https://www.euro-online.org/enog/inoc2007/Papers/mac-slots.html https://www.euro-online.org/enog/inoc2007/Papers/m https://www.euro-online.org/enog/inoc2007/Papers/mac-slots.html

Differences between NIR Instruments

mp08's picture
Forums: 

I'm fairly new to NIR and have been having some issues that I was hoping someone on here may be able to give some guidance.

We have two different NIR machines (same brand, model, and age) - one in our QC lab and one in our Production control room. We use the NIR for product quality. We developed curves a few years ago using the QC NIR and then transferred that calibration curve to the Production NIR. The Production NIR also has a higher baseline than the QC one.

The problem we are seeing is that the readings from the Production NIR is very different then the ones on the QC NIR as well as much more variation is seen on the Production NIR. We tested this by taking a sample from a few batches and running the same sample on both NIRs 10 times for replication. Would anyone know where to first start looking as to what is wrong with our NIR instruments?

My other question is whether or not we did the correct thing and transfer the calibration curves between the instruments? From reading other posts on here, it seems like its possible but extra steps are needed and I know we didn't do anything except transfer the files from one machine to another.

 

Thanks

russell's picture

Calibration transfer between two nearly identical instruments often works if baseline correction was utilized in the original calibration.

In your case, if it was not, since you've observed a difference in the backgound, the results would simply be wrong.

First step would be to examine the spectra than you ran on both instruments to see if there are other differences besides the baseline change.

If a pre-treatment can be identified that minimizes the differences, the orignal calibration could be redone utilizing that pre-treatment.

The calibration should then transfer to the production instrument sucessfully.

jcg2000's picture

First all, no two instrument are the same even if they are assembled by the same person using the same parts. Calibration trasnfer (or instrument standardization) is required before you use calibration model obtained from one instrument on another. Google "NIR calibration transfer" and you will find a lot.

Don't expect to see the same spectra when you measure the same sample on a  lab instrument and production-floor instrument. The environment, lamp, sample presentation all affect photometric response (the Y axis of your spectrum), but you should see all peak locations are all the same (wavelength accuracy) between instruments. Otherwsie, talk with the customer support about instrument  qualification. Once you have both instrument qualified, calibration transfer should resolve your problem.

 

Jerry Jin

 

shileyda's picture

Could be nothing "the matter with" either of your instruments, just differences caused by differences in components. I recommend that you look at the spectral residual measurement for the predictions and also correlation of reference to the predictions as well as standard error of prediction. These should help you understand whether the spectra are really different or not. If the average spectral distance for both systems is the same then there is no real difference between the two systems and the correlation and SEP are similar you may just be able to use a slope/bias adjustment to the predicted results. If the metrics are different then the instruments are not spectrally similar and you would either need to optically clone the instruments so they are more similar, or create a calibration for each device.

regards,

Dan

gabiruth's picture

Hi,

All the advice that you have been given is well and good. The main question remains - how do I fix the issue quickly. Well, here is what I would do -

I would collect 30 to 40 samples on the floor instrument and on the lab instrument. These samples shall best cover as much as possible of the range of values for all relevant measured parameters.

I would give the lab these samples and have them provide me with high quality ref values.

I would then insert these samples into the chemometric software, add them to the existing set of data and run a new regression and create a new model - that uses both instruments. Since the differences between the instruments are constant and not maple dependent - the new correlation factors and loading weights will actually zero these differences.

Now you will have one model that is good for the two instruments.

I have done such procedures many times and they work better than many elaborate schemes. The power behind this approach is:

1. The differences between instruments are constant and sample independent.

2. The power of good chemometric software to create the needed "zeroing" factors to "represent" these constant differences is very high.

Obviously, it will most likely require maybe one more PC to achieve that - so what?

People know that I am a practitioner - not a theoretician - but on the floor what matters is if it works, works well, reliably and robust enough to endure the variability of the product.

Gabi Levin

Brimrose Corp. of America

td's picture

Hello MPo8,

Welcome to the Forum,

Before you follow any of the good advice above I think you should check the "Production" spectrometer. My question is "Why is the Production spectromer producing such variable results?" You should check the noise of this instument and compare it with your QC instrument.

Of course I do not know your location but if you are in an area of high humdity this could be a source of the problem. This is especially true if you have air conditioning in your QC lab but not in the production area. Atmospheric water vapour is source of variable noise. If you were not told about this I can give you some references.

I would also be intersted to see spectra of the same samples recorded on each of your spectrometers. You can upload them to this area.

 

Best wishes,

Tony

mp08's picture

Thanks for the comments. We have been thinking about re-doing the calibration scans on both instruments since there is the difference. We're also looking into purchasing a new NIR, but don't want to until we have a better understanding on the instruments and how to transfer the calibration data sets.

We're located in New Jersey so high humidity isn't a problem right now. 

Right before we did the tests, we did replace the infrarometer (?) on the QC NIR. We did a before repair and after repair test (same sample 10 times) and saw that the variability did go down after the repair. We're not sure if this helped us since the technician said it shouldn't cause the issues/improvements we've seen.

We had some scans sent to a technical resource from the manufacturer and he did a PLS1 calibration  and plotted a residual spectrum. He basically said that the M-Distance and Residual Ratio were extremely high for Production scans and the residual scans between QC and Production were completely different. I've attached the report to show some of the information. 

I'll need to get the scans first to upgrade them. 

Uploaded Files: 
td's picture

Hello again,

Thank you for the extra information.

It is very strange! The short term reproduciblity of the production instrument looks better than the QC instrument but the long term reproducibility is terrible.

I think that there is something wrong with the production instrument and since you have told us that the inferometer has been replaced this should be invesigated. I suggest you ask the manufacturer to check that it was installed correctly.

There is more information that you could give that would be helpful:-

1) Spectra of a standard sample from both instruments.

2) What is the analyte you are measuring?

3) The report mentions a vial. Are the samples being measured in transmittance mode? Are samples being position correctly?

Best wishes,

Tony

mp08's picture

It was the QC instrument that was repaired. The Production has not had any repairs.

I do have some scans but was wondering how they should be upload. Does Word work or are the original scans in the Spectrum format better?

ianm's picture

Probably the easiest is copy and paste them into Word, although the WYSIWYG editor does allow images to be inserted, but this may mean they end up a bit small.

Any problems, get in touch: http://www.impublications.com/users/ianm

Ian

mp08's picture

So these are going to have to come up over a few posts.

All these scans are from the same sample. Thist post has the scans for the first test on Production NIR.

mp08's picture

Scans from second production NIR test

 

mp08's picture

Scans for first QC NIR test

ianm's picture

It sounds as if you should start over. The steps are described in the attached report, assuming that the sample is presented to the instrument in a sample cell. If the sample is presented in some other way there are different steps.

(Posted by Ian Michael on behalf of Phil Williams)

mp08's picture

Scans from second QC NIR test

Again, any help is appreciated. All we know is that the two machines are giving different readings. We plan on changing some setpoints (ex:wavelength) based on the manufacturer's recommendation but we're not sure how that will effect either machine

We're also looking at buying a new NIR machine, so I have to work on developing a plan to make sure we don't run into this problem again.

 

Thanks.

jcg2000's picture

Can your software overlay spectra of the same sample measured from two instruments on the same screen? It is difficult to see how these individually attached spectra differ from each other. Or you could upload raw data and we can plot the spectra.

We can make suggestion only if we understand where the difference is. 

 

Jerry Jin