| Summary: | 博士 === 國立臺灣大學 === 電機工程學系研究所 === 86 === The fast-acting drug sodium nitroprusside (SNP) is often administered to l
ower mean arterial blood pressure (MABP) in hospitalized patients. It is tedio
us, time-consuming and may yield undesirable or even hazardous oscillations in
the level of MABP due to lack of timely adjustment of infusion or over-correc
tion for the manual adjustment of the SNP infusion rate. Thus, closed-loop fee
dback controllers is necessary to maintain MABP near a desired level because o
f disturbances that perturb blood pressure, the changing condition of patient
and the wide rangeof response characteristics among patients. The automatic ar
terial blood pressurecontrol system is composed with a pressure/voltage piezoe
lectric transducer mounted on femoral or brachial cannula, a blood pressure po
lygraph or patient monitor, a personal computer (PC) as controller, and a medi
cation drug infusion pump. The mathematical model of MABP of a patient und
er the influence of SNP infusionrate is a time-delay, time-varying, nonlinear
single-input/single-output (SISO) system and corrupted with much noise. The ga
in of patient characteristic can vary as much as 36 fold from one patient to t
he next. Furthermore, a patient''s characteristic also change during the course
of therapy. Thus, the traditional control theory, such as nonadaptive control
ler, optimal controller, single-model adaptive controller or multi-model contr
oller, is difficult to achieve good and robust performance, and meet the clini
cal constraints. In this thesis, a new hybrid intelligent control strategy
is proposed by combining neural network and fuzzy-logic algorithms to control
the time-varying single-input/single-output (SISO) system. A model with an aut
oregressive moving average, representing the dynamics of the system, and a mod
ified back-propagation training algorithm are used to design the control syste
m to meet specified objectives of design (settling time and undershoot/oversho
ot) and clinical constraints. We present a parallel two-model multilayer neura
l network (MNN) controller structure to approximate the large dynamic range of
parameter gains and time-varying plant. One MNN controller is to map the lear
ned range of large-gain, and the other is for the range of small-gain, functio
n of the system characteristics. The two-model MNN controller is also associat
ed with a weighting determinant (WDU), such as fuzzy-logic unit (FLU) or rule-
based unit, to determine an incremental value and update the output weighting
factor of the parallel two-model MNN controller for adequate control action.
Extensive computer simulations indicate satisfactory performance and robustn
ess of the proposed controller in the presence of much noise, over the full ra
nge of plant parameters, uncertainties and large variation of parameters, and
no requirement ofsystem parameters identification a poriori, and good signal t
racking capacity.
|
|---|
