CN101576485A - Analytical method of multi-source spectrum fusion water quality - Google Patents

Abstract

The invention discloses an analytical method of multi-source spectrum fusion water quality. The method takes an ultraviolet/visible absorption spectrum signal and a multi-dimensional fluorescence emission spectrum signal of a water sample to be tested as input, and respectively adopts independent component pretreatment method to extract characteristic signals, remove interference information and obtain characteristic signals of the spectrums. According to the contribution of the characteristic signals of the spectrums to a water quality analyzing model, and aiming at avoiding information hiding, the combined characteristic signal number of the two types of spectrums is determined, the characteristic signals are combined into a spectrum characteristic signal federated data set, and the optimal spectrum configuration combination is solved. Based on an improved support vector machine algorithm, a Boosting method is adopted for modeling, and the optimal spectrum fusion water quality analyzing model is obtained by combining a plurality of modeling results. Based on the spectrum characteristic signal federated data set, the computation model is adopted to compute the comprehensive organic substance pollution index value of the water sample to be tested. The method has the remarkable advantages of high analyzing accuracy, high analysis speed, no pollution of chemical agent, simple operation and maintenance and the like.

Description

A kind of analytical method of multi-source spectrum fusion water quality

Technical field

The invention belongs to technical field of resource environments, be used for the fast detecting of water-quality guideline, be meant a kind of analytical method of multi-source spectrum fusion water quality especially.

Background technology

Comprehensive organic index (as TOC, COD, BOD, DOC, PAHs and permanganate index) is to weigh the important state definiteness mark of water pollution degree.

Traditional chemical analysis method is with a long history, technology maturation, analysis precision height, reliable results.But its analytical cycle length, secondary pollution, need shortcomings such as specialty analysis personnel operation, make this method or do not have to use under the situations such as specialty analysis personnel in the open air, more can cause second environmental pollution.

Have remarkable advantages such as no chemical reagent pollutes, analysis speed is fast based on the comprehensive organic index analysis method of the water quality of spectral analysis, developed fast in recent years.Existing these class methods mainly are based on ultraviolet/visible (UV/Vis) absorption spectrum or fluorescence excitation spectrum.

Ultra-violet absorption spectrum is used to estimate that organism comprehensive parameters such as TOC, COD have a lot of achievements in research, and existing portable analytical equipment comes into the market.China had issued uv absorption water quality automatic on-line monitoring instrument technical requirement (HJ/T 191-2005) in 2005, when pointing out that in water quality monitoring absorbance and COD or permanganate index have correlativity, UV absorption spectrum absorbance can be converted to COD or permanganate index, for road has been paved in the application of such optical analysis method in water analysis.But in fact,, disturb very big to absorption spectrum because inorganic suspension has stronger absorptivity to the UV/Vis light source; As adopt isolated by filtration, and can remove suspension simultaneously again, can't monitor true pollution condition; In addition, the UV absorptiometry is bigger to the extremely low organism analytical error of specific absorption coefficient, only is applicable to the water sample object that change of component is little.These shortcomings hinder the widespread use of water analysis instrument in extensive environmental monitoring system based on the UV absorption spectrum.

Fluorometry comprises synchronous fluorimetry, three-dimensional fluorescence method, derivative fluorescence method, time-resolved fluorescence method, differentiates fluorescence method, fluorescence kinetics method etc. mutually, what wherein have outstanding performance is that three-dimensional excites-the emitting fluorescence method, can intactly describe the fluorescent characteristics of material, be known as the spectral fingerprint technology of fluorescent material.American scholar utilized three-dimensional fluorescence that the dissolved organic carbon of NJ USA rainwash is monitored in 2004, and effect is better than single wavelength UV absorbance method [6].Beginning in 2003, the Chinese scholar also launched the research that fluorescence method is used for water analysis in succession, and organism has humic acid, dissolved organic matter, chlorophyll-a in the branch bleed.The shortcoming of fluorometry is easily affected by environment, there are labile factors such as cancellation, self-absorption, interior optical filtering in fluorescence itself, the spectral information (such as Raman spectrum) that directly except that fluorescence information, has also comprised water itself in the fluorescence data of measuring, and some other interference and noise information.

For this reason, the subject matter that solves in the subject study comprises the pre-service of high strength interfere information in the analysis water-like spectrum measuring data, the ultraviolet/visible absorption spectra of analysis water-like and the information fusion and the rapid modeling algorithm of multidimensional fluorescence emission spectrum, utilizes the organism parameter index of model solution water sample to be measured.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, a kind of analytical method of multi-source spectrum fusion water quality is provided.

Analytical method of multi-source spectrum fusion water quality comprises the steps:

1),, produces ultraviolet/visible absorption spectra signal respectively by uviol lamp and integrated semiconductor laser irradiation for given water sample $x_1=\{x_{\mathrm{1,1}},x_{\mathrm{1,2}},\·\·\·,x_{1,s_1}\}$ With multidimensional fluorescence emission spectrum signal

$x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\·\·\·,x_{2,s_2}\};$

2) adopt the statistical independence that comprises higher order statistical information to weigh the characteristic signal of extraction and the proximity of reference signal, by dual independent component analysis, structure denoise algorithm IICA-R carries out feature extraction to the training sample that is used for model modeling;

3) ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum signal with water sample is input, uses denoise algorithm, obtains the spectrum characteristics signal;

4) adopt k-to roll over the calculated performance of cross validation method evaluate spectra analytical model, find the solution best spectral signature signal configures combination;

5) construct algorithm of support vector machine LS-SVM fast, obtain the basic model of water analysis method;

6) adopting the Boosting method, is basic modeling algorithm with the LS-SVM algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges the water analysis model;

7) adopt spectrum to merge the water analysis model, on the basis of spectral signature combined signal data set, calculate the comprehensive organic pollutants desired value of water sample to be measured.

Described employing comprises the statistical independence of higher order statistical information and weighs the characteristic signal of extraction and the proximity of reference signal, by dual independent component analysis, structure denoise algorithm IICA-R, the training sample that is used for model modeling is carried out characteristic extraction step: adopt the statistical independence that comprises higher order statistical information to weigh the characteristic signal of extraction and the proximity of reference signal, use independent component to analyze ICA raw sample data X is extracted d characteristic signal: F=BX ^TConstitute set of variables F=[f ₁..., f _d] ^TBe the given new set of variables V=[F of organism overall target chemical analysis value y composition with F and reference signal then, y] ^T=[v ₁..., v _D+1] ^T, and V reused ICA, and: S=WV; Because at d characteristic signal f ₁..., f _dIn, exist a characteristic signal relevant at least with reference signal y, therefore ICA can only extract d independent component at the most for the second time; If the separation matrix that the second time, ICA tried to achieve is W=[w _Ij] ∈ R ^{D * (d+1)}, the new independent component of extraction is S=[s ₁..., s _d] ^TDefine independence measure index vector a: T=[t ₁..., t _i... t _d] ^T, $t_i=\underset{k=1,\·\·\·,d}{\arg \max }\left\{{\stackrel{\‾}{w}}_{\mathrm{ik}}\right\}$ Set independence metric threshold θ ∈ (0,1); If

t _g＞θ then thinks individual features variable f _gWith y approach separate, can be with f _gRegard the interfere information irrelevant as and rejected that remaining characteristic variable can be thought merging the effective characteristic information of modeling with merging modeling.

Described ultraviolet/visible absorption spectra and multidimensional fluorescence emission spectrum signal with water sample is input, uses denoise algorithm, obtains the spectrum characteristics signals step: with the ultraviolet/visible absorption spectra signal of water sample $x_1=\{x_{\mathrm{1,1}},x_{\mathrm{1,2}},\·\·\·,x_{1,s_1}\}$ With multidimensional fluorescence emission spectrum signal $x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\·\·\·,x_{2,s_2}\}$ Be input, use ICA tentatively to extract characteristic signal to X, order:

F＝B·X＝[f ₁，…，f _d] ^T

B ∈ R wherein ^{D * p}The spectral signature separation matrix that obtains when using ICA for the first time for the spectrum water analysis model modeling stage, the characteristic signal of F for therefrom just selecting; Independence metric threshold θ is set, and calculates independence measure index t _i, i=1,2 ..., d.All are satisfied t _k＜θ, k=k ₁, k ₂..., k _n, the n of a n≤d characteristic signal constitutes water sample spectral signature signal vector $F^{*}={[{\mathrm{<figure-callout\; id="1"\; label="f\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">f</figure-callout>}}_{k_{}},\&CenterDot;\&CenterDot;\&CenterDot;,f_{k_n}]}^T.$

The calculated performance of described employing k-folding cross validation method evaluate spectra analytical model, find the solution best spectral signature signal configures combination step: be located at n and be used for setting up the sample that spectrum merges the water analysis model, the ultraviolet/visible absorption spectra of k water sample and the characteristic signal set of multidimensional fluorescence emission spectrum extraction are respectively $z_{k1}\left(n_1\right)=\{{\mathrm{<figure-callout\; id="1"\; label="z\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="z\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>\; </figure-callout>}}_k,{\mathrm{<figure-callout\; id="1"\; label="z\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>}}_k,\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{<figure-callout\; id="1"\; label="z\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>}}_k\}$ With $z_{k2}\left(n_2\right)=\{{\mathrm{<figure-callout\; id="1"\; label="z\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>}}_k,z_{k\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,z_{k2,n_2}\},$ N wherein ₁And n ₂Be respectively uv-visible absorption spectra and multidimensional fluorescence emission spectrum characteristic signal quantity; K the spectrum fusion water analysis model input vector that two kinds of spectral signature combined signals are constituted is z _k(n ₁, n ₂)=[z _K1(n ₁), z _K2(n ₂)] ^T, remember that the water-quality guideline chemical analysis value of this water sample is y _kBe located at sample set Θ (n ₁, n ₂)={ z _k(n ₁, n ₂), y _k} _{K=1,2 ..., n}The basis on, the spectrum of foundation merges the water analysis model and is:

$\hat{y}=f(z,\mathrm{\&Theta;}({\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">n</figure-callout>}}_1,n_2))$

Adopt the calculated performance of k-folding cross validation method assessment analysis of spectrum model, the set that is about to n sample be divided into randomly k mutually disjoint, big or small subclass about equally, set up the water analysis model with k-1 subclass wherein, utilize the performance of the analytical error root mean square assessment models of last remaining subclass.Repeat k time according to above process, so each subclass all has an opportunity to test, estimate to expect extensive error according to the mean value RMSEP of the RMSEP that obtains after k the iteration; For given n modeling water sample, obviously spectrum merges water analysis modular form RMSEP index and n ₁And n ₂Relevant; Therefore, can select best characteristic signal configuration combination by finding the solution of following optimization problem: $\underset{n_{}}{\mathrm{<figure-callout\; id="1"\; label="min\; n"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">min</figure-callout>}}\stackrel{\&OverBar;}{\mathrm{RMSEP}}({\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">n</figure-callout>}}_1,n_2)$

Described utilization is algorithm of support vector machine LS-SVM fast, obtain the basic model step of water analysis method: with the ultraviolet/visible absorption spectra of water quality sample and the associating characteristic signal vector of multidimensional fluorescence emission spectrum is the input sample, with corresponding water sample organism overall target chemical analysis value is output sample, at sample set { z _k, y _k} _{K=1,2 ..., n}The basis on the LS-SVM algorithm can provide the water analysis basic model of following form:

$y=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{k}K(z,z_k)+b$

Wherein z is the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of water sample to be analyzed, z _kBe the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of training water sample, y is the Model Calculation value of given organism overall target, and K is the kernel function that satisfies the Mercer condition.

Described employing Boosting method is basic modeling algorithm with the LS-SVM algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges water analysis model step:

(1) imports the sample set (z of nominalization ₁, y ₁) ..., (z _n, y _n); And initialization r=y, combination regression model F _c=0, the combination regression model is to the match value of r ${\hat{r}}_c=0,$ Iterations m=1;

(2) to data set (z ₁, r ₁) ..., (z _n, r _n) use the modeling of LS-SVM algorithm, obtain basic regression model:

$F\left(z\right)=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{m,k}K(z,z_k)+b_m$

A wherein _{M, k}(k=1 ..., n) and b _mBy following Solving Linear:

${{\stackrel{\&RightArrow;}{1}}_n}^T{a}_m=0$

${{\stackrel{\&RightArrow;}{1}}_n}^T{b}_m+(\mathrm{\&Omega;}+{\mathrm{\&gamma;}}^{-1}I)a_m=r$

Wherein ${\stackrel{\&RightArrow;}{1}}_n={[1,\&CenterDot;\&CenterDot;\&CenterDot;,1]}^T,$ a _m＝[a _m1，…，a _mn] ^T，

k，l＝1，…，n，K(x _k，x _l)

Be kernel function;

(3) calculate the weights β of basic regression model in this iteration _m:

${\mathrm{\&beta;}}_m=\frac{\mathrm{\&epsiv;}{\left|\right|r\left|\right|}_2}{{\left|\right|\hat{r}\left|\right|}_2}$

Wherein

Be the match value of basic regression model F to r, and

$\mathrm{\&epsiv;}=\frac{r\&CenterDot;\hat{r}}{{\left|\right|r\left|\right|}_2{\left|\right|\hat{r}\left|\right|}_2}$

(4) upgrade combination regression model F _c

$F_c\left(z\right)\&DoubleLeftArrow;F_c\left(z\right)+{\mathrm{\&beta;}}_m\&CenterDot;F\left(z\right)$

(5) upgrade $r=y-{\hat{r}}_c,$ Wherein

Be combination regression model F _cMatch value to r;

(6) calculate iteration stopping desired value C _m:

$C_m=\frac{1}{n}{\left|\right|y-{\hat{r}}_c\left|\right|}^2+\frac{1}{2}|{\left|\right|\frac{1}{n}{\mathrm{\&beta;}}_m\underset{m}{\mathrm{\&Sigma;}}{a}_m\left|\right|}^2$

If satisfy iteration stopping criterion C _m＞C _M-1, (m＞1) then algorithm stops; Otherwise return step (2), just can obtain best model.

The comprehensive organic pollutants desired value step of described calculating water sample to be measured: calculate the comprehensive organic pollutants desired value y of water sample to be measured on the basis of spectral signature combined signal data set z, computing formula is as follows:

$y=\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}\{{\mathrm{\&beta;}}_m\&CenterDot;[\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{\mathrm{mk}}K(z,z_k)+b_m\left]\right\}$

Wherein z is the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of water sample to be analyzed, z _kBe the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of training water sample, y is the Model Calculation value of given organism overall target, and K is the kernel function that satisfies the Mercer condition, wherein β _mBe the basic model weights that obtain after the m time Boosting iteration, a and b are model constants.

The present invention has set up mathematical model accurately, has realized that UV, visible light (UV/Vis) spectral analysis and emission spectrographic analyses such as fluorescence, Raman unite two into one, and information fusion has improved analysis precision.Can satisfy the on-the-spot express-analysis demand in daily and the outburst surroundings pollution monitoring work, can be used for the development research of corresponding instrument instrument.

Description of drawings

Fig. 1 is the analytical method of multi-source spectrum fusion water quality process flow diagram;

Fig. 2 is the Model Calculation value of multi-source optical spectrum convergence analysis model and the comparison synoptic diagram of measured value;

Fig. 3 is the Model Calculation value of multi-source optical spectrum convergence analysis model and the degree of correlation synoptic diagram of measured value.

Embodiment

Water analysis of the present invention is to liking surface water.Because the sampling position is changeable, also variation is bigger along with contaminated situation is different for the composition of water sample and each components contents.From the analytic target angle, there is the interference of suspension, colloidal materials and inorganic ions in the water sample to be analyzed, can have a strong impact on analysis precision.And these interference can not physical separation be removed, and also do not have clear and definite statistical nature, can't remove by general filter, must adopt other effective means to reduce and disturb the influence that brings.Existing denoising method or priori that need be more, there is irrationality in the setting of perhaps denoising mechanism, is not suitable for water quality spectral analysis research object; And feature extracting method is not to design for signal denoising specially, can not guarantee that the feature of extracting is effective to regression modeling.At this problem, the present invention proposes a kind of new denoising method that is applicable to this subject study object on the basis of ICA algorithm, be used for solving the denoising problem of water analysis higher-dimension spectroscopic data.

From the analysis tool angle, though single ultraviolet/visible absorption spectra or multidimensional fluorescence emission spectrometry can both divide the overall target of the organism in the bleed, but because existence and the intrinsic separately weak point of single spectrum disturbed are difficult to obtain desirable analysis precision.On the other hand, it is few and lack the characteristics of the priori of modeling that the water sample spectroscopic data has a sample size, therefore must select suitable modeling algorithm, to obtain the best model precision.It is theoretical at the best of small sample statistical estimate and Model Calculation study at present that Statistical Learning Theory is considered to, the support vector machine (SVM) of coming based on the Statistical Learning Theory development is to be specifically designed to a kind of modeling algorithm of small sample data, has good extensive performance, and solved the dimension disaster problem, can effectively handle high dimensional data.The LS-SVM algorithm is based on the improvement of regularization theory to standard SVM, be converted into the problem of finding the solution system of linear equations and don't change characteristics such as original kernel function mapping relations and global optimum finding the solution quadratic programming problem among the SVM, therefore when guaranteeing model performance, greatly reduce the computation complexity of SVM, have arithmetic speed faster.The present invention proposes a kind of water analysis method based on the multi-source optical spectrum information fusion, adopts the Boosting method, is basic modeling algorithm with the LS-SVM algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges the water analysis model.So, can realize the selection and the information balance of spectral information amalgamation mode, thereby improve the Model Calculation precision of organic parameter.

Analytical method of multi-source spectrum fusion water quality comprises the steps:

1),, produces ultraviolet/visible absorption spectra signal respectively by uviol lamp and integrated semiconductor laser irradiation for given water sample $x_1=\{x_{\mathrm{1,1}},x_{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,x_{1,s_1}\}$ With multidimensional fluorescence emission spectrum signal

$x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{2,s_2}\};$

2) adopt the statistical independence that comprises higher order statistical information to weigh the characteristic signal of extraction and the proximity of reference signal, by dual independent component analysis, structure denoise algorithm IICA-R carries out feature extraction to the training sample that is used for model modeling;

3) ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum signal with water sample is input, uses denoise algorithm, obtains the spectrum characteristics signal;

4) adopt k-to roll over the calculated performance of cross validation method evaluate spectra analytical model, find the solution best spectral signature signal configures combination;

5) construct algorithm of support vector machine LS-SVM fast, obtain the basic model of water analysis method;

6) adopting the Boosting method, is basic modeling algorithm with the LS-SVM algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges the water analysis model;

7) adopt spectrum to merge the water analysis model, on the basis of spectral signature combined signal data set, calculate the comprehensive organic pollutants desired value of water sample to be measured.

Described employing comprises the statistical independence of higher order statistical information and weighs the characteristic signal of extraction and the proximity of reference signal, by dual independent component analysis, structure denoise algorithm IICA-R, the training sample that is used for model modeling is carried out characteristic extraction step: adopt the statistical independence that comprises higher order statistical information to weigh the characteristic signal of extraction and the proximity of reference signal, use independent component to analyze ICA raw sample data X is extracted d characteristic signal: F=BX ^TConstitute set of variables F=[f ₁..., f _d] ^TBe the given new set of variables V=[F of organism overall target chemical analysis value y composition with F and reference signal then, y] ^T=[v ₁..., v _D+1] ^T, and V reused ICA, and: S=WV; Because at d characteristic signal f ₁..., f _dIn, exist a characteristic signal relevant at least with reference signal y, therefore ICA can only extract d independent component at the most for the second time; If the separation matrix that the second time, ICA tried to achieve is W=[w _Ij] ∈ R ^{D * (d+1)}, the new independent component of extraction is S=[s ₁..., s _d] ^TDefine independence measure index vector a: T=[t ₁..., t _i... t _d] ^T, $t_i=\underset{k=1,\&CenterDot;\&CenterDot;\&CenterDot;,d}{\arg \max }\left\{{\stackrel{\&OverBar;}{w}}_{\mathrm{ik}}\right\}$ Set independence metric threshold θ ∈ (0,1); If

t _g＞θ then thinks individual features variable f _gWith y approach separate, can be with f _gRegard the interfere information irrelevant as and rejected that remaining characteristic variable can be thought merging the effective characteristic information of modeling with merging modeling.

Described ultraviolet/visible absorption spectra and multidimensional fluorescence emission spectrum signal with water sample is input, uses denoise algorithm, obtains the spectrum characteristics signals step: with the ultraviolet/visible absorption spectra signal of water sample $x_1=\{x_{\mathrm{1,1}},x_{\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{<figure-callout\; id="1"\; label="x"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">x</figure-callout>}}_{1,s_1}\}$ With multidimensional fluorescence emission spectrum signal $x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{2,s_2}\}$ Be input, use ICA tentatively to extract characteristic signal to X, order:

F＝B·X＝[f ₁，…，f _d] ^T

B ∈ R wherein ^{D * p}The spectral signature separation matrix that obtains when using ICA for the first time for the spectrum water analysis model modeling stage, the characteristic signal of F for therefrom just selecting; Independence metric threshold θ is set, and calculates independence measure index t _i, i=1,2 ..., d.All are satisfied t _k＜θ, k=k ₁, k ₂..., k _n, the n of a n≤d characteristic signal constitutes water sample spectral signature signal vector $F^{*}={[{\mathrm{<figure-callout\; id="1"\; label="f\; k"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">f</figure-callout>}}_{k_{}},\&CenterDot;\&CenterDot;\&CenterDot;,f_{k_n}]}^T.$

The calculated performance of described employing k-folding cross validation method evaluate spectra analytical model, find the solution best spectral signature signal configures combination step: be located at n and be used for setting up the sample that spectrum merges the water analysis model, the ultraviolet/visible absorption spectra of k water sample and the characteristic signal set of multidimensional fluorescence emission spectrum extraction are respectively $z_{k1}\left(n_1\right)=\{z_{k\mathrm{1,1}},z_{k\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,z_{k1,n_1}\}$ With $z_{k2}\left(n_2\right)=\{z_{k\mathrm{2,1}},z_{k\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,z_{k2,n_2}\},$ N wherein ₁And n ₂Be respectively uv-visible absorption spectra and multidimensional fluorescence emission spectrum characteristic signal quantity; K the spectrum fusion water analysis model input vector that two kinds of spectral signature combined signals are constituted is z _k(n ₁, n ₂)=[z _K1(n ₁), z _K2(n ₂)] ^T, remember that the water-quality guideline chemical analysis value of this water sample is y _kBe located at sample set Θ (n ₁, n ₂)={ z _k(n ₁, n ₂), y _k} _{K=1,2 ..., n}The basis on, the spectrum of foundation merges the water analysis model and is:

$\hat{y}=f(z,\mathrm{\&Theta;}({\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">n</figure-callout>}}_1,n_2))$

Adopt the calculated performance of k-folding cross validation method assessment analysis of spectrum model, the set that is about to n sample be divided into randomly k mutually disjoint, big or small subclass about equally, set up the water analysis model with k-1 subclass wherein, utilize the performance of the analytical error root mean square assessment models of last remaining subclass.Repeat k time according to above process, so each subclass all has an opportunity to test, estimate to expect extensive error according to the mean value RMSEP of the RMSEP that obtains after k the iteration; For given n modeling water sample, obviously spectrum merges water analysis modular form RMSEP index and n ₁And n ₂Relevant; Therefore, can select best characteristic signal configuration combination by finding the solution of following optimization problem: $\underset{n_{}}{\mathrm{<figure-callout\; id="1"\; label="min\; n"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">min</figure-callout>}}\stackrel{\&OverBar;}{\mathrm{RMSEP}}({\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">n</figure-callout>}}_1,n_2)$

Described utilization is algorithm of support vector machine LS-SVM fast, obtain the basic model step of water analysis method: with the ultraviolet/visible absorption spectra of water quality sample and the associating characteristic signal vector of multidimensional fluorescence emission spectrum is the input sample, with corresponding water sample organism overall target chemical analysis value is output sample, at sample set { z _k, y _k} _{K=1,2 ..., n}The basis on the LS-SVM algorithm can provide the water analysis basic model of following form:

$y=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{k}K(z,z_k)+b$

Wherein z is the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of water sample to be analyzed, z _kBe the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of training water sample, y is the Model Calculation value of given organism overall target, and K is the kernel function that satisfies the Mercer condition.

Described employing Boosting method is basic modeling algorithm with the LS-SVM algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges water analysis model step:

(1) imports the sample set (z of nominalization ₁, y ₁) ..., (z _n, y _n); And initialization r=y, combination regression model F _c=0, the combination regression model is to the match value of r ${\hat{r}}_c=0,$ Iterations m=1;

(2) to data set (z ₁, r ₁) ..., (z _n, r _n) use the modeling of LS-SVM algorithm, obtain basic regression model:

$F\left(z\right)=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{m,k}K(z,z_k)+b_m$

A wherein _{M, k}(k=1 ..., n) and b _mBy following Solving Linear:

${{\stackrel{\&RightArrow;}{1}}_n}^T{a}_m=0$

${{\stackrel{\&RightArrow;}{1}}_n}^T{b}_m+(\mathrm{\&Omega;}+{\mathrm{\&gamma;}}^{-1}I)a_m=r$

Wherein ${\stackrel{\&RightArrow;}{1}}_n={[1,\&CenterDot;\&CenterDot;\&CenterDot;,1]}^T,$ a _m=[a _M1..., a _Mn] ^T,

K, l=1 ..., n, K (x _k, x _l) be kernel function;

(3) calculate the weights β of basic regression model in this iteration _m:

${\mathrm{\&beta;}}_m=\frac{\mathrm{\&epsiv;}{\left|\right|r\left|\right|}_2}{{\left|\right|\hat{r}\left|\right|}_2}$

Wherein

Be the match value of basic regression model F to r, and

$\mathrm{\&epsiv;}=\frac{r\&CenterDot;\hat{r}}{{\left|\right|r\left|\right|}_2{\left|\right|\hat{r}\left|\right|}_2}$

(4) upgrade combination regression model F _c

$F_c\left(z\right)\&DoubleLeftArrow;F_c\left(z\right)+{\mathrm{\&beta;}}_m\&CenterDot;F\left(z\right)$

(5) upgrade $r=y-{\hat{r}}_c,$ Wherein

Be combination regression model F _cMatch value to r;

(6) calculate iteration stopping desired value C _m:

$C_m=\frac{1}{n}{\left|\right|y-{\hat{r}}_c\left|\right|}^2+\frac{1}{2}|{\left|\right|\frac{1}{n}{\mathrm{\&beta;}}_m\underset{m}{\mathrm{\&Sigma;}}{a}_m\left|\right|}^2$

If satisfy iteration stopping criterion C _m＞C _M-1(m＞1) then algorithm stops; Otherwise return step (2), just can obtain best model.

The comprehensive organic pollutants desired value step of described calculating water sample to be measured: calculate the comprehensive organic pollutants desired value y of water sample to be measured on the basis of spectral signature combined signal data set z, computing formula is as follows:

$y=\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}\{{\mathrm{\&beta;}}_m\&CenterDot;[\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{\mathrm{mk}}K(z,z_k)+b_m\left]\right\}$

Wherein z is the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of water sample to be analyzed, z _kBe the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of training water sample, y is the Model Calculation value of given organism overall target, and K is the kernel function that satisfies the Mercer condition, wherein β _mBe the basic model weights that obtain after the m time Boosting iteration, a and b are model constants.

By a large amount of experiments, checked the validity of above-mentioned analytical method of multi-source spectrum fusion water quality.Be noted that: for the analysis of each given water quality organism overall target, all need on same water sample ultraviolet/visible absorption spectra and multidimensional fluorescence emission spectrum measuring-signal basis, according to corresponding organism overall target chemical analysis value, setting up independently, spectrum merges the water analysis model.But say from methodology, above-mentioned analytical method of multi-source spectrum fusion water quality merges the water analysis model to the spectrum of each given organism overall target and all is suitable for, the concrete adjustment but the various algorithm parameters that relate in modeling and the analytical calculation process then need according to different analysis purposes.

It is example that the multi-source optical spectrum of one of water quality organism overall target---TOC merges water analysis, and the experiment results about analytical method of multi-source spectrum fusion water quality validity is described.Surface water water sample with one group of Hangzhou is an example, and following table has been listed the TOC chemical analysis value of this group water sample:

The water sample numbering	TOC (mg/L)	The water sample numbering	TOC (mg/L)	The water sample numbering	TOC (mg/L)	The water sample numbering	TOC (mg/L)
								1	6.93	9	82.46	17	49.17	25	29.35
2	16.20	10	27.74	18	43.25	26	80.43
								3	14.69	11	25.67	19	29.07	27	26.94
4	11.77	12	27.79	20	14.98	28	51.07
								5	17.06	13	33.54	21	76.79	29	64.74
6	3.41	14	53.74	22	95.49	30	67.09
								7	8.04	15	78.29	23	125.30	31	63.98
8	13.10	16	54.58	24	49.46	32	59.48

Use the multi-source optical spectrum analytical approach then, simple steps is as follows:

Step 1: the spectrum that generates water sample to be measured.

For given water sample,, produce ultraviolet/visible absorption spectra and multidimensional fluorescence emission spectrum signal respectively by the light source irradiation of certain intensity and wavelength $x_1=\{x_{\mathrm{1,1}},x_{\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{<figure-callout\; id="1"\; label="x"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">x</figure-callout>}}_{1,s_1}\}$ With $x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{2,s_2}\}.$

Step 2: the spectral signature signal that extracts water sample to be measured.

To water sample ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum signal that obtains, adopt the IICA-R Signal Pre-Processing Method to extract characteristic signal respectively, remove interfere information, obtain the spectral signature signal $z_{}$ ${}_1=\{{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="z"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>\; </figure-callout>}}_{\mathrm{1,1}},{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>}}_{\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>}}_{1,n_1}\}$ With $z_2=\{{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20091009933600171.png"\; state="\{\{state\}\}">z</figure-callout>}}_{\mathrm{2,1}},z_{\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,z_{2,n_2}\}.$

Step 3: the spectral signature signal that makes up water sample to be measured.

According to the contribution of pretreated water sample ultraviolet/visible absorption spectra to be measured and multidimensional fluorescence emission spectrum characteristic signal to the water analysis model, be the definite two class spectral signature number of signals that make up of target to avoid information to cover, and be combined into spectral signature combined signal data set $z={[{\stackrel{\&OverBar;}{z}}_1,{\stackrel{\&OverBar;}{z}}_2]}^T\&Element;R^{{\stackrel{\&OverBar;}{n}}_1+{\stackrel{\&OverBar;}{n}}_2}.$

Step 4: the spectrum convergence analysis value of calculating water sample to be measured.

Adopt spectrum to merge the water analysis model, on the basis of spectral signature combined signal data set z, calculate the comprehensive organic pollutants desired value y of water sample to be measured.

Fig. 3 and Fig. 4 provide respectively adopt the IICA-R method that two kinds of spectrum independently are extracted feature and the characteristic signal number combinations of the best is set after the The model calculation of Boosting-LS-SVM institute established model, and the correlative relationship figure of Model Calculation value and assay value.As can be seen, the IICA-R method that adopts this problem to propose is extracted feature and the spectrum of setting up after the best characteristic signal number combinations is set and merges the water analysis model, Model Calculation value to water sample TOC parameter can be good at approaching the chemical measurement value, the linear dependence degree height of the two.This shows that the water analysis method that multi-source optical spectrum merges has excellent application value, is the another kind of method of analyzing water pollution, has remarkable advantages such as analysis precision height, fast, the no chemical reagent pollution of analysis speed, Operation and Maintenance be simple.

Claims (7)

1, a kind of analytical method of multi-source spectrum fusion water quality is characterized in that comprising the steps:

1),, produces ultraviolet/visible absorption spectra signal respectively by uviol lamp and integrated semiconductor laser irradiation for given water sample $x_1=\{x_{\mathrm{1,1}},x_{\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{1,s_1}\}$ With multidimensional fluorescence emission spectrum signal

$x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{2,s_2}\};$

2) adopt the statistical independence that comprises higher order statistical information to weigh the characteristic signal of extraction and the proximity of reference signal, by dual independent component analysis, structure denoise algorithm IICA-R carries out feature extraction to the training sample that is used for model modeling;

3) ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum signal with water sample is input, uses denoise algorithm, obtains the spectrum characteristics signal;

4) adopt k-to roll over the calculated performance of cross validation method evaluate spectra analytical model, find the solution best spectral signature signal configures combination;

5) construct algorithm of support vector machine LS-SVM fast, obtain the basic model of water analysis method;

6) adopting the Boosting method, is basic modeling algorithm with the LS-SVM algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges the water analysis model;

7) adopt spectrum to merge the water analysis model, on the basis of spectral signature combined signal data set, calculate the comprehensive organic pollutants desired value of water sample to be measured.

2. a kind of analytical method of multi-source spectrum fusion water quality according to claim 1, it is characterized in that the statistical independence that described employing comprises higher order statistical information weighs the characteristic signal of extraction and the proximity of reference signal, by dual independent component analysis, structure denoise algorithm IICA-R, the training sample that is used for model modeling is carried out characteristic extraction step: adopt the statistical independence that comprises higher order statistical information to weigh the characteristic signal of extraction and the proximity of reference signal, use independent component to analyze ICA raw sample data X is extracted d characteristic signal: F=BX ^TConstitute set of variables F=[f ₁..., f _d] ^TBe the given new set of variables V=[F of organism overall target chemical analysis value y composition with F and reference signal then, y] ^T=[v ₁..., v _D+1] ^T, and V reused ICA, and: S=WV; Because at d characteristic signal f ₁..., f _dIn, exist a characteristic signal relevant at least with reference signal y, therefore ICA can only extract d independent component at the most for the second time; If the separation matrix that the second time, ICA tried to achieve is W=[w _Ij] ∈ R ^{D * (d+1)}, the new independent component of extraction is S=[s ₁..., s _d] ^TDefine independence measure index vector a: T=[t ₁..., t _i... t _d] ^T, $t_i=\underset{k=1,\&CenterDot;\&CenterDot;\&CenterDot;,d}{\arg \max }\left\{{\stackrel{\&OverBar;}{w}}_{\mathrm{ik}}\right\}$ Set independence metric threshold θ ∈ (0,1); If t _g＞θ then thinks individual features variable f _gWith y approach separate, can be with f _gRegard the interfere information irrelevant as and rejected that remaining characteristic variable can be thought merging the effective characteristic information of modeling with merging modeling.

3. a kind of analytical method of multi-source spectrum fusion water quality according to claim 1, it is characterized in that described ultraviolet/visible absorption spectra and multidimensional fluorescence emission spectrum signal with water sample is input, use denoise algorithm, obtain the spectrum characteristics signals step: with the ultraviolet/visible absorption spectra signal of water sample $x_1=\{x_{\mathrm{1,1}},x_{\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{1,s_1}\}$ With multidimensional fluorescence emission spectrum signal $x_2=\{x_{\mathrm{2,1}},x_{\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{2,s_2}\}$ Be input, use ICA tentatively to extract characteristic signal to X, order:

F＝B·X＝[f ₁，…，f _d] ^T

B ∈ R wherein ^{D * p}The spectral signature separation matrix that obtains when using ICA for the first time for the spectrum water analysis model modeling stage, the characteristic signal of F for therefrom just selecting; Independence metric threshold θ is set, and calculates independence measure index t _i, i=1,2 ..., d.All are satisfied t _k＜θ, k=k ₁, k ₂..., k _n, the n of a n≤d characteristic signal constitutes water sample spectral signature signal vector $F^{*}={[f_{k_1},\&CenterDot;\&CenterDot;\&CenterDot;,f_{k_n}]}^T.$

4. a kind of analytical method of multi-source spectrum fusion water quality according to claim 1, the calculated performance that it is characterized in that described employing k-folding cross validation method evaluate spectra analytical model, find the solution best spectral signature signal configures combination step: be located at n and be used for setting up the sample that spectrum merges the water analysis model, the ultraviolet/visible absorption spectra of k water sample and the characteristic signal set of multidimensional fluorescence emission spectrum extraction are respectively $z_{k1}\left(n_1\right)=\{z_{k\mathrm{1,1}},z_{k\mathrm{1,2}},\&CenterDot;\&CenterDot;\&CenterDot;,z_{k1,n_1}\}$ With $z_{k2}\left(n_2\right)=\{z_{k\mathrm{2,1}},z_{k\mathrm{2,2}},\&CenterDot;\&CenterDot;\&CenterDot;,z_{k2,n_2}\},$ N wherein ₁And n ₂Be respectively uv-visible absorption spectra and multidimensional fluorescence emission spectrum characteristic signal quantity; K the spectrum fusion water analysis model input vector that two kinds of spectral signature combined signals are constituted is z _k(n ₁, n ₂)=[z _K1(n ₁), z _K2(n ₂)] ^T, remember that the water-quality guideline chemical analysis value of this water sample is y _kBe located at sample set Θ (n ₁, n ₂)={ z _k(n ₁, n ₂), y _k} _{K=1,2 ..., n}The basis on, the spectrum of foundation merges the water analysis model and is:

$\hat{y}=f(z,\mathrm{\&Theta;}(n_1,n_2))$

Adopt the calculated performance of k-folding cross validation method assessment analysis of spectrum model, the set that is about to n sample be divided into randomly k mutually disjoint, big or small subclass about equally, set up the water analysis model with k-1 subclass wherein, utilize the performance of the analytical error root mean square assessment models of last remaining subclass.Repeat k time according to above process, so each subclass all has an opportunity to test, estimate to expect extensive error according to the mean value RMSEP of the RMSEP that obtains after k the iteration; For given n modeling water sample, obviously spectrum merges water analysis modular form RMSEP index and n ₁And n ₂Relevant; Therefore, can select best characteristic signal configuration combination by finding the solution of following optimization problem:

5. a kind of analytical method of multi-source spectrum fusion water quality according to claim 1, it is characterized in that described utilization algorithm of support vector machine LS-SVM fast, obtain the basic model step of water analysis method: with the ultraviolet/visible absorption spectra of water quality sample and the associating characteristic signal vector of multidimensional fluorescence emission spectrum is the input sample, with corresponding water sample organism overall target chemical analysis value is output sample, at sample set { z _k, y _k} _{K=1,2 ..., n}The basis on the LS-SVM algorithm can provide the water analysis basic model of following form:

$y=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{k}K(z,z_k)+b$

Wherein z is the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of water sample to be analyzed, z _kBe the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of training water sample, y is the Model Calculation value of given organism overall target, and K is the kernel function that satisfies the Mercer condition.

6. a kind of analytical method of multi-source spectrum fusion water quality according to claim 1, it is characterized in that described employing Boosting method, with the LS-SVM algorithm is basic modeling algorithm, obtains best spectrum by the result who makes up repeatedly modeling and merges water analysis model step:

(1) imports the sample set (z of nominalization ₁, y ₁) ..., (z _n, y _n); And initialization r=y, combination regression model F _c=0, the combination regression model is to the match value of r ${\hat{r}}_c=0,$ Iterations m=1;

(2) to data set (z ₁, r ₁) ..., (z _n, r _n) use the modeling of LS-SVM algorithm, obtain basic regression model:

$F\left(z\right)=\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{m,k}K(z,z_k)+b_m$

A wherein _{M, k}(k=1 ..., n) and b _mBy following Solving Linear:

${{\stackrel{\&RightArrow;}{1}}_n}^T{a}_m=0$

${{\stackrel{\&RightArrow;}{1}}_n}^T{b}_m+(\mathrm{\&Omega;}+{\mathrm{\&gamma;}}^{-1}I)a_m=r$

Wherein ${\stackrel{\&RightArrow;}{1}}_n={[1,\&CenterDot;\&CenterDot;\&CenterDot;,1]}^T,$ a _m=[a _M1..., a _Mn] ^T,

K, l=1 ..., n, K (x _k, x _l) be kernel function;

(3) calculate the weights β of basic regression model in this iteration _m:

${\mathrm{\&beta;}}_m=\frac{\mathrm{\&epsiv;}{\left|\right|r\left|\right|}_2}{{\left|\right|\hat{r}\left|\right|}_2}$

Wherein

Be the match value of basic regression model F to r, and

$\mathrm{\&epsiv;}=\frac{r\&CenterDot;\hat{r}}{{\left|\right|r\left|\right|}_2{\left|\right|\hat{r}\left|\right|}_2}$

(4) upgrade combination regression model F _c

$F_c\left(z\right)\&DoubleLeftArrow;F_c\left(z\right)+{\mathrm{\&beta;}}_m\&CenterDot;F\left(z\right)$

(5) upgrade $r=y-{\hat{r}}_c,$ Wherein Be combination regression model F _cMatch value to r;

(6) calculate iteration stopping desired value C _m:

$C_m=\frac{1}{n}{\left|\right|y-{\hat{r}}_c\left|\right|}^2+\frac{1}{2}|{\left|\right|\frac{1}{n}{\mathrm{\&beta;}}_m\underset{m}{\mathrm{\&Sigma;}}{a}_m\left|\right|}^2$

If satisfy iteration stopping criterion C _m＞C _M-1, (m＞1) then algorithm stops; Otherwise return step (2), just can obtain best model.

7. a kind of analytical method of multi-source spectrum fusion water quality according to claim 1, the comprehensive organic pollutants desired value step that it is characterized in that described calculating water sample to be measured: calculate the comprehensive organic pollutants desired value y of water sample to be measured on the basis of spectral signature combined signal data set z, computing formula is as follows:

$y=\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}\{{\mathrm{\&beta;}}_m\&CenterDot;[\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{\mathrm{mk}}K(z,z_k)+b_m\left]\right\}$

Wherein z is the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of water sample to be analyzed, z _kBe the ultraviolet/visible absorption spectra and the multidimensional fluorescence emission spectrum associating characteristic signal vector of training water sample, y is the Model Calculation value of given organism overall target, and K is the kernel function that satisfies the Mercer condition, wherein β _mBe the basic model weights that obtain after the m time Boosting iteration, a and b are model constants.

Images