JP7801766B2 - Chemical component determination device, chemical component determination method, chemical component determination program, and recording medium - Google Patents

Description

The present invention relates to a chemical component determination device, a chemical component determination method, a chemical component determination program, and a recording medium.

Prior art for determining the odor contained in gas involves using a PID (Photo Ionization Detector) and a semiconductor gas sensor as detectors, and separating and monitoring odor components using a chromatographic column (see, for example, Patent Document 1). The technology disclosed in Patent Document 1 inputs the obtained waveform into a neural network to determine the odor.

Patent Publication No. 2021-139914

The technology disclosed in Patent Document 1 determines the quality and intensity of an odor using separate neural networks. This means that this technology cannot determine both the quality and intensity of an odor at the same time. Furthermore, the intensity output here is a stepped (discontinuous) discrete value, not a continuous value like concentration.

The present invention was made in light of these circumstances, and its purpose is to simultaneously determine the chemical components and concentrations contained in a sample using the sensor signal of the sample detected by a chemical sensor.

In order to solve the above problems, a chemical component determination device according to one embodiment of the present invention includes an input layer to which sensor signals detected by a plurality of chemical sensors from a sample containing n candidate chemical components are input, an intermediate layer, an output layer that performs machine learning to obtain a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample and outputs the result, and a concentration calculation unit that calculates the concentration s _t (t=1 to n) of the chemical component associated with the sample,
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
is.

In one embodiment, the chemical sensor may be a gas sensor.

In some embodiments, the chemical sensor may detect the odor of the sample.

In one embodiment, the chemical component determination device may further include a pre-processing unit that logarithms and/or standardizes the sensor signal values detected by the chemical sensor.

In one embodiment, the multiple chemical sensors may be selected from the candidate chemical sensors based on the largest standard deviation of weights from the input layer to the hidden layer nodes.

Another aspect of the present invention is a chemical component determination method for determining the chemical components and concentrations of the chemical components of a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit, the method including the steps of: inputting sensor signals detected by a plurality of chemical sensors from a sample containing n candidate chemical components to the input layer; outputting a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample from the output layer through machine learning; and calculating the concentration s _t (t=1 to n) of the chemical component associated with the sample using the concentration calculation unit.
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
is.

Yet another aspect of the present invention is a chemical component determination program that causes a computer to execute a method for determining the chemical components and concentrations of the chemical components of a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit, the program including the steps of: inputting sensor signals detected by a plurality of chemical sensors from a sample containing n candidate chemical components to the input layer; outputting a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample from the output layer through machine learning; and calculating the concentration s _t (t=1 to n) of the chemical component associated with the sample using the concentration calculation unit.
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
The method is performed by a computer.

Yet another aspect of the present invention is a recording medium storing a program for causing a computer to execute a method for determining chemical components and the concentrations of the chemical components of a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit, the method including the steps of: inputting sensor signals detected by a plurality of chemical sensors from a sample containing n candidate chemical components to the input layer; outputting a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample from the output layer through machine learning; and calculating the concentration s _t (t=1 to n) of the chemical component associated with the sample using the concentration calculation unit.
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
A program for causing a computer to execute the method is recorded.

In addition, any combination of the above components, and any transformation of the present invention into an apparatus, method, system, recording medium, computer program, etc., are also valid aspects of the present invention.

According to the present invention, the chemical components contained in a sample and their concentrations can be determined simultaneously from the sensor signal of the sample detected by the chemical sensor.

1 is a functional block diagram of a chemical component determining device according to a first embodiment; FIG. 10 is a functional block diagram of a chemical component determining device according to a fourth embodiment. 13 is a flowchart showing the process of a chemical component determination method according to a sixth embodiment. FIG. 1 is a schematic diagram of a verification experiment system. 10 is a graph showing the sensor response to odor 2 in experiment 1.

The present invention will be described below based on preferred embodiments with reference to the drawings. In the embodiments and variations, identical or equivalent components and parts will be given the same reference numerals, and duplicate explanations will be omitted where appropriate. The dimensions of the components in each drawing will be enlarged or reduced as appropriate to facilitate understanding. Some components that are not important for explaining the embodiments will be omitted from the drawings. Furthermore, terms including ordinal numbers such as first and second are used to describe various components, but these terms are used only for the purpose of distinguishing one component from another, and do not limit the components.

[First embodiment]
FIG. 1 shows a functional block diagram of a chemical component determination device 1 according to a first embodiment. The chemical component determination device 1 includes an input layer 10, an intermediate layer 20, an output layer 30, and a concentration calculation unit 40. The input layer 10, the intermediate layer 20, and the output layer 30 form a neural network. In the following, unless there is confusion, the same symbol may be used to represent data and nodes in a neural network. For example, the same symbol _x1 may be used to represent data _x1 and node _x1 .

In the example shown in Figure 1, a sample with n candidate chemical components is being considered. m different chemical sensors 101, 102, 103, ..., 10m are connected to the chemical component determination device 1. It is desirable that each of the chemical sensors 101, 102, 103, ..., 10m has a certain sensitivity to the above components 1, 2, 3, ..., n.

Sensor signals _x1 , _x2 , _x3 , ..., _xm detected by chemical sensors 101, 102, 103, ..., 10m are input to the input layer 10. Note that in Fig. 1, the node indicated by _x0 in the input layer 10 is a bias node. Excluding the bias node, the number m of nodes in the input layer is equal to the number m of chemical sensors (i.e., the number m of sensor signals from the chemical sensors).

The sensor signals _x1 , _x2 , _x3 , ..., _xm input to the input layer 10 are passed to each node _b1 , _b2 , _b3 , ..., _bk in the hidden layer 20. The number of layers in the hidden layer 20 and the number of nodes in each hidden layer (k in this example) may be selected arbitrarily and suitably depending on the accuracy of machine learning and the calculation cost. Each node in the hidden layer 20 may convert the values _b1 , _b2 , _b3 , ..., _bk into _z1 , _z2 , _z3 , ..., _zk , respectively, using an activation function such as ReLU. _z0 in the figure is a bias node.

The output layer 30 outputs chemical component values _y1 , _y2 , _y3 , ..., yn and a reference value y0 associated with the sample, which are machine-learned from the sensor signals _x1 , _x2 , _x3 , _... , _xm input to the input layer _10. The reference value _y0 will be described later. Note that the number of nodes in the output layer (n+1) is the number n of candidate chemical components plus 1.

The concentration calculation unit 40 calculates concentrations s ₁ , s ₂ , s ₃ , ..., s _n of chemical components associated with the sample, where the chemical component values y _t (t=1 to n) and concentrations s _t (t=1 to n) associated with the sample satisfy the following:
Σ _i=0 ⁿ y _i =1 (1)
_yt / _y0 = _st (2)
It should be noted that y ₀ and y _t (t=1 to n) can be obtained from (1) and (2) as follows:
y ₀ =1/(1+Σ _t=1 ⁿ s _t ) (3)
y _t = s _t /(1+Σ _t=1 ⁿ s _t ) (4)

In this way, the chemical component determination device 1 simultaneously determines the chemical component values _yt (t=1 to n) and the concentrations _st (t=1 to n) of the chemical components contained in the sample from the sensor signals _x1 , _x2 , _x3 , ..., _xm of the sample detected by the chemical sensors 101, 102, 103, ..., 10m.

One of the features of this embodiment is that it is a neural network that not only performs classification and identification by placing an activation function in the output layer, but also simultaneously performs regression. In conventional general regression models, the concentration values input as training data correspond directly to the concentration values output. In this case, it is not possible to identify the type of chemical component. In contrast, this embodiment has the significant advantage of being able to simultaneously identify the type of chemical component and predict the concentration by adding a new reference value node to the output layer.

This embodiment makes it possible to realize a device that simultaneously determines the chemical components and concentrations contained in a sample using the sensor signal of the sample detected by a chemical sensor.

Second Embodiment
In the second embodiment, the sample is a gas. In this case, the candidate chemical components are gas components, such as candidate 1 being ammonia, candidate 2 being hydrogen sulfide, candidate 3 being LP gas, ..., and candidate n being ethanol. The sample gas may be a pure gas containing only one of these gas components, or a mixed gas containing multiple of these gas components. In this case, chemical sensors 101, 102, 103, ..., 10m may be gas sensors. For example, the gas sensors may be general commercially available gas sensors. It is desirable that each of gas sensors 101, 102, 103, ..., 10m has a certain sensitivity to sample gas components 1, 2, 3, ..., n.

According to this embodiment, the component gases contained in the sample gas and their concentrations can be determined simultaneously using the sensor signal of the sample gas detected by the gas sensor.

[Third embodiment]
In the third embodiment, the chemical sensors detect the odor of the sample. For example, in the example of the second embodiment, each of the chemical sensors 101, 102, 103, ..., 10m connected to the chemical component determination device 1 may detect the odor of ammonia, hydrogen sulfide, LP gas, ..., ethanol.

This embodiment can be effectively applied to fields where odor identification is required, such as disease screening using breath measurement, evaluating food freshness using odor measurement, and indoor air quality evaluation.

[Fourth embodiment]
2 shows a functional block diagram of a chemical component determination apparatus 2 according to a fourth embodiment. The chemical component determination apparatus 2 includes an input layer 10, an intermediate layer 20, an output layer 30, a concentration calculation unit 40, and a preprocessing unit 50. That is, the chemical component determination apparatus 2 further includes the preprocessing unit 50 in addition to the components of the chemical component determination apparatus 1 in FIG. The remaining configuration of the chemical component determination apparatus 2 is the same as that of the chemical component determination apparatus 1. The following description will focus on the differences from the chemical component determination apparatus 1.

The pre-processing unit 50 logarithms and/or standardizes the sensor signal values detected by the chemical sensors 101, 102, 103, ..., 10m.

Some chemical sensors 101, 102, 103, ..., 10m exhibit extremely large sensor response values for only a portion of the data. Using such data as is can result in large calculation errors. However, by converting the original data to a logarithm, for example, using 10 as the base, it is possible to mitigate extremely large response values.

Furthermore, the average response of chemical sensors 101, 102, 103, ..., 10m varies from sensor to sensor. In this case, the output obtained may be biased toward sensors with large responses. In response to this, for example, by converting the original data into standardized data with a mean of 0 and a standard deviation of 1, it is possible to mitigate the bias of extremely large responses.

By converting the data obtained from the chemical sensor into logarithmic data, standardized data, or both, the machine learning results can be made more accurate. In other words, this embodiment makes it possible to more accurately determine the chemical components and their concentrations contained in a sample.

Fifth Embodiment
In the fifth embodiment, a chemical sensor is selected from among candidate chemical sensors based on the largest standard deviation of the weights from the input layer to the nodes in the intermediate layer. For example, a neural network is first constructed using a large number of candidate chemical sensors, and then the weights of each sensor data from the input layer to the nodes in the intermediate layer are compared. As a result, a predetermined number of chemical sensors with larger standard deviations of the weights can be selected as the chemical sensors to be actually used, and the remaining chemical sensors can be left unused. In this case, the larger the standard deviation of the weights, the greater the contribution of the corresponding chemical sensor. Therefore, by selecting chemical sensors with larger standard deviations of the weights as the chemical sensors to be actually used, the number of chemical sensors can be reduced without significantly sacrificing accuracy.

This embodiment allows for a reduction in the number of chemical sensors, making the overall configuration more compact and reducing costs.

Sixth Embodiment
3 shows a process flow of a chemical component determination method according to a sixth embodiment. This method determines the chemical components and concentrations of the chemical components in a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit. This method includes steps S10, S20, and S30.

In step S10, the method inputs sensor signals x ₁ , x ₂ , x ₃ , . . . , x _m detected by a plurality of chemical sensors from a sample containing n candidate chemical components to the input layer.

In step 20, the method outputs from the output layer the chemical component values _y1 , _y2 , _y3 , ..., _yn and the reference value _y0 associated with the sample, which have been machine-learned from the sensor signals _x1 , _x2 , _x3 , ..., _xm input in step S10.

In step S30, the method uses a concentration calculation unit to calculate concentrations s ₁ , s ₂ , s ₃ , ..., s _n of chemical components associated with the sample, where the chemical component values y _t (t=1 to n) and concentrations s _t (t=1 to n) associated with the sample satisfy the following:
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
y ₀ =1/(1+Σ _t=1 ⁿ s _t )
however,
y _t = s _t /(1+Σ _t=1 ⁿ s _t )

According to this embodiment, the chemical component determination device can be used to simultaneously determine the chemical components contained in a sample and the concentrations of those chemical components.

[Seventh embodiment]
This embodiment is a chemical component determination program that causes a computer to execute a method for determining the chemical components and concentrations of the chemical components in a sample using a chemical component determination device that includes an input layer, an intermediate layer, an output layer, and a concentration calculation unit. This program causes a computer to execute a method that includes steps S10, S20, and S30.

In step S10, the method inputs sensor signals x ₁ , x ₂ , x ₃ , . . . , x _m detected by a plurality of chemical sensors from a sample containing n candidate chemical components to an input layer.

In step 20, the method outputs from the output layer the chemical component values _y1 , _y2 , _y3 , ..., _yn and the reference value _y0 associated with the sample, which have been machine-learned from the sensor signals _x1 , _x2 , _x3 , ..., _xm input in step S10.

In step S30, the method uses a concentration calculation unit to calculate concentrations s ₁ , s ₂ , s ₃ , ..., s _n of chemical components associated with the sample, where the chemical component values y _t (t=1 to n) and concentrations s _t (t=1 to n) associated with the sample satisfy the following:
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
however,
y ₀ =1/(1+Σ _t=1 ⁿ s _t )
y _t = s _t /(1+Σ _t=1 ⁿ s _t )

According to this embodiment, a program can be implemented in software that executes the process of simultaneously determining the chemical components contained in a sample and the concentrations of those chemical components.

Eighth Embodiment
This embodiment is a recording medium having a chemical component determination program recorded thereon that causes a computer to execute a method for determining the chemical components and concentrations of the chemical components in a sample using a chemical component determination device that includes an input layer, an intermediate layer, an output layer, and a concentration calculation unit. The program recorded on this recording medium causes the computer to execute a method that includes steps S10, S20, and S30.

In step S10, the method inputs sensor signals x ₁ , x ₂ , x ₃ , . . . , x _m detected by a plurality of chemical sensors from a sample containing n candidate chemical components to an input layer.

In step 20, the method outputs from the output layer the chemical component values _y1 , _y2 , _y3 , ..., _yn and the reference value _y0 associated with the sample, which have been machine-learned from the sensor signals _x1 , _x2 , _x3 , ..., _xm input in step S10.

In step S30, the method uses a concentration calculation unit to calculate concentrations s ₁ , s ₂ , s ₃ , ..., s _n of chemical components associated with the sample, where the chemical component values y _t (t=1 to n) and concentrations s _t (t=1 to n) associated with the sample satisfy the following:
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
however,
y ₀ =1/(1+Σ _t=1 ⁿ s _t )
y _t = s _t /(1+Σ _t=1 ⁿ s _t )

According to this embodiment, software that implements a program that executes a process to simultaneously determine the chemical components contained in a sample and the concentrations of those chemical components can be recorded on a recording medium.

[Verification experiment]
The inventors conducted the following five experiments to verify the effects of this technology. Figure 4 shows a schematic diagram of the experimental system used. In these experiments, the chemical component determination device was implemented using a notebook PC. Eight gas sensor modules manufactured by Figaro Giken Co., Ltd. were used as chemical sensors. The specifications are shown in Table 1.
[Experiment 1]
The samples used were the substances and concentrations shown in Table 2. In the following, unless there is any confusion, odors may be referred to as odor 1, odor 2, ..., odor 9, etc. instead of substance 1, substance 2, ..., substance 9.
In this experiment, the sample was adjusted to the concentration shown in Table 2 by adding 60% RH (water vapor) to the sample, and the sample gas was constantly flowing at a total flow rate of 500 mL/min. The gas sensor system equipped with the eight sensors described above aspirated 200 mL/min from the flow path and exposed the gas to the sensors. The resistance value of the sensor signal was obtained every 5 seconds.

Figure 5 shows an example of the sensor response to odor 2. The 13 data points from the last minute of the flow section for each gas concentration were used to determine Rg, and the data immediately before the introduction of each gas concentration was used as Ra, resulting in 13 sensor response values Ra/Rg for each gas concentration.

The neural network used in this experiment is as follows: The number of nodes in the input layer was 8 (the number of sensors, excluding bias nodes). There were two intermediate layers, with the first layer having 20 nodes (excluding bias) and an activation function of ReLU, and the second layer having 30 nodes (excluding bias) and the same activation function of ReLU. The number of nodes in the output layer was 10 (the number of odor components plus 1), the activation function was softmax, the number of epochs was 10,000, and the batch size was 100.

All 429 data points were randomly divided into an 8:2 ratio, with the former used as training data and the latter as test data, and the test data was judged using a classifier created from the training data.

The following three criteria were set as the criteria for judgment.
1) The gas species with the maximum concentration of the output value matches the original gas.
2) Even if it is output that a different gas species coexists, the concentration of the different gas species is 1/10 or less of the original gas species.
3) The concentration of the original gas species is within ±20% of the correct value.
1) is a determination of whether the odor type is correct, and corresponds to the classification of odor types. Hereinafter, the rate of success of 1) will be referred to as the "gas type determination rate." 2) is a criterion indicating whether different odor types are present at concentrations that cannot be ignored. 3) is a criterion indicating whether the concentration values of the original odor types are correctly reflected. In particular, 3) within ±20% is a value that is often used as an indicator of the accuracy of the gas sensor itself.

The results of the evaluation are shown in Table 3.

The gas type identification rate, which corresponds to the classification of gas types, was 96.5%, indicating that identification was generally accurate. The concentration prediction accuracy, which corresponds to the regression of gas concentration, was 53.5%, and again, roughly more than half of the data met the gas sensor accuracy index. The concentration prediction accuracy was lower than the gas type identification rate because a large proportion did not meet criterion 3). The test data was divided into those that met 3) and those that did not, and the average concentration was calculated by excluding the top and bottom 20% of the data for each. The former was 2.3 ppm and the latter was 0.39 ppm. Generally, as the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, which is thought to have had an impact.

Furthermore, four concentrations of each odor (excluding odor 7) were evaluated. Next, fictitious sensor responses corresponding to the median values of the sensor responses at each concentration were created, and these were assumed to be the sensor responses of the median concentrations. We then input these values into the discriminator and examined whether the output concentration values matched the median concentrations. As an example, Table 4 shows the averages of the sensor responses x1 to x8 for each concentration of odor 4 used in the test data.

Table 5 shows the concentrations of odor 4 corresponding to the intermediate values and the sensor response values x1 to x8, assuming that the sensor response is proportional in the range of odor 4 concentrations from 0.18 ppm to 0.72 ppm.

Table 6 shows the results of substituting the values in Table 5 into the discriminator, and then substituting the output y values into equations (1) to (4) to convert them into density values. It can be seen that all three points satisfy all of the criteria 1), 2), and 3).

[Experiment 2]
The measurement data was the same as in Experiment 1, but we attempted to improve the results by changing some of the neural network conditions as follows.

The number of nodes in the input layer was 8 (the number of sensors, excluding bias nodes), the number of nodes in the output layer was 10 (the number of odor components plus 1), and the activation function was softmax, which was the same as in Experiment 1. The only changes were made to the two intermediate layers, with the number of nodes in both the first and second layers being 64 (excluding bias nodes), the number of epochs being 5000, and the batch size being 32 (the activation function was also ReLU).

The sensor response values were preprocessed by logarithmizing them to the base 10 before being input to the input layer.

The results are shown in Table 7.

The gas type identification rate, which corresponds to the classification of gas types, was 95.3% (application rate for criterion 1), and the concentration prediction accuracy, which corresponds to the regression of gas concentrations, was 41.9% (application rate for criteria 1), 2), and 3). Although some of the conditions of the neural network developed in this way were changed and preprocessing was performed, results roughly equivalent to those in Experiment 1 were obtained. Furthermore, by performing logarithmic conversion to base 10, the amount of calculations can be reduced.

The reason why the concentration prediction accuracy is lower than the gas type identification rate is because the proportion that does not meet criterion 3) is high. The test data was divided into those that met criterion 3) and those that did not, and the average concentration was calculated by excluding the top and bottom 20% of the data for each. The former was 2.6 ppm, and the latter was 0.51 ppm. Generally, as the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, so this is thought to have had an impact.

Furthermore, in Experiment 2, we also created a fictitious sensor response value corresponding to the median value of the sensor response values at each concentration, assumed this to be the sensor response value of the median concentration, and examined whether the concentration value output from the classifier matched the median concentration. However, in this experiment, we examined all odors except for odor 7, which only had one concentration. The judgment results are shown in Table 8.

The gas type identification rate, which corresponds to the gas type classification, was 91.7% (applicable rate for criterion 1), and the concentration prediction accuracy, which corresponds to the gas concentration regression, was 50.0% (applicable rate for criteria 1), 2), and 3), which are roughly the same results as those in Table 7.

In this case too, the reason why the concentration prediction accuracy is lower than the gas type identification rate is because the proportion that does not meet criterion 3) is high. When data was divided into those that met criterion 3) and those that did not, and the average concentration was calculated by excluding the top and bottom 20% of the data for each, the former was 2.7 ppm and the latter was 0.52 ppm. As the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, so it is thought that this also affected the acquisition of the median value.

[Experiment 3]
The measurement data was the same as in Experiment 1, and the neural network conditions were the same as in Experiment 2.

The sensor response values were standardized as a preprocessing step before being input to the input layer.

The results are shown in Table 9.

The gas type identification rate, which corresponds to the classification of gas types, was 96.5% (application rate for criterion 1), and the concentration prediction accuracy, which corresponds to the regression of gas concentrations, was 64.0% (application rate for criteria 1), 2), and 3). By changing some of the conditions of the neural network developed in this way and performing preprocessing, it can be seen that the results were improved compared to Experiment 1.

The reason why the concentration prediction accuracy is lower than the gas type identification rate is because the proportion that does not meet criterion 3) is high. The test data was divided into those that met criterion 3) and those that did not, and the average concentration was calculated by excluding the top and bottom 30% of the data for each. The former was 1.1 ppm, and the latter was 0.46 ppm. Generally, as the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, so this is thought to have had an impact.

Furthermore, in Experiment 3, fictitious sensor response values corresponding to the median values of the sensor response values at each concentration were created, and these were assumed to be the sensor response values of the median concentration. It was then input into the discriminator, and it was examined whether the output concentration value matched the median concentration. However, in this experiment, all odors were examined except for odor 7, which had only one concentration. The judgment results are shown in Table 10.

The gas type identification rate, which corresponds to the gas type classification, was 100% (applicable rate for criterion 1), and the concentration prediction accuracy, which corresponds to the gas concentration regression, was 54.2% (applicable rate for criteria 1), 2), and 3), which are roughly the same results as those in Table 9.

In this case too, the reason why the concentration prediction accuracy is lower than the gas type identification rate is because the proportion that does not meet criterion 3) is high. When data was divided into those that meet criterion 3) and those that do not, and the average concentration was calculated by excluding the top and bottom 20% of the data for each, the former was 2.2 ppm and the latter was 0.42 ppm. As the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, so it is thought that this also affected the acquisition of the median value.

[Experiment 4]
The measurement data was the same as in Experiment 1, but we attempted to improve the results by changing some of the neural network conditions as follows.

The number of nodes in the input layer was 8 (the number of sensors, excluding bias nodes), the number of nodes in the output layer was 10 (the number of odor components plus 1), and the activation function was softmax, which was the same as in Experiment 1. The only changes were made to the two intermediate layers, with the number of nodes in both the first and second layers being 64 (excluding bias nodes), the number of epochs being 5000, and the batch size being 32 (the activation function was also ReLU).

The sensor response values were preprocessed by logarithmizing them to the base 10 and then standardizing them before being input to the input layer.

The results are shown in Table 11.

The gas type identification rate, which corresponds to the classification of gas types, was 100% (application rate for criterion 1), and the concentration prediction accuracy, which corresponds to the regression of gas concentrations, was 76.7% (application rate for criteria 1), 2), and 3). By changing some of the conditions of the neural network developed in this way and performing preprocessing, it can be seen that the results were improved compared to Experiment 1.

The reason why the concentration prediction accuracy is lower than the gas type identification rate is because the proportion that does not meet criterion 3) is high. The test data was divided into those that met criterion 3) and those that did not, and the average concentration was calculated by excluding the top and bottom 20% of the data for each. The former was 1.6 ppm, and the latter was 0.20 ppm. Generally, as the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, so this is thought to have had an impact.

Furthermore, in Experiment 4, fictitious sensor response values corresponding to the median values of the sensor response values at each concentration were created, and these were assumed to be the sensor response values of the median concentration. It was then examined whether the concentration values output from the classifier matched the median concentration. However, in this experiment, all odors were examined except for odor 7, which only had one concentration. The judgment results are shown in Table 12.

The gas type identification rate, which corresponds to the gas type classification, was 100% (applicable rate for criterion 1), and the concentration prediction accuracy, which corresponds to the gas concentration regression, was 66.7% (applicable rate for criteria 1), 2), and 3), which are roughly the same results as those in Table 6.

In this case too, the reason why the concentration prediction accuracy is lower than the gas type identification rate is because the proportion that does not meet criterion 3) is high. When data was divided into those that met criterion 3) and those that did not, and the average concentration was calculated by excluding the top and bottom 20% of the data for each, the former was 1.9 ppm and the latter was 0.34 ppm. As the concentration decreases, the sensor signal tends to become smaller and accuracy tends to decrease, so it is thought that this also affected the acquisition of the median value.

[Experiment 5]
The conditions for the developed neural network were the same as in Experiment 4, and the measurement data were the same as in Experiment 1. However, this experiment differed from Experiments 1 and 4 in that five out of eight sensors were selected for execution.

The five selected sensors were TGS2600, TGS2602, TGS2603, TGS2620, and TGS2444, which had large standard deviations of weights from each node in the input layer to the 64 nodes in one hidden layer in the neural network of Experiment 4. Table 13 shows the judgment results.

The gas type determination rate, which corresponds to the classification of gas types, was 96.5% (application rate for criterion 1), and the concentration prediction accuracy, which corresponds to the regression of gas concentrations, was 74.4% (application rate for criteria 1), 2), and 3), which are comparable to the determination results obtained with the eight sensors in Table 6.

Furthermore, in Experiment 5, a fictitious sensor response value corresponding to the median value of the sensor response values at each concentration was created, and this was assumed to be the sensor response value of the median concentration. It was then input to the discriminator, and it was examined whether the output concentration value matched the median concentration. In this experiment, all odors were examined except for odor 7, which only had one concentration. The judgment results are shown in Table 14.

The gas type determination rate, which corresponds to the classification of gas types, was 100% (application rate for criterion 1), and the concentration prediction accuracy, which corresponds to the regression of gas concentrations, was 75% (application rate for criteria 1), 2), and 3), which are comparable to the determination results obtained with the eight sensors in Table 7.

The above has been a description based on embodiments of the present invention. These embodiments are merely examples, and those skilled in the art will understand that various modifications are possible in the combination of each component and each treatment process, and that such modifications are also within the scope of the present invention.

It can be widely used in industrial fields where it is necessary to determine chemical components. It is particularly useful in fields where there is a need to identify odors, such as disease screening using breath measurement, assessing food freshness using odor measurement, and evaluating indoor air quality.

1. Chemical component determination device,
2. Chemical component determination device,
10. Input layer,
20. Middle class,
30: Output layer unit,
40...concentration calculation section,
50: Pre-processing section
101: Chemical sensor
102: Chemical sensor
103: Chemical sensor
10m... Chemical sensor,
S10: A step of inputting a sensor signal to the input layer;
S20: A step of outputting the reference value and the chemical component value from the output layer;
S30: A step of calculating the concentration of the chemical component using the concentration calculation unit.

Claims (8)

an input layer to which sensor signals detected by a plurality of chemical sensors from a sample having n candidate chemical components are input;
The middle class and
an output layer that performs machine learning to output a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample;
a concentration calculation unit for calculating the concentration s _t (t=1 to n) of a chemical component associated with the sample;
Including,
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
A chemical component determining device characterized by: The chemical component determination device described in claim 1, characterized in that the chemical sensor is a gas sensor. The chemical component determination device described in claim 2, characterized in that the chemical sensor detects the odor of the sample. The chemical component determination device according to claim 1, further comprising a pre-processing unit that logarithmizes and/or standardizes the sensor signal value detected by the chemical sensor. The chemical component determination device described in claim 1, characterized in that the multiple chemical sensors are selected from candidate chemical sensors based on the largest standard deviation of the weights from the input layer to the nodes in the hidden layer. A method for determining chemical components and concentrations of the chemical components of a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit, comprising:
inputting sensor signals detected by a plurality of chemical sensors from a sample having n candidate chemical components into the input layer;
a step of machine learning and outputting a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample from the output layer;
calculating, using a concentration calculation unit, concentrations s _t (t=1 to n) of chemical components associated with the sample;
Including,
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
A chemical component determination method characterized by: A program for causing a computer to execute a method for determining chemical components and concentrations of the chemical components of a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit, the program comprising:
The method comprises:
inputting sensor signals detected by a plurality of chemical sensors from a sample having n candidate chemical components into the input layer;
a step of machine learning and outputting a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample from the output layer;
calculating, using a concentration calculation unit, concentrations s _t (t=1 to n) of chemical components associated with the sample;
Including,
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
A chemical component determination program characterized by: A recording medium storing a program for causing a computer to execute a method for determining chemical components and concentrations of the chemical components of a sample using a chemical component determination device including an input layer, an intermediate layer, an output layer, and a concentration calculation unit,
The method comprises:
inputting sensor signals detected by a plurality of chemical sensors from a sample having n candidate chemical components into the input layer;
a step of machine learning and outputting a reference value y ₀ and a chemical component value y _t (t=1 to n) associated with the sample from the output layer;
calculating, using a concentration calculation unit, concentrations s _t (t=1 to n) of chemical components associated with the sample;
Including,
Σ _i=0 ⁿ y _i =1
y _t / y ₀ =s _t
A recording medium characterized by: