CN102192927B - Air-quality monitoring system based on electronic nose technique, and monitoring method thereof - Google Patents

Abstract

The invention discloses an air-quality monitoring system based on electronic nose technique, and a monitoring method thereof. The system is characterized by being provided with a sensor signal conditioning circuit, wherein a quality detection input end group of the sensor signal conditioning circuit is connected to a gas sensor group; the output end of the sensor signal conditioning circuit is connected to a sensor array signal preprocessor; the output end of the sensor array signal preprocessor is connected to a central processing unit; and a display end of the central processing unit is connected to a display. The system has the advantages that the discreteness problem of a metallic oxide sensor and the mutual interference problem of various gases are better solved by utilizing a virtual resistance method and a specific algorithm provided by the invention, thereby greatly promoting the output precision of the system, being capable of calculating the concentration values of different outputted gases, and determining the type of the gases, and simultaneously finishing the qualitative and quantitative identification of gases, and one output of an artificial neural network represents one gas.

Description

A kind of air quality monitoring system and monitoring method thereof based on Electronic Nose Technology

Technical field

The present invention relates to a kind of air borne sensor monitoring technology, specifically, be based on the air quality monitoring system and the monitoring method thereof of Electronic Nose Technology

Background technology

The air quality Electronic Nose of traditional based semiconductor metal oxide sensor array is often only carried out qualitative identification or odorousness is carried out grade identification, and this has all influenced the practicability of Electronic Nose.Even possess the ability of qualitative and quantitative identification, also be to realize: at first use the qualitative identification of network to re-use another network and quantitatively discern through the mode that two neural networks are connected.It is not strong that this makes Electronic Nose have practicality, and the complexity of algorithm is high, poor anti jamming capability, and precision is relatively poor, and arithmetic speed waits shortcoming slowly.

Because Detection of Air Quality Electronic Nose product has run into the problem of metal oxide sensor discreteness in mass production: the discreteness that is included in sensitive resistance Ro in the air and sensitivity Rs/Ro.Even such result has caused our sensor under the situation of equivalent environment and identical tested gas concentration, also to have different responses, thereby make algorithm output result also different, influenced the production of product lot quantityization.The method of generally using at present is to use same batch sensor to reduce the discreteness of sensor and equate to reduce the discreteness influence with Ro through on hardware, pull-up resistor being made as with this, but does not all well address this problem.

Summary of the invention

The present invention provides a kind of air quality monitoring system and monitoring method thereof based on Electronic Nose Technology, can screen out sensor and under the situation of varying environment and tested gas concentration, have different responses, can solve the mutual interference problem of multiple gases.Propose the virtual resistance method simultaneously and solved the sensor discrete property problem of Electronic Nose in mass production.

For achieving the above object, the present invention is based upon on the basis of Electronic Nose Technology.

A kind of air quality monitoring system based on Electronic Nose Technology; Its key is: be provided with sensor signal conditioning circuit; The quality testing input end group of this sensor signal conditioning circuit is connected with sensor array; The output terminal of this sensor signal conditioning circuit is connected with the sensor array signal pretreater, and the output terminal of this sensor array signal pretreater is connected with central processing unit, and the display end of this central processing unit is connected with display;

Theoretical foundation of the present invention is pattern-recognition, and pattern-recognition need be carried out certain pre-service to data, just can obtain best recognition effect.If the sensor array of electric nasus system has n sensor; Wherein the number of gas sensor is Q; Tested toxic and harmful is measured p time; Will obtain the sample set data matrix X of a p * n dimension, all pattern classifications and qualitative identifying all are based on this sample data collection matrix, so the quality of data set X directly has influence on the result of identification.

Have humiture module and gas sensor group in the said sensor array: said humiture module obtains ambient temperature and humidity digital signal g, sends in the transducer in the central processing unit; Through the variation of sensitive resistance value in the gas sensor, obtain one group of gas information a, send to sensor signal conditioning circuit;

Sensor signal conditioning circuit: be used for changing gas information a into one group of voltage signal b that changes;

Sensor array signal pretreater: be used for converting the voltage signal b of said variation into set of number signal b ' through AD, and this group digital signal b ' is carried out feature extraction, obtain the digital signal d of the stack features extraction of Q gas sensor;

The mode of feature extraction is more, can extract all data in the output response curve, extracts the voltage max in each sensor curve of output, minimum value, or slope variation maximal value, minimum value etc.

Be provided with transducer, neural network and one group of corrector in the said central processing unit;

Said each corrector uses the virtual resistance method that each road gas sensor digital signal among the digital signal d is carried out sensitivity compensation respectively, is compensated the pairing set of number signal correction of the output voltage values value e of the said gas sensor group in back;

Said transducer carries out the normalization processing to said digital signal modified value e and said ambient temperature and humidity digital signal g, obtains one group of normalized signal f;

Because the output difference of each sensor is bigger in the sensor array, the same difference of variance separately is very big.Based on pattern recognition theory, have only the covariance between each independent variable very high, just can obtain good classifying quality.And to obtain big covariance coefficient, sample data just needs through some pre-service.Pretreated purpose is for the contribution of each variable of balance to the overall variance of sample, thereby makes sample to separate more significantly.The pretreated method of electric nasus system data generally comprises the conversion of baseline correction, data and the extraction of eigenwert, and purpose all is in order to carry out pattern-recognition better.

Can know through article " based on the application of temperature and humidity compensation method in gasmetry of artificial intelligence " and " the temperature and humidity compensation new method in the gasmetry " simultaneously, the humiture data can be compensated the influence of humiture to gas sensor as the input of neural network; Temperature and humidity compensation method based on artificial neural network is the arranged side by side input of the response of the humiture of environment and gas sensor as measuring system, as 3 preconditions that the status is identical inferring gas concentration; Ambient temperature and humidity is reflected on the connection power of network the influence of gas sensor output, becomes the indivisible ingredient of system.During actual measurement, through the computing of network, the influence of ambient temperature and humidity is fallen by compensation impliedly, thereby obtains the precise information of gas concentration.

The input neural unit number of said neural network promptly is the dimension of said normalized signal f, and neural network is handled normalized signal f, and is qualitative, identify the kind and the concentration of dusty gas quantitatively, and sends to said display and show.

Sensor normalization a kind of effectively data conversion method of can yet be regarded as; Sensor normalization can make the output of single-sensor be positioned at [0; 1] between, so not only can reduce the error of calculation in the stoichiometry identification, and can prepare suitable data for the input space in the neural network recognizer; Improve convergence of algorithm speed, this method is particularly suitable for the quantitative identification of electric nasus system.

The sensor method for normalizing that the present invention adopts is the linear function transformation approach:

y＝(x-Min)/(Max-Min)

X, y are respectively the value of normalization front and back in the formula.

In traditional method for normalizing, need to seek maximal value Max and minimum M in the sample data, and among the present invention the maximal value Max of sample in the formula is confirmed as the full scale value x that electric nasus system is exported _Full, sample minimum M in confirms as 0.The normalization formula is write as:

Y=x/x _Full

The utilization following formula carries out data normalization to be handled, and has simplified computation process greatly.

The input neural unit number of said neural network promptly is the dimension of said normalized signal f, and neural network is qualitative, identify the kind and the concentration of dusty gas quantitatively, and sends to said display and show.

The warning end of said central processing unit is connected with audible-visual annunciator.

Said central processing unit also is connected with wireless transmitter module, and this wireless transmitter module is connected with wireless mode with wireless receiving module on the air purifier.

4, according to the said air quality monitoring system based on Electronic Nose Technology of claim 1, it is characterized in that: the circuit structure of all gas sensor is consistent in the said gas sensor group, and the model of each gas sensor is different;

Gas sensor is provided with heating resistor R _HWith sensor sensing resistance R s, heating resistor R _HTwo ends be connected with heating voltage V _HThe end of sensor sensing resistance R s is connected the positive power source terminal of DC voltage Vc; Behind the other end polyphone pull-up resistor RL of sensor sensing resistance R s, be connected the negative power end of DC voltage Vc, the two ends of this pull-up resistor RL connect said sensor signal conditioning circuit.

The quantity of said gas sensor group; The quantity of gas information a among one group of gas information a; The quantity of the voltage signal b that changes among the voltage signal b that a group changes; The quantity of digital signal b ' among the set of number signal b ', the quantity of the digital signal d of feature extraction among the digital signal d that a stack features extracts, the quantity of digital signal modified value e all equates among the set of number signal correction value e.The circuit structure of all gas sensor is consistent in the said gas sensor group, and the model of each gas sensor is different;

Gas sensor is provided with heating resistor R _HWith sensor sensing resistance R s, heating resistor R _HTwo ends be connected with heating voltage V _HThe end of sensor sensing resistance R s is connected the positive power source terminal of DC voltage Vc; Behind the end polyphone pull-up resistor RL of sensor sensing resistance R s, be connected the negative power end of DC voltage Vc, the two ends of this pull-up resistor RL connect said sensor signal conditioning circuit.

Using neural network algorithm that gas to be measured is carried out in the quantitative identifying, algorithm the convergence speed is slow.Under the big situation of training sample; The training time of algorithm more extends, and this raising for work efficiency is very disadvantageous, through research; The method of weighted is carried out in the target output that has proposed quantitatively to discern neural network; Promptly, make the data of gas with various output be in the same order of magnitude, improved the speed of convergence of neural network greatly multiply by corresponding coefficient in the corresponding network objectives output of every kind of gas.

Suppose that air quality detector mixes m kind poisonous and harmful or single gas detects, overall measurement p time, the neural network target that then is used for quantitatively identification is output as:

$T_{m\&times;p}=\left[\begin{array}{cccc}{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>}}_{\mathrm{1,1}}& {\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& ...& {\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>}}_{1,\mathrm{<figure-callout\; id="2"\; label="p\; T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">p</figure-callout>}}& \\ T_{}{}_{\mathrm{2,1}}& {\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& ...& {\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>}}_{2,p}\mathrm{}\\ .& .& \mathrm{}& .\\ .& .& ...& .\\ .& .& \mathrm{}& .\\ T_{m,1}& T_{m,2}& ...& T_{m,p}\end{array}\right]$

T _{M * p}In the measurement concentration value of the corresponding a kind of gas of each row when at every turn measuring; Because national air quality standard standard is different to the regulation of gas with various; Have not even at the same order of magnitude, the output concentration that causes gas with various in the testing process is not on the same order of magnitude, in matrix T _{M * p}In to show as the data bulk level of different rows different, the normalization method that the present invention adopts is to neural network target output T _{M * p '}Carry out weighting, the weighting coefficient of establishing m kind gas is respectively α ₁, α ..., α _m, the objective matrix after the weighting is:

${T^{\&prime;}}_{m\&times;p}=\left[\begin{array}{cccc}{\mathrm{\&alpha;}}_1{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{1,1}}& {\mathrm{\&alpha;}}_1{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& ...& {\mathrm{\&alpha;}}_1{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>}}_{1,p}\\ {\mathrm{\&alpha;}}_2{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{2,1}}& {\mathrm{\&alpha;}}_2{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& ...& {\mathrm{\&alpha;}}_2{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>}}_{2,p}\mathrm{}\\ .& .& \mathrm{}& .\\ .& .& ...& .\\ .& .& \mathrm{}& .\\ {\mathrm{\&alpha;}}_m{T}_{m,1}& {\mathrm{\&alpha;}}_m{T}_{m,2}& ...& {\mathrm{\&alpha;}}_m{T}_{m,p}\end{array}\right]$

Weighted value α ₁, α ₂α _mMake matrix T after the conversion ' _{M * p}Middle data are in the same order of magnitude.Experiment showed, with the matrix T after the weighting ' _{M * p}As the output of neural network target, net training time significantly reduces.

The monitoring method of a kind of air quality monitoring system, its key is to comprise the steps:

Step 1: utilize the humiture module to obtain ambient temperature and humidity digital signal g, and directly send into the transducer in the central processing unit;

Utilize the sensitive resistance value of Q gas sensor to change, obtain Q gas information a that changes, send to sensor signal conditioning circuit;

Step 2: sensor signal conditioning circuit is used for changing gas information a into Q voltage signal b that changes;

Step 3: the sensor array signal pretreater is transformed into Q digital signal b ' with Q voltage signal b that changes, and Q digital signal b ' carried out feature extraction, obtains Q feature extraction value d;

Step 4: a Q corrector uses the virtual resistance method that each road gas sensor digital signal among the digital signal d is carried out sensitivity compensation respectively, is compensated the pairing Q of output voltage values digital signal correction value e of the said gas sensor group in back;

Step 5: transducer carries out normalization and handles said digital signal modified value e and ambient temperature and humidity digital signal g, obtains n normalized signal f, and wherein n equals the dimension sum of e and g;

Step 6: neural network is calculated n normalized signal f, and is qualitative, identify the kind and the concentration of dusty gas quantitatively, and sends to said display and show.

Wherein the virtual resistance method in the step 4 is:

At first because digital signal d is one group of digital signal data after the process conversion, the inside comprises a plurality of gas sensor information, at first analyzes the j road sensing data d among the digital signal d so understand for ease _jBe example, release the sensing data of whole digital signal d with this.

For d _jAccording to the virtual resistance ratio juris, we can release:

s _1j＝a _j*s _2j+c _j；

s _1jBe the sensitivity of pairing this type standard transducer of j circuit-switched data; s _2jBe the sensitivity that needs this type sensor of rectification; a _i, c _jIt is the pairing sensor of j circuit-switched data and its standard transducer parameter value through the definite correction relationship of test.If

v _outj＝d _j*D

V _OutjBe the output voltage of reality of sensor this moment, d _jBe the digital signal value of this road signal, D is the analog-to-digital resolution of AD in the sensor array signal pretreater, and for example: the reference voltage of establishing AD converter is 5V, and this AD is 12, the resolution of this AD then, and promptly step-length is 5/2 ¹²=0.00122V/bit,

${\mathrm{Rs}}_{2j}=(\frac{\mathrm{Vc}}{V_{\mathrm{OUTj}}}-1){\mathrm{RL}}_j$

Rs _2jBe the resistance value of sensitive resistance Rs of j sensor this moment, Vc is a probe power voltage, RL _jResistance value for the actual pull-up resistor RL of this sensor; Order

$S_{2j}=\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Ro}}_{2j}}$

Ro _2jBe the sensor sensing resistance value of this sensor under the standard nonpollution environment, the pull-up resistor Rl in virtual circuit reconstruct _j=Ro _2j, this is in order to let the circuit after the virtual reconstruct under standard environment, be output as Vc/2, because this moment, sensor resolution was best, to have realized that also the sensor that is corrected all has the consistance of output in standard environment simultaneously.

Following formula is brought into:

${V_{\mathrm{OUTj}}}^{\&prime;}=\frac{\mathrm{Vc}}{\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Rl}}_j}+1}=\frac{\mathrm{Vc}}{S_{2j}+1}$

V _OUTj' be the output voltage values when not carrying out sensitivity compensation after the virtual circuit reconstruct, carry out the V as a result behind the sensitivity compensation _OUTj" as follows:

${V_{\mathrm{OUTj}}}^{\&prime;\&prime;}=\frac{\mathrm{Vc}}{a_j*\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Ro}}_{2j}}+c_j+1}=\frac{\mathrm{Vc}}{a_j*\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Rl}}_j}+{\mathrm{<figure-callout\; id="1"\; label="c\; j\; +"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">c</figure-callout>}}_j+1}$

So following according to the formula of algorithm reasoning, this formula is applicable to the correction of a way word.

$e_j=\frac{{V_{\mathrm{OUTj}}}^{\&prime;\&prime;}}{D}=\frac{\mathrm{Vc}}{a_j*\frac{(\frac{\mathrm{Vc}}{d_j*D}-1)*{\mathrm{RL}}_j}{{\mathrm{Rl}}_j}+c_j+1}*\frac{1}{D}=\frac{\mathrm{Vc}}{a_j*(\frac{\mathrm{Vc}}{d_j*D}-1)*\frac{{\mathrm{RL}}_j}{{\mathrm{Rl}}_j}+c_j+1}*\frac{1}{D}$

Wherein, e _jBe this road signal d _jDigital signal after revising, D is the resolution of A D analog to digital converter, RL _jIt is the pull-up resistor value in the corresponding side circuit of this sensor; Rl _jBe the pull-up resistor value in the reconstruct of sensor virtual circuit, it equals the sensitive resistance value of the sensor under the standard nonpollution environment.

So, can release converting into of whole gas sensor group digital signal vector d and e thus:

$e=\frac{\mathrm{Vc}}{A(\frac{\mathrm{Vc}}{d*D}-1)*\frac{\mathrm{Rz}}{\mathrm{Rl}}+C+1}*\frac{1}{D}$

This formula is applicable to the correction of whole gas sensor group.

Wherein, A, d, R _Z, Rl, C, e are multi-C vector and dimension and equate, and A, C are gas sensor arrays through testing the correction parameter vector between definite and the standard transducer array susceptibility; R _ZBe the resistance RL of each sensor load resistance of gas sensor group _jThe vector of being formed, Rl are the pull-up resistor value Rl of gas sensor group each sensor in virtual circuit reconstruct _jThe vector of being formed, d are the digital signals that need proofread and correct through the step 1 group, and e is through the corrected set of number signal of virtual resistance method, and their dimension equals the dimension Q of gas sensor output signal.Calculate arbitrary way word signal, only need the formula above the respective components substitution in the above-mentioned vector is got final product.

The circuit structure of a said Q gas sensor is consistent, independent model of each gas sensor;

Be provided with heating resistor R _HWith sensor sensing resistance R s, heating resistor V _HTwo ends be connected with heating voltage V _HThe end of sensor sensing resistance R s is connected the positive power source terminal of DC voltage Vc; Behind the end polyphone pull-up resistor RL of sensor sensing resistance R s, be connected the negative power end of DC voltage Vc, the two ends of this pull-up resistor RL connect said sensor signal conditioning circuit.

In the said step 6, each gas sensor is different to the response sensitivity of gas with various, and after normalization was handled, central processing unit was judged the kind and the concentration of gas according to the sensitivity of all gas sensor's response through algorithm process;

The input neural unit number of neural network algorithm is the signal number (dimension n of signal) of sensor; The output nerve unit number promptly is the gaseous species (m) that we need detect; Can find out here; Each output unit representative of output neuron needs a kind of gas of detection, and for example first represents formaldehyde.Represent CO for second ...And the data of output neuron output are exactly the concentration value of its corresponding such gas.

In the experiment poisonous and harmful mixed or single gasmetry p time, the structure of establishing neural metwork training input matrix P is following:

$P=\left[\begin{array}{cccc}{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">S</figure-callout>}}_{\mathrm{1,1}}& {\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="S"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">S</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& ...& {\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">S</figure-callout>}}_{1,\mathrm{<figure-callout\; id="2"\; label="p\; S"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">p</figure-callout>}}& \\ S_{}{}_{\mathrm{2,1}}& {\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="S"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">S</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& ...& {\mathrm{<figure-callout\; id="2"\; label="S"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">S</figure-callout>}}_{2,p}\mathrm{}\\ .& .& \&CenterDot;& .\\ .& .& .& .\\ .& .& \&CenterDot;& .\\ S_{n,1}& S_{n,2}& ...& S_{n,p}\end{array}\right]$

Wherein, n representes number of sensors

The structure of training objective matrix T does $T=\left[\begin{array}{cccc}{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">t</figure-callout>}}_{\mathrm{1,1}}& {\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="t"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">t</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& ...& {\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">t</figure-callout>}}_{1,\mathrm{<figure-callout\; id="2"\; label="p\; t"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">p</figure-callout>}}& \\ t_{}{}_{\mathrm{2,1}}& {\mathrm{<figure-callout\; id="2"\; label="t"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="t"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">t</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& ...& {\mathrm{<figure-callout\; id="2"\; label="t"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">t</figure-callout>}}_{2,p}\\ .& .& \mathrm{}& .\\ .& .& ...& .\\ .& .& \mathrm{}& .\\ \mathrm{}& \mathrm{}& .& \mathrm{}\\ t_{m,1}& ...& .& t_{m,p}\\ \mathrm{}& \mathrm{}& .& \mathrm{}\end{array}\right]$

The concentration value of the element representation gas of objective matrix T, respectively corresponding certain gas of each row shared concentration in tested gas, such as, the concentration value of the first line display gas 1 in each the measurement.According to training input matrix P and training objective matrix T, can train through neural network, according to neural network BP training algorithm, thereby obtain the weight matrix W of input layer to hidden layer ₁, threshold value B ₁, and hidden layer is to the weight matrix W of output layer ₂, threshold value B ₂If the transport function of hidden layer is f ¹, the transport function of output layer is f ²But, be output as by network structure computational grid hidden layer

Y ₁＝f ¹(W ₁*P+B ₁)

The network output layer is output as so

Y ₂＝f ²(W ₂*Y ₁+B ₂)

Output valve Y ₂Be the gas concentration value of prediction, its each row is represented the concentration value of certain gas under each the measurement respectively.In practical application, according to the sensor's response value, through the weights and the threshold value of neural network, just can calculate the concentration value of output, and can confirm the kind of gas, a kind of gas of one tunnel output expression.So far, we have accomplished the process of the qualitative and quantitative identification of gas.

Remarkable result of the present invention is:

Can screen out sensor and under the situation of varying environment and tested gas concentration, have different responses; Solve the discreteness problem of metal oxide sensor preferably, the output accuracy of system is improved greatly, solved the mutual interference problem of multiple gases; Can calculate the concentration value of gas with various output; And can confirm the kind of all gases, a kind of gas of one tunnel output expression can be accomplished the qualitative and quantitative identification to gas simultaneously.

Description of drawings

Fig. 1 is a system architecture block scheme of the present invention;

Fig. 2 is the circuit structure diagram of gas sensor;

Fig. 3 is a neural network qualitative, quantitative recognition structure block diagram;

Fig. 4 is the concrete multiple gases qualitative, quantitative of a neural network recognition structure block diagram;

Fig. 5 is the topology view of neural metwork training input matrix P;

Fig. 6 is the topology view of training objective matrix T.

Embodiment

Below in conjunction with accompanying drawing and specific embodiment the present invention is explained further details:

Embodiment 1; As shown in Figure 1: a kind of air quality monitoring system of the present invention based on Electronic Nose Technology; Be provided with sensor signal conditioning circuit 1, the quality testing input end group of this sensor signal conditioning circuit 1 is connected with the gas sensor group, and the output terminal of this sensor signal conditioning circuit 1 is connected with sensor array signal pretreater 2; The output terminal of this sensor array signal pretreater 2 is connected with central processing unit 3, and the display end of this central processing unit 3 is connected with display 4;

Theoretical foundation of the present invention is pattern-recognition, and pattern-recognition need be carried out certain pre-service to data, just can obtain best recognition effect.The sensor array of supposing electric nasus system has n sensor; Tested volatile organic matter is measured p time; Will obtain the sample set data matrix X of a p * n dimension; All pattern classifications and qualitative identifying all are based on this sample data collection matrix, so the quality of data set X directly has influence on the result of identification.Said sensor array comprises humiture module and gas sensor group: utilize the humiture module to obtain ambient temperature and humidity digital signal g, directly send into the transducer 32 in the central processing unit; Utilize the variation of the sensitive resistance value of gas sensor, obtain one group of gas information a, send to sensor signal conditioning circuit 1;

Sensor signal conditioning circuit 1: be used for changing gas information a into one group of voltage signal b that changes;

Sensor array signal pretreater 2: be used for converting the voltage signal b of said variation into set of number signal b ' through AD, and this group digital signal b ' is carried out feature extraction, obtain the digital signal d of the stack features extraction of Q sensor;

The mode of feature extraction is more, can extract the voltage max in each sensor curve of output, minimum value, or slope variation maximal value, minimum value etc.

Be provided with transducer 32, neural network 33 and one group of corrector 31 in the said central processing unit 3;

Said each corrector 31 uses the virtual resistance method that each way word signal among the digital signal d is carried out sensitivity compensation respectively, is compensated the pairing set of number signal correction of the output voltage values value e of the said gas sensor group in back;

32 couples of said digital signal modified value e of said transducer and humiture digital signal g carry out normalization to be handled, and obtains one group of normalized signal f;

Because the output difference of each sensor is bigger in the sensor array, sometimes maybe be not on an one magnitude, the same difference of variance separately is very big.Based on pattern recognition theory, have only the covariance between each independent variable very high, just can obtain good classifying quality.And to obtain big covariance coefficient, sample data just needs through some pre-service.Pretreated purpose is for the contribution of each variable of balance to the overall variance of sample, thereby makes sample to separate more significantly.The pretreated method of electric nasus system data generally comprises the conversion of baseline correction, data and the extraction of eigenwert, and purpose all is in order to carry out pattern-recognition better.

Sensor normalization a kind of effectively data conversion method of can yet be regarded as; Sensor normalization can make the output of single-sensor be positioned at [0; 1] between, makes response vector be in the same order of magnitude, so not only can reduce the error of calculation in the stoichiometry identification; And can prepare suitable data for the input space in the neural network recognizer, this method is particularly suitable for the quantitative identification of electric nasus system.

The sensor method for normalizing that the present invention adopts is the linear function transformation approach:

y＝(x-Min)/(Max-Min)

X, y are respectively the value of normalization front and back in the formula.

In traditional method for normalizing, need to seek maximal value Max and minimum M in the sample data, and among the present invention the maximal value Max of sample in the formula is confirmed as the full scale value x that electric nasus system is exported _Full, sample minimum M in confirms as 0.The normalization formula is write as:

Y=x/x _Full

The utilization following formula carries out data normalization to be handled, and has simplified computation process greatly.

The input neural unit number of said neural network 33 promptly is the dimension of said normalized signal f, neural network 33 through qualitative, identify the kind and the concentration of dusty gas quantitatively, and send to said display 4 and show.

The warning end of said central processing unit 3 is connected with audible-visual annunciator 5.

Said central processing unit 3 also be connected with wireless transmitter module 6, the wireless receiving module wireless connections on this wireless transmitter module 6 and the air purifier 7.

As shown in Figure 2: the circuit structure of all gas sensor is consistent in the said gas sensor group, and the model of each gas sensor is different;

Be provided with heating resistor R _HWith sensor sensing resistance R s, heating resistor R _HTwo ends be connected with heating voltage V _HThe end of sensor sensing resistance R s is connected the positive power source terminal of DC voltage Vc; Behind the end polyphone pull-up resistor RL of sensor sensing resistance R s, be connected the negative power end of DC voltage Vc, the two ends of this pull-up resistor RL connect said sensor signal conditioning circuit 1.

The quantity of said gas sensor group; The quantity of gas information a among one group of gas information a; The quantity of the voltage signal b that changes among the voltage signal b that a group changes; The quantity of digital signal b ' among the set of number signal b ', the quantity of the digital signal d of feature extraction among the digital signal d that a stack features extracts, the quantity of digital signal modified value e all equates among the set of number signal correction value e.

Like Fig. 3, shown in 4: using neural network algorithm that gas is carried out in the quantitative identifying, algorithm the convergence speed is slow.Under the big situation of data volume; The training time of algorithm more extends, and this raising for work efficiency is very disadvantageous, through research; Proposition will quantitatively be discerned the target of neural network and export the method for carrying out weighted; Promptly in the corresponding network objectives output of every kind of gas, multiply by corresponding coefficient, make the data of gas with various output be in the same order of magnitude, improved the speed of convergence of neural network greatly.

Suppose that air quality detector mixes m kind poisonous and harmful or single gas detects, every kind of gasmetry p time, the neural network target that then is used for quantitatively discerning is output as:

$T_{m\&times;p}=\left[\begin{array}{cccc}{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>}}_{\mathrm{1,1}}& {\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& ...& {\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>}}_{1,\mathrm{<figure-callout\; id="2"\; label="p\; T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">p</figure-callout>}}& \\ T_{}{}_{\mathrm{2,1}}& {\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& ...& {\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>}}_{2,p}\mathrm{}\\ .& .& \mathrm{}& .\\ .& .& ...& .\\ .& .& \mathrm{}& .\\ T_{m,1}& T_{m,2}& ...& T_{m,p}\end{array}\right]$

T _{M * p}In the output concentration value of the corresponding a kind of gas of each row because national air quality standard standard is different to the regulation of gas with various, have not even at the same order of magnitude, the output concentration that causes gas with various in the testing process is not on the same order of magnitude, in matrix T _{M * p}In to show as the data bulk level of different row different, the normalization method that the present invention adopts is to neural network target output T _{M * p}Carry out weighting, the weighting coefficient of establishing m kind gas is respectively α ₁, α ₂..., α _m:

${T^{\&prime;}}_{m\&times;p}=\left[\begin{array}{cccc}{\mathrm{\&alpha;}}_1{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{1,1}}& {\mathrm{\&alpha;}}_1{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{1,2}}& ...& {\mathrm{\&alpha;}}_1{\mathrm{<figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>}}_{1,p}\\ {\mathrm{\&alpha;}}_2{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="T"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{2,1}}& {\mathrm{\&alpha;}}_2{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">\; <figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>\; </figure-callout>}}_{\mathrm{2,2}}& ...& {\mathrm{\&alpha;}}_2{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="HSA00000335056500041.png"\; state="\{\{state\}\}">T</figure-callout>}}_{2,p}\mathrm{}\\ .& .& \mathrm{}& .\\ .& .& ...& .\\ .& .& \mathrm{}& .\\ {\mathrm{\&alpha;}}_m{T}_{m,1}& {\mathrm{\&alpha;}}_m{T}_{m,2}& ...& {\mathrm{\&alpha;}}_m{T}_{m,p}\end{array}\right]$

Value α ₁, α ₂α _mMake matrix T after the conversion ' _{M * p}Middle data are in the same order of magnitude.Experiment showed, with the matrix T after the weighting ' _{M * p}When exporting as the neural network target, net training time significantly reduces.

Embodiment 2, the monitoring method of a kind of air quality monitoring system, and its key is to comprise the steps:

Step 1: utilize one group; Promptly utilize the humiture module to obtain ambient temperature and humidity digital signal g; The sensitive resistance value of sending into transducer 32 and Q gas sensor in the central processing unit changes, and obtains the gas information a of Q variation, sends to sensor signal conditioning circuit 1;

Step 2: sensor signal conditioning circuit 1 is used for changing gas information a into Q voltage signal b that changes;

Step 3: sensor array signal pretreater 2 is transformed into Q digital signal b ' with Q voltage signal b that changes, and Q digital signal b ' carried out feature extraction, obtains Q feature extraction value d;

Step 4: a Q corrector 31 uses the virtual resistance method that a digital signal d is carried out sensitivity compensation respectively, is compensated the pairing Q of output voltage values digital signal correction value e of the said gas sensor group in back;

Step 5: 32 couples of said digital signal modified value e of transducer and humiture digital signal g, carry out normalization and handle, obtain n normalized signal f, wherein n equals the dimension sum of e and g;

Step 6: neural network 33 is through calculating n normalized signal f, and is qualitative, identify the kind and the concentration of dusty gas quantitatively, and sends to said display 4 and show.

Wherein the virtual resistance method in the step 4 is:

At first because digital signal d is one group of digital signal data after the process conversion, the inside comprises a plurality of gas sensor information, at first analyzes the j road sensing data d among the digital signal d so understand for ease _jBe example, release the sensing data of whole digital signal d with this.

For d _jAccording to the virtual resistance ratio juris, we can release:

s _1j＝a _j*s _2j+c _j；

s _1jBe the sensitivity of pairing this type standard transducer of j circuit-switched data; s _2jBe the sensitivity that needs this type sensor of rectification; a _j, c _jIt is the pairing sensor of j circuit-switched data and its standard transducer parameter value through the definite correction relationship of test.If

V _OUTj＝d _j*D

V _OUTjBe the output voltage of reality of sensor this moment, d _jBe the digital signal value of this road signal, D is the analog-to-digital resolution of AD in the sensor array signal pretreater 2, and for example: the reference voltage of establishing AD converter is 5V, and this AD is 12, and then the resolution of this AD (being step-length) is 5/2 ¹²=0.00122V/bit;

${\mathrm{Rs}}_{2j}=(\frac{\mathrm{Vc}}{V_{\mathrm{OUTj}}}-1){\mathrm{RL}}_j$

As shown in Figure 2: Rs _2jBe the resistance value of sensitive resistance Rs of j sensor this moment, Vc is a probe power voltage, RL _jResistance value for the actual pull-up resistor RL of this sensor; Order

$S_{2j}=\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Ro}}_{2j}}$

Ro _2jBe the sensor sensing resistance value of this sensor under the standard nonpollution environment, the pull-up resistor Rl in virtual circuit reconstruct _j=Ro _2jIn order to let the circuit after the virtual reconstruct under standard environment, be output as Vc/2,, realized that also the sensor that is corrected all has the consistance of output in standard environment simultaneously here because this moment, sensor resolution was best.

Following formula is brought into:

${V_{\mathrm{OUTj}}}^{\&prime;}=\frac{\mathrm{Vc}}{\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Rl}}_j}+1}=\frac{\mathrm{Vc}}{S_{2j}+1}$

V _OUTj' be the output voltage values when not carrying out sensitivity compensation after the virtual circuit reconstruct, carry out the V as a result behind the sensitivity compensation _OUTj" as follows:

${V_{\mathrm{OUTj}}}^{\&prime;\&prime;}=\frac{\mathrm{Vc}}{a_j*\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Ro}}_{2j}}+c_j+1}=\frac{\mathrm{Vc}}{a_j*\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Rl}}_j}+{\mathrm{<figure-callout\; id="1"\; label="c\; j\; +"\; filenames="HSA00000335056500041.png,HSA00000335056500042.png"\; state="\{\{state\}\}">c</figure-callout>}}_j+1}$

So following according to the formula of algorithm reasoning, this formula is applicable to the correction of a way word.

$e_j=\frac{{V_{\mathrm{OUTj}}}^{\&prime;\&prime;}}{D}=\frac{\mathrm{Vc}}{a_j*\frac{(\frac{\mathrm{Vc}}{d_j*D}-1)*{\mathrm{RL}}_j}{{\mathrm{Rl}}_j}+c_j+1}*\frac{1}{D}=\frac{\mathrm{Vc}}{a_j*(\frac{\mathrm{Vc}}{d_j*D}-1)*\frac{{\mathrm{RL}}_j}{{\mathrm{Rl}}_j}+c_j+1}*\frac{1}{D}$

Wherein, e _jBe this road signal d _jDigital signal after revising, D is the resolution of A D analog to digital converter, RL _jIt is the pull-up resistor value in the corresponding side circuit of this sensor; Rl _jBe the pull-up resistor value in the reconstruct of sensor virtual circuit, it equals the sensitive resistance value of the sensor under the standard nonpollution environment.

So, can release converting into of whole gas sensor group digital signal vector d and e thus:

$e=\frac{\mathrm{Vc}}{A(\frac{\mathrm{Vc}}{d*D}-1)*\frac{\mathrm{Rz}}{\mathrm{Rl}}+C+1}*\frac{1}{D}$

This formula is applicable to the correction of whole gas sensor group.

Wherein, A, d, Rz, Rl, C, e are multi-C vector and dimension and equate, A, C are gas sensor arrays through testing the correction parameter vector between definite and the standard transducer array susceptibility; R _ZBe the resistance RL of each sensor load resistance of gas sensor group _jThe vector of being formed, Rl are the pull-up resistor value Rl of gas sensor group each sensor in virtual circuit reconstruct _jThe vector of being formed, d are the digital signals that need proofread and correct through the step 1 group, and e is through the corrected set of number signal of virtual resistance method, and their dimension equals the dimension Q of gas sensor output signal.

As shown in Figure 2:

The circuit structure of a said Q gas sensor is consistent, independent model of each gas sensor;

Be provided with heating resistor R _HWith sensor sensing resistance R s, heating resistor R _HTwo ends be connected with heating voltage V _HThe end of sensor sensing resistance R s is connected the positive power source terminal of DC voltage Vc; Behind the end polyphone pull-up resistor RL of sensor sensing resistance R s, be connected the negative power end of DC voltage Vc, the two ends of this pull-up resistor RL connect said sensor signal conditioning circuit 1.

In the said step 6, each gas sensor is different to the response sensitivity of gas with various, and after normalization was handled, central processing unit 3 was judged the kind and the concentration of gas according to the sensitivity of all gas sensor's response through algorithm process;

As shown in Figure 3:

1, the qualitative model of cognition of gasometry:

Can find out by Fig. 3; The input neural unit number of artificial neural network algorithm is the signal number of sensor; Be the dimension of signal, the output nerve unit number promptly is the gaseous species m that we need detect, and can find out here; Each output unit representative of output neuron needs a kind of gas of detection, and for example first represents formaldehyde.Represent CO for second ...And the data of output neuron output are exactly the concentration value of its corresponding such gas.

2, the training of this algorithm and prediction and calculation process:

If the structure of neural metwork training input matrix P is as shown in Figure 5, the structure of training objective matrix T such as as shown in Figure 6.

According to training input matrix P and training objective matrix T, can train through neural network, according to neural network BP training algorithm, thereby obtain the weight matrix W of input layer to hidden layer ₁, threshold value B ₁, and hidden layer is to the weight matrix W of output layer ₂, threshold value B ₂If the transport function of hidden layer is f ¹, the transport function of output layer is f ²But, be output as by network structure computational grid hidden layer

Y ₁＝f ¹(W ₁*P+B ₁)

The network output layer is output as so

Y ₂＝f ²(W ₂*Y ₁+B ₂)

The structure of Fig. 4 is the present example that we use: output valve Y ₂Be the gas concentration value of prediction, its each row is represented gas 1 respectively, gas 2 ..., the concentration value of gas m.In practical application, according to the sensor's response value, through the weights and the threshold value of neural network, just can calculate the concentration value of output, and can confirm the kind of gas, a kind of gas of one tunnel output expression.So far, we have accomplished the process of the qualitative and quantitative identification of gas.

Claims (8)

1. air quality monitoring system based on Electronic Nose Technology; It is characterized in that: be provided with sensor signal conditioning circuit (1); The quality testing input end group of this sensor signal conditioning circuit (1) is connected with sensor array; The output terminal of this sensor signal conditioning circuit (1) is connected with sensor array signal pretreater (2), and the output terminal of this sensor array signal pretreater (2) is connected with central processing unit (3), and the display end of this central processing unit (3) is connected with display (4);

Have humiture module and gas sensor group in the said sensor array: said humiture module obtains ambient temperature and humidity digital signal g, sends in the transducer (32) in the central processing unit; Through the variation of sensitive resistance value in the gas sensor, obtain one group of gas information a, send to sensor signal conditioning circuit (1);

Sensor signal conditioning circuit (1): be used for changing gas information a into one group of voltage signal b that changes;

Sensor array signal pretreater (2): be used for converting the voltage signal b of said variation into set of number signal b ' through AD, and this group digital signal b ' is carried out feature extraction, obtain the digital signal d of the stack features extraction of Q gas sensor;

Be provided with transducer (32), neural network (33) and one group of corrector (31) in the said central processing unit (3);

Said each corrector (31) uses the virtual resistance method that each road gas sensor digital signal among the digital signal d is carried out sensitivity compensation respectively, is compensated the pairing set of number signal correction of the output voltage values value e of the said gas sensor group in back;

Said transducer (32) carries out the normalization processing to said digital signal modified value e and said ambient temperature and humidity digital signal g, obtains one group of normalized signal f;

The input neural unit number of said neural network (33) promptly is the dimension of said normalized signal f, and neural network (33) is handled normalized signal f, and is qualitative, identify the kind and the concentration of dusty gas quantitatively, and sends to said display (4) and show.

2. according to the said air quality monitoring system based on Electronic Nose Technology of claim 1, it is characterized in that: the warning end of said central processing unit (3) is connected with audible-visual annunciator (5).

3. according to the said air quality monitoring system of claim 1 based on Electronic Nose Technology; It is characterized in that: said central processing unit (3) also is connected with wireless transmitter module (6), and this wireless transmitter module (6) is connected with wireless mode with wireless receiving module on the air purifier (7).

4. according to the said air quality monitoring system based on Electronic Nose Technology of claim 1, it is characterized in that: the circuit structure of all gas sensor is consistent in the said gas sensor group, and the model of each gas sensor is different;

Gas sensor is provided with heating resistor R _HWith sensor sensing resistance R s, heating resistor R _HTwo ends be connected with heating voltage V _HThe end of sensor sensing resistance R s is connected the positive power source terminal of DC voltage Vc; Behind the other end polyphone pull-up resistor RL of sensor sensing resistance R s, be connected the negative power end of DC voltage Vc, the two ends of this pull-up resistor RL connect said sensor signal conditioning circuit (1).

5. according to the said air quality monitoring system of claim 1 based on Electronic Nose Technology; It is characterized in that: the quantity of said gas sensor group; The quantity of gas information a among one group of gas information a; The quantity of the voltage signal b that changes among the voltage signal b that a group changes; The quantity of digital signal b ' among the set of number signal b ', the quantity of the digital signal d of feature extraction among the digital signal d that a stack features extracts, the quantity of digital signal modified value e all equates among the set of number signal correction value e.

6. the monitoring method of the said air quality monitoring of claim 5 system is characterized in that comprising the steps:

Step 1: utilize the humiture module to obtain ambient temperature and humidity digital signal g, and send into the transducer (32) in the central processing unit;

Utilize the sensitive resistance value of Q gas sensor to change, obtain Q gas information a that changes, send to sensor signal conditioning circuit (1);

Step 2: sensor signal conditioning circuit (1) is used for changing gas information a into Q voltage signal b that changes;

Step 3: sensor array signal pretreater (2) is transformed into Q digital signal b ' with Q voltage signal b that changes, and Q digital signal b ' carried out feature extraction, obtains Q feature extraction value d;

Step 4: a Q corrector (31) uses the virtual resistance method that each road gas sensor digital signal among the digital signal d is carried out sensitivity compensation respectively, is compensated the pairing Q of output voltage values digital signal correction value e of the said gas sensor group in back;

Step 5: transducer (32) carries out normalization and handles said digital signal modified value e and ambient temperature and humidity digital signal g, obtains n normalized signal f, and wherein n equals the dimension sum of e and g;

Step 6: neural network (33) is calculated n normalized signal f, and is qualitative, identify the kind and the concentration of dusty gas quantitatively, and sends to said display (4) and show.

7. monitoring method according to claim 6 is characterized in that:

Virtual resistance method in the said step 4 is:

At first because digital signal d is one group of digital signal data after the process conversion, with the j road sensing data d among the digital signal d _jBe example, release the sensing data of whole digital signal d with this;

For d _jRelease according to the virtual resistance ratio juris:

s _1j＝a _j*s _2j+c _j；

s _1jBe the sensitivity of pairing this type standard transducer of j circuit-switched data; s _2jBe the sensitivity that needs this type sensor of rectification; a _j, c _jIt is respectively the pairing sensor of j circuit-switched data and its standard transducer parameter value through the definite correction relationship of variable concentrations alcohol testing; If

V _OUT?j＝d _j*D

V _{OUT j}Be the output voltage of reality of sensor this moment, D is the analog-to-digital resolution of AD in the sensor array signal pretreater (2);

${\mathrm{Rs}}_{2j}=(\frac{\mathrm{Vc}}{V_{\mathrm{OUT\; j}}}-1){\mathrm{RL}}_j$

Rs _2jBe the resistance value of sensitive resistance Rs of j sensor this moment, Vc is a probe power voltage, RL _jResistance value for the actual pull-up resistor RL of this sensor; Order

$S_{2j}=\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Ro}}_{2j}}$

Ro _2jBe the sensor sensing resistance value of this sensor under the standard nonpollution environment, the pull-up resistor Rl in virtual circuit reconstruct _j=Ro _2j

With following formula substitution following formula:

${V_{\mathrm{OUTj}}}^{\&prime;}=\frac{\mathrm{Vc}}{\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Rl}}_j}+1}=\frac{\mathrm{Vc}}{S_{2j}+1}$

V _{OUT j}' be the output voltage values when not carrying out sensitivity compensation after the virtual circuit reconstruct, carry out the V as a result behind the sensitivity compensation _{OUT j}" as follows:

${V_{\mathrm{OUTj}}}^{\&prime;\&prime;}=\frac{\mathrm{Vc}}{a_j*\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Ro}}_{2j}}+c_j+1}=\frac{\mathrm{Vc}}{a_j*\frac{{\mathrm{Rs}}_{2j}}{{\mathrm{Rl}}_j}+c_j+1}$

Because revised signal e _jBe digital signal, and V _{OUT j}" be the compensated voltage value, so, need carry out analog to digital conversion to it, following according to the formula of algorithm reasoning

The analog-to-digital resolution of AD in number pretreater (2), RL _jIt is the pull-up resistor value in the corresponding side circuit of this sensor; Rl _jBe the pull-up resistor value in the reconstruct of sensor virtual circuit, it equals the sensitive resistance value Ro of the sensor under the standard nonpollution environment _2j

So, can release converting into of gas sensor group set of number signal vector d and e thus:

$e=\frac{\mathrm{Vc}}{A*(\frac{\mathrm{Vc}}{d*D}-1)*\frac{\mathrm{Rz}}{\mathrm{Rl}}+C+1}*\frac{1}{D}$

Wherein, A, d, R _Z, Rl, C, e are multi-C vector and dimension and equate, and A, C are gas sensor arrays through testing the correction parameter vector between definite and the standard transducer array susceptibility; R _ZBe the resistance RL of each sensor load resistance of gas sensor group _jThe vector of being formed, Rl are the pull-up resistor value Rl of gas sensor group each sensor in virtual circuit reconstruct _jThe vector of being formed, d are the one group of digital signals that need proofread and correct, and e is through the corrected digital signal of virtual resistance method, and their dimension equals the dimension Q of gas sensor output signal.

8. monitoring method according to claim 6 is characterized in that:

In the said step 6; Each gas sensor is different to the response sensitivity of gas with various; After data normalization was handled, central processing unit (3) was based on the sensitivity of all gas sensor's response, through kind and the concentration of judging gas behind the algorithm for pattern recognition;

The input neural unit of neural network is counted the signal output number that n is a sensor array; The output nerve unit number promptly is the gaseous species m that needs detection; Each output unit representative of output neuron needs a kind of gas of detection, and the data of output neuron output are exactly the concentration value of this kind gas;

In the experiment poisonous and harmful is mixed perhaps single gasmetry p time of planting, the structure of establishing neural metwork training input matrix P is following:

$P=\left[\begin{array}{cccc}{S}_{\mathrm{1,1}}& S_{\mathrm{1,2}}& ...& S_{1,p}\\ S_{\mathrm{2,1}}& S_{\mathrm{2,2}}& ...& S_{2,p}\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ S_{n,1}& S_{n,2}& ...& S_{\text{n,p}}\end{array}\right]$

Wherein, n representes number of sensors

The structure of training objective matrix T does

$T=\left[\begin{array}{cccc}{t}_{\mathrm{1,1}}& t_{\mathrm{1,2}}& ...& t_{1,p}\\ t_{\mathrm{2,1}}& t_{\mathrm{2,2}}& ...& t_{2,p}\\ .& .& & .\\ .& .& ...& .\\ .& .& & .\\ & & .& \\ t_{m,1}& ...& .& t_{m,p}\\ & & .& \end{array}\right]$

The concentration value of the element representation gas of objective matrix T, respectively corresponding certain gas of each row shared concentration in tested gas, such as, the concentration value of the first line display gas 1 in each the measurement; According to training input matrix P and training objective matrix T, can train through neural network, according to neural network BP training algorithm, thereby obtain the weight matrix W of input layer to hidden layer ₁, threshold value B ₁, and hidden layer is to the weight matrix W of output layer ₂, threshold value B ₂, the transport function of establishing hidden layer is f ¹, the transport function of output layer is f ²But, be output as by network structure computational grid hidden layer

Y ₁＝f ¹(W ₁*P+B ₁)

The network output layer is output as so

Y ₂＝f ²(W ₂*Y ₁+B ₂)

Output valve Y ₂Be the gas concentration value of prediction, its each row is represented gas 1 respectively, gas 2 ..., the concentration value of gas m; In practical application,, through the weights and the threshold value of neural network, just can calculate the concentration value of output, and can confirm the kind of gas according to the sensor's response value.

Images