PCR/PLS: Calibration using the spectr... Log Out | Topics | Search
Moderators | Register | Edit Profile

NIR Discussion Forum » Bruce Campbell's List » Chemometrics » PCR/PLS: Calibration using the spectra of pure components « Previous Next »

Author Message
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
Intermediate Member
Username: vilasmechgmailcom

Post Number: 19
Registered: 5-2010
Posted on Monday, August 30, 2010 - 5:47 pm:   

Hello All,

Thanks a lot for your responses.

Francesco - I know that GC is a well-established technique for these types of applications. However, we are trying to find a low-cost alternative to GC. The NIR sensor will be definitely less expensive than GC, if prodeced in sufficient quantity.

Gabi - We have experimentall proved that there is significant interaction amongst the considered species and the spectra of such mixtures can be very accurately computed by taking the weighted sum of spectra of individual species. We are using the same method, as described by you, for calibration and validation.

Thanks Tony and Johan. Your comments were really useful. I am not an expert in chemometrics, so, I think, I will need some time to understand O-PLS method.

Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

Gustavo Figueira de Paula (gustavo)
Member
Username: gustavo

Post Number: 12
Registered: 6-2008
Posted on Monday, August 30, 2010 - 6:38 am:   

Hello Guys,

There's some software package free for academic use capable of OPLS calibrations? If not, a paid one, but with educational discounts?

Thanks,
Gustavo.
Top of pagePrevious messageNext messageBottom of page Link to this message

Johan Trygg (trygg)
New member
Username: trygg

Post Number: 4
Registered: 1-2006
Posted on Monday, August 30, 2010 - 3:01 am:   

Hi again,

Hi,
agree w Tonys last comment, "use the number of scores that gave a minimum SEC" when you have a representative calibration set.

It's always interesting, and I regard as an obligation for every scientist to investigate why the number of PCR/PLS components exceeds the number of mixed gases. Often it boils down to experimental errors, non-linearity or unexpected interaction effects. The only methods that highlight these types of variation, called Orthogonal variation are preprocessing methods such as the orthogonal signal correction techniques (Wold, Fearn, Andersson, H�skuldsson, Westerhuis) or the PLS method's recent development, OPLS (Trygg et al).
Using these methods, you highlight that type of variation (not correlated to your response matrix), making it much easier to understand their origin.

After that, you can decide what the next step will be, redo measurements or change profiling method to GC for example.

These types of methods represent a new tool for understanding variations in your data, especially the unexpected one :-)

See,

Trygg J, Wold S. Orthogonal projections to latent structures, OPLS. JOURNAL OF CHEMOMETRICS, 2002; 16: 119-128
http://onlinelibrary.wiley.com/doi/10.1002/cem.695/pdf

or

Trygg J, Prediction and spectral profile estimation in multivariate calibration
JOURNAL OF CHEMOMETRICS 18 (3-4): 166-172 MAR-APR 2004
http://www3.interscience.wiley.com/journal/109572688/abstract?CRETRY=1&SRETRY=0

regards,
Johan Trygg,Chemometrics group, Ume� University, Sweden
Top of pagePrevious messageNext messageBottom of page Link to this message

Tony Davies (td)
Moderator
Username: td

Post Number: 247
Registered: 1-2001
Posted on Friday, August 27, 2010 - 3:20 pm:   

Hello Vilas,

The answer to your question is that you use the number of scores that gave a minimum SEC on you calibration data and expect the performance you obtained on the validation data.

You need to ask yourself the question "How could any of the scores from the calibration set contain information about unknown contaminates?" I'm sure you realise that they cannot; so retaining more than the optimum number of scores will only reduce the performance of the calibration and make it more unstable.

I agree with Gabi that your proposed procedure is unsafe. You need to calibrate with mixtures sets which contain all of the possible gases in your application. If this is too difficult (or impossible) then I agree with Franz you need a method, such as GC. "The man who only has a hammer sees all problems to look like a nail"! We have to resist the urge to use NIR spectroscopy to solve EVERY problem.

Best wishes,

Tony
Top of pagePrevious messageNext messageBottom of page Link to this message

Gabi Levin (gabiruth)
Senior Member
Username: gabiruth

Post Number: 40
Registered: 5-2009
Posted on Thursday, August 26, 2010 - 4:01 am:   

Hi guys, I see this thread is going on for 3 months or so. As some of you know me - while I have respect for theories, I don't have complete confidence in them. This is to say that:
1. When you collect spectrum of pure components you only take into consideration the interactions of one type of molecules with themselves.
2. When you have mixtures, you now have interactions of different molecules with each other. This can not be taken care of by weighted spectra calculations.
3. Having just 4 or 5 elements in a regression means that the regression remains susceptible to bizzare effects when any perturbation takes place.
4. I believe that the possible errors due to creation of gas mixtures are small compared with the errors that can arise from having a small number of samples.
5. I believe that not having sufficent number of real mixtures, with various levels of moisture that maybe encountered in the "REAL LIFE" will yield regressions that are sensitive to so many perturbations, that when put to test, will fail.
5. As a devout practitioner, I believe that Murphy's law applies - anything that can go wrong will go wrong, and in the worst possible time. Besides water contamination, there can be other contaminants that vary in concnetration from one lot of gas sample to anohter, these will also affect the over all result.
Another issue - when testing the regression in real cases - how will you validate it? by doing what kind of analysis on the actual gas mixture? If this is to be done by a GC - then you need to understand the entailed error of the GC method. The best way to allow for the reference error is by using gas mixtures, have them flow through the NIR instrument, and from there directly into a GC for analysis. Use this data for the regression and it will make the regression robust and better comparable with the reference method.
I know it sounds like a lot of work - but if the saying that there are no free meals applies, then this is one of those cases.
I think that regardless of how much this is attractive to achieve, there are no short cuts that lead to the promised land.
Yours sincerely,

Gabi Levin
Brimrose
Top of pagePrevious messageNext messageBottom of page Link to this message

Francesco Davini (franz)
Junior Member
Username: franz

Post Number: 10
Registered: 2-2009
Posted on Thursday, August 26, 2010 - 3:26 am:   

Are you sure that NIR is more economical than GC for this application?
The GC analysis of such gases is a well established and very reliable technology. Calibrations are straighforward and can easily "resist" the presence of unknowns or changes in the mix composition over the limits of a NIR calibration. GC instruments exist from decades for this kind of application in a range of arrangements, from bench top to on-line ex-proof to portable units and from many vendors. A normal lab technician can manage the apparatus and all calculations, no need for chemometrics experts.

I have been working with both NIR and GC techniques in my professional life and I like both them. I'd have no doubts which one to select for this application if I need affordable data.

Sorry if this post can be perceived as polemical or disruptive. It is not my intention. I only have the practical point of view in mind.

Best regards
Franz
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
Intermediate Member
Username: vilasmechgmailcom

Post Number: 18
Registered: 5-2010
Posted on Wednesday, August 25, 2010 - 8:10 pm:   

Hello All,

This is regarding the number of scores that I should use in this application. I will perform calibration using mixtures prepared in the lab, which will contain only 6 components: methane, ethane, propane, butane, carbon dioxide and nitrogen. Out of these, all except nitrogen are active in NIR. However, when I go for actual field measurements, there is a possibility that the mixtures (to be tested) can contain a few more NIR active components. And I am not sure about the number of these components. In such a case, what is the number of scores/components that I should use to compute regression coefficients?

Thank you,
Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
Junior Member
Username: vilasmechgmailcom

Post Number: 9
Registered: 5-2010
Posted on Sunday, May 30, 2010 - 10:17 pm:   

Hello Jittima,

I think if we want to use PCR/PLS, it is better to use mixtures for calibration. However, my goal is to predict the mixture composition using pure species. Hence, CLS/GLS/SBC is better than PCR/PLS.

Thank you,
Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

Jittima Weeranantanaphan (ladypalm)
New member
Username: ladypalm

Post Number: 1
Registered: 5-2010
Posted on Sunday, May 30, 2010 - 9:58 am:   

Hello,

Regarding the post on Fri 28th, would it be possible that the mixtures actually provide more variations to the model than the pure components do, meaning that the model is actually learning and recognising more of those sources of variation hence better calibration??
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
Junior Member
Username: vilasmechgmailcom

Post Number: 8
Registered: 5-2010
Posted on Friday, May 28, 2010 - 1:23 am:   

Can anyone please help me in understanding why PCR/PLS work better only when calibration is done with mixtures? Why the results are like this, when calibration is done with pure species?
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
Junior Member
Username: vilasmechgmailcom

Post Number: 7
Registered: 5-2010
Posted on Wednesday, May 26, 2010 - 6:49 pm:   

Thanks Scott and Johan for suggesting the articles. I will try this method and check if it works.

David - We are able to achive high prediction accuracy using PCR/PLS on the NIR spectra of these 4 hydrocarbons.

Thank you,
Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

David Russell (russell)
Senior Member
Username: russell

Post Number: 47
Registered: 2-2001
Posted on Wednesday, May 26, 2010 - 3:19 pm:   

I would question the wisdom of attempting this with NIR. The four aliphatic hydrocarbons would have nearly identical NIR spectra.
Top of pagePrevious messageNext messageBottom of page Link to this message

Johan Trygg (trygg)
New member
Username: trygg

Post Number: 3
Registered: 1-2006
Posted on Wednesday, May 26, 2010 - 2:19 pm:   

The latest comments mentioned CLS (classical least squares) as a good alternative, especially the hybrid versions that could cope with some interferences. In that respect, I would like to add a similar alternative using the PLS method, that shows both good predictions but equally good interpretations as CLS presents. The idea behind it is a simple linear transformation of the regression coefficient matrix K=B*inv(B'*B) that yiels CLS like pure profile estimates.

Please read, "Prediction and spectral profile estimation in multivariate calibration" published in Journal of Chemometrics that you will find at
http://www3.interscience.wiley.com/journal/109572688/abstract?CRETRY=1&SRETRY=0

I will be happy to send it as an electronic pdf to anyone who is interested,
regards,
Johan Trygg,Chemometrics group, Ume� University, Sweden
Top of pagePrevious messageNext messageBottom of page Link to this message

Scott Ramos (lsramos)
Junior Member
Username: lsramos

Post Number: 8
Registered: 1-2007
Posted on Wednesday, May 26, 2010 - 2:04 pm:   

Vilas,

Regarding CLS, I suggest you search for articles by David Haaland who, while working at Sandia Labs, did some very important work early, before PLS caught on, and recently, where he showed how CLS could be enhanced to correct for interferences.

Scott Ramos
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 324
Registered: 9-2001
Posted on Wednesday, May 26, 2010 - 4:27 am:   

Vlas - yes, in condensed phases (solids & liquids) the intermolecular interactions perturb the molcular energy levels so that the fine stucture is smeared out by the interactions, and therefore the fine structure is never seen.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
Junior Member
Username: vilasmechgmailcom

Post Number: 6
Registered: 5-2010
Posted on Tuesday, May 25, 2010 - 10:23 pm:   

Hello Howard, Jerry and Richard,

Thanks a lot for your responses. Also, thank you for suggesting to use CLS. It will be great if someone suggests me any book/reference text that can be used to get in detail information about this method.

Howard - I did not understand the meaning of the term "with samples other than gases". Is this applicable for solids/liquids?

Jerry - Our spectra look like broad absorption bands. The primary reason for this is low resolution of the spectrometer that we are using. About your second question, as the C1 to C4 hydrocarbons have strong spectral intereference in the NIR, we can not assign the peaks directly to the components. Also, as the pressure increases, the spectral valleys among different absorption bands starts getting filled in, called "pressure broadening", as Howard said.

Thank you,
Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 323
Registered: 9-2001
Posted on Tuesday, May 25, 2010 - 6:29 pm:   

Jerry - the spectra of gases in general, depend on the measurement conditions. What we see in liquids and solids are the result of unavoidable broadening of the underlying absorption peaks due to intermolecular interactions.

In gases, if the pressure is low enough so that the molecules don't interact with each other (i.e., don't collide very often. Collisions perturb the molecular energy levels and thus change the absorbtion frequencies.) then the vibrational bands that we're used to seeing are split into their rotational fine structure. The lines from these can be very sharp, sharp enough so that Doppler effect and quantum uncertainty can be the limiting factors on their width. When lines are this narrow, it is difficult get enough instrument resolution to separate them, so that they are rarely seen, especially in the NIR.

This is exacerbated by the fact that the larger the molecule is, the closer together the rotational fine structure lines are, and then they tend to merge into the "lumps" we are used to seeing.

Increasing the pressure creates "pressure broadening" on the lines, because there are more intermolecular collisions. If the lines are close enough to each other then they overlap, and this is the fundamental limit on the resolution you can obtain.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Jerry Jin (jcg2000)
Senior Member
Username: jcg2000

Post Number: 28
Registered: 1-2009
Posted on Tuesday, May 25, 2010 - 4:57 pm:   

I agree with Richard Kramer. You already have the reference spectra, then you can even use the simple standard addition to quantify the components. I am curious what your spectra of gases look like. Do they have sharp line absorption peak or the commonly-seen broad absorption band? If your testing samples have known composition, and each component has their own characteristic peak, perhaps you could assign the peak directly to the components.

Jerry Jin
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 322
Registered: 9-2001
Posted on Tuesday, May 25, 2010 - 4:46 pm:   

Vilas - the May issue of Spectroscopy magazine, which just came out, discusses Richard's method of CLS calibration in some detail. It should work very well in your case, and by reorganizing your data suitably you can use an MLR program to do the calculations. The answers you get should be directly in mole percent.

It turns out, however, that with samples other than gases there are some hidden pitfalls. We have submitted an article to Applied Spectroscopy that discusses this issue. We just got it back from the reviewers and so hopefully should be out within another couple of months.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Richard Kramer (kramer)
Advanced Member
Username: kramer

Post Number: 25
Registered: 1-2001
Posted on Tuesday, May 25, 2010 - 3:23 pm:   

If you have quantitative reference spectra for every species which is present and absorbing over the spectral range being measured, the PLS/PCR are not the most appropriate techniques.

The classical way of handling such a situation is to use CLS to directly build a matrix of calibration coefficients for each spectrally active gas. In such cases, CLS is a one-step process which yields the calibration vectors for each analyte directly since in is not necessary to use the traditional "first step" of CLS to estimate the pure component spectra which are already available.
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
New member
Username: vilasmechgmailcom

Post Number: 5
Registered: 5-2010
Posted on Tuesday, May 25, 2010 - 12:38 pm:   

Hello Jerry and David,

Thank you for your response. David is right. When I added the spectrum of a mixture of 50:50 methane:nitrogen, I got 100% accurate results.

I am trying to perform prediction testing in order to check accuracy of the model at predicting the composition. The following procedure:

1) I measure the spectra of 4 pure gases. These spectra make the calibration set. The data matrix consists of two columns - wavelengths and absorbances at each wavelength. I calculate regression coefficients using PCR and PLS with 4 principal components. Now if I plot the regression coefficients against wavelength, the graphs look very much similar to the pure spectra for each gas.

2) Say, I have a mixture of 75:10:5:1 methane:ethane:propane:butane (remaining can be non-absorbing gases, such as, nitrogen with no spectrum). I can choose these concentrations randomly between 0 to 100, such that the total sum<=100.

3) Now, I want to check how accurate the calibration model is at predicting the composition of this known gas mixture. To check this, I can prepare the mixture, measure its spectrum, and use this spectrum as an input to the PCR/PLS to predict the composition. However, measuring the spectrum is not necessary. I compute the spectrum of the mixture by taking the weighted sum of pure spectra at each wavelength in 900 to 1700 nm region. (Because of the linearity of pure spectra and absence of interaction amongst the gases, we do not need to actually prepare the mixture and measure its spectrum. We can do it mathematically).

4) I use this calculated mixture spectrum as an input to PCR/PLS. Predict the comsposition and comapre the actual and predicted concentrations.

I am getting 100% accurate results, when the sum of concentrations of absorbing components is 100. However, if the sum<100, (for example, for the above mentioned mixture, sum<100), the predicted concetrations are such that their sum is exactly equal to 100. This forces some of the predicted concentrations to be greater than 100 and some of them less than zero (which is not expected/desirable). I am sure I need to make a very small change somewhere in the regression coefficients, however, I am not able to figure out.

Also, the goal is to predict mixture composition using the spectra of pure gases. It is not necessary to use PCR/PLS. I can use any other statistical method, such as, MLR etc, if possible. I hope my question is clear. As I am more into spectroscopy, I am not aware of all statistical methods used for NIR data analysis. I tried to use PCR/PLS as they are very widely used and accepted. Please suggest me any other methods, which can perform equally well as PCR/PLS/PCA.

Thank you,
Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

Jerry Jin (jcg2000)
Senior Member
Username: jcg2000

Post Number: 27
Registered: 1-2009
Posted on Tuesday, May 25, 2010 - 11:25 am:   

Hi Vilas,

I am afraid you have no choice but to use gas mixtures as your calibration samples if you want to predict mixture composition.

You are concerned about errors associated mixture
preparation. There is always a big error in the nominal concentration of gas mixture, considering the variation in volume measurement. 1 liter of methane mixed with 1 liter of ethane does not mean the percentage concentration of methane will be 50%. Nevertheless, you can always quantify precisely the composition of a gas mixture by using GC. Even for methane you can choose suitable column to get it separated from other alkanes.

Jerry Jin
Top of pagePrevious messageNext messageBottom of page Link to this message

David W. Hopkins (dhopkins)
Senior Member
Username: dhopkins

Post Number: 143
Registered: 10-2002
Posted on Tuesday, May 25, 2010 - 11:23 am:   

Hi Vilas

I am confused by your statement "I am presently calculating the spectra of mixtures from the spectra of pure gases and trying to predict the composition of these mixtures using the calculated spectra (in other words, I am not measuring the spectra of mixtures experimentally, to reduce errors in calculations). In order to calculate the spectrum of a mixture, I take the weighted sum of the spectra of pure gases." How are you calculating the spectrum of an unknown sample from the pure components? Curve fitting might be an option here, but you did not mention that.

If you are observing the 120 -20 effect with the 4 factor calibration with 4 calibration samples, then I think you have some problem with the low number of samples, as someone mentioned earlier. I wonder if it would be fixed by including duplicate scans of the calibration samples, so you would have 8 or more samples in your calibration set. It also might help to stabilize the calibration to have some binary mixtures in your calibration set, such as 50:50 mixes of c1 and c2, etc, and c1 and N2, etc. When these are added to the calibration set under standard pressure and temperature, they might help the regression achieve the proper results, but you still should only use 4 factors and no mean centering.

Best wishes,
Dave
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
New member
Username: vilasmechgmailcom

Post Number: 4
Registered: 5-2010
Posted on Tuesday, May 25, 2010 - 10:36 am:   

Hello Gabi, Hovard and David,

Thank you for your responses. We are developing this real-time optical gas sensor as a low-cost alternative to gas chromatographs. We do compare our results with those of GC and we are able to achieve comparable accuracy using spectra of mixtures during calibration. Now I am trying to improve and minimize calibration requirements, by using only the spectra of pure gases (to avoid errors associated mixture preparation, as already mentioned).

We have experimentally found that, in the considered pressure and temperature range, the NIR absorption of hydrocarbons and carbon dioxide increases linearly with density. There are no significant spectral effects associated with line broadening and/or shift in the line centers. All the spectra, that we measure, are corrected to standard conditions of pressure and temperature (14.696 psia and 20 degree C), before performing any other mathematical computations on these spectra. Also, as carbon dioxide has a very low absorption in NIR, thus, I am concentrating only on hydrocarbons and their mixtures to minimize prediction errors. Thus, my calibration set now has 4 pure spectra (C1 to C4 hydrocarbons). I use the number of factors 4. Calculate regression coefficients using PCR and PLS. Then, I am trying to predict the composition of mixtures, which can have 5 component gases - C1 to C4 hydrocarbons and nitrogen. The pressure remains fairly constant (14.5 to 14.7 psia) during calibration and prediction testing.

All the gases, we use, are CP grade and 99% pure. I know that water has a very strong absorption near 1900 nm. However, in order to avoid such errors, I am presently calculating the spectra of mixtures from the spectra of pure gases and trying to predict the composition of these mixtures using the calculated spectra (in other words, I am not measuring the spectra of mixtures experimentally, to reduce errors in calculations). In order to calculate the spectrum of a mixture, I take the weighted sum of the spectra of pure gases.

Also, we have experimentally verified that there are no chemical interactions between different gas molecules. Or, in other words, all the gases in our mixtures are linear binary structures. Hence, in the considered wavelength range, the total absorption at each wavelength is equal to the sum of (absorption of each gas * its molar fraction).

Raman - I think it will be good to start a new thread in order to discuss your question.
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 321
Registered: 9-2001
Posted on Tuesday, May 25, 2010 - 8:05 am:   

Gabi - I have some spectra of "synthetic natural gas" (99+% methane) at pressures up to 1,000 psi. These show fairly large spectral effects due to pressure, and the effets are visible at pressures well below that value. I agree that these spectral effects are fairly small, but when the calibration calculations are on the edge of being singular (as I think you and I agree these might be), the (mathematical) system is instable and small changes to the data can have large effects on the calculations.

Good point about the water, though, although I think it was mentioned previously.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Gabi Levin (gabiruth)
Senior Member
Username: gabiruth

Post Number: 32
Registered: 5-2009
Posted on Tuesday, May 25, 2010 - 5:54 am:   

Howrad is right - when working with gases - the pressure needs to be regulated carefully - the degree of absorption depends on molecules density in the path of light - critical to have same pressure.
Temperature is also essential, particularly if the chamber is sealed - temp changes will chage pressure.
If I understood correctly - Vilas indicated that the pressure is well below the pressure where molecular collisions can affect spectral changes, but even at very high pressures these effects are third order - not first order (a friend of mine had his Ph.D. dissertation done on spectra of diatomic molecules under extreme pressures, many years ago, 1974)
On the other hand, considering the high absorption coefficients of water molecules, even small amounts of moisture can cause significant errors in predictions. Creating a calibration with variations in conditions, which include variability of moisture will eliminate the risk from such influences.
Still, in my opinion the 120-20=100 case is due to the small amount of samples in the regression, and to the perfect interdependence of the values of the consituents.
Top of pagePrevious messageNext messageBottom of page Link to this message

Raman NVS (raman_nvs)
New member
Username: raman_nvs

Post Number: 1
Registered: 5-2010
Posted on Tuesday, May 25, 2010 - 5:27 am:   

Dear All
I just joined this group. I am working in a pharamceutical industry. I am planning NIR for my both formulations and API units. When I started exercising, I found two techniques are unsing in this feild. One is FTNIR and another one is Dispersive NIR.

Can anybody suggest which one is suitable for our applications.

Raman
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 319
Registered: 9-2001
Posted on Tuesday, May 25, 2010 - 5:17 am:   

Vilas - two more conditions, due to physical effects, that you should take into account:

1) Unlike liquids and solids, with gases you can have different amounts of material even though the composition is the same, since the amount of material will change with pressure. Therefore, if your calibration gases are measured at one pressure, and the mixtures at a different pressure, there may well be less of the gas present than the calibration spectra can accomodate. Therefore, your total pressure should be the same for calibration and measurement.

2) Similarly, the presence of nitrogen, with no spectral features, will dilute the other gases so that their absorbances are lower than when the N2 is not present. If you add nitrogen to a given mixture, while keeping the total pressure constant, you should be able to see this effect. If, on the other hand, the partial pressure of the mixture is kept constant and nitrogen is added so as to increase the total pressure, then see #3 below.

3) Another factor is the possibility of interactions between the different gases changing their spectra. Pressure can also change the spectra of gases, due to broadening of the rotational fine structure (not all the gases in your mixture are linear binary structures).

That's three out of two.

Howard

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Gabi Levin (gabiruth)
Senior Member
Username: gabiruth

Post Number: 31
Registered: 5-2009
Posted on Tuesday, May 25, 2010 - 4:34 am:   

Hi guys,

The subject is very interesting - but what is the purpose - if it is to measure synthetic mixtures - where no other inteferences of contaminants will be observed, it is one thing, if it is to measure real process caes then in my opinion it is a futile effort. In such a case there is no substitute to the tedious but useful work to obtain spectra from process condtions and analyze each process condition for composition by GC - and use these values for the regression. The scope of variability of each gas should reflect the anticipated range of variability under real process conditions. This will also eliminate most of the problem with the interdependence between the Y variables and prevent such a condition of 120 - 20 being equal to 100.
I know it is not a great help to those who would like to find shortcuts, but shortcuts sometimes are longer then the long cuts.

Gabi
Top of pagePrevious messageNext messageBottom of page Link to this message

David W. Hopkins (dhopkins)
Senior Member
Username: dhopkins

Post Number: 141
Registered: 10-2002
Posted on Tuesday, May 25, 2010 - 1:32 am:   

Hi Vilas,

I agree with Howard, you are lucky your calibrations gave you anything at all.

Because only 4 of the 6 gases have an NIR spectrum of any appreciable magnitude, I expect that you should obtain a good calibration with only 4 factors (and don't use mean centering-if you mean center, you could only use 3 factors). You have to accept that you will not be able to obtain values for N2 and CO2, but your values for methane, ethane, propane and butane should be reasonably good, if the samples are dry. You could be having some trouble if there is any water vapor in the samples.

Please let us know how the 4-factor calibrations perform.

Best regards,
Dave
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
New member
Username: vilasmechgmailcom

Post Number: 3
Registered: 5-2010
Posted on Monday, May 24, 2010 - 8:03 pm:   

To be clearer, there are only five pure spectra in the calibration data set - methane, ethane, propane, butane and carbon dioxide. And, mixtures can have SIX components - the above mentioned five and nitrogen.

As nitrogen does not have absorption in the considered wavelength range, it does not make sense to add its spectra to the calibration data set. Its spectrum will be a flat line with Absorbance=0.
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
New member
Username: vilasmechgmailcom

Post Number: 2
Registered: 5-2010
Posted on Monday, May 24, 2010 - 7:56 pm:   

Hello Jerry,

Thanks a lot for your response. Nitrogen does not absorb light in the NIR region at least upto 1000 bar. Our considered pressures are well below this maximum pressure. (A homonuclear diatomic molecule, such as, nitrogen (N2), oxygen (O2), does not have electric dipole moment. Thus neither rotational nor vibrational transitions produce an oscillating dipole moment and associated dipole radiation. Hence, homonuclear diatomic molecules do not absorb light in NIR spectral region.) Hence, the fact that "the absorption increases linearly with concentration" is not applicable to nitrogen.

Yes, every calibration sample is a pure gas and for prediction testing, I use synthetic mixtures (prepared using desired concentrations of pure gases).

My goal is to use pure gases for calibration and to predict the composition of mixtures of those pure gases using any statistical method. Calibration with pure gases eliminates the need to prepare synthetic mixtures and hence, the error associated with mixture preparation.

I request you to suggest me the possible methods that can be used for this purpose.

Thank you,
Vilas
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 318
Registered: 9-2001
Posted on Monday, May 24, 2010 - 7:43 pm:   

Vilas - you're lucky you were able to get answers at all, without a divide-by-zero error or some such, since the all the samples sum to 100% and therefore there is perfect intercorrelation. The trick to making that sort of situation work, is to do the regression with the B0 term (the constant of the calibration) set equal to zero. Draper & Smith discuss how to set this up; it's a minor modification of the standard regression equations, but is critical to avoiding those amomalies. I'm sure that there are other books on regression that discuss that, also, but I'm familiar with D & S.

Howard

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Jerry Jin (jcg2000)
Senior Member
Username: jcg2000

Post Number: 26
Registered: 1-2009
Posted on Monday, May 24, 2010 - 7:25 pm:   

Vilas,

When you said "the NIR spectra of these components are linear with their concentration", does that hold for N2?

I assume all samples you used are synthetic, and every sample in your calibration set is pure gas. How can you use a 5-factor PCR model to describe a single-component pure gas sample?

If you want to predict gas mixture composition, you should use mixture as your calibration sample.

Jerry Jin
Top of pagePrevious messageNext messageBottom of page Link to this message

Vilas Vyankatrao Jangale (vilasmechgmailcom)
New member
Username: vilasmechgmailcom

Post Number: 1
Registered: 5-2010
Posted on Monday, May 24, 2010 - 6:31 pm:   

Hello,

I am trying to find the composition of multi-component natural gas mixtures, which contain methane, ethane, propane, butane, carbon dioxide and nitrogen using NIR spectroscopy. Out of these gases, only nitrogen is not active in NIR. The wavelength range I use is 900 to 1700 nm. For the spectral resolution (FWHM 12.5 nm) I am using, the NIR spectra of these components are linear with their concentration in the considered pressure and temperature range. Hence, I am able to reproduce the spectra of mixtures from the spectra of pure components using the molar fractions of individual components in the mixture. Or, in other words, the weighted sum of pure spectra gives the spectrum of a mixture. This suggests me that, if I use the spectra of these 5 pure components in the calibration data set, the PCR/PLS methods should be able to predict the composition of a mixture. However, my observations are as follows:

1) For a mixture, which does not contain nitrogen, the PCR/PLS predictions are 100% accurate.

2) For mixtures, which contain nitrogen, some of the predicted concentrations are positive and the rest negative, in such a way that the sum of the concentrations is 100. (I use percentage concentrations during calibration). For example, for a mixture of 50% methane, 25% propane and 25% nitrogen, the predicted concentrations will be something like these: 120% methane and -20% propane; so that sum=120+(-20)=100. This is perhaps happening because, for each spectrum in the calibration data set, the concentration is 100% (pure gas).

I use the number of components = 5 for both methods, as there are 5 absorbing components in the mixtures. I use SVD while implementing PCR. I am not able to understand the exact statistical reason behind this.

Any help/comments/suggestions would be highly appreciated.

Thank you,
Vilas

Add Your Message Here
Posting is currently disabled in this topic. Contact your discussion moderator for more information.