Professional Documents
Culture Documents
Adaptive Aortic Pressure Observer For The Penn State Electric Ventricular Assist Device
Adaptive Aortic Pressure Observer For The Penn State Electric Ventricular Assist Device
Adaptive Aortic Pressure Observer For The Penn State Electric Ventricular Assist Device
4, APRIL 1990
INTRODUCTION
Fig. 3. The Penn State mock circulatory system and the EVAD with its
controller.
Model Formulation
berger observer [3] is stable. The objective function is
defined as a weighted sum of the quadratic deviations be- A Luenberger observer [3] utilizes a mathematical
tween the system and the observer states. The optimal model to estimate the values of the state variables. This
gains of the observer are independent of the (often un- section addresses the model development of The Penn
known) system state initial conditions. This deterministic State EVAD and the mock circulatory system (shown in
optimal observer design has been introduced by Jacquot Fig. 3) during systole. A bond graph model of the EVAD
[4] for a single-state/single-output model, and is extended and the circulatory system is shown in Fig. 4. The no-
to accommodate multistatelsingle-output models in the menclature and the numerical values of the system param-
present paper. A deterministic optimal observer design is eters are listed in Table I. The motor voltage e, is the
attractive, when the signal to noise ratio is high and the input variable and the pusher plate position x is the mea-
system covariance matrices are not known. The covari- sured output. The system states are the pusher plate po-
ance matrices of a population of implant candidates may sition ( x ) the pusher plate velocity ( U ) , and the aortic
have a time dependent range of numerical values. 'The ( P a o )and atrial ( P a t )pressures. The system state equa-
covariance matrice values are required for the formulation tions are
of a Kalman filter [ 5 ] .
A major contribution of this paper is the development
of an adaptation scheme that improves the state estimates
by adjusting the observer initial conditions. Previous
schemes of adaptation [6], [7] concentrate on identifying
the system parameters and updating the observer equa-
tions. To ensure convergence, these schemes require that
the system inputs be sufficiently rich in frequencies. In
the case of the EVAD, the biological range of the model
parameters is known; and instead of improving the state
estimates through a fine tuning of the model parameters,
the state initial conditions are adjusted on a beat by beat
basis. The pulsatory nature of The Penn State EVAD fits
the initial conditions adaptation scheme well, and its real-
time implementation is straight forward.
The paper is structured as follows: a time invariant, lin-
ear state space model of The Penn State EVAD is devel-
(4)
oped and a Luenberger observer is formulated. The ob-
server gains are deterministically optimized using an
extended version of Jacquot's method, and an adaptation By substituting the numerical values of the parameters
scheme for the aortic pressure observer is developed. In (Table I) into (1)-(4) the following state space model is
376 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 37. NO. 4. APRIL 1990
SYSTOLE
--+ l +
P
ce t
Rm J
I
Rsys
Fig. 4. A bond graph model of The Penn State EVAD and the mock cir-
culatory system.
TABLE I
NOMENCLATURE
A N D NUMERICAL
VALUES OF SYSTEM PARAMETERS
(7)
(5)
y(n) = D X ( n ) = [ l 0 01
LaoiU(.)
N i
J = X ( 0 ) ’ DTKT[cY]K D values Pa, ( 0 ) :
dCR3,,/aKi = 0 where i = 1, 2, 3 . (32)
+ n= 1 { [ @ 1 ( n ) l T [ P 1[ m , c n ) l , ] X ( 0 ) . (26)
K I , K 2 , and K3 that satisfy the three nonlinear equations
(32) and result in a stable closed loop observer, are the
The optimal gain values have to equate the s partial de- desired gains. The entry CR3,3 for various N values ( N is
rivatives of (26) to zero: the number of discrete instants at which the deviations
i
between the system and observer states are minimized) is
aJ/aKi = X(0)‘ 2K,DT[a]D listed in Appendix A, and the resultant optimal gains and
N
their corresponding closed loop eigenvalues are summa-
+ n= 1 { ~ [ & I ( ~ ) ] ‘ / ~ K , [ P I [@i,(n)] rized in Table 11. The gain values obtained when N = 4,
are used for estimating the aortic pressure of the mock
circulatory system. When the EVAD’s beat rate is 60 bpm
I]
+ [ & l ( n ) l T [ a la [ a l ( n > l / a ~ i ~ ( 0 ) and the initial state values of the observer are zero the
estimated pressure turns out to be unsatisfactory (see Fig.
(27) 6). To improve the pressure estimates, an adaptation
scheme that adjusts the observer’s initial pressure value
- -
where i = 1 , 2, * , s and Ki is gain value that operates
2a,(0), is utilized.
on the ith state deviation. The partial derivative of
[(RI ( n ) ]with respect to Ki can be expressed as An Adaptation Scheme
The performance of an observer that has to estimate the
.e [
n-1 n-j-2
a [ a l ( n > l / a ~=i j = O r = O {([+I - K D Y system states in a finite (rather short) time is highly de-
pendent on the initial state deviations [ X ( O ) - X ( O ) ] .
-
k i D ( [ 4 ]- K D ) n - Z - r - ’ K D [ 4 ] j ) The observer was evaluated by comparing the estimated
aortic pressure with the measured mock loop values. Fig.
- ( [ 4 ]- K D ) . + j k i D [ 4 ] ’ ]
7 depicts a satisfactory pressure estimate of the observer
when the system beat rate is 60 (same as in Fig. 6), and
the initial pressure deviation [ P , , ( O ) - P , , ( O ) ] , is -4
(28)
mmHg (in Fig. 6 the initial pressure deviation is 69
where ki is an s x 1 unit vector having a unit entry at the mmHg). Since the EVAD is a pulsatory device, one can
ith row and zeros elsewhere: adjust the observer’s state initial conditions on-line to ob-
k T = [O 0 - . * 0 1 0 * . * 0 01. tain satisfactory aortic pressure estimates. However, to
quantify the initial pressure deviation, the value of the
Equations (27) and (28) can be combined into the follow- unmeasured aortic pressure is required. Therefore a sec-
ing compact form ond pressure estimate has to be used. Equation (7) can be
recast and a static relationship between the aortic pres-
dJ/aKi = X(O)‘a[CR]/aKi X ( 0 ) = 0 (29) sure, the pusher plate velocity, and the motor voltage can
where i = 1 , 2 , - , s, and [a] is an s X s matrix. The be written as
optimal gains that are independent of the system state ini-
tial conditions have to equate the partial derivatives of
+
$,,(T) = - 9 . 4 ~ ( 7 ) 5 8 . 3 e , ( ~ ) . (33)
matrix [ CR] to zero where T is the time at which the plate acceleration d v / d t
is zero. The real time implementpion of (33) is simple,
d[CR]/aKi = 0 where i = 1, 2, * * s. (30) and the second pressure estimate Pa, ( T), obtained for var-
7
In the general case, (30) results in s 3 equations and s ious beat rates are listed in Table 111. These pressure es-
-
unknowns ( K l , K2, * , K s ) , i.e., the problem is over- timates pa,( T), match their corresponding measured val-
determined. However, in the case of the EVAD the ob- ues well.
TASCH et ul. : PENN STATE ELECTRIC VENTRICULAR ASSIST DEVICE 379
TABLE I1
THERESULTANTOPTIMAL GAINVALUES, AND THEIR CORRESPONDING CLOSED LOOP EIGENVALUES WHEN
THE DEVIATIONS AT 2, 3, A N D 4 DISCRETE
BETWEEN THE SYSTEM A N D OBSERVER STATES ARE MINIMIZED
INSTANTS
Minimization Kl K2 K3 Eigenvalues
Points [s-ll Is-*] mmHg/( m m . s)1 [s-’l
TABLE 111
MEASURED
A N D SECONDPRESSURE
ESTIMATES
‘peak =
Second
Beat Measured Pressure
Rate Pressure Estimate Error ( mmHg )
bPm Pa,( 7 ) mmHg 7 ) mmHg
pa,,,( Pa,,( 7)--Pa0 ( 7 )
Estimated
P
e 60 106 105
70 110 109
80 122 121
90 136 135
100 141 14 1
110 155 155
120 153 152
0
0.0
I”, 0.2
‘
1 ’
0.4
1
I
’ 1
0.6
I
‘
I
I
0.8
i
I .o Note: 7 is the time at which the pusher plate acceleration d u / d r is zero.
Time [sec]
Fig. 6. The measured and estimated aortic pressures when the beat rate of
the EVAD is 60 bpm. Here the initial pressure deviation [ P , , ( O ) - adapting the observer state initial conditions. Yet, at these
P a O ( 0 ) is
] 69 mmHg and the resultant pressure estimate is unsatisfac- high beat rates the observer consistently underestimates
tory.
the aortic pressure measured at the mock circulatory loop.
It is therefore postulated that the system in its systolic
mode contains significant nonlinear dynamic effects that
Ppeak = 127
I 5 O T the linear model equations [see (7)] do not capture.
When the EVAD is in its systolic mode, the pusher plate
moves forward to eject blood through the aortic valve.
During this process, the flexible blood sac is reshaped and
blood is displaced radially, as schematically shown in Fig.
10. This phenomenon can be incorporated into the model
equations by varying the pusher plate area. Fig. 11 shows
the experimentally determined function of the corrected
blood ejection area plotted as a function of the pusher plate
position. This function was constructed by translating the
1 discrepancy between the measured and estimated aortic
0 1 I I I I I I #+I pressures into an increased blood ejection area. This non-
0.0 0.2 0.4 0.6 0.8 1 .o
Time [sec]
linear effect was incorporated into the observer equations
and the resultant satisfactory pressure predictions for 90
Fig. 7. The measured and estimated aortic pressures when the beat rate of and 120 bpm are depicted in Fig. 12.
the EVAD is 60 bpm. Here the initial pressure deviation [ P a o ( 0 )-
paO(0)] is -4 mmHg and the resultant pressure estimate is satisfactory.
Parametric Study of the Observer’s Robustness
An adaptation procedure of the aortic pressure observer To exploit the observer robustness to parameter varia-
is shown in*Fig. 8. According to the deviation between tions of the circulatory system, in vitro test runs were per-
Pa, ( T ) and Pa, ( T ) , the initial condition Pa,( 0 ) of the suc- formed. In these runs the systemic resistance, aortic com-
ceeding systole is adjusted. This adapiation procedure pliance, and atrial compliance were physically changed in
continues until the deviation [Pa,( T ) - Pa,( 7 )1, is within the loop while the observer reduced order model [see (8)
preset bounds. and (lo)] was not corrected. As a result, actual system
parameter variations were not reflected in the observer
Effects of Nonlinear Hemodynamics in the Blood Sac model equations.
The estimated pressure values for 90 and 120 bpm are The pressure estimate of the observer when the sys-
shown in Fig. 9. These pressure values were obtained after temic resistance is reduced to 0.0007 mmHg s/mm3 and
380 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 37, NO. 4. APRIL 1990
h
, 5 0 7
c
Pao(r) = 105
c)
-
3E
I8O T1 Ppeak = 162
T Ppeak = 185
I BO
E
Y
5 180 'peak = \
VI
v)
e
2
a v)
U 2
.m
Y
L
a
0
4
Time [sec]
(b)
Fig. 9. The measured and estimated aortic pressures when the EVAD's
beat rate is 90 bpm in (a) and 120 bpm in (b). These estimates are ob-
tained after adaptation. Due to nonlinear effects the observer consistently
underestimates the pressure peaks.
the beat rate 120 bpm is depicted in Fig. 13(a). Fig. 13(b) Table IV summarizes the deviations of the peak pres-
depicts the pressure estimate when the systemic resistance sure obtained when the aortic and atrial compliances were
is 0.00136 mmHg s/mm3 and the beat rate 60 bpm. Fig. changed from 1000 to 1800, and 4000 to 5500
13 shows that the pressure is estimated well and the max- mm3/mmHg, respectively. These pressure values were
imum deviation in the peak pressure is 4 mmHg which obtained when the EVAD was operating in 60, 90, and
represents an error of less than 5 .O% . 120 bpm. Table 111 shows that the maximum deviation in
TASCH et d.:PENN STATE ELECTRIC VENTRICULAR ASSIST DEVICE 381
m 0
lno T Ppeak = 164
2E
Y
E
E
160
I40 f 'peak
;
p.
E
g
Y
E l.oi
1 6 0 t
Ppeak = 185
0
c)I
.
.I
W
2
(c)
Fig. IO. During ejection the flexible blood sac is reshaped and blood is 0.0 0.1 0.2 0.3 0.4 0 5
displaced radially as shown in (bj and (c). This results in a variable
effective ejection area as depicted in Fig. 1 1 . Time [sec]
(bj
Fig. 12. The measured and estimated aortic pressures when the EVAD's
<
N beat rate is 90 bpm in (a) and 120 bpm in (b). These estimates are ob-
I 5200 tained after adaptation having the nonlinear hemodynamic effects incor-
5 porated.
5000
U
W
K
U 4800 function is defined such that the resultant values of the
z optimal gains are independent of the (often unknown) sys-
4600
0
F
tem state initial conditions. An adaptation scheme that
0
W 4400 improves the observer pressure estimates is introduced. In
3
w this scheme, the observer initial pressure values of the
4200
10 20 30 succeeding systolic cycle is adjusted, when the pressure
POSITION (MM) estimate does not match its corresponding second esti-
Fig. 1 1 . An experimentally determined function of the effective ejection mated value. This adaptation scheme suits the pulsatile
area plotted as a function of the pusher plate position. nature of the EVAD and its real-time implementation is
straightforward.
the peak pressure is 6 mmHg which represents an error of The aortic pressure observer was tested in vitro and after
4.0%. In general, Fig. 13 and Table I11 indicate that the introducing the nonlinear hemodynamic effects of the
developed aortic pressure observer performs well when blood sac, the aortic pressure estimates were very satis-
the values of the system parameters are changed substan- factory. The robustness of the observer to system param-
tially within their biological range. eter variations was studied. When the values of the sys-
temic resistance, aortic compliance, and atrial compliance
CONCLUSION are within their biological range, the maximum deviation
A formulation of an aortic pressure observer for The of the estimated peak pressure is less than 5.0%. This
Penn State EVAD is presented and its gains are determin- indicates that the developed pressure observer is robust to
istically optimized. In this formulation, the objective system parameter variations.
382 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 31, NO. 4. APRIL 1990
(4 (b)
Fig. 13. The measured and estimated aortic pressures. In (a) the systemic
resistance is reduced to 0.0007 mmHg s/mm3 and the beat rate is 120
bpm, and in (b) Rrys is increased to 0.00136 mmHg s/mm' and the beat
rate is 60 bpm. In both cases the pressure estimates are obtained after
adaptation having the nonlinear hemodynamic effects incorporated.
TABLE IV
SUMMARY
OF THE ROBUSTNESS
OF THE AORTICPRESSUREESTIMATES
WHEN
VARYING THE AORTICAND ATRIALCOMPLIANCES
Deviation in / Measured
Peak Pressure / Peak Pressure
Pa,-Pao [ m d g l 1 Pm [ m d g l
Aortic compliance
[mm3/mmHg]
1000 1/ 125 0/116 0/214
1330 2/119 3/163 1/185
1800 2/111 4/156 3/165
Atrial compliance
[ mm3/mmHg]
4000 1 / 108 6/150 0/181
4400 2/119 3/162 1/185
5500 1/120 5/151 2/191
@3.3(N = 4) = 1.03E-3 K:Ki - 3.93E-3 K I K : electric assist device through the design of an optimal output control-
ler,” Trans. Amer. Soc. Artif. Intern. Organs, vol. IO, no. 3 , pp. 714-
719, July-Sept. 1987.
- 8.21E-4 K:K2Ki D. G . Luenberger, “An introduction to observers,” IEEE Trans. Au-
+ 3.1OE-3 K,K2K: tornat. Conrr., vol. AC-16, pp. 596-602, Dec. 1971.
R. G. Jacquot, “Optimal deterministic state estimation-A first step,”
J . DynamicSyst., Measure., Contr., vol. 106, pp. 176-178, June 1984.
+ 2.98E-2 K;K: 2.28E-1 KiK:
- R. E. Kalman and R. S . Bucy, “New results in linear filtering and
prediction theory,” J . Basic Eng., vol. 8 3 , p p . 95-105, Mar. 1961.
+ i . i i ~ - i K:K: - ~ K , K : R. L. Carroll and D. P. Lindorff, “An adaptive observer for single-
inputisingle-output linear systems,’’ IEEE Trans. Automar. Contr., vol.
+ 5.15E-5 K:KiK3 AC-18, pp. 428-435, Oct. 1973.
G . Kreisselmeier, “Adaptive observers with exponential rate of con-
vergence,” IEEE Trans. Automat. Contr., vol. AC-22, pp. 2-8, Feb.
- 1.21E-4 K,KiK3 1917.
- 2.208-3 K:K2K3
+ i . 4 8 ~ - 2 K:K,K, Uri Tasch (M’86) was born in Tel-Aviv, Israel
He received the B S degree from the Israel Insti-
- 3.816-2 K,K2K3 + 4.12E-4 K;K2 tute of Technology, Tel-Aviv, in 1976, the M S
degree from the Illinois Institute of Technology,
- 2.43E-3 K:K3 + 1.99K:K3 Chicago, in 1978, and the Ph D degree from the
Massachusetts Institute of Technology, Cam-
- 9.86K,K3 - 4.11E-5 KiK; bridge, in 1983
He served on the faculty of the Pennsylvania
+ 1.21E-4 K,K; + 1.49E-3 K:K$ State University, University Park, from 1984-
1989, and is currently with the University of
- 8.79E-3 K:Ki + 2.2lE-2 K:K:
Maryland, Baltimore
Dr. Tasch I$ a member of ASME, ASEE, and the American Society of
Artificial Internal Organs
- 3.04E-2 K , K i - 3.29E-4 K:K2
+ 1.17E-3 K:K2 - 1.46E-2 K:K2
Jeffrey W. Koontz (S’89) was born in Bedford,
- 3.43E-2 K1K2 + 5.96E-3 K! PA, on July 23, 1964. He received the B.S de-
gree in electncal engineering in 1987 and the M S
- 4.11E-2 K ; + 1.6lE-1 Kt degree in bioengineering in 1989 from the Penn-
sylvania State University, University Park, PA
- 3.59E-1 K , - 3.46E-1 K: In the courw of graduate studies his research
interests included control systems, adaptive ob-
+ 4.50E-1 K: - 1.42E-5 K2Ki servers, and left ventricular assist devices Cur-
rently, he is employed as a ClinicaliBiomedical
+ 6.1OE-6 K:K: - 4.19E-3 K2K: Engineer in the Fairfax Hospital System in Falls
Church. VA
- 1.09E-1 K:K3 + 5.71E-5 K$K3
+ 1.33E-2 K2K3 + 2.83E-7 K i Michael A. lgnatoski (S’89) was born October 2 ,
+ 1.14E-4 K ; + 1.92E-2 K i
1965 in Reed City, MI. He received the B S de-
gree in electrical engineering from the Univerwy
of Detroit, Detroit. MI, in August, 1988
+ 2.5lE-1 K2 + 8.88E-6 K i He is currently pursuing the M S degree at the
Pennsylvania State University, University Park,
+ 5.39E-3 K i + 1.12K: + 19.44K3 in the Bioengineering program His research there
involve\ developing and implementing an aortic
pressure ob\erver for the Penn State Electrical
REFERENCES Ventricular A\\ist Device
[ I ] U. Tasch, D. B. Geselowitz, A. Sinha, and H . K. Hsu, “A novel
output feedback pusher plate controller for the Penn State electric ven-
tricular assist device,” J . Dynamic Syst., Measure., Conrr., Mar. 1989.
121 U. Tasch, D. B. Geselowitz, A. Sinha, J . Tirinato, H. K. Hsu. G . David B. Geselowitz (S’51-A’54-M’61-SM’62-F’78) for a photograph
Rosenberg, and W . S . Pierce, “Minimum power consumption of the p. 145.
and biography, see the February issue of this TRANSACTIONS,