JP7408136B2 - Gas analysis systems and programs - Google Patents

Description

The present invention relates to a gas analysis system and program for analyzing gas.

Conventional gas analysis systems are known that include a gas chromatograph that generates a gas chromatogram from gas and a computer that uses a neural network to determine the odor of the gas based on the gas chromatogram generated by the gas chromatograph (for example, , see Patent Document 1).

JP2019-056664A

However, in conventional gas analysis systems, the detection intensities at all sampling times in the gas chromatogram are input into the neural network, resulting in a large number of detection intensities that are unnecessary for gas odor determination to be input into the neural network. As a result, there is a problem in that the accuracy rate of gas odor determination is not high.

Therefore, an object of the present invention is to provide a gas analysis system and program that can improve the accuracy rate of gas analysis using a neural network.

The gas analysis system of the present invention generates simplified data as simplified data by simplifying gas chromatogram data of a sample gas using information indicating the retention time of each gas component. a simplification unit; and a gas analysis unit that executes an analysis of the sample gas based on the simplified data using a gas analysis neural network as a neural network that executes an analysis of the sample gas based on the simplified data. The simplified data includes only the detection intensity corresponding to the retention time indicated in the information as the detection intensity based on the gas chromatogram.

With this configuration, the gas analysis system of the present invention allows the gas analysis neural network to perform supervised learning using simplified data that does not include data of low importance in the analysis of sample gas by the gas analysis neural network. Therefore, it is possible to improve the accuracy rate of gas analysis by the gas analysis neural network.

In the gas analysis system of the present invention, the simplified data may include a relative detection intensity based on a plurality of detection intensities included in the gas chromatogram, as the detection intensity based on the gas chromatogram.

With this configuration, the gas analysis system of the present invention can generate simplified data that reduces the influence of baseline variations in gas chromatograms that depend on the individual gas chromatographs that generate gas chromatograms for sample gas. Therefore, the accuracy rate of gas analysis by the gas analysis neural network can be improved.

The gas analysis system of the present invention automatically corrects the simplified data into the appropriate simplified data using a data correction neural network as a neural network that automatically corrects the simplified data into the appropriate simplified data. The gas analysis unit may include a data correction unit for correction, and the gas analysis unit may execute analysis of the sample gas by the gas analysis neural network based on the simplified data corrected by the data correction unit.

With this configuration, the gas analysis system of the present invention reduces the influence of variation in retention time depending on the individual of the gas chromatograph that generates the gas chromatogram for the sample gas on the simplified data by correction using the data correction neural network. Therefore, it is possible to improve the accuracy rate of gas analysis by the gas analysis neural network.

The gas analysis system of the present invention includes a gas analysis learning unit that causes the gas analysis neural network to perform learning, and the gas analysis learning unit is configured to simplify data of the gas chromatogram of the sample gas by the data simplification unit. The gas analysis neural network may be caused to perform supervised learning using a combination of the simplified data corrected by the data correction unit and the results of human analysis of the sample gas.

With this configuration, the gas analysis system of the present invention uses simplified data in which the influence of variation in retention time depending on the individual gas chromatograph is reduced by correction by the data correction neural network to train the gas analysis neural network. Since the presence learning is executed, it is possible to improve the accuracy rate of gas analysis by the gas analysis neural network.

The program of the present invention generates simplified data as simplified data by simplifying data of a gas chromatogram of a sample gas using information indicating the retention time of each gas component. The simplified data includes only the detection intensity corresponding to the retention time indicated in the information as the detection intensity based on the gas chromatogram.

With this configuration, the computer executing the program of the present invention uses simplified data that does not include data of low importance in the analysis of the sample gas by the gas analysis neural network as the neural network that performs the analysis of the sample gas. Since it is possible to cause the gas analysis neural network to perform supervised learning, it is possible to improve the accuracy rate of gas analysis by the gas analysis neural network.

The gas analysis system and program of the present invention can improve the accuracy rate of gas analysis using a neural network.

1 is a configuration diagram of a gas analysis system according to an embodiment of the present invention. 2 is a block diagram of the computer shown in FIG. 1. FIG. 3 is a diagram showing an example of the gas dictionary shown in FIG. 2. FIG. 3 is a flowchart of the operation of the computer shown in FIG. 2 when analyzing sample gas. 5 is a flowchart of the data simplification process shown in FIG. 4. 2 is a diagram showing an example of a gas chromatogram generated by the gas chromatograph shown in FIG. 1. FIG. 2 is a diagram showing an example of various data generated based on data of a gas chromatogram generated by the gas chromatograph shown in FIG. 1. FIG. 6 is a diagram showing an example of bar graph data generated in the data simplification process shown in FIG. 5. FIG. 6 is a diagram showing an example of simplified data generated in the data simplification step shown in FIG. 5. FIG. FIG. 5 is a diagram showing an example of the data correction process shown in FIG. 4; 5 is a diagram showing an example of the gas analysis process shown in FIG. 4. FIG. 3 is a flowchart of a data modification learning method for causing the data modification neural network shown in FIG. 2 to perform learning. 13 is a diagram showing a part of an example of bar graph data displayed in the data correction learning method shown in FIG. 12. FIG. (a) A diagram showing an example of simplified data before modification by the data modification learning method shown in FIG. 12. (b) A diagram showing an example of simplified data after modification by the data modification learning method shown in FIG. 12. 3 is a flowchart of a gas analysis learning method for causing the gas analysis neural network shown in FIG. 2 to perform learning.

Embodiments of the present invention will be described below with reference to the drawings.

First, the configuration of the gas analysis system according to this embodiment will be explained.

FIG. 1 is a configuration diagram of a gas analysis system 10 according to this embodiment.

As shown in FIG. 1, the gas analysis system 10 includes a gas chromatograph 20 that generates a gas chromatogram from a sample gas, which is a gas to be analyzed, and a PC (Personal Computer), etc., for displaying the analysis results of the sample gas. computer 30. The gas analysis system 10 is an odor determination system that determines the odor of a sample gas as an analysis of the sample gas.

The gas chromatograph 20 includes a compression tank 21 in which carrier gas is compressed and stored, and an injection part 22 connected to the compression tank 21 and into which sample gas is injected from a syringe 90 for injecting sample gas. and a gas separation column 23 which is connected to the injection unit 22 and separates the mixed gas, which is a mixture of carrier gas and sample gas in the injection unit 22, into components (hereinafter referred to as "component gas"). and a detector 24 that is connected to the gas separation column 23 and detects component gas flowing from the gas separation column 23.

FIG. 2 is a block diagram of computer 30.

As shown in FIG. 2, the computer 30 includes an operation unit 31, which is an operation device such as a keyboard and a mouse, through which various operations are input, and a display device, such as an LCD (Liquid Crystal Display), which displays various information. and a communication section 33, which is a communication device that communicates with an external device via a network such as a LAN (Local Area Network) or the Internet, or directly by wire or wirelessly without going through a network. , a storage section 34 that is a non-volatile storage device such as a semiconductor memory or an HDD (Hard Disk Drive) for storing various information, and a control section 35 that controls the entire computer 30.

The computer 30 can display the results of the analysis of the sample gas on the display unit 32.

The storage unit 34 stores a gas analysis program 34a as a program for analyzing gas. The gas analysis program 34a may be installed in the computer 30 at the time of manufacturing the computer 30, or may be stored in an external storage such as a CD (Compact Disk), a DVD (Digital Versatile Disk), or a USB (Universal Serial Bus) memory. It may be additionally installed on the computer 30 from a medium or may be additionally installed on the computer 30 from a network.

The storage unit 34 stores a gas dictionary 34b indicating various types of gas information.

FIG. 3 is a diagram showing an example of the gas dictionary 34b.

The gas dictionary 34b shown in FIG. 3 uses a combination of the retention time of each component gas and the ID of the component gas (hereinafter referred to as "gas ID") to the ID of the gas category (hereinafter referred to as "category ID"). ) is stored for each. Here, the gas categories are divided according to the type of odor, such as gas containing the odor of coffee and gas containing the odor of a poultry farm. For example, a category whose category ID is "34" is a gas category that includes the odor of any of fruits such as "apple," "tangerine," "banana," and "watermelon."

As shown in FIG. 2, the storage unit 34 stores category information 34c indicating the correspondence between the gas category ID and the contents of the gas category. The storage unit 34 also stores component gas information 34d that indicates the correspondence between the gas ID of the component gas and the name of the component gas for each gas category.

The control unit 35 includes, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory) that stores programs and various data, and a RAM (Random Memory) as a memory used as a work area for the CPU of the control unit 35. Access Memory). The CPU of the control unit 35 executes a program stored in the storage unit 34 or the ROM of the control unit 35.

The control unit 35 executes the gas analysis program 34a to simplify the gas chromatogram data generated by the gas chromatograph 20, thereby generating simplified data (hereinafter referred to as "simplified data"). a data simplification unit 35a that automatically corrects the simplified data generated by the data simplification unit 35a into appropriate simplified data; A data correction unit 35c automatically corrects the simplified data into appropriate simplified data by a data correction neural network 35b, and executes analysis of the sample gas based on the simplified data after correction by the data correction unit 35c. A gas analysis neural network 35d as a neural network, a gas analysis section 35e that executes analysis of a sample gas by the gas analysis neural network 35d based on the simplified data corrected by the data correction section 35c, and a data correction neural network 35b. A data correction learning section 35f that causes the learning to be executed and a gas analysis learning section 35g that causes the gas analysis neural network 35d to execute the learning are realized.

Next, the operation of the computer 30 when analyzing sample gas will be explained.

FIG. 4 is a flow chart of the operation of computer 30 when analyzing sample gas.

As shown in FIG. 4, the data simplification unit 35a executes a data simplification step of generating simplified data by simplifying the gas chromatogram data generated by the gas chromatograph 20 for the target sample gas. (S101).

Next, the data modification unit 35c executes a data modification step in which the data modification neural network 35b automatically modifies the simplified data generated in the data simplification step of S101 into appropriate simplified data (S102).

Next, the gas analysis unit 35e executes a gas analysis step in which the gas analysis neural network 35d analyzes the sample gas based on the simplified data corrected by the data correction step of S102 (S103). Finish the indicated action.

FIG. 5 is a flowchart of the data simplification process shown in FIG. FIG. 6 is a diagram showing an example of a gas chromatogram generated by the gas chromatograph 20. FIG. 7 is a diagram showing an example of various data generated based on data of a gas chromatogram generated by the gas chromatograph 20.

Note that the gas chromatogram shown in FIG. 6 and the data shown in FIG. 7 are separate examples. That is, the data shown in FIG. 7 is data generated based on data of a gas chromatogram different from the gas chromatogram shown in FIG. 6. In addition, although the data shown in Figure 7 is drawn with some information corresponding to retention times omitted, in reality, in addition to the information shown, there is a large amount of information corresponding to retention times. Contains.

As shown in FIG. 5, the data simplification unit 35a stores the data of the gas chromatogram generated by the gas chromatograph 20 and the category ID of the target gas of this gas chromatogram (S121). For example, a gas chromatogram generated by the gas chromatograph 20 is as shown in FIG. 6, for example. The baseline value and retention time value in the gas chromatogram shown in FIG. 6 are values that depend on the individual gas chromatograph 20. That is, the gas chromatograph 20 and gas chromatographs other than the gas chromatograph 20 may have different baseline values and retention time values in the gas chromatogram. In the process of S121, the data simplification unit 35a stores data consisting of retention time and detection intensity for each retention time as gas chromatogram data, as shown in FIG. 7, for example. Furthermore, in the process of S121, the data simplification unit 35a stores, for example, the category ID input via the operation unit 31 as the category ID of the gas to be subjected to the gas chromatogram.

After the process in S121, the data simplification unit 35a adds the gas ID that is associated in the gas dictionary 34b with the category ID stored in S121 to the data stored in S121 (S122). In the process of S122, the data simplification unit 35a stores the gas ID in the gas dictionary 34b in association with the retention time associated with this gas ID in the gas dictionary 34b, for example, as shown in FIG. 7.

After the processing in S122, the data simplification unit 35a detects a retention time at which a mountain top exists as a local upper peak in the gas chromatogram data stored in S121 (S123). For example, the data simplification unit 35a targets retention times one by one, and the detection strength at the target retention time is determined by adding a specific value to the detection strength at the two retention times before the target retention time. It is determined that a mountain peak exists at the target retention time when it is larger than either of the value and the value obtained by adding a specific value to the detection intensity at the retention time two after the target retention time. Here, the specific value is, for example, 50 in the data shown in FIG. In the process of S123, the data simplification unit 35a stores, for example, among the data shown in FIG. 7, the max value indicating the detection strength of the mountaintop in association with the retention time. The max value shown in FIG. 7 is associated with the detection strength associated with the retention time for a retention time where a peak exists, and is associated with 0 for a retention time where a peak does not exist.

After the processing in S123, the data simplification unit 35a calculates the relative detection strength, that is, the density value, at the retention time where the peak exists (S124). For example, the data simplification unit 35a targets retention times where the peak exists, that is, retention times where the max value is greater than 0, one by one, and from the retention time two positions before the target retention time, The existence of a mountain peak is determined by subtracting the minimum value of detection strength in the range of all retention times before the target retention time from the maximum value of detection strength in the range of retention times up to two retention times after the target retention time. Obtain the concentration value at the retention time. In the process of S124, the data simplification unit 35a stores, for example, the density value at each retention time among the data shown in FIG. 7 in association with the retention time. Here, the data simplification unit 35a sets the concentration value to 0 at the retention time where the peak does not exist.

After the process of S124, the data simplification unit 35a calculates data obtained by converting the gas chromatogram into a bar graph (hereinafter referred to as "bar graph data") (S125). For example, the data simplification unit 35a targets retention times one by one, and calculates the range of retention times from two retention times before the target retention time to two retention times after the target retention time. If the total value of the detection strength in is greater than 0, the maximum value of the detection strength in the retention time range from the two retention times before the target retention time to the two retention times after the target retention time. , obtained as a concentration value in bar graph data of the target retention time. On the other hand, the data simplification unit 35a determines that the total value of detection intensities in the range of retention times from two retention times before the target retention time to two retention times after the target retention time is not greater than 0. In this case, 0 is determined as the concentration value in the bar graph data of the target retention time. In the process of S125, the data simplification unit 35a stores, for example, the density value in the bar graph data of each retention time among the data shown in FIG.

FIG. 8 is a diagram showing an example of bar graph data generated in the data simplification process shown in FIG. 5.

The data simplification unit 35a generates bar graph data shown in FIG. 8 based on the gas chromatogram shown in FIG. 6 in the data simplification process shown in FIG. That is, the gas chromatogram showing the analog waveform shown in FIG. 6 is converted into bar graph data showing the digital waveform shown in FIG. 8 by removing the sloped portion. Note that the data of the inclined portion is data that changes depending on, for example, the type and temperature of the gas separation column 23 used in the gas chromatograph 20 and the type and flow rate of the carrier gas used in the gas chromatograph 20, and is data that changes depending on the type and flow rate of the carrier gas used in the gas chromatograph 20. This data is of low importance in the analysis of the sample gas by the neural network 35d.

As shown in FIG. 5, after the S125 process, the data simplification unit 35a generates simplified data by simplifying the gas chromatogram data (S126). That is, in S126, the data simplification unit 35a creates data in which the gas ID added to the data in S122, the retention time associated with this gas ID, and the concentration value associated in S125 are associated. Generate as simplified data.

FIG. 9 is a diagram showing an example of simplified data 81 generated in the data simplification process shown in FIG.

The simplified data 81 shown in FIG. 9 is data consisting of the gas ID added in S122, the concentration value in the bar graph data associated with this gas ID, and the data stored in S121 among the data shown in FIG. This data consists of the category ID. Note that the simplified data 81 shown in FIG. 9 is drawn with information corresponding to some gas IDs omitted, but in reality, in addition to the information shown, many gas IDs Contains corresponding information.

As shown in FIG. 5, the data simplification unit 35a ends the data simplification process shown in FIG. 5 after the S126 process.

In addition, in the processing of S123, S124, and S125, the retention time two times before the target retention time and the retention time two times after the target retention time are used as processing standards. The number "two" is determined based on the retention time error in the gas chromatograph 20, for example. That is, when the error in retention time in the gas chromatograph 20 is ±N%, the retention time equivalent to N% is two retention times. Therefore, how many retention times are used as a processing standard is actually determined based on the retention time error in the gas chromatograph 20.

FIG. 10 is a diagram showing an example of the data modification process shown in FIG. 4.

As shown in FIG. 10, the data modification unit 35c automatically modifies the simplified data 81 shown in FIG. 9, generated by the data simplification process, into appropriate simplified data using the data modification neural network 35b. The setting items of the data correction neural network 35b include the number of input nodes indicating the number of nodes on the input side, the number of hidden nodes indicating the number of hidden nodes, the number of output nodes indicating the number of nodes on the output side, and the number of activated nodes. What to use for the function, what to use for optimization, whether to use Batch normalization, whether to use Weight decay, whether to use Dropout, input data and The number of supervised learning repetitions indicating the number of repetitions of supervised learning for each learning data consisting of output correct data is included. The data correction neural network 35b shown in FIG. 10 has a total of n+1 input nodes (n from gas 1 to gas n and 1 category), 20 hidden nodes, and 20 output nodes. is a total of n+1 items, including n items from gas 1 to gas n and 1 item from the category, the activation function is ReLU, the optimization is Adam, Batch normalization is used, and Weight decay is used. , Dropout is used, and the number of teacher learning iterations is 10,000.

In the example shown in FIG. 10, the simplified data is modified so that the value of gas 1 and the value of gas 2 are exchanged with each other.

FIG. 11 is a diagram showing an example of the gas analysis process shown in FIG. 4.

As shown in FIG. 11, the gas analysis section 35e uses the gas analysis neural network 35d to analyze the sample gas based on the simplified data corrected by the data correction process. The setting items for the gas analysis neural network 35d include the number of input nodes, the number of hidden nodes, the number of output nodes, what to use for the activation function, and what to use for optimization. , whether Batch normalization is used, whether Weight decay is used, whether Dropout is used, and the number of teacher learning repetitions. The gas analysis neural network 35d shown in FIG. 11 has a total of n+1 input nodes including n from gas 1 to gas n and one category, the number of hidden nodes is 20, and the number of output nodes is n+1. are "apple", "tangerine", "banana", and "watermelon", the activation function is ReLU, the optimization is Adam, Batch normalization is used, Weight decay is used, and Dropout is used. is used, and the number of teacher learning iterations is 10,000.

In the example shown in FIG. 11, "mandarin orange" is output as the result of the odor determination of the sample gas.

Next, a method for causing the data correction neural network 35b to perform learning (hereinafter referred to as "data correction learning method") will be described.

FIG. 12 is a flowchart of the data correction learning method.

As shown in FIG. 12, the data correction learning unit 35f uses data to simplify the gas chromatogram data generated by the gas chromatograph 20 for the target sample gas, similarly to the process in S101 (see FIG. 4). A simplification step is executed (S141).

Next, the data correction learning unit 35f displays the bar graph data generated in the data simplification process in S141 on the display unit 32 with the gas ID associated with the bar graph data in the data simplification process in S141. (S142).

FIG. 13 is a diagram showing a part of an example of bar graph data displayed in the data correction learning method.

The bar graph data shown in FIG. 13 has a gas ID associated with the portion corresponding to the mountain top. Note that the gas ID associated with the portion corresponding to the valley bottom does not need to be displayed.

As shown in FIG. 12, after the process of S142, an expert determines whether or not an inappropriate component gas exists in the bar graph data displayed in S142 (S143).

If it is determined in S143 that there is no inappropriate component gas, the data correction learning unit 35f uses the simplified data generated in the data simplification process in S141 as input data, and also uses it as output correct data. By doing so, the data correction neural network 35b is caused to perform supervised learning (S144), and the data correction learning method shown in FIG. 12 is ended.

When it is determined in S143 that an inappropriate component gas exists, the expert determines whether or not the inappropriate component gas can be corrected (S145). Here, for all inappropriate component gases present in the bar graph data displayed in S142, the expert selects component gases corresponding to retention times near the retention times corresponding to the target inappropriate component gases. If there is a suitable component gas to replace the target inappropriate component gas, it is determined that the inappropriate component gas can be corrected. On the other hand, for at least one of the inappropriate component gases present in the bar graph data displayed in S142, the expert selects a component gas corresponding to a retention time near the retention time corresponding to the target inappropriate component gas. If there is no suitable component gas to replace the target inappropriate component gas, it is determined that the inappropriate component gas cannot be corrected.

If it is determined in S145 that it is not possible to correct the inappropriate component gas, the data correction learning method shown in FIG. 12 ends.

If it is determined in S145 that inappropriate component gas can be corrected, correction is performed by an expert (S146). Here, for all inappropriate component gases present in the bar graph data displayed in S142, the expert selects among the component gases corresponding to retention times near the retention times corresponding to the target inappropriate component gases. , correct the target inappropriate component gas by replacing it with the target inappropriate component gas.

FIG. 14A is a diagram showing an example of simplified data before modification by the data modification learning method. FIG. 14(b) is a diagram showing an example of simplified data after correction by the data correction learning method.

The simplified data shown in FIG. 14(a) is corrected to the simplified data shown in FIG. 14(b), for example, by the process of S146. The simplified data shown in FIG. 14(b) has been modified so that the values of gas 3 and gas 4 in the simplified data shown in FIG. 14(a) are interchanged. Note that the simplified data shown in Figure 14 is drawn with information corresponding to some gas IDs omitted, but in reality, in addition to the information shown, there is information corresponding to many gas IDs. Contains information to

As shown in FIG. 12, when the process of S146 is completed, the data correction learning unit 35f uses the simplified data generated in the data simplification process of S141 as input data, and also obtains data from the bar graph data after correction in S146. By using the simplified data obtained as output correct data, the data correction neural network 35b is caused to perform supervised learning (S147), and the data correction learning method shown in FIG. 12 is ended.

Next, a method for causing the gas analysis neural network 35d to perform learning (hereinafter referred to as "gas analysis learning method") will be described.

FIG. 15 is a flowchart of the gas analysis learning method.

As shown in FIG. 15, the gas analysis learning unit 35g uses data to simplify the gas chromatogram data generated by the gas chromatograph 20 for the target sample gas, similar to the process in S101 (see FIG. 4). A simplification step is executed (S161).

Next, the gas analysis learning unit 35g automatically converts the simplified data generated in the data simplification process in S161 into appropriate simplified data using the data correction neural network 35b, similar to the process in S102 (see FIG. 4). A data modification step is executed to modify the data (S162).

Next, the odor of the target sample gas is determined by an expert (S163).

When the process in S163 is completed, the gas analysis learning unit 35g uses the simplified data corrected by the data correction process in S162 as input data, and also uses the result of the determination in S163 as the correct data to output. The analysis neural network 35d is caused to perform supervised learning (S164), and the gas analysis learning method shown in FIG. 15 is ended.

As explained above, the simplified data in the gas analysis system 10 includes only the detection intensity corresponding to the retention time shown in the gas dictionary 34b as the detection intensity based on the gas chromatogram. With this configuration, the gas analysis system 10 causes the gas analysis neural network 35d to perform supervised learning using simplified data that does not include data of low importance in the analysis of the sample gas by the gas analysis neural network 35d. S164), for example, it is possible to prevent overlearning by the gas analysis neural network 35d, and as a result, it is possible to improve the accuracy rate of sample gas analysis by the gas analysis neural network 35d.

The gas analysis system 10 reduces the influence of individual-dependent retention time variations of the gas chromatographs 20 that generate gas chromatograms for the sample gas on the simplified data by correction in S102 by the data correction neural network 35b. Therefore, it is possible to improve the accuracy rate of gas analysis by the gas analysis neural network 35d.

In the gas analysis system 10, the simplified data includes a relative detection intensity based on a plurality of detection intensities included in the gas chromatogram, that is, a concentration value as the detection intensity based on the gas chromatogram. With this configuration, the gas analysis system 10 generates simplified data that reduces the influence of baseline variations in gas chromatograms that depend on the individual gas chromatographs 20 that generate gas chromatograms for sample gas (S101). Therefore, the accuracy rate of gas analysis by the gas analysis neural network 35d can be improved. Note that the gas analysis system 10 uses the gas chromatogram as the detection intensity corresponding to the retention time shown in the gas dictionary 34b, rather than the concentration value, which is a relative detection intensity based on a plurality of detection intensities included in the gas chromatogram. The included detection strength may be included in the simplified data as is.

The gas analysis system 10 uses the simplified data in which the influence of variation in retention time depending on the individual gas chromatograph 20 is reduced by the correction in S162 by the data correction neural network 35b, and the gas analysis neural network 35d is supervised. Since learning is performed (S164), it is possible to improve the accuracy rate of gas analysis by the gas analysis neural network 35d.

Since the gas analysis system 10 can improve the accuracy rate of gas analysis by the gas analysis neural network 35d, it can reduce the number of times the supervised learning needs to be performed by the gas analysis neural network 35d. As a result, the number of times the sample gas is measured by the gas chromatograph 20, which is necessary for supervised learning of the gas analysis neural network 35d, can be reduced. Here, measuring the sample gas using the gas chromatograph 20 takes time and cost. Therefore, the gas analysis system 10 can save time and money.

Since the gas analysis system 10 executes the sample gas analysis by the gas analysis neural network 35d based on the simplified data, the calculation time for analyzing the sample gas in the gas analysis neural network 35d can be shortened.

In the present embodiment, the gas analysis system 10 employs online learning in which learning is performed for each piece of learning data as shown in FIGS. 12 and 15 as a supervised learning method. However, the supervised learning method may be other than online learning. For example, the gas analysis system 10 may employ batch learning in which a large number of learning data are learned at once as a supervised learning method.

The gas analysis system 10 in this embodiment is an odor determination system that determines the odor of a sample gas. However, the gas analysis system 10 may be a system that performs analysis other than odor determination as analysis of sample gas.

10 Gas analysis system 30 Computer 34a Gas analysis program (program)
34b Gas dictionary (information indicating retention time for each gas component)
35a Data simplification section 35b Data correction neural network 35c Data correction section 35d Gas analysis neural network 35e Gas analysis section 35g Gas analysis learning section

Claims (6)

a data simplification unit that generates simplified data as simplified data by simplifying data of a gas chromatogram of a sample gas using information indicating retention times of each gas component;
A gas analysis unit that executes analysis of the sample gas based on the simplified data by a gas analysis neural network as a neural network that executes analysis of the sample gas based on the simplified data,
The simplified data includes only the value corresponding to the retention time of the component indicated in the information as a value based on the detection intensity based on the gas chromatogram ,
The simplified data includes the maximum value of detection intensity in a specific time range including the retention time of the component included in the gas chromatogram, and the minimum value of detection intensity in all time ranges before this retention time. A gas analysis system characterized in that the relative detection intensity obtained by subtraction is included as the value based on the detection intensity based on the gas chromatogram . A data correction unit that automatically corrects the simplified data to the appropriate simplified data using a data correction neural network as a neural network that automatically corrects the simplified data to the appropriate simplified data,
The gas analysis system according to claim 1, wherein the gas analysis section analyzes the sample gas using the gas analysis neural network based on the simplified data corrected by the data correction section. a data simplification unit that generates simplified data as simplified data by simplifying data of a gas chromatogram of a sample gas using information indicating retention times of each gas component;
a gas analysis section that executes an analysis of the sample gas based on the simplified data using a gas analysis neural network as a neural network that executes an analysis of the sample gas based on the simplified data;
Equipped with
The simplified data includes only the value corresponding to the retention time of the component indicated in the information as a value based on the detection intensity based on the gas chromatogram,
at least the simple comprising a data correction unit that automatically corrects the simplified data to the appropriate simplified data,
The gas analysis system is characterized in that the gas analysis section analyzes the sample gas using the gas analysis neural network based on the simplified data corrected by the data correction section. comprising a gas analysis learning unit that causes the gas analysis neural network to perform learning,
The gas analysis learning unit is configured to use the simplified data obtained by simplifying the gas chromatogram data of the sample gas by the data simplification unit and correcting it by the data correction unit, and the results of human analysis of the sample gas. 4. The gas analysis system according to claim 2 or 3, wherein the gas analysis neural network is caused to perform supervised learning using a combination of . A data simplification unit that generates simplified data as simplified data by simplifying data of a gas chromatogram of a sample gas using information indicating the retention time of each gas component is implemented in a computer. ,
The simplified data includes only the value corresponding to the retention time of the component indicated in the information as a value based on the detection intensity based on the gas chromatogram ,
The simplified data includes the maximum value of detection intensity in a specific time range including the retention time of the component included in the gas chromatogram, and the minimum value of detection intensity in all time ranges before this retention time. A program characterized in that the relative detection intensity obtained by subtraction is included as the value based on the detection intensity based on the gas chromatogram . a data simplification unit that generates simplified data as simplified data by simplifying data of a gas chromatogram of a sample gas using information indicating retention times of each gas component;
a gas analysis section that executes an analysis of the sample gas based on the simplified data using a gas analysis neural network as a neural network that executes an analysis of the sample gas based on the simplified data;
A program that causes a computer to realize
The simplified data includes only the value corresponding to the retention time of the component indicated in the information as a value based on the detection intensity based on the gas chromatogram,
The program includes at least a data correction neural network as a neural network that automatically corrects the simplified data to the appropriate simplified data by exchanging the values corresponding to the retention times of the two components with each other. causing the computer to implement a data correction unit that automatically corrects the simplified data to the appropriate simplified data,
The program is characterized in that the gas analysis section executes analysis of the sample gas using the gas analysis neural network based on the simplified data corrected by the data correction section.

Images