WO2019117099A1

WO2019117099A1 - Fragrance quality identification system, high performance portable terminal, and program

Info

Publication number: WO2019117099A1
Application number: PCT/JP2018/045363
Authority: WO
Inventors: 広明松岡; 哲彰小出
Original assignee: 株式会社レボーン
Priority date: 2017-12-11
Filing date: 2018-12-10
Publication date: 2019-06-20
Also published as: JPWO2019117099A1; JP7074365B2

Abstract

Provided is a fragrance quality identification system that, for an object that emits a fragrance, can make it easy to identify the object emitting the fragrance by improving, beyond the conventional, the certainty of an association between an "object" and a "fragrance". A fragrance quality identification system (100) comprises: an input/output unit (70) which accepts input of a noun for an object (1) and an adjective representing the noun; a measurement unit (30) that measures a numerical value based on a fragrance component released from the object (1); a data accumulation unit (51) that accumulates, as combination data, a combination of the noun and the adjective input to the input/output unit (70) and the numerical value measured by the measurement unit (30); and a data collation unit (52) that, when a retrieval object noun for an object (1) to be retrieved and a retrieval object adjective representing the retrieval object noun are input, and a retrieval object numerical value for the object (1) to be retrieved is received, collates the retrieval object noun, the retrieval object adjective and the retrieval object numerical value against the combination data, and on the basis of the combination data, identifies the object (1) having the retrieval object noun indicating the fragrance quality of the retrieval object adjective and the retrieval object numerical value.

Description

Aroma quality specification system, high-performance mobile terminal and program

The present invention relates to an aroma quality identification system, an intelligent portable terminal, and a program for identifying an object having an aroma based on data in which noun information, adjective information, and aroma numerical information are associated with each other.

2. Description of the Related Art Conventionally, a technique as disclosed in Patent Document 1 is disclosed which records information related to a detected scent together with captured image data. Patent Document 1 includes a RAM that detects a scent component and records a detection value of the scent component as scent information, and a CPU that performs image processing on an imaging signal obtained by imaging subject light and generates an image file. However, the invention regarding the camera which links and records fragrance information and an image file is disclosed.

In particular, in the invention described in Patent Document 1, the fragrance name corresponding to the scent information, the name of the scent component included in the fragrance name corresponding to the scent information, and the image file are associated. Then, when the image file is reproduced, the fragrance is sprayed by selecting the displayed fragrance name. From this, according to the invention described in Patent Document 1, "image file" and "name of scent component" are recorded together.

JP, 2010-87940, A

However, with such a technique, it may not be possible to identify which object in the image of the image file releases the scent. For example, if there is an object that emits a scent in the image, it may be possible to visually identify the object that emits a scent, but if there is no object that emits a scent in the image, the object that emits a scent is It can not be identified visually. In this case, it is not clear at first what "an object that emits this scent". Also, in this case, even if the fragrance name is known, only the association between "image data unrelated to the object that emits the aroma" and "the name of the fragrance" is recorded, and "the object that emits the aroma" It will not be associated with "the fragrance name". As a result, as described above, an object that emits a scent is not identified.

On the other hand, when there is an object that emits a scent, there is also a need to know what this object is and what the scent of this object is. For example, when you go to a general store and pass through a good scented area, if you pick up the scented object and look at it well, "this scented object is a soap" and "the scent of this object is It may not be known that it is "Rouge royal" of the rose variety. In this case, the user should want to know the information that “what is this object” is “soap” and “what is the smell of this object” is “the rouge royal scent” .

Therefore, in the present invention, when there is an object that emits a scent, a scent quality specification system that can improve the certainty of the association between "object" and "aroma" more than before and make it easier to identify the object that emits a scent Intended to provide.

In order to achieve the above object, the scent quality specification system of the present invention comprises a noun of an object, an input unit capable of inputting an adjective representing the noun, and a measuring unit for measuring a numerical value based on a scent component emitted from the object A data storage unit configured to combine and store at least two of the noun input to the input unit, the adjective, and the numerical value measured by the measurement unit, and a search noun for an object to be searched, When a search object adjective representing the search pronoun is input and a search object numerical value which is the numerical value of the object to be searched is received, the search pronoun, the search object adjective and the search object numerical value are combined Data for matching with data and identifying the object having the search pronoun indicating the quality of the search object adjective and the smell of the search object value based on the combination data Characterized in that it comprises a verification unit.

According to the present invention, it is possible to improve the certainty of the association between the "object" and the "aroma" more than in the past, and to make it easier to identify the object that emits the aroma.

It is a schematic diagram which shows the structure of the aroma quality identification system which concerns on one Example of this invention. It is a schematic diagram showing composition of a crystal oscillator sensor. It is a graph which shows the change of the resonant frequency of a quartz oscillator sensor. It is a graph which shows the change of the resonant frequency of a quartz oscillator sensor. It is a graph which shows the change of the resonant frequency of a quartz oscillator sensor. It is a table showing the reference value of the resonance frequency for every smell of rose detected by each quartz oscillator sensor, and the average value of the reference value of the resonance frequency for each scent of rose detected by a plurality of quartz oscillator sensors. . It is a block diagram which shows the path | route which transmits the noun of the object which a flavor quality identification system processes, an adjective, and the smell of an object. It is a block diagram which shows the procedure of deep learning of the relationship between the object of a noun and an adjective. It is a block diagram which shows the procedure of deep learning of the relationship between a noun and the object of a fragrance | flavor numerical value. It is combination data of noun information, adjective information, and scent numerical information. It is a search screen displayed on the input / output unit before the user inputs a search term. It is a search screen displayed on the input / output unit after the user inputs a search term. FIG. 5 is a diagram for explaining a recursive neural network in which an input layer is a noun, an intermediate layer is a scent numerical value, and an output layer is an adjective. FIG. 5 is a diagram for explaining a recursive neural network in which an input layer is an adjective, an intermediate layer is a scent numerical value, and an output layer is a noun. FIG. 5 is a diagram for explaining a recursive neural network in which an input layer is a noun, an intermediate layer is an adjective, and an output layer is a scent value. It is a figure for demonstrating a recurrent neural network in which an input layer is an aroma numerical value, an intermediate layer is an adjective, and an output layer is a noun. It is a figure for demonstrating a recurrent neural network in which an input layer is an aroma numerical value, an intermediate layer is a noun, and an output layer is an adjective. It is a figure for demonstrating a recurrent neural network in which an input layer is an adjective, an intermediate layer is a noun, and an output layer is an aroma value.

Hereinafter, the scent quality identification system of the present invention will be described in detail with reference to the drawings. The embodiment described below is an example of the embodiment of the present invention, and the present invention is not construed as being limited to the following embodiment. Further, in the drawings referred to in the present embodiment, parts having the same configuration or function may be given the same reference numerals and descriptions thereof may be omitted.

As shown in FIG. 1A, the scent quality identification system 100 includes a scent component transfer device 10, a measurement unit 30, an artificial intelligence unit 50, and an input / output unit 70.

(Scent component transfer device)
The scent component transfer device 10 includes a housing 11 capable of containing the object 1 having a scent, a tube 12 connecting the measurement unit 30, and a pump 13 disposed in the middle of the tube 12. The scent component transfer device 10 is a device capable of transferring the gas inside the housing 11 to the measurement unit 30. For example, the aroma component transfer device 10 moves the gas to the measurement unit 30 by reducing the pressure inside the measurement unit 30.

(Measurement section)
The measurement unit 30 is a part that measures a numerical value based on the scent component emitted from the object 1 and includes a plurality of gas sensors 31, a plurality of quartz crystal vibrator sensors 32, and a scent numerical value calculation unit 33.

Each of the plurality of gas sensors 31 is a sensor that detects a gas released from the object 1 such as carbon dioxide, carbon monoxide, methane, butane, or ammonia. When animals and plants release these gases, some may emit a scent while others may not emit a scent, but a plurality of gas sensors 31 are provided in order to be able to detect a scent containing any gas component. It is done. Each gas sensor 31 detects a different gas such that one of the gas sensors 31 detects ammonia (with odor) and one of the gas sensors 31 detects carbon dioxide (no odor). These gas sensors 31 detect the gas released together with the scent molecules emitted from the object 1. For example, an electrochemical sensor is used as the gas sensor 31. An electrochemical sensor has a detection electrode, a counter electrode, and an ion conductor provided between the detection electrode and the counter electrode, and is configured to measure a short circuit current in which the detection electrode and the counter electrode are electrically connected. When it is exposed to the gas to be detected, it detects a change in current value. The data of this current value change is used for gas detection.

Each of the plurality of quartz oscillator sensors 32 is a sensor that detects a scent component emitted from the object 1. Each of the quartz oscillator sensors 32a to 32g has a quartz oscillator 34 (see FIG. 1B) formed of a thin film having nonspecific adsorption. Each of the plurality of quartz oscillator sensors 32 detects a change in resonant frequency with respect to time from when the scent component is not attached to the quartz oscillator 34 to when the scent component is attached to the quartz oscillator 34.

On each of the quartz oscillators 34 included in the quartz oscillator sensors 32a to 32g, a specific compound capable of detecting a plurality of scent components is vapor deposited. The compounds deposited on each of the quartz oscillators 34 included in the quartz oscillator sensors 32a to 32g are different from each other. Therefore, the scent components that can be detected by each of the quartz oscillator sensors 32a to 32g are also different from one another. Examples of the compound deposited on each of the quartz oscillators 34 included in the quartz oscillator sensors 32 a to 32 g include D phenylalanine, D-tyrosine, DL-sethidine, D glucose, adenine, and polyethylene.

For example, the quartz crystal sensor 32a detects the scent components "A, B, C, D, E", and the quartz crystal sensor 32b detects the scent components "A, B, F, G, H", and the quartz crystal is vibrated. An example in which the child sensor 32c detects the scent components "A, B" will be described. And the scent component of the rose variety "Rouge royal" is "A, B", the scent component of the rose variety "Rosemary" is "C, D, E" and the rose variety "Denti Bes" It is assumed that the scent component of “F, G, H”.

In this case, the aroma components "A, B" of "rouge royal" can be detected by the

quartz crystal sensors

32a, 32b, 32c. However, although the aroma components "C, D, E" of "rosemary" can be detected by the quartz oscillator sensor 32a, they can not be detected by the

quartz oscillator sensors

32b and 32c. Also, the scent components "F, G, H" of "Denti Beth" can be detected by the quartz oscillator sensor 32b, but can not be detected by the

quartz oscillator sensors

32a, 32c. As described above, the quartz crystal sensors 32a to 32c may detect the same scent component or may not detect the same scent component. In addition, even if the same scent component is detected, each of the quartz crystal sensors 32a to 32c is detected at different resonance frequencies. Further, even if the same scent component can not be detected, each of the quartz crystal sensors 32a to 32c will detect different scent components.

The odor numerical value calculation unit 33 receives current value signals based on the gas detected by the plurality of gas sensors 31, and holds numerical values based on the current value signals as current value data. The scent numerical value calculation unit 33 receives a resonant frequency signal based on the scent component detected by the plurality of crystal oscillator sensors 32, and a numerical value based on the resonant frequency signal (hereinafter sometimes referred to as "scent number") is a resonant frequency It is held as data (hereinafter sometimes referred to as "scent numerical value information"). The resonance frequency data indicates the change in resonance frequency with respect to the time from the state in which the scent molecules are not attached to the quartz oscillator 34 to the state in which the scent molecules are attached to the quartz oscillator 34.

(I / O unit)
The input / output unit 70 is a portion to which the user can input the noun of the object 1 and the adjective represented by the noun, and the user can use the noun, the adjective, the aroma numerical value, or the noun, the adjective, the aroma numerical value of the object 1 It is also a portion capable of outputting an image or the like of the object 1 having a feature.

As shown in FIG. 2A, the odor numerical value calculation unit 33 calculates resonant frequency data (data indicating a change in resonant frequency with respect to time) shown in graphs K1 to K5 based on the detection result detected by the quartz crystal sensor 32a. Possess. As shown in FIG. 2B, the scent numerical value calculation unit 33 calculates resonant frequency data (data indicating a change in resonant frequency with respect to time) shown in graphs K1 to K5 based on the detection result detected by the quartz crystal sensor 32b. Possess. As shown in FIG. 3, the scent numerical value calculation unit 33 calculates resonant frequency data (data indicating a change in resonant frequency with respect to time) shown in graphs K1 to K5 based on the detection result detected by the quartz oscillator sensor 32c. Possess.

K1 is data of the scent component of Rouge Royal, K2 is data of the scent component of Jeanne d'Arc, K3 is data of the scent component of Shedrachable, K4 is data of the scent component of Sadark lady, K5 is It is data of the scent component of wishing. Rouge Roway, Jeanne d'Arc, Shedradable, Sadark Lady, and Wishing are varieties of roses. The data of the scent component was similarly measured for the quartz oscillator sensors 32c to 32g (see FIG. 4).

FIG. 4 is a table showing reference values of graphs K1 to K5 of FIGS. 2A and 2B and graphs K1 to K3 of FIG. Here, a numerical value that stably indicates the resonance frequency after 100 seconds is referred to as a reference value. For example, the reference value of the resonant frequency of the rouge royal detected by the quartz oscillator sensor 32a is 1990760 Hz. The reference value of the resonant frequency of the rouge royal detected by the quartz oscillator sensor 32b is 1977230 Hz. The reference value of the resonant frequency of the rouge royal detected by the quartz oscillator sensor 32c is 1994785 Hz. The odor numerical value calculation unit 33 holds the detection results detected by the quartz oscillator sensors 32a to 32g as an aroma numerical value represented by a vector. The table in FIG. 4 looks similarly for Jeanne d'Arc, Shedrachable, Sadark Lady, and Wishing.

(Artificial Intelligence Department)
Here, it returns to the explanation of FIG. 1A. As shown in FIG. 1A, the artificial intelligence unit 50 includes a data storage unit 51 and a data collating unit 52. The data storage unit 51 includes information on nouns (noun labels) of an object input from the input / output unit 70 (input unit / output unit), information on adjectives (adjective expressions) representing the nouns, and scents measured by the measurement unit 30. Receive numerical information. Then, the data storage unit 51 combines the information of these nouns, the information of adjectives, and the information of scent value, and stores them as combination data (see FIG. 6). The data storage unit 51 learns nouns, adjectives and scent values with a recursive neural network among nouns, adjectives and scent values, and creates and accumulates combination data. In many cases, the data storage unit 51 combines all the nouns and adjectives input to the input / output unit 70 and the scent value measured by the measurement unit 30 and stores them as combination data, but two of them are combined and combined In some cases, data may be accumulated (see ID3 and ID4 in FIG. 6). This is because not all nouns, adjectives and scent values are always consistent.

Noun information includes, for example, cooking name (curry), varieties such as flowers (rose, rose), producer's name (有佳子), lot number (product number), confectionery name (Pringles sour cream & onion), etc. Is the case. Information of adjectives includes information of olfactory senses, tastes, and expressions obtained by sight on noun objects.

(Data storage unit) (On deep learning between noun information and adjective information)
When the noun of the object 1 and the adjective are simultaneously input and the noun information and the adjective information are received, the data storage unit 51 creates combination data combining the noun information and the adjective information. The data storage unit 51 uses the noun information in the input layer, uses the adjective information in the output layer, and performs odor learning using a recursive neural network using the aroma numerical information in the intermediate layer when the aroma numerical information is received. The data accumulation unit 51 accumulates data obtained by deep learning by the recursive neural network as combination data as needed.

This content will be described with reference to FIGS. 6, 8A and 8B. For example, the data storage unit 51 receives the noun information "rouge royale" and the adjective information "fruity" described in ID1 of FIG. In this case, the data storage unit 51 creates combination data from the noun information "rouge royal" and the adjective information "fruity". Then, the data storage unit 51 uses the noun "Rouge Royal" in the input layer, uses the adjective "Fruity" in the output layer, and uses the aroma numerical information in the middle layer when the aroma numerical information is received, and the noun information Conduct deep learning from "Rouge royal" to adjective information "Fruity". In this way, the probability of the adjective expressing the noun is improved between the noun and the adjective by performing deep learning by the flow of the noun → the numerical value of the aroma → the adjective.

In addition, the data storage unit 51 uses the noun "rose, fresh flower, love" in the input layer, uses the adjective "fruity, good smell, sweet, thick, good smell" in the output layer, and smell value information in the middle layer. Deep learning is performed using noun information to adjective information. In this way, deep learning is performed by the flow of noun → aroma numerical value → adjective. Although the above-mentioned example is an explanation about ID1 of middle class being scent numerical value information, it carries out similarly about ID2, ID3, and ID4.

As described above, the data storage unit 51 performs deep learning by means of a recursive neural network having a noun as an input unit, an adjective as an output unit, and an aroma value as an intermediate layer (FIG. 8A). Deep learning is performed by a recursive neural network having a part, a noun as an output part, and an aroma value as an intermediate layer (FIG. 8B), and the relationship between the noun and the adjective is learned.

(Data storage unit) (About deep learning between noun information and scent numerical information)
The data storage unit 51 receives the noun of the object 1 and receives the noun information, and when receiving the aroma numerical information measured by the measuring unit 30, creates the combination data combining the noun information and the aroma numerical information Do. The data storage unit 51 performs deep learning by means of a recursive neural network using noun information in the input layer, using odor numerical information in the output layer, and using adjective information in the middle layer when adjective information is received. The data accumulation unit 51 accumulates data obtained by deep learning by the recurrent neural network as needed as combination data.

This content will be described with reference to FIGS. 6, 8C, and 8D. For example, the data storage unit 51 receives the input information of the noun information "rouge royale" described in the ID1 of FIG. 6, and receives the aroma numerical value information "1990760". In this case, the data storage unit 51 creates combination data from the noun information "rouge royal" and the aroma numerical value information "1990760". Then, the data storage unit 51 uses noun "Rouge royal" in the input layer, uses aroma numerical value information "1990760" in the output layer, and adjective information in the middle layer when adjective information is received, noun information We perform deep learning from "Rouge royal" to aroma numerical value information "1990760". As described above, by performing deep learning with the flow of noun → adjective → scent numerical value, the probability that the scent numerical value expresses the noun is improved between the noun and the scent numerical value.

Also, the data storage unit 51 uses the noun "rose, fresh flower, love" in the input layer, and the scent numerical value information "1977230, 1977230, 1994785, ... 1996620, 1988760" in the output layer (here, id1 in FIG. 6). In the middle layer, adjective information is used, and deep learning from noun information to adjective information is performed using the following numerical value information of scent). As odor numerical value information provided in the output layer, id2 and id3 in FIG. 6 may be used. In the above example, the output layer is the ID1 of the scent numerical value information, but the same applies to ID2, ID3 and ID4.

As described above, the data storage unit 51 performs deep learning by means of a recursive neural network having a noun as an input unit, an aroma value as an output unit, and an adjective as an intermediate layer (FIG. 8C). Deep learning is performed by a recursive neural network having an input unit, a noun as an output unit, and an adjective as an intermediate layer (FIG. 8D), and the relationship between the noun and the scent value is learned.

(Data storage unit) (About deep learning between scent numerical information and adjective information)
The data storage unit 51 receives combined odor numerical information measured by the measurement unit, and when the adjective information of the object 1 is input and receives the adjective information, creates combined data combining the aroma numerical value information and the adjective information . The data storage unit 51 performs deep learning using a recurrence neural network by using aroma numerical value information in the input layer, using adjective information in the output layer, and using noun information in the intermediate layer when noun information is received. The data accumulation unit 51 accumulates data obtained by deep learning by the recursive neural network as combination data as needed.

This content will be described with reference to FIGS. 6, 8E, and 8F. For example, the data storage unit 51 receives the aroma numerical value information “1990760” (see FIG. 6) described in the ID1 of FIG. 6, and receives the adjective information “fruity” (see FIG. 6). In this case, the data storage unit 51 creates combination data from the scent numerical value information "1990760" and the adjective information "Fruity". Then, the data storage unit 51 uses the scent numerical value information "1990760" in the input layer, uses the adjective information "Fruity" in the output layer, and uses noun information in the middle layer when the noun information is received. We perform deep learning from information "1990760" to adjective information "fruity". In this way, by performing deep learning with the flow of the scent value → noun → adjective, the probability that the adjective expresses the scent value between the scent value and the adjective is improved.

Further, the data storage unit 51 uses the scent numerical value information “1977230, 1977230, 1994785,... 1996620, 1988760” (here, the scent numerical value information of id1 in FIG. 6 is illustrated) for the input layer, and Using adjective information “fruity, good smell, sweet, thick, good smell” and using noun information in the middle layer, we perform deep learning from scent numerical information to adjective information. In this way, deep learning is performed by the flow of the numerical value information of the scent → noun information → adjective information. As odor numerical value information provided in the input layer, id2 and id3 in FIG. 6 may be used. In the above-mentioned example, the input layer is the description about ID1 of the scent numerical value information, but the same goes for ID2, ID3 and ID4.

As described above, the data storage unit 51 performs deep learning by means of a recursive neural network having an aroma value as an input unit, an adjective as an output unit, and a noun as an intermediate layer (FIG. 8E) and, conversely, inputs an adjective Deep learning is performed by a recursive neural network in which the part, the scent value is an output part, and the noun is an intermediate layer (FIG. 8F), and the relationship between the scent value and the adjective is learned.

The data storage unit 51 collects the information between the nouns, the adjectives and the scent value as described above from the Internet information, creates combination data, and stores the combination data as needed. The data storage unit 51 collects combination information of noun information and adjective information from combinations searched by Twitter (registered trademark), texts tweeted by Twitter, etc. There is. The data storage unit 51 is based on the information from the noun, the adjective and the scent value as described above, and from the system starting from the response history of the scent quality identification system 100 including the auto answer system (chatbot) in others. collect.

Next, the operation of the scent quality identification system 100 will be described using FIGS. 5A to 5C. As shown in FIG. 5A, the user inserts the object 1 whose aroma information is to be measured into the aroma component transfer device 10, and the aroma of the object 1 fills the inside of the housing 11 (M1). The user drives the pump 13 (M2), and sends the air containing the aroma component in the aroma component transfer device 10 to the measurement unit 30.

In the measuring unit 30, the gas sensor 31 detects carbon monoxide, carbon dioxide, ammonia and the like in the air (M31 of M3). At the same time, the quartz oscillator sensor 32 detects the resonance frequency of the scent component in the air (M32 of M3). Here, it is assumed that the quartz oscillator sensor 32 detects a rouge royal as a scent component.

On the other hand, nouns and adjectives (M4) are transmitted to the data storage unit 51 of the artificial intelligence unit 50 through the Internet (M5), and a feature learning model (LSTM which is a recurrent neural network) is executed (M6) ). Alternatively, the user transmits the noun of the object 1 and the adjective (M4) to the data storage unit 51 of the artificial intelligence unit 50 using the input / output unit 70 of the high-performance mobile terminal (smartphone etc.), Execute the recursive neural network (LSTM) (M6). The relationship between the noun of the learned object 1 and the adjective is sent to the data collating unit 52 of the artificial intelligence unit 50 (M7).

The artificial intelligence unit 50 causes the data storage unit 51 to store combination data in which noun information, adjective information and odor numerical value information are combined.

As shown in FIG. 5B, the data storage unit 51 of the artificial intelligence unit 50 of FIG. 5A uses the noun (M11) for the language input layer (M12) and the adjective represented by the adjective for the language output layer (M16). The relationship between a noun and an adjective is deep-learned by a feature learning model (LSTM which is a recursive neural network) (M17) using information (M15). In addition, the data storage unit 51 uses an adjective (M13) in the language input layer (M14) and information (M15) of an object represented by a noun in the language output layer (M16) to provide a feature learning model (recursive neural network Deeply learn the relationship between nouns and adjectives by LSTM (M17) which is a network.

As shown in FIG. 5C, the data storage unit 51 of the artificial intelligence unit 50 of FIG. 5A uses nouns (M21) for the language input layer (M22) and is represented by odor numerical values for the scent numerical value output layer (M26). Using the object information (M25), deep learning is performed on the relationship between the noun and the scent value by a feature learning model (LSTM which is a recursive neural network) (M27). In addition, the data storage unit 51 uses the scent numerical value (M23) for the scent numerical value input layer (M24), uses the information (M25) of the object represented by the noun for the language output layer (M26), Deep learning of the relationship between scent value and noun by LSTM (M 27), which is a neural network.

As shown in FIG. 6, the combination data is information in which a plurality of noun information, a plurality of adjective information, and a plurality of scent numerical value information are combined. The noun information of ID 1 includes "rouge royal", "rose", "flower" and "love". The data storage unit 51 classifies “Rouge royal”, “rose”, “floral flower”, and “love” as being strongly associated with each other according to the frequency of use. This relationship arises because "Rouge royal" is a variety of "rose", is used for "fresh flowers", and has the flower language of "love". Further, the input noun information is related to each other according to the frequency of the input for each input content of the noun information or the adjective information or the aroma numerical information input to the data storage unit 51, the adjectives are related to each other, and the aroma numerical value information Relationships arise between one another. When performing deep learning, "Rouge royale, roses" that may be frequently presented as combination data are deeply learned to have a strong relationship, and for combination data with a low frequency of input such as "rouge royale, rice" Deep learning is performed so as to have only a weak relationship.

Next, the adjective information of ID 1 includes "fruity", "good smell", "sweet", "rich" and "good smell". The data storage unit 51 determines that “fruity”, “good smell”, “sweet”, “rich”, “good smell” is used as an adjective that expresses “rouge royal”. As adjective information, "fruity" and "good smell" are expressed by the first expressor, "sweet" and "rich" are expressed by the second expressor, and "good smell" by the third expressor It is also recorded that it was expressed.

Next, in the aroma adjective information of ID1, ID, time stamp (aroma measurement time), resonance frequency detected by the quartz crystal sensors 32a to 32g, average (average value of resonance frequency), ondition_id (air temperature, humidity, etc.) An id indicating an environmental classification is included. In FIG. 6, s1 corresponds to the quartz crystal sensor 32a, s2 corresponds to the quartz crystal sensor 32b, s3 corresponds to the quartz crystal sensor 32c, and s4 corresponds to the quartz crystal sensor 32d. S5 corresponds to the quartz oscillator sensor 32e, s6 corresponds to the quartz oscillator sensor 32f, and s7 corresponds to the quartz oscillator sensor 32g.

(Data collation unit)
The data collating unit 52 receives input information of the noun of the object 1 for which the user has searched for the object 1 (searching noun) and the adjective representing this noun (searching adjective), and the user measures it to search the object 1 The scent value (the search item value) of the object 1 measured by the unit 30 is received. For example, the data collating unit 52 receives input information of the noun “rose” and the adjective “fruity” searched by the user, and the measuring unit 30 receives the scent value “1990760”.

Then, the data collating unit 52 collates the noun, the adjective and the aroma value of the retrieved object with the stored combination data. The data collating unit 52 specifies the object 1 having the noun of the search object indicating the adjective of the search object and the scent value of the search object based on the combination data. For example, in the above example, it is as follows. The data collating unit 52 collates the noun "rose" of the search object with the noun information "rose" of ID 1 and the noun information "rose" of ID 2 in FIG. Further, the data collating unit 52 collates the adjective "fruity" of the search object with the adjective "fruity" of ID 1 and the adjective "fruity" of ID 2 in FIG. Furthermore, the data collating unit 52 collates the scent value “1990760” of the search item with the average value “1990760” of the scent values of FIG.

Although the data collating unit 52 can determine that it is “rose” from the noun and the adjective of the search object, it can not determine whether it is “rouge royal” or “Janne d'arc”. However, the data collating unit 52 determines that it is "rouge royal" from the average value of the scent value. The data collating unit 52 outputs the rouge royal roses to the input / output unit 70.

Example 1
Next, the user went to the flower shop to find "something that has a good smell", but it is assumed that a new soap that has the same smell as this smell goes to a general store and searches for a scent quality specification system The usage mode of the user regarding 100 and the operation of the scent quality specification system 100 will be described. Also, in reality, if it is a general store, the soap component is often described in the packaging paper of the soap, so "a scent equivalent to a good smell" is specified in the packaging paper, Here, it is assumed that soaps in which the content of “a scent equivalent to a good scent” is not described in the packaging sheet are lined up in the store.

As the user walks through the florist, find something that smells like yours. The user activates the "scent application" on the screen of the input / output unit 70 of the smartphone 80. The user stores the scented component in the casing 11 of the scent component transfer device 10. The user drives the pump 13 to flow the air of the housing 11 into the measurement unit 30. The measuring unit 30 detects the scent of soap by the gas sensor 31 and the quartz crystal sensor 32, measures the scent value of the soap by the scent value calculating unit 33, and transmits the scent value information "1990760" to the data storage unit 51. The data storage unit 51 creates combination data of the adjective information “fruity, good smell, sweet, thick, good smell” and the numerical value information of aroma “1990760”.

Then, the user goes to the grocery store and looks at the soap. As shown in FIG. 7A, the user has the sentences "What are you looking for?" And "What is the smell like that? Are there other features?" Is displayed. As shown in FIG. 7B, the user inputs the noun "soap" in the "What are you looking for?" Input field, and "What is the smell like that? Is there any other feature?" Enter the adjective "fruity sweet" in the entry field of. The input / output unit 70 transmits the noun information “soap” and the adjective information “fruity sweet” to the data collating unit 52.

The data collating section 52 is a part of combination data of adjective information “fruity, good smell, sweet, rich, good smell” collated by florist and numerical value information of smell “1990760”, and noun information “soap” searched by general store And the adjective information "Fruity Sweet" and the combination data part is further combined. Then, the data collating unit 52 derives the noun information “soap”, the adjective information “fruity, sweet”, and the aroma numerical value information “1990760”. The data collating unit 52 specifies from the noun information that the noun is "soap" and from the adjective information and the numerical value information that the smell is "rouge royal". The artificial intelligence unit 50 displays the soap of “Rouge Royale's Soap” on the screen, and also displays the soap with the word “Rouge Royale” written on the packaging sheet. The user recognizes that the soap is "rouge royal soap".

(Example 2)
Next, the user goes to a general store to find “something that smells good” and finds out what it is, but explains the usage pattern of the user and the operation of the scent quality specification system 100 regarding the scent quality identification system 100 To go. In this case, an object is identified because something with a good smell is often described on the packaging paper in practice, but here, it is not packaged in the packaging paper I will explain assuming the case where is lined up in the store.

While the user is walking in the grocery store, find something that smells like yours. The user stores in the casing 11 of the scent component transfer device 10 what smells his / her favorite. The user drives the pump 13 to flow the air of the housing 11 into the measurement unit 30. The measuring unit 30 detects what smell is detected by the gas sensor 31 and the quartz crystal sensor 32, measures the odor numerical value of the smell with the aroma numerical value calculating unit 33, and detects the odor numerical value information “1990760 Send ".

On the other hand, the user activates the "scent application" on the screen of the input / output unit 70 of the smartphone. As shown in FIG. 7A, on the screen, the sentences "What are you looking for?" And "What is the smell like, what other features do you have?" Is displayed. As shown in FIG. 7B, the user inputs the noun "unknown" in the "What are you looking for?" Input field, and "What is the smell like that? Is there any other feature?" Enter "Fruity" in the entry field of. The input / output unit 70 transmits the noun information “unknown” and the adjective information “fruity” to the data collating unit 52.

The data collating unit 52 collates the noun information “unknown”, the adjective information “fruity”, and the scent numerical information “1990760” with the combination data of the data storage unit 51. Then, the artificial intelligence unit 50 determines that the noun is not specified from the noun information, the smell is not specified from the adjective information, and the smell is “rouge royal” from the smell numerical information. Although the noun information is not known at the search stage, if the data collating unit 52 has the combination data of the noun information “soap”, the adjective information “fruity”, and the aroma numerical information “1990760”, It may be concluded that it is a "roial soap". However, it may also conclude that it is a "Rouge royal rose".

(Example 3)
In the third embodiment, in the case of the second embodiment, the data comparison unit 52 assumes that there is data in which adjective information and odor numerical value information are combined in combination data but there is no combination data in which these are combined with noun information To explain. In this case, in order to identify the noun information, "is it a square?", "Is it hard?", "Is it slippery?", "It is white," on the screen The sentence is displayed. In the data storage unit 51, the user's answer "Square" to "Is it a square?" The user's answer to "Is it hard?" The answer "Hard" or "Is it slippery?"", Is it white?" The user's answer to the adjective information such as "white" is the middle layer, the scent numerical information "1990760" is the input layer, the noun information "unknown" is the output layer, and the noun by the recursive neural network Identify "soap". The user recognizes that what he wanted to know is "Rouge Royale's Soap".

According to the scent quality identification system 100 described above, it is possible to improve the certainty of the association between the "object" and the "aroma" more than in the past and to easily identify the object 1 that emits the scent. For example, when there is a rouge royale soap, the certainty of the association between the noun "soap" and the aroma number "1990760" is improved by the adjective "fruity, good smell, sweet, thick, good smell" etc. An object 1 to be emitted can be easily identified.

(Modification)
In the above embodiment, the crystal oscillator sensor is used to detect the scent component. However, the present invention is not limited to this, a configuration in which even scent molecules are finely detected, or a semiconductor sensor, contact combustion type A sensor, an electrochemical sensor, or an optical sensor may be used to detect a gas emitted from an odor component or an odor molecule or an object having an odor.

When the surface of the oxide semiconductor is exposed to a flammable gas, the semiconductor sensor detects a change in the resistance value of the oxide semiconductor because the electric resistance of the oxide semiconductor changes. The data of this resistance value change is used for detection of the scent molecule. The contact combustion type sensor is a bridge by raising the resistance of only the detection piece among the detection piece that reacts to the combustible gas and the compensation piece that does not react to the combustible gas when exposed to the flammable gas. The balance of the circuit is broken and a change in the unbalanced voltage value is detected. The data of this voltage value change is used for detection of the scent molecule.

In the above embodiment, the data storage unit 51 is a unique system developed by the applicant named IINIOI AI but may be a system based on other deep learning or database.

In the above embodiment, seven crystal oscillator sensors 32 are used, so the scent data has a seven-dimensional configuration, but the present invention is not limited to the above embodiment. That is, a configuration of eight or more dimensions in which eight or more quartz oscillator sensors 32 are used may be used. As the number of types of the quartz oscillator sensor 32 increases, the data storage unit 51 can hold the scent numerical value information more finely. In addition, the data collating unit 52 can collate the odor numerical value more finely as the number of types of the quartz oscillator sensor 32 increases.

In the above embodiment, the data storage unit 51 is configured to use noun information in the input layer, adjective information in the output layer, and scent numerical value information in the intermediate layer, but the configuration is not limited to this configuration. In this case, in the middle layer training, a learned network between scent numerical information and noun information is used for transfer learning, or a learned network between scent numerical information and adjective information is used for transfer. You may learn it. This is the same even if the input layer uses adjective information, the output layer uses noun information, and the middle layer uses scent numerical information.

In the above embodiment, the data storage unit 51 is configured to use noun information in the input layer, scent numerical value information in the output layer, and adjective information in the middle layer, but the configuration is not limited to this. In this case, in the training of the middle layer, a learned network between adjective information and noun information is used for transfer learning, or a learned network between adjective information and odor numerical information is used for transfer learning You may do it. The same applies to a configuration in which the numerical value information of the scent is used for the input layer, the noun information for the output layer, and the adjective information for the intermediate layer.

In the above embodiment, the data storage unit 51 is configured to use the scent numerical value information for the input layer, the adjective information for the output layer, and the noun information for the intermediate layer, but the configuration is not limited to this configuration. In this case, in the middle layer training, a learned network between noun information and scent numerical information is used for transfer learning, or a learned network between noun information and adjective information is used for transfer learning You may do it. The same applies to a configuration using adjective information in the input layer, scent numerical information in the output layer, and noun information in the middle layer.

In the said embodiment, although the fragrance quality identification system 100 was a structure which has the fragrance component transfer apparatus 10, it may not be limited to this structure. The scent quality specifying system 100 may be provided with a scent suction mechanism for sucking a scent.

In the above embodiment, the scent quality specification system 100 has the configuration in which the scent component transfer device 10, the scent measurement unit 30, the artificial intelligence unit 50, and the input / output unit 70 are separately configured, but the present invention is limited to this configuration. It does not have to be. All configurations may be built in high-performance communication terminals such as smartphones.

Although the data collating unit 52 collates the scent numerical value information of the object 1 to be searched with the average value of the scent numerical value information of the data storage unit 51 in the above embodiment, the present invention is not limited to this configuration. For example, instead of the average value of the odor numerical value information of the data storage unit 51, a value obtained by adding a weighted numerical value to the aroma value detected by each of the quartz crystal sensors 32a to 32g, each quartz oscillator sensor 32a to A value obtained by multiplying the scent value detected by 32 g by a weighted value may be used.

In the above embodiment, deep layer learning processing in which the scent quality identification system 100 learns the relationship between nouns, adjectives and scent values of the object 1, and object identification processing in which the object 1 is identified using combination data created by deep layer learning Although two processes were performed, the scent quality specification system 100 may be configured as a server / client system in which the server apparatus and the client perform the two processes. Specifically, the server device performs deep learning to create combination data, distributes the "scent application" (program) and combination data to the user's high-performance portable terminal via a public communication network such as the Internet, and installs the combination. Let The high-performance portable terminal receives an input of nouns and adjectives of the object 1 to be searched from the user, receives an aroma value of the object 1 to be searched from the measuring unit 30, and collates it with combination data to specify the object 1 . As described above, the hardware for performing deep learning (the data storage unit 51) and the hardware for identifying the object 1 (the data collating unit 52) may be different.

It should be noted that the server device holds the combination data without installing the combination data in the high-performance mobile terminal, and the high-performance mobile terminal accesses the server device when the object 1 is specified to perform collation with the combination data. Of course it is also good.

DESCRIPTION OF SYMBOLS 1 object 10 aroma component movement apparatus 100 aroma quality identification system 11 housing | casing 12 tube 13 pump 30 measurement part 31 gas sensor 32 quartz oscillator sensor 32a-32g quartz oscillator sensor 33 numerical calculation part 34 quartz oscillator 50 artificial intelligence part 51 data Storage unit 52 Data collation unit 70 Input / output unit K1 to K5 graph

Claims

An input part to which an object noun and an adjective representing the noun can be input;
A measurement unit that measures a numerical value based on the scent component emitted from the object;
A data storage unit for combining and storing at least two of the noun input to the input unit, the adjective, and the numerical value measured by the measurement unit; When a search object adjective representing a search pronoun is input and a search object numerical value which is the numerical value of the object to be searched is received, the search pronoun, the search object adjective and the search object numerical value are combined data An object having the search pronoun indicating the quality of the search object adjective and the smell of the search object value based on the combination data, or the search pronoun indicating the quality of the smell of the search object value A data collating unit for specifying an object to be possessed;
Aroma quality identification system characterized in that it comprises.
The data storage unit is characterized in that the noun, the adjective and the numerical value are learned by a recursive neural network among the noun, the adjective and the numerical value, and the combination data is created and accumulated. The scent quality identification system according to Item 1.
The data storage unit collects the noun, the adjective and the numerical value from Internet information or a response history of an automatic response system, and creates and stores the combination data. Aroma quality identification system described in 2.
The high functional portable terminal provided with the fragrance quality identification system of any one of Claim 1 thru | or 3.
Accept input of a search pronoun of an object to be searched and a search adjective representing the search pronoun,
Receiving a search item numerical value that is a numerical value based on a scent component emitted from the object to be searched;
The search pronoun, the search object adjective, and the search object numerical value are collated with combination data combining at least two of the object noun, the adjective representing the noun, and the numerical value, and based on the combination data Identifying the object having the search pronoun indicating the quality of the search object adjective and the smell of the search object value, or the object having the search pronoun indicating the quality of the search object value The program to run.