Continuum Regresion in Process Modeling Based on Plant Data

Važne procesne veličine koje daju informaciju o kakvoći izlaznog proizvoda često nije moguće mjeriti senzorom nego se njihov iznos utvrđuje laboratorijskom analizom. Kako bi se omogućilo kontinuirano praćenje tijeka procesa te efikasnije upravljanje proizvodnim procesom, ovu teško mjerljivu procesnu veličinu je potrebno estimirati, tj. odrediti na temelju matematičkog modela. Za izgradnju odgovarajućeg modela procesa vrlo često su na raspolaganju samo procesni mjerni podaci pohranjeni u procesnu bazu podataka. U ovom se radu prikazuje prikladna metodologija za modeliranje procesa na temelju pogonskih podataka. Za izgradnju modela pri tome se predlažu regresijske metode zasnovane na preslikavanju ulaznog prostora u latentni potprostor. U radu se posebno istražuju svojstva kontinuum regresije (CR). Budući da neuronske mreže predstavljaju dobru osnovu za izgradnju modela na podacima, dopunski se istražuje mogućnost hibridizacije višeslojne perceptronske (MLP) neuronske mreže i CR metode, s ciljem iskorištavanja dobrih svojstava obiju metoda te izbjegavanja njihovih nedostataka u izgradnji modela procesa na pogonskim podacima. Prednosti predloženih metoda izgradnje modela procesa nad uobičajeno korištenim regresijskim metodama prikazane su na primjeru modeliranja procesa destilacije nafte na raspoloživim mjernim podacima.Important process variables which give information about the final product quality cannot often be measured by a sensor but their value is determined based on laboratory analysis. In order to perform a continuous monitoring of a process variable and an efficient process control, it is necessary to estimate this difficult-to-measure process variable, i.e. to determine it on the basis of a mathematical model. However, to build an appropriate process model in many cases there are available only process measurement data stored in a process data base. This paper gives appropriate methodology for process modeling based on plant data. Regression methods based on input space projection into a latent subspace are proposed to build a model. The paper investigates, in particular, properties of continuum regression (CR). As neural networks present a good basis for data based model building, possibility of hybridization of multilayer perceptron (MLP) neural network with CR method is additionally investigated. The aim of that is to use good properties of both methods and to avoid their weaknesses in process model building based on plant data. Advantages of the proposed methods for process model building as compared to the usually used regression methods are demonstrated by the modeling of crude oil distillation process based on the measuring data available