| Author | Message | ||
     Henk-Jan Ramaker (pieranja) New member Username: pieranja Post Number: 1 Registered: 7-2006 |
Dear all, In the meanwhile, the calibration transfer has been completed. The results are satisfying. The creation/collection of a representative transfer set was difficult. We used the PDS algorithm to do the job. Based on a visual inspection, the transferred spectra look quite similar. Also, by comparing residuals, the majority of the spectra show good results. There were two regions that gave problems to some degree. This is probably due to the higher resolution of the new instrument. As a results, some (rather small though) peaks appear in one spectrum but not the other. I quess no algorithm can correct for that. But still, these residuals were within the spectral noise region. The calibration model that was constructed for the transferred spectra performed equally well, and sometimes better, compared to the old calibration model. In those terms, the calibration transfer can be called succesful. Thank you all for your input. Regards, Henk-Jan | ||
| hlmark (Unregistered Guest) Unregistered guest |
Henk - it seems likely that the possibility of transferring calibrations was one of the selling points the salesman promised, and used to convince [whoever it was that bought the instrument] to buy that one. Maybe it's time "to hold his feet to the fire" as they say, and let them make good on the promise, by transferring the calibration for you. Howard \o/ /_\ | ||
| Henk-Jan Ramaker (Unregistered Guest) Unregistered guest |
Hello Tony, The application concerns agricultural samples. In this case, the samples consists of compost. The current method primarily intends to monitor weekly trends of 12 quality variables. In that sense, the academic worrys we might have (concerning this particular calibration transfer) might be less relevant. It is more like a semi-qualitative method we're dealing with. Maybe that will partly saves us and we don't have to worry too much about loosing prediction strength? I am not involved in the replacement of the instrument. Probably the salesman of the FT apparatus did a good job. I think that the economic advantage of debitting for the grating instrument does not exist anymore and thus it makes sense for this company to buy a new instrument. Besides, who doesn't want to have new toys! What exactly do you think will be the most concerning? I think the subset is the biggest worry since it will not come from the calibration database. The best thing they can do is by taking samples the next weeks that reflects extreme process conditions at best. | ||
| Tony Davies (Td) Moderator Username: Td Post Number: 125 Registered: 1-2001 |
Hello Henk-Jan, I am concerned about your proposal but if Pierre says that he can do it; I believe him! There is a remaining question: What are your samples? If they are agrig/food then success is more likely than with chemical/pharma samples because of the nature of the spectra. Actually there is another question: Why? Why are you replacing a grating instrument in a system that you have used for years (and I assume successfully) with an FT instrument? Do you enjoy the challenge? Good luck, Tony | ||
| mpdc (Unregistered Guest) Unregistered guest |
It is also possible to standardise instruments. You create a virtual instrument and after every measurement calculate how the spectrum would have looked on this virtual master. This has the advantage you do not need to recalibrate after every minor change. If you run a large organisation you can standardise a spare instrument in the lab and do a "drop-in replacement" of your production machine in case of problems. Usually you will define your standardisation algorithm by scanning some well known standards. I remember Foss Electric patented melting grain in perspex blocks which fitted the instrument sample holder. Unfortunately I don't know any off-the-shelf software which will do something like this (maybe I am wrong?). It seems this is still a stronghold of the consultants and some instrument vendors. We use it in my company on a fluorescence spectrometer. The method is not perfect though, and I have no experience with copy/paste-ing a 6 factor model... | ||
| David Russell (Russell) New member Username: Russell Post Number: 26 Registered: 2-2001 |
I'm a member of the Process Spectrocopy Team at DuPont. We have the nice problem of having to migrate successful applications to new software/instruments. In most cases, if the application uses a complex model (ie: 6 factor PLS) we try to make as subtle a change as possible to avoid loss of precision/accuracy. Even a simple lamp change can result in calibration shift. | ||
| Henk-Jan Ramaker (Unregistered Guest) Unregistered guest |
Howard - Thank you for the interesting paper from Peter Griffith. As I understand, he uses quite a different apporach to the problem compared to e.g. the following people: Lin, Lo, Brown, Calibration transfer from a scanning NIR spectrophotometer to a FT-near-IR spectrophotometer, Analytica Chmimica Acta (1997) Leion, Folestad, Josefson,Sparen, Evaluation of basic algorithms for transferring quantitative multivariate calibrations between scanning grating and FT NIR spectrometers, Journal of Pharmaceutical and biomediacal analysis (2005) Especially the first paper deals with the issues that are mentioned by you (different wavelength scales and resolution). The authors use some kind of cubic funtion to match the spectrum. Then, they apply piecewise direct standardization to standardize the spectrum. This also seems like a good strategy to me. David - Indeed, it is true that this calibration transfer is required in the first place in order to use the new instrument quickly. Are you guys more involved in transfers between scanning spectrophotometers? | ||
| Pierre Dardenne (Dardenne) New member Username: Dardenne Post Number: 12 Registered: 3-2002 |
Henk-Jan, We are doing these kind of transfer for years. We transfered models from a dispersive instrument to 3 other brands of dispersive instruments and to 3 brands of FT. We use sealed cups for the transfer and we validate the SEP with real new samples. Email me if you want to know more. Pierre | ||
| David Russell (Russell) New member Username: Russell Post Number: 25 Registered: 2-2001 |
I do not want to speculate w/r/t your level of expertise, but as a practical matter calibration transfer methods usually provide only a good starting point in the process of deploying the new instrument. The difficulty is a function of how different the instruments are. Transferring a model between two identical instruments requires some work. Transferring between instruments that measure the same spectral range using similar technology (ie new model grating instrument vs old model) requires more work. Transferring between instruments measuring the same spectral range using different technology is even more difficult. The other significant factor is the complexity of the original model. If it did indeed take years to develop a satisfactory calibration set, it could take just as long to accomplish comparable results on the new instrument. However, if from the old data you can identify the most important members of the calibration set, then you have a basis for doing calibration transfer. How close to satisfactory the results of the calibration transfer will be can only be determined by doing the work. | ||
| hlmark (Unregistered Guest) Unregistered guest |
Henk-Jan - I haven't done that sort of calibration, but you may be interested in Peter Griffiths' work in transferring wavelength standard data between FTIR and grating instruments. It was reported in JNIRS; 11(4); p.229-240 (2003). The biggest instrumental problems in transferring data between instruments using the different technologies are the changes in both the wavelength scale and the spectral resolution across the wavelength range. Peter worked out the theoretically correct transfer functions and applied them to the FTIR data to simulate what the corresponding grating data would have been. Briefly, for every data point he took the FTIR spectrum, converted it to the Fourier domain, applied the weighting function that gave the theoretically correct corresponding value for the wavelength being analyzed, and reconverted back to a spectrum, so that for a spectrum consisting of n wavelengths he performed the algorithm n times, each time with a different weighting function. It was very computationally intensive, but in principle at least, not very complicated. The details are well-described in the above reference. That was part of a larger study on wavelength calibration of FTIR spectrometers, and you may wish to obtain the other related papers as well. Howard \o/ /_\ | ||
| Henk-Jan Ramaker (Unregistered Guest) Unregistered guest |
Dear all, I have the following situation: There exists a database of spectral + reference measurements. Based on this database, a calibration model is developed in the past. The spectroscopic instrument that is used concerns a grating instrument. The database is the result of years of measurements. Now, the instrument is being replaced with a FT instrument. There is no time to collect years of data on this new instrument. So, some sort of calibration transfer is required in order to continue with the new instrument. As it seems, the only option that is available to do so reads as follows: A new subset set of approximately 30 samples is being measured on both instruments. Ideally, a subset is created from the database. However, the samples themselves are not available anymore. The spectra from the grating instrument is taken to be the slave instrument. This makes the FT instrument the master instrument. By e.g. PDS the most suited transformation F is determined. Once F is available, all grating spectra from the large database are transformed using F so that they 'match' the FT spectra. Then, a new calibration model is developed. I was wondering if anybody has followed this strategy before? If so, what was your experience? Or maybe you have some comments or remarks for me based on the above? Regards, Henk-Jan |