| Author | Message | ||
     dumin (dumin) New member Username: dumin Post Number: 1 Registered: 11-2011 |
Hi Fernando , You refered that"Some years ago I make a model with samples milled and in some momment the customer need sample of the same type but without milled. Then, we make the follow : take a sample without milled and registry the result, after I mill the same sample and registry the result, the same was made for 20 samples covering all the analitical range. After that we observe a constant bias in the prediction between the 2 groups ( milled and not milled), we adjust the bias of the model and now the customer measure all without milled." I was wondering how to "adjust the bias of the model"to guarantee the accuracy of the calibration model? Regards, Dumin | ||
| Fernando Morgado (fmorgado) Senior Member Username: fmorgado Post Number: 27 Registered: 12-2005 |
Hello : A last time comment. My opinion is if you make the model with samples milled you need masure the problem sample in the same conditions. Some years ago I make a model with samples milled and in some momment the customer need sample of the same type but without milled. Then, we make the follow : take a sample without milled and registry the result, after I mill the same sample and registry the result, the same was made for 20 samples covering all the analitical range. After that we observe a constant bias in the prediction between the 2 groups ( milled and not milled), we adjust the bias of the model and now the customer measure all without milled. Of course maybe this not work for all type of samples. But, for me, always the models created with milled samples have better performance. regards Fernando | ||
| Marijana Maslovaric (vidra) Junior Member Username: vidra Post Number: 6 Registered: 5-2011 |
Thanks everyone! Sorry for the delay! Unfortunately, I had many other duties that are not related to NIR. I will surely take into consideration all of the advices! Best regards, Marijana | ||
| Gustavo Figueira de Paula (gustavo) Senior Member Username: gustavo Post Number: 27 Registered: 6-2008 |
My two cents: I'm measuring diffuse reflectance of roasted and ground coffee for about two years, without any sample pre-treatment. Particle sizes range from very coarse to very fine, with mixtures of coarse and fine in the same sample. I already tried a lot of data pre-processing and found the best one being SNV to classify quality index by PLS and MLR. For soybean I don't believe things will be much different. Gustavo. | ||
| David W. Hopkins (dhopkins) Senior Member Username: dhopkins Post Number: 192 Registered: 10-2002 |
Hi Jim, I've been thinking about your observation.... My observation about the spectra of Soybean Meal (SBM) certainly does not rule out a possible contribution of specular reflection, if it were sometimes, or even always present at say 10%. It would just be another source of non-linearity in the spectra, which would complicate the regression results. How small are the pixels in your images? Would you be able to detect specular reflection contributions in coarsely ground wheat? Do you have access to any SBM? Best regards, Dave | ||
| David W. Hopkins (dhopkins) Senior Member Username: dhopkins Post Number: 191 Registered: 10-2002 |
Hi Mike and Jim, You raise interesting points. Mike, the size of the sample the DA7200 is interrogating is indeed large, the size of a normal petri dish, I think, and about 10 cm diam. Because Marijana is measuring milled soybean meal of various coarseness, the number of particles being interrogated is very large. The big problem she faces is, the depth of penetration varies with the particle size. She may find that her best results will be obtained with an MSC or SNV pretreatment. Jim, I think your observations are very good, the specular reflectance on whole grain wheat can be severe, and I'm surprised that you only see the effect on 10-15 % of the kernels. However, I don't think this effect is seen with soybean particles. It may be that the grinding coats the particles with a small amount of oil and fine particles, so that the specular reflectance component is negligible, but I have never observed a problem with SB meal samples measured in reflection. The problem is certainly seen with whole SB seed, I expect. I think the most successful whole grain measurements use transmission measurements, which avoid the specular reflection problem. Best regards, Dave | ||
| Michael C Mound (mike) Senior Member Username: mike Post Number: 62 Registered: 7-2007 |
Guys: I would like to offer an analogy: In evaluating the accuracy of a measurement using other techniques, such as XRD, it is an issue as to how many crystallites (down to the unit crystallographioc cell, if possible) one can create per sample interrogated. That is, the increased number of specific sample members in a population increases the statistics of the results and always improves the accuracy. Spot analyses are subject to question as they may not represent truly the mass being evaluated. That is, the specific "spot" being measured is only as accurate as the number of results per crystallite population. So it would seeem to be at least relevant to NIR. This principle holds true for XRF, XRD, PGNAA, and most other analytical comparometrics. Think of the numbers of available valid reflectance molecules providing the vibrational and how many can provide the modeled "fits" to a patterned database, and you can imagine how the math helps improve the vagueness of the actual reported signals. However, it may be that it is better to be using as great a population of responses than a neurally constructed network or otherwise corrected spectral report might represent true and reliable results. Thus, milling might well make a difference. Just a thought. Mike | ||
| Jim Burger (jburger) New member Username: jburger Post Number: 5 Registered: 11-2010 |
Hi Marijana, Does milling make a difference? The most informative approach to YOUR samples, would be to measure a set of training samples from both populations AND a set of independent samples. Then mill all samples and measure everything again. Construct two different models. Which approach gives better results? Btw... Yes, the NIR spot probe can provide remarkable results. However as a slight precaution: I recently performed hyperspectral NIR imaging (HSI) of a set of wheat samples (100 x 100 micron pixels of whole wheat kernels in a petri dish). About 10 - 15 % of the individual spectra showed some kind of 'saturation' due to specular reflection. With HSI, I can filter out any saturated spectra - but keep in mind if I had used a spot probe for the same analysis, the 'average' spectrum would not be saturated, however some of the contributing signal may in fact be saturated, without my knowing it! Such an effect 'may' be reduced if the wheat had been milled. Does anyone have any experience with this? -Jim | ||
| Marijana Maslovaric (vidra) New member Username: vidra Post Number: 5 Registered: 5-2011 |
Dear David, Thank you for clarifying this, and for the ideas. I also thought that I could develop both one and two calibrations. I'm measuring moisture, protein, fat, fiber and ash. Kind regards. Marijana | ||
| David W. Hopkins (dhopkins) Senior Member Username: dhopkins Post Number: 190 Registered: 10-2002 |
Hi Marijana, The large field of view or the DA7200 is designed to do the averaging over sample inhomogeneity. Of course, it would be best if you had samples all of the same type of particle size distribution, but I would advise trying to include all of your samples, without any extra milling, in the same calibration. After all, NIR is a technique that touts its ability to measure samples "without preparation". You can look at the PCA scores plots for your samples, to see whether the population is badly distributed, or just widely distributed, and decide whether a single calibration, or possibly 2 calibrations will be appropriate. Try it, and see what accuracy you are able to achieve with one calibration for the population, for each quality you want to measure. I expect that the moisture calibration will work fine with a single calibration, and probably also protein. Maybe oil. Are you measuring any other constituents? Good luck. Best wishes, Dave | ||
| Marijana Maslovaric (vidra) New member Username: vidra Post Number: 4 Registered: 5-2011 |
I have a set of soybean meal samples. Some of them are pretty homogeneous, small particle size, but some of them have both big and small particles. I'm using PERTEN DA 7200, which doesn't require milling. But, I was wondering if these samples could be in the same calibration set. If I mill these non-homogeneous, should I mill them all? Regards, Marijana |