The parameter acquisition unit 204 acquires, as parameters, values that determine the properties of the plurality of fitting functions used for fitting as described above. That is, the parameter acquisition unit 204 acquires, as parameters, at least one of the center frequency, the amplitude, the width, and the area of the n normal distribution functions used for fitting by the second fitting processing unit 203C. In the case of creating a learning model using the neural network illustrated in
The parameters acquired by the parameter acquisition unit 204 are stored in, for example, a removable storage medium and input to the sample analysis apparatus 100A illustrated in
Note that here, a description has been given of an example in which parameters are transferred from the parameter calculation apparatus 200 to the sample analysis apparatus 100A via the removable storage medium. However, the invention is not limited thereto. For example, the parameter calculation apparatus 200 and the sample analysis apparatus 100A may be connected via a wired or wireless communication network, and parameters may be transmitted from the parameter calculation apparatus 200 to the sample analysis apparatus 100A via the communication network.
In addition, here, a description has been given of an example in which the parameter calculation apparatus 200 and the sample analysis apparatus 100A are separately configured. However, the invention is not limited thereto. For example, the sample analysis apparatus 100A may have an integrated configuration incorporating a function of the parameter calculation apparatus 200.
In the case of configuring the sample analysis apparatus 100A′ as illustrated in
In addition, at the time of prediction, the parameter calculation unit 200′ acquires a frequency spectrum obtained by spectroscopic measurement using a terahertz wave for the prediction sample as prediction data, and analyzes the acquired frequency spectrum to calculate a parameter. The sample information prediction unit 22A predicts information about the prediction sample by applying the parameter calculated by the parameter calculation unit 200′ to the learning model.
Note that in the examples of
As described in detail above, in the first embodiment, a learning model for predicting information about a sample is prepared using a parameter that determines a property of each of a plurality of fitting functions corresponding to generation sources of a composite waveform fit to a frequency spectrum obtained by spectroscopic measurement of the sample using a terahertz wave, and a parameter obtained for a prediction sample is applied to the learning model, thereby predicting information about the prediction sample.
According to the first embodiment configured as described above, it is possible to predict information about a prediction sample using a learning model in which a feature quantity of a frequency spectrum obtained by performing spectroscopic measurement on a sample is represented by parameters that determine properties of a plurality of fitting functions. That is, a composite waveform of the plurality of fitting functions is fit to a frequency spectrum of a terahertz wave of the sample. For this reason, the parameters that determine the properties of the fitting functions reflect a property of the sample hidden in the frequency spectrum. Therefore, even when the frequency spectrum obtained by performing spectroscopic measurement on the sample using the terahertz wave is one in which a difference in property of the sample is unlikely to clearly appear as a feature of a waveform such as a peak of a specific frequency, information about the prediction sample can be accurately predicted by the learning model.
In addition, a learning model that can accurately predict information about the prediction sample in this way (for example, a pattern of a type of information to be predicted with respect to a liquid sample) and information about a liquid sample corresponding to an index value output from the learning model can contribute to standardization of a liquid state, which has been difficult to realize. Further, by accumulating the learning, the liquid state can be specified as a numerical value, and the liquid state can be standardized. In other words, an object used to rely on the sense, feelings, etc. of craftsmen can now be obtained as a standardized index without any individual difference, which can beneficially contribute to a liquid-using industry and further contribute to a functional design of the liquid, which has been difficult until now.
Next, a second embodiment of the invention will be described with reference to the drawings.
Also in the second embodiment, a parameter calculation apparatus 200 (see
In the second embodiment, the learning model storage unit 30B stores a learning model for predicting information about a sample using a frequency spectrum of a terahertz wave and a parameter of a fitting function. And, the prediction device 20B (sample information predicting unit 22B) applies a frequency spectrum obtained by spectroscopic measurement of a prediction sample using the terahertz wave and a parameter obtained for the prediction sample to the learning model, thereby predicting information about the prediction sample.
The learning data input unit 11B inputs a parameter and a frequency spectrum obtained for a learning sample and information about the learning sample as learning data for each of a plurality of learning samples. Here, each of the plurality of learning samples, the information about the learning sample (teacher data), and the parameter obtained for the learning sample is similar to that described in the first embodiment.
In addition, the frequency spectrum obtained for the learning sample is a frequency spectrum acquired by the frequency spectrum acquisition unit 201 when spectroscopic measurement is performed on the learning sample using the terahertz wave in the spectroscopic apparatus 300 of
The learning model creation unit 12B creates a learning model using the learning data (parameter, frequency spectrum, and teacher data) input by the learning data input unit 11B, and causes the learning model storage unit 30B to store the created learning model. Also in the second embodiment, the learning model creation unit 12B creates a learning model by applying a machine learning algorithm using a known neural network using the above-described learning data.
In addition, in an example illustrated in
The learning model creation unit 12B provides spectrum data related to a frequency spectrum of a learning sample acquired in the parameter calculation apparatus 200 and parameters related to n fitting functions used to generate a composite waveform to the input layer, and provides information about the learning sample (teacher data) to the output layer, thereby performing supervised learning. In this way, a neural network in which a degree of binding between nodes is weighted such that the same information as the teacher data is obtained as an index value from the output layer for the frequency spectrum and parameters input to the input layer is created.
Note that a configuration of the learning model created by the learning model creation unit 12B is not limited to the form of the neural network illustrated in
The prediction data input unit 21B inputs a parameter and a frequency spectrum obtained for the prediction sample as prediction data. Here, the parameter obtained for the prediction sample is the same as that described in the first embodiment. In addition, the frequency spectrum obtained for the prediction sample is a frequency spectrum acquired by the frequency spectrum acquisition unit 201 when spectroscopic measurement is performed on the prediction sample by the terahertz wave in the spectroscopic apparatus 300 of
The sample information prediction unit 22B applies the prediction data (parameter and frequency spectrum) input by the prediction data input unit 21B to the learning model stored in the learning model storage unit 30B, thereby predicting information about the prediction sample. In the case of mixing of the foreign substance described above as an example, information predicted with respect to the prediction sample corresponds to presence or absence of mixing of the foreign substance in a liquid used as the prediction sample, the type of the foreign substance mixed in the liquid, the amount of the foreign substance mixed in the liquid, etc.
Note that also in the second embodiment, a description has been given of an example in which the parameter calculation apparatus 200 and the sample analysis apparatus 100B are separately configured. However, as illustrated in
In the case of configuring the sample analysis apparatus 100B′ as illustrated in
In addition, at the time of prediction, the parameter calculation unit 200′ acquires a frequency spectrum obtained by spectroscopic measurement using a terahertz wave for the prediction sample as prediction data, and analyzes the acquired frequency spectrum to calculate a parameter. The sample information prediction unit 22B applies the frequency spectrum related to the prediction sample acquired by the parameter calculation unit 200′ and the parameter calculated by the parameter calculation unit 200′ to the learning model, thereby predicting information about the prediction sample.
Note that in the examples illustrated in
As described in detail above, in the second embodiment, a learning model for predicting information about the sample is created using the spectrum data of the frequency spectrum in addition to the parameter of the fitting function obtained by analyzing the frequency spectrum of the terahertz wave, and the parameter and the frequency spectrum obtained for the prediction sample is applied to the learning model, thereby predicting information about the prediction sample. According to the second embodiment configured as described above, since information about the prediction sample can be predicted using a larger feature quantity, information about the sample can be more accurately predicted.
Next, a third embodiment of the invention will be described with reference to the drawings.
In the third embodiment, a process of calculating a parameter by the parameter calculation apparatus 200 (parameter calculation unit 200′) is executed as a part of a neural network. Therefore, in the third embodiment, the parameter calculation apparatus 200 is not present separately from the sample analysis apparatus 100C. However, only a function of the frequency spectrum acquisition unit 201 of
In the third embodiment, the learning model storage unit 30C stores a learning model for obtaining a parameter of a fitting function from a frequency spectrum of a terahertz wave, and for predicting information about a sample using the obtained parameter. Then, the prediction device 20C (sample information prediction unit 22C) applies a frequency spectrum obtained by spectroscopic measurement of the prediction sample using the terahertz wave to the learning model to predict information about the prediction sample.
The learning data input unit 11C inputs a frequency spectrum obtained for a learning sample and information about the learning sample as learning data for each of a plurality of learning samples. Here, each of the plurality of learning samples, the information about the learning samples (teacher data), and the frequency spectra obtained for the learning samples are the same as that described in the second embodiment.
The learning model creation unit 12C creates a learning model using the learning data (frequency spectra and teacher data) input by the learning data input unit 11C, and causes the learning model storage unit 30C to store the created learning model. Also in the third embodiment, the learning model creation unit 12C creates a learning model by applying a machine learning algorithm using a known neural network using the above-described learning data.
In addition, in an example illustrated in
As described above, the partial learning model 200C including the first layer and the second layer of the intermediate layer is a model for predicting a plurality of parameters by applying the spectrum data of the frequency spectrum to the partial learning model 200C. Each node of a third layer is connected from each node of the second layer, and has a role for guiding an index value of the output layer from a value of each parameter output to each node of the second layer.
The learning model creation unit 12C provides the spectrum data related to the frequency spectrum of the learning sample acquired by the frequency spectrum acquisition unit 201 to the input layer, and provides the information about the learning sample (teacher data) to the output layer, thereby performing supervised learning. In this way, a neural network in which a degree of binding between nodes is weighted such that the same information as the teacher data is obtained as an index value from the output layer for the frequency spectrum input to the input layer is created.
Note that here, a description has been given of an example in which learning is performed for the entire neural network including the partial learning model 200C. However, after learning only the partial learning model 200C for predicting a parameter from the frequency spectrum in a separate frame, the learned partial learning model 200C may be incorporated in the entire neural network.
That is, to learn the partial learning model 200C, as illustrated in
Note that to increase the amount of learning data used when the partial learning model 200C is learned, it is possible to create another frequency spectrum obtained by adding noise to a frequency spectrum of an actually existing learning sample, and calculate a parameter by the parameter calculation apparatus 200 using the another frequency spectrum, thereby adding a set of the another frequency spectrum and the parameter (teacher data) as learning data.
When the entire learning model is learned by incorporating the partial learning model 200C learned as illustrated in
Note that a configuration of the learning model created by the learning model creation unit 12C (including a configuration of the partial learning model 200C) is not limited to the form of the neural network illustrated in
The prediction data input unit 21C inputs a frequency spectrum obtained for the prediction sample as prediction data. Here, the frequency spectrum obtained for the prediction sample is the same as that described in the second embodiment.
The sample information prediction unit 22C predicts information about the prediction sample by applying the prediction data (frequency spectrum) input by the prediction data input unit 21C to the learning model stored in the learning model storage unit 30C.
As described above in detail, according to the third embodiment, it is unnecessary to separately perform a process analyzing a frequency spectrum obtained by spectroscopic measurement of a sample using a terahertz wave to calculate a parameter and a process of performing prediction by applying a parameter to a neural network. That is, it becomes possible to perform all the processes related to prediction using a computer dedicated to the neural network.
Note that in the third embodiment, content described in the second embodiment may be applied in combination. In more detail, learning and prediction may be performed using a parameter extracted from the frequency spectrum in the compression layer of the neural network in addition to the parameter predicted in the partial learning model 200C.
In addition, in the first to third embodiments, descriptions have been given of examples in which the sample analysis apparatus 100A to 100C are configured as including both the learning devices 10A to 10C and the prediction devices 20A to 20C, respectively. However, the invention is not limited thereto. That is, the learning devices 10A to 10C and the prediction devices 20A to 20C may be configured as separate devices, and the prediction devices 20A to 20C may be used as the sample analysis apparatuses 100A to 100C.
In addition, in the above embodiments, as a property of a sample to be analyzed, a property associated with mixing of a foreign substance in a liquid has been given as an example. However, the invention is not limited thereto. That is, when a frequency spectrum of a terahertz wave can be obtained by performing spectroscopic measurement on a sample using the spectroscopic apparatus 300, and known information can be provided as teacher data for the sample, each one can be used as an object to be analyzed according to the embodiments.
For example, when a name of a sample, a chemical formula, etc. is used as “information about a learning sample (teacher data)” corresponding to one piece of learning data, and learning of a learning model is performed using the teacher data and a parameter and/or a frequency spectrum obtained for the learning sample as the learning data, it is possible to predict the sample name or the chemical formula from a parameters and/or a frequency spectrum obtained for an unknown sample.
In addition, when predetermined evaluation information for an index related to taste or smell of a beverage is used as “information about a learning sample (teacher data)” corresponding to one piece of learning data, and a learning model is learned using the teacher data and a parameter and/or a frequency spectrum obtained for the learning sample as learning data, it is possible to predict whether a beverage matching the index can be manufactured from a parameter and/or a frequency spectrum obtained for a manufactured beverage.
For example, taste or smell of a specific beverage (liquor, soft drink, milk, and other various beverages) is provided with a certain index for each manufacturer in many cases. That is, when such beverages are manufactured, evaluations are performed as uniformly as possible by human senses so as to satisfy a predetermined index. In this case, for example, when a learning model is created using evaluation information created by a person skilled in the evaluation (whether the index is matched, a degree of matching indicating a degree at which the index is matched, etc.) as teacher data, and a parameter and/or a frequency spectrum obtained for a manufactured beverage (prediction sample) is applied to the learning model, it is possible to obtain, from the learning model, an evaluation result corresponding to a degree at which the beverage matches the index of taste or smell regardless of the skilled person.
In addition, when a pH value of a sample is used as “information about a learning sample (teacher data)” corresponding to one piece of learning data, and a learning model is learned using the teacher data and a parameter and/or a frequency spectrum obtained for the learning sample as learning data, it is possible to predict the pH value from a parameter and/or a frequency spectrum obtained for an unknown sample. A normal pH meter corresponds to a contact type, and corresponds to a destructive inspection since liquid seeps out from a sensor unit. On the other hand, spectroscopic measurement corresponds to a non-destructive non-contact inspection. Therefore, according to the present embodiment, it is possible to predict a pH value by a non-destructive non-contact inspection of a sample.
In addition, when a micelle or vesicle state of a liquid (whether or not the liquid is in a micelle state, whether or not the liquid is in a vesicle state, critical micelle concentration, critical vesicle concentration, etc.) is used as “information about a learning sample (teacher data)” corresponding to one piece of learning data, and a learning model is learned using the teacher data and a parameter and/or a frequency spectrum obtained for the learning sample as learning data, it is possible to predict a micelle or vesicle state from a parameter and/or a frequency spectrum obtained for an unknown sample. An object conventionally empirically determined using various measuring devices in combination can be predicted from a result of spectroscopic measurement using a terahertz wave.
In addition, when a body liquid such as blood, urine, breast milk, or sweat is used as a sample, information about presence or absence of a disease or progress of a medical condition is used as “information about a learning sample (teacher data)” corresponding to one piece of learning data, and a learning model is learned using the teacher data and a parameter and/or a frequency spectrum obtained for the learning sample as learning data, it is possible to predict presence or absence of a disease or progress of a medical condition from a parameter and/or a frequency spectrum obtained for an unknown body liquid.
Possible examples are listed above. When the present embodiment can be applied, it is possible to predict various other properties.
In addition, in the first to third embodiments, the neural network is given as an example of the learning model. However, the learning model is not limited thereto. For example, the learning model may be configured as a regression model such as a linear regression, a logistic regression, or a support vector machine. Alternatively, the learning model may be configured as a tree model such as a decision tree or a random forest. Alternatively, the learning model may be configured as a Bayes model or a clustering model such as a k-nearest neighbor method.
In addition, in the first to third embodiments, a description has been given of an example of performing supervised learning. However, when a large amount of learning data can be prepared, unsupervised learning may be performed.
In addition, in the first to third embodiments, a description has been given of an example in which after performing the first fitting process on frequency spectra related to a plurality of samples, n center frequencies are specified to perform the second fitting process. However, only the first fitting process may be performed. Note that it is preferable to perform the second fitting processing as in the above-described embodiments since noise can be suppressed.
In addition, in the first to third embodiments, a normal distribution function (Gaussian function) is used as an example of the function used for fitting. However, implementation is allowed using a Lorentz function. In addition, it is possible to use a probability distribution function such as a Poisson distribution function (probability mass function or cumulative distribution function) or a chi-square distribution function (probability density function or cumulative distribution function) that is not centrally symmetric and is asymmetric, and it is possible to use another function whose waveform has a mountain shape. When a probability distribution function is used, fitting is performed using a value representing a property of a probability distribution (for example, a median or a mode of the amplitude, a frequency at which an amplitude value is obtained, a frequency width at which the amplitude is equal to or greater than a predetermined value or equal to or less than a predetermined value, etc.) as a parameter. In the case of using a mountain-shaped function, fitting is performed using the maximum amplitude corresponding to an apex, a frequency at which the maximum amplitude is obtained, a frequency width at which the amplitude is equal to or greater than a predetermined value or equal to or less than a predetermined value, etc. as a parameter.
In addition, in the first to third embodiments, a description has been given of an example in which an absorbance is used as a property value of a terahertz wave signal, and a frequency spectrum representing an absorbance with respect to a frequency is obtained. However, another property value such as a transmittance may be used.
In addition, each of the first to third embodiments described above is merely an example of a concrete embodiment for carrying out the invention, and the technical scope of the invention should not be interpreted in a limited manner by the embodiment. That is, the invention can be implemented in various forms without departing from a gist or a main feature thereof.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2018-126551 | Jul 2018 | JP | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2018/029351 | 8/6/2018 | WO | 00 |
