| Summary: | A near infrared (NIR) spectrometer produces, from a single sample, a spectrum formed from several hundred absorbance readings at a range of wavelengths in the NIR region. Using regression approaches and a large number of samples for which reference values and spectra are known, the instrument can be calibrated to predict reference values from spectra. A problem with NIR spectrometers is that no two instruments produce exactly the same output, as a result of which a calibration developed on one instrument cannot be transferred to a second instrument unless the second instrument has been standardised first. Our aim in this thesis is to explore and assess improved methods of standardising NIR spectrometers. The main line of attack is to use standard models but incorporate prior information through Bayesian techniques. The main commercially used standardisation techniques adjust the spectra wavelength by wavelength without any use being made of the fact that the spectra and therefore the appropriate adjustment varies smoothly. By the use of suitable priors within a Bayesian analysis we produce a better solution. The analysis is very time-consuming, involving inverting large matrices and MCMC or some other process for determining parameters. A second attempt using the same assumptions uses dynamic linear modelling, treating the spectra as time series. While theoretically slightly inferior, this method is very much quicker and produces comparable results. A third solution, while using the same basic model, makes an estimate of the wavelength shift in the wavelet domain. Our final, non-Bayes, method is intended to standardise a number of similar instruments simultaneously. This is achieved by projecting spectra onto a subspace orthogonal to the space spanned by between-instrument variation and calibrating on the subspace to produce a robust calibration.
|
|---|
