| Summary: | 碩士 === 國立成功大學 === 化學工程研究所 === 81 === In this thesis, we show by several examples that the multi-
layer percetron(MLP) is a very powerful tool for systems
identification. In particular, MLP is suitable for identifing
static systems and dynamical systems without time delay. We
propose an MLP-based control structure which can control both
linear and nonlinear systems without time delay, can accomodate
external distrubances and can be extended to MIMO systems. In
general, the characters of the proposed control structure can
be listed as follows: (1) MLPs are used both the system model
and controller. (2) The system model is trained off-line before
the controller taking action, while the controller learns on-
line; (3) It employs the back-propagation algorithm, and
computes control command through pre-trained system model; (4)
It applies an IMC-like loop to compensate external
distrubances; (5) It adjusts the conserved factor $\beta$ to
control the time-response characteristics; (6) It is suitable
for linear and nonlinear systems without time delay; (7) It is
sutiable for SISO and MIMO systems.
|
|---|
