| Author | Message | ||
     Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 50 Registered: 5-2009 |
This shows for the who knows how many times that there is really no value in the visible range for NIR calibrations - I have seen it happen many times when I work with people that have a lab Foss or any other instrument that covers the visible range along with the NIR. In many cases I had to suggest to them to cut off the visible range - and calibrations were substantailly better. Gabi | ||
| Rong Zhou (rongz) New member Username: rongz Post Number: 3 Registered: 12-2010 |
Dear all: Thank you for providing helpful suggestions. With Dr. Ole Rasmussen's help, the calibration is back to normal. What we did was to cut 400-700 nm spectra from the calibration file and used the shortened spectra to make a new equation. The new equition has very good global H and neighbourhood H. Best regards Rong | ||
| Fernando Morgado (fmorgado) Member Username: fmorgado Post Number: 14 Registered: 12-2005 |
Dear Dafna : I know, but sometime the users forgot make this test ( for Instruments who have drifting) . I know the �new� program winisi say � please run the test, time expirate� but..... not always the users make that. I am accord with you absolutely, is a bad practice not make that. Regars Fernando | ||
| dafna barkan (dafna) Member Username: dafna Post Number: 15 Registered: 4-2003 |
Fernando, I hope I am not being rude, but not making linearizations and not checking an instrument periodically for other possible problems is actually bad NIR practice. A lot of work is put into the process of developing good calibrations. Then, if you expect them to work well over a long period of time, good instrument maintenance is unavoidable. | ||
| Fernando Morgado (fmorgado) Member Username: fmorgado Post Number: 13 Registered: 12-2005 |
Hello : David have the reason, Foss say make K and Phi adjust after change the lamp. But that supose the customer make linearization each some time. If the customer never make that change K and Phi values without registry first the old K and Phi value can be a problem, if the result obtained are bad. My comment is not only teory, is a real experience , not one time, any times with my customer. For know if realy some change ocurr after and before adjust, the better is plot spectra, one ole spectra and some one new, of the same type of sample. You can check x axis position for the peak and you can make a sustraction of the spectra. If not change in X and Y, your problem with H is in other part, maybe the sample. Please consider scaterrin if you make this spectra sustraction, the better is use 1 derivate spectra for comparate. Regards Fernando | ||
| David von Boisman (david_von_boisman) Advanced Member Username: david_von_boisman Post Number: 23 Registered: 2-2006 |
Working with this instrument family for more than 10 years, I fully support the view of Dafna. Id like to add, that adjusting K and Phi values through linearization, is a routine procedure for these instruments after any service or repair, and in routine use, its performed at least each time the lamp is replaced. | ||
| dafna barkan (dafna) Member Username: dafna Post Number: 14 Registered: 4-2003 |
Fernando and Rong, From my personal experience, I know that repairing an instrument does not necessarily incur in big changes to your predictions. This is especially true when your application is robust enough. In more than 10 years dealing with calibrations and instruments, I can assure you that in many cases the FOSS instruments can be recalibrated and potential damages can be significantly reduced. K and Phi values change along time also during routine work and need periodically to be adjusted. And still it does not mean that every time we update them we need to come up with new calibrations. As for a solution to your problem, I think that many contributors have already mentioned the standardization possibilty, provided you still have some of your original spectra . As for the future, backing up your data and updating your applications with significant spectra can make them quite robust, and this is a good insurance policy against possible instrument changes. | ||
| Fernando Morgado (fmorgado) Member Username: fmorgado Post Number: 12 Registered: 12-2005 |
Dear Rong : About high H value. You say your Foss was repaired. Normaly technical service repair the instrument and run Wavelenght Test, after that the system indicate actual K and Phi and sugested K and Phi. The service without think in the models the customer have, adjust K and Phi. After that normaly all the result change, including H. Foss have drifting in the time, and probably your samples was collected with some X position and Y axis Intensity , maybe for example your OH is in 1943 nm, and with K and Phi adjustment move to 1940 nm. After that all the spectra was moved to the left and all the result change. I recomend always have a registry about K and Phi, if the instrument fail, in any moment you can input this value manualy. If you don�t have the old value you have a real problem for use your old models again. If you need use this models you have two options 1.- Look for some check cell spectra or sample stored before the fail, read again ( same phisical sample) and created a STD file. 2.- If you don�t have some stored spectra and the phisical sample, the only posibility is move manualy the grating until the new spectra reach the old spectra, but, this is a not easy procedure. ( excuse mi English) Regards Fernando Morgado | ||
| venkatarman (venkynir) Senior Member Username: venkynir Post Number: 124 Registered: 3-2004 |
To all NIRs lovers Wish you Happy New Year Dear Gabi Please avoid the word some in your answers. You are expertise in NIRS and your knowledge should help for solution . Few months before I asked same question to my friend Mark , he thrown it Instrument DAS manufacturer. But I found the Dark and Reference are not same .It fails t test and repeatability . Next is how to ratify it . We did it by means of tuning it with regression . Later samples , I think more samples will good result , so we went for more samples but it difficult to get by inferential method like .1% FFA to .15% FFA. So it becomes very difficult task . But we collect samples from random source of the same we have written source code for calibration. Thanks to Paul for it.We use Unscarmbler/Grams/AI /PLS Tool box as versification for methods for test our source code .We provide calibration facility in the instrument it self. We found that we are able to go close to 75 % to 85 % of the true value only. We have tried MLR with SPA and PLS too. In long range MLR better others close range PLS look better. Main draw back of NIRS Instrument manufactures is not providing calibration facility in the system . When ask samples the user ran way from us . For new user NIRS calibration is "like blind man looking at elephant ". | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 391 Registered: 9-2001 |
Gabi - those are all possibilities, but I figured that instead of trying to guess, it made more sense to find out for sure, which I could do just by asking!! \o/ /_\ | ||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 49 Registered: 5-2009 |
Hi Howard, True, but I assumed that what really bothers them is that predictions are now off the mark. I am not really an expert on issues such as H etc. being a real practitioner I am mostly concerned with quality of predictions. If a "fixed" change occured in the spectrum due to the repair (for example - changing a lense or a window in the optical path of the sample can cause a fixed change in the spectrum as a result of the difference in spectral transmittance across the spectrum of the new lens coatings) while the internal reference optics remains the same, so now the actual "transmittance" curve is slightly different - this should cause both a change in the H value and the predictions. Thus, I would expect that the procedure I suggested will "correct" both aspects. In my experience - doing this and using classification criteria such as sample to model distance (obtained from Unscrambler PCA methods) it is possible to "bring" the STMD values back to where they were before. Regarding possibility that samples have changed - I doubt it - but hopefully they kept some "old" samples somewhere under suitable conditions so they can check that possibility. Gabi Levin | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 389 Registered: 9-2001 |
Well, Gabi's suggestion is certainly worth a try. But a change in the H value is not the same as a bias in the analytical prediction - and the cause and cure may not be the same either. In the extreme, you may need to do a whole new calibration. BTW, Rong - How are the analytical predictions? Are you observing larger errors (larger SEP), or a bias, or any other systematic effects, besides the larger H values? Maybe it's the samples that have changed. \o/ /_\ | ||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 48 Registered: 5-2009 |
Hi every one, All the best wishes for a new better happier and successful year. To the question - I am not familair with the details of the instrument as I come from a different technology - but if you have the ability to perform the necessary reference analysis on about 15 to 20 samples with values across the range - you can scan them, do the analysis and add these new spectra with new values to your calibration set and it should fix the problem. There must be some difference somewhere in the spectrum that is a constant difference, therefore introducing this difference into the PLS method, will allow the regression to compensate for this, and enable you obtain good predictions. I have used this approach several times when a repair is done on our AOTF spectromerters, and it always worked well. Gabi Levin Brimrose | ||
| Rong Zhou (rongz) New member Username: rongz Post Number: 2 Registered: 12-2010 |
Thank you, Howard. I do not think FOSS will be able to do it. We usually use NIR to scan canola seed samples and mustard seed samples for the period October to December. The machine worked well in 2009. It broke down in the mid of October/2010. It was fixed. However, if we apply the equation, most of the results have a global H value higher than 3 (the average distance from mean is high, at least 4-5). Due to that this machine was not standardized before, FOSS suggests me to use the spectrum of my check cell can107 before breakdown as the master and the spectrum of my check cell can107 after fixation as the host. I tried it and got better results (more samples with global H less than 3, the lowest Global H is about 1.5 and the highest global H is about 4.5-5). I have a home made check cell. When I applied the equation to it, its spectrum before the breakdown is good while its spectrum after breakdown is about 3.5. Best regards Rong | ||
| Howard Mark (hlmark) Senior Member Username: hlmark Post Number: 386 Registered: 9-2001 |
Rong - since you just had the instrument repaired by the manufacturer's service, changes in the performance should be the responsibility of the manufacturer, under the warranty for the service. I suggest you contact FOSS. \o/ /_\ | ||
| Rong Zhou (rongz) New member Username: rongz Post Number: 1 Registered: 12-2010 |
Happy New Year, everyone: I guess I could not find where I posted my message. I have an old FOSS NIRSystem 6500 for canola seed and mustard seed quality analyses. Recently it was repaired by FOSS for the order sorter failure. The problem is that the equation pridicts with higher global H for newly generated spectra. I used the old spectrum of the check cell can107 as the master and the new spectrum as the host and did not get satisfied results (i.e. global H is still over 3). I was wondering whether you experienced similar situation. Any suggestions? Best regards Rong [email protected] | ||
| Razaq Balogun (razbalogun) New member Username: razbalogun Post Number: 3 Registered: 1-2007 |
Thanks to everyone that have responded to my inquiries about NIR. Yes, we believe in our business that developing our calibrations will provide opportunity to learn and acquire skill in the technology. As far as cost is concerned, we actually looked at the option and found that it will be cheaper for us on the long run than buying licence to use pre-developed global calibrations. My apologies that my email is hidden. I am a first time user of this discussion forum and do not know how it operates. If anyone wants to send me email personally, my email address is [email protected] I am looking forward to read from anyone with usefull information regarding my initial inquiries. Razaq | ||
| Pierre Dardenne (dardenne) Intermediate Member Username: dardenne Post Number: 20 Registered: 3-2002 |
Razaq, Apologies for having sent you to Google. I am sure you are right to build your own calibrations if you want to learn about NIR. It will be just more costly for your company. As Richard said, with your email address, I could suggest you some humble advises on the subject, if you agree. Pierre | ||
| Richard Kramer (kramer) Junior Member Username: kramer Post Number: 7 Registered: 1-2001 |
Razaq, I tried to send you some info which might be helpful but your email address is "hidden." I'd suggest to you, and to others posting questions to the list, that if you are going to post such general questions which might encourage detailed answers, it would be more appropriate if your email address is not hidden. That way people could send you attachments or lengthy responses which would not be appropriate to post directly to the list. | ||
| Gavriel Levin (levin) Senior Member Username: levin Post Number: 31 Registered: 1-2006 |
While I am not sure that Dafna's remarks about the recommendation to go to the web are fully jsutified, I agree with her that in this case, particularly for people that need to learn the trade and become skilled, it is highly important to learn how to shave on his own beard. The experience and understanding he is going to gain by it are worth a lot. It will later help him understand problems if he runs into, and be able to fix them rather than start asking everyone questions. By all means, get your feet wet, gain the experience, and become a good skilled NIR user - it will last you years. Gabi Levin | ||
| dafna barkan (dafna) New member Username: dafna Post Number: 9 Registered: 4-2003 |
I don't think that somebody who is trying to develop his own NIR calibrations is "reinventing the wheel", as Mr. Dardenne says... With all due respect to his enormous professional skills and merits, redirecting someone to Google to find the INGOT calibration packages as the ultimate solution for calibration-making, has a commercial flavor which I understand is not so welcome at this forum. As someone who worked very hard to develop his own calibrations, I think that Razaq has a lot to profit from his own trials. INGOT can absolutely be a wonderful solution for anyone interested in NIR assessments. And still it does not mean that Razaq's questions should go unanswered. | ||
| Razaq Balogun (razbalogun) New member Username: razbalogun Post Number: 2 Registered: 1-2007 |
Thanks Pierre, We are interested in developing our own calibrations. Starting from the scratch so to say. As you stated, "reinvent the wheel." So far it is working at one of the sites. | ||
| Pierre Dardenne (dardenne) Intermediate Member Username: dardenne Post Number: 18 Registered: 3-2002 |
Hi, If you googled "nir global calibration package feed", you would find what you need. But if you want to reinvent the wheel, this is up to you. Pierre | ||
| Razaq Balogun (razbalogun) New member Username: razbalogun Post Number: 1 Registered: 1-2007 |
I am not an expert in NIR technology but have some knowledge of the technology. I am leading a project on NIR calibration for a feed company. We have three mills and we plan to purchase three NIR instruments (same) and use one as the master instrument for developing calibrations. Two are already purchased mid 2006 (only one being used currently). The other will be purchased soon. My questions are: 1. Will there be any age effect on test results from these instruments? 2. The plan is to collect spectra from the three instruments and transfer to a software for calibration development. How best to implement this? I intend to use PLSI method for developing calibrations. 3. I do not intend to use equal number of samples for each of the instrument. Is this an issue? 4. For practical purposes, what is the rule of thumb for the number of data points required to generate a reasonable calibrations for feed ingredients and mixed feed? 5. I also like to have a book on NIR technology that will be useful for feed industry and not for academic purpose. 6. Are there any factors I need to be aware of or take into consideration to enable allow proper use of calibrations? Razaq | ||
| Lewis Kansas |
We've developed an NIR calibration model using an equal number of scans from each of 10 instruments. The instruments are all of the same type. With an independent validation set, the 1st group of 4 instruments shows a significantly lower SEP than the 2nd group of 6. We have seen this with other calibrations. By increasing the number of factors from 5 to 6, the SEP of the 2nd group drops to that of the 1st . Is it acceptable to increase the number of factors to adjust for instrument variability ? As for the instruments, the 1st 4 were purchased at about the same time, the other 6 about a year later. Bandwidth and other critical specs were closely matched. Diagnostics show no significant differences between instruments. | ||
| Bruce H. Campbell (Campclan) |
Lewis, I would think it not only would be "permissable" to increase the number of PCs to six, but almost required. My thinking is that you have introduced another variable in the total effort, that of old and new, i.e., age. This would require another vector space to properly use the calibration; which it appears you have found to be true. Bruce | ||
| Gabi Levin |
I will agree with Bruce. I assume PLS1 was used. The small but existing difference between the instruments is something the calibration needs to learn, by adding a PC. One important thing, looking into the loading weights will also tell us if by using additional PC we are fitting something of value or just force fitting anything. Gabi | ||
| Nuno Matos (Nmatos) |
Have you tried to use the same amount of factors but making a slope and bias correction? |