JP5403621B2 - Odor identification method - Google Patents

Description

  The present invention relates to a method for identifying an odor substance using a plurality of odor sensors.

In the fields of food, cosmetics, and fragrance manufacturing, the odor of a substance is detected and identified. Conventionally, a sensory test using a human nose has been performed for odor identification. However, the sensory test has individual differences, and there are many problems such as physical condition and the influence of fatigue on the nose.
For this reason, a detection method using an odor detection chemical sensor for objectively evaluating the odor has been put into practical use instead of the sensory test (see Patent Document 1).

  According to the above-mentioned document, a neural group that uses a sensor group composed of a plurality of crystal resonators each formed with a different polymer film layer, has a learning function, and performs pattern identification of an output signal of the sensor group. Measurements are made using an odor detection system equipped with a circuit (for example, backpropagation (BP) method), and the consistency of the output pattern data obtained for various known substances is observed to determine the odor of unknown substances. I try to identify.

  In addition, a plurality of detection means (sensors) are used for odor detection, and the concentration of the odor component in the measurement gas introduced into these detection means is set to a predetermined value (for example, 1 time, 2 times, 4 times). Are sequentially detected, and an index characterizing the odor component is calculated by using principal component analysis and various approximate expressions for the detection signal to identify the odor ( Patent Document 2).

JP-A-1-244335 Japanese Patent Laid-Open No. 11-352088

In general, in odor identification, a feature amount (feature vector) suitable for odor identification is obtained from measurement data, and pattern identification is performed based on this feature amount.
The measurement system that detects odor components using a semiconductor sensor or a vibrator sensor includes a stationary measurement system that measures the odor sample by confining the odor sample in the cell in which the sensor is placed, and the measurement while the odor sample flows at a constant flow rate in the cell. There is a flow measurement system. Since the static measurement system measures under a constant concentration, measurement with high reproducibility is possible, but it is necessary to set strict measurement conditions, and the measurement time is also long. Therefore, a flow measurement system that can be measured easily and in a short time is practically suitable.

  In this flow measurement system, there is a problem of how to set feature quantities used for odor identification. For example, FIG. 2 is a diagram showing temporal changes in sensor output when a certain coffee odor sample is measured by 14 types of odor sensors, and FIG. 3 is an enlarged view of the rising portion of the output signal. Specifically, it is a time change of the output voltage value when the semiconductor sensor is an odor sensor (when a vibrator sensor is used, a frequency fluctuation value is used instead of the output voltage value). The output voltage value of each sensor before flowing the odor sample is set to 0 (reference value). When the odor sample reaches the sensor (the arrival time is 0 second), an odor component is detected, and the output voltage value increases. Then, after the inflow amount of the odor component reaches a peak, when the odor component starts to decrease, the output voltage value also attenuates. When all the odor component flows out, the output voltage value returns to zero.

When handling measurement data that changes with time so that the output signal forms a peak in this way, the method of determining the feature value is: data at a specific point in time, all sampling data, slope of rising characteristics, maximum change amount (maximum value) It is conceivable to set various values such as the difference between the reference value and the reference value as the feature value, but the result obtained differs depending on which of these values is used as the feature value.
Further, not only how to determine the feature amount but also the same odor substance, when measured at different concentrations, the sensor shows a completely different time change (see FIG. 4 described later). Therefore, when performing odor identification, it is desirable to perform measurement using feature quantities that are not affected by density.

  Accordingly, an object of the present invention is to provide a method for performing odor identification using a feature quantity that is less influenced by the concentration of an odor sample.

In order to solve the above problem, in the present invention, from measurement data measured using a plurality of odor sensors, an invariant that is not easily affected by the concentration of the odor sample is calculated as a feature value, and based on this feature value, Odor identification is performed.
That is, in the odor identification method according to the first invention, a plurality of odor sensors (s) having different characteristics with respect to odors are installed in the container, and N odor samples (d ) having different concentrations with the same type of odor collected. ) was carried out respectively constant flow rate at the odor measurement by the plurality of odor sensors over a predetermined time by supplying to the vessel, wherein for N odor samples, detection of the concentration obtained by the plurality of odor sensor a smell identification method for identifying an odor of odor samples based on signal group, the maximum is the difference between the maximum signal value after flowing the previous reference signal value and odor samples the flow of N odor samples respectively A maximum value vector calculating step of calculating a displacement amount and calculating a maximum value vector (ΔV ^max _{d, s} ) based on the maximum displacement amount of each detection signal, and the maximum value vector (for each odor sample) ΔV ^max _{d, s} ) norm ‖ΔV ^max _d ‖ and the average value <‖ΔV ^max _d ‖> of norm ‖ΔV ^max _d Ｎ for N odor samples with different odors of the same species Then, the density including the density information is normalized by setting the norm of the maximum value vector to 1 and multiplying by the average value according to the equation (1) to obtain the density normalized feature value (ΔV ^reg _{d, s} ). A normalization feature amount calculation step and an identification step for performing an odor identification calculation process based on the acquired density normalization feature amount (ΔV ^reg _{d, s} ).
_{^{ΔV reg d, s = {(}} ΔV max d, s) / ‖ΔV max d ||} · <‖ΔV ^max _d ‖> (1)

Here, ‖ΔV ^max _d ‖ is a norm, and <‖ΔV ^max _d ‖> is a norm average value, which is a value defined by equations (2) and (3).

‖ΔV ^max _d ‖ = √ (Σ _s (ΔV ^max _{d, s} ) ² ) (2)

<‖ΔV ^max _d ‖> = (Σ _d (‖ΔV ^max _d ‖)) / d (3)

The second invention is to install a plurality of odor sensor characteristics to odors differ (s) in the container, the odor concentration of the same species collected N different odor samples (d) is respectively constant flow rate in performs odor measurement by the plurality of odor sensors over a predetermined time by supplying to the vessel, wherein for N smell samples, based on the detection signal group concentration obtained by the plurality of odor sensor a odor odor identifying method for identifying the odor samples, calculates the maximum amount of displacement said is a difference between the maximum signal value after flowing the previous reference signal value and odor sample of N odor sample flow respectively , norm of the maximum value vector calculation step of calculating the maximum value vector (ΔV ^max _{d, s)} based on the maximum displacement amount of the detection signal, the maximum value vector of each odor sample (ΔV ^max _{d, s)} [Delta] V ^max _d ‖, and calculates a norm ‖ΔV ^max _d ‖ average value for the concentration in the same type of smell N different odor samples <‖ΔV ^max _d ‖>, the maximum value by equation (1) A density normalization feature value calculating step of performing density normalization feature value (ΔV ^reg _{d, s} ) by performing normalization including density information by multiplying the average value after setting the norm of the vector to 1; For the density normalized feature value (ΔV ^reg _{d, s} ), the norm ‖ΔV ^reg _s特 徴 of the maximum value vector for each odor sensor is calculated, and the maximum value vector for each odor sensor is set to 1 according to Equation (4). The normalization feature amount calculation process for each sensor which performs normalization including the magnitude information of the response of each sensor to obtain the normalized feature amount for each sensor (C _{d, s} ), and the acquired normalized feature amount for each sensor ( It performs arithmetic processing of odor discrimination based on C _{d, s)} Consisting of a separate step.

C _{d, s} = {(ΔV ^reg _{d, s} ) / ‖ΔV ^reg _s・ ・ ・ (4)

here,
‖ΔV ^reg _s ‖ = √ (Σ _d (ΔV ^reg _{d, s} ) ² ) (5)

According to the odor identification method of the present invention, measurement is performed until the peak of the odor signal is obtained in the odor measurement by the odor sensor group, and the maximum displacement is calculated.
By calculating the maximum value vector based on the maximum displacement of the detection signal of each odor sensor as the feature quantity used for odor identification, setting the norm of the maximum value vector to 1, and then multiplying the norm average value for each odor sample By adopting the density normalized feature that normalizes in consideration of the density, it is possible to reduce the variation in density, and the odor identification calculation process that uses the feature that is less affected by density Can be realized.
Furthermore, instead of the density normalized feature value, the norm ‖ΔV ^reg _sの of the maximum value vector for each odor sensor is calculated for the density normalized feature value (ΔV ^reg _{d, s} ), and the maximum value for each odor sensor is calculated. By adopting a normalized feature value for each sensor that includes normalization including the magnitude information of the response of each sensor by setting the value vector to 1, even a signal from a sensor with a small response is smelled It becomes possible to reflect in identification.

It is a figure which shows an example of the measurement system used with the smell identification method of this invention. It is a figure which shows an example of the measurement data when an odor sample is measured. It is the figure which expanded the rising part of the odor signal in the measurement data of FIG. It is a figure which shows the measurement data when measuring about the different density | concentration in the odor sample of the same kind. It is the figure which arranged the odor vector in the horizontal-axis direction in order with the small density | concentration. It is a figure explaining BP method. It is a figure which shows the result of having identified 10 types of coffee and tea by BP method.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. Note that the present invention is not necessarily limited to only the embodiments described below, and it goes without saying that the object of the present invention is achieved and can be appropriately modified and changed without departing from the spirit of the present invention.

(Measurement system)
A flow measurement system used in the present invention will be described. FIG. 1 is a diagram showing an example of a flow measurement system for collecting measurement data from an odor sample.
The measurement system 10 mainly includes a gas cylinder 11 filled with a carrier gas (dry air, N _2, etc.), a permeator 12, and a sampling box 13 in which 14 different odor sensors S (S1 to S14) are installed. It is configured to be connected to the flow path. In the present embodiment, 14 different semiconductor sensors are used as odor sensors. However, the number of sensors S may be increased or decreased depending on the object to be measured, and a quartz resonator sensor or the like may be used instead of the semiconductor sensor. Good.

The flow path R1 from the gas cylinder 11 branches and is divided into a first flow path R2 connected directly to the sampling box and a second flow path R3 connected to the sampling box via the permeator 12.
A pressure regulating valve 14 and a pressure system 15 are disposed in the flow path R1, and a flow meter 16 and a needle valve 17 are attached to the first flow path R2. A pressure adjusting valve 21, a pressure gauge 22, a needle valve 23, a flow meter 24, and a tube holder 25 for connecting an odor sample are attached to the second flow path R 3, and a sampling box 13 is connected to the downstream side of the tube holder 25. Yes. The tube holder 25 is installed in the constant temperature water tank 26 in order to keep the temperature constant.

In the measurement system 10, the odorous substance is supplied to the sensor S by pushing the odorous sample connected to the tube holder 25 to the sampling box 13 with the carrier gas filled in the gas cylinder 11. The needle valves 17 and 23 are used to select whether the gas supplied to the sensor S is only a carrier gas or an odor sample. When only the carrier gas is allowed to flow, only the needle valve 17 is opened, and the amount of gas injected by the flow meter 16 is determined. When flowing the odor sample gas, only the needle valve 23 is opened, and the amount of gas (concentration of the odor sample) to be injected is determined by the flow meter 24. The signal level in the state where only the carrier gas is supplied is used as a reference value, and the change in the output signal (voltage value or frequency value) from the reference value is used as the sensor response.
By performing flow measurement in this way, measurement can be performed easily and in a short time.

(Measurement data example)
As described above, FIG. 2 is an example of measurement results obtained with 14 types of odor sensors S when the odor of coffee beans is measured using the measurement system 10. FIG. 3 is a diagram showing measurement results when the rising portions of the signals in FIG. 2 are enlarged and the reference points are aligned.
During the period in which no odor gas is detected, the reference value is set by the carrier gas, but when the odor gas reaches the sampling box 13, it rises rapidly. Then, after the signal reaches a peak, it gradually attenuates. This is because the collected odor samples decrease and gradually wither.

From this measurement data, the feature quantity used for odor identification will be extracted. The feature quantities are “data at a certain point in time”, “all collected data”, “slope of rising part”, “maximum value” (Maximum displacement from the reference value) "can be considered.
Among these, when “all collected data” and “slope of the rising portion” are used as feature amounts, they change depending on the concentration of odor and the inflow speed into the sampling box. Even if “data at a certain time point” is used as a feature amount, it changes depending on which time point is selected.
Therefore, the “maximum value (maximum displacement from the reference value)”, which is more robust than the other three data, is used for the subsequent calculations.

However, even when the maximum value is selected as the feature amount, the concentration of the odor sample is strongly affected.
The three measurement data in FIG. 4 are measurement results when measuring the same type of coffee beans and different odors. According to this, when the concentration is different, even if the odor is the same type, the time changes completely differently.
Therefore, it is important to select a feature quantity that does not change in density as a feature quantity for odor identification in order to increase the reliability of odor identification.
In view of this, a feature amount that is less affected by changes in density is separately determined using the calculation described below.

(Feature 1)
The odor of coffee beans is measured a plurality of times (in this case, 35 times) with 14 kinds of odor sensors, and an odor vector (maximum value vector) having a measurement result (ie, maximum value) of each sensor as a component is obtained. In the measurement, the odors of N coffee beans with the same odor and different concentrations are measured.
FIG. 5 is a diagram in which the obtained 35 odor vectors are arranged in the horizontal axis direction in ascending order of the concentration (average value), and the vertical axis represents the output value of each sensor.
From this data, it is predicted that the output signal value of each sensor changes almost exponentially due to the concentration change. Therefore, as a result of logarithm plotting these measurement data, it was found that the measurement data changed almost linearly. This means that by calculating the difference on the logarithmic plot, a constant value independent of the concentration can be obtained.
Therefore, the odor vector (maximum value vector) for each measurement of a certain concentration is normalized so that the norm is 1 and adjustment is performed in consideration of the concentration so as to correspond to calculating the difference in calculation. This calculation was performed and defined as “density normalized feature value (ΔV ^reg _{d, s} )”. Specifically, the calculations represented by the following formulas (1), (2), and (3) are performed. This density normalized feature quantity (ΔV ^reg _{d, s} ) can be used as an input quantity for the later-described odor identification calculation as a feature quantity that is hardly influenced by the density.

The calculation of the density normalized feature value (ΔV ^reg _{d, s} ) will be described.
It is assumed that an odor sample having a concentration d is measured by the sensor S (S1 to S14). The odor vector (maximum value vector) obtained at this time is expressed as a component, and {ΔV ^max _{d, s} }.
The norm ‖ΔV ^max _dの of the odor vector {ΔV ^max _{d, s} } is given by equation (2).
An average value <値 ΔV ^max _d ‖> of N odor vectors (maximum value vectors) having the same odor and different concentrations is given by Expression (3).
At this time, the density normalized feature value (ΔV ^reg _{d, s} ) is calculated using Expression (1).

_{^{ΔV reg d, s = {(}} ΔV max d, s) / ‖ΔV max d ||} · <‖ΔV ^max _d ‖> (1)

here,
‖ΔV ^max _d ‖ = √ (Σ _s (ΔV ^max _{d, s} ) ² ) (2)

<‖ΔV ^max _d ‖> = (Σ _d (‖ΔV ^max _d ‖)) / N (3)

(Feature 2)
By the way, as seen in FIGS. 3 and 4, some of the 14 sensors have a larger or smaller output value than others, and the density normalized feature value (ΔV ^reg _{d, s} ) is used as it is. When used as a quantity, a sensor with a small signal is not sufficiently reflected in identification.
Therefore, since the large feature amount is reduced and the small feature amount is enlarged with respect to the density normalized feature amount (ΔV ^reg _{d, s} ), the norm of the odor vector (maximum value vector) having a different concentration is obtained for each odor sensor. By calculating ΔV ^reg _{s 、 and} setting the odor vector (maximum value vector) for each odor sensor to 1, normalization including the response magnitude information of each sensor is performed, and “normalized feature value for each sensor ( C _{d, s} ) ”. Specifically, the calculations shown in equations (4) and (5) were performed.

This sensor-specific normalized feature value (C _{d, s} ) is a feature value that is not easily influenced by the concentration, and also includes an information from a sensor having a small signal value. Can be used as

The calculation of the normalized feature amount (C _{d, s} ) for each sensor will be described.
It is assumed that an odor sample having a concentration d is measured by the sensor S (S1 to S14). The density normalized feature (ΔV ^reg _{d, s} ) odor vector (maximum value vector) obtained at this time is expressed as a component, and is represented as {ΔV ^max _{d, s} }.

The norm ‖ΔV ^max _dの of the odor vector {ΔV ^max _{d, s} } is given by equation (2).

C _{d, s} = {(ΔV ^reg _{d, s} ) / ‖ΔV ^reg _s・ ・ ・ (4)

here,
‖ΔV ^reg _s ‖ = √ (Σ _d (ΔV ^reg _{d, s} ) ² ) (5)

(Odor identification)
In order to identify the odor sample, here, the pattern recognition method is used for identification. There are various pattern identification methods such as a method using a neural network and a method using multivariate analysis. However, the above-described concentration normalized feature (ΔV ^reg _{d, s} ) or sensor-specific normalized feature (C _{d). , s} ) is not particularly limited as long as it is a calculation method capable of inputting as feature quantities.

Here, an identification method using the BP method (Back Propagation method) will be described. The BP method has a calculation structure shown in FIG. 6 and is a supervised learning algorithm of a hierarchical neural network. When a certain value x _i is given to the input layer, the output y _k is output by equations (6) and (7).

Here, f (x) is a sigmoid function, which is a differentiable continuous function. A random small value is set as an initial value for the coupling coefficient v _jk between the intermediate layer and the output layer and the coupling coefficient w _ij between the input layer and the intermediate layer. Further, the error E between the teacher signal d _k as the target output and the actual output y _k is defined as in the following equation (8).

In order to decrease this value, the coupling coefficient v _jk between the intermediate layer and the output layer and the coupling coefficient w _ij between the input layer and the intermediate layer are sequentially updated. If the error E finally becomes a certain value or less, learning is performed. It judges that it has converged and ends. Otherwise, repeat from equation (6). v _jk and w _ij are updated by the following equations (9) to (12) by the steepest descent method.

Here, η is a learning coefficient, and 0 <η <1.
By repeating these updates, the output y _k approaches the teacher signal d _k .

(Calculation example)
Actually, 10 types of coffee and tea were learned and identified by the BP method. Here, the number of odor sensors is 22 and the number of input layer units is 22. Then, the value of the density normalized feature value (ΔV ^reg _{d, s} ) obtained by Expression (1) is given as an input. The number of intermediate layer units was 10, the number of output layers was 10, and the learning coefficient was 0.05. The initial value of the coupling coefficient was given as a random number, and learning was repeated until the error was 5 × 10 ⁻² or less.
Ten measurements were performed for each sample, 4 pieces were used as learning data, and the remaining 6 pieces were used as identification data. Learning data was randomly selected and repeated 3000 times.
The calculation results are shown in FIG. Overall, it could be identified with an accuracy of 91%.

  This result is an improvement over the conventional calculation in which identification is performed using “maximum value (maximum displacement from the reference value)” without normalization.

In the above embodiment, the calculation of odor identification is performed using the “density normalized feature value (ΔV ^reg _{d, s} )”, but instead of this, “sensor-specific normalized feature value (C _{d, s} )” is used. If the calculation of odor identification is performed using this, the identification result is further improved.

  The present invention can be used as an odor identification method.

10 Measurement System 11 Gas Cylinder 12 Permeator 13 Sampling Box S Odor Sensor

Claims (2)

With installing multiple odor sensor characteristics to odors differ (s) in the container, the concentration is different at the same kind of odors taken the N smell samples (d) is respectively at a constant flow rate can be supplied to the vessel the odor measurement by a plurality of odor sensors to perform more than a predetermined time, the for N odor samples, odor identifying the odor of odor samples based on the detection signal group concentration obtained by the plurality of odor sensor by An identification method,
Calculating a maximum displacement the which is the difference between the maximum signal value after flowing the previous reference signal value and odor sample of N odor sample flow, respectively, the maximum value vector based on the maximum displacement amount of the detection signals A maximum value vector calculating step of calculating (ΔV ^max _{d, s} );
The norm ‖ΔV ^max _d ‖ of the maximum value vector (ΔV ^max _{d, s} ) for each odor sample, and the average value of norms ‖ΔV ^max _dに つ い て for N odor samples having different concentrations with the same kind of odor < calculating a ‖ΔV ^max _d ‖> formula (1) by the maximum value density normalization feature value by performing a normalization including density information by multiplying said average value on which the 1-norm of the vector ([Delta] V ^reg _{d, s} ), a density normalization feature calculation step,
An odor identification method comprising an identification step of performing an odor identification calculation process based on the acquired density normalization feature quantity (ΔV ^reg _{d, s} ).

_{^{ΔV reg d, s = {(}} ΔV max d, s) / ‖ΔV max d ||} · <‖ΔV ^max _d ‖> (1)

here,
‖ΔV ^max _d ‖ = √ (Σ _s (ΔV ^max _{d, s} ) ² ) (2)

<‖ΔV ^max _d ‖> = (Σ _d (‖ΔV ^max _d ‖)) / N (3)
With installing multiple odor sensor characteristics to odors differ (s) in the container, the concentration is different at the same kind of odors taken the N smell samples (d) is respectively at a constant flow rate can be supplied to the vessel the odor measurement by a plurality of odor sensors to perform more than a predetermined time, the for N odor samples, odor identifying the odor of odor samples based on the detection signal group concentration obtained by the plurality of odor sensor by An identification method,
Calculating a maximum displacement the which is the difference between the maximum signal value after flowing the previous reference signal value and odor sample of N odor sample flow, respectively, the maximum value vector based on the maximum displacement amount of the detection signals A maximum value vector calculating step of calculating (ΔV ^max _{d, s} );
The norm ‖ΔV ^max _d ‖ of the maximum value vector (ΔV ^max _{d, s} ) for each odor sample, and the average value of norms ‖ΔV ^max _dに つ い て for N odor samples having different concentrations with the same kind of odor < calculating a ‖ΔV ^max _d ‖> formula (1) by the maximum value density normalization feature value by performing a normalization including density information by multiplying said average value on which the 1-norm of the vector ([Delta] V ^reg _{d, s} ), a density normalization feature calculation step,
The norm ‖ΔV ^reg _sの of the maximum value vector for each odor sensor is calculated with respect to the density normalized feature value (ΔV ^reg _{d, s} ), and the maximum value vector for each odor sensor is set to 1 by Equation (4). Normalization feature amount calculation process for each sensor which performs normalization including the magnitude information of the response of each sensor to obtain a normalization feature value for each sensor (C _{d, s} ),
An odor identification method comprising an identification step of performing an odor identification calculation process based on the acquired sensor-specific normalized feature value (C _{d, s} ).

_{^{ΔV reg d, s = {(}} ΔV max d, s) / ‖ΔV max d ||} · <‖ΔV ^max _d ‖> (1)

here,
‖ΔV ^max _d ‖ = √ (Σ _s (ΔV ^max _{d, s} ) ² ) (2)

<‖ΔV ^max _d ‖> = (Σ _d (‖ΔV ^max _d ‖)) / N (3)

C _{d, s} = {(ΔV ^reg _{d, s} ) / ‖ΔV ^reg _s・ ・ ・ (4)

here,
‖ΔV ^reg _s ‖ = √ (Σ _d (ΔV ^reg _{d, s} ) ² ) (5)