Razvoj modela softverskih senzora primjenom "bootstrapping" metode

master's thesis

• Title (english): Soft sensor model development based on bootstrapping method
• Author: Matija Cvetnić
• Mentor: Nenad Bolf (mentor)
• Committee member: Nenad Bolf (predsjednik povjerenstva)
• Committee member: Igor Dejanović (član povjerenstva)
• Committee member: Željka Ujević Andrijić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Chemical Engineering and Technology; Zagreb
• Defense date and country: 2015-07-20, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Chemical Engineering; Analysis, Synthesis and Management of Chemical Processes

Abstract

U rafinerijskim postrojenjima ključne procesne veličine kao što su sastavi procesnih struja i svojstva proizvoda mjere se pomoću procesnih analizatora. Budući da su procesni analizatori često nedostupni, ključna svojstva moraju se određivati rijetkim i dugotrajnim laboratorijskim analizama. U radu je prikazan razvoj modela softverskih senzora za procjenu sadržaja benzena u lakom reformatu. Prikazana je usporedba više modela softverskih senzora razvijenih na temelju malog skupa eksperimentalnih podataka. Softverski senzori razvijeni su primjenom viševeličinske linearne regresije, viševeličinske adaptivne regresije (MARSpline) te primjenom neuronskih mreža. Svi razvijeni modeli pokazali su dobro vladanje na skupu podataka za vrednovanje modela uz visoke korelacijske faktore i male pogreške modela. Kao najbolji modeli među razvijenima pokazale su se neuronske mreže. Nešto lošije, ali i dalje zadovoljavajuće vladanje dali su modeli viševeličinske adaptivne regresije te linearni model. Linearni viševeličinski model razvijen na generiranim podacima daje donekle lošije rezultate uz zadovoljavajuće vladanje pri određenim procesnim uvjetima.

Abstract (english)

In refinery plants, key process variables like contents of process stream and fuel properties needs to be continuously monitored using adequate measuring devices. Measuring devices are often unavailable or malfunction, and laboratory assays, which are irregular and therefore not suitable for process control, are inevitable alternative. In this study, soft sensors for the estimation of benzene content of light reformate are developed. This research shows a comparison of soft sensor models developed on the basis of a small data sets. Soft sensors have been developed using multivariable linear regression, multivariable adaptive regression splines (MARSpline) and neural networks. Within the validation data set, all models have shown good results with high correlation factors and small errors. Best results are achieved by neural network-based models. Models developed on small data set by multivariable adaptive regression splines and by multivariable linear regression are satisfactory and still applicable for benzene content estimation while multivariable linear regression developed on resampled data set showed slightly worse behavior on particular data range.

• Keywords: softverski senzori; mali skupovi podataka; modeliranje procesa; neuronska mreža; bootstrap metoda
• Keywords (english): soft sensor; small data set; process modeling; neural network; bootstrapping
• Language: croatian
• URN:NBN: urn:nbn:hr:149:102823
• Study programme: Title: Chemical Engineering - Graduate study; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka kemijskog inženjerstva (magistar/magistra inženjer/inženjerka kemijskog inženjerstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2017-03-06 10:59:21