Information processing method, information processing device, and program

Abstract

An information processing method for training an identification model for identifying an odor, wherein an information processing device: acquires a plurality of output signals that are obtained, from a plurality of sensor elements having different detection characteristics, when an odor sensor provided with the plurality of sensor elements has detected a gas containing an odor component that is a detection target; calculates a feature amount indicating the relationship related to the plurality of output signals; and trains the identification model by using training data including the feature amount.

Description

Information Processing Method, Information Processing Apparatus, and Program

[0001] The present disclosure relates to an information processing method, an information processing apparatus, and a program.

[0002] The following Patent Document 1 discloses a sensing system according to the background art. The sensing system includes a first odor sensor including at least one detection element that detects the amount of odor-causing substances present in the air, a filter that removes the odor-causing substances present in the air, a second odor sensor including at least one detection element that detects the amount of odor-causing substances present in the air that has passed through the filter, at least one first output signal detected by at least one detection element included in the first odor sensor, and at least one second output signal detected by the at least one detection element included in the second odor sensor. A difference calculation unit that calculates a difference between the signals to generate at least one detection signal, and a collection unit that collects information generated based on the at least one detection signal as odor information.

[0003] In the background art, there has been no consideration of executing each process of learning an identification model, identifying an odor, or generating learning data using a feature amount indicating the relationship between a plurality of output signals from a plurality of sensor elements included in an odor sensor.

[0004] International Publication No. 2019 / 187671

[0005] An object of the present disclosure is to obtain an information processing method, an information processing apparatus, and a program that can reduce the calculation cost of processing and shorten the required time of processing as compared with the case of executing processing using output signals from sensor elements.

[0006] An information processing method according to an aspect of the present disclosure is an information processing method for learning an identification model for identifying an odor, in which an information processing apparatus acquires a plurality of output signals from a plurality of sensor elements when a gas containing an odor component to be detected is detected by an odor sensor including a plurality of sensor elements having different detection characteristics, calculates a feature amount indicating the relationship between the plurality of output signals, and learns the identification model using the learning data including the feature amount.

[0007] This is a simplified diagram showing the configuration of a measuring device according to an embodiment of this disclosure. This is a simplified diagram showing the configuration of a gas generator. This is a simplified diagram showing the external configuration of an odor sensor. This is a schematic diagram showing a first example of the detection characteristics of a certain odor component for a certain sensor element. This is a schematic diagram showing a second example of the detection characteristics of a certain odor component for a certain sensor element. This is a flowchart showing the processing flow executed by the information processing device. This is a simplified diagram showing the configuration of a learning device according to an embodiment of this disclosure. This is a flowchart showing the processing flow executed by the information processing device. This is a diagram showing a first example of training data. This is a diagram showing a second example of training data. This is a diagram showing a simplified configuration of an odor identification device according to an embodiment of this disclosure. This is a flowchart showing the processing flow executed by the information processing device. This is a schematic diagram showing an example of how the odor identification result is displayed on a display device. This is a simplified diagram showing the configuration of a learning data generation device. This is a flowchart showing the processing flow executed by the information processing device. This is a schematic diagram showing a first example of training data. This is a schematic diagram showing a second example of training data. This is a flowchart showing a first example of the processing flow executed by the information processing device. This is a flowchart showing a second example of the processing flow executed by an information processing device. This is a diagram showing a first example of dividing an output signal into multiple subsignals. This is a diagram showing a second example of dividing an output signal into multiple subsignals. This is a diagram showing a third example of dividing an output signal into multiple subsignals. This is a diagram showing a fourth example of dividing an output signal into multiple subsignals.

[0008] (Knowledge forming the basis of this disclosure) Research is underway on odor identification devices used in artificial olfactory sensors and the like. The odor identification device acquires the output signal from the odor sensor when it detects an odor, and identifies the odor using a trained identification model that identifies odors.

[0009] Odor sensors are equipped with multiple sensor elements that have different detection characteristics. Conventionally, a discrimination model is trained using deep learning or the like with training data that includes multiple output signals from multiple sensor elements. However, because the amount of training data is enormous due to the inclusion of multiple output signals, the computational cost of processing increases, and the processing time becomes longer.

[0010] To solve these problems, the inventors have found that by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to when processing is performed using multiple output signals from multiple sensor elements, and have come up with this disclosure.

[0011] Next, we will describe each aspect of this disclosure.

[0012] An information processing method according to a first aspect of the present disclosure is an information processing method for learning an identification model for identifying odors, wherein an information processing device acquires multiple output signals from a plurality of sensor elements, which are equipped with a plurality of sensor elements having different detection characteristics, when an odor sensor detects a gas containing an odor component to be detected, calculates a feature quantity that shows the relationship between the plurality of output signals, and learns the identification model using training data including the feature quantity.

[0013] According to the first embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to when the learning process of the discrimination model is performed using multiple output signals.

[0014] An information processing method according to a second aspect of the present disclosure is an information processing method for identifying an odor using a trained identification model, wherein the information processing device acquires the identification model, acquires a plurality of output signals from a plurality of sensor elements having different detection characteristics when an odor sensor detects a gas containing an odor component, calculates a feature quantity that shows the relationship between the plurality of output signals, and identifies the odor component detected by the odor sensor based on the identification model and the feature quantity.

[0015] According to the second embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to the case in which odor identification processing is performed using multiple output signals.

[0016] An information processing method according to a third aspect of the present disclosure is an information processing method for generating training data to be used for training an identification model for identifying odors, wherein the information processing device generates a gas containing odor components to be detected in a gas generator, has an odor sensor equipped with a plurality of sensor elements having different detection characteristics detect the gas generated by the gas generator, acquires a plurality of output signals from the plurality of sensor elements when the gas is detected, calculates a feature quantity that shows the relationship between the plurality of output signals, and generates training data including the feature quantity.

[0017] According to the third embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to the case in which the data generation process for training is performed using multiple output signals.

[0018] In the information processing method according to the fourth aspect of this disclosure, in any one of the first to third aspects, the relationship may include the similarity or ratio of the plurality of output signals.

[0019] According to the fourth embodiment, appropriate feature quantities can be calculated by including the similarity or ratio between multiple output signals in the relationship.

[0020] In the information processing method according to the fifth aspect of this disclosure, in the fourth aspect, the feature quantity may be calculated with respect to all combinations of the plurality of output signals, or with respect to a plurality of combinations selected from all combinations of the plurality of output signals.

[0021] According to the fifth embodiment, appropriate feature quantities can be calculated according to the required accuracy or time, etc.

[0022] In the sixth aspect of the information processing method of this disclosure, in any one of the first to third aspects, the relationship may include the similarity or ratio between the plurality of output signals and the plurality of reference signals.

[0023] According to the sixth embodiment, appropriate feature quantities can be calculated by including the similarity or ratio between multiple output signals and multiple reference signals in the relationship.

[0024] In the information processing method according to the seventh aspect of this disclosure, in the sixth aspect, the feature quantity may be calculated with respect to all combinations of the plurality of output signals and the plurality of reference signals, or with respect to a plurality of combinations selected from all combinations of the plurality of output signals and the plurality of reference signals.

[0025] According to the seventh embodiment, appropriate feature quantities can be calculated according to the required accuracy or time, etc.

[0026] In the information processing method according to the eighth aspect of this disclosure, in any one of the first to third aspects, in calculating the feature quantity, the output signal is divided into a plurality of subsignals corresponding to a plurality of intervals, and the feature quantity is calculated for each subsignal.

[0027] According to the eighth aspect, the accuracy of learning or classification can be improved by dividing the output signal into multiple subsignals and calculating feature quantities for each subsignal.

[0028] In the information processing method relating to the ninth aspect of this disclosure, in the eighth aspect, the interval may be a certain period of time.

[0029] According to the ninth aspect, the output signal can be easily divided into multiple subsignals.

[0030] In the information processing method according to the tenth aspect of this disclosure, in the eighth aspect, the interval is preferably a variable period set based on control conditions relating to the control of causing the odor sensor to detect the gas.

[0031] According to the tenth embodiment, the output signal can be appropriately divided into multiple subsignals according to the control conditions.

[0032] The information processing method according to the eleventh aspect of this disclosure further involves performing preprocessing on the acquired plurality of output signals in any one of the first to third aspects, and in calculating the feature quantities, it is preferable to calculate the feature quantities with respect to the plurality of output signals on which the preprocessing has been performed.

[0033] According to the eleventh embodiment, the stability of learning or classification can be improved by calculating feature quantities for multiple output signals that have undergone preprocessing.

[0034] An information processing device according to a twelfth aspect of the present disclosure is an information processing device for learning an identification model for identifying odors, comprising a circuit configuration, the circuit configuration acquires multiple output signals from a plurality of sensor elements, which have different detection characteristics, when an odor sensor detects a gas containing odor components to be detected, calculates feature quantities that show the relationships between the plurality of output signals, and learns the identification model using training data including the feature quantities.

[0035] According to the twelfth embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to the case in which the learning process of the discrimination model is performed using multiple output signals.

[0036] An information processing device according to a thirteenth aspect of the present disclosure is an information processing device that identifies odors using a trained identification model, and comprises a circuit configuration, the circuit configuration acquires the identification model, acquires a plurality of output signals from a plurality of sensor elements having different detection characteristics when an odor sensor detects a gas containing odor components, calculates a feature quantity that shows the relationship between the plurality of output signals, and identifies the odor components detected by the odor sensor based on the identification model and the feature quantity.

[0037] According to the 13th embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to the case in which odor identification processing is performed using multiple output signals.

[0038] An information processing device according to a fourteenth aspect of the present disclosure is an information processing device that generates training data for training an identification model for identifying odors, and comprises a circuit configuration, the circuit configuration generates a gas containing odor components to be detected in a gas generator, the gas generated by the gas generator is detected by an odor sensor having a plurality of sensor elements with different detection characteristics, a plurality of output signals from the plurality of sensor elements when the gas is detected is acquired, a feature quantity showing the relationship between the plurality of output signals is calculated, and training data including the feature quantity is generated.

[0039] According to the 14th embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to the case in which the data generation process for training is performed using multiple output signals.

[0040] A program according to a 15th aspect of the present disclosure is a program for causing an information processing device that learns an identification model for identifying odors to perform a process, wherein the process involves acquiring multiple output signals from a plurality of sensor elements, which are equipped with a plurality of sensor elements having different detection characteristics, when an odor sensor detects a gas containing an odor component to be detected, calculates a feature quantity that shows the relationship between the plurality of output signals, and learns the identification model using training data including the feature quantity.

[0041] According to the 15th embodiment, by using feature quantities that show the relationships between multiple output signals from multiple sensor elements, the computational cost of processing can be reduced and the processing time can be shortened compared to the case in which the learning process of the discrimination model is performed using multiple output signals.

[0042] The program according to the 16th aspect of the present disclosure is a program for causing an information processing apparatus that identifies odors using a learned identification model to execute a process, the process including: obtaining the identification model; obtaining a plurality of output signals from a plurality of sensor elements when a gas containing an odor component is detected by an odor sensor including a plurality of sensor elements having different detection characteristics; calculating a feature amount indicating a relationship regarding the plurality of output signals; and identifying the odor component detected by the odor sensor based on the identification model and the feature amount.

[0043] According to the 16th aspect, by using a feature amount indicating a relationship regarding a plurality of output signals from a plurality of sensor elements, the computational cost of the process can be reduced and the required time of the process can be shortened as compared with the case of executing an odor identification process using the plurality of output signals.

[0044] The program according to the 17th aspect of the present disclosure is a program for causing an information processing apparatus that generates learning data used for learning an identification model for identifying an odor to execute a process, the process including: causing a gas generator to generate a gas containing an odor component to be detected; causing the odor sensor including a plurality of sensor elements having different detection characteristics to detect the gas generated by the gas generator; obtaining a plurality of output signals from the plurality of sensor elements when the gas is detected; calculating a feature amount indicating a relationship regarding the plurality of output signals; and generating learning data including the feature amount.

[0045] According to the 17th aspect, by using a feature amount indicating a relationship regarding a plurality of output signals from a plurality of sensor elements, the computational cost of the process can be reduced and the required time of the process can be shortened as compared with the case of executing a process for generating learning data using the plurality of output signals.

[0046] The present disclosure can also be realized as a program for causing a computer to execute each characteristic configuration included in such a method or apparatus, or as a system operated by this program. Needless to say, such a computer program can be distributed via a computer-readable non-temporary recording medium such as a CD-ROM or a communication network such as the Internet.

[0047] (Embodiments of the Present Disclosure) Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. Elements denoted by the same reference numerals in different drawings indicate the same or corresponding elements. Also, the components, the arrangement positions of the components, the connection forms, the order of operations, etc. shown in the following embodiments are merely examples and are not intended to limit the present disclosure. The present disclosure is limited only by the claims. Therefore, among the components in the following embodiments, the components not described in the independent claims indicating the most general concept of the present disclosure are not necessarily required to achieve the problems of the present disclosure, but are described as constituting a more preferable form.

[0048] (Odor Signal Measurement Phase) FIG. 1 is a diagram schematically showing the configuration of a measurement device 10 according to an embodiment of the present disclosure. The measurement device 10 acquires an odor signal by measuring a gas containing an odor component to be detected prior to the learning of an identification model 27 described later for identifying an odor. The measurement device 10 includes a personal computer, an edge device, a cloud server, or the like. The measurement device 10 includes an information processing device 11 and a storage device 12. The information processing device 11 includes a circuit configuration such as a processor. The storage device 12 includes a storage medium such as an HDD, SSD, or semiconductor memory. The measurement device 10 may be physically divided into a plurality of devices, and the plurality of devices may be connected to each other via a wired cable or a wireless communication network.

[0049] The storage device 12 stores an output signal set S1 and a program 16. The storage device 12 includes a computer-readable non-transitory storage medium (such as a ROM), and the program 16 is stored in the storage medium. As functions realized by the processor executing the program 16 read from the storage device 12, the information processing device 11 has an acquisition unit 13 and a control unit 14. Details of the processing contents of each unit will be described later. An odor sensor 1 and a gas generation device 2 are connected to the measurement device 10.

[0050] Figure 2 is a simplified diagram showing the configuration of the gas generator 2. The gas generator 2 comprises a controller 50, a liquid tank 51, a valve 54A provided in the piping connecting the liquid tank 51 and the carrier gas supply source 52, a valve 54B provided in the piping connecting the liquid tank 51 and the chamber 53, and a valve 55 provided in the piping connecting the supply source 52 and the chamber 53. The carrier gas is nitrogen gas or the like. The chamber 53 is a constant temperature bath or the like, and the odor sensor 1 is placed inside the chamber 53.

[0051] The liquid tank 51 stores liquefied gas of the odor components to be detected by the odor sensor 1. If there are multiple types of odor components to be detected, multiple liquid tanks 51 are provided according to the number of types.

[0052] The controller 50 controls the flow rate of gas passing through each valve by controlling the opening degree of valves 54A, 54B, and 55 based on the control signal P1 input from the measuring device 10. This allows gas G1, in which the odor component to be detected has been adjusted to the desired concentration, to be supplied into the chamber 53.

[0053] Although Figure 2 shows an example of a bubbling-type gas generator 2, other methods such as direct vaporization or baking may also be used.

[0054] Figure 3 shows a simplified external configuration of the odor sensor 1. The odor sensor 1 comprises multiple sensor elements X (16 sensor elements Xa to Xp in the example of Figure 3) of different materials, arranged in a matrix. Note that the number of sensor elements X is not limited to 16; any number is acceptable. Each sensor element X is also referred to as a channel, and the number of sensor elements X is also referred to as the number of channels.

[0055] Figure 4 schematically shows a first example of the detection characteristics K1 for a certain odor component for a certain sensor element X. Figure 4 shows an example of the detection characteristics K1 for a fluid-controlled odor sensor 1. In the control profile of the fluid-controlled method, the control conditions for controlling the odor sensor 1 to detect gas G1 are as follows: before time T11, a reference gas is supplied into the chamber 53; from time T11 to time T12, a sample gas is supplied into the chamber 53; and after time T12, a reference gas is supplied into the chamber 53. The reference gas is a gas that does not contain the odor component to be detected, and may be a carrier gas. The sample gas is a gas in which the odor component to be detected is adjusted to a desired concentration.

[0056] During the period from time T11 to time T12, sensor elements Xa to Xp detect odor components as resistance values ​​that increase or decrease according to the amount of odor molecules adhering to the surface of the sensor elements. Since the ease with which odor molecules adhere to the surface of the sensor elements differs depending on the material, the detection characteristics K1a to K1p, which represent the maximum value or the manner of change of the resistance value, differ for each sensor element Xa to Xp. The odor sensor 1 outputs multiple output signals that represent multiple detection characteristics K1a to K1p related to the multiple sensor elements Xa to Xp. Output signal set S1 includes multiple output signals output from the multiple sensor elements Xa to Xp.

[0057] Figure 5 schematically shows a second example of the detection characteristics K2 for a certain odor component with respect to a certain sensor element X. Figure 5 shows an example of the detection characteristics K2 for a thermally controlled odor sensor 1. In the thermally controlled control profile, the control conditions for controlling the odor sensor 1 to detect gas G1 are that a sample gas is supplied into the chamber 53 for the entire period, and a heating period or a cooling period is set by controlling the on or off of the heater built into the sensor 1. The period before time T21 is set as the heating period, the period from time T21 to time T22 is set as the cooling period, and the period after time T22 is set as the heating period. Sensor elements Xa to Xp detect odor components as resistance values ​​that increase or decrease according to the amount of odor molecules attached to the surface of the sensor elements. Since the ease with which odor molecules attach to the surface of the sensor elements differs depending on the material and temperature, the detection characteristics K2a to K2p, which represent the maximum value or change in resistance value, differ for each sensor element Xa to Xp. The odor sensor 1 outputs multiple output signals that represent multiple detection characteristics K2a to K2p related to multiple sensor elements Xa to Xp. The output signal set S1 includes multiple output signals output from the multiple sensor elements Xa to Xp.

[0058] Figure 6 is a flowchart showing the processing flow executed by the information processing device 11.

[0059] First, in step SP01, the acquisition unit 13 reads the measurement setting file. The measurement setting file includes, for example, various parameters related to the operation control of the gas generator 2, gas generation conditions such as the type or concentration of gas to be generated by the gas generator 2, and environmental conditions such as the temperature or humidity of the chamber 53. The parameters related to the operation control of the gas generator 2 include heating time, cooling time, reference gas supply time, and sample gas supply time.

[0060] Next, in step SP02, the control unit 14 sets the environmental conditions of the chamber 53 based on the measurement setting file acquired in step SP01, and waits until the chamber 53 stabilizes at the set temperature or set humidity.

[0061] Next, in step SP03, the control unit 14 sets a control signal P1 based on the measurement setting file acquired in step SP01, and generates gas G1 in the gas generator 2 by inputting the set control signal P1 to the gas generator 2. Gas G1 contains components to be detected by the odor sensor 1. Gas G1 may further contain components that are not detected by the odor sensor 1. By including components that are not detected by the odor sensor 1 in gas G1, the range of adjustment for the concentration of the components to be detected in gas G1 is widened. Gas G1 is supplied from the gas generator 2 into the chamber 53.

[0062] Next, in step SP04, the control unit 14 sets a control signal Q1 based on the measurement setting file acquired in step SP01, and inputs the set control signal Q1 to the odor sensor 1, causing the odor sensor 1 to detect the gas G1. The odor sensor 1 outputs an output signal set S1 related to the detection result of the gas G1. The output signal set S1 includes multiple output signals output from multiple sensor elements Xa to Xp. The multiple output signals include multiple characteristics K1a to K1p or multiple characteristics K2a to K2p related to the multiple sensor elements Xa to Xp when the gas G1 is detected.

[0063] Next, in step SP05, the acquisition unit 13 acquires the output signal set S1 from the odor sensor 1.

[0064] Next, in step SP06, the acquisition unit 13 stores the output signal set S1 acquired in step SP05 in the storage device 12. In the storage device 12, the output signal set S1 is configured as a database having multiple records. The acquisition unit 13 adds the output signal set S1 acquired in step SP05 to the database as a new record.

[0065] Next, in step SP07, the control unit 14 determines whether or not the measurement has been completed for all gas generation conditions specified in the measurement setting file.

[0066] If there are unmeasured gas generation conditions (step SP07: NO), the information processing device 11 executes the following processes with respect to the unmeasured conditions.

[0067] If there are no unmeasured gas generation conditions (Step SP07: YES), the information processing device 11 terminates the measurement process.

[0068] In the above embodiment, the output signal set S1 was a measured signal output from the odor sensor 1. However, as a modification, the output signal set S1 may also include a generated signal generated using a simulation model of the odor sensor 1. The simulation model is a model for mathematically reproducing the detection operation of the odor sensor 1 according to gas generation conditions such as the type and concentration of gas generated by the gas generator 2. The simulation model generates and outputs the output signal set S1 from the odor sensor 1 when the odor sensor 1 detects the gas G1 generated by the gas generator 2 based on the control signal P1, and the odor sensor 1 detects the gas based on the control signal Q1. The generated signal, output signal set S1, is acquired by the acquisition unit 13. By including both measured and generated signals in the output signal set S1, a large amount of training data necessary for training the identification model 27 can be efficiently collected.

[0069] (Generating the Identification Model) Figure 7 is a simplified diagram showing the configuration of the learning device 20 according to the embodiment of this disclosure. The learning device 20 learns an identification model 27 for identifying odors by machine learning using the training data D1. The learning device 20 may be the same device as the measuring device 10, or it may be a separate device.

[0070] The learning device 20 includes a personal computer, an edge device, or a cloud server. The learning device 20 comprises an information processing device 21 and a storage device 22. The information processing device 21 includes a circuit configuration such as a processor. The storage device 22 includes a storage medium such as an HDD, SSD, or semiconductor memory. The learning device 20 may be physically divided into multiple devices, and these multiple devices may be interconnected via wired cables or a wireless communication network.

[0071] The storage device 22 stores the output signal set S1, learning data D1, identification model 27, and program 28. If the learning device 20 is a common device with the measurement device 10, the output signal set S1 stored in the storage device 12 can be used as is with the output signal set S1 stored in the storage device 22. If the learning device 20 is a separate device from the measurement device 10, the measurement device 10 transfers the output signal set S1 stored in the storage device 12 to the learning device 20, and the learning device 20 stores the transferred output signal set S1 in the storage device 22. The storage device 22 includes a computer-readable non-temporary storage medium (ROM, etc.), and the program 28 is stored in this storage medium. As a function realized by the processor executing the program 28 read from the storage device 22, the information processing device 21 has an acquisition unit 23, a pre-processing unit 24, a calculation unit 25, and a learning unit 26. Details of the processing content of each unit will be described later.

[0072] Figure 8 is a flowchart showing the processing flow executed by the information processing device 21.

[0073] First, in step SP11, the acquisition unit 23 reads the training configuration file. The training configuration file contains, for example, various parameters necessary for machine learning of the discrimination model 27.

[0074] Next, in step SP12, the acquisition unit 23 acquires the output signal set S1 from the storage device 22.

[0075] Next, in step SP13, the preprocessing unit 24 performs preprocessing on the output signal set S1 acquired in step SP12. Preprocessing includes temperature correction, humidity correction, or normalization processing. Note that preprocessing may be omitted.

[0076] Next, in step SP14, the calculation unit 25 calculates feature quantities that indicate the relationship between channels, that is, feature quantities that indicate the relationship between multiple output signals included in the output signal set S1, based on the output signal set S1. The relationship includes the similarity or ratio between the multiple output signals.

[0077] When the output signal of the first channel is X and the output signal of the second channel is Y, the calculation unit 25 calculates the similarity S(X,Y) between the two channels using the following formula (1). The similarity S(X,Y) includes cosine similarity or Euclidean distance, etc.

[0078]

[0079] Furthermore, when the output signal of the first channel is X and the output signal of the second channel is Y, the calculation unit 25 calculates the ratio R(X,Y) between the two channels using the following formula (2).

[0080]

[0081] Next, in step SP15, the calculation unit 25 generates training data D1 that includes the features calculated in step SP14. The calculation unit 25 may also store the generated training data D1 in the storage device 22.

[0082] Figure 9 shows a first example of training data D1. Training data D1 includes similarity S(X,Y) for all combinations of two output signals included in output signal set S1. For example, similarity S(a,b) indicates the similarity between the output signal of the channel corresponding to sensor element Xa and the output signal of the channel corresponding to sensor element Xb, and similarity S(d,p) indicates the similarity between the output signal of the channel corresponding to sensor element Xd and the output signal of the channel corresponding to sensor element Xp. Note that the calculation unit 25 may calculate similarity S(X,Y) not only for all combinations, but also for any multiple combinations selected from all combinations of multiple output signals (for example, combinations in which the similarity is characteristic).

[0083] Figure 10 shows a second example of the training data D1. The training data D1 includes ratios R(X,Y) for all combinations of two output signals included in the output signal set S1. For example, ratio R(a,b) represents the ratio of the output signal of the channel corresponding to sensor element Xa to the output signal of the channel corresponding to sensor element Xb, and ratio (d,p) represents the ratio of the output signal of the channel corresponding to sensor element Xd to the output signal of the channel corresponding to sensor element Xp. Note that the calculation unit 25 may calculate ratios R(X,Y) not only for all combinations, but also for any multiple combinations selected from all combinations of multiple output signals (for example, combinations with characteristic similarity).

[0084] Next, in step SP16, the learning unit 26 learns the discrimination model 27 using machine learning with the training data D1. As a learning algorithm, for example, a deep neural network can be used, but any learning method is acceptable, not limited to this example.

[0085] Next, in step SP17, the learning unit 26 stores the learned identification model 27 in the storage device 22.

[0086] According to the learning device 20 of this embodiment, the learning process of the discrimination model 27 is performed using feature quantities that show the relationships between multiple output signals from multiple sensor elements X. This reduces the computational cost of the process and shortens the processing time compared to the case in which the learning process of the discrimination model is performed using multiple output signals.

[0087] (Odor Identification Phase) Figure 11 is a simplified diagram showing the configuration of an odor identification device 30 according to an embodiment of the present disclosure. The odor identification device 30 identifies the odor detected by the odor sensor 1 using a learned identification model 27. The odor identification device 30 may be a device common to the measuring device 10 or the learning device 20, or it may be a separate device.

[0088] The odor identification device 30 includes a personal computer, an edge device, or a cloud server. The odor identification device 30 comprises an information processing device 31, a storage device 32, and a display device 39. The information processing device 31 includes a circuit configuration such as a processor. The storage device 32 includes a storage medium such as an HDD, SSD, or semiconductor memory. The display device 39 includes an LCD display or an organic EL display. The odor identification device 30 may be physically divided into multiple devices, and these multiple devices may be interconnected via wired cables or a wireless communication network.

[0089] The storage device 32 stores the identification model 27 and the program 37. If the odor identification device 30 is a common device with the learning device 20, the learned identification model 27 stored in the storage device 22 can be used as is, as the identification model 27 stored in the storage device 32. If the odor identification device 30 is a separate device from the learning device 20, the learning device 20 transfers the identification model 27 stored in the storage device 22 to the odor identification device 30, and the odor identification device 30 stores the transferred identification model 27 in the storage device 32. The storage device 32 includes a computer-readable non-temporary storage medium (ROM, etc.), and the program 37 is stored in this storage medium. As a function realized by the processor executing the program 37 read from the storage device 32, the information processing device 31 has an acquisition unit 33, a pre-processing unit 34, a calculation unit 35, and an identification unit 36. The identification unit 36 ​​includes an inference unit 38. Details of the processing content of each unit will be described later.

[0090] Figure 12 is a flowchart showing the processing flow executed by the information processing device 31.

[0091] First, in step SP21, the acquisition unit 33 reads the identification setting file. The identification setting file contains, for example, various parameters necessary for odor identification using the identification model 27 by the identification unit 36.

[0092] Next, in step SP22, the acquisition unit 33 acquires the trained identification model 27 from the storage device 32. If the odor identification device 30 is an edge device and the training of the identification model 27 is performed by a cloud server, the odor identification device 30 may receive the trained identification model 27 from the cloud server and store it in the storage device 32.

[0093] Next, in step SP23, the acquisition unit 33 acquires an output signal set S2 from the odor sensor 1, which relates to the detection result of a gas containing odor components detected by the odor sensor 1. The output signal set S2 includes a plurality of characteristics K related to the plurality of sensor elements X when the gas is detected.

[0094] Next, in step SP24, the preprocessing unit 34 performs preprocessing on the output signal set S2 acquired in step SP23. Preprocessing includes temperature correction, humidity correction, or normalization processing. Note that preprocessing may be omitted.

[0095] Next, in step SP25, the calculation unit 35 calculates feature quantities that indicate the relationship between channels, that is, feature quantities that indicate the relationship between the multiple output signals included in the output signal set S2, based on the output signal set S2. Similarly, the relationship includes the similarity or ratio between the multiple output signals.

[0096] Next, in step SP26, the identification unit 36 ​​inputs the feature quantities calculated in step SP25 into the identification model 27 acquired in step SP22, and performs inference using the identification model 27 to identify the odor components detected by the odor sensor 1.

[0097] The identification unit 36 ​​includes an inference unit 38. Based on the feature quantities calculated in step SP25, the inference unit 38 infers the type and concentration of the odor component detected by the odor sensor 1. If there are multiple types of odor components to be detected, the inference unit 38 performs inference for each type.

[0098] Next, in step SP27, the identification unit 36 ​​outputs data D10 indicating the odor identification result based on inference. Data D10 is input from the information processing device 31 to the display device 39, and the display device 39 displays the odor identification result based on data D10.

[0099] When inference is performed on a cloud server, the cloud server has an identification unit 36 ​​and stores the identification model 27. In this case, the odor identification device 30 may transmit the output signal set S2 acquired by the acquisition unit 33 from the odor sensor 1 to the cloud server and receive data D10 indicating the odor identification result from the cloud server.

[0100] Figure 13 schematically shows an example of how the odor identification results are displayed in the display device 39. In the example shown in Figure 13, the concentrations of each of the eight odor components A to H are shown as concentration points, where the distance from the center of the circle increases as the concentration increases. The inside of the figure obtained by connecting adjacent concentration points with straight lines is highlighted by coloring or the like. Note that the visualization method for the odor identification results is not limited to the example in Figure 13, and any visualization method such as numerical display of concentration values, line graph display, or histogram display may be used.

[0101] According to the odor identification device 30 of this embodiment, the odor component detected by the odor sensor 1 is identified based on a feature quantity that shows the relationship between multiple output signals from multiple sensor elements X. This reduces the computational cost of processing and shortens the processing time compared to when odor identification processing is performed using multiple output signals.

[0102] (First Modification) The learning data generation device 40 may be configured by combining the measuring device 10 with a part of the learning device 20. The learning data generation device 40 generates learning data D1 used for training the identification model 27 for identifying odors. The learning data generation device 40 includes a personal computer, an edge device, or a cloud server.

[0103] Figure 14 is a simplified diagram showing the configuration of the learning data generation device 40. The learning data generation device 40 comprises an information processing device 41 and a storage device 42. The information processing device 41 includes a circuit configuration such as a processor. The storage device 42 includes a storage medium such as an HDD, SSD, or semiconductor memory. The learning data generation device 40 may be physically divided into multiple devices, and these multiple devices may be interconnected via wired cables or a wireless communication network.

[0104] The storage device 42 stores the output signal set S1, the learning data D1, and the program 47. The storage device 42 includes a computer-readable non-temporary storage medium (such as ROM), and the program 47 is stored in this storage medium. The information processing device 41 has an acquisition unit 43, a control unit 44, a pre-processing unit 45, and a calculation unit 46, which are functions realized when the processor executes the program 47 read from the storage device 42. Details of the processing content of each unit will be described later. The learning data generation device 40 is connected to an odor sensor 1 and a gas generator 2.

[0105] Figure 15 is a flowchart showing the processing flow executed by the information processing device 41.

[0106] First, in step SP31, the acquisition unit 43 reads the measurement setting file.

[0107] Next, in step SP32, the control unit 44 sets the environmental conditions of the chamber 53 based on the measurement setting file acquired in step SP31, and waits until the chamber 53 stabilizes at the set temperature or set humidity.

[0108] Next, in step SP33, the control unit 44 sets a control signal P1 based on the measurement setting file acquired in step SP41, and inputs the set control signal P1 to the gas generator 2, thereby generating gas G1 in the gas generator 2.

[0109] Next, in step SP34, the control unit 44 sets a control signal Q1 based on the measurement setting file acquired in step SP41, and inputs the set control signal Q1 to the odor sensor 1, causing the odor sensor 1 to detect the gas G1.

[0110] Next, in step SP35, the acquisition unit 43 acquires the output signal set S1 from the odor sensor 1. The acquisition unit 43 may also store the acquired output signal set S1 in the storage device 12.

[0111] Next, in step SP36, the preprocessing unit 45 performs preprocessing on the output signal set S1. Preprocessing includes temperature correction, humidity correction, or normalization processing. Note that preprocessing may be omitted.

[0112] Next, in step SP37, the calculation unit 46 calculates feature quantities that indicate the relationship between channels, that is, feature quantities that indicate the relationship between the multiple output signals included in the output signal set S1, based on the output signal set S1. Similarly, the relationship includes the similarity or ratio between the multiple output signals.

[0113] Next, in step SP38, the calculation unit 46 generates training data D1 that includes the features calculated in step SP37. The calculation unit 46 stores the generated training data D1 in the storage device 22.

[0114] Next, in step SP39, the control unit 44 determines whether or not the measurement has been completed for all gas generation conditions specified in the measurement setting file.

[0115] If there are unmeasured gas generation conditions (step SP39: NO), the information processing device 41 executes the following processes with respect to the unmeasured conditions.

[0116] If there are no unmeasured gas generation conditions (step SP39: YES), the information processing device 41 terminates the process of generating the learning data D1.

[0117] According to this modified example, the process of generating training data D1 is performed using feature quantities that show the relationships between multiple output signals from multiple sensor elements X. This reduces the computational cost of the process and shortens the processing time compared to the case where the training data generation process is performed using multiple output signals.

[0118] (Second Modification) In the above embodiment, the similarity or ratio of the multiple output signals was calculated as a feature quantity indicating the relationship between the multiple output signals. However, a reference signal z may be set in advance for each channel, and the similarity or ratio between the multiple output signals and the multiple reference signals z may be calculated. The reference signal z may be, for example, the output signals from sensor elements Xa to Xp when the odor sensor 1 detects the smell of a good product.

[0119] The storage device 22 of the learning device 20 and the storage device 32 of the odor identification device 30 store a set of reference signals, including reference signals za to zp set for each channel corresponding to the sensor elements Xa to Xp.

[0120] In the model generation phase, the calculation unit 25 calculates the similarity or ratio between multiple output signals a to p and multiple reference signals za to zp as features based on the output signal set S1. The calculation unit 25 generates training data D1 that includes the calculated features.

[0121] Figure 16 shows a first example of training data D1. Training data D1 includes the similarity S(a, za) to S(p, zp) between output signals a to p included in output signal set S1 and reference signals za to zp included in reference signal set. Note that the calculation unit 25 may calculate the similarity not only for all channels, but also for any multiple channels selected from all channels (for example, channels whose similarity is characteristic).

[0122] Figure 17 shows a second example of the training data D1. The training data D1 includes the ratios R(a, za) to R(p, zp) between the output signals a to p included in the output signal set S1 and the reference signals za to zp included in the reference signal set. Note that the calculation unit 25 may calculate the ratios not only for all channels, but also for any multiple channels selected from all channels (for example, channels whose similarity is characteristic).

[0123] Figure 18 is a flowchart showing a first example of the processing flow executed by the information processing device 31 of the odor identification device 30 during the odor identification phase.

[0124] In step SP41, following step SP22, the acquisition unit 33 acquires a reference signal set when the odor sensor 1 detects, for example, the odor of a good product.

[0125] Next, in step SP23, the odor sensor 1 detects, for example, the odor of the item to be inspected, and the acquisition unit 33 acquires the output signal set S2.

[0126] In step SP25, the calculation unit 35 calculates the similarity or ratio between the multiple output signals a to p included in the output signal set S2 and the multiple reference signals za to zp included in the reference signal set, based on the output signal set S2 and the reference signal set, as feature quantities.

[0127] Figure 19 is a flowchart showing a second example of the processing flow executed by the information processing device 31 of the odor identification device 30 during the odor identification phase.

[0128] By providing multiple odor sensors 1, it is possible to simultaneously and in parallel detect the odor of good products and the odor of products under inspection. In this case, the acquisition of the reference signal set in step SP41 and the acquisition of the output signal set S2 in step SP23 may be performed simultaneously and in parallel.

[0129] (Third Modification) In calculating the feature quantities, the output signal from the odor sensor 1 may be divided into multiple subsignals corresponding to multiple intervals, and the feature quantities may be calculated for each subsignal.

[0130] Figure 20 shows a first example of dividing an output signal into multiple sub-signals. The output signal from the sensor element X exhibiting detection characteristics K is divided into multiple intervals P1 to P9, which are fixed periods of time.

[0131] Figure 21 shows a second example of dividing the output signal into multiple sub-signals. The output signal from the sensor element X exhibiting detection characteristics K is divided into multiple intervals P1 to P7, which are variable periods set based on control conditions related to controlling the odor sensor 1 to detect gas. During the period before time T11, a reference gas is supplied into the chamber 53, and since the identification performance of the odor components to be detected is low, a long interval P1 is set. During the period from time T11 to time T12, a sample gas is supplied into the chamber 53, and since the identification performance of the odor components to be detected is high, intervals P4 to P6, which are shorter than interval P1, are set. During the period after time T12, a reference gas is supplied into the chamber 53, and since the identification performance of the odor components to be detected is low, an interval P7, which is longer than intervals P4 to P6, is set.

[0132] Figure 22 shows a third example of dividing the output signal into multiple sub-signals. Periods before time T11 and after time T12, during which the identification performance of the odor components to be detected is low, may be excluded from the detection period by omitting the setting of intervals P1 and P7.

[0133] Figure 23 shows a fourth example of dividing an output signal into multiple subsignals. The signal values ​​V4 to V6 of the subsignals divided into intervals P4 to P6 may be representative values ​​such as the maximum value, average value, or median value within each interval P4 to P6. The average value may be the average value within a small period including the boundary with the next interval. Alternatively, the signal values ​​V4 to V6 of the subsignals divided into intervals P4 to P6 may be the change from the previous interval.

[0134] According to this modified example, the information content of the training data D1 can be increased by dividing the output signal into multiple subsignals and calculating feature quantities for each subsignal. As a result, the accuracy of training or classification can be improved.

[0135] This disclosure is broadly applicable to information processing methods, information processing devices, information processing systems, or programs that identify odors based on output signals from odor sensors and learned identification models.

Claims

1. An information processing method for learning an identification model for identifying odors, wherein an information processing device acquires multiple output signals from multiple sensor elements, which are equipped with multiple sensor elements having different detection characteristics, when an odor sensor detects a gas containing odor components to be detected, calculates feature quantities that show the relationships between the multiple output signals, and learns the identification model using training data that includes the feature quantities.

2. An information processing method for identifying odors using a trained identification model, wherein the information processing device acquires the identification model, acquires multiple output signals from a plurality of sensor elements having different detection characteristics when an odor sensor detects a gas containing odor components, calculates a feature quantity indicating the relationship between the plurality of output signals, and identifies the odor components detected by the odor sensor based on the identification model and the feature quantity.

3. An information processing method for generating training data to be used for training a recognition model for identifying odors, wherein an information processing device generates a gas containing odor components to be detected in a gas generator; the gas generated by the gas generator is detected by an odor sensor having a plurality of sensor elements with different detection characteristics; a plurality of output signals are obtained from the plurality of sensor elements when the gas is detected; a feature quantity showing the relationship between the plurality of output signals is calculated; and training data including the feature quantity is generated.

4. The information processing method according to any one of claims 1 to 3, wherein the relationship includes the similarity or ratio of the plurality of output signals.

5. The information processing method according to claim 4, wherein, in calculating the feature quantities, the feature quantities are calculated with respect to all combinations of the plurality of output signals, or with respect to a plurality of combinations selected from all combinations of the plurality of output signals.

6. The information processing method according to any one of claims 1 to 3, wherein the relationship includes the similarity or ratio between the plurality of output signals and the plurality of reference signals.

7. The information processing method according to claim 6, wherein, in calculating the feature quantities, the feature quantities are calculated with respect to all combinations of the plurality of output signals and the plurality of reference signals, or with respect to a plurality of combinations selected from all combinations of the plurality of output signals and the plurality of reference signals.

8. The information processing method according to any one of claims 1 to 3, wherein, in calculating the feature quantities, the output signal is divided into a plurality of subsignals corresponding to a plurality of intervals, and the feature quantities are calculated for each of the subsignals.

9. The information processing method according to claim 8, wherein the interval is a fixed period of time.

10. The information processing method according to claim 8, wherein the interval is a variable period set based on control conditions relating to the control of causing the odor sensor to detect the gas.

11. The information processing method according to any one of claims 1 to 3, further comprising performing preprocessing on the acquired plurality of output signals, and in calculating the feature quantities, calculating the feature quantities with respect to the plurality of output signals on which the preprocessing has been performed.

12. An information processing device for learning an identification model for identifying odors, comprising a circuit configuration, the circuit configuration acquiring multiple output signals from a plurality of sensor elements having different detection characteristics when an odor sensor detects a gas containing odor components to be detected, calculating feature quantities that show the relationships between the plurality of output signals, and learning the identification model using training data including the feature quantities.

13. An information processing device for identifying odors using a trained identification model, comprising a circuit configuration, the circuit configuration acquiring the identification model, acquiring multiple output signals from multiple sensor elements when an odor sensor having multiple sensor elements with different detection characteristics detects a gas containing odor components, calculating feature quantities that show the relationships between the multiple output signals, and identifying the odor components detected by the odor sensor based on the identification model and the feature quantities.

14. An information processing device for generating training data to be used for training a recognition model for identifying odors, comprising a circuit configuration, the circuit configuration comprising: generating a gas containing odor components to be detected in a gas generator; having an odor sensor equipped with a plurality of sensor elements having different detection characteristics detect the gas generated by the gas generator; acquiring a plurality of output signals from the plurality of sensor elements when the gas is detected; calculating a feature quantity that shows the relationship between the plurality of output signals; and generating training data including the feature quantity.

15. A program for causing an information processing device that learns an identification model for identifying odors to execute a process, wherein the process involves: acquiring multiple output signals from multiple sensor elements when an odor sensor, which has multiple sensor elements with different detection characteristics, detects a gas containing an odor component to be detected; calculating feature quantities that show the relationships between the multiple output signals; and learning the identification model using training data that includes the feature quantities.

16. A program for causing an information processing device that identifies odors using a trained identification model to perform processing, wherein the processing includes: acquiring the identification model; acquiring multiple output signals from multiple sensor elements when an odor sensor having multiple sensor elements with different detection characteristics detects a gas containing odor components; calculating a feature quantity that shows the relationship between the multiple output signals; and identifying the odor components detected by the odor sensor based on the identification model and the feature quantity.

17. A program for causing an information processing device that generates training data to be used for training a recognition model for identifying odors to execute a process, wherein the process involves: generating a gas containing odor components to be detected in a gas generator; having an odor sensor equipped with a plurality of sensor elements having different detection characteristics detect the gas generated by the gas generator; acquiring a plurality of output signals from the plurality of sensor elements when the gas is detected; calculating a feature quantity that shows the relationship between the plurality of output signals; and generating training data including the feature quantity.

Images