JP2000292403A - Gas detecting method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately determine a gas kind by a neutral net while accurately determining the concn. of gas by multiple regression analysis. SOLUTION: A sine wave is applied to the heater 6 of a gas sensor 2 and the signal waveform of the sensor 2 is subjected to Fourier transform. The obtained Fourier series is inputted to a neutral net 22 to determine a gas kind and the concn. of gas is determined on the basis of the regression coefficient calculated by the multiple regression analysis from the Fourier series to the concn. of gas with respect to the determined gas kind.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to gas detection using a metal oxide semiconductor.

[0002]

2. Description of the Related Art Yoshikawa et al. Propose that a heater voltage such as a sine wave is applied to a metal oxide semiconductor gas sensor, and a signal waveform of the gas sensor is analyzed to determine a gas type and a gas concentration. No. 2867,474). In this method, the sinusoidal heater voltage is a stimulus to the gas sensor, and the signal waveform of the gas sensor in response to the stimulus not only has the same frequency component as the original sine wave but also harmonic components such as the second harmonic and the third harmonic. Contains. For this reason, it can be said that this technique detects a non-linear response to stimulation applied as a change in the heater voltage.

[0003] Yoshikawa et al. Show that the signal waveform of the above-described gas sensor is Fourier-transformed, and the gas type and gas concentration are determined from the characteristics of the obtained Fourier series. However, the relationship between Fourier series and gas species and concentration is complicated.

[0004]

An object of the present invention is to determine the type and concentration of gas accurately and easily.

[0005]

According to the gas detection method of the present invention, the power of the heater is periodically changed by using a gas sensor having a heater and a metal oxide semiconductor for gas detection, and the metal oxide semiconductor for this is changed. In the method of detecting a gas by Fourier transforming the signal waveform of the above, the obtained Fourier series is input to a neural network to determine the gas type, and the regression coefficient from the Fourier series to the gas concentration is stored for each gas type. The gas concentration is determined from the regression coefficient and the obtained Fourier series. Preferably, a Fourier series is obtained from signal waveforms of a plurality of gas sensors at a known gas concentration, and a regression coefficient is obtained by a multiple regression analysis using the value of the Fourier series. That is, the regression coefficient is determined using the signals of a plurality of sensors including another sensor used for the actual measurement.

[0006]

According to the present invention, the heater power is periodically changed, and the signal waveform of the metal oxide semiconductor corresponding thereto is Fourier-transformed to obtain a Fourier series. In this way, for example, a large number of signals can be obtained from one gas sensor. Then, the obtained Fourier series is processed by a neural network to determine a gas type. Next, multiple regression analysis is performed on the relationship of the Fourier series to the gas concentration for each gas type, and the regression coefficient from each component of the Fourier series to the gas concentration is stored. The gas concentration is determined using the regression coefficient and the Fourier series actually measured at the time of measurement. Where the gas type is
The gas type determined by the neural network is used, and the corresponding regression coefficient is used. Therefore, the neural network only needs to determine the gas type, and the gas concentration can be easily determined by multiple regression analysis.

Here, in order to obtain a regression coefficient, a plurality of sets of Fourier series values are required. Therefore, for example, using a plurality of gas sensors of the same kind, preferably of the same lot, which are actually used for gas detection, and obtaining a Fourier series from a signal waveform at a known gas concentration, the data of these plural sensors is obtained. Is subjected to multiple regression analysis to determine a regression coefficient. In this way, the regression coefficient can be easily determined without repeatedly measuring the signal waveform with one gas sensor.

[0008]

1 to 13 show an embodiment. FIG. 1 shows a configuration of a gas detection device. Reference numeral 2 denotes a gas sensor which heats a metal oxide semiconductor 4 whose resistance value changes by a gas such as SnO 2 by a heater 6. Reference numeral 8 denotes a sine wave power supply, which generates a sine wave heater voltage for one cycle, for example, about 10 to 60 seconds. Reference numeral 10 denotes a constant voltage power supply that applies a constant voltage to the series side of the metal oxide semiconductor 4 and the load resistor 11 and uses the output as a power supply for each part of the detection device. Here, the value of the resistor 11 is set to be sufficiently small so that the output to the resistor 11 is inversely proportional to the resistance value of the metal oxide semiconductor 4.

Reference numeral 12 denotes an AD converter, and reference numeral 14 denotes a memory, which stores a signal waveform for one cycle from the gas sensor 2. Here, the cycle means the cycle of the output signal from the sine wave power supply 8. In addition, the memory 14 starts storing with a timing signal from the sine wave power supply 8, and the timing to start storing is, for example, the maximum value or the minimum value of the heater voltage. This is because the amount of information in the signal waveform is small near the maximum value or the minimum value of the heater voltage, and the state of the gas sensor 2 changes during one cycle. This is because, even if the values of the signal waveforms are different, the influence is small.

Reference numeral 16 denotes a shaping unit which processes the signal waveform for one cycle stored in the memory 14 and checks whether the difference between the first and last values of the signal waveform for one cycle is within an allowable range. If this value is outside the allowable range, no detection is performed and the process waits until the difference between the first and last values of the signal waveform falls within the allowable range. Next, even if this value is within the allowable range, if there is a difference between the sensor signal values at the beginning and end of one cycle, a correction signal that changes linearly with time is added so as to cancel this difference. 1. The values of the sensor signals are matched at the beginning and end of one cycle. As a result, the output from the shaping section 16 becomes a completely periodic signal.

Reference numeral 18 denotes a memory which stores sensor signals for one cycle, for example, and repeatedly reads out the value of the memory 18 a plurality of times, for example, 4 to 16 times, so that the Fourier transform unit 20 provides four or 16 cycles. Input signal waveform such as minute. That is, if Fourier transform is performed on the sensor signal for one cycle, the conversion accuracy is reduced due to so-called reflection or the like at both ends of the cycle.
It is repeatedly stretched to a plurality of cycles such as a cycle, particularly preferably a power-of-two cycle, and Fourier-transformed.

Reference numeral 22 denotes, for example, a three-layer neural network, which comprises an input layer, an intermediate layer, and an output layer. The number of input neurons is determined by the number of the Fourier series output from the Fourier transform unit 20. For example, here, in addition to the zero-order component (R0) of the real number, the sixth component from the fundamental component of the real part and the imaginary part is used. A total of 13 neurons of R1 to R6 and I1 to I6 are arranged. In addition, a humidity sensor, a temperature sensor, an airflow sensor, and the like may be provided, and these signals may be input to the neural network 22. The intermediate layer is composed of, for example, about 8 to 10 neurons, and the output layer is composed of, for example, 8 neurons for the purpose of discriminating eight kinds of gases. The output of each neuron takes a sigmoid function ranging from 0 to 1, with a value of 1 for the output neuron being 100% probability of gas being present for that neuron and a value of 0 being 0 being gas being present for that neuron. %.

A signal from the Fourier transform unit 20 is input to a quantitative unit 26, and a memory 24 stores regression coefficients obtained by multiple regression analysis. Then, in the quantitative unit 26,
A read address for reading a regression coefficient corresponding to which gas is determined based on the gas type obtained by the neural network 22, and a regression coefficient corresponding to the read address is sequentially output. The output regression coefficient is supplied with the value of the corresponding Fourier series from the Fourier transformer 20, multiplied by these, and sequentially added to obtain the gas concentration. This gas concentration is in logarithmic units.

FIG. 2 shows an example of the output voltage of the sine wave power supply 8. Here, one cycle is 40 seconds, the minimum sine wave voltage is 2 V, the maximum voltage is 5.5 V, and the storage in the memory 14 is started when the minimum voltage 2 V is output.

FIG. 3 shows a signal waveform for one cycle for 100 ppm of ethanol and 100 ppm of methanol. As is evident from the figure, ethanol is
The harmonic components in the Fourier series are large, forming a more complex signal waveform. In FIG. 3, the signal values at the beginning and end of one cycle completely match, but if they do not match, the signal is processed by the shaping unit 16 and the first and last values of one cycle are artificially processed. To match.

FIG. 4 shows the real part of the Fourier series for the eight gases, and FIG. 5 shows the imaginary part of the Fourier series for the same eight gases. Each gas concentration is 100ppm
The gas used was eight kinds of ethanol, methanol, diethyl ether, acetone, ethylene, ammonia, isobutane, and benzene. In addition to this, the Fourier series has a real constant part R0, the value of which is shown in Table 1.

[0017]

[Table 1] Fourier series (real part 0th order component) of eight kinds of gases ethanol 2182 methanol 2421 diethyl ether 2234 acetone 3525 ethylene 3687 ammonia 7224 isobutane 6398 benzene 3697

In this embodiment, since six components (R1 to R6) are considered as real components and six components (I1 to I6) are considered as imaginary components in addition to the constant part R0 of the real number as the Fourier series, the sum of one signal waveform is considered. Thirteen signals can be extracted. These 13 signals are input to the input layer of the neural network 22, and the gas type is output from the output layer. The learning of the neural network 22 and the determination of the regression constant are performed as follows. For example, three gas sensors of the same type and of the same lot are used, and these gas sensors are simultaneously used under the same conditions, and eight kinds of gases (1) of a known concentration are used.
(0, 30, 100 ppm), and the output corresponding thereto was subjected to Fourier transform to obtain a Fourier series. As a result, 1
Three Fourier series values were obtained for each gas concentration of one gas type. Then, these values were input to the neural network 22, and learning was performed until the neural network 22 converged. The obtained coupling constant between neurons was copied and used for all gas detectors.

Since the type of gas is known, multiple regression analysis is performed using the Fourier series value of signals at three gas concentrations for one gas (three sensors for each gas concentration), and the regression constant is calculated. Were determined. Then, the determined regression constant was stored in the memory 24 of all the gas detectors. The reason why the number of gas sensors is set to three here is due to the convenience of experiments, and is preferably set to a larger number such as 10 or 60. Also preferably, using a gas sensor of the same kind and the same lot including a gas sensor actually used for measurement, a coupling constant between neurons in the neural network 22 and a regression constant in multiple regression analysis are obtained.
These are stored in each gas detector. In this way, instead of performing a plurality of measurements for each gas detection device and individually learning the neural network 22 or finding a regression constant for each sensor, a large number of gas sensors are used at once. Neural network 22
Can be used to determine the coupling constant between neurons and the regression constant. As a result, the labor for calibrating the gas detection device can be reduced.

Tables 2 to 9 show the values of the regression constants for the above eight types of gases. The regression coefficient is obtained by a multiple regression analysis from the values of the Fourier series obtained for the three gas sensors. Next, the conversion from the regression coefficient to the gas concentration is, for example, in the case of ethanol in Table 2, the equation
Do as in (1). That is, the output is a natural logarithmic unit of the gas concentration, with the constant term placed at the top, and then multiplication of the regression coefficient for the real 0th order component and multiplication of the regression coefficient for the real number 2nd order component.

In the determination of the regression coefficient, for example, the number of signal components used in the multiple regression analysis is limited to a maximum of four, and a combination that minimizes Wilkes' lambda within this range is obtained.
This corresponds to finding, for 13 possible signals, up to four components that can best explain the gas concentration.
Next, when a maximum of four components are obtained, a combination in which one or two components are reduced is examined. If the error due to the multiple regression analysis increases only within a predetermined range, the combination is used for the multiple regression analysis. The components (factors) were reduced. In the examples, the factors used for multiple regression analysis are 2 to
There are four components.

[0022]

[Table 2] Ethanol Non-standardized regression coefficient constant 8.82460 Real 0-0.0048 Real 2-0.01566

LnC (ppm) = 8.8246-0.0004
8.R0-0.01566R2 (1) The constant term, 8.824
Reference numeral 6 denotes an offset, which is a regression coefficient −0 to the real zero-order component R0.
By multiplying R0 by [00048] and multiplying the regression coefficient to the real quadratic component R2 by -0.01566R2, the ethanol concentration is determined.

[0024]

[Table 3] Methanol Non-standardized regression coefficient constant 12.47281 Real number 0-0.000005 Real number 2-0.02281 Real number 1-0.000072

[0025]

[Table 4] Diethyl ether Non-standardized regression coefficient constant 7.1648 Real number 1-0.00308 Real number 0-0.0005

[0026]

[Table 5] Acetone Non-standardized regression coefficient constant 7.3010 Real number 0-0.00059 Real number 6 0.047837

[0027]

[Table 6] Ethylene Non-standardized regression coefficient constant 4.66242 Real number 3 0.017473 Real number 6 0.05363

[0028]

Table 7: Ammonia non-standardized regression coefficient constant 18.16242 Imaginary number 2 -0.0329 Imaginary number 3 0.041732 Real number 0 -0.000023

[0029]

[Table 8] Isobutane Non-standardized regression coefficient constant 11.94946 Real 1-0.000068 Imaginary 1 0.0000795 Imaginary 6-0.111536 Imaginary 2-0.000045

[0030]

[Table 9] Benzene Non-standardized regression coefficient constant 6.59884 Real number 4-0.04553 Real number 2-0.01176 Real number 0-0.00015

FIGS. 6 to 13 show that eight kinds of gases are used.
The gas concentration obtained by regression analysis using the data of Tables 2 to 9 is shown. Actual gas concentrations are 10ppm, 30ppm, 100
From the data of the three types of ppm and the measurement different from the measurement used to determine the regression coefficient, the predicted values by the regression analysis of the gas concentrations in FIGS. 6 to 13 were obtained. That is, for the three gas sensors used to determine the regression coefficients, the value of the Fourier series of the signal waveform is determined by another measurement, and the gas type output by the neural network 22 is stored as the correct gas type and stored. Was used to determine the gas concentration by equation (1) and the like.
In addition, completely correct values were obtained for the eight kinds of gases. This is because the neural network 22
This means that eight types of gases can be identified.

As shown in FIGS. 6 to 13, the gas concentration predicted by the regression analysis agrees very well with the actual gas concentration. Therefore, it is clear that the type of gas can be specified by using the neural network 22, and the gas concentration can be accurately determined by using regression analysis.

[Brief description of the drawings]

FIG. 1 is a block diagram of a gas detection device according to an embodiment.

FIG. 2 is a characteristic diagram showing a heater waveform used in the embodiment.

FIG. 3 is a characteristic diagram showing an output waveform of a gas sensor in 100 ppm of methanol and 100 ppm of ethanol in Examples.

FIG. 4 shows an example in which each of ethanol (EtOH), methanol (MeOH), diethyl ether, acetone, ethylene, ammonia, isobutane, and benzene is 100%.
Characteristic diagram showing the real part of the Fourier series when the output waveform of the gas sensor in ppm is Fourier transformed

FIG. 5 shows an example in which each of ethanol (EtOH), methanol (MeOH), diethyl ether, acetone, ethylene, ammonia, isobutane, and benzene is 100%.
Characteristic diagram showing the imaginary part of Fourier series when the output waveform of gas sensor in ppm is Fourier transformed

FIG. 6 shows an example of ethanol 10, 30, 10
Characteristic diagram showing the results of multiple regression analysis for 0 ppm

FIG. 7 shows methanol, 10, 30, and 10 in Examples.
Characteristic diagram showing the results of multiple regression analysis for 0 ppm

FIG. 8 shows an example of diethyl ether 10,3.
Characteristic diagram showing the results of multiple regression analysis for 0,100 ppm

FIG. 9 shows acetone, 10, 30, and 100 in Examples.
Characteristic diagram showing the results of multiple regression analysis for ppm

FIG. 10 shows examples of ammonia 10, 30, 10
Characteristic diagram showing the results of multiple regression analysis for 0 ppm

FIG. 11 shows an example of ethylene 10, 30, 100.
Characteristic diagram showing the results of multiple regression analysis for ppm

FIG. 12: Isobutane 10, 30, 10 in Examples
Characteristic diagram showing the results of multiple regression analysis for 0 ppm

FIG. 13: Benzene 10, 30, 100 in Examples
Characteristic diagram showing the results of multiple regression analysis for ppm

[Explanation of symbols]

 Reference Signs List 2 gas sensor 4 metal oxide semiconductor 6 heater 8 sine wave power supply 10 constant voltage circuit power supply 12 AD converter 14 memory 16 shaping unit 18 memory 20 Fourier transform unit 22 neural network 24 regression coefficient memory 26 quantitative unit

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Toshihiro Utaka 1-3-5 Senba Nishi, Minoh City Inside Figaro Giken Co., Ltd. (72) Inventor Kazuko Takamatsu 1-3-5 Senba Nishi Mino City, Figaro Giken F term in the company (reference) 2G046 AA10 AA18 AA24 AA26 DA03 DC07 DC16 DC17 DC18 DD02 EB06 FB02 FE39

Claims (2)

[Claims] 1. A gas sensor including a heater and a metal oxide semiconductor for gas detection, wherein the power of the heater is periodically changed, and a signal waveform of the metal oxide semiconductor corresponding thereto is Fourier-transformed. In the method of detecting gas, the obtained Fourier series is input to a neural network to determine the gas type, and the regression coefficient from the Fourier series to the gas concentration is stored for each gas type, and the regression coefficient and the obtained Determining a gas concentration from the Fourier series obtained. 2. The gas detection method according to claim 1, wherein the regression coefficient is obtained by a multiple regression analysis using a Fourier series obtained from signal waveforms of a plurality of gas sensors at a known gas concentration.