Integration of design of experiment techniques and intelligent systems for modelling and prediction
Loading...
Date
2009
Authors
Kho, Hiaw San
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In this thesis, hybrid models comprising Design of Experiment (DOE) techniques
and intelligent systems for tackling data-based modelling and prediction problems are
presented. Traditionally, DOE techniques employ statistical regression methods to study
the relationship between the explanatory variables and response variables in observational
studies, and an empirical model is then built for prediction purposes.
In this research, the fuzzy linear regression (FLR) approach is proposed as an
alternative for model building in DOE whereby statistical regression methods are
inappropriate. The usefulness of the FLR model is first evaluated using two benchmarks,
i.e. the injection moulding process and wire electrical discharge machining (WEDM) process,
and the performance is evaluated using the R2 value. The results show that FLR is able
to improve R2 by at least 27.13% as compared with those from statistical regression
methods. However, the limitation of FLR is that a lot of parameters need to be pre-set
empirically in order to achieve good results. Thus, the FLR is integrated with the GA
(Genetic Algorithm) to produce a novel FLR-GA model. The GA is used to facilitate
parameter tuning in FLR and to improve its robustness and prediction capabilities.
The effectiveness of the proposed FLR-GA model is demonstrated using another
benchmark problem, i.e., the epoxy dispensing process, in addition to the injection moulding
and WEDM processes. As compared with FLR without GA, FLR-GA is able to improve
the R2 values by 39.33% (injection moulding), 24.21% (WEDM process), and to reduce
the prediction error percentage from 6.24% to 0.24% (epoxy dispensing process). Besides
benchmark studies, two real modelling and prediction tasks, i.e., a stereo vision system and a
soda pulping process, are studied. Performance comparisons between FLR-GA and LSR
(Least Square Regression) are conducted, the improvement achieved is 4.77% and 35.87%
respectively. These results positively demonstrate the potentials of FLR-GA as an effective
model building tool in DOE techniques for undertaking data-based modelling and prediction
tasks.
Description
Master
Keywords
Science Physic , Design , Intelligent systems