Data-Driven Evolving Fuzzy Models - Algorithms and Applications
Sprache des Vortragstitels:
In nowadays industrial systems there are quite a lot requirements for an incremental modelling of system behaviors
from data. These requirements arise due to a fast tracking of dependencies between system variables
with online recorded measurements or due to huge data-bases which cannot be loaded into virtual memory at once.
Improving the process security by extending already available models to new operating conditions or by adjusting them
on the basis of some feedback from operators is an important issue as well.
In this talk algorithms for a data-driven incremental learning of
fuzzy models are demonstrated. The main focus is given to fuzzy basis function networks (as a
specific type of Takagi-Sugeno fuzzy models), but some aspects how to extend the approach to other types
of fuzzy models are illuminated as well.
A modified version of vector quantization is exploited for rule evolution and an incremental learning of the rules'
premise parts. The modifications include the generation of new clusters due to the nature, distribution and quality of new
data, point-wise update of already existing clusters and an alternative distance strategy for selecting the most adjacent
clusters for each new incoming sample. The premise part learning is connected in a stable manner
with a recursive learning of rule consequent functions possessing linear parameters.
Stability can be achieved by preventing
the 'unlearning' effect in case of steady states and by ensuring that a suboptimal solution of the parameters
can be achieved which is close to the optimal one in the least squares sense
(i.e. with respect to least squares as underlying optimization function).
Some aspects about maintaining the interpretability of the trained fuzzy models round off the algorithmic part
of the talk.
The second part is focussed on applications of data-driven evolving fuzzy models in industrial systems.