| Author | Message | ||
     Dr.K.Balasubramanian (drkbala) New member Username: drkbala Post Number: 4 Registered: 9-2006 |
Mr.rajpoot how limited are the field samples the electronic signature of the sample is important | ||
| MSRajpoot (msarajpoot) New member Username: msarajpoot Post Number: 4 Registered: 7-2006 |
Dear Lee thanks for your comment, actually here the cost is not important the important is time. The classical method more often take 3 to 4 hours tahts why we want to go for NIR. Now the problem is number of samples, so what i did i took the real sample as a bulk sample and introduced the intrested component externaly and prepared 20 samples, then i caliberated it and include some of the real samples and then checked the reproducibility it was really good. I checked for real samples and for the lab samples both were given quite good result. Our future plane is for passage of time we will replace all lab prepared samples with the real process samples but for starting point we did this practise. | ||
| kathryn lee (kathrynlee) New member Username: kathrynlee Post Number: 2 Registered: 2-2004 |
Dear MSRajpoot, I was just wondering why, if you have such a perfect process such that you have only 5 process samples that are very similar, you need a NIR method? If you are not manufacturing very often, a NIR method may not be cost-justified. Presumably you are testing the samples already with another analytical method. You may want to check that method to determine its real precision and accuracy, and whether the NIR can even detect changes on that level. Sometimes people think the method is more accurate and reproducible than it is. There are other chemometric methods for determining whether your process is going well without doing quantitative analysis. Sometimes people take years developing reliable methods of analysis, yet they are worth while. Have you done a cost justification, including instrument cost, method development, lab analysis costs, maintenance, etc.? In addition, a risk analysis of what could happen if the chemometric method fails either false negative or false positive would be worthwhile. Those two things might be worth doing to help you decide how to proceed. In addition, start up and shut down are good times to get process samples that are beyond the range of good samples. Also, other grades of product may expand the range. | ||
| Bruce H. Campbell (campclan) Moderator Username: campclan Post Number: 87 Registered: 4-2001 |
Adding pure components is asking for trouble. If you absolutely need to see if chemometric NIR MIGHT work, you may want to mix samples in different proportions ONLY if the range of the desired constituent is very large. By very large I mean the value of the ratio of the range divided by the standard deviation of the reference method. This ratio should be above 12 and preferably above 20. If the samples are liquid you should be able to get uniform mixing. If they are solids, you have to be very careful to get a good mixing. Further, if the samples are powders, differences in particle size could mislead you, especially if the differences are related to the concentration of the constituent of interest. If the results of the study show promise, there is no assurance it will work in real life with a multitude of calibration samples, but it does suggest there is a chance it will. If the calibration doesn't work, forget it and go to a separation scheme and use classical spectroscopy or some other technique. Finally, with only five samples, you don't have many degrees of freedom, a dangerous approach. Bruce | ||
| David Russell (russell) New member Username: russell Post Number: 28 Registered: 2-2001 |
In my group we have found that lab samples are usually only useful to determine if there is enough signal in the spectroscopic data to justify the attempt at process analysis. Of course, a lot depends on how difficult the measurement is in the first place. For a difficult measurement even the best designed laboratory experiment yields a calibration model which would be described as "a good start". Long term success in process analysis requires ultimately replacing that model with one that uses spectra collected on real samples under controlled conditions and reference analyses for the concentrations of interest that can be relied on. I too agree that mixing lab and real samples is a bad idea. Mixing real samples or spiking them will yield a better result, but you won't do as well as you can until all of the samples in the training set are real process samples. | ||
| MSRajpoot (msarajpoot) New member Username: msarajpoot Post Number: 3 Registered: 7-2006 |
thanks to all for their comments, actually most of the time it is difficult to get large number of samples with different concentration of the intrested component. My question is very simple if i have only few samples from process which have very little variation and this variation lies for long time so how we can establish the calibration with these limited samples, or we can add some laboratory prepared sample from the process sample as bulk. | ||
| Donald J. Dahm (dahm) New member Username: dahm Post Number: 2 Registered: 1-2006 |
We are getting the advice we are from Howard because it is good advice. Since he asked for other perspectives, I will chime in. The people who claim to do NIR Spectroscopy frequently are not doing much spectroscopy. They are just using Chemometrics on Spectroscopic data. That has been a powerful combination. If you need to violate the �Rules of Chemometrics Procedure�, you must be prepared to do Spectroscopy for real. And, in my experience, you then should abandon PLS. I can remember being presented with a beautiful set of samples covering a wide concentration range of the analytes, and all the current process samples I wanted. Reference analysis by NMR was there for the asking. The problem was that the matrix of the �extended range� samples had changed over time; and there was no was they were going to regenerate those samples. By using MLR and carefully selecting wavelengths, the analysis could be made insensitive to the effects of the matrix change. The model wound up being a single wavelength to predict concentration, and a ratio of the 1st derivative of two other wavelengths to remove the matrix effects. Based on what I hear at various Conferences, the art of Spectroscopy in NIR Spectroscopy is dying. As I�ve said before, every time Karl Norris used MLR in a Chemometric Shootout at Chambersburg and got better results than folks did with PLS, I did a silent cheer. It�s not that I�m against the Chemometric tools: I just am saddened that our first response to people with questions is �Get Better Samples� rather than �Do Better Spectroscopy�. | ||
| Gavriel Levin (levin) Junior Member Username: levin Post Number: 8 Registered: 1-2006 |
Howard, You know I will agree with you. A possible comparison - people can learn how to drive in a simulated environment, simulating traffic etc. But when they go first time to the real streets, they learn how to handle the real traffic based on the skills they acquired, and in accordance to the capcity of their own brain to learn. The chemometric software, as good as it is, does not have a brain of its own, and what it was not taught (sometimes people refer to the calibration set as the training set, and it really fits) it can not learn by itself. Hence - an accident. I want to ask Rajpoot - what is the difficulty in obtaining spectra from the process and obtaining a corresponding sample - if I know the issue I might be able to suggest some idea based on experience we have gained over time. I would truly wish to help. Thanks, Gabi Levin Brimrose | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 35 Registered: 9-2001 |
Well, I don't really have a whole lot more to add at this point. The basic problem with doing it all (or even mostly) in the lab is just that those samples do not have the process "signature" on them: whatever variations there might be in temperature, compression (very important for a powder), inhomogneity, averaging of the inhomogeneity if the data are collected while the sample is moving, or any other effects that will affect the dynamic process, as contrasted with the relatively static environment of a laboratory. Sometimes you can get away with it, but that's a rare case. Far better to expand the sample set with "real" samples. Anyone else have any comments? Don't be afraid to chime in. \o/ /_\ | ||
| MSRajpoot (msarajpoot) New member Username: msarajpoot Post Number: 2 Registered: 7-2006 |
thanks Howard, Actually normally in the industries it is difficult to find many samples for NIR caliberation and we know it is ideal to have real samples for caliberation, if we collect a bulk real sample in which we know the concentration of the component can we use this bulk real sample for further sample preparation by varying the component concentration by addition externally. By this way at least the reference samples could be of same matrix. I will be waiting for your kind comments. thanks, MS Rajpoot | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 33 Registered: 9-2001 |
MSRajpoot - the problem with lab samples alone is exactly the fact that they do not give you a good representation of what you are calling "matrix effect". The usual recommendation for NIR calibration is to use a large number of "real" samples for calibration; the reason is just to make sure you include representation of all the ways that the samples (i.e., the "matrix effect") can vary. If you do a good job of that, then the training process for the instrument/algorithm will be able to recognize and compensate for all the changes in the matrix that you are likely to encounter. Having only 5 real samples and trying to extend the sample set with lab samples is very risky, since the model will not be able to compensate for the various changes in the matrix you will encounter in routine use, and your results will be inacurrate. We normally recommend that you not only include real samples in the calibration data set, but you should also have "extra" samples, not part of the calibration data, to reserve and use as "unknowns", to test the model you obtain to see how well it can analyze other samples, that are more representative of the ones you will be measuring when you deploy the calibration for routine use. Howard \o/ /_\ | ||
| MSRajpoot (msarajpoot) New member Username: msarajpoot Post Number: 1 Registered: 7-2006 |
I am participating first time to this very informative forum, my question is that if we have very limited field reference samples i.e. 5 can we include the samples which have been prepared in the lab which contain the same component whcih we are looking for in real sample, and what will be the matrix effect on the calibration or any body have any other sugestion. |