WO2024069919A1 - Machine learning device and carbon dioxide concentration measurement device - Google Patents

Machine learning device and carbon dioxide concentration measurement device Download PDF

Info

Publication number
WO2024069919A1
WO2024069919A1 PCT/JP2022/036655 JP2022036655W WO2024069919A1 WO 2024069919 A1 WO2024069919 A1 WO 2024069919A1 JP 2022036655 W JP2022036655 W JP 2022036655W WO 2024069919 A1 WO2024069919 A1 WO 2024069919A1
Authority
WO
WIPO (PCT)
Prior art keywords
value
estimated value
correction
unit
machine learning
Prior art date
Application number
PCT/JP2022/036655
Other languages
French (fr)
Japanese (ja)
Inventor
一紀 中田
Original Assignee
Tdk株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tdk株式会社 filed Critical Tdk株式会社
Priority to PCT/JP2022/036655 priority Critical patent/WO2024069919A1/en
Publication of WO2024069919A1 publication Critical patent/WO2024069919A1/en

Links

Images

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01NINVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/17Systems in which incident light is modified in accordance with the properties of the material investigated
    • G01N21/59Transmissivity
    • G01N21/61Non-dispersive gas analysers

Definitions

  • the present invention relates to a machine learning machine and a carbon dioxide concentration measurement device.
  • Carbon dioxide concentration measuring devices that measure the carbon dioxide concentration in a space are used in various places.
  • the carbon dioxide concentration in a vinyl greenhouse is measured with a CO2 sensor, and the greenhouse is controlled to have an environment suitable for plant growth.
  • CO2 sensors By measuring the carbon dioxide concentration indoors with a CO2 sensor, it is possible to promote ventilation in the room and prevent infectious diseases.
  • air conditioning systems often measure the carbon dioxide concentration to control the indoor environment.
  • the non-dispersive infrared absorption (ND-IR) method is known, which calculates carbon dioxide concentration based on the amount of infrared radiation absorbed.
  • CO2 sensors using the ND-IR method can measure CO2 concentration with relatively high accuracy. Even with CO2 sensors using the ND-IR method, as described in Patent Documents 1 and 2, the reference value may shift due to long-term changes over time, and regular calibration is required.
  • CO2 sensors using the ND-IR method can measure CO2 concentration with relatively high accuracy, but are expensive.
  • CO2 sensors using the ND-IR method have many components such as infrared emitters, filters, and light receiving elements, and the package is large, which is a problem.
  • the present invention has been made in consideration of the above circumstances, and aims to provide a machine learning machine for realizing an inexpensive, highly accurate CO2 sensor.
  • a machine learning machine includes an estimation unit, a timing detection unit, and a correction unit.
  • the estimation unit estimates a carbon dioxide concentration based on an input signal.
  • the timing detection unit monitors an estimated value output from the estimation unit and determines the timing of correction of the estimated value.
  • the correction unit determines a correction value for correcting the estimated value each time a correction instruction is received from the timing detection unit.
  • the timing detection unit instructs the correction unit to make a correction when the change width between a first estimated value at a first time and a second estimated value at a second time subsequent to the first time is greater than the first change width and the second estimated value is equal to or smaller than a first specified value, or when the change width between the first estimated value and the second estimated value is greater than the second change width and the second estimated value is equal to or larger than a second specified value.
  • the timing detection unit may further instruct the correction unit to perform correction when the estimated value output from the estimation unit is equal to or less than a third specified value over a first period.
  • the timing detection unit may further instruct the correction unit to perform correction when the estimated value output from the estimation unit is equal to or greater than a fourth specified value over a second period.
  • the timing detection unit may further instruct the correction unit to perform correction when the change range between the first estimated value and the second estimated value becomes larger than a third change range.
  • the correction unit may add the correction value to the estimated value or subtract the correction value from the estimated value.
  • the correction unit may add a value to the estimated value or subtract a value from the estimated value. The value increases stepwise until it reaches the correction value.
  • the estimation unit may be a recurrent neural network.
  • the estimation unit may perform a learning process and an estimation process, and the learning process may be performed in an interval in which the timing detection unit does not determine that correction is necessary.
  • the carbon dioxide concentration measuring device according to the second aspect is equipped with the machine learning device according to the above aspect.
  • the machine learning machine and carbon dioxide concentration measuring device can measure carbon dioxide concentration with high accuracy even with an inexpensive sensor.
  • FIG. 1 is a schematic diagram of a carbon dioxide concentration measuring device according to a first embodiment.
  • FIG. FIG. 1 is a conceptual diagram of a reservoir network, which is an example of a recurrent neural network.
  • FIG. 11 is a schematic diagram for explaining a first condition.
  • FIG. 11 is a schematic diagram for explaining a second condition.
  • FIG. 13 is a schematic diagram for explaining a fourth condition.
  • FIG. 13 is a schematic diagram for explaining a fifth condition.
  • 11A and 11B are diagrams illustrating a correction operation by a correction unit.
  • 13A and 13B are diagrams illustrating another example of the correction operation by the correction unit.
  • 4 shows an example of a measurement result of the carbon dioxide concentration measuring device according to the first embodiment. 4 shows an example of the measurement result of a carbon dioxide concentration measuring device that does not have a correction unit.
  • FIG. 11 is a schematic diagram of a carbon dioxide concentration measuring device according to a first modified example.
  • First embodiment 1 shows a carbon dioxide concentration measuring device 100 according to the first embodiment.
  • the carbon dioxide concentration measuring device 100 includes, for example, a machine learning machine 10 and a sensor 20.
  • the sensor 20 is any sensor.
  • the sensor 20 is, for example, a CO2 sensor.
  • the sensor 20 measures the carbon dioxide concentration in the environment in which the carbon dioxide concentration measuring device 100 is placed, for example, with reference to temperature, humidity, air pressure, etc.
  • the accuracy of the sensor 20 may be lower than that of a CO2 sensor using the ND-IR method.
  • the machine learning machine 10 is, for example, a microcontroller or a processor.
  • the machine learning machine 10 has, for example, a CPU that performs arithmetic processing, a ROM that stores programs, and a RAM that is a working memory when the CPU performs arithmetic processing.
  • the machine learning machine 10 operates by executing the stored programs.
  • the machine learning machine 10 outputs an output signal based on an input signal from the sensor 20. Even if the measurement accuracy of the carbon dioxide concentration of the sensor 20 is low, the measurement accuracy of the carbon dioxide concentration measuring device 100 is improved by using the machine learning machine 10.
  • the machine learning device 10 has, for example, an estimation unit 1, a timing detection unit 2, a correction unit 3, a comparison unit 4, and a memory unit 5.
  • the machine learning device 10 performs a learning process and an estimation process.
  • the learning process is performed by the estimation unit 1, the comparison unit 4, and the memory unit 5.
  • the estimation process is performed by the estimation unit 1, the timing detection unit 2, and the correction unit 3.
  • the machine learning device 10 improves the measurement accuracy of the carbon dioxide concentration measuring device 100 through the learning process.
  • the machine learning device 10 outputs an estimated value of the carbon dioxide concentration in the environment in which the carbon dioxide concentration measuring device 100 is placed.
  • the estimation unit 1 is, for example, a part of the microcontroller or processor of the machine learning device 10.
  • the estimation unit 1 includes, for example, a CPU, ROM, RAM, etc.
  • the estimation unit 1 estimates the carbon dioxide concentration of the environment in which the carbon dioxide concentration measuring device 100 is placed based on the input signal from the sensor 20.
  • the estimation unit 1 includes a neural network.
  • a neural network is a mathematical model that mimics the network of nerve cells in the brain.
  • the estimation unit 1 is, for example, a recurrent neural network.
  • a recurrent neural network can handle nonlinear time series data. Nonlinear time series data is data whose values change over time, and examples include stock prices and the number of influenza epidemics.
  • a recurrent neural network is suitable for processing time series data by returning the results of processing in neurons in a later layer to neurons in an earlier layer.
  • Fig. 2 is a conceptual diagram of a reservoir network, which is an example of a recurrent neural network.
  • the neural network NN shown in Fig. 2 has an input layer L in , a reservoir layer R, and an output layer L out .
  • the reservoir layer R includes a plurality of nodes n i .
  • the number of nodes n i is not particularly limited. Hereinafter, the number of nodes n i is set to N.
  • Each of the nodes n i may be replaced with, for example, a physical device.
  • the physical device is, for example, a device that can convert an input signal into vibration, an electromagnetic field, a magnetic field, a spin wave, or the like.
  • connection weights are defined between each node n i .
  • the number of connection weights defined is equal to the number of combinations of connections between nodes n i .
  • Each connection weight between nodes n i is defined in principle and does not change through learning.
  • Each connection weight between nodes n i is arbitrary and may be the same as or different from each other. Some of the connection weights between multiple nodes n i may change through learning.
  • An input signal is input to the reservoir layer R from the input layer L in .
  • the input signal is input, for example, from an external sensor 20.
  • the input signal interacts with the nodes n i while propagating between them in the reservoir layer R. Signals interact with each other when a signal propagating to a node n i affects a signal propagating to another node n i . For example, when the input signal propagates between the nodes n i , a connection weight is applied to the input signal, and the input signal changes.
  • the reservoir layer R projects the input signal into a multidimensional nonlinear space.
  • the input signal input to the reservoir layer R is replaced with another signal. At least a portion of the information contained in the input signal is retained in a changed form.
  • One or more signals S i are sent to the output layer L out from the reservoir layer R.
  • a connection weight x i is applied to each of the signals S i output from the reservoir layer R.
  • the output layer L out performs a multiplication operation in which the connection weight x i is applied to the signal S i , and a summation operation in which the results of each product operation are added together.
  • the connection weight x i is updated in a learning process, and inference is performed based on the updated connection weight x i .
  • the comparison unit 4 is, for example, a part of a microcontroller or a processor.
  • the comparison unit 4 includes, for example, a CPU, a ROM, a RAM, etc.
  • the comparison unit 4 compares the teacher data stored in the storage unit 5 with the estimation result from the estimation unit 1.
  • the comparison unit 4 calculates the degree of agreement (correct answer rate) between the teacher data and the estimation result. If the accuracy rate of the comparison unit 4 is poor, the machine learning device 10 adjusts the connection weight x i of the estimation unit 1 so as to increase the accuracy rate.
  • the storage unit 5 is, for example, a memory.
  • the storage unit 5 may be an external server.
  • Teacher data is stored in the storage unit 5.
  • the teacher data is, for example, a measurement result obtained by a highly accurate carbon dioxide concentration measuring device.
  • the teacher data is, for example, a measurement result obtained by a CO2 sensor using the ND-IR method.
  • the comparison unit 4 compares the estimation result from the estimation unit 1 with the teacher data stored in the memory unit 5 to increase the degree of agreement between the estimation result and the teacher data.
  • the degree of agreement between the estimation result and the teacher data in the learning process, even when the carbon dioxide concentration is measured using a low-accuracy sensor 20, it is possible to measure the carbon dioxide concentration with the same accuracy as when the carbon dioxide concentration is measured using a highly accurate sensor that serves as teacher data.
  • the learning process is preferably performed in an interval in which the timing detection unit 2, described later, does not determine that correction is necessary.
  • the timing detection unit 2 is, for example, a part of a microcontroller or a processor.
  • the timing detection unit 2 includes, for example, a CPU, a ROM, a RAM, etc.
  • the timing detection unit 2 operates during the estimation process of the machine learning device 10. During the estimation process, the timing detection unit 2 monitors the estimated value output from the estimation unit 1 and determines the timing for correcting the estimated value.
  • the timing detection unit 2 instructs the correction unit 3 to correct the estimated value output from the estimation unit 1 when, for example, at least the first condition or the second condition is satisfied.
  • the timing detection unit 2 may instruct the correction unit 3 to correct the estimated value output from the estimation unit 1 when, for example, the first condition and the second condition are satisfied.
  • FIG. 3 is a schematic diagram for explaining the first condition.
  • the graph shown in FIG. 3 shows the change over time of the estimated value output from the estimation unit 1.
  • the horizontal axis of the graph in FIG. 3 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
  • the estimated value is output at regular intervals and changes over time.
  • the carbon dioxide concentration fluctuates based on a reference value Vb.
  • the reference value Vb can be set arbitrarily according to the environment in which the carbon dioxide concentration measuring device 100 is installed.
  • the reference value Vb is, for example, the CO2 concentration in the atmosphere, and is approximately 400 ppm.
  • the first condition is that the change range ⁇ V between the first estimated value V1 and the second estimated value V2 is greater than the first change range ⁇ V1, and the second estimated value V2 is equal to or smaller than the first specified value Vs1.
  • the first estimated value V1 is an estimated value at the first time t1.
  • the second estimated value V2 is an estimated value at the second time t2.
  • the first time t1 and the second time t2 are consecutive in time, and the second time t2 is the time at which the estimation unit 1 outputs an estimated value after the first time t1.
  • the first change width ⁇ V1 can be set arbitrarily in accordance with the measurement environment.
  • the first specified value Vs1 can be set arbitrarily in accordance with the measurement environment.
  • the first condition may be met.
  • the reference value Vb is set to the CO2 concentration of the outside air, and the estimated value should approach the reference value Vb by opening the window for ventilation or the like. If the estimated value is equal to or less than the first specified value Vs1 that is lower than the reference value Vb, it is highly likely that the accuracy rate of the estimated value has decreased.
  • the machine learning device 10 can maintain a high accuracy rate when the data changes continuously over time, but the accuracy rate tends to decrease when the data changes suddenly. Therefore, when the first condition is met, correcting the estimated value increases the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100.
  • FIG. 4 is a schematic diagram for explaining the second condition.
  • the graph shown in FIG. 4 shows the change over time of the estimated value output from the estimation unit 1.
  • the horizontal axis of the graph in FIG. 4 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
  • the second condition is that the change range ⁇ V' between the first estimated value V1' and the second estimated value V2' is greater than the second change range ⁇ V2, and the second estimated value V2' is greater than or equal to the second specified value Vs2.
  • the first estimated value V1' is an estimated value at the first time t1'.
  • the second estimated value V2' is an estimated value at the second time t2'.
  • the first time t1' and the second time t2' are consecutive in time, and the second time t2' is the time at which the estimation unit 1 outputs an estimated value after the first time t1'.
  • the second change width ⁇ V2 can be set arbitrarily in accordance with the measurement environment.
  • the second change width ⁇ V2 may be the same as the first change width V1.
  • the second specified value Vs2 can be set arbitrarily in accordance with the measurement environment.
  • the absolute value of the difference between the second specified value Vs2 and the reference value Vb may be the same as the absolute value of the difference between the first specified value Vs1 and the reference value Vb.
  • the second condition may be met. If the CO2 concentration that is unlikely to be reached in the environment in which the carbon dioxide concentration measuring device 100 is installed is set as the second specified value Vs2, the fact that the estimated value is equal to or greater than the second specified value Vs2 means that the accuracy rate of the estimated value is likely to be low. In addition, as described above, the machine learning device 10 is not good at responding to sudden changes in data. When the second condition is met, correcting the estimated value increases the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100.
  • the timing detection unit 2 may also instruct the correction unit 3 to make a correction when the third condition is satisfied.
  • the third condition is that the change range between the first estimated value and the second estimated value is greater than the third change range.
  • the third condition is a condition that focuses only on the change range and is unrelated to the first specified value Vs1 and the second specified value Vs2.
  • the third change range may be the same as or different from the first change range ⁇ V1 or the second change range ⁇ V2. As described above, the machine learning device 10 has difficulty responding to sudden changes in data. If the estimated value is corrected when the third condition is satisfied, the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100 is improved.
  • the timing detection unit 2 may also instruct the correction unit 3 to make a correction when the fourth condition is satisfied.
  • FIG. 5 is a schematic diagram for explaining the fourth condition.
  • the graph shown in FIG. 5 shows the change over time in the estimated value output from the estimation unit 1.
  • the horizontal axis of the graph in FIG. 5 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
  • the fourth condition is that the estimated value output from the estimation unit 1 is equal to or less than the third specified value Vs3 over the first period ⁇ t1.
  • the first period ⁇ t1 can be set arbitrarily according to the measurement environment.
  • the third specified value Vs3 can be set arbitrarily according to the measurement environment.
  • the third specified value Vs3 may be the same as the first specified value Vs1.
  • the lower limit of the CO2 concentration that may be reached in the environment in which the carbon dioxide concentration measuring device 100 is installed is set as the third specified value Vs3
  • the fact that the estimated value is maintained at or below the third specified value Vs3 for a certain period of time means that the accuracy rate of the estimated value is likely to have decreased. Therefore, if the estimated value is corrected when the fourth condition is satisfied, the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100 is improved.
  • the timing detection unit 2 may also instruct the correction unit 3 to make a correction when the fifth condition is satisfied.
  • FIG. 6 is a schematic diagram for explaining the fifth condition.
  • the graph shown in FIG. 6 shows the change over time in the estimated value output from the estimation unit 1.
  • the horizontal axis of the graph in FIG. 6 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
  • the fifth condition is that the estimated value output from the estimation unit 1 is equal to or greater than the fourth specified value Vs4 over the second period ⁇ t2.
  • the second period ⁇ t2 can be set arbitrarily to suit the measurement environment.
  • the second period ⁇ t2 may be the same as or different from the first period ⁇ t1.
  • the fourth specified value Vs4 can be set arbitrarily to suit the measurement environment.
  • the fourth specified value Vs4 may be the same as the second specified value Vs2.
  • the upper limit of the CO2 concentration that may be reached in the environment in which the carbon dioxide concentration measuring device 100 is installed is set as the fourth specified value Vs4
  • the fact that the estimated value is maintained at or above the fourth specified value Vs4 for a certain period of time means that the accuracy rate of the estimated value is likely to have decreased. Therefore, if the estimated value is corrected when the fifth condition is satisfied, the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100 is improved.
  • the correction unit 3 is, for example, a part of a microcontroller or a processor.
  • the correction unit 3 includes, for example, a CPU, a ROM, a RAM, etc.
  • the correction unit 3 operates during the estimation process of the machine learning device 10. During the estimation process, the correction unit 3 determines a correction value for correcting the estimated value each time a correction instruction is received from the timing detection unit 2. The correction value is rewritten, for example, each time a correction instruction is received from the timing detection unit 2.
  • FIG. 7 is a diagram showing the correction operation by the correction unit 3. As shown in FIG. 7, for example, when the first condition is satisfied, the absolute value of the difference between the reference value Vb and the second estimated value V2 is set as the correction value. For example, the correction unit 3 adds the correction value to the estimated value output at or after the third time t3, when the estimated value is output next after the second time t2.
  • FIG. 8 is a diagram showing another example of the correction operation by the correction unit 3.
  • the correction unit 3 may gradually increase the value added to the estimated value until it reaches the correction value.
  • the correction unit 3 adds gradually larger values to the estimated value during a third period ⁇ t3 from the third time t3 to the fourth time t4. Then, the correction value is added to the estimated value that is output after the fourth time t4. By gradually increasing the value added to the estimated value until it reaches the correction value, it is possible to prevent the output estimated value from changing suddenly.
  • the absolute value of the difference between the reference value Vb and the second estimated value V2' is set as the correction value.
  • the correction unit 3 subtracts the correction value from the estimated value that is output, for example, at or after the third time t3 when the estimated value is output next after the second time t2.
  • the correction unit 3 may gradually increase the value that is subtracted from the estimated value until the value that is subtracted from the estimated value becomes the correction value.
  • the absolute value of the reference value Vb and the second estimated value is set as the correction value.
  • the correction unit 3 adds the correction value to the estimated value output at or after the third time t3 when the estimated value is output next after the second time t2, or subtracts the correction value from the estimated value. If the second estimated value is smaller than the reference value Vb, the correction value is added to the estimated value, and if the second estimated value is larger than the reference value Vb, the correction value is subtracted from the estimated value. The value added to or subtracted from the estimated value may be increased stepwise until it reaches the correction value.
  • the absolute value of the difference between the reference value Vb and the estimated value that is first output after the first period ⁇ t1 has elapsed is set as the correction value.
  • the correction unit 3 adds the correction value to the estimated value that is output after the first period ⁇ t1 has elapsed.
  • the value added by the correction unit 3 may be increased stepwise until it reaches the correction value.
  • the absolute value of the difference between the reference value Vb and the estimated value that is first output after the second period ⁇ t2 has elapsed is set as the correction value.
  • the correction unit 3 subtracts the correction value from the estimated value that is output after the second period ⁇ t2 has elapsed.
  • the value that the correction unit 3 subtracts may be increased stepwise until it reaches the correction value.
  • the estimation unit 1 has a higher degree of agreement with the teacher data through a learning process, and the accuracy of the estimated value output from the estimation unit 1 is higher than that of the raw data output by the sensor 20.
  • the machine learning device 10 outputs the estimated value from the estimation unit 1 as a predicted value of the carbon dioxide concentration.
  • the correction unit 3 corrects the estimated value.
  • the carbon dioxide concentration measuring device 100 according to this embodiment can improve the measurement accuracy of the carbon dioxide concentration by using the machine learning device 10. Furthermore, the carbon dioxide concentration measuring device 100 according to this embodiment can further improve the measurement accuracy of the carbon dioxide concentration by correcting the estimated value from the estimation unit 1 using the correction unit 3.
  • Fig. 9 shows an example of the measurement result of the carbon dioxide concentration measuring device according to the first embodiment.
  • Fig. 10 shows an example of the measurement result of the carbon dioxide concentration measuring device without the correction unit 3.
  • the reference in Fig. 9 and Fig. 10 is teacher data, which is the output value of a CO2 sensor using the ND-IR method.
  • the carbon dioxide concentration measuring device shown in Figures 9 and 10 performs learning processing in the first stage and estimation processing in the second stage.
  • the learning processing is performed in an interval where the timing detection unit 2 does not determine that correction is necessary.
  • the timing detection unit 2 and the correction unit 3 are turned off.
  • the estimated value using the output value from the less accurate sensor 20 has a high degree of agreement with the reference, and it can be confirmed that the measurement accuracy of the carbon dioxide concentration of the carbon dioxide concentration measuring device of the first embodiment is as high as that of a CO2 sensor using the ND-IR method.
  • Figure 9 shows a higher degree of agreement between the estimated value and the reference. This is because a correction process that satisfies the first condition was performed in the area surrounded by the dotted line. In other words, it can be confirmed that the functioning of correction unit 3 improves the measurement accuracy of the carbon dioxide concentration by the carbon dioxide concentration measuring device 100.
  • FIG. 11 is a schematic diagram of a carbon dioxide concentration measuring device 101 according to the first embodiment.
  • the comparison unit 4 and memory unit 5 are arranged outside the machine learning device 11. In other words, the comparison unit 4 and memory unit 5 are not installed in the edge terminal.
  • the learning process can be performed in advance at the time of shipment, or can be removed from the edge terminal.
  • the comparison unit 4 and memory unit 5 can also be stored in an external server, and the edge terminal and the external server can communicate to update the learning process.

Landscapes

  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

This machine learning device comprises: an estimation unit that estimates a carbon dioxide concentration on the basis of an input signal; a timing detection unit that monitors an estimated value output from the estimation unit, and that determines a timing for correction of the estimated value; and a correction unit that determines a correction value for correcting the estimated value every time a correction instruction is received from the timing detection unit. The timing detection unit issues a correction instruction to the correction unit if a variation width between a first estimated value at a first time and a second estimated value at a second time which is continuous with the first time is greater than a first variation width and the second estimated value is equal to or less than a first defined value, or if a variation width between the first estimated value and the second estimated value is greater than a second variation width and the second estimated value is equal to or greater than a second defined value.

Description

機械学習器及び二酸化炭素濃度測定装置Machine learning machine and carbon dioxide concentration measuring device
 本発明は、機械学習器及び二酸化炭素濃度測定装置に関する。 The present invention relates to a machine learning machine and a carbon dioxide concentration measurement device.
 空間内の二酸化炭素濃度を測定する二酸化炭素濃度測定装置(COセンサー)は様々なところで利用されている。例えば、ビニールハウス内の二酸化炭素濃度をCOセンサーで測定し、ビニールハウス内を植物の育成に適した環境に制御することが行われている。また例えば、室内の二酸化炭素濃度をCOセンサーで測定することで、室内の換気を促し、感染症を予防することができる。また例えば、空調システムも二酸化炭素濃度を測定して、室内環境をコントロールしている場合が多い。 Carbon dioxide concentration measuring devices ( CO2 sensors) that measure the carbon dioxide concentration in a space are used in various places. For example, the carbon dioxide concentration in a vinyl greenhouse is measured with a CO2 sensor, and the greenhouse is controlled to have an environment suitable for plant growth. For example, by measuring the carbon dioxide concentration indoors with a CO2 sensor, it is possible to promote ventilation in the room and prevent infectious diseases. For example, air conditioning systems often measure the carbon dioxide concentration to control the indoor environment.
 COセンサーを用いて二酸化炭素濃度を測定する方法の一つとして、赤外線の吸収量を基に二酸化炭素濃度を算出する非分散型赤外線吸収(ND-IR)法が知られている。ND-IR法を利用したCOセンサーは、比較的高い精度でCO濃度を測定できる。ND-IR法を利用したCOセンサーであっても、特許文献1及び特許文献2に記載のように、長期的な経時変化によって基準値がずれる場合があり、定期的な較正が必要である。 As one method for measuring carbon dioxide concentration using a CO2 sensor, the non-dispersive infrared absorption (ND-IR) method is known, which calculates carbon dioxide concentration based on the amount of infrared radiation absorbed. CO2 sensors using the ND-IR method can measure CO2 concentration with relatively high accuracy. Even with CO2 sensors using the ND-IR method, as described in Patent Documents 1 and 2, the reference value may shift due to long-term changes over time, and regular calibration is required.
特開2014-115175号公報JP 2014-115175 A 特許第6786817号公報Patent No. 6786817
 ND-IR法を利用したCOセンサーは、比較的高い精度でCO濃度を測定できるが、高価である。またND-IR法を用いたCOセンサーは、赤外線発光体、フィルター、受光素子等の部品が多く、パッケージが大きくなるという問題もある。より安価なCOセンサーで、ND-IR法を用いたCOセンサーと同等の精度を有するセンサーが求められている。 CO2 sensors using the ND-IR method can measure CO2 concentration with relatively high accuracy, but are expensive. In addition, CO2 sensors using the ND-IR method have many components such as infrared emitters, filters, and light receiving elements, and the package is large, which is a problem. There is a demand for a cheaper CO2 sensor that has the same accuracy as a CO2 sensor using the ND-IR method.
 本発明は上記事情に鑑みてなされたものであり、安価で精度の高いCOセンサーを実現するための機械学習器を提供することを目的とする。 The present invention has been made in consideration of the above circumstances, and aims to provide a machine learning machine for realizing an inexpensive, highly accurate CO2 sensor.
(1)第1の態様に係る機械学習器は、推定部とタイミング検出部と補正部とを有する。推定部は、入力信号に基づき、二酸化炭素濃度を推定する。タイミング検出部は、前記推定部から出力される推定値をモニターし、前記推定値の補正のタイミングを判断する。補正部は、前記タイミング検出部から補正の指示を受ける毎に、前記推定値を補正するための補正値を決定する。前記タイミング検出部は、第1時刻における第1推定値と前記第1時刻に連続する第2時刻における第2推定値との変化幅が第1変化幅より大きく、かつ、前記第2推定値が第1規定値以下の場合、又は、前記第1推定値と前記第2推定値との変化幅が第2変化幅より大きく、かつ、前記第2推定値が第2規定値以上の場合に、前記補正部に補正の指示を行う。 (1) A machine learning machine according to a first aspect includes an estimation unit, a timing detection unit, and a correction unit. The estimation unit estimates a carbon dioxide concentration based on an input signal. The timing detection unit monitors an estimated value output from the estimation unit and determines the timing of correction of the estimated value. The correction unit determines a correction value for correcting the estimated value each time a correction instruction is received from the timing detection unit. The timing detection unit instructs the correction unit to make a correction when the change width between a first estimated value at a first time and a second estimated value at a second time subsequent to the first time is greater than the first change width and the second estimated value is equal to or smaller than a first specified value, or when the change width between the first estimated value and the second estimated value is greater than the second change width and the second estimated value is equal to or larger than a second specified value.
(2)上記態様に係る機械学習器において、前記タイミング検出部は、前記推定部から出力される前記推定値が第1期間に亘って第3規定値以下となる場合に、前記補正部に補正の指示をさらに行ってもよい。 (2) In the machine learning device according to the above aspect, the timing detection unit may further instruct the correction unit to perform correction when the estimated value output from the estimation unit is equal to or less than a third specified value over a first period.
(3)上記態様に係る機械学習器において、前記タイミング検出部は、前記推定部から出力される前記推定値が第2期間に亘って第4規定値以上となる場合に、前記補正部に補正の指示をさらに行ってもよい。 (3) In the machine learning device according to the above aspect, the timing detection unit may further instruct the correction unit to perform correction when the estimated value output from the estimation unit is equal to or greater than a fourth specified value over a second period.
(4)上記態様に係る機械学習器において、前記タイミング検出部は、前記第1推定値と前記第2推定値との変化幅が第3変化幅より大きくなる場合に、前記補正部に補正の指示をさらに行ってもよい。 (4) In the machine learning device according to the above aspect, the timing detection unit may further instruct the correction unit to perform correction when the change range between the first estimated value and the second estimated value becomes larger than a third change range.
(5)上記態様に係る機械学習器において、前記補正部は、前記推定値に前記補正値を追加してもよいし、前記推定値から前記補正値を差し引いてもよい。 (5) In the machine learning device according to the above aspect, the correction unit may add the correction value to the estimated value or subtract the correction value from the estimated value.
(6)上記態様に係る機械学習器において、前記補正部は、前記推定値に値を追加してもよいし、前記推定値から値を差し引いてもよい。前記値は、前記補正値に至るまで段階的に大きくなる。 (6) In the machine learning machine according to the above aspect, the correction unit may add a value to the estimated value or subtract a value from the estimated value. The value increases stepwise until it reaches the correction value.
(7)上記態様に係る機械学習器において、前記推定部は、再帰型ニューラルネットワークでもよい。 (7) In the machine learning device according to the above aspect, the estimation unit may be a recurrent neural network.
(8)上記態様に係る機械学習器において、前記推定部は、学習処理と推定処理とを行い、前記学習処理は、前記タイミング検出部が、補正が必要であると判断していない区間で行ってもよい。 (8) In the machine learning device according to the above aspect, the estimation unit may perform a learning process and an estimation process, and the learning process may be performed in an interval in which the timing detection unit does not determine that correction is necessary.
(9)第2の態様に係る二酸化炭素濃度測定装置は、上記態様に係る機械学習器を備える。 (9) The carbon dioxide concentration measuring device according to the second aspect is equipped with the machine learning device according to the above aspect.
 上記態様にかかる機械学習器及び二酸化炭素濃度測定装置は、安価なセンサーでも高い精度で二酸化炭素濃度を測定できる。 The machine learning machine and carbon dioxide concentration measuring device according to the above aspect can measure carbon dioxide concentration with high accuracy even with an inexpensive sensor.
第1実施形態に係る二酸化炭素濃度測定装置の模式図である。1 is a schematic diagram of a carbon dioxide concentration measuring device according to a first embodiment. FIG. 再帰型ニューラルネットワークの一例であるリザボアネットワークの概念図である。FIG. 1 is a conceptual diagram of a reservoir network, which is an example of a recurrent neural network. 第1条件を説明するための模式図である。FIG. 11 is a schematic diagram for explaining a first condition. 第2条件を説明するための模式図である。FIG. 11 is a schematic diagram for explaining a second condition. 第4条件を説明するための模式図である。FIG. 13 is a schematic diagram for explaining a fourth condition. 第5条件を説明するための模式図である。FIG. 13 is a schematic diagram for explaining a fifth condition. 補正部による補正動作を示す図である。11A and 11B are diagrams illustrating a correction operation by a correction unit. 補正部による補正動作の別の例を示す図である。13A and 13B are diagrams illustrating another example of the correction operation by the correction unit. 第1実施形態に係る二酸化炭素濃度測定装置の測定結果の一例を示す。4 shows an example of a measurement result of the carbon dioxide concentration measuring device according to the first embodiment. 補正部を有さない二酸化炭素濃度測定装置の測定結果の一例を示す。4 shows an example of the measurement result of a carbon dioxide concentration measuring device that does not have a correction unit. 第1変形例に係る二酸化炭素濃度測定装置の模式図である。FIG. 11 is a schematic diagram of a carbon dioxide concentration measuring device according to a first modified example.
 以下、本実施形態について、図を適宜参照しながら詳細に説明する。以下の説明で用いる図面は、本発明の特徴をわかりやすくするために便宜上特徴となる部分を拡大して示している場合があり、各構成要素の寸法比率などは実際とは異なっていることがある。以下の説明において例示される材料、寸法等は一例であって、本発明はそれらに限定されるものではなく、本発明の効果を奏する範囲で適宜変更して実施することが可能である。 The present embodiment will now be described in detail with reference to the drawings as appropriate. The drawings used in the following description may show enlarged characteristic parts for the sake of convenience in order to make the features of the present invention easier to understand, and the dimensional ratios of each component may differ from the actual ones. The materials, dimensions, etc. exemplified in the following description are merely examples, and the present invention is not limited to them. Appropriate modifications can be made within the scope of the effects of the present invention.
「第1実施形態」
 図1は、第1実施形態に係る二酸化炭素濃度測定装置100である。二酸化炭素濃度測定装置100は、例えば、機械学習器10とセンサー20とを有する。
"First embodiment"
1 shows a carbon dioxide concentration measuring device 100 according to the first embodiment. The carbon dioxide concentration measuring device 100 includes, for example, a machine learning machine 10 and a sensor 20.
 センサー20は、任意のセンサーである。センサー20は、例えば、COセンサーである。センサー20は、例えば、温度、湿度、気圧等を参考に、二酸化炭素濃度測定装置100が配置される環境の二酸化炭素濃度を測定する。センサー20の精度は、ND-IR法を用いたCOセンサーの精度より低くてもよい。 The sensor 20 is any sensor. The sensor 20 is, for example, a CO2 sensor. The sensor 20 measures the carbon dioxide concentration in the environment in which the carbon dioxide concentration measuring device 100 is placed, for example, with reference to temperature, humidity, air pressure, etc. The accuracy of the sensor 20 may be lower than that of a CO2 sensor using the ND-IR method.
 機械学習器10は、例えば、マイクロコントローラー、プロセッサである。機械学習器10は、例えば、演算処理を行うCPU、プログラムを格納するROM、CPUの演算処理を行う際の作業用のメモリであるRAM等を有する。機械学習器10は、記憶されたプログラムを実行することにより動作する。 The machine learning machine 10 is, for example, a microcontroller or a processor. The machine learning machine 10 has, for example, a CPU that performs arithmetic processing, a ROM that stores programs, and a RAM that is a working memory when the CPU performs arithmetic processing. The machine learning machine 10 operates by executing the stored programs.
 機械学習器10は、センサー20からの入力信号を基に、出力信号を出力する。センサー20の二酸化炭素濃度の測定精度が低い場合でも、機械学習器10を用いることで二酸化炭素濃度測定装置100の測定精度が高まる。 The machine learning machine 10 outputs an output signal based on an input signal from the sensor 20. Even if the measurement accuracy of the carbon dioxide concentration of the sensor 20 is low, the measurement accuracy of the carbon dioxide concentration measuring device 100 is improved by using the machine learning machine 10.
 機械学習器10は、例えば、推定部1とタイミング検出部2と補正部3と比較部4と記憶部5とを有する。機械学習器10は、学習処理と推定処理とを行う。学習処理は、推定部1と比較部4と記憶部5とで行う。推定処理は、推定部1とタイミング検出部2と補正部3とで行う。機械学習器10は、学習処理で、二酸化炭素濃度測定装置100の測定精度を高める。機械学習器10は、推定処理で、二酸化炭素濃度測定装置100が配置される環境の二酸化炭素濃度の推定値を出力する。 The machine learning device 10 has, for example, an estimation unit 1, a timing detection unit 2, a correction unit 3, a comparison unit 4, and a memory unit 5. The machine learning device 10 performs a learning process and an estimation process. The learning process is performed by the estimation unit 1, the comparison unit 4, and the memory unit 5. The estimation process is performed by the estimation unit 1, the timing detection unit 2, and the correction unit 3. The machine learning device 10 improves the measurement accuracy of the carbon dioxide concentration measuring device 100 through the learning process. Through the estimation process, the machine learning device 10 outputs an estimated value of the carbon dioxide concentration in the environment in which the carbon dioxide concentration measuring device 100 is placed.
 推定部1は、例えば、機械学習器10のマイクロコントローラー、プロセッサの一部である。推定部1は、例えば、CPU、ROM、RAM等を含む。推定部1は、センサー20からの入力信号に基づき、二酸化炭素濃度測定装置100が配置される環境の二酸化炭素濃度を推定する。 The estimation unit 1 is, for example, a part of the microcontroller or processor of the machine learning device 10. The estimation unit 1 includes, for example, a CPU, ROM, RAM, etc. The estimation unit 1 estimates the carbon dioxide concentration of the environment in which the carbon dioxide concentration measuring device 100 is placed based on the input signal from the sensor 20.
 推定部1は、ニューラルネットワークを含む。ニューラルネットワークは、脳内の神経細胞のネットワークを模倣した数学モデルである。推定部1は、例えば、再帰型ニューラルネットワークである。再帰型ニューラルネットワークは、非線形な時系列のデータを扱うことができる。非線形な時系列のデータは、時間の経過とともに値が変化するデータであり、株価やインフルエンザの流行者数はその一例である。再帰型ニューラルネットワークは、後段の階層のニューロンでの処理結果を前段の階層のニューロンに戻すことで、時系列のデータの処理に適している。 The estimation unit 1 includes a neural network. A neural network is a mathematical model that mimics the network of nerve cells in the brain. The estimation unit 1 is, for example, a recurrent neural network. A recurrent neural network can handle nonlinear time series data. Nonlinear time series data is data whose values change over time, and examples include stock prices and the number of influenza epidemics. A recurrent neural network is suitable for processing time series data by returning the results of processing in neurons in a later layer to neurons in an earlier layer.
 図2は、再帰型ニューラルネットワークの一例のであるリザボアネットワークの概念図である。図2に示すニューラルネットワークNNは、入力層Linとレザバー層Rと出力層Loutとを有する。 Fig. 2 is a conceptual diagram of a reservoir network, which is an example of a recurrent neural network. The neural network NN shown in Fig. 2 has an input layer L in , a reservoir layer R, and an output layer L out .
 レザバー層Rは、複数のノードnを備える。ノードnの数は、特に問わない。以下、ノードnの数をN個とする。ノードnのそれぞれは、例えば、物理的なデバイスに置き換えてもよい。物理デバイスは、例えば、入力された信号を振動、電磁場、磁場、スピン波等に変換できるデバイスである。 The reservoir layer R includes a plurality of nodes n i . The number of nodes n i is not particularly limited. Hereinafter, the number of nodes n i is set to N. Each of the nodes n i may be replaced with, for example, a physical device. The physical device is, for example, a device that can convert an input signal into vibration, an electromagnetic field, a magnetic field, a spin wave, or the like.
 それぞれのノードnは、周囲のノードnと相互作用している。それぞれのノードnの間には、例えば、結合重みが規定されている。規定される結合重みの数は、ノードn間の接続の組み合わせの数だけある。ノードnの間の結合重みのそれぞれは、原則、規定されており、学習により変動するものではない。ノードnの間の結合重みのそれぞれは、任意であり、互いに一致していても、異なっていてもよい。複数のノードnの間の結合重みの一部は、学習により変動してもよい。 Each node n i interacts with surrounding nodes n i . For example, connection weights are defined between each node n i . The number of connection weights defined is equal to the number of combinations of connections between nodes n i . Each connection weight between nodes n i is defined in principle and does not change through learning. Each connection weight between nodes n i is arbitrary and may be the same as or different from each other. Some of the connection weights between multiple nodes n i may change through learning.
 レザバー層Rには、入力層Linから入力信号が入力される。入力信号は、例えば、外部に設けられたセンサー20から入力される。入力信号は、レザバー層R内で複数のノードn間を伝搬しながら、相互作用する。信号が相互作用するとは、あるノードnに伝搬した信号が他のノードnを伝搬する信号に影響を及ぼすことをいう。例えば、入力信号は、ノードn間を伝搬する際に結合重みが印加され、変化していく。レザバー層Rは、入力された入力信号を多次元の非線形空間に射影する。 An input signal is input to the reservoir layer R from the input layer L in . The input signal is input, for example, from an external sensor 20. The input signal interacts with the nodes n i while propagating between them in the reservoir layer R. Signals interact with each other when a signal propagating to a node n i affects a signal propagating to another node n i . For example, when the input signal propagates between the nodes n i , a connection weight is applied to the input signal, and the input signal changes. The reservoir layer R projects the input signal into a multidimensional nonlinear space.
 レザバー層Rに入力された入力信号は、別の信号に置き換わる。入力された入力信号に含まれる情報の少なくとも一部は形を変えて保有される。 The input signal input to the reservoir layer R is replaced with another signal. At least a portion of the information contained in the input signal is retained in a changed form.
 出力層Loutには、レザバー層Rから1つ以上の信号Sが送られる。レザバー層Rから出力された信号Sのそれぞれには、結合重みxが印加される。出力層Loutは、信号Sに結合重みxを印加する積演算と、それぞれの積演算結果を足し合わせる和演算とを行う。結合重みxは、学習処理で更新され、更新された結合重みxに基づいて推論が行われる。 One or more signals S i are sent to the output layer L out from the reservoir layer R. A connection weight x i is applied to each of the signals S i output from the reservoir layer R. The output layer L out performs a multiplication operation in which the connection weight x i is applied to the signal S i , and a summation operation in which the results of each product operation are added together. The connection weight x i is updated in a learning process, and inference is performed based on the updated connection weight x i .
 比較部4は、例えば、マイクロコントローラー、プロセッサの一部である。比較部4は、例えば、CPU、ROM、RAM等を含む。比較部4は、記憶部5に記憶された教師データと推定部1からの推定結果とを比較する。比較部4は、教師データと推定結果との一致度(正答率)を求める。比較部4の正答率が悪い場合は、機械学習器10は、正答率が高くなるように、推定部1の結合重みxを調整する。 The comparison unit 4 is, for example, a part of a microcontroller or a processor. The comparison unit 4 includes, for example, a CPU, a ROM, a RAM, etc. The comparison unit 4 compares the teacher data stored in the storage unit 5 with the estimation result from the estimation unit 1. The comparison unit 4 calculates the degree of agreement (correct answer rate) between the teacher data and the estimation result. If the accuracy rate of the comparison unit 4 is poor, the machine learning device 10 adjusts the connection weight x i of the estimation unit 1 so as to increase the accuracy rate.
 記憶部5は、例えば、メモリである。記憶部5は、外部サーバーでもよい。記憶部5には、教師データが格納されている。教師データは、例えば、精度の高い二酸化炭素濃度測定装置による測定結果である。教師データは、例えば、ND-IR法を用いたCOセンサーの測定結果である。 The storage unit 5 is, for example, a memory. The storage unit 5 may be an external server. Teacher data is stored in the storage unit 5. The teacher data is, for example, a measurement result obtained by a highly accurate carbon dioxide concentration measuring device. The teacher data is, for example, a measurement result obtained by a CO2 sensor using the ND-IR method.
 ここで機械学習器10の学習処理について説明する。学習処理では、推定部1からの推定結果と、記憶部5に保存された教師データと、を比較部4で比較し、推定結果と教師データとの一致度を高める。学習処理で推定結果と教師データとの一致度を高めることで、精度が低いセンサー20を用いて二酸化炭素濃度を測定した場合でも、教師データとなる精度が高いセンサーを用いて二酸化炭素濃度を測定した場合と、同等の精度で二酸化炭素濃度を測定できる。学習処理は、後述するタイミング検出部2が、補正が必要であると判断していない区間で行うことが好ましい。 The learning process of the machine learning device 10 will now be described. In the learning process, the comparison unit 4 compares the estimation result from the estimation unit 1 with the teacher data stored in the memory unit 5 to increase the degree of agreement between the estimation result and the teacher data. By increasing the degree of agreement between the estimation result and the teacher data in the learning process, even when the carbon dioxide concentration is measured using a low-accuracy sensor 20, it is possible to measure the carbon dioxide concentration with the same accuracy as when the carbon dioxide concentration is measured using a highly accurate sensor that serves as teacher data. The learning process is preferably performed in an interval in which the timing detection unit 2, described later, does not determine that correction is necessary.
 タイミング検出部2は、例えば、マイクロコントローラー、プロセッサの一部である。タイミング検出部2は、例えば、CPU、ROM、RAM等を含む。 The timing detection unit 2 is, for example, a part of a microcontroller or a processor. The timing detection unit 2 includes, for example, a CPU, a ROM, a RAM, etc.
 タイミング検出部2は、機械学習器10の推定処理時に動作する。タイミング検出部2は、推定処理時に、推定部1から出力される推定値をモニターし、推定値の補正のタイミングを判断する。 The timing detection unit 2 operates during the estimation process of the machine learning device 10. During the estimation process, the timing detection unit 2 monitors the estimated value output from the estimation unit 1 and determines the timing for correcting the estimated value.
 タイミング検出部2は、例えば、少なくとも第1条件又は第2条件を満たす場合に、推定部1から出力される推定値を補正するように、補正部3に指示する。タイミング検出部2は、例えば、第1条件及び第2条件を満たす場合に、補正部3に推定部1から出力される推定値を補正するように指示してもよい。 The timing detection unit 2 instructs the correction unit 3 to correct the estimated value output from the estimation unit 1 when, for example, at least the first condition or the second condition is satisfied. The timing detection unit 2 may instruct the correction unit 3 to correct the estimated value output from the estimation unit 1 when, for example, the first condition and the second condition are satisfied.
 図3は、第1条件を説明するための模式図である。図3に示すグラフは、推定部1から出力される推定値の時間変化を示す。図3のグラフの横軸は時間tであり、縦軸は推定部1から出力される二酸化炭素濃度の推定値Vである。推定値は、一定間隔で出力されており、時間ごとに変化している。二酸化炭素濃度は、基準値Vbを基準に変動する。基準値Vbは、二酸化炭素濃度測定装置100が設置される環境に合わせて任意に設定できる。基準値Vbは、例えば、大気中のCO濃度であり、およそ400ppmである。 FIG. 3 is a schematic diagram for explaining the first condition. The graph shown in FIG. 3 shows the change over time of the estimated value output from the estimation unit 1. The horizontal axis of the graph in FIG. 3 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1. The estimated value is output at regular intervals and changes over time. The carbon dioxide concentration fluctuates based on a reference value Vb. The reference value Vb can be set arbitrarily according to the environment in which the carbon dioxide concentration measuring device 100 is installed. The reference value Vb is, for example, the CO2 concentration in the atmosphere, and is approximately 400 ppm.
 第1条件は、第1推定値V1と第2推定値V2との変化幅ΔVが第1変化幅ΔV1より大きく、かつ、第2推定値V2が第1規定値Vs1以下であるという条件である。 The first condition is that the change range ΔV between the first estimated value V1 and the second estimated value V2 is greater than the first change range ΔV1, and the second estimated value V2 is equal to or smaller than the first specified value Vs1.
 第1推定値V1は、第1時刻t1における推定値である。第2推定値V2は、第2時刻t2における推定値である。第1時刻t1と第2時刻t2とは時間的に連続し、第2時刻t2は第1時刻t1の次に推定部1が推定値を出力する時刻である。第1変化幅ΔV1は、測定環境に合わせて、任意に設定できる。第1規定値Vs1は、測定環境に合わせて、任意に設定できる。 The first estimated value V1 is an estimated value at the first time t1. The second estimated value V2 is an estimated value at the second time t2. The first time t1 and the second time t2 are consecutive in time, and the second time t2 is the time at which the estimation unit 1 outputs an estimated value after the first time t1. The first change width ΔV1 can be set arbitrarily in accordance with the measurement environment. The first specified value Vs1 can be set arbitrarily in accordance with the measurement environment.
 例えば、換気等で窓を開け、室内の二酸化炭素濃度が急激に下がった場合は、第1条件に該当する場合がある。この場合、基準値Vbは外気のCO濃度に設定されており、換気等で窓を開けることで、推定値は基準値Vbに近づくはずである。推定値が、基準値Vbより下回る第1規定値Vs1以下となるということは、推定値の正答率が下がっている可能性が高い。機械学習器10は、データが時間経過とともに連続的に変化する場合には高い正答率を維持できるが、データが急激に変化する場合に正答率が下がる傾向にある。そのため、第1条件を満たす場合に、推定値を補正すると、二酸化炭素濃度測定装置100から出力される二酸化炭素濃度の精度が高まる。 For example, when the window is opened for ventilation or the like and the carbon dioxide concentration in the room drops suddenly, the first condition may be met. In this case, the reference value Vb is set to the CO2 concentration of the outside air, and the estimated value should approach the reference value Vb by opening the window for ventilation or the like. If the estimated value is equal to or less than the first specified value Vs1 that is lower than the reference value Vb, it is highly likely that the accuracy rate of the estimated value has decreased. The machine learning device 10 can maintain a high accuracy rate when the data changes continuously over time, but the accuracy rate tends to decrease when the data changes suddenly. Therefore, when the first condition is met, correcting the estimated value increases the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100.
 図4は、第2条件を説明するための模式図である。図4に示すグラフは、推定部1から出力される推定値の時間変化を示す。図4のグラフの横軸は時間tであり、縦軸は推定部1から出力される二酸化炭素濃度の推定値Vである。 FIG. 4 is a schematic diagram for explaining the second condition. The graph shown in FIG. 4 shows the change over time of the estimated value output from the estimation unit 1. The horizontal axis of the graph in FIG. 4 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
 第2条件は、第1推定値V1’と第2推定値V2’との変化幅ΔV’が第2変化幅ΔV2より大きく、かつ、第2推定値V2’が第2規定値Vs2以上であるという条件である。 The second condition is that the change range ΔV' between the first estimated value V1' and the second estimated value V2' is greater than the second change range ΔV2, and the second estimated value V2' is greater than or equal to the second specified value Vs2.
 第1推定値V1’は、第1時刻t1’における推定値である。第2推定値V2’は、第2時刻t2’における推定値である。第1時刻t1’と第2時刻t2’とは時間的に連続し、第2時刻t2’は第1時刻t1’の次に推定部1が推定値を出力する時刻である。第2変化幅ΔV2は、測定環境に合わせて、任意に設定できる。第2変化幅ΔV2は、第1変化幅V1と同じとしてもよい。第2規定値Vs2は、測定環境に合わせて、任意に設定できる。第2規定値Vs2と基準値Vbとの差の絶対値は、第1規定値Vs1と基準値Vbとの差の絶対値と同じでもよい。 The first estimated value V1' is an estimated value at the first time t1'. The second estimated value V2' is an estimated value at the second time t2'. The first time t1' and the second time t2' are consecutive in time, and the second time t2' is the time at which the estimation unit 1 outputs an estimated value after the first time t1'. The second change width ΔV2 can be set arbitrarily in accordance with the measurement environment. The second change width ΔV2 may be the same as the first change width V1. The second specified value Vs2 can be set arbitrarily in accordance with the measurement environment. The absolute value of the difference between the second specified value Vs2 and the reference value Vb may be the same as the absolute value of the difference between the first specified value Vs1 and the reference value Vb.
 例えば、何かしらの理由で室内の二酸化炭素濃度が急激に上昇した場合は、第2条件に該当する場合がある。二酸化炭素濃度測定装置100が設置される環境において到達する可能性が低いCO濃度を第2規定値Vs2として設定した場合、推定値が第2規定値Vs2以上になるということは、推定値の正答率が下がっている可能性が高い。また機械学習器10は、上述のようにデータの急激な変化に対応することが苦手である。第2条件を満たす場合に、推定値を補正すると、二酸化炭素濃度測定装置100から出力される二酸化炭素濃度の精度が高まる。 For example, if the indoor carbon dioxide concentration rises suddenly for some reason, the second condition may be met. If the CO2 concentration that is unlikely to be reached in the environment in which the carbon dioxide concentration measuring device 100 is installed is set as the second specified value Vs2, the fact that the estimated value is equal to or greater than the second specified value Vs2 means that the accuracy rate of the estimated value is likely to be low. In addition, as described above, the machine learning device 10 is not good at responding to sudden changes in data. When the second condition is met, correcting the estimated value increases the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100.
 またタイミング検出部2は、第3条件を満たす場合に、補正部3に補正の指示を行ってもよい。第3条件は、第1推定値と第2推定値との変化幅が第3変化幅より大きくなるという条件である。第3条件は、変化幅のみに注目した条件であり、第1規定値Vs1及び第2規定値Vs2とは無関係である。第3変化幅は、第1変化幅ΔV1又は第2変化幅ΔV2と同じでもよいし、異なってもよい。上述のように、機械学習器10は、データの急激な変化に対応することが苦手である。第3条件を満たす場合に、推定値を補正すると、二酸化炭素濃度測定装置100から出力される二酸化炭素濃度の精度が高まる。 The timing detection unit 2 may also instruct the correction unit 3 to make a correction when the third condition is satisfied. The third condition is that the change range between the first estimated value and the second estimated value is greater than the third change range. The third condition is a condition that focuses only on the change range and is unrelated to the first specified value Vs1 and the second specified value Vs2. The third change range may be the same as or different from the first change range ΔV1 or the second change range ΔV2. As described above, the machine learning device 10 has difficulty responding to sudden changes in data. If the estimated value is corrected when the third condition is satisfied, the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100 is improved.
 またタイミング検出部2は、第4条件を満たす場合に、補正部3に補正の指示を行ってもよい。図5は、第4条件を説明するための模式図である。図5に示すグラフは、推定部1から出力される推定値の時間変化を示す。図5のグラフの横軸は時間tであり、縦軸は推定部1から出力される二酸化炭素濃度の推定値Vである。 The timing detection unit 2 may also instruct the correction unit 3 to make a correction when the fourth condition is satisfied. FIG. 5 is a schematic diagram for explaining the fourth condition. The graph shown in FIG. 5 shows the change over time in the estimated value output from the estimation unit 1. The horizontal axis of the graph in FIG. 5 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
 第4条件は、推定部1から出力される推定値が第1期間Δt1に亘って第3規定値Vs3以下となるという条件である。 The fourth condition is that the estimated value output from the estimation unit 1 is equal to or less than the third specified value Vs3 over the first period Δt1.
 第1期間Δt1は、測定環境に合わせて、任意に設定できる。第3規定値Vs3は、測定環境に合わせて、任意に設定できる。第3規定値Vs3は、第1規定値Vs1と同じでもよい。 The first period Δt1 can be set arbitrarily according to the measurement environment. The third specified value Vs3 can be set arbitrarily according to the measurement environment. The third specified value Vs3 may be the same as the first specified value Vs1.
 例えば、二酸化炭素濃度測定装置100が設置される環境において到達する可能性のあるCO濃度の下限を第3規定値Vs3として設定した場合、推定値が第3規定値Vs3以下である状態が一定期間維持されるということは、推定値の正答率が下がっている可能性が高い。そのため、第4条件を満たす場合に、推定値を補正すると、二酸化炭素濃度測定装置100から出力される二酸化炭素濃度の精度が高まる。 For example, if the lower limit of the CO2 concentration that may be reached in the environment in which the carbon dioxide concentration measuring device 100 is installed is set as the third specified value Vs3, the fact that the estimated value is maintained at or below the third specified value Vs3 for a certain period of time means that the accuracy rate of the estimated value is likely to have decreased. Therefore, if the estimated value is corrected when the fourth condition is satisfied, the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100 is improved.
 またタイミング検出部2は、第5条件を満たす場合に、補正部3に補正の指示を行ってもよい。図6は、第5条件を説明するための模式図である。図6に示すグラフは、推定部1から出力される推定値の時間変化を示す。図6のグラフの横軸は時間tであり、縦軸は推定部1から出力される二酸化炭素濃度の推定値Vである。 The timing detection unit 2 may also instruct the correction unit 3 to make a correction when the fifth condition is satisfied. FIG. 6 is a schematic diagram for explaining the fifth condition. The graph shown in FIG. 6 shows the change over time in the estimated value output from the estimation unit 1. The horizontal axis of the graph in FIG. 6 is time t, and the vertical axis is the estimated value V of the carbon dioxide concentration output from the estimation unit 1.
 第5条件は、推定部1から出力される推定値が第2期間Δt2に亘って第4規定値Vs4以上となるという条件である。 The fifth condition is that the estimated value output from the estimation unit 1 is equal to or greater than the fourth specified value Vs4 over the second period Δt2.
 第2期間Δt2は、測定環境に合わせて、任意に設定できる。第2期間Δt2は、第1期間Δt1と同じでも異なってもよい。第4規定値Vs4は、測定環境に合わせて、任意に設定できる。第4規定値Vs4は、第2規定値Vs2と同じでもよい。 The second period Δt2 can be set arbitrarily to suit the measurement environment. The second period Δt2 may be the same as or different from the first period Δt1. The fourth specified value Vs4 can be set arbitrarily to suit the measurement environment. The fourth specified value Vs4 may be the same as the second specified value Vs2.
 例えば、二酸化炭素濃度測定装置100が設置される環境において到達する可能性のあるCO濃度の上限を第4規定値Vs4として設定した場合、推定値が第4規定値Vs4以上である状態が一定期間維持されるということは、推定値の正答率が下がっている可能性が高い。そのため、第5条件を満たす場合に、推定値を補正すると、二酸化炭素濃度測定装置100から出力される二酸化炭素濃度の精度が高まる。 For example, if the upper limit of the CO2 concentration that may be reached in the environment in which the carbon dioxide concentration measuring device 100 is installed is set as the fourth specified value Vs4, the fact that the estimated value is maintained at or above the fourth specified value Vs4 for a certain period of time means that the accuracy rate of the estimated value is likely to have decreased. Therefore, if the estimated value is corrected when the fifth condition is satisfied, the accuracy of the carbon dioxide concentration output from the carbon dioxide concentration measuring device 100 is improved.
 補正部3は、例えば、マイクロコントローラー、プロセッサの一部である。補正部3は、例えば、CPU、ROM、RAM等を含む。 The correction unit 3 is, for example, a part of a microcontroller or a processor. The correction unit 3 includes, for example, a CPU, a ROM, a RAM, etc.
 補正部3は、機械学習器10の推定処理時に動作する。補正部3は、推定処理時に、タイミング検出部2から補正の指示を受ける毎に、推定値を補正するための補正値を決定する。補正値は、例えば、タイミング検出部2からの補正の指示があるたびに書き換えられる。 The correction unit 3 operates during the estimation process of the machine learning device 10. During the estimation process, the correction unit 3 determines a correction value for correcting the estimated value each time a correction instruction is received from the timing detection unit 2. The correction value is rewritten, for example, each time a correction instruction is received from the timing detection unit 2.
 図7は、補正部3による補正動作を示す図である。図7に示すように、例えば、第1条件を満たす場合は、基準値Vbと第2推定値V2との差の絶対値を補正値とする。補正部3は、例えば、第2時刻t2の次に推定値が出力される第3時刻t3以降に出力される推定値に補正値を追加する。 FIG. 7 is a diagram showing the correction operation by the correction unit 3. As shown in FIG. 7, for example, when the first condition is satisfied, the absolute value of the difference between the reference value Vb and the second estimated value V2 is set as the correction value. For example, the correction unit 3 adds the correction value to the estimated value output at or after the third time t3, when the estimated value is output next after the second time t2.
 また図8は、補正部3による補正動作の別の例を示す図である。例えば、図8に示すように、補正部3は、推定値に追加する値が補正値に至るまで、段階的に値を大きくしてもよい。例えば図8に示すように、補正部3は第3時刻t3から第4時刻t4に至るまでの第3期間Δt3において、徐々に大きな値を推定値に追加する。そして、第4時刻t4以降に出力される推定値に対して補正値を追加する。推定値に追加する値を補正値に至るまで徐々に大きくすると、出力される推定値が急激に変化することを防ぐことができる。 FIG. 8 is a diagram showing another example of the correction operation by the correction unit 3. For example, as shown in FIG. 8, the correction unit 3 may gradually increase the value added to the estimated value until it reaches the correction value. For example, as shown in FIG. 8, the correction unit 3 adds gradually larger values to the estimated value during a third period Δt3 from the third time t3 to the fourth time t4. Then, the correction value is added to the estimated value that is output after the fourth time t4. By gradually increasing the value added to the estimated value until it reaches the correction value, it is possible to prevent the output estimated value from changing suddenly.
 また例えば、タイミング検出部2が第2条件を満たすと判断した場合は、基準値Vbと第2推定値V2’との差の絶対値を補正値とする。補正部3は、例えば、第2時刻t2の次に推定値が出力される第3時刻t3以降に出力される推定値から補正値を差し引く。補正部3は、推定値から差し引く値が補正値に至るまで、推定値から差し引く値を段階的に大きくしてもよい。 For example, if the timing detection unit 2 determines that the second condition is satisfied, the absolute value of the difference between the reference value Vb and the second estimated value V2' is set as the correction value. The correction unit 3 subtracts the correction value from the estimated value that is output, for example, at or after the third time t3 when the estimated value is output next after the second time t2. The correction unit 3 may gradually increase the value that is subtracted from the estimated value until the value that is subtracted from the estimated value becomes the correction value.
 また例えば、タイミング検出部2が第3条件を満たすと判断した場合は、基準値Vbと第2推定値との絶対値を補正値とする。補正部3は、例えば、第2時刻t2の次に推定値が出力される第3時刻t3以降に出力される推定値に補正値を追加する、又は、推定値から補正値を差し引く。第2推定値が基準値Vbより小さい場合は、推定値に補正値を追加し、第2推定値が基準値Vbより大きい場合は、推定値から補正値を差し引く。推定値に追加または差し引く値は、補正値に至るまで段階的に大きくしてもよい。 Furthermore, for example, if the timing detection unit 2 determines that the third condition is satisfied, the absolute value of the reference value Vb and the second estimated value is set as the correction value. For example, the correction unit 3 adds the correction value to the estimated value output at or after the third time t3 when the estimated value is output next after the second time t2, or subtracts the correction value from the estimated value. If the second estimated value is smaller than the reference value Vb, the correction value is added to the estimated value, and if the second estimated value is larger than the reference value Vb, the correction value is subtracted from the estimated value. The value added to or subtracted from the estimated value may be increased stepwise until it reaches the correction value.
 また例えば、タイミング検出部2が第4条件を満たすと判断した場合は、基準値Vbと第1期間Δt1経過後に最初に出力される推定値との差の絶対値を補正値とする。補正部3は、第1期間Δt1経過後に出力される推定値に対して補正値を追加する。補正部3が追加する値は、補正値に至るまで段階的に大きくしてもよい。 For example, if the timing detection unit 2 determines that the fourth condition is satisfied, the absolute value of the difference between the reference value Vb and the estimated value that is first output after the first period Δt1 has elapsed is set as the correction value. The correction unit 3 adds the correction value to the estimated value that is output after the first period Δt1 has elapsed. The value added by the correction unit 3 may be increased stepwise until it reaches the correction value.
 また例えば、タイミング検出部2が第5条件を満たすと判断した場合は、基準値Vbと第2期間Δt2経過後に最初に出力される推定値との差の絶対値を補正値とする。補正部3は、第2期間Δt2経過後に出力される推定値から補正値を差し引く。補正部3が差し引く値は、補正値に至るまで段階的に大きくしてもよい。 For example, if the timing detection unit 2 determines that the fifth condition is satisfied, the absolute value of the difference between the reference value Vb and the estimated value that is first output after the second period Δt2 has elapsed is set as the correction value. The correction unit 3 subtracts the correction value from the estimated value that is output after the second period Δt2 has elapsed. The value that the correction unit 3 subtracts may be increased stepwise until it reaches the correction value.
 ここで機械学習器10の推定処理について説明する。推定部1は学習処理により教師データとの一致度が高められており、推定部1から出力される推定値の精度は、センサー20が出力する生データより高められている。機械学習器10は、推定部1からの推定値を二酸化炭素濃度の予測値として出力する。ただ、教師データ自体も較正が必要なように、推定部1から出力される推定値も経時変化等で基準がずれる場合もある。この場合、補正部3は推定値を補正する。 The estimation process of the machine learning device 10 will now be described. The estimation unit 1 has a higher degree of agreement with the teacher data through a learning process, and the accuracy of the estimated value output from the estimation unit 1 is higher than that of the raw data output by the sensor 20. The machine learning device 10 outputs the estimated value from the estimation unit 1 as a predicted value of the carbon dioxide concentration. However, just as the teacher data itself requires calibration, the estimated value output from the estimation unit 1 may also deviate from the standard due to changes over time, etc. In this case, the correction unit 3 corrects the estimated value.
 本実施形態に係る二酸化炭素濃度測定装置100は、機械学習器10を用いることで、二酸化炭素濃度の測定精度を高めることができる。また本実施形態に係る二酸化炭素濃度測定装置100は、補正部3が推定部1からの推定値を補正することで、二酸化炭素濃度の測定精度をより高めることができる。 The carbon dioxide concentration measuring device 100 according to this embodiment can improve the measurement accuracy of the carbon dioxide concentration by using the machine learning device 10. Furthermore, the carbon dioxide concentration measuring device 100 according to this embodiment can further improve the measurement accuracy of the carbon dioxide concentration by correcting the estimated value from the estimation unit 1 using the correction unit 3.
 図9は、第1実施形態に係る二酸化炭素濃度測定装置の測定結果の一例を示す。また図10は、補正部3を有さない二酸化炭素濃度測定装置の測定結果の一例を示す。図9及び図10におけるリファレンスは、教師データであり、ND-IR法を用いたCOセンサーの出力値である。 Fig. 9 shows an example of the measurement result of the carbon dioxide concentration measuring device according to the first embodiment. Also, Fig. 10 shows an example of the measurement result of the carbon dioxide concentration measuring device without the correction unit 3. The reference in Fig. 9 and Fig. 10 is teacher data, which is the output value of a CO2 sensor using the ND-IR method.
 図9及び図10に示す二酸化炭素濃度測定装置は、前段で学習処理を行い、後段で推定処理を行っている。学習処理は、タイミング検出部2が、補正が必要であると判断していない区間で行われている。学習処理では、タイミング検出部2及び補正部3をOFFとしている。 The carbon dioxide concentration measuring device shown in Figures 9 and 10 performs learning processing in the first stage and estimation processing in the second stage. The learning processing is performed in an interval where the timing detection unit 2 does not determine that correction is necessary. During the learning processing, the timing detection unit 2 and the correction unit 3 are turned off.
 図9及び図10に示すように、精度に劣るセンサー20からの出力値を用いた推定値は、リファレンスとの一致度が高く、第1実施形態に係る二酸化炭素濃度測定装置の二酸化炭素濃度の測定精度は、ND-IR法を用いたCOセンサーと同等に高いことが確認できる。 As shown in Figures 9 and 10, the estimated value using the output value from the less accurate sensor 20 has a high degree of agreement with the reference, and it can be confirmed that the measurement accuracy of the carbon dioxide concentration of the carbon dioxide concentration measuring device of the first embodiment is as high as that of a CO2 sensor using the ND-IR method.
 また図9及び図10に示す測定結果を比較すると、点線で囲まれた領域において、図9の方が推定値とリファレンスとの一致度が高い。これは、点線で囲まれた領域において、第1条件を満たす補正処理が行われたためである。すなわち、補正部3が機能することで、二酸化炭素濃度測定装置100の二酸化炭素濃度の測定精度がより高くなっていることが確認できる。 Comparing the measurement results shown in Figures 9 and 10, in the area surrounded by the dotted line, Figure 9 shows a higher degree of agreement between the estimated value and the reference. This is because a correction process that satisfies the first condition was performed in the area surrounded by the dotted line. In other words, it can be confirmed that the functioning of correction unit 3 improves the measurement accuracy of the carbon dioxide concentration by the carbon dioxide concentration measuring device 100.
 ここまで、第1実施形態を基に、本発明の好ましい態様を例示したが、本発明はこれらの実施形態に限られるものではない。 Up to this point, preferred aspects of the present invention have been illustrated based on the first embodiment, but the present invention is not limited to these embodiments.
 図11は、第1実施形態に係る二酸化炭素濃度測定装置101の模式図である。二酸化炭素濃度測定装置101は、比較部4と記憶部5とが機械学習器11の外部に配置されている。すなわち、比較部4と記憶部5とが、エッジ端末には設置されていない。学習処理は出荷時に事前に行っておくことも可能であり、エッジ端末から除くこともできる。また比較部4及び記憶部5を外部サーバーに保存し、エッジ端末と外部サーバーとが通信し、学習処理を更新してもよい。 FIG. 11 is a schematic diagram of a carbon dioxide concentration measuring device 101 according to the first embodiment. In the carbon dioxide concentration measuring device 101, the comparison unit 4 and memory unit 5 are arranged outside the machine learning device 11. In other words, the comparison unit 4 and memory unit 5 are not installed in the edge terminal. The learning process can be performed in advance at the time of shipment, or can be removed from the edge terminal. The comparison unit 4 and memory unit 5 can also be stored in an external server, and the edge terminal and the external server can communicate to update the learning process.
1…推定部、2…タイミング検出部、3…補正部、4…比較部、5…記憶部、10,11…機械学習器、20…センサー、100,101…二酸化炭素濃度測定装置、Lin…入力層、Lout…出力層、n…ノード、NN…ニューラルネットワーク、R…レザバー層、t1…第1時刻、t2…第2時刻、t3…第3時刻、t4…第4時刻、Δt1…第1期間、Δt2…第2期間、Δt3…第3期間、V1,V1’…第1推定値、V2,V2’…第2推定値、ΔV,ΔV’…変化幅、ΔV1…第1変化幅、ΔV2…第2変化幅、Vs1…第1規定値、Vs2…第2規定値、Vs3…第3規定値、Vs4…第4規定値、Vb…基準値 1...Estimation unit, 2...Timing detection unit, 3...Correction unit, 4...Comparison unit, 5...Memory unit, 10, 11...Machine learning machine, 20...Sensor, 100, 101...Carbon dioxide concentration measurement device, L in ...Input layer, L out ...Output layer, n i ...Node, NN...Neural network, R...Reservoir layer, t1...First time, t2...Second time, t3...Third time, t4...Fourth time, Δt1...First period, Δt2...Second period, Δt3...Third period, V1, V1'...First estimated value, V2, V2'...Second estimated value, ΔV, ΔV'...Change width, ΔV1...First change width, ΔV2...Second change width, Vs1...First specified value, Vs2...Second specified value, Vs3...Third specified value, Vs4...Fourth specified value, Vb...Reference value

Claims (9)

  1.  入力信号に基づき、二酸化炭素濃度を推定する推定部と、
     前記推定部から出力される推定値をモニターし、前記推定値の補正のタイミングを判断するタイミング検出部と、
     前記タイミング検出部から補正の指示を受ける毎に、前記推定値を補正するための補正値を決定する補正部と、を備え、
     前記タイミング検出部は、
    第1時刻における第1推定値と前記第1時刻に連続する第2時刻における第2推定値との変化幅が第1変化幅より大きく、かつ、前記第2推定値が第1規定値以下の場合、又は、
    前記第1推定値と前記第2推定値との変化幅が第2変化幅より大きく、かつ、前記第2推定値が第2規定値以上の場合に、
    前記補正部に補正の指示を行う、機械学習器。
    an estimation unit that estimates a carbon dioxide concentration based on an input signal;
    a timing detection unit that monitors the estimated value output from the estimation unit and determines a timing for correcting the estimated value;
    a correction unit that determines a correction value for correcting the estimated value every time a correction instruction is received from the timing detection unit,
    The timing detection unit
    A change range between a first estimate value at a first time and a second estimate value at a second time subsequent to the first time is greater than a first change range, and the second estimate value is equal to or smaller than a first specified value, or
    When a change range between the first estimated value and the second estimated value is larger than a second change range and the second estimated value is equal to or larger than a second specified value,
    A machine learning machine that instructs the correction unit to make corrections.
  2.  前記タイミング検出部は、
    前記推定部から出力される前記推定値が第1期間に亘って第3規定値以下となる場合に、前記補正部に補正の指示をさらに行う、請求項1に記載の機械学習器。
    The timing detection unit
    The machine learning machine according to claim 1 , further comprising: an instruction for correction to the correction unit when the estimated value output from the estimation unit is equal to or smaller than a third specified value over a first period.
  3.  前記タイミング検出部は、
    前記推定部から出力される前記推定値が第2期間に亘って第4規定値以上となる場合に、前記補正部に補正の指示をさらに行う、請求項1に記載の機械学習器。
    The timing detection unit
    The machine learning device according to claim 1 , further comprising: an instruction for correction to the correction unit when the estimated value output from the estimation unit is equal to or greater than a fourth specified value over a second period.
  4.  前記タイミング検出部は、前記第1推定値と前記第2推定値との変化幅が第3変化幅より大きくなる場合に、前記補正部に補正の指示をさらに行う、請求項1に記載の機械学習器。 The machine learning device according to claim 1, wherein the timing detection unit further instructs the correction unit to perform correction when the change range between the first estimated value and the second estimated value becomes larger than a third change range.
  5.  前記補正部は、前記推定値に前記補正値を追加する、又は、前記推定値から前記補正値を差し引く、請求項1に記載の機械学習器。 The machine learning device of claim 1, wherein the correction unit adds the correction value to the estimated value or subtracts the correction value from the estimated value.
  6.  前記補正部は、前記推定値に値を追加する、又は、前記推定値から値を差し引き、
     前記値は、前記補正値に至るまで段階的に大きくなる、請求項1に記載の機械学習器。
    The correction unit adds a value to the estimated value or subtracts a value from the estimated value,
    The machine learning machine of claim 1 , wherein the value increases in steps until it reaches the correction value.
  7.  前記推定部は、再帰型ニューラルネットワークである、請求項1に記載の機械学習器。 The machine learning device according to claim 1, wherein the estimation unit is a recurrent neural network.
  8.  前記推定部は、学習処理と推定処理とを行い、
     前記学習処理は、前記タイミング検出部が、補正が必要であると判断していない区間で行う、請求項1に記載の機械学習器。
    The estimation unit performs a learning process and an estimation process,
    The machine learning device according to claim 1 , wherein the learning process is performed in an interval in which the timing detection unit has determined that correction is not necessary.
  9.  請求項1に記載の機械学習器を備える、二酸化炭素濃度測定装置。 A carbon dioxide concentration measuring device equipped with the machine learning device according to claim 1.
PCT/JP2022/036655 2022-09-30 2022-09-30 Machine learning device and carbon dioxide concentration measurement device WO2024069919A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/JP2022/036655 WO2024069919A1 (en) 2022-09-30 2022-09-30 Machine learning device and carbon dioxide concentration measurement device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2022/036655 WO2024069919A1 (en) 2022-09-30 2022-09-30 Machine learning device and carbon dioxide concentration measurement device

Publications (1)

Publication Number Publication Date
WO2024069919A1 true WO2024069919A1 (en) 2024-04-04

Family

ID=90476956

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2022/036655 WO2024069919A1 (en) 2022-09-30 2022-09-30 Machine learning device and carbon dioxide concentration measurement device

Country Status (1)

Country Link
WO (1) WO2024069919A1 (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20010013026A1 (en) * 1998-04-17 2001-08-09 Ronald E. Shaffer Chemical sensor pattern recognition system and method using a self-training neural network classifier with automated outlier detection
WO2017002245A1 (en) * 2015-07-01 2017-01-05 三菱電機株式会社 Air-conditioning system control device and air-conditioning system
CN112683836A (en) * 2021-01-12 2021-04-20 杭州麦乐克科技股份有限公司 Calibration method and system of carbon dioxide sensor based on BP neural network
JP2021512322A (en) * 2018-03-28 2021-05-13 アイセンス,インコーポレーテッド Measurement method, device, learning method and system utilizing artificial neural network deep learning technique
US20220114485A1 (en) * 2020-10-09 2022-04-14 Metre, Inc. Systems and methods for predicting analyte concentrations via machine learning techniques

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20010013026A1 (en) * 1998-04-17 2001-08-09 Ronald E. Shaffer Chemical sensor pattern recognition system and method using a self-training neural network classifier with automated outlier detection
WO2017002245A1 (en) * 2015-07-01 2017-01-05 三菱電機株式会社 Air-conditioning system control device and air-conditioning system
JP2021512322A (en) * 2018-03-28 2021-05-13 アイセンス,インコーポレーテッド Measurement method, device, learning method and system utilizing artificial neural network deep learning technique
US20220114485A1 (en) * 2020-10-09 2022-04-14 Metre, Inc. Systems and methods for predicting analyte concentrations via machine learning techniques
CN112683836A (en) * 2021-01-12 2021-04-20 杭州麦乐克科技股份有限公司 Calibration method and system of carbon dioxide sensor based on BP neural network

Similar Documents

Publication Publication Date Title
AU2017317610B2 (en) Method and device for detecting equivalent load of wind turbine
US5554273A (en) Neural network compensation for sensors
KR101898629B1 (en) Method and device for post-correction of forecasted meteorological parameters by using a H-Infinity filter
EP0592796A1 (en) Method and apparatus for determining an air quality level
US20190064753A1 (en) Intelligent control method and system, and intelligent monitoring system
US11711429B2 (en) Method for calibrating networks of environmental sensors
CN111461909B (en) Short-term prediction method for farmland evapotranspiration
US20070019704A1 (en) Temperature sensor apparatus and method
Takruri et al. Spatio-temporal modelling-based drift-aware wireless sensor networks
KR20200063338A (en) Prediction system based on parameter improvement through learning and method thereof
WO2024069919A1 (en) Machine learning device and carbon dioxide concentration measurement device
CN111458395B (en) Kalman filtering method and device for changing Q value and readable storage medium
Outanoute et al. Synthesis of an Optimal Dynamic Regulator Based on Linear Quadratic Gaussian (LQG) for the Control of the Relative Humidity Under Experimental Greenhouse.
US20220279731A1 (en) Information processing device and method
CN111726407B (en) Fog calculation monitoring technology for famous flowers and medicinal plants cultivation in intelligent plant factory environment
Bontsema et al. On-line estimation of the transpiration in greenhouse horticulture
KR20180024171A (en) Server and method for determining actuator parameter of greenhouse
CN116373539A (en) Logistics information intelligent monitoring system and method based on Internet of things
CN108802153B (en) Calibration method for period fading, electronic device and storage medium
US20220373980A1 (en) Dymamic control of a manufacturing process using deep reinforcement learning
CN110530400A (en) Gyroscopic drift modification method, device, photoelectric nacelle and aircraft
Skufca Analysis Still Matters: A Surprising Instance of Failure of Runge--Kutta--Felberg ODE Solvers
JPH05307018A (en) Carbon dioxide gas concentration sensing device
Salsbury et al. Two new normalized EWMA-based indices for control loop performance assessment
US8781168B2 (en) Device for updating a photometric model

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22960978

Country of ref document: EP

Kind code of ref document: A1