bitjungle

(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:

M.Sc. (eng) the­sis: On-line NIR ana­ly­sis in a high den­sity poly­et­hene plant, eva­lua­tion of samp­ling sys­tem and opti­mal cali­bra­tion strategy

Abs­tract

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.