CN101949826A - Positive model and inverse model-based quantitative spectrometric analysis and calibration method of multi-component gas - Google Patents
Positive model and inverse model-based quantitative spectrometric analysis and calibration method of multi-component gas Download PDFInfo
- Publication number
- CN101949826A CN101949826A CN 201010270822 CN201010270822A CN101949826A CN 101949826 A CN101949826 A CN 101949826A CN 201010270822 CN201010270822 CN 201010270822 CN 201010270822 A CN201010270822 A CN 201010270822A CN 101949826 A CN101949826 A CN 101949826A
- Authority
- CN
- China
- Prior art keywords
- gas
- model
- sample
- spectral line
- input
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000004458 analytical method Methods 0.000 title claims abstract description 31
- 230000003595 spectral effect Effects 0.000 claims abstract description 64
- 238000013528 artificial neural network Methods 0.000 claims abstract description 24
- 238000010183 spectrum analysis Methods 0.000 claims abstract description 6
- 238000004364 calculation method Methods 0.000 claims abstract description 4
- 239000013598 vector Substances 0.000 claims description 40
- 238000012549 training Methods 0.000 claims description 35
- 238000001228 spectrum Methods 0.000 claims description 16
- 238000005316 response function Methods 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 9
- 230000035945 sensitivity Effects 0.000 claims description 7
- 238000012821 model calculation Methods 0.000 claims description 3
- 230000003287 optical effect Effects 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 claims 1
- 230000009194 climbing Effects 0.000 claims 1
- 230000002068 genetic effect Effects 0.000 claims 1
- 238000001514 detection method Methods 0.000 abstract description 2
- 238000003745 diagnosis Methods 0.000 abstract description 2
- 230000007613 environmental effect Effects 0.000 abstract description 2
- 238000004445 quantitative analysis Methods 0.000 abstract 1
- 239000007789 gas Substances 0.000 description 85
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical compound C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 description 66
- ATUOYWHBWRKTHZ-UHFFFAOYSA-N Propane Chemical compound CCC ATUOYWHBWRKTHZ-UHFFFAOYSA-N 0.000 description 28
- OTMSDBZUPAUEDD-UHFFFAOYSA-N Ethane Chemical compound CC OTMSDBZUPAUEDD-UHFFFAOYSA-N 0.000 description 26
- 239000001294 propane Substances 0.000 description 14
- 238000002835 absorbance Methods 0.000 description 11
- 238000006243 chemical reaction Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- NNPPMTNAJDCUHE-UHFFFAOYSA-N isobutane Chemical compound CC(C)C NNPPMTNAJDCUHE-UHFFFAOYSA-N 0.000 description 4
- 238000000862 absorption spectrum Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- CURLTUGMZLYLDI-UHFFFAOYSA-N Carbon dioxide Chemical compound O=C=O CURLTUGMZLYLDI-UHFFFAOYSA-N 0.000 description 2
- HSFWRNGVRCDJHI-UHFFFAOYSA-N alpha-acetylene Natural products C#C HSFWRNGVRCDJHI-UHFFFAOYSA-N 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 229910052799 carbon Inorganic materials 0.000 description 2
- 125000002534 ethynyl group Chemical group [H]C#C* 0.000 description 2
- 238000004868 gas analysis Methods 0.000 description 2
- 239000001282 iso-butane Substances 0.000 description 2
- 238000012113 quantitative test Methods 0.000 description 2
- IJDNQMDRQITEOD-UHFFFAOYSA-N sec-butylidene Natural products CCCC IJDNQMDRQITEOD-UHFFFAOYSA-N 0.000 description 2
- 238000010187 selection method Methods 0.000 description 2
- 238000005033 Fourier transform infrared spectroscopy Methods 0.000 description 1
- 229910018503 SF6 Inorganic materials 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 150000001335 aliphatic alkanes Chemical class 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 229910002092 carbon dioxide Inorganic materials 0.000 description 1
- 239000001569 carbon dioxide Substances 0.000 description 1
- 239000012467 final product Substances 0.000 description 1
- 238000002329 infrared spectrum Methods 0.000 description 1
- 238000002386 leaching Methods 0.000 description 1
- 239000000047 product Substances 0.000 description 1
- 239000011435 rock Substances 0.000 description 1
- 238000004611 spectroscopical analysis Methods 0.000 description 1
- SFZCNBIFKDRMGX-UHFFFAOYSA-N sulfur hexafluoride Chemical compound FS(F)(F)(F)(F)F SFZCNBIFKDRMGX-UHFFFAOYSA-N 0.000 description 1
- 229960000909 sulfur hexafluoride Drugs 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Landscapes
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention discloses a positive model and inverse model-based quantitative spectrometric analysis and calibration method of a multi-component gas. The method comprises the following steps of: establishing a positive model by taking concentrations of a target gas and an interference gas as input and a spectral line value as output; generating an additional sample by the positive model, wherein the additional sample is obtained by performing calculation by an interpolation method; extracting characteristic variable from a sample set consisting of the manufactured samples and the additional sample so as to form a new sample set; and finally establishing an analysis model of each target gas by a neural network modeling method by taking the characteristic variable as input and the concentration of the target gas as output. The method greatly reduces the number of samples required by the adoption of the conventional black box calibration method during the quantitative spectrometric analysis of the multi-component gas, so that the spectral analysis can be independently applied to the fields related to the quantitative analysis of the multi-component gas, such as gas logging, mine safety, environmental protection, fault diagnosis, product quality detection and the like.
Description
Technical field:
The invention belongs to the spectral analysis field, relate to a kind of multicomponent gas quantitative spectrochemical analysis scaling method, especially a kind of based on just, the multicomponent gas quantitative spectrochemical analysis scaling method of inversion model.
Background technology:
Quantitative gas analysis is the work that numerous areas such as environmental protection, fault diagnosis, safety monitoring, gas detection logging and exploration need be carried out.For example, factory's emission gases need be carried out quantitative test to compositions such as sulphuric dioxide wherein, environment structure is threatened determining whether; The exploration of oil, rock gas is to study and judge by the content of various alkane in the monitor layer gassing; The colliery also needs gases such as gas in the monitoring mine, acetylene, sulfur hexafluoride to judge the safety of mine; Transformer fault can be diagnosed by the content of gases such as methane, ethane, acetylene, carbon dioxide in the monitoring insulating oil.One of technology that realizes quantitative gas analysis is spectral analysis.The spectral analysis of gas has such characteristics: on the one hand, often composition is more in the combination gas to be analyzed, and the characteristic spectrum of some composition overlaps mutually, even it is serious to overlap; On the other hand, the spectral value and the relation between the gas concentration of nearly all spectrum line are non-linear, particularly big when gas concentration, when spectrometer resolution is low, all the more so, and the absorbance spectrum value of each spectral line often is not the stack of each component gas absorbance in the mixed gas yet.Therefore, adopt conventional multisensor measuring multiple parameters technology, as polynomial expression approach, neural network, support vector base etc., spectroscopic analysis system demarcated just has been absorbed in predicament: if adopt polynomial expression to approach, then because approximation capability is limited, the analysis result deviation is bigger; If adopt support vector machine, then or the kernel function complexity, be difficult to choose, or be absorbed in the predicament of dimension calamity; If adopt neural network,, otherwise be easy to occur training then because need the magnanimity sample.For example, methane, ethane, propane, isobutane and normal butane five component gas are owing to all contain-CH
3Molecule group, its mid infrared absorption spectrum overlaps mutually, and relatively more serious, if to its quantitative test, according to the quantitative spectrochemical analysis of multicomponent gas, a kind of gas is represented the one dimension direction with the fourier-transform infrared absorption spectrometer, each dimension is made 10 calibration points, then needs 10
5=100000 groups of data, each dimension is made 5 calibration points, also needs 5
5=3125 groups of data.Because the mixed gas sample is made quite wastes time and energy, cost is also high, make so that multisample has also just brought very large trouble to the demarcation of system, make the application of quantitative spectrochemical analysis of multicomponent gas be subjected to very big restriction.
Summary of the invention:
The objective of the invention is to, a kind of scaling method of multicomponent gas quantitative spectrochemical analysis is provided, adopt this scaling method, can obtain higher accuracy of analysis, and needed sample reduces greatly.Analysis for methane, ethane, propane, isobutane and normal butane five component gas obtains same accuracy of analysis, and required sample is reduced to hundreds of by thousands of groups of required samples of traditional multisensor scaling method.
To achieve these goals, the technical solution used in the present invention is, at first generate appended sample with positive model, then, the sample set training analysis model that forms by sample and appended sample, it is characterized in that: at first be input with the gas concentration, each spectral line value of spectrogram is output, utilize uncorrelated characteristic between each component gas concentration, make up rectangular coordinate system, and in this coordinate system, adopt simple relatively, curve interpolation method and surface interpolation method that result of calculation is relatively accurate are set up the spectral analysis positive model, and adopt positive model to calculate the spectrogram of the gas concentration correspondence between the sample point, obtain appended sample; Make up each component gas analytical model with nerual network technique then, and train each component gas analytical model, obtain the parameter of analytical model with the sample set that sample of making and appended sample constitute.Because the appended sample that positive model generates is abundant, therefore can overcome the training problem of crossing of neural network.Thought of the present invention is summarized as shown in Figure 1, mainly may further comprise the steps:
1) sample gas spectrum obtains.At first determine the L component target gas of required analysis, and the M that may exist in the combination gas to be analyzed group is disturbed gas, concentration range according to this L+M group gas, and proportioning, according to the sample gas of certain interval making some, perhaps buy corresponding standard gas, and these sample gas are full of the spectrometer air chamber successively, use the spectrometer scanning optical spectrum, obtain the spectrogram of this sample gas;
2) characteristic variable is extracted.From spectrogram N bar spectral line, extract S characteristic variable (S 〉=L+M), these characteristic variables can be the spectral line values of certain bar spectral line, the methane of locating also can be certain value that the combinatorial operation of some spectral line value obtains, can also be the area of a certain wave number section.Each characteristic variable is to a kind of sensitivity maximum in L+M kind target gas or the interference gas, and is less relatively to the sensitivity of other gas, and the smaller the better.The method of extracting can be that forward direction is chosen method, pivot analysis method or the bridge Return Law;
3) space makes up with appended sample and makes.Regard S every related spectral line value of characteristic variable as a signal of sensor, regard each component concentrations in the gas to be analyzed as influence this output input signal.Like this, L+M component concentrations value in certain gas to be analyzed, and the spectral line value of a spectral line, can be regarded as a point in L+M+1 dimension rectangular coordinate system space, allow this L+M component gas change in whole analyst coverage, then the spectral line value of this spectral line is exactly a curved surface in L+M+1 dimension rectangular coordinate system space.Between two or more consecutive point of this curved surface, selected one group of new concentration value carries out interpolation arithmetic, so just obtain a new spectral line value of this spectral line as input.With this group concentration value is input, increase the spectral line sequence number successively, and this spectral line value carried out interpolation arithmetic as output, so just obtain the individual new spectral line value of N (N is the quantity of the spectral line relevant with S characteristic variable), this N new spectral line value constituted a new spectrogram, and the concentration value of spectrogram that this is new and corresponding L+M component gas thereof is referred to as an appended sample;
4) generation of sample set.Change each concentration of component value in the gas to be measured, repeating step 3), then can obtain many arbitrarily appended sample, the gas concentration value of appended sample correspondence also can be set arbitrarily.The sample of making and just constitute a sample set P by the appended sample that method of interpolation obtains.According to 2) in the array mode of each characteristic variable, spectral line among the sample set P is made up.So the characteristic variable vector has constituted a new sample set Q again with target gas concentration value;
5) analytical model makes up.Each component in the L component target gas is made up an analytical model, and this model can be the BP neural network, also can be the RBF neural network.The input of model is the vector that S characteristic variable forms, and output is the vector that the concentration of certain or all target gas is formed.The hidden node number of neural network is set according to the quantity of input vector and the degree of correlation between the vector.Input vector is many more, and the degree of correlation between the input vector is big more, and required the number of hidden nodes is just many more.Initial hidden node quantity should be greater than S+T, and wherein T represents the quantity of two pairwise correlations between the input vector.The hidden layer response function is a nonlinear response function, adopts tansig function or logsig function.Output layer adopts linear response function;
6) analytical model training.Train the analytical model of each target gas with the sample set Q in the step 4).If model can't meet the demands, promptly the experience error is inadequately little, then returns 5), revise model parameter, as neural network hidden node number, the hidden node response function, the duplicate step of laying equal stress on is up to satisfying training requirement.If the experience error meets the demands, reduce a hidden node, repeat this step, can't meet the demands up to the experience error, the model of getting last training is as the final analysis model.
7) structure of analytical model calculating formula
After training analytical model, according to the parameter and the response function of analytical model, rebuild the calculation expression of analytical model, can obtain complete analytical model.The model of setting up for step 5), creation analysis Model Calculation expression formula as follows:
(1) calculates the hidden node input
In
m＝M
input×Input
S+b
m
M in the formula
InputBe the connection weight value matrix between input and the hidden node, Input
SFor having the input vector value of S component, In
mInput vector for hidden node; b
mThreshold vector for hidden node.
(2) calculate hidden node output
Out
m(i)＝f
m(In
m(i)))
F in the formula
m() is the hidden layer response function; Out
mBe the hidden layer output vector.
(3) calculate output layer output
In
O＝M
output×Out
m+b
O
M in the formula
OutputBe the weight matrix that links between hidden layer and the output layer; b
OBe the output layer threshold vector; In
OBe the output layer input vector.Because what output layer adopted is linear response function, so the output vector of output layer is In
O
Description of drawings
Fig. 1 synoptic diagram of the present invention;
Fig. 2 sample gas spectrum is made synoptic diagram;
Fig. 3 (a) is one-component 10% methane, 2% ethane, 1% propane, and the high wave number section of the middle infrared spectrum spectrum of the combination gas of 10% methane and 2% ethane;
Fig. 3 (b) is the curve that the conversion absorbance of 3002.2 place's spectral line methane changes with methane concentration for wave number;
Fig. 4 is 3002.2 place's methane, the trigonometric interpolation synoptic diagram of ethane conversion absorbance in rectangular coordinate system for wave number;
Fig. 5 is used for the neural network structure figure of gas concentration analysis modeling.
Embodiment
For the purpose of drawing, illustrate conveniently, be without loss of generality, the present invention is example explanation the specific embodiment of the present invention with methane, ethane and propane three component gas, supposes that here methane, ethane are target gas, and propane is to disturb gas, and the output valve of spectrum is a transmissivity.This method adopts triangular interpolation method to obtain appended sample, and sets up the concentration analysis model of each component gas with neural network.
1) sample gas spectrum obtains
Need to analyze methane, ethane two component target gas in this embodiment, the interference gas propane that exists in the combination gas to be analyzed, wherein the methane concentration scope is 0～100%, the ethane concentration range is 0～5%, propane concentration scope 0～1%.The one-component gas sample disposes voluntarily with flow controller, and wherein the sample point concentration value of methane is [0.02%, 0.1%, 0.2%, 0.5%, 1%, 3%, 10%, 30%, 50%, 70%, 100%], ethane sample point concentration value is [0.02%, 0.1%, 0.2%, 0.5%, 1%, 2%, 5%], propane sample point concentration value is [0.02%, 0.1%, 0.2%, 0.5%, 1%]; The combination gas sample obtains by buying corresponding standard gas, and the calibrating gas concentration of corresponding partially mixed gas is respectively:
These sample gas are full of the spectrometer air chamber successively, use the spectrometer scanning optical spectrum, obtain the spectrogram of this sample gas.The horizontal ordinate of spectrogram is generally wave number, and ordinate mainly contains two kinds: a kind of is absorbance, and another kind is a transmissivity.The absorbance at certain spectral line place is corresponding to the negative of the natural logarithm of this place's transmissivity.Therefore, no matter the ordinate of spectrogram adopts any form, be identical in essence.Spectrogram in this example is output with the transmissivity, spectrometer is made the spectrum samples synoptic diagram as shown in Figure 1, one-component gas 10% methane, 1% ethane, 1% propane that obtain, and the combination gas spectrum of 10% methane and 2% ethane is shown in accompanying drawing 3 (a), and wave number is that the curve that changes with methane concentration of the conversion absorbance of 3002.3 place's spectral line methane is shown in accompanying drawing 3 (b).From accompanying drawing 2 (a) as can be seen the absorption spectrum of these three kinds of gases overlap serious, from accompanying drawing 3 (b) as can be seen, along with the increase of concentration, the conversion absorptance of gas reduces, that is to say that the conversion absorbance of gas is more and more littler with the amplitude that gas concentration increases.Wherein converting absorbance is the product of absorptance and light path, and the conversion absorptance is the ratio of conversion absorbance and gas concentration;
2) characteristic variable is extracted
In the characteristic variable leaching process, can extract the characteristic variable of a certain spectral line as certain gas, area that also can a certain section spectrum can also extract the characteristic variable of the combination of many spectral line values as gas with various as characteristic variable.For different application scenarios, the extracting method difference of employing, selected characteristic variable are also different.For example, observe accompanying drawing 3 (a) as can be known, for the quantitative spectrochemical analysis of the combination gas of methane, ethane, propane three components, the spectral line value val at wave number 3016.5 places
3016.5Can be used as the characteristic variable of methane, wave number 3002.2 is to all the spectral line value sums between the wave number 3022.3, and the combination of the spectral line value at wave number 3002.2,3016.5 and 3022.3 places: log (v
3002.2)+log (val
3022.3)-2 * log (val
3016.5) (log () represents natural logarithm), also can be used as the characteristic variable of methane.But anyway, in learning sample, the spectral line value of the spectral line that use under the gas with various concentration combination need be known.For the sake of simplicity, adopt forward selection procedures here, the difference of natural logarithm of choosing wave number and be two spectral line values of 3022.3 and 3016.5 is as the characteristic variable of methane:
v
m＝ln(val
3022.3)-ln(val
3016.5) (1a)
Wave number is the characteristic variable of the natural logarithm difference of 3029.2 and 3081.2 spectral line value as ethane:
v
e＝ln(val
3081.2)-ln(val
3029.2) (1b)
Wave number is the characteristic variable of the natural logarithm difference of 3002.2 and 3028.0 spectral line value as propane:
v
p＝ln(val
3028.0)-ln(val
3002.2) (1c)
V in the formula
m, v
eAnd v
pThe characteristic variable of representing methane, ethane and propane respectively, ln (val
n) represent that wave number is the natural logarithm of the spectral line value of n.Here just by observing accompanying drawing 3 (a), relatively the characteristic variable of the difference of which two spectral line formation is higher to the remolding sensitivity of a certain gas, and lower to other gas sensitivity for the forward selection procedures of Cai Yonging.Adopting natural logarithm is because the spectral line value is a spectral-transmission favtor, takes from the linearity of the later characteristic variable that forms of right logarithm and wants high relatively.Getting the spectral line difference is because the characteristic variable that this method forms helps eliminating the influence that the translation of spectrum baseline brings as characteristic variable.If what the ordinate of spectrogram adopted is the absorbance form, then do not need to ask natural logarithm in the formula (1), directly get final product with the spectral line value;
3) space makes up with appended sample and makes
Here be the output of the natural logarithm of 3002.2 spectral line value as sensor with wave number, methane and ethane concentration are the sensor input, make up rectangular coordinate system, rectangular coordinate system just, and be that example is made " appended sample " with the triangular interpolation method.With the output of the natural logarithm of spectral line value as sensor, be because compare with the transmissivity with the spectral line that is output, the linearity between it and the gas concentration is far better.As shown in Figure 4, the absorbance of known this wave number spectral line under 10% methane, 2% ethane situation is respectively 0.0126 and 0.2913, spectral value is respectively: 0.90892 and 0.66372, and its natural logarithm is respectively :-0.095498 and-0.40989, the coordinate in accompanying drawing 4 is used respectively
"+" indicates; The spectral value of mixed gas at wave number 3002.3 places that is made of 10% methane, 2% ethane is: 0.5298, and natural logarithm is-0.65204, the coordinate in accompanying drawing 4 indicates with " ◇ ".If methane, ethane and propane all do not have in the tested gas, then the spectral value at wave number 3002.3 places is 0, and coordinate indicates with " O " in accompanying drawing 3.According to triangular interpolation method, given any input (c
m, c
e), c wherein
mAnd c
eThe concentration of representing methane and ethane respectively, output z is the weighted sum of the most contiguous three samples of input coordinate:
z＝w
1z
1+w
2z
2+w
3z
3 (2)
Its weight w
i(i=1,2,3) can be determined by the input coordinate and the distance of the most contiguous three samples:
In the formula
The distance of the input coordinate of i sample in expression input coordinate and the most contiguous three samples.For example, for the input coordinate in the accompanying drawing 3 (4,1.5), the input coordinate of its three adjacent sample is respectively (0,0), (0,2) and (10,2), so can calculate w by formula (3)
1=0.3611, w
2=0.3827 and w
3=0.2562.Because the output of these three adjacent sample is respectively 0 ,-0.40989 and-0.65204.So the output that input coordinate (4,1.5) is corresponding can calculate-0.32157 by formula (2).Adopting and use the same method, can be-0.22803 in the hope of the value of input coordinate (4,1.5) at wave number 3028.0 places.So, corresponding v
pCalculate-0.22803-(0.32157)=0.09354 by formula (1c).By that analogy, can calculate input coordinate (4,1.5) characteristic of correspondence variate-value v
mAnd v
e
4) sample set generates
Allow the input component, just methane concentration and ethane concentration begin to increase by a fixed step size from detecting lower limit in sensing range, calculate the v of each input coordinate correspondence
m, v
eAnd v
pInput coordinate is write a matrix:
I＝[C
m，C
e，C
p] (4)
C in the formula
m=[c
M1, c
M2..., c
MN]
T, C
e=[c
E1, c
E2..., c
EN]
T, C
p=[c
P1, c
P2..., c
PN]
T, represent the concentration value vector of methane, ethane and propane respectively, wherein N represents the logarithm of sample in the sample set.A matrix is write in the characteristic variable input:
V＝[V
m，V
e，V
p] (5)
V in the formula
m=[v
M1, v
M2..., v
MN]
T, V
e=[v
E1, v
E2..., v
EN]
T, V
p=[v
P1, v
P2..., v
PN]
T, represent respectively characteristic variable v
m, v
eAnd v
pVector.
5) analytical model makes up
Analytical model is to be input with the characteristic variable, and object gas concentration is the model of output.In this example, model is input as three characteristic variables that formula (1) provides, and output is respectively methane concentration and ethane concentration.Constructed model can be the BP neural network, also can be the RBF neural network, and structure as shown in Figure 5.Because two pairwise correlations between three input vectors, so the hidden layer start node number of neural network can be made as 6.Owing to spectrum to the sensitivity of gas concentration generally along with the increase of concentration reduces, the characteristic variable of being extracted is by the combining of the value of some spectral lines, and therefore has same characteristic.So for the BP neural network, the response function of hidden node selects the S type function proper.
6) analytical model training
The training of analytical model is a determining step 5) weights and the threshold parameter of constructed analytical model.Owing among the Matlab Neural Network Toolbox is arranged, therefore can directly realize the training of analytical model by this tool box.
For the training of BP neural network, can in Matlab, finish by following source program:
Net=newff (V, C, 6, ' tansig ', ' purelin ' }); % creates analytical model, and model name is net.
Net.trainParam.epochs=500; It is 500 that % is provided with training algebraically, also bigger value can be set.
Net.trainParam.goal=0.000001; % is provided with the network training target, and it can be set according to the user,
The square error of acquiescence training sample, just experience error.
In the process of training, training objective reaches requirement or instruction
Practice algebraically and reach setting value, network training finishes.
[net, TR, Y, E]=train (net, V, C); % training analysis model.In the output parameter, net represents to train
Network; Comprise training record among the TR, comprising instruction
Practice parameters such as generation, experience deviation and study step-length; Y is
The actual output of the training sample of network; E network training sample
Actual output and the deviation between the desired output;
If Matlab creates network and training for the first time, then after training process finishes, check that TR.perf can see, the experience error of analytical model is about 7.5 * 10
-3Because the initial value of the experience sum of errors parameter that neural network is last is closely related, and the initial value of neural network parameter produces at random.Therefore, single training is determining step 5 fully) whether constructed analytical model can meet the demands, and needs multiple authentication.Re-enter above-mentioned code, once more training.If continuous 5 times all can't meet the demands, then increase a hidden node.If training just meets the demands for the first time, then reduce by a hidden node, then training again.
7) structure of analytical model calculating formula
For this embodiment, 3 input: v are arranged
m, v
eAnd v
p, 6 hidden nodes, output layer have two nodes, the concentration of namely for methane and ethane.After training finishes, for step 5) analytical model structure net that set up, that step 6) trains, creation analysis Model Calculation expression formula as follows:
(1) calculates the hidden node input
In
m＝net.IW{1}×Input
S+net.b{1} (6)
Input in the formula
SBe input vector, and Input
S=[v
m, v
e, v
p]
TNet.IW{1} is the connection weight value matrix between input layer and the hidden layer in the analytical model, and it is one 6 * 3 a matrix; Net.b{1} is 6 * 1 hidden node threshold vector; In
mThe row that are 6 hidden nodes are to input vector.
(2) calculate hidden node output
Out
m(i)＝f
m(In
m(i))) (7)
F in the formula
m() is the hidden layer response function, and it is the tansig function in this example, and for tansig (In), it embodies formula and is: 2/ (1+exp (2 * In))-1; Out
mBe the hidden layer output vector.
(3) calculate output layer output
In
O＝net.LW{2}×Out
m+net.b{2} (8)
Net.LW{2} is 2 * 6 hidden layer and the connection weight value matrix between the output layer in the formula; Net.b{2} is 2 * 1 output layer threshold vector; In
OBe the output vector of analytical model, the concentration of methane and ethane just.
Above content is to further describing that the present invention did in conjunction with concrete preferred implementation; can not assert that the specific embodiment of the present invention only limits to this; for the general technical staff of the technical field of the invention; without departing from the inventive concept of the premise; can also make some simple deduction or replace, all should be considered as belonging to the present invention and determine scope of patent protection by claims of being submitted to.
Claims (4)
- One kind based on just, the multicomponent gas quantitative spectrochemical analysis scaling method of inversion model, it is characterized in that: at first be that input, each spectral line value of spectrogram are output with the gas concentration, utilize uncorrelated characteristic between each component gas concentration, make up rectangular coordinate system, and in this rectangular coordinate system, adopt curve interpolation method and surface interpolation method to set up the spectral analysis positive model, and adopt positive model to calculate the spectrogram of the gas concentration correspondence between the initial sample point, obtain appended sample; Make up each component gas analytical model with neural network then, and train each component gas analytical model, obtain the parameter of analytical model with the sample set training sample that initial sample of making and appended sample constitute.
- 2. a kind of according to claim 1 based on just, the multicomponent gas quantitative spectrochemical analysis scaling method of inversion model, it is characterized in that, comprise the steps:(1) sample gas spectrum obtains: the L component target gas of at first determining required analysis, and the M that may exist in the combination gas to be analyzed group is disturbed gas, concentration range and proportioning thereof according to this L+M group gas, make the sample gas of some according to certain interval, and these sample gas are full of the spectrometer air chamber successively, use the spectrometer scanning optical spectrum, obtain the spectrogram of this sample gas;(2) characteristic variable is extracted: from spectrogram N bar spectral line, extract S characteristic variable, and S 〉=L+M, these characteristic variables are certain value that obtains of the combinatorial operation of spectral line value, some spectral line value of certain bar spectral line or the area of a certain wave number section; Each characteristic variable is to a kind of sensitivity maximum in L+M kind target gas or the interference gas, and is less relatively to the sensitivity of other gas, and the smaller the better;(3) space makes up with appended sample and makes: regard S every related spectral line value of characteristic variable as a signal of sensor, regard each component concentrations in the gas to be analyzed as influence this output input signal; Like this, L+M component concentrations value in certain gas to be analyzed, and the spectral line value of a spectral line, can be regarded as a point in L+M+1 dimension rectangular coordinate system space, allow this L+M component gas change in whole analyst coverage, then the spectral line value of this article one spectral line is exactly a curved surface in L+M+1 dimension rectangular coordinate system space; Between two or more consecutive point of this curved surface, selected one group of new concentration value carries out interpolation arithmetic, so just obtain a new spectral line value of this spectral line as input; With this group concentration value is input, increase the spectral line sequence number successively, and this spectral line value carried out interpolation arithmetic as output, so just obtain N new spectral line value, this N new spectral line value constituted a new spectrogram, and spectrogram that this is new and corresponding N component concentrations value thereof are referred to as an appended sample;(4) generation of sample set: change each concentration of component value in the gas to be measured, repeating step (3) then can obtain many arbitrarily appended sample, and the gas concentration value of appended sample correspondence also is to set arbitrarily; The sample of making and just constitute a sample set P by the appended sample that method of interpolation obtains; Array mode according to each characteristic variable in the step (2) makes up spectral line among the sample set P; So the characteristic variable vector has constituted a new sample set Q again with target gas concentration value;(5) analytical model makes up: to the analytical model of each structure in the L kind target gas, this model is BP neural network or RBF neural network; The input of model is the vector that S characteristic variable forms, and output is the vector that the concentration of certain or all target gas is formed; The hidden node number of neural network can be set according to the quantity of input vector and the degree of correlation between the vector; Input vector is many more, and the degree of correlation between the input vector is big more, and required the number of hidden nodes is just many more; Initial hidden node quantity should be greater than S+T, and wherein T represents the quantity of two pairwise correlations between the input vector;(6) analytical model training: the analytical model of training each target gas with the sample set Q in the step (4); If model can't meet the demands, promptly the experience error is enough little, then returns (5), revises model parameter, and the duplicate step of laying equal stress on is up to satisfying training requirement; Described model parameter is neural network hidden node number or hidden node response function; If the experience error meets the demands, reduce a hidden node, repeat this step, can't meet the demands up to the experience error, the model of getting last training is as the final analysis model;(7) structure of analytical model calculating formulaAfter training analytical model, according to the parameter and the response function of analytical model, rebuild the calculation expression of analytical model, can obtain complete analytical model.The model of setting up for step (5), creation analysis Model Calculation expression formula as follows:(a) calculate the hidden node inputIn m＝M input×Input S+b mM in the formula InputBe the connection weight value matrix between input and the hidden node, Input SFor having the input vector value of S component, In mInput vector for hidden node; b mThreshold vector for hidden node.(b) calculate hidden node outputOut m(i)＝f m(In m(i)))F in the formula m() is the hidden layer response function; Out mBe the hidden layer output vector.(c) calculate output layer outputIn O＝M output×Out m+b OM in the formula OutputBe the weight matrix that links between hidden layer and the output layer; b OBe the output layer threshold vector; In OBe the output layer input vector.Because what output layer adopted is linear response function, so the output vector of output layer is In O
- As described in the claim 2 a kind of based on just, the multicomponent gas quantitative spectrochemical analysis scaling method of inversion model, it is characterized in that: sample gas is calibrating gas in the described step (1).
- As described in the claim 2 a kind of based on just, the multicomponent gas quantitative spectrochemical analysis scaling method of inversion model, it is characterized in that: described neural network hidden node response function is index function or s type function; Described exponential function comprises logarithmic s type function and tangential type s type function, and its training method is method of steepest descent, momentum method, climbing method or genetic algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010102708223A CN101949826B (en) | 2010-09-02 | 2010-09-02 | Positive model and inverse model-based quantitative spectrometric analysis and calibration method of multi-component gas |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010102708223A CN101949826B (en) | 2010-09-02 | 2010-09-02 | Positive model and inverse model-based quantitative spectrometric analysis and calibration method of multi-component gas |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101949826A true CN101949826A (en) | 2011-01-19 |
| CN101949826B CN101949826B (en) | 2012-09-05 |
Family
ID=43453401
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2010102708223A Expired - Fee Related CN101949826B (en) | 2010-09-02 | 2010-09-02 | Positive model and inverse model-based quantitative spectrometric analysis and calibration method of multi-component gas |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101949826B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102678100A (en) * | 2012-03-21 | 2012-09-19 | 西安交通大学 | Remote determinator for well head gas |
| CN104535528A (en) * | 2014-11-26 | 2015-04-22 | 东南大学 | Method for real time extraction of TDLAS gas absorption spectrum absorbance by BP neural network |
| CN104833649A (en) * | 2015-02-26 | 2015-08-12 | 内蒙古科技大学 | Method of detecting pollutants with computer-assisted Fourier transform infrared spectroscopy |
| CN109612686A (en) * | 2018-12-06 | 2019-04-12 | 华中科技大学 | A kind of dispersion confocal measuring apparatus scaling method neural network based |
| CN111289455A (en) * | 2020-03-25 | 2020-06-16 | 欧梯恩智能科技(苏州)有限公司 | Distributed peculiar smell evaluation method and device, terminal and readable storage medium |
| CN112413855A (en) * | 2020-11-17 | 2021-02-26 | 珠海格力电器股份有限公司 | Air conditioning method, device and system |
| CN115860056A (en) * | 2023-02-17 | 2023-03-28 | 电子科技大学 | Sensor array neural network method for mixed gas concentration prediction |
| CN116519622A (en) * | 2023-02-03 | 2023-08-01 | 湖北工业大学 | Complex mixed gas detection device and method based on optical path adjustable spectrum detection |
| CN117517240A (en) * | 2024-01-08 | 2024-02-06 | 新仟意能源科技(成都)集团有限责任公司 | Light hydrocarbon component on-line detection method and system based on infrared light |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040035183A1 (en) * | 2000-01-25 | 2004-02-26 | The State Of Oregon Acting And Through The State Board Of Higher Education On Behalf Of Portl | Method and apparatus for sample analysis |
| CN1204375C (en) * | 2003-07-23 | 2005-06-01 | 西安交通大学 | Multi dimension curved surface disintegrating and interpolating subsidiary specimen method in multiple sensor system indentification |
| WO2006126011A2 (en) * | 2005-05-27 | 2006-11-30 | University Of Strathclyde | Spectral nose |
| CN101059425A (en) * | 2007-05-29 | 2007-10-24 | 浙江大学 | Method and device for identifying different variety green tea based on multiple spectrum image texture analysis |
| CN101750145B (en) * | 2009-12-02 | 2011-03-09 | 中国科学院高能物理研究所 | Spectrum calibration method for variable-included-angle monochromator |
-
2010
- 2010-09-02 CN CN2010102708223A patent/CN101949826B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040035183A1 (en) * | 2000-01-25 | 2004-02-26 | The State Of Oregon Acting And Through The State Board Of Higher Education On Behalf Of Portl | Method and apparatus for sample analysis |
| CN1204375C (en) * | 2003-07-23 | 2005-06-01 | 西安交通大学 | Multi dimension curved surface disintegrating and interpolating subsidiary specimen method in multiple sensor system indentification |
| WO2006126011A2 (en) * | 2005-05-27 | 2006-11-30 | University Of Strathclyde | Spectral nose |
| CN101059425A (en) * | 2007-05-29 | 2007-10-24 | 浙江大学 | Method and device for identifying different variety green tea based on multiple spectrum image texture analysis |
| CN101750145B (en) * | 2009-12-02 | 2011-03-09 | 中国科学院高能物理研究所 | Spectrum calibration method for variable-included-angle monochromator |
Non-Patent Citations (2)
| Title |
|---|
| 《光谱学与光谱分析》 20100331 李玉军 等 基于粒子群优化的最小二乘支持向量机在混合气体定量分析中的应用 774-778 1-4 第30卷, 第3期 2 * |
| 《西安交通大学学报》 20030831 汤晓君 等 用多维曲面插值法改善多组分气体临界值识别的鲁棒性 795-799 1-4 第37卷, 第8期 2 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102678100A (en) * | 2012-03-21 | 2012-09-19 | 西安交通大学 | Remote determinator for well head gas |
| CN102678100B (en) * | 2012-03-21 | 2014-11-05 | 西安交通大学 | Remote determinator for well head gas |
| CN104535528A (en) * | 2014-11-26 | 2015-04-22 | 东南大学 | Method for real time extraction of TDLAS gas absorption spectrum absorbance by BP neural network |
| CN104833649A (en) * | 2015-02-26 | 2015-08-12 | 内蒙古科技大学 | Method of detecting pollutants with computer-assisted Fourier transform infrared spectroscopy |
| CN109612686A (en) * | 2018-12-06 | 2019-04-12 | 华中科技大学 | A kind of dispersion confocal measuring apparatus scaling method neural network based |
| CN109612686B (en) * | 2018-12-06 | 2020-01-14 | 华中科技大学 | Calibration method of dispersion confocal measuring device based on neural network |
| CN111289455A (en) * | 2020-03-25 | 2020-06-16 | 欧梯恩智能科技(苏州)有限公司 | Distributed peculiar smell evaluation method and device, terminal and readable storage medium |
| CN112413855A (en) * | 2020-11-17 | 2021-02-26 | 珠海格力电器股份有限公司 | Air conditioning method, device and system |
| CN116519622A (en) * | 2023-02-03 | 2023-08-01 | 湖北工业大学 | Complex mixed gas detection device and method based on optical path adjustable spectrum detection |
| CN116519622B (en) * | 2023-02-03 | 2023-10-10 | 湖北工业大学 | Complex mixed gas detection device and method based on optical path adjustable spectrum detection |
| CN115860056A (en) * | 2023-02-17 | 2023-03-28 | 电子科技大学 | Sensor array neural network method for mixed gas concentration prediction |
| CN117517240A (en) * | 2024-01-08 | 2024-02-06 | 新仟意能源科技(成都)集团有限责任公司 | Light hydrocarbon component on-line detection method and system based on infrared light |
| CN117517240B (en) * | 2024-01-08 | 2024-03-19 | 新仟意能源科技(成都)集团有限责任公司 | Light hydrocarbon component on-line detection method and system based on infrared light |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101949826B (en) | 2012-09-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101949826B (en) | Positive model and inverse model-based quantitative spectrometric analysis and calibration method of multi-component gas | |
| Bigi et al. | Performance of NO, NO 2 low cost sensors and three calibration approaches within a real world application | |
| CN101881727B (en) | Multicomponent gas concentration quantitative analysis method based on absorption spectrogram reconstruction | |
| Esposito et al. | Dynamic neural network architectures for on field stochastic calibration of indicative low cost air quality sensing systems | |
| Norton et al. | Extending OSSE beyond numerical weather prediction to new areas in Earth observing science | |
| Gandomi et al. | A hybrid computational approach to derive new ground-motion prediction equations | |
| Mahecha et al. | Comparing observations and process‐based simulations of biosphere‐atmosphere exchanges on multiple timescales | |
| CN104020135B (en) | Calibration model modeling method based near infrared spectrum | |
| CN102608061B (en) | Improved method for extracting Fourier transformation infrared spectrum characteristic variable of multi-component gas by aid of TR (Tikhonov regularization) | |
| CN103913432A (en) | Near infrared spectrum wavelength selecting method based on particle swarm optimization | |
| CN102680425B (en) | Multiple analysis model information fusion method for multicomponent gas Fourier transform spectral analysis | |
| Hobbs et al. | Simulation-based uncertainty quantification for estimating atmospheric co _2 from satellite data | |
| CN109521437B (en) | Multispectral laser radar wavelength selection method for vegetation biochemical parameter detection | |
| CN114021436A (en) | Near-surface ozone inversion method based on near-surface ultraviolet radiation | |
| Rudd et al. | An inverse method for determining source characteristics for emergency response applications | |
| Si et al. | Hierarchical temperature imaging using pseudoinversed convolutional neural network aided TDLAS tomography | |
| Manepalli et al. | Emulating numeric hydroclimate models with physics-informed cGANs | |
| CN114429077A (en) | Time sequence multi-scale analysis method based on quantum migration | |
| Viskari et al. | Calibrating the soil organic carbon model Yasso20 with multiple datasets | |
| Kaminski et al. | Reviews and syntheses: guiding the evolution of the observing system for the carbon cycle through quantitative network design | |
| Mekhaimr et al. | Sources of uncertainty in atmospheric dispersion modeling in support of Comprehensive Nuclear–Test–Ban Treaty monitoring and verification system | |
| CN111089850B (en) | Multi-component concentration estimation method based on single-component absorption spectrum | |
| CN111272680B (en) | Multi-component concentration distribution estimation method based on single-component absorption spectrum | |
| CN105241823B (en) | Coal steam-electric plant smoke quantitative analysis method of spectrum based on rarefaction representation | |
| Busby et al. | A workflow for decision making under uncertainty |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20120905 |