Method for basic odor selection, method and device for representing, suggesting or synthesizing an odor by combining basic odors

Abstract

The present application can decompose any odor into a relatively small combination of odors as well as synthesize, as with the decomposition of color into three primary colors. Since the "basic odor" corresponding to the primary color is not yet clear, in the present application, a method and apparatus are provided to enable the above decomposition, synthesis, etc. by selecting a subset of odors from a set of multiple odors that can approximate other odors as closely as possible by mixing, instead of predefining fixed basic odors.

Description

Technical Field

[0001] This invention relates to selecting a subset of relatively few odors from a set of odors consisting of multiple odors, wherein the subset is a subset of the set of odors that, when combined, at least approximately represents any odor in the set of odors (hereinafter, each odor in this subset is referred to as a basic odor). Furthermore, this invention relates to representing, suggesting, or synthesizing multiple odors at least approximately through combinations of such selected basic odors. Further, this invention relates to representing or suggesting multiple odors in a form that can be perceived by senses other than smell, such as color. Further, this invention relates to an apparatus for this purpose. Background Technology

[0002] As is well known, any color can be decomposed into three primary colors and represented by the weights of each of these primary colors. Furthermore, by mixing the light of each primary color with its corresponding weight, in principle, only three colors are needed to synthesize any color in a way that is visually reproducible to humans, replacing the need for light sources and colorants for all colors. As an analogy to this primary color decomposition and synthesis, if any odor could be decomposed into a relatively small number of odors—equivalent to the "basic odors" of the primary colors—and then synthesized by weighted mixing of these basic odors, a wide range of applications could be expected.

[0003] Additionally, it's important to note regarding basic odors here that breaking down an odor into basic odors or synthesizing an odor by mixing some basic odors does not imply a decomposition or synthesis of odors in the sense of their composition. It's crucial to understand that a basic odor refers to any odor that, when mixed, can be synthesized into one that is identical or very similar to any odor perceived by the sense of smell (human or other animal olfaction, and more generally, the response of various odor sensors). For example, even if an odor X containing acetic acid can be synthesized by mixing basic odors A and B, neither of the basic odors A nor B necessarily contains acetic acid; it simply means that the gas obtained by mixing these basic odors in a certain proportion is perceived as having the same or very similar odor to odor X.

[0004] Furthermore, by performing basic odor decomposition, odors can be transformed into forms that can be identified through senses other than smell, making them easier to represent. For example, consider the application of representing odors with color. If odors could be given color, the effect of "scent" could be used in comics and movies. Additionally, if perfumes, cosmetics, whiskey, wine, etc., could be represented by color, their scent could be visually understood through the use of that color on product packaging, potentially contributing to sales. Since color has a psychological effect, the applications of representing odors with color are limitless. Furthermore, since there is currently no established method for representing odors themselves, representing odors with color could allow for a more objective representation than before, potentially making odor memorization, learning, and understanding easier.

[0005] However, as mentioned above, representing odors based on the decomposition of basic odors, synthesizing and generating odors based on mixing, and converting odors into forms that other senses can recognize are generally difficult. This is because the basic odors in smell, corresponding to the three primary colors of light and the four primary flavors (there are also ideas of five, six, or more primary flavors), have not been scientifically determined. Attempts to define basic odors have a long history, with well-known examples such as Henning's olfactory triangle, which identifies six basic odors: floral, fruity, putrid, spicy, resinous, and burnt. However, even if samples considered to be basic odors are collected, it is impossible to represent all odors with mixtures of these basic odors. Unlike light or flavor, which can be reproduced simply by preparing a reference sample, it is not possible to represent any other light or flavor. For example, in the case of wanting to represent odors with a single color, it would be sufficient to identify the three basic odors corresponding to the three primary colors of light, but for the reasons mentioned above, this is practically impossible to determine.

[0006] On the other hand, in recent years, the development of artificial olfactory systems using sensor arrays has been remarkable. By utilizing these systems, it has become increasingly possible to represent odors numerically. Furthermore, the visualization of odors using sensor arrays has been conducted through various methods, typically representing odors as a set of multiple colors. Among these, the use of optoelectronic noses is the most mainstream. In this method, various dyes or pigments, etc., whose colors change upon contact with odor components, are prepared, and the color changes when the target odor is sprayed onto them are detected. However, these strategies do not perform basic odor searches, nor can they represent odors with a single color.

[0007] For example, existing technologies disclose "absolute value performance analysis," which quantifies comparison results with a variety of pre-defined reference gases and displays them on a radar chart. Examples of such quantification and radar chart representation are shown in existing technologies. However, in the absolute value performance described here, the provider or user of the measurement system needs to pre-set odors (reference gases) as references, but it does not investigate whether the odor of the object being measured can be well approximated by combinations of these reference gases. Therefore, the absolute value performance analysis disclosed here does not provide any guidance on representing odors based on basic odor decomposition, synthesizing and generating odors based on mixing, or converting odors into forms that can be recognized by other senses. Summary of the Invention

[0008] The problem the invention aims to solve

[0009] The problem solved by this invention is to systematically select basic odors from a set of odors. Furthermore, it involves using the selected basic odors to decompose, synthesize, and generate arbitrary odors. Additionally, it uses the selected basic odors to represent or indicate odors with a single color. Furthermore, it involves creating a device capable of converting odors into colors in real time.

[0010] means for solving problems

[0011] According to one aspect of the present invention, a method for selecting basic odors is provided, which detects each odor in a set of odors and selects a subset consisting of basic odors of the odors in the set of odors based on the detection results of each odor.

[0012] Here, the detection of each odor includes: finding a feature vector corresponding to each odor, which is composed of multiple features obtained from the output signals of the sensor for each odor. The selection of the subset composed of the basic odor may include: selecting the basic odor from the set of feature vectors corresponding to each odor in the odor set based on the subset of feature vectors selected by endpoint detection in the feature space containing the feature vectors.

[0013] Alternatively, the rows of the feature matrix, which is constructed by using the feature vectors corresponding to each of the odors as the vertical vectors, can be scaled.

[0014] Alternatively, the number of times a point is located at an end in any direction of the feature space (i.e., an endpoint) can be counted, and feature vectors can be extracted from the set of feature vectors in descending order of the number of occurrences, thereby obtaining a subset of the feature vectors.

[0015] Alternatively, the sensor can be a surface stress sensor.

[0016] The sensor is a sensor array having multiple sensor elements, and the output signal may include a separate output signal from each of the multiple sensor elements.

[0017] According to another aspect of the present invention, a basic odor selection apparatus is provided, comprising a sensor, a gas supply mechanism for alternately supplying a sample gas and a reference gas having the respective odors to the sensor, and an information processing device for inputting an output signal from the sensor, wherein the odor of the sample gas corresponding to each element of the subset of the feature vector obtained by the information processing device through the any basic odor selection method is selected as the basic odor.

[0018] According to another aspect of the present invention, a method for odor synthesis is provided, wherein a given odor is synthesized in the form of an odor formed by mixing the selected basic odors based on a plurality of basic odors selected by the aforementioned basic odor selection method.

[0019] According to another aspect of the present invention, an odor synthesis apparatus is provided, which includes a mechanism for mixing the sample gas with each of the odors in the basic odors selected by the basic odor selection device.

[0020] According to another aspect of the present invention, a method for representing an odor is provided, which represents a given odor as a combination of the selected basic odors based on a plurality of basic odors selected by the aforementioned basic odor selection method.

[0021] Here, the combination of the basic odors can also be represented by a linear combination of the feature vectors of the basic odors.

[0022] According to another aspect of the present invention, an odor representation apparatus is provided, comprising a sensor, a gas supply mechanism for alternately supplying a sample gas and a reference gas having the respective odors to the sensor, and an information processing device for inputting an output signal from the sensor, and a mechanism for storing or providing to the outside the odor representation result obtained by the information processing device through any of the odor representation methods.

[0023] According to another aspect of the invention, an odor prompting method is provided, which can prompt the combination of said basic odors through sensory recognition other than olfaction.

[0024] Here, the cues, which can be identified by senses other than smell, can establish a correspondence between different colors and each of the basic odors, and the odors can be indicated by the result of mixing the different colors.

[0025] In addition, the different colors that correspond to the various odors in the basic odor can be the three primary colors.

[0026] According to another aspect of the present invention, an odor prompting device is provided, comprising a sensor, a gas supply mechanism that alternately supplies the sensor with a sample gas and a reference gas having the respective odors, an information processing device that inputs an output signal from the sensor, and an output device that outputs an odor prompting result obtained by the information processing device through any of the odor prompting methods.

[0027] The effects of the invention

[0028] According to the present invention, a sample gas, i.e., an odor (a gas produced when the provided sample is a liquid or a solid), which may contain a variety of chemical substances, is analyzed, and a sample is selected as the basic odor. By utilizing the basic odor thus selected, any odor can be represented by a combination of basic odors. This representation can be a representation that can be recognized by human senses, such as color (e.g., using a form that can be recognized by senses other than smell). In addition, any odor can be synthesized at least approximately. Attached Figure Description

[0029] Figure 1 This is a conceptual diagram of a measuring device to which the present invention can be applied.

[0030] Figure 2 This is a graph showing an example of a response signal measured by a membrane surface pressure sensor (hereinafter referred to as MSS). The horizontal axis represents time (in seconds), and the vertical axis represents the output (in mV). A, B, C, and D are used when representing the response signal using characteristic quantities.

[0031] Figure 3 This is an example of a repetitive step in endpoint detection. In this example, the endpoint score is increased by one point for each of the two points enclosed by the circular marker.

[0032] Figure 4A This is a graph showing the response signals obtained from channels 1-6 of the 12 channels of the MSS array. The horizontal axis represents time (in seconds), and the vertical axis represents output (in mV). Results are shown using 12 different seasonings as samples.

[0033] Figure 4B This is a graph showing the response signals obtained from channels 7 through 12 of the 12 channels of the MSS array. The horizontal axis represents time (in seconds), and the vertical axis represents output (in mV). Results are shown using 12 different seasonings as samples.

[0034] Figure 5This represents a color map of the seasonings obtained through the examples. The values ​​in parentheses represent the RGB values. The results show the selection of fish sauce, sake, and pure water as three basic aromas using endpoint detection, and the assignment of the colors of each basic aroma to red, green, and blue.

[0035] Figure 6A This graph compares the raw response signals measured in channels 1-6 of the MSS array for barbecue sauce with a mixed signal created by superimposing the response signals of fish sauce, sake, and pure water using coefficients (w1, w2, w3) in the embodiment. The horizontal axis represents time (in seconds), and the vertical axis represents the output (in mV).

[0036] Figure 6B This graph compares the raw response signals measured in channels 7-12 of the MSS array for barbecue sauce with a mixed signal created by superimposing the response signals of fish sauce, sake, and pure water using coefficients (w1, w2, w3). The horizontal axis represents time (in seconds), and the vertical axis represents the output (in mV).

[0037] Figure 6C This graph compares the raw response signals measured in channels 1-6 of the MSS array for soy sauce with a mixed signal created by superimposing the response signals of fish sauce, sake, and pure water using coefficients (w1, w2, w3) in the embodiment. The horizontal axis represents time (in seconds), and the vertical axis represents the output (in mV).

[0038] Figure 6D This graph compares the raw response signals measured in channels 7-12 of the MSS array for soy sauce with a mixed signal created by superimposing the response signals of fish sauce, sake, and pure water using coefficients (w1, w2, w3) in the embodiment. The horizontal axis represents time (in seconds), and the vertical axis represents the output (in mV).

[0039] Figure 7A This is a diagram illustrating an example of the color of the odor output in real time in an embodiment.

[0040] Figure 7B This is a graph showing the time dependence of the color of the odor output in real time and the RGB values ​​in the embodiment. The horizontal axis represents time (in seconds), and the vertical axis represents the RGB values. Detailed Implementation

[0041] In one aspect of the invention, nanomechanical sensors and machine learning are used to select a subset of odors from a set of multiple sample gases (hereinafter, sample gases are also often referred to as "odors"). (When the provided sample is a liquid or solid, the gas generated from it is referred to as the sample gas). In this application, each odor in the subset of odors thus selected is also referred to as a basic odor. It should be noted here that, as can be seen from the above description, the basic odor is not provided in advance (i.e., it is not provided in advance before the application of the invention), but is a point dynamically determined during the implementation of the invention. Of course, in the initial measurement, the basic odor is selected as a subset from multiple odors, and in subsequent measurements, the initially selected basic odor can be reused instead of each subsequent selection of basic odor, and the temporarily selected basic odor can also be reused. In this way, by determining the basic odor, any odor in the provided set of odors can be represented by a combination of odors as basic odors. The representation of any odor based on the basic odor can be simple data (e.g., basic odor 1, basic odor 2... basic odor n are c1, c2... c... respectively). n Data such as percentages (e.g., %) can be converted into information that allows for more intuitive representation of odors through senses other than smell. For example, different colors can be assigned to each basic odor, and the odor can be represented as a color created by mixing these assigned colors. Typically, three basic odors can be selected, and the three primary colors can be assigned to them. Furthermore, not only can odors be represented as combinations of basic odors, but also odors that are the same as or similar to those represented by combinations of basic odors can be synthesized by mixing these basic odors.

[0042] In other words, this invention provides a novel and practical technical idea that does not target an almost infinite totality of information such as all odors (the sensor's response to all odors), but rather extracts a computable range from it. This technical idea can be implemented even when a logically defined "true basic odor" is not found within this range. Furthermore, even if a true basic odor can be identified, the substances that give it that odor are not necessarily easy to obtain, process, or safe for living organisms. In addition, there may be a vast number of basic odors. Therefore, it is also considered that, in reality, decomposing odors based on true basic odors is not a practical solution. This invention is also fully applicable to such situations.

[0043] To illustrate one embodiment of the invention more specifically, firstly, in order to convert odors into response signals that can be processed numerically, multiple sensors exhibiting different responsiveness to various odors are used. As such sensors, for example, there are nanomechanical sensors that detect changes in physical parameters on a defined surface. Surface stress sensors are one type of nanomechanical sensor. In a surface stress sensor, a membrane (sensing membrane) adsorbs a detector such as a gas on the surface, and the expansion and contraction caused by adsorption in the sensing membrane manifests as changes in surface stress on that surface. Since the surface stress changes of various patterns are measured by measuring the combination of the target substance and the material of the sensing membrane (sensing membrane material), qualitative or quantitative analysis of various substances in the sample can be performed by supplying a sample to one or more surface stress sensors and measuring their responses.

[0044] Various sensors for odor detection have been proposed and fabricated. Hereinafter, a nanomechanical sensor known as MSS (Patent Document 1, Non-Patent Document 3) is used as an example, employing piezoelectric resistance-based electrical readout. This sensor is reported to have several advantages over conventional cantilever nanomechanical sensors, achieving high sensitivity, a compact system, and high stability. Furthermore, multiple channels with sensor channel dimensions less than 1 mm and coated with different sensing film materials can be easily prepared. This allows for the simultaneous acquisition of a large amount of odor information and the collection of reliable response signals, which can be considered as the "fingerprint" of the odor sample. Practical applications exist for quantitative odor analysis or odor identification using the feature quantities extracted from this response signal (hereinafter, the vector defined by multiple feature quantities will be called the feature vector). Of course, other types of sensors can also be used.

[0045] To discover the fundamental odor, odor samples are first collected and converted into response signals using multiple MSS (Micro-Sensitive Sensors) of various sensing membranes. Then, high-dimensional feature vectors are extracted from the response signals based on physicochemical knowledge. Next, endpoint detection using machine learning is performed in the high-dimensional feature space to rank all collected odor samples as candidates for fundamental odor. The number of odors to be selected as fundamental odors (the fundamental odor count) is determined, and odors are selected sequentially starting from the highest rank. Other odors are mixtures of these fundamental odors. If the target odor can be represented or approximated by a linear combination of the feature vectors of these fundamental odors, the coefficients of this linear combination can be calculated, for example, using quadratic programming. Thus, any odor can be represented (approximated) as a mixture of a few fundamental odors.

[0046] Here, the term "endpoint detection" mentioned above is given in general. Endpoint detection as used in this application refers to any data analysis method that selects samples located at the ends in the feature space. Various such machine learning algorithms have been proposed (e.g., Non-Patent Document 4). In this application, as an example of such a method, the following describes a method using the method of Non-Patent Document 5 to find points located at the ends in various directions of the feature space, count the frequency of each point, select points in descending order of frequency, and select odors that correspond to these points as basic odors.

[0047] Furthermore, in the field of machine learning, "end" is also called an endpoint or endmember. Therefore, in this application, a point located at the end of the feature space is defined as an endpoint, and the data analysis method for detecting endpoints is called endpoint detection. As mentioned above, there are various data analysis methods for detecting endpoints, and the specific method described below is given as an example in this application. However, it should be noted that endpoint detection is not limited to this specific method, and any data analysis method commonly used for endpoint detection can be used.

[0048] It should be noted that, in addition to dynamically determining which odor is selected as the basic odor, the number of basic odors can also be dynamically set. For example, instead of setting the number of basic odors to a predetermined fixed number, if the approximation of odors other than the basic odors synthesized when a certain number of basic odors is selected is lower than the expected value, the approximation accuracy can be improved by sequentially adding the superior odors of the remaining odors in the above sorting to the basic odors. That is, by determining the intensity (weight) of each basic odor and mixing them, odors other than the basic odors can be synthesized (in most cases, approximately). This synthesis can be achieved by actually mixing the basic odors according to the proportions of the aforementioned weights. Alternatively, instead of actually synthesizing odors, odors other than the basic odors can be represented in data as a combination of the selected basic odors. This representation can be achieved, for example, by reconstructing the feature vector of the odor being approximated as a linear combination of the feature vectors of the basic odors, thus numerically approximating a linear combination of the basic odors. The approximation of the synthesized and approximated odors to the synthesized and approximated object, i.e., the target odor, can be determined by a sensing experiment, or by calculating the difference between the feature vectors of the two odors and the difference in the sensor's response signal itself.

[0049] The number of basic odors can be arbitrarily set according to the intended use of the basic odors. For example, when displaying odors by establishing a correspondence between odors and colors, the top three ranked samples are designated as three basic odors. They are assigned to the three primary colors of red, green, and blue, respectively. However, it is not necessary to use the three primary colors as the colors corresponding to the basic odors; any three colors can be assigned as the colors of the three basic odors. In this case, the position of the target odor on the chromaticity coordinates is calculated, and the color of the target odor is obtained based on its position on the chromaticity coordinates. Devices for outputting odors in real time in a single color can be made using, for example, MSS (Microcontroller System), compact computers equipped with machine learning methods, LED lights, etc.

[0050] The following explains the specific method for selecting a basic scent.

[0051] [Extracting features from the response signal]

[0052] Odor is measured using a sensor array consisting of MSS (hereinafter, each MSS is also referred to as a channel) coated with different sensing membrane materials. In surface stress sensors such as MSS, a sample to be measured and a reference fluid (a reference gas for odor measurement) are alternately supplied to the sensor at a predetermined period. When the odor to be measured is supplied to the sensor, the components of the odor are adsorbed onto the sensing membrane, but the adsorbed components are desorbed during the flow of the reference gas, hence it is also called a purge gas in this sense. Nitrogen (N2) is used in this example, thereby obtaining a response signal based on the periodic adsorption and desorption that occurs in the sensing membrane.

[0053] A conceptual illustration of an example of the configuration of a measuring device used to perform this measurement is shown in [the diagram]. Figure 1Here, reference gas is supplied to the gas flow paths of both systems from the left side of the diagram. Each flow path contains an MFC (mass flow controller) that alternates between gas delivery operations (i.e., gas delivery operations are performed in opposite phases), delivering a specified flow rate of gas downstream during operation. Thus, during the operation of the MFC in the upper gas flow path, reference gas is directly supplied to the vial on the right. In the lower gas flow path, reference gas from the MFC is supplied to the top space of the upstream vial, and sample gas accumulated in the top space from the evaporation of the sample (here, a liquid sample) contained in that vial is delivered downstream through the gas flow path. Therefore, during the operation of the MFC in the lower flow path, sample gas diluted with reference gas is delivered to the aforementioned vial on the downstream side. Thus, reference gas and sample gas are alternately supplied to the MSS (a sensor array comprising multiple MSSs) shown in the diagram. In response, response signals from each MSS in the sensor array are sent to an information processing device for various processing described later. It should be noted that, in order to perform measurements on multiple samples, the samples entering the vials need to be changed appropriately. This sample change can be done manually or automatically by adding and removing the samples into the vials. Alternatively, odor generating containers containing each sample can be prepared in advance, and the samples can be changed manually or automatically. Furthermore, all of the aforementioned odor generating containers can be connected to multiple gas flow paths, and the flow paths can be switched sequentially using MFC operation and valve control to perform a series of odor measurements. In the information processing device that supplies the measurement results of the multiple odors obtained in this way, basic odor selection, odor expression based on the selected basic odor, and odor prompts are performed. The information processing device can be equipped with various storage devices or interfaces that can be used to store the selected basic odor and the odor information expressed by the basic odor, or to supply it to other devices. Additionally, output devices such as display devices can be provided to prompt the user about the odor in any form. Furthermore, odor synthesis using the basic odor can be performed, for example, by appropriately switching the gas flow paths connected to the multiple odor generating containers through the information processing device and mixing the resulting sample gases. Of course, Figure 1 The configuration shown is merely an example and can be modified as needed.

[0054] To explain this information processing more specifically, firstly, information is obtained from various channels coated with different sensing film materials, such as... Figure 2 The response signals are shown. In each response signal, a peak appears at each of the aforementioned defined periods; here, each peak is considered independent. Furthermore, the following four parameters representing the peak shape are extracted for each peak. It should be noted that the method for extracting the feature quantities is not necessarily limited to this.

[0055] Parameter 1: by Figure 2The gradient between point A, corresponding to the moment of switching from the reference gas to the sample gas (odor), and point B, where the response signal approaches saturation due to continuous odor supply, is defined. This is related to the odor adsorption process. Specifically, point B can be defined as the point on the response signal at the moment a specified time has elapsed since point A. Alternatively, point B can be defined as the point where the response signal value increases to the minimum value of the response signal + (maximum value - minimum value) × a specified ratio (e.g., 95%).

[0056] Parameter 2: Defined by the gradient between point B and point C, the point at which the supplied gas is switched from the sample gas (odor) to the reference gas. It includes information about the quasi-equilibrium state.

[0057] Parameter 3: Defined by the gradient between point C, which serves as the starting point of the desorption process for the sample gas (odor), and point D, where the value of the response signal approaches its minimum value due to continuous supply of reference gas. As for point B, point D can be determined by setting a point on the response signal at a time elapsed from point C. Alternatively, point D can be set by setting the value of the response signal to decrease to the minimum value of the response signal + (maximum value - minimum value) × a specified ratio (e.g., 5%).

[0058] Parameter 4: Defined by the peak height. It includes information on the adsorption capacity of each sensing membrane material. Figure 2 In the peak shape shown, the height of point C is parameter 4. However, typically, the output signal does not increase monotonically during sample supply. Sometimes, the peak height reaches its maximum value between points B and C, and the output signal decreases from that point onwards.

[0059] It should be noted that the shape of the aforementioned peaks is not actually constant, and sometimes changes gradually immediately after the alternating supply of the sample gas (odor) and the reference gas begins. Therefore, as the peaks for extracting the above parameters, it is sometimes preferable to appropriately select, for example, the peaks that appear after a sufficient number of alternating supply cycles from the start of the alternating supply.

[0060] Furthermore, the sample gas (odor) can be a gas from the outset, or the odor source can be a liquid or a solid, but the gas evaporated from it can also be the subject of the measurement. It should be noted that the examples are described with the odor source being a liquid, but this description does not lose its generality.

[0061] [Endpoint detection method for selecting basic odors]

[0062] To identify a set of basic odors from the collected odor samples, the endpoint detection method shown below is performed on the high-dimensional feature vectors of the odors obtained by measurement using MSS. Of course, as mentioned above, the endpoint detection method shown below is only one example, and any other method can certainly be used. For simplicity, the number of basic odors is set to 3. Furthermore, the feature quantities extracted from each odor sample are reference values ​​obtained from 12 channels coated with different sensing films. Figure 2 The parameters 1 through 4 are described above. Therefore, 12 × 4 = 48 feature quantities are provided for each odor sample, and the feature vector is 48-dimensional. Furthermore, this explanation of the number of basic odors and the number of feature quantities as specific values ​​does not lose its generality.

[0063] Let N be the number of odor samples. Let X be the vertical vector containing 48-dimensional features for the i-th odor sample. Then, the feature matrix X is defined as follows.

[0064]

[0065] The rows of the feature matrix X are standardized to have a mean and standard deviation of 0 and 1, respectively.

[0066] If we express the above standardization mathematically, it is the value x for each row of the above feature matrix. j1 x j2 , ..., x ji , ..., x jN , will x ji Convert to (x) ji The treatment of -μ) / σ. Here, the subscripts i and j, and the constants μ and σ are as follows.

[0067] i: The index of the column of the feature matrix X, as mentioned above, is the number of each odor sample from 1 to N.

[0068] j: The row index of the feature matrix X, which establishes a correspondence with the MSS number used to determine the value and the parameter number obtained from the response signal of the MSS. For example, when using 12 MSS (MSS1 to MSS12) as in the example above, and extracting 4 parameters (parameter 1 to parameter 4) from each MSS, j can be set as j = (MSS number - 1) × 4 + parameter number.

[0069] μ:x j1 x j2 , ..., x ji , ..., x jN The arithmetic mean.

[0070] σ:x j1 x j2 , ..., x ji , ..., x jN The standard deviation.

[0071] This standardization refers to shifting the eigenvectors in the feature space in parallel to distribute them around the origin of the feature space by setting the average of each row of the feature matrix to 0. Furthermore, setting the standard deviation of each row to 1 means unifying the "utility" of each MSS and parameter—that is, the degree to which they contribute to the selection of the basic odor—by making the magnitude and deviation of the MSS used and some of the eigenvalues ​​(parameters) derived from these MSS consistent.

[0072] Furthermore, setting the mean of each row to 0 in the above standardization process does not fundamentally affect endpoint detection. In machine learning, standardization with the mean set to 0 and the standard deviation set to 1 (also known as z-score) is frequently used; therefore, this standardization is adopted here as well. Of course, the normalized values ​​of each row can be used, or the rows themselves can be used instead of standardization; standardization is not always necessary. Here, standardization refers to setting the variance to 1 and the mean to 0, while normalization refers to transforming the rows by setting the maximum to 1 and the minimum to 0. It should be noted that scaling transformations as part of data preprocessing are not necessarily limited to standardization or normalization; any scaling transformation can be appropriately applied. However, it should also be noted that such scaling transformations are not mandatory.

[0073] First, in the endpoint detection method, a value K is set as large as possible (although it also depends on the required precision of the basic odor selection, but is preferably more than 200 times the number of features) to generate a set of K random unit vectors. .

[0074] Here, "random" means that directions within the feature space (in this case, a 48-dimensional vector space) exhibit the same randomness. Each unit vector... The dimension of the feature vector is the same as the dimension of the feature vector, which is 48 dimensions in this case.

[0075] Additionally, prepare vectors .

[0076] This vector is used to store endpoint-related scores, also known as endpoint scores.

[0077] In this endpoint detection method, the following steps are repeated K times.

[0078] (1) The feature vector of all standardized odor data Projected onto each unit vector That is, calculating the characteristic vectors of each odor after standardization. (i=1~N) and a given unit vector The inner product is the vector of the element. This refers to the characteristic vectors of each odor sample 1 to N in the unit vector. The coordinates of the direction (note that the feature vector starts from the origin of the feature space).

[0079] (2) Obtain the vector as follows The indices I of the smallest and largest features in the features (coordinates) - and I + .

[0080]

[0081] (3) Index I for the selected elements - and I + Update the vector as follows The component. That is, the endpoint score of the element with the selected index is increased by 1.

[0082]

[0083] An example of this repetitive step is shown below. Figure 3 When K is sufficiently large, it is considered to have a large endpoint score n. i The odor sample is positioned at the end of a high-dimensional feature space (feature vector space). Therefore, n i Larger odor samples are considered to be the samples at the very end of the odor sample range, and n is considered to be... i The order from largest to smallest is the ranking of the candidate basic odors. The top three samples here are designated as the three basic odors, and their indices are defined as e1, e2, and e3, respectively. When representing other odors with color, these basic odors are assigned to the vertices of the chromaticity triangle.

[0084] [Any method of representing a scent using basic scents]

[0085] The above-described [Endpoint Detection Method for Selecting Basic Odors] describes a method for selecting multiple basic odors from an odor sample (three basic odors were exemplified in the above description; the following description will also use three basic odors). To represent other odors through combinations of basic odors, the feature vectors of other odors are represented by linear combinations of the feature vectors of the three basic odors. At this point, a linear combination of feature vectors that best represents the target odor is searched. (w1, w2, w3) is set as the normalized feature vector X representing the three basic odors. e1 X e2 and X e3The coefficients of the linear combination. Furthermore, the following explains how to establish a correspondence between one of the three primary colors of color and each of the three basic odors, thereby representing any odor as a mixture of the three primary colors, without losing universality.

[0086] The coefficients representing the above linear combination are obtained by performing a quadratic programming method. That is, when the standardized feature vector of the target odor is set as X, the coefficients (w1, w2, w3) are found when the difference Δ is minimized, as defined by the following formula.

[0087]

[0088] The following two conditions are set.

[0089]

[0090] The former condition requires that the target odor be represented as a positive mixture of three basic odors. This is because basic odors cannot be mixed at negative concentrations. The latter condition, on the other hand, requires that w... i The conditions required to mix the basic odors correspond to the concentrations needed. When using color to represent other odors, the position of the chromaticity triangle for the target odor is provided below.

[0091]

[0092] The color in the chromaticity triangle corresponding to that location is used to determine the color that establishes a correspondence with the target odor. It should be noted that although the mixed concentration of the basic odor is determined here to approximate a normalized feature vector, the approximation is not limited to this; normalized feature vectors, unnormalized feature vectors, or direct response signals can also be processed.

[0093] Furthermore, the top three samples as candidates for basic odors are considered as three basic odors, and used as vertices of a color triangle, much like the three primary colors of a color, thus representing all odors with color. As an example of a sample, seasonings are used in the embodiments. However, as described below, these samples or test objects are merely examples, and it is evident that a wide variety of other samples or chemical substances can be arbitrarily selected. Furthermore, it is also evident that this method can be applied to liquid or solid samples that produce odors.

[0094] Example

[0095] [Feature extraction from response signals and the use of sensing film materials]

[0096] The above references were extracted. Figure 2The parameters 1 to 4 are explained below. Here, the repetition period for the supply of the sample (odor) and the reference gas is set to 10 seconds, and the time difference between point A and point B is determined to be 0.5 seconds. Similarly, the time difference between point C and point D is also set to 0.5 seconds.

[0097] Here, as 12 sensing film materials, various surface-functionalized silica / titanium dioxide composite nanoparticles, various polymers, and SiO2-C16TAC (silica / hexadecyltrimethylammonium composite particles) were prepared. These were coated onto the MSS surface using various coating methods. The materials coated on each channel and the coating methods are described later in [Details of the Examples - Fabrication of MSS Chips Coated with Various Materials].

[0098] As mentioned above, an odor sample is represented by 48 parameters (the product of 4 features and 12 channels). This is a high-dimensional feature quantity, i.e., a feature vector, for each odor sample.

[0099] [Basic aroma selection from seasonings]

[0100] To illustrate, a basic odor was selected from the aromas of seasonings using MSS and machine learning. Twelve seasonings were prepared: pure water, ketchup, mayonnaise, lemon juice, oyster sauce, hot sauce, sake, tsuna sauce, barbecue sauce, vinegar, soy sauce, and fish sauce. These liquid samples were placed in vials, and the gas accumulated in the headspace of the vials was measured using MSS. In this embodiment, N2 gas, the same as the reference gas, was used as the carrier gas to expel the gas (odor) accumulated in the headspace and introduce it into the MSS. The dependence of the response signal on the sensing membrane material is shown below. Figure 4A and Figure 4B Except for channel 5, which is coated with aminopropyl-modified silica / titanium dioxide composite nanoparticles, the signal shape does not vary much across channels. However, peak heights differ between samples, with sake consistently exhibiting the largest peak and fish sauce the smallest. Therefore, from the perspective of the raw response signal, sake and fish sauce are considered candidates for basic odors. On the other hand, for other odor samples, it is difficult to select basic odor candidates solely by observing the raw response signal. It should be noted that the intensity of the response signal obtained in each channel is different for each channel. Figure 4A and Figure 4B In this approach, visibility is improved by making the peak values ​​of the response signals in each channel almost the same height.

[0101] 48-dimensional features were extracted from the response signal obtained from the MSS, and endpoint detection was performed to select three basic odors from the prepared odor data. The example described in [Endpoint Detection Method for Selecting Basic Odors] was used as the endpoint detection method. Here, the parameter used in endpoint detection was set to K=10000, and the total endpoint score was 20000. The table below shows the ranking of the endpoint scores. Here, the top three samples are fish sauce, sake, and pure water. More than half of the total endpoint scores are concentrated in the top two odors, indicating that fish sauce and sake are indeed at the endpoints in the prepared condiments. This is consistent with the fact confirmed by the original response signal. On the other hand, compared to these two odors, the endpoint score of pure water, which ranks third, is relatively small. In this case, it is difficult to directly determine the order by visual inspection or simple comparison between feature vectors. The present invention has the feature that it can also handle this situation by performing endpoint detection. Furthermore, it can be seen that in the data of the embodiment, since the endpoint scores of barbecue sauce, oyster sauce, and hot sauce are quite small, these odors are located deep inside the region of feature vector distribution in the 48-dimensional feature space. Thus, by performing endpoint detection, the relative positions of the prepared odors within the feature space can be determined. Based on the endpoint detection results, the basic odors in the prepared seasoning dataset can be identified as fish sauce, sake, and pure water.

[0102] [Color representation of odor]

[0103] To convert the target odor into color, coefficients (w1, w2, w3) are calculated using quadratic programming (details are shown above in [Representation of Arbitrary Odors Using Basic Odors]). These coefficients represent the normalized feature vector X of the three basic odors extracted from the response signal. e1 X e2 and X e3 The coefficients of the linear combination are given. Here, fish sauce, sake, and pure water are used as the three basic odors selected above. The table below shows the results of obtaining these coefficients using quadratic programming. Additionally, the difference Δ between the linear combination and the standardized eigenvector of the target odor is also shown in the table.

[0104] Table 1

[0105]

[0106] The Δ value represents the accuracy when the target odor sample is represented using the feature vectors of the three basic odors. Therefore, samples with small Δ values, such as hot sauce and barbecue sauce, can be accurately represented by a mixture of the three basic odors. On the other hand, it was confirmed that when odors other than fish sauce, sake, and pure water are used as basic odors, the average Δ value increases. This fact means that endpoint detection is effectively functioning to achieve better color representation. Furthermore, using these coefficients, the colors of each odor sample, their positions in the chromaticity triangle, and the RGB values ​​of each odor are summarized in... Figure 5 However, in the original color chart, the colors of fish sauce, sake, and pure water were assigned to the three primary colors of red, green, and blue. Figure 5 It is a diagram converted into a black and white illustration.

[0107] This color chart shows the components of soy sauce and dipping sauce that do not contain pure water, represented as a mixture of fish sauce and sake. Note that the colors themselves are meaningless in this representation. This is because colors unrelated to the original odor samples are assigned to the three basic odors, and other odor samples are thus transformed into colors based on their relative evaluation using these colors. On the other hand, with a prepared dataset, it is also possible to apply the liquid sample's own color or a color suitable for human perception to the three basic odors, creating a chromaticity triangle through the mixing of these colors, thereby giving the colors themselves meaning.

[0108] In addition, Figures 6A to 6D In this example, using the coefficients (w1, w2, w3) obtained for soy sauce and barbecue sauce, a mixed signal created by superimposing the response signals of fish sauce, sake, and pure water is compared with the response signals obtained from actual barbecue sauce and soy sauce. Thus, it can be seen that these peaks are very consistent across all channels. This indicates that, much like how a given color can be decomposed into its three primary colors, the obtained coefficients (w1, w2, w3) can be used to approximately decompose the original response signal of the target odor into the response signals of the three basic odors. In other words, this also shows that, as explained above, although only 48-dimensional features were used to estimate the coefficients for the linear combination of odors, from the perspective of the original response signal, even if the obtained coefficients are considered as the concentrations when mixing the basic odors, it is still a good approximation. Therefore, this information is also helpful when creating new odors from three basic odors. It should be noted that the intensity of the response signal obtained in each channel is different for each channel, but as mentioned above regarding... Figure 4A and Figure 4B As mentioned above, in Figures 6A to 6D In this approach, visibility is improved by making the peak values ​​of the response signals in each channel have almost the same height.

[0109] [A device that changes the color of odors in real time]

[0110] With three basic odors predetermined, various odors can be instantly converted into forms that can be recognized by senses other than smell, such as color, simply by performing quadratic programming. Therefore, the color of, for example, the target odor (obtained by converting the odor into color as described above) can be output in real time. Thus, a device for displaying the color of odors in real time is created using an MSS (Microcontroller System), a compact computer equipped with a machine learning information processing unit, and LED lights. Here, LED lights are used as the device for outputting the color, but any output device can be used. Here, as an example, fish sauce, sake, and pure water are used as the three basic odors. Furthermore, the colors of each basic odor are set to red, green, and blue. Here, since N2 gas is used as the reference gas and carrier gas when measuring the response signal of the basic odors, N2 gas is also used as the reference gas and carrier gas when measuring the target odor. However, air can also be used as the reference gas and carrier gas.

[0111] Display the real-time measurement results Figure 7A and Figure 7B . Figure 7A The first part of the measurements shown is the real-time sequential measurement of fish sauce, sake, pure water, barbecue sauce, and soy sauce. Additionally, in Figure 7B The text summarizes the time dependence of the LED light output color and RGB values. In the fabricated device, during sample switching, the operation of temporarily removing the vial containing the completed sample from the gas flow path and connecting the vial containing the next sample to be measured introduces ambient air into the vial, mixing with the sample odor. Due to this effect, the coefficients change significantly during sample switching, resulting in an output color that differs from the color the target odor should represent. Next, it is confirmed that when the response signal reaches equilibrium, the output and... Figure 5 The target odor should be a color close to the color shown. Figure 7B (An example is shown on the lower side). Additionally, in the latter part of the measurement, the aroma of three mixtures of fish sauce and sake at certain concentrations was measured. Mixed samples with concentrations of 4:1, 2:1, and 1:1 were prepared. It was confirmed that the output color was between that of fish sauce and sake, that is, a mixture of red and green.

[0112] [Details of the Example - Fabrication of MSS Chips Coated with Various Materials]

[0113] The MSS chip used in this embodiment has a total of 12 channels (hereinafter also referred to as Ch), and the materials and coating conditions for each channel are as follows.

[0114] Ch1: Aminopropyl modified silica / titanium dioxide composite nanoparticles (spray coating).

[0115] Ch2: Octadecyl-modified silica / titanium dioxide composite nanoparticles (inkjet coating, 1 g / L, 400 times).

[0116] Ch3: Phenyl-modified silica / titanium dioxide composite nanoparticles (inkjet coating, 1 g / L, 200 times).

[0117] Ch4: Polymethyl methacrylate (inkjet coating, 1 g / L, 300 times);

[0118] Ch5: Aminopropyl modified silica / titanium dioxide composite nanoparticles (inkjet coating, 1 g / L, 600 times).

[0119] Ch6: Octadecyl-modified silica / titanium dioxide composite nanoparticles (inkjet coating, 1 g / L, 800 times).

[0120] Ch7: Phenyl-modified silica / titanium dioxide composite nanoparticles (inkjet coating, 1 g / L, 500 times).

[0121] Ch8: Silica / hexadecyltrimethylammonium composite particles (inkjet coating, 1g / L, 1000 times);

[0122] Ch9: TenaxTA (20 / 35 mesh) (inkjet coating, 1 g / L, 300 cycles, stage temperature 100°C);

[0123] Ch10: TenaxTA (20 / 35 mesh) (inkjet coating, 1 g / L, 300 cycles, stage temperature 20°C);

[0124] Ch11: TenaxTA (60 / 80 mesh) (inkjet coating, 1 g / L, 300 cycles, stage temperature 100°C);

[0125] Ch12: TenaxTA (60 / 80 mesh) (inkjet coating, 1 g / L, 300 times, stage temperature 20°C).

[0126] Various silica / titanium dioxide composite nanoparticles coated on Ch1, 2, 3, 5, 6, and 7 were synthesized according to methods known to those skilled in the art as described in Non-Patent Documents 6 and 7. The silica / hexadecyltrimethylammonium composite particles coated on Ch8 were synthesized according to methods also known to those skilled in the art as described in Non-Patent Document 8. The materials coated on Ch4, 9, 10, 11, and 12 were all commercially available, and the purchased products were used directly without purification.

[0127] The coating of various materials on the MSS chip is carried out as follows. Ch1 is coated by spray coating, and the rest is coated by inkjet coating. These coating methods are well known to those skilled in the art. For details on spray coating, please refer to Non-Patent Document 7, and for details on inkjet coating, please refer to Non-Patent Document 9. It should be noted that for any of the materials in Ch2 to Ch12, a dispersion or solution with a concentration of 1 g / L is prepared for coating. It should be noted that the number of coats, stage temperature, and other conditions vary for each channel.

[0128] Industrial applicability

[0129] The application of this invention is not limited to odors; it can be used in any field where a basic odor is useful, selected from all gaseous, liquid, and solid samples. For example, any odor can be represented, suggested, or synthesized using a relatively small amount of basic odor. Furthermore, as an application example, a basic odor can be selected from various samples such as breath, sweat, saliva, tears, other bodily fluids, or gases and odors emitted from the body. The odor of any sample being measured can be used as a combination of the basic odors selected by the individual, and this combination can be represented in a form that can be recognized by other senses, such as color. Moreover, this invention is expected to be widely used in various fields that generate odors or gases, such as food manufacturing, product design, entertainment, storage, distribution and security, or pharmaceuticals.

[0130] Existing technical documents

[0131] Patent documents

[0132] Patent Document 1: International Publication No. 2011 / 148774;

[0133] Non-patent literature

[0134] Non-patent literature 3: G. Yoshikawa, T. Akiyama, S. Gautsch, P. Vettiger, and H. Rohrer, “Nanomechanical Membrane-type Surface Stress Sensor” Nano Letters 11, 1044-1048 (2011).

[0135] Non-patent document 4: https: / / ieeexplore.ieee.org / document / 1411995;

[0136] Non-patent document 5: https: / / ieeexplore.ieee.org / abstract / document / 1576691;

[0137] Non-patent literature 6: K. Shiba, T. Sugiyama, T. Takei and G. Yoshikawa, Chem. Commun., 2015, 51, 15854-15857.

[0138] Non-patent literature 7: K. Shiba, R. Tamura, G. Imamura and G. Yoshikawa, Sci. Rep., 2017, 7, 3661.;

[0139] Non-patent literature 8: K. Kambara, N. Shimura and M. Ogawa, J. Ceram. Soc. Jpn., 2007, 115, 315-318.;

[0140] Non-patent literature 9: K. Shiba, R. Tamura, T. Sugiyama, Y. Kameyama, K. Koda, E. Sakon, K. Minami, HTNgo, G. Imamura, K. Tsuda and G. Yoshikawa, ACS Sensors, 2018, 3, 1592-1600.

Claims

1. A basic odor selection method, It uses sensors to detect each odor in a collection of odors. Based on the detection results of each odor, a subset is selected from the odor set, consisting of the basic odors of the odors in the odor set. This subset represents any odor in the odor set through combinations of the odors in this subset, i.e., the basic odors. In the basic odor selection method described above Each odor in the odor set is a sample gas that may contain multiple chemical substances. The detection of each odor includes: Calculate the feature vector corresponding to each odor, which is composed of multiple features obtained from the sensor's output signals for each odor. The selection of the subset consisting of the basic odors includes: determining the number of odors selected as basic odors; and selecting the determined number of basic odors from the set of feature vectors corresponding to each odor in the odor set, based on the subset of feature vectors selected by endpoint detection in the feature space containing the feature vectors. The number of times a point is located at an end in any direction of the feature space is counted, and the determined number of feature vectors are taken from the set of feature vectors in descending order of the number of times the feature vectors are counted, thereby obtaining the subset of the feature vectors.

2. The basic odor selection method as described in claim 1, wherein, The rows of the feature matrix, which is constructed by taking the feature vectors corresponding to each of the odors as the vertical vectors, are scaled.

3. The basic odor selection method as described in claim 1, wherein, The sensor is a surface stress sensor.

4. The basic odor selection method as described in claim 1, wherein, The sensor is a sensor array with multiple sensor elements. The output signal includes a separate output signal from each of the plurality of sensor elements.

5. The basic odor selection method as described in claim 1, wherein, The number of scents selected as the basic scents is 3.

6. A basic odor selection device, wherein, It is equipped with: sensors; A gas supply mechanism that alternately supplies sample gas and reference gas having the respective odors to the sensor; and An information processing device that receives the output signal from the sensor. The odor of the sample gas corresponding to each element of the subset of the feature vector obtained by the information processing device by the basic odor selection method according to any one of claims 1 to 5 is selected as the basic odor.

7. A method for synthesizing an odor, wherein, Based on a plurality of basic odors selected by the basic odor selection method according to any one of claims 1 to 5, a given odor is synthesized in the form of an odor formed by mixing the selected basic odors.

8. An odor synthesis apparatus, wherein, It is provided with a mechanism for mixing the sample gas with each of the basic odors selected by the basic odor selection device of claim 6.

9. A method for representing odor, wherein, Based on a plurality of basic odors selected by the basic odor selection method according to any one of claims 1 to 5, a given odor is represented as a combination of the selected basic odors.

10. The odor representation method as described in claim 9, wherein, The combination of the basic odors is represented by a linear combination of the feature vectors of the basic odors.

11. An odor expressing device, wherein, It is equipped with: sensors; A gas supply mechanism that alternately supplies sample gas and reference gas having the respective odors to the sensor; and An information processing device that receives the output signal from the sensor. It also includes a mechanism for storing or providing to the outside the odor representation results obtained by the information processing device through the odor representation method of claim 9 or 10.

12. An odor prompting method, wherein, The combination of basic odors as described in claim 9 or 10 can be indicated by means of sensory recognition other than smell.

13. The odor prompting method as described in claim 12, wherein, The method establishes a correspondence between different colors and each of the basic odors through cues that can be recognized by senses other than smell, and provides odor cues by mixing the different colors.

14. The odor prompting method as described in claim 13, wherein, The different colors that correspond to the various odors in the basic odors are the three primary colors.

15. An odor indication device, wherein, It is equipped with: sensors; A gas supply mechanism that alternately supplies sample gas and reference gas with the respective odors to the sensor; An information processing device that receives the output signal from the sensor; as well as An output device that outputs the odor prompting result obtained by the information processing device using the odor prompting method according to any one of claims 12 to 14.

Images