M.Sc. (eng) thesis

Om oppgava (english below)

Jeg skrev en masteroppgave (Master of Science in engineering) ved institutt for prosessteknikk på Høgskolen i Telemark i 1999. For å få den oppgava jeg hadde mest lyst på, måtte jeg inngå en avtale om fortrolighet med bedriften jeg samarbeidet med. Avtalen var at oppgava skulle være unntatt offentlighet i 10 år. I ettertid har jeg angret på dette, akademiske arbeider bør være tilgjengelig for alle. Men nå har det i alle fall gått 10 år, og her er oppgava for nedlasting:


This thesis deals with the calibration of an on-line near infra-red (NIR) instrument measuring on high density polyethene (HDPE) powder. The goal was to recommend an “optimal” calibration strategy for the dynamic prediction system. Important aspects in relation to the calibration were thorough understanding of the NIR spectrum and of the sampling system. The physical polymer properties of interest in this application were density and melt flow rate (MFR).

Because of the collinearity and lack of selectivity in the NIR region, it is necessary to use many absorption frequencies and multivariate regression techniques like principal component regression (PCR) or partial least squares regression (PLSR).

The relation between polymer density and the NIR spectrum looks simple and direct using PLSR; a higher co-monomer content in the polymer lowers the density and give origin for an increased methyl content. Methyl (and methylene) is known to have overtone absorption bands in the NIR spectral region. The relation between the MFR property and the NIR spectra, is not as clear as for density. MFR predictions were found infeasible in this application.

It was found that particle properties (size a.o.) constitute the major variations in the NIR spectra, in addition to the problems with varying polymer concentration and a polymer film on the sampling system window. These variations should be reduced to a minimum prior to calibration.

The recommended calibration strategy for density predictions, is:

  • Careful selection of the calibration data.
  • Abandon the upper part of the acquired NIR spectrum.
  • Pre-processing to reduce all baseline effects.
  • A rigid outlier detection system.
  • Continuous calibration model updating.

The sampling system under study suffers from many weaknesses, resulting in very unstable and noisy spectra. Variations in polymer concentration in the sampling system and a polymer film on the sampling system (sapphire) window are probably the two major problems.

Another problem with the system under study is the introduction of several measurement biases of unknown magnitude. These make it impossible to give reliable estimates of the prediction accuracy.

A rebuild of the sampling system is probably necessary to overcome the bias problem, and also for improving the spectrum quality.