JP2986582B2 - Odor identification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an odor discriminating apparatus for discriminating odors in an environment.

[0002]

2. Description of the Related Art Researches for identifying the type of odor using an odor sensor have been actively conducted. Identification of odor
It is not possible with simple equipment, and is detected as a molecular weight by gas chromatography, mass spectrometer, etc.
Further, a gas sensor using ordinary stannic oxide SnO ₂ has almost no selectivity to odor.

[0003]

SUMMARY OF THE INVENTION The present invention improves the selectivity of a conventionally used odor sensor.
It is an object of the present invention to obtain an odor discriminating apparatus capable of accurately discriminating a wide variety of odors by a simpler method than conventionally performed.

[0004]

Therefore, according to the present invention, a plurality of types of different odor sensors and sensor output values from these odor sensors are identified to perform grouping of odors in the environment. Calculating an odor level for each of the grouped odor types, e.g., an odor identification unit composed of a neural network, and calculating the odor level for each of the grouped odor types from the odor identification unit An odor discriminating apparatus comprising: a determination unit that determines a meaningful odor based on a result.

[0005]

The odor sensor has a specific relationship between a certain odor and the amount of odor generated and the output of the odor sensor. Since the relationship is also different, if a plurality of odor sensors having different responses to the odor are used and the response of each odor sensor is comprehensively determined, the odor level for each type of odor, Alternatively, it is possible to classify the odor into several groups and identify the odor level for each type of each group. The odor discriminating unit, which is preferably constituted by a neural network,
Such identification is performed based on sensor output values from a plurality of different odor sensors. The determining unit determines and outputs, for example, an odor that is meaningful to human health, a odor that is meaningful for monitoring a dangerous environment, or another meaningful odor, based on the identification content of the odor identification unit. .

[0006]

Hereinafter, embodiments of the present invention will be described.
Prior to that, the contents of the test and research on which the present invention is based will be described.

Types of sensors used in the test First, SnO ₂ was deposited on an alumina substrate to a thickness of 1000 ° (sensor 1) and 2500 ° (sensor 2), respectively, and two types of odors were formed by sputtering a Pt heater on the back surface. Sensors were prepared, and these were heated to 350 ° C. with a heater before use. The measurement circuit of the odor sensor is shown in FIG. The measurement circuit of FIG. 1 is provided for each of the sensor 1 and the sensor 2.

[0008] The types of tests performed were Test 1 for measuring the sensitivity to burning odor, Test 2 for measuring the sensitivity of environmental odors other than burning odor, and high sensitivity smoke in the state of smoking. Sensor (IE
Company I's VESDA [very early smoke detection appar
atus] and a test 3 for measuring the sensitivity comparison with the odor sensor.

[0009] Burning odor sensitivity measurement method (for tests 1 and 3
Case) Eighteen kinds of materials to be heated were prepared, heated at a heating rate of 10 ° C./min, and the sensor sensitivity to odor generated at that time was measured. At the same time, the smoke sensitivity was 0.01.
The smoke density was measured with both IEI's VESDA, a sensitive smoke detector and a particle counter capable of detecting% / m smoke, and the smoke temperature was measured with a thermometer.
As shown in FIG. 2, the measurement was performed using a HEPA filter (high eff
An iciency particle filter (high-efficiency particle filter) and a deodorizing filter are used to control the number of particles in the air to 10 ⁴ particles / m ³ or less with a size of 0.3 μm and placed in a clean duct from which odors have been removed. The material was heated to 330 ° C. in a closed container.

[0010] Two kinds of containers were properly used. In Test 1, a 1.8 L (liter) glass container was used. In this case, two odor sensors were mounted in the bell jar, and the air inside was guided to the particle counter at a flow rate of 300 mL / min. In Test 3, a stainless steel 65 L (liter) container was used. At the time of test 3, the air inside the container was pumped at a flow rate of 3 L / min by VES.
Led to DA and odor sensors.

Method for measuring sensitivity of environmental odor (in the case of test 2 ) At the time of measurement in test 2, an odor sensor was placed in a 1.8 L (liter) bell jar, and an odor-generating substance was placed inside the bell jar in a sealed state. Measured.

Test Results (a) Sensitivity to Burning Odor (Test 1) As shown in the table of FIG. 3, the sensitivity of the thick film sensor 2 to cellulosic materials (copy paper, newsprint, filter paper, pine) is high. However, the relatively low sensitivity of the same family of materials to carton paper is due to poor paper quality,
It is presumed that the fixing agent became smoke and flew off at a low temperature of ° C. Further, the temperature of the odor and the generation of smoke became higher as the paper contained less impurities. The cotton did not smoke even when the temperature reached 330 ° C., and generated only odor. The sensor sensitivity to cotton showed a high value. As the electric circuit board, a glass epoxy board and a paper epoxy board were heated to measure the sensor sensitivity. The paper epoxy board emits odor at 150 ° C.
It smoked at 80 ° C. 2 for epoxy board (with resist)
At 10 ° C. and 300 ° C., respectively, odor and smoke were generated. As a result, the sensor sensitivity to the glass epoxy substrate showed a high value equal to or higher than the burning smell of the cellulosic material, while the paper epoxy substrate showed a low value. The sensor sensitivity of plastic and rubber materials to odor was higher in the following order.

Acrylic> Polyethylene = Styrofoam>Rubber> ABS Resin From these facts, it was found that the sensitivity of the sensor to the burning odor of a material having a large difference between the odor generation temperature and the smoke generation temperature was high. When the cellulosic material and the glass epoxy substrate are heated at a rate of 10 ° C./min, the odor sensor can sufficiently detect the smoke before it emits.

When the sensor temperature is obtained as a parameter,
The sensitivity to the burning odor of the seven materials is shown in the table of FIG.
In the case of copy paper, newsprint, filter paper, and ABS resin, when the heater temperature was lowered, the sensor sensitivity of both the sensor 1 and the sensor 2 was improved, but the response speed was reduced. Conversely, for the acrylic resin, the sensor sensitivity was greatly improved when the heater temperature was increased. For glass epoxy substrates, a film thickness of 10
The thin sensor 1 having a thickness of 00 ° showed the same tendency as the acrylic resin. In the case of the thick sensor 2 having a thickness of 2500 °, the heater temperature 3
It showed maximum sensitivity at 17 ° C.

(B) Sensitivity to odors generated in a normal environment (Test 2) There are very many types of odors generated in a normal environment. In this experiment, a sensor for odors of the materials shown in the table of FIG. 5 was used. 1 and the sensor sensitivity of the sensor 2 were measured. Very high sensitivity to coffee powder and cigarette butts. Since these materials have been burnt before, it may be natural that the sensitivity will increase. In addition, the sensor responded highly sensitively and quickly to the odor of a material containing alcohol or a volatile solvent. The sensor also had high sensitivity to fragrances containing higher alcohols. On the other hand, the sensitivity of the sensor to confections with considerable odor was low.

(C) Comparison of sensitivity between VESDA and odor sensor (Test 3) To compare the sensitivity between VESDA and odor sensor
The material was heated and smoked in an L (liter) test box, producing odor and smoke. However, the smoke density is max.
The amount of the heating material was controlled so as to be 15% / m. Various materials are smoked, and the output change value of the odor sensor when the smoke density is 0.1% / m is shown in the table of FIG. The main materials in the order of higher sensitivity of the sensor 2 having a film thickness of 2500 ° are as follows.

Cotton> Copy paper> Glass epoxy board (with resist)>Pine>Newspaper> Glass epoxy board (No resist)> Acrylic Suppose the threshold level of the odor sensor is 300m
When set to V, these materials are smoked and 0.1% / m
The odor sensor will be able to alert until the smoke density is reached. However, when the vinyl chloride and the ABS resin were heated, the odor sensor felt little. The sensor 1 having a film thickness of 1000 ° has higher sensitivity to the smell of acrylic and cotton than the sensor 2 having a large film thickness.

(D) Output trajectories of two different sensors with respect to odor The output of the thin-film sensor 1 with a thickness of 1000 ° of SnO ₂ is graduated on the horizontal axis, and the output of the thick-film sensor 2 with a thickness of 2500 ° on the vertical axis. Figure 7 shows the scale. It can be seen from FIG. 7 that the odor of various materials when heated and smoked and the odor of an organic solvent such as alcohol and thinner which are often present in the normal environment can be clearly distinguished. That is, in almost all the materials, the sensor trajectory for the odor generated when the material was smoked was concentrated in the same region (the thick film sensor 2 side). On the other hand, the sensor 1 of the thin film type showed a fast response to the odor of ethyl alcohol, IPA, cinna and perfume, and thereafter the sensor having a large thickness responded slowly. In particular, among organic solvents, the lower the boiling point of the material, the greater the tendency. The sensor locus of acrylic for smell showed a behavior close to that of the alcohol group, but it was considered that acrylonitrilo with a low boiling point was generated by thermal decomposition, and this was sensitive. The response trajectories of coffee, fragrance, and cigarette butts to the odor were located between the two groups. Thus, FIG. 7 indicates that the odor generated in the event of a fire can be distinguished from the odor such as alcohol often appearing in a normal environment.

From the above experimental results, it was found that response trajectories to odors drawn by two odor sensors having different film thicknesses can be grouped according to odor types as shown in FIG. Therefore, based on the above experimental results, as one embodiment of the present invention, a neural network which is good at pattern matching analog data with the above-mentioned two SnO ₂ odor sensors 1 and 2 having different film thicknesses will be described below. A system for determining the type of odor configured using the above will be described.

FIG. 8 is a diagram schematically showing a block circuit configuration of a signal processing unit PR which may be constituted by a microcomputer or the like as a system for determining the type of odor. Is connected to the _two SnO ₂ thin film odor sensors S ₁ and S ₂ having different film thicknesses as described above. Memory RAM1 and RAM2, etc. are connected (or their may be a built-in memory) odor identifying unit OD performs the identification of the odor receiving analog sensor signal from the odor sensor S ₁ and S ₂ The identification result is output to the determination unit DE, and the determination unit DE performs an appropriate comparison determination based on the identification result. For example, when the signal processing unit PR is used as a fire detection device or a fire alarm device that makes a fire determination, the determination unit DE focuses only on the burning odor from the signal output from the odor discriminating unit OD, and detects the burning odor. When the intensity is equal to or higher than a certain reference value, a signal for judging a fire can be output.

For example, when the signal processing unit PR is used as an odor discriminating device or the like for discriminating the quality of odor, the odor discriminating unit OD performs necessary discrimination judgment processing based on FIGS. Based on the results, the determination unit DE performs a discrimination between a comfortable odor and an unpleasant odor, a discrimination between an exciting odor and a mild odor, a discrimination between an exciting odor and a calming odor, and outputs a signal regarding the type. Is also possible.

Further, the type and the number of sensors connected to the signal processing unit PR are appropriately changed to those suitable for the odor to be identified, and the odor identification unit OD performs an odor determination process to make the determination unit DE. May output a signal relating to the type of odor.

The odor discrimination unit OD may be associated with a memory storing a table representing the experimental results described with reference to FIG. 7 in order to make a judgment on the type of odor. Then, a neural network learning and memorizing the contents of FIG. 7 is used. The so-called neural network of the back propagation method is described in detail in, for example, Japanese Patent Application Laid-Open No. 2-105299 filed by the present applicant, but the neural network used here has a similar effect. In the present embodiment, as a learning content to be memorized in the neural network, based on the graph of FIG. 7, a definition table as shown in FIG. 9 that grasps the characteristics of the graph is created. As can be seen from FIG. 9, this definition table is
Twenty-six patterns of combinations of the input values of the _two sensors S ₁ and S ₂ are selected from the graph of FIG. 7, and two definition values (ie, the definition value D ₁ of organic odor and the burning smell) for each combination pattern are selected. the defined values D ₂₎ of the relationship is used to represent extracted from FIG. When the content of this definition table is learned and learned by the neural network, the neural network recognizes input values that are not in the definition table, that is, combinations of inputs of the _two sensors S ₁ and S ₂ other than the 26 patterns. Even when input, an accurate calculation value according to the contents of the definition table can be output. The calculated values C ₁ and C ₂ are also shown in FIG.

As an example, the definition table shown in FIG. 9 shows that, for a point on an arc line having the same distance from the origin in FIG. 7, the smaller the angle between the line drawn from the origin to that point and the vertical axis, the smaller the angle. Make sure that the accuracy of the burning smell is large, that the smaller the angle between the line drawn from the origin to that point and the horizontal axis is, the more accurate the odor of the organic solvent is, and that the line drawn from the origin and the vertical For a point at which the angle formed by the axis is constant, it can be set so that the greater the distance from the origin to the point, the greater the accuracy of the burning smell and the odor of the organic solvent. . In addition to such a principle, it may be made to take into account human experience. Further, the distance from the origin and the vertical length are set not for the points apart from the origin as described above but for each of the 200 mV squares shown in FIG. 7 (the smaller the squares, the higher the accuracy). A definition value of accuracy may be given according to the angle deviation from the axis and the horizontal axis.

The neural network in the present embodiment converts analog data from the _two sensors S ₁ and S ₂ , as conceptually shown in the odor discriminating section OD in the signal processing section PR in FIG. The two input layers INi (i
= 1, 2), five intermediate layers IMj (j = 1 to 5),
Outputs calculated values of organic odor and burnt odor, respectively 2
And ten output layers OTk (k = 1, 2). Ten lines connecting each input layer and each intermediate layer and ten lines connecting each intermediate layer and each output layer include: Each of them depends on the degree of contribution to the output value in the output layers OT ₁ and OT _2.
Weight values ω _ij and ν _jk are given. It should be noted that the net structure shown in the odor discriminating unit OD in FIG. 8 is conceptually shown for easy understanding in explaining the operation process in the odor discriminating unit OD. Is not a division that actually exists in the odor identification unit OD.

In the definition table of FIG. 9, the column of "input" indicates that the first sensor S having a thickness of SnO ₂ 1000
₁ (thin sensor) and the analog sensor levels from the _second sensor S ₂ (thick sensor) with a thickness of 2500 mm SnO ₂ are shown converted to values of 0 to 1 (zero of the sensor output). １０００1000 mV corresponds to 0-1), and the four columns on the right include the defined value D ₁ and the calculated value C _{1 of} the organic odor according to the sensor levels from the two sensors, and the defined value D of the burning odor. ₂ and calculated value C ₂ are shown respectively converted to values of 0-1.

In learning the contents of the definition table of FIG. 9, first, the relationship between the sensor input and the definition values of organic odor and burning odor in the definition table is stored in the definition table memory RAM 1 of FIG. Next, FIG.
The first and second sensor inputs of the definition table of FIG. 9 are given to the input layers IN ₁ and IN ₂ , respectively, and the output from the output layers OT ₁ and OT ₂ is performed based on those inputs. Is compared with the defined value D ₁ of organic odor and the defined value D _{2 of} burnt odor in FIG. 9, respectively, and the weighting value given to each connection line so that the error of each comparison result is minimized. change ω _ij and ν _jk ,
Finally, these weighting values for which the error of the comparison result is minimum or zero are stored in the weighting value table memory RAM2. In this way, it is possible to instruct the net structure of FIG. 8 a very close approximation to the entire function of the definition table of FIG. 9, which is shown only at 26 points. When the contents of the definition table are taught in the net structure in this way, the input of the sensors S ₁ and S ₂ is thereafter operated by using the weighting values stored in the memory RAM 2, thereby obtaining an organic information. the calculated values C ₁ and calculated value C ₂ of the burnt odor of odor can be obtained accurately.

Now, a weight ω _ij between the input layer INi and the intermediate layer IMj and a weight ν _jk between the intermediate layer IMj and the output layer OTk (i = 1, 2, j = 1 to 5) , K = 1, 2)
Have positive, zero, and negative values, respectively. If the input value in the input layer INi is represented by INi, the intermediate layer I
The total sum NET ₁ (j) of inputs to Mj is expressed by the following equation 1.

[0029]

(Equation 1)

This value NET ₁ (j) is converted to, for example, sigmoid (s
igmoid) function to convert the value to 0-1 and convert it to IMj
Is expressed by the following equation (2).

[0031]

(Equation 2)

Similarly, the total sum NET ₂ (k) of inputs to the output layer OTk is expressed by the following equation (3).

[0033]

(Equation 3)

If this value NET ₂ (k) is converted to a value of 0 to 1 by the same sigmoid function and is represented by OTk, it is expressed by the following equation (4).

[0035]

(Equation 4)

As described above, the relationship between the input values IN ₁ and IN ₂ and the output values OT ₁ and OT ₂ in the net structure shown in FIG. 8 is expressed by Equations 1 to 4 using weighting values. Here, γ ₁ and γ ₂ are adjustment coefficients of the sigmoid curve, and in this embodiment, γ ₁ = 1.0 and γ = 1.2 are appropriately selected.

In the net structure creation program shown in FIG. 10, first, when one of the 26 combinations of sensor inputs in the definition table of FIG. 9 stored in the RAM 1 is given to the input layer, Of the actual outputs OT ₁ , O calculated from Equations 1 to 4 and output from the output layer
Below T ₂ is compared respectively with the defined values D ₂ defined values D ₁ and burnt smell of organic odor of FIG. 9, the sum Em of each error at each output layer at that time (m = 1 to 26) Equation 5

[0038]

(Equation 5)

Here, OTk is a value obtained by the above equation (4). The value E obtained by summing the sum of errors Em for all 26 combinations in the table of FIG.

[0040]

(Equation 6)

Finally, an operation is performed to adjust the weights of the signal lines one by one so that the value E in Equation 6 becomes zero or minimum. Then, the weight values stored in the memory RAM 2 are updated with the adjusted new weight values, and are used when odor determination is performed.

When the education of the definition table of FIG. 9 for the net structure conceptually shown in FIG. 8 is completed, that is, the adjustment of the weight value of each one is completed, and the weight value table memory RAM 2 is suitable. Once the weight is stored,
At the time of actual odor discrimination, the memory RAM is executed by a net structure calculation program shown in the flowchart of FIG.
It is possible to determine the odor in the environment using the weight value stored in 2. That is, the input value from each sensor section is given to the net structure, and the value obtained from each output layer is obtained by calculation using the above-described equations 1 to 4, thereby calculating the organic odor calculation value C ₁ and the burning odor. calculated C ₂ are obtained, respectively.

As described above, in the present embodiment, the odor discriminating unit O
D shows a case where odors are not classified into individual odors, but are classified into two groups, namely, organic odors and burnt odors. Based on the output from the odor discriminating unit OD, the determination unit DE can determine whether the environment is good or bad for human health, for example, by focusing only on the odor of the organic solvent, or based on only the burning odor. It is also possible to make a fire discrimination. Further, the type of grouping is not limited to the above-described embodiment, and for example, explosive gas may be used as the type of grouping using another odor sensor.

FIGS. 12 and 13 show the accuracy output when various combinations of sensor input patterns are given using the neural network for which the education of the contents of the definition table in FIG. 9 has been completed. The axis represents the value obtained by converting the output of the thick film sensor S2 from 0 to 1, the Y axis represents the value obtained by converting the output of the thin film sensor S1 from 0 to 1, and the Z axis represents:
In the case of FIG. 12, values of 0 to 1 of the degree of accuracy of the burning smell are shown, and in the case of FIG. 13, values of 0 to 1 of the degree of accuracy of the odor of the organic solvent are shown. The mesh in FIG. 12 and FIG.
The degree of accuracy calculated by a neural network for an arbitrary sensor input is shown, and a line or a dotted line on a mesh represents a burning smell by a specific substance or an odor of a specific organic solvent.

FIG. 14 shows the maximum accuracy value with respect to the odor of each substance, that is, the maximum value on the Z axis in FIGS. 12 and 13 in%. Numbers 1-1
1 is the maximum accuracy value extracted from FIG. 12 in the case where odor is generated by heating and smoking at a rate of 10 ° C./min. On the other hand, it can be seen that the accuracy is very high except for acrylic. In the case of odors of coffee and aromatics, the accuracy of odors at the time of smoking is only slightly increased. Smell of cigarette butts shows moderate accuracy. On the other hand, IPA, ethyl alcohol, acetone, thinner,
The perfume group does not show much accuracy, with the highest accuracy being about 15%.

Similarly, when the maximum accuracy value of the odor of the organic solvent, which is an environmental odor, is extracted from FIG. 13, the accuracy of the odor of the organic solvent is exceptionally exceptionally high although the odor of the acrylic solvent is a burnt odor. High accuracy was recorded, up to 90%, but the burn odor accuracy of the other materials numbered 1-11 was zero. With the materials of Nos. 14 to 26, the accuracy of the organic solvent as the odor was high, and it was found that the difference in the quality of the odor could be distinguished from the accuracy of the odor at the time of smoking and the accuracy of the odor often occurring in the environment.

Further, according to the present embodiment using two odor sensors and neural nets having different SnO ₂ film thicknesses, the sensitivity to various odors is summarized as follows. (A) High sensitivity to burning smell of cellulosic and glass epoxy substrate materials. (B) The sensitivity varies greatly depending on the thickness of the sensor. A sensor with a thick film thickness has higher sensitivity to a burning smell, but a sensor with a thin film thickness has higher sensitivity to an odor of an organic solvent. (C) The lower the heater temperature, the higher the sensitivity to the burning smell. However, the opposite tendency is observed for the burning smell of acrylic.

In the above embodiment, the case where two odor sensors having different film thicknesses are used has been described. However, if a neural network is formed by using a larger number of odor sensors,
A subtle difference in odor in the environment can be more clearly identified, and the determination unit DE can appropriately make a necessary comparison determination according to the application.

[0049]

As described above, according to the present invention, based on the sensor output values from a plurality of different types of odor sensors, for example, the odor discriminating section preferably constituted by a neural network is used for the type of odor or the odor. The grouped types are identified and the judgment unit is made to judge meaningful odors, so even though the configuration is simple, the selectivity of odors can be increased, and a wide range of types of odors can be accurately identified. There is an effect that can be.

[Brief description of the drawings]

FIG. 1 is a diagram showing a measurement circuit for measuring an odor in an environment by an odor sensor used in an embodiment of the present invention.

FIG. 2 is a diagram showing a method for measuring odor sensitivity for conducting an experiment based on the present invention.

FIG. 3 is a diagram showing a table of a relationship between a material and sensitivity as an experimental result.

FIG. 4 is a diagram showing a table of sensitivity to sensor temperature as an experimental result.

FIG. 5 is a diagram showing a table of sensitivity to odors generated in a normal environment as experimental results.

FIG. 6 is a diagram showing a table of sensor sensitivity at a smoke density of 0.1% / m as an experimental result.

FIG. 7 is a diagram showing a graph for comprehensively determining responses of two sensors as an experimental result.

FIG. 8 is a block circuit diagram showing an odor discriminating apparatus according to an embodiment of the present invention.

9 is a diagram showing a definition table used in the odor identification device of FIG.

FIG. 10 is a flowchart showing a net structure creation program used in the odor identification device of FIG. 8;

11 is a flowchart showing a net structure calculation program used in the odor identification device of FIG.

FIG. 12 is a diagram showing the accuracy of burnt odor that can be determined by the odor discriminating apparatus of FIG. 8;

FIG. 13 is a diagram showing the accuracy of the odor of the organic solvent which can be determined by the odor discriminating apparatus of FIG. 8;

FIG. 14 is a table showing the maximum values of the accuracy of the burning smell and the accuracy of the odor of the organic solvent in FIGS. 12 and 13.

[Explanation of symbols]

 S1 Thin film type first sensor S2 Thick film type second sensor PR signal processing unit OD odor identification unit RAM1 memory for definition table RAM2 memory for weight value table DE determination unit

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G01N 27/12

Claims (3)

(57) [Claims] 1. A method of identifying a plurality of types of different odor sensors and sensor output values from these odor sensors,
An odor discriminating unit that performs grouping of odors in the environment and calculates an odor level for each of the classified odor types; and an odor level for each of the classified odor types from the odor discriminating unit. A determination unit for determining a significant odor based on the calculation result , wherein the plurality of types of different odor sensors have different thicknesses.
Sensor with a metal oxide film of the same material
Odor identifying apparatus characterized by some. 2. The type of smell classified as said is burnt.
Claims: Two types of odors: odor and organic solvent odor
1 odor identification device. 3. The neural network according to claim 1, wherein the odor discriminating unit is a neural network.
3. The odor identification device according to claim 1, further comprising: