master's thesis
Design of RBF Neural Network for Process Multiple Model Control

Robert Šojo (2014)
Josip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
Department of Computer Engineering and Automation
Chair of Automation and Robotics
Metadata
TitleProjektiranje RBF neuronske mreže za upravljanje višestrukim modelom procesa
AuthorRobert Šojo
Mentor(s)Dražen Slišković (thesis advisor)
Ratko Grbić (thesis advisor)
Abstract
Kako bi se estimirala teško-mjerljiva procesna veličina potrebno je raspolagati modelom procesa. Procesi su po svojoj prirodi nelinearni, a izgradnja nelinearnog modela i njihova implementacija je bitno složenija u odnosu na linearne modele. S obzirom na pretpostavku da se nelinearni proces u određenom dijelu radnog područja može dovoljno dobro opisati s linearnim modelom, cijelo radno područje procesa se rastavlja na nekoliko manjih dijelova, kako bi se vladanje procesa opisalo s više linearnih modela. Za tu svrhu koriste se metode klasteriranja podataka, kojima se od ulaznog skupa mjernih podataka, prikupljenih kroz duže vremensko razdoblje stvaraju manje grupe podataka, koje se zovu klasteri. Svrha grupiranja podataka leži u tome da se dobiju klasteri koji sadrže međusobno slične podatke, koji su međusobno različiti nasuprot podataka koji su smješteni u druge klastere. U slučaju nailaska novog mjernog uzorka za koji se treba odrediti izlaz modela procesa potrebno je omogućiti aktivaciju odgovarajućeg modela iz ovog skupa linearnih modela, čime je ovaj višestruki model upravljan. U ovom radu upravljanje višestrukim modelom procesa obavlja se RBF neuronskom mrežom čiji se broj neurona i receptivna polja pojedinog neurona određuju predstavljenim metodama klasteriranja. Ako testni podaci pripadnu samo jednom linearnom modelu, izlaz RBF neuronske mreže predstavlja izlaz tog linearnog modela. U slučaju da testni podaci pripadaju u više linearnih modela, tada bi se koristilo meko određivanje težina za svaki linearni model, a iznos težine ovisi o raspodjeli testnih podataka unutar klastera. Ovako svi linearni modeli imaju doprinos u estimaciji teško-mjerljive veličine. Ovi modeli opisuju vladanje procesa u pojedinom dijelu radnog područja.
KeywordsProcess model neural network data clustering multiple linear process model
Parallel title (English)Design of RBF Neural Network for Process Multiple Model Control
Committee MembersRobert Cupec (committee chairperson)
Dražen Slišković (committee member)
Emmanuel Karlo Nyarko (committee member)
GranterJosip Juraj Strossmayer University of Osijek
Faculty of Electrical Engineering, Computer Science and Information Technology Osijek
Lower level organizational unitsDepartment of Computer Engineering and Automation
Chair of Automation and Robotics
PlaceOsijek
StateCroatia
Scientific field, discipline, subdisciplineTECHNICAL SCIENCES
Computing
Process Computing
Study programme typeuniversity
Study levelgraduate
Study programmeGraduate University Study Programme in Computer Engineering
Academic title abbreviationmag.ing.comp.
Genremaster's thesis
Language Croatian
Defense date2014-10-17
Parallel abstract (English)
In order to estimate hard-measured process, it is necessary to have a model of the process. The processes are by their nature non-linear and building non-linear models and their implementation is rather more complicated compared to linear models. Given the assumption that the nonlinear process in a certain part of the work area can be sufficiently well described by a linear model, the entire operating range of the process breaks down into several smaller parts in order to describe the process with multiple linear models. For this purpose data clustering method are used for creating smaller groups of data which are called clusters. Each cluster contains similar data, which are different from other data that are located in other clusters. If a new measurement sample arrives, for which is needed to determine the output of the process, it is necessary to allow the activation of appropriate model from this set of linear models. In this paper, the management of multiple process model is done by RBF neural network whose number of neurons and receptive fields of each neuron is determined by presented clustering methods. If the test data belongs to a single linear model, the output of RBF neural network represents the output of that linear model. In case that the test data belongs to the several linear models, then the soft determination of weights is used for each linear model and the amount of weight depends on the distribution of test data within a cluster. In that case all linear models have a contribution to the estimation of hard-measured value. These models describe the behavior of the process in a particular part of the work area.
Parallel keywords (Croatian)Modeliranje procesa neuronske mreže klasteriranje podataka višestruki linearni model procesa
Resource typetext
Access conditionOpen access
Terms of usehttp://rightsstatements.org/vocab/InC/1.0/
URN:NBNhttps://urn.nsk.hr/urn:nbn:hr:200:075109
CommitterAnka Ovničević