JPH04186139A - Smell identifying device - Google Patents

Abstract

PURPOSE:To allow the management of smell quality in place of a man by providing a switching means for preliminarily passing a cleaning gas to a gas sensor and a flow system and an arithmetic means for holding a reference pattern and determining the difference with the pattern output of the gas sensor. CONSTITUTION:Prior to setting each sample in a sample vapor generating part 53, sets of solenoid valves (55a, 56a)... (55e, 56e) are first successively opened plural times to conduct the cleaning of a flow system for removing a smell material adhered to each solenoid valve. After each sample is set in the generating part 53, a standard air is sent to a 5th sample, for example, to recover a quartz oscillator sensor array 10. Then, a gas containing the sample vapor is passed through the sensor array 10. The output pattern of the sensor array 10 is detected by a reply pattern detecting part. This detected pattern is held in a memory as a reference signal. The difference signal between the reply pattern similarly detected to a target sample and the reference pattern is determined and inputted to an neural network to identify the smell.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an odor identification device that achieves a high odor identification function.

"Prior Art" Many gas sensors have been proposed in the past, and some of them are currently in practical use and widely used.

However, although these sensors have excellent sensitivity stability, it has been difficult to realize a sensor with good selectivity. It is said that living organisms identify odors by recognizing the response patterns of a large number of olfactory cells. Therefore, recently, a type of odor detection force sensor that imitates the biological sense of smell and recognizes multiple gas sensor response patterns has been proposed in order to improve selectivity. One example of this is the prior art proposed in Japanese Patent Application Laid-Open No. 1-244335.

Because biological olfaction is non-specific due to the characteristics of the receptors, it is said that the olfactory nervous system is able to identify odors by recognizing the odor stimulus response patterns of a large number of olfactory cells. Therefore, in the conventional technology, as shown in FIG. 6, a crystal oscillator sensor array 10 having various types of adsorption membranes to which odor molecules are adsorbed instead of olfactory cells, and a neural network 41 instead of the intracerebral olfactory nervous system are used. It is used to identify odors. In other words, when the odor-containing residue from the sample 53 is passed through the sensor array 10, the odorants are adsorbed by the adsorption film stretched over the surfaces of the crystal oscillator sensors Oa, IOb, and 10c of the sensor array 10. Then, the surface load mass of the crystal resonator is changed, and the oscillation frequency of each of the oscillation circuits 2a, 2b, and 2C is changed.

The adsorption films of each of the crystal oscillator sensors 10a, 10b, and 10c are of different types with different adsorption characteristics, and respond differently depending on the type of odor (frequency change Δf2.
△f7. Δf3). This response pattern is recognized by a neural network 41 to identify the odor. In addition, the neural network [4] is equipped with a learning function, and its odor recognition rate can be improved by accumulating experience.

[Problem to be solved by the invention] Currently, in the fields of food and beverages, cosmetics, environmental products, etc., sensory tests by experts (springs) are used to perform odor quality tests, product development, and abnormality detection. ing. However, there are individual differences among humans, and their physical condition may differ from day to day, so there is a need for a device that can objectively identify odors.

However, the odor detection sensors in the conventional technology described above do not have a sufficient odor identification rate to replace the odor quality control (sensory test) using the human sense of smell, which is required not to identify similar odors. Therefore, it is currently impossible to meet the above requirements, and the problem to be solved is to improve the identification rate.

- The present invention is proposed to solve the above problems, and its purpose is to improve odor quality control (
An object of the present invention is to provide an odor identification device that has a high odor identification rate that can perform sensory tests.

[Means for Solving the Problems] The structure of the odor identification device of the present invention for achieving the object described in item 1 is as follows: Sample vapor from a sample vapor supply system is passed through a plurality of gas sensors, and pattern outputs of those gas sensors are detected. In an odor identification device that identifies odors by pattern recognition,
a switching means for not passing a cleaning gas through the gas sensor and the flow system before passing the sample vapor; a calculation means for holding a reference pattern and calculating a difference between the pattern output of the gas sensor; It is characterized by recognition.

Similarly, other configurations of the present invention include, in addition to the above configuration, providing a constant temperature means for maintaining the ambient temperature of the gas sensor higher than that of the sample vapor supply system, or when the sample is a liquid, the sample vapor supply system is characterized in that a standard gas is blown onto the sample to generate sample vapor.

1. Effect: The present invention improves the identification rate by pattern recognition by emphasizing the difference between similar sample vapors by calculating the difference between the pattern outputs of a plurality of gas sensors and a reference signal. Also, by providing a step of cleaning the gas sensor and flow system before passing the sample vapor through the gas sensor, the identification rate is improved by eliminating the influence between samples. Furthermore, the circumference 12 of the gas sensor
Maintaining Bf higher than the sample vapor supply system to prevent condensation of the sample vapor and ensure accurate sample measurements or, if the sample is a liquid, by blowing standard vapor over the sample. The identification rate is improved by ensuring the homogeneity of the sample vapor supply system by generating sample vapor at different times.

[Example] Hereinafter, an example of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention. In the configuration of this embodiment, 1 is a sensor cell that accommodates a sensor array consisting of eight crystal oscillator sensors 10a, 10b, -, 10h having various types of adsorption films and measures odor;
is a Colpitts-type oscillator circuit with 8 independent circuits, 3 is a 8-circuit
4 is a computer constituting a neural network; 5 is a sample steam supply system; 6 is a gas discharge system; and 7 is a water circulation pump for supplying circulating water for constant temperature to the sensor cell 1.

FIGS. 2(a) and 2(b) are a perspective view (a) and a side view (b) showing the structure of the sensor cell 1, including an exploded part. Sensor cell 1 is made of metal such as stainless steel.
The two portions 11a, l], b, and IIC have flat surfaces, and a fluorine resin plate 11d has a middle chamber.

By sealing and sandwiching the lie, a path for debris is created as shown by the arrow in (b), and a crystal oscillator sensor array 10 is configured in the middle part of each fluorocarbon resin plate 1], d, and lle. Crystal oscillator sensors 10a, 10b, . . . have various types of adsorption films.

] Oh is divided into two parts (only IOa to IOa to 10d are shown) and installed. Of the above three parts, the upper part 11a and the lower part 11c are made of stainless steel plates, and the middle part 1 ],
b is provided with an inlet passage 12, an outlet passage 3, and a vent 14 for sample vapor and cleaning gas (hereinafter collectively referred to as gas).

The inlet passage 12 is opened at one end of the middle chamber portion of one fluoro-resin plate 11d, and the outlet passage 13 is opened at one end of the middle chamber portion of one fluoro-resin plate 11d.
The vent hole 14 is opened at one end of the middle chamber portion of the ie, and communicates between the other ends of each middle chamber portion.

In addition, there are pipes 11f on the upper and lower parts 11a and Ilc.

11g and a water reservoir 11h are provided in the middle part 1.1b, and circulating water for constant temperature is circulated from the water circulation pump 7 shown in FIG.

FIG. 3 is an explanatory diagram including a cross-sectional view showing the structure of the above-mentioned two crystal oscillator sensors. 101 is a crystal oscillator, 1.0
2, 103 are electrodes that sandwich this crystal resonator 101; 104;
, 105 is an odorant adsorption film. In this example,
AT-cut crystal resonator 101 (for example, 10.1/
A sensor is fabricated by coating the IMH electrode 102, ], 03J2 with a stationary phase material used in gas chromatography, a cellulose material, a lipid bilayer membrane, or the like. If an adsorption film is applied in this way, the elastic loss of the crystal resonator 101 will increase, and if too much is applied, oscillation will stop. Therefore, the above-mentioned coating is performed while keeping in mind that the Q value does not become smaller than 5400 using an impedance analyzer, and also taking into account the change in the resonance frequency. Each electrode 102
, 103 are connected to each oscillation circuit of the Colpitts type oscillation circuit 2.

Returning to the configuration shown in FIG. 1, the eight crystal oscillator sensors 10a, lOb, -, 10h of the sensor cell 1 are each connected to each circuit of the Colpitts type oscillation circuit 2. The oscillation outputs of each of these circuits are input to each channel of the 8-channel frequency counter 3, respectively. This frequency counter 3 measures eight crystal oscillator sensors simultaneously in parallel with frequency changes from Oa to 10h, and samples at intervals of, for example, 1 second. On the computer 4 side, this measurement data is read from the I10 port via the interface board 1. Furthermore, automatic measurement is made possible by controlling the valves (electromagnetic valves) of the sample vapor supply system 5, which will be described later, by software through the valve control circuit via the interface port.

Next, the components of the sample steam supply system 5 are as follows:
51 is a standard air cylinder that generates sample vapor carrier scum and cleaning gas, 51a is a mass flow controller installed after 51 to the standard air horn, 52 is a carrier scum distributor, 53a, 53b, . . , 53e
(denoted as 53 when representing) is a sample steam generation section, 54 is a constant temperature chamber that accommodates each sample steam generation section, 5
5a, 55b, -, 55e and 56a, 56b, -,
56e is a solenoid valve for selecting one sample vapor;
57 is a distributor for sending selected sample vapor into the sensor cell 1, 58 is an electromagnetic valve for sending standard air as a cleaning gas into the sensor cell, and 59 is a control signal line for each electromagnetic valve.

The gas outlet of the distributor 57 is connected to the gas port 12 of the sensor cell 1 .

Conventionally, dry air was obtained by passing it through a popular carrier gas, but in this example, standard air was used as the carrier gas, and data reliability was improved by always supplying the same carrier gas. At the same time, the pressure of the standard air cylinder 51 is set to flow drive. Further, in this embodiment, under the control of the computer 4, before sending the sample vapor into the sensor cell 1, the electromagnetic valve 58 is opened to send standard air as it is into the sensor cell 1, thereby cleaning the adsorption films of each sensor 10a to loh. After this, sample vapor is supplied to the sensor cell 1. For example, when measuring sample 3, solenoid valve 5
8 and close the solenoid valve 55c.

Only 56c is opened, and standard air as a carrier gas is introduced into the sample vapor generating section 53c, and then into the sensor cell 1 through the distributor 57. Here, the mass flow controller 51a has a flow rate control function so that a preset constant flow rate of standard air is supplied to the sample steam generating section 53c even if there is a load change, etc., and improves reproducibility. Contribute to

FIG. 4(a) is a configuration diagram of the two sample steam generating sections 53. For comparison, FIG. 4(b) shows
1 is a configuration diagram showing a conventional sample vapor generation method. In each figure, 531 is a sample container, 53
2 is a carrier gas injection nozzle, 533 is an outlet nozzle for gas containing sample vapor, 534 is a sample liquid, and 535
fixes each nozzle 532, 533 and connects the sample container 531.
For example, a Hyton plug is used to make it airtight. The conventional sample vapor generation method is the bubble method, in which a carrier gas injection nozzle 532 is inserted into the sample liquid 534 to generate bubbles, as shown in (b), or the bubble method in which the carrier gas is generated in the sample liquid 534, as shown in FIG. It has a structure in which sample vapor is generated by spraying it onto the surface of the liquid 534 (non-bubble method).

Next, the configuration of the gas exhaust system 6 shown in FIG. 1 will be explained. In its configuration, 61 is Hasofa, 62 is Trap, 63
is a flow meter. The buffer 61 has a structure in which a hard tube is passed through a sample tube, and is inserted between the outlet passage 13 of the sensor cell 1 and the trap 62 to prevent backflow of water, etc. from the trap 62. The trap 62 consists of a container containing water or a solvent, and the gas from the trap 62 is passed through the water or the like to cause the sample vapor vented to the sensor cell 1 to bubble, thereby dissolving odorous substances and preventing the smell from flowing to the outside. do. The flow meter 63 has a sensor cell l
A volumetric flowmeter is preferred.

The volumetric flow meter has its outlet open to the atmosphere (or no load)
If this condition is not met, the correct flow rate cannot be measured, so the trap 62 is provided at the outlet of the trap 62. The flow rate adjustment based on this measurement is performed by adjusting the pressure reducing valve of the standard air cylinder 51. In this way, the exhaust gas from the sensor cell 1 is released into the atmosphere from the flow meter 63.

Next, the configuration of the neural network of this example will be explained in the fifth section.
As shown in the figure. In this embodiment, a layered connected neural network 4 built into the computer 4 as a calculation specification is used.
1 is used. This layered connected neural network 4
1 has eight input layer units 41a and seven intermediate layers/
+lb and five output layer units l-4]c. The number of units in the input layer unit l-41a is equal to the number of sensors, and the number of units in the output layer unit 4.1b is equal to the number of samples. This neural network 4
1, 42 is a response pattern detection unit that calculates the response pattern of the sensor cell 1 from the output of the frequency counter 3;
numeral 44 is a memory for holding the reference signal; numeral 45 is a controller for controlling each electromagnetic valve of the sample steam supply system 5 through the control signal line 59 shown in FIG. , a valve control unit that automatically measures odor. As the reference signal pattern, a response pattern of one sample can be used, and the difference signal of the calculation unit 43 is
is input to each input layer unit y t·4 ]a.

The operation and effect of the embodiment configured as above will be described.

First, the operating principle of the crystal resonator sensor will be explained using FIG. If odor molecules exist in the atmosphere, the adsorption film 1
04,105 is adsorbed. Here, if the crystal resonator is an AT-cut crystal resonator with a fundamental resonant frequency f, the oscillation frequency change Δ[ when a substance is adsorbed to the crystal resonator with a weight of 6 M is derived by 5auerbrey. There is.

A Electrode area (cm") From formula (1), it can be seen that when a substance is adsorbed on the surface of the crystal resonator 101, the oscillation frequency changes in proportion to the mass of the adsorbed substance. This mass loading effect causes oscillation. A decrease in the oscillation frequency occurs in the circuit 2. This frequency change is taken as the sensor output.

In this example, after the odor measurement, a desorption step of the odor substance is provided, and standard air is passed through the sensor, so that the adsorption membrane 1
04. ．． 105 odor molecules are desorbed and returned to the original oscillation frequency.

Next, a method of odor pattern recognition will be explained using FIG. The initial value of the connection weight coefficient between each layer of the neural network 41 is the value after setting five samples in advance, taking data for 10 times, and performing backpropagation learning (initial learning) 20,000 times. Use. and,
In order to adapt to data fluctuations due to changes in the crystal resonator sensor array 10 over time, the weighting coefficients are changed by performing backpropagation learning (adaptive learning) 500 times every time five samples are measured. As an evaluation of odor discrimination, the number of times the odor was correctly identified in 10 sets of discrimination was calculated based on the total number of trials of 50.
Identification by dividing by (=5 samples XIO measurements);
Seek ne snow. The reason why neural networks are suitable for this example is that they are more flexible in identification compared to existing pattern recognition algorithms, and can be expected to improve the recognition rate. This is because it is easy to change the smell, and because it mimics the information processing in the brain, there is a possibility that it will be able to judge smells in the same way as humans.

However, for a substance like the one used here, where the main components are water and alcohol, and only a small amount of the remaining components are used as samples, there is no difference in the output pattern of the sensor array 10 between the samples. , it is difficult to identify patterns. Therefore, in this embodiment, the sensor array output cover pattern of one sample is used as a reference signal pattern, and the difference between the output cover pattern of the sensor array 10 and its reference signal is taken to emphasize the subtle differences between the samples. Signal-
By using it for pattern recognition described in 1-, the identification rate is improved.

Next, in Section 1, the measurement procedure in the learned example will be described using numbers as an example. First, before loading each sample into the sample steam generating section 53, set the solenoid valves (55a, 5
6a), (55b, 56b).

(55c, 56c), (55d, 56d), (
55e, 56e) in the same amount in sequence to clean the flow system, such as removing odor substances attached to each electromagnetic valve.

Subsequently, after seven samples are poured into the sample vapor generating section 53, for example, for the fifth sample, (1) standard air is flowed for, for example, 60 seconds to recover the crystal resonator sensor array 10;

(2) Pass the dregs containing sample vapor through the crystal resonator sensor array 10 for 30 seconds. .

(3) The response pattern detection unit 42 detects the output pattern of the sensor array 10.

(4) The detected pattern is held in the memory 44 as a reference signal pattern.

(5) Next, apply (1) above to the first sample.

And perform (2) and (3).

(6) The response pattern detected in (3) above and
) is obtained, and the odor identification is performed by manually inputting it into the neural network/II.

(7) Below, from the second sample to the fifth sample -
Similarly to (5) and (6), -1- (1).

and repeat (2), (3), and (6).

(8) In the evaluation, 10 sesotators are taken with the following five types of samples as one set, and the identification rate is determined as described in the pattern recognition method described above.

In the following, improvements in the measurement system for increasing the identification rate in this embodiment will be described.

In this embodiment, the sample steam generation section 53 of the sample steam supply system 5 described above is placed in a constant temperature bath 54, and constant temperature water (variation within ±0.degree. 5.degree. C.) is circulated through the sensor cell 1 from the constant temperature bath. As a result, the temperature distribution is set as follows in the sample vapor generating section 53, the intermediate route, and the sensor cell 1.

Steam generation part ≦ Temperature of intermediate passage ≦ Temperature of sensor cell part
(Outside temperature) Temperature (approx. 16°C) (approx. 18°C) (approx. 20°
C) This prevents condensation of the sample vapor, reduces the temperature influence of the adsorption phenomenon on the adsorption films 104 and 105 of the sensor array IO, and reduces the coefficient of variation of the measurement (the standard deviation of the sensor output is the average of the sensor outputs). Keep the divided value) small. As a result of the experiment, the coefficient of variation in the conventional technology without the above measures was 2 to 3%, whereas in this example, the coefficient of variation was 1%.
It can be about %.

Next, the problem with the identification rate is the homogeneity of the systems that supply each sample vapor. In the past, the homogeneity of the strains was not sufficiently ensured, and even if samples 1 to 5 were the same, they could be recognized as different samples because they passed through different strains. In the conventional sample vapor generation method, as shown in FIG. 4(b), dry air is introduced into the sample and bubbled, thereby generating sample vapor and sending it to the sensor cell. However, the homogeneity of the system can be achieved by sending the odor vapor evaporated above the liquid surface to the sensor cell 1 (referred to as non-bubble) as shown in FIG. was revealed as a result of principal component analysis and discriminant analysis. Regarding the nozzle in the figure, L7. L3
It is also necessary to make the length of all sample containers 531 the same. This allows measurement accuracy to be maintained and identification rate to be improved.

Furthermore, we investigated the relationship between the time for sending the sample vapor to the sensor cell 1, the stability of data, and the homogeneity of each system. The method was to put the same sample into each system, supply the sample vapor at a flow rate of 25 m1/min for 1 minute, and
0 seconds later, 20 seconds later...

Data was read after 60 seconds and evaluated at what stage it was possible to obtain sufficiently stable data. As a result of multivariate analysis of the results, it was found that data fluctuations were small at the early stage and then increased, and that the homogeneity of the strains was initially poor and then decreased. Both were set to 30 seconds. In this way, by optimizing the sample vapor supply time, it is possible to maintain the homogeneity of the system and contribute to improving the identification rate.

In addition, by performing the flow system cleaning process before sample setting and the sensor cleaning process before measurement in which air flows through the solenoid valve 58 shown in Figure 1, the influence between measurements can be eliminated and identification can be made. It is possible to improve the rate. In addition, in the flow system cleaning process, the solenoid valves 55a to 55e and the solenoid valves 56a to 56e
It is effective to wrap a flexible heater around each solenoid valve to make it easier to desorb odor adsorbed components.

By improving the input processing to the neural network 41 and the measurement system as described below, it is possible to significantly improve the classification rate compared to the conventional technology in the case of samples that have strong similarities and are difficult to identify. can.

Note that even when a lipid bilayer membrane is not used as the adsorption membrane of the sensor, the identification rate can be significantly improved compared to the conventional technology. As described above, the present invention can be applied in various ways and can take various embodiments in accordance with its gist.

[Effects of the Invention] As is clear from the above explanation, according to the odor identification device of the present invention, the odor identification rate can be greatly improved.
It has the advantage of achieving an odor identification rate high enough to replace the odor quality control that is performed by humans.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention, and FIG. 2 (
a) and (b) are the configuration diagram of the sensor cell according to the first embodiment, FIG. 3 is the configuration diagram of the crystal resonator sensor according to the second embodiment, and FIGS. FIG. 5 is a block diagram of a neural network according to the embodiment described above, and FIG. 6 is an explanatory diagram of an odor detection sensor in a conventional example. Sensor cell, 2 Colpitts type transmission time K, 3 - Frequency counter, 4 Computer, 5 - Sample steam supply system, 6... Waste discharge system, 7 - Water circulation circuit, 53
...Sample steam generation section, 54 Constant crystal tank, 58 Control valve. (b) Figure

Claims (3)

[Claims] (1) In an odor identification device that passes sample vapor from a sample vapor supply system through a plurality of gas sensors and recognizes the pattern output of those gas sensors to identify odors, the sample vapor is passed through the gas sensor and the flow system before passing the sample vapor. An odor identification device comprising: switching means for passing a cleaning gas; arithmetic means for holding a reference pattern and determining a difference between the pattern output of the gas sensor; and performing the pattern recognition from a signal of the determined difference. (2) The odor identification device according to claim 1, further comprising constant temperature means for maintaining the ambient temperature of the gas sensor higher than that of the sample vapor supply system. (3) The odor identification device according to claim 1 or 2, wherein when the sample is a liquid, the sample vapor supply system sprays a standard gas onto the sample to generate sample vapor.