As we know a NIR instrument is costly to calibrate. Laboratory reference analysis are expensive, and it requires a long time to get the necessary variability, for a certain product, to get a calibration we can trust to start analyzing in routine, that is the reason why some users buy the NIR instruments with a calibration package for different products.
Anyway we have to test this calibration with “our own test set” to check the performance of the calibration, and check if we have a bias, or slope and intercept problems.
In this case to get a good test set is important. It has to cover the variability of our suppliers (raw ingredients) or the variability of our final product.
One way to do it is to analyze all the income samples we are receiving for some time, storing the spectra, and after keep these samples in plastic bags or containers. We can analyze also samples of our process (at the start, in the middle or at the end of the batch, several batches,) and store the samples and the spectra.
After 15 or 30 days we can have more than 100 samples, for example, so now we can send to the Lab the samples with more variability per product.
The way to do it is to create a Principal Component Space for these samples and select the spectra with more variability. We can look also to the spectra looking for the bands of water, protein, stark or fat and try to find some relation between the scores and the spectra.
Win Isi has a function to select the number of samples with more variability, that is the one I show you, but sure other programs have similar tools. We can analyze again the selected samples in the NIR instrument, and send them immediately to the Lab.
This way we can evaluate better the calibrations we have acquired with the instruments and make better adjustments.
With the 100 spectra create the PCA and LIB files with the function "Create a Score File from a Spectra File".....after use the function "Select Samples from a Spectra file", and as Input use the 100 samples NIR file and in Options, say to the software that we are looking for 10 (or more) samples with most variability. In the output file we get this file telling us which samples we have to send to the Lab.