Procwdlngsot the 2Sth Conference

on Dwldon and Control

Honolulu. Hawall Decemberlee0 TA-11-1 = 8:40
Practical Repetitive Control
System Design

Tadashi Inoue

Department of Electrical Engineering

National Defense Academy
Hashirimizu 1-10-20, Yokosuka, Japan 239

ABSTRACT

Considering two basic problems with repetitive con-

t r o l -tight stability condition and poor noise charac-
teristics, I present three techniques f o r a practical
repetitive controller design: stabilizing compensation
by an inverse phase system, band limitation of the con-
t r o l , and smoothing of control signal over periods of
repetitive operation.
1. INTRODUCTION

A repetitive control is not an all-purpose tool. I t

improves accuracy of the transient and steady-state re-
sponses of a control system when reference input signals
and disturbances are periodic, consisting of the harmon-
ic components of a common fundamental frequency. It is I -1
somewhat troublesome t o use because of tight stability
condition, which may cause difficulty i n planning a sta-
bilizing compensation, and because of t h e poor response Fig. 1. Equivalent configurations of
t o non-periodic or non-harmonic disturbances. basic repetitive control system.

Such problems must be w e l l understand and properly

handled i n a repetitive control system design. The aim of I
d
the paper is t o present points f o r a successful design,
and t o report t h e results of the experimental studies
through which the practical way of repetitive control
has been established.
2. PROBLEMS WITH REPE3ITJ.VE CONTROL
I I
II I

I
’

T \

A linear SISO repetitive control system w i l l be c v \I

discussed, so t h a t the basic problems and t h e i r solutions repetitive stabilizing basic

can be demonstrated in a simple manner. The repetitive compensator compensator servo
control system design method ever presented is based

-
upon one of the three basically equivalent representa-
tions shown i n Fig.1 (a) (c). Here, P stands f o r a
controlled object, and D f o r a dead-time element. For a
Fig.2. repetitive control system
discussed in the paper.

high-accuracy control, t h e common fundamental fre-

quency among exogenous signals must be known, and the
dead time length L must be equal t o t h e period of the
r+d -4 G
hndamental component. A signal will be called harmonic
if it is periodic and has no frequency components other
than 2 71. k/L rad/s (k=O, & 1, &2, Else t h e signal is
a-).

non-harmonic. The reference input is always harmonic.

The configurations (b) and (c) allow a powerful way

of stabilization as w i l l be discussed in t h e next section. Fig.3. Equivalent system with simplified loop.
The s t r u c t u r e (c) can be generalized t o Fig.2. Here, C(z)
is the transfer function f o r a closed-loop servo, which because an exact dead-time element is required f o r an
substitutes f o r the inner loop around P i n Fig.l(c). T h i s accurate repetitive control, and a digital implementation
will be called the basic servo i n this paper. The control of the controller is preferred. In Fig.2, M is the number
loop in the basic servo may be time-continuous o r dis- of the stages of unit delay i n the repetitive compen-
crete. Any design procedure may be applied t o make its sator, and is given by
response accurate, quick and well-damped. There is no
reason t h a t a good basic servo is bad f o r a repetitive M = L/T, (1)
control. The transfer function C ( z ) stands f o r a stabi- where T denotes the sampling interval of t h e control.
lizing compensator. The signal d ( t ) represents the The system i n Fig.2 is equivalent t o t h e block diagram in
disturbance appearing a t the output of the basic servo. Fig. 3. Here, e,(t) has a physical meaning a s the control
e r r o r t h a t would appear i f t h e basic servo were sub-
A discrete-time repetitive control is considered, jected t o r(t) and d ( t ) without t h e repetitive control. If

CH2917-3/90/0000-1673$1.00 @ 1990 IEEE 1673

Authorized licensed use limited to: Columbia University Libraries. Downloaded on November 01,2020 at 03:01:54 UTC from IEEE Xplore. Restrictions apply.
 Fig.4. Example of stable Nyquist diagram.

the reference input and the disturbance a r e harmonic, Fig.5. Gain f o r non-harmonic disturbance
t h e steady-state response in e,(t) has period L and the plotted on complex plane f o r C G ( e x p ( j u ) ) .
difference 6 (t)=es(t)-e,(t-L) disappears. T h i s explains
the effect of the repetitive control. The loop i n t h e last To simplify t h e discussion, consider a disturbance with a
part of the diagram governs the stability. To apply constant power spectrum S," like that of a white noise.
Nyquist's stability condition t o this loop, consider the Then the power of the control e r r o r averaged over one
frequency response period of etp(-juM) is given by
-
G,(exp(ju)) = exp(-jul\l)-Il C G ( e x p ( j u ) ) l , (2)
which is given by substituting z i n the loop gain
G,(z) = z - ~ l{- C ( z ) . G ( z ) } (2)' For a quantitative discussion, assume that CG(exp(ju)) is
by e x p ( j u ) , where U is a normalized frequency. The constant over the averaging interval. T h i s makes a close
Nyquist's criterion then tells us that, i f the loop trans- approximation if M is large because the interval length
f e r function (2)' is stable, the amplitude of G , ( e x p ( j u ) ) becomes small. Then we have
must become less than unity each time its phase angle
crosses ( 2 k + l )x (k=O, + 1 , &2, -) w h i l e U is varied from - K K2(u)= A a 2 ( u ) - S a 2 ,
t o X . In the case t h a t the dead time L (and conse- where
quently, M) is very large, the phase varies very rapidly,
crossing ( 2 k + l )x very many times. T h i s makes a sufficient
condition on t h e amplitude,
The contour plot of AdZ(u)on the complex plane for
I 1 - CG(exp(ju)) I < 1, ( -E < U I X) (3) C G ( e r p ( j u ) ) is shown i n Fig.5. The power is greater than
nearly necessary for stability. In many applications unity i n the stable region, which means t h a t the control
including t h e examples cited in this paper, t h e dead time e r r o r is increased by t h e repetitive control if distur-
is expected t o be much larger than the dynamical lags bance is non-harmonic. If C G ( e x p ( j u ) ) = l , the average
and delays in the basic servo. Thus, a stabilizing com- power of t h e control e r r o r is doubled. If such increase
pensation based upon (3) is meaningful. is not tolerable, the amplitude ICG(exp(j u ) ) I must be
decreased. A t an fixed amplitude, the average e r r o r is
The condition requires t h a t the Nyquist plot f o r minimized by making the phase shift zero.
CG(exp(ju)) be contained i n a circle of radius 1 centered
a t l + j O , as is shown i n Fig.4. Most basic servo systems 3. STABIUZATION WITH AN INVERSE PHASE SYSTEM
alone do not satisfy this. Thus, one has t o stabilize the The results obtained i n the preceding section sug-
repetitive system with C(z). One may think of using an gest t h a t the phase shift i n G ( e z p ( j u ) ) should be
inverse system f o r G(z) as C(z) t o let the l e f t hand side cancelled out by C(exp(j u ) ) . The amplitude ICG(exp(j u ) )1
of (3) vanish. Then a deadbeat repetitive control takes can be used t o find a compromise between the response t o
place. It is easily seen t h a t the tracking e r r o r t o r(t) harmonic and non-harmonic components. If the amplitude
vanishes a f t e r t h e first cycle of r(t)has been applied t o is identical t o unity, harmonic components in control
t h e system. However, t h e discrete-time transfer function e r r o r decay very quickly. If the amplitude is small, the
G ( z ) may have unstable zeros. The differentiating nature amplification of non-harmonic disturbance is insignifi-
of a inverse system may amplify high-frequency distur- cant. The importance of t h e two factors may differ from
bances. Hence, the stabilizing compensator should be one frequency range t o another. The signal d(t) stands
sought from a wider class of components f o r a disturbance appearing a t the output of a
closed-loop servo. Thus its spectrum, which may have
Along with the tight stability condition, t h e re- non-harmonic components, would concentrate a t high fre-
sponse t o non-harmonic disturbances is another problem quencies whereas the reference input, which is always
t o consider i n designing a repetitive control system. The harmonic, would have its principal components a t low
frequency response of the control e r r o r e(t) t o t h e frequencies. Then, low-pass characteristics would be
dibturbance d( t) is given by agreeable. Low-pass gain characteristics together with a
zero phase shift may seem t o be unreasonable.
1-exp(- juM)
Gd(exp(Ju)) = l+{CG(erp(j u ) ) - 1} e x p (-juM) ' (4)
The s t r u c t u r e of the control system that C ( z ) is
As we expect, l G d l becomes zero a t a harmonic frequency, placed i n series with the dead-time element makes this
u=2k z /M. The response t o a non-harmonic signal differ possible. The configuration of a stabilizing compensator
from frequency t o frequency. In (4), exp(-juM) is a peri- is shown i n Fig.5. Here, t h e output signal is synthesized
odic function of U. The period length is given by 2x/M. from v ( t ) and its future values available i n the dead-time

1674

Authorized licensed use limited to: Columbia University Libraries. Downloaded on November 01,2020 at 03:01:54 UTC from IEEE Xplore. Restrictions apply.
 element. The transfer function from v ( t ) t o u , ( t ) is then IF(exp(ju))l < 1, t h e radius is larger than 1 and allows
the Nyquist plot f o r CG t o encircle the origin. The
effect of t h e repetitive control is weakened because t h e
final value f o r the harmonic components in t h e control
This brings about leading phase characteristics in e r r o r do not become zero. Usually, uncertainty in t h e
C(ezp(ju)), which can be used t o cancel out t h e phase lag basic servo appears i n its high-frequency character-
in C(ezp(ju)) so t h a t condition (3) is satisfied with the istics. To avoid unnecessary loss of accuracy, one should
widest possible margin, and the effect of non-harmonic use a low-pass filter with dc gain 1 and a suitable
disturbances is minimized. bandwidth. I have recommended t o use a linear phase
filter[3]. Hara e t al. have pointed out t h a t the repet-
The weighting sequence {wk} can be determined in itive compensator has t o be band-limited in a
the manner that C(ezp(ju)) approximate G,(u)/G(exp(ju)), continuous-time control system so t h a t t h e system be
where a real function G,(u) represents a desired fre- stable[l], and discussed a case in which a 1st-order lag
quency response f o r C(z)-G(z),
in the sense t h a t filter was used i n t h e place of F(z).

l!C(ezp(ju)) - G,(u)/G(ezp(ju)) I Z - d u - Min.

The effect of the repetitive control is weakened by
t h e f i l t e r F not only in t h e attenuation band but in the
pass band also, mainly because of t h e lagging phase in
The solution becomes t h e f i l t e r . The harmonic signals have their frequency
components at
t)k = i l E { e Z p ( jku)-G,(-u)/G(ezp(-ju))}-du. (10) Uk = 2ak/M (k=O,k1,+2,-- ) (12)
A t these harmonic frequencies, the frequency response
If the desired frequency response G,(u) were t o be for z - ~ is equal t o l + j O , making the closed-loop gain of
specified theoretically, one would have t o know the t h e repetitive compensator infinitely high when no f i l t e r
power spectra f o r both r ( t ) and d ( t ) . As such knowledge exists(Fig.6(a)). Instead of z - ~ ,the band-limited compen-
is not available in general, one usually has t o make a s a t o r i n Fig.6(b) has zM.F(z). Its frequency response at
trial and e r r o r approach. The gain function f o r t h e basic uk is given by F(ezp(ju,)) instead of l + j O .
servo, lG(ezp(ju))l, seems t o be a good starting point f o r
G,(u) since it makes harmonic e r r o r decay fairly quickly, This means t h a t no higher accuracy in harmonic
and the natural low-pass characteristics i n CG make a components can be expected with the band-limited
precaution against high-frequency disturbance. repetitive compensator than with a simpler servo
compensator shown in Fig.6(c). The stability condition (3)
The number of t h e terms N in the summation (8) must applies also t o this servo compensator because this is a
be increased until G,(u) is well approximated. The deter- special case of a repetitive compensator that L=O.
mination is easy because each weight is calculated However, one would be able t o design t h e control system
independently of other terms. The compensator C(z) much easier on t h e basis of the necessary and sufficient
designed in this way is called the 'inverse phase system' condition. The controller would be so good a t
f o r G(z) in this paper. non-harmonic frequencies as at harmonic frequencies. For
this reason, both of the following conditions must be
The repetitive control has been first applied t o a satisfied so t h a t t h e band-limited repetitive control
power supply f o r a 12 GeV proton synchrotron in pulsed still be meaningful from t h e practical point of view:
operation[l]. In t h i s application, t h e basic servo was a
2nd-order system, and an inverse phase system was i ) the f i l t e r F(z) can not be implemented independent of
designed with G,(u)=IG(ezp(ju))I . A six-term summation was t h e dead-time element and be used in a servo
used, and phase e r r o r in CG(exp(ju)) was within 11 compensator shown i n Fig.6(b);
degrees. ii) some m e r i t of a repetitive control is expected.

The inverse phase system is applicable t o a wide Following is such an example.

class of basic servo with any order, any relative degree,
and with o r without dead time within itself. The diffi-
culty in building the inverse system is evaded even i f
the basic servo has a non-minimum phase. The dead-time (a)
element in the repetitive compensator places a quanti-
tative limit: N cannot exceed M.

The idea has been theoretically generalized t o a

multivariable case[ 21. (b) @
*-
4. BAND LIWITATION OF REPETITIVE CONTROL
If the basic servo has an unmodeled dynamics, it may
become difficult t o guarantee t h a t condition (3) is
satisfied. In such a case, t h e effect of the repetitive
control must be 'weakened'. If an element F(z) is placed
in the repetitive compensator a s shown in Fig.6(b),
condition (3) is replaced by Fig.6. Original (a) and band-limited (b)
repetitive compensators
Il/F(ezp(ju))l-ll - CG(ezp(ju))I < 1. ( -.n < U 6 a ) (11) and simpler servo compensator(c).
The radius of the circular limit becomes ll/F(ezp(ju))l. If

1675

Authorized licensed use limited to: Columbia University Libraries. Downloaded on November 01,2020 at 03:01:54 UTC from IEEE Xplore. Restrictions apply.
 Consider a FIR f i l t e r of order 2P+1 with linear phase
shift characteristics. The phase characteristics are
equal t o that of a P-stage delay. This f i l t e r and a M-P

-I-
stages of unit delay in series connection produce a
transfer function ~ - ~ e F ( z )where
,
P

F(z) = a, + C ak.( zk + z - ~
) (13) ?. 5 10
k=l

The frequency response f o r F(z) becomes a real

function
of U [degree]
F ( e x p ( j u ) )= a, + 2 ~ a k c o s ( k u ) . (14)
90
Without any phase shift, t h e frequency response i n t h e
pass band can approximate l + j O t o the desired degree if
t h e order P is sufficiently large. The only Limit on P is
that P+N cannot exceed M. (Note that one should apply
the stabilizing compensation t o the known part of the
dynamics i n order t o “ i z e the loss of accuracy due t o Fig.8. Frequency response of
the band Limitation.) The repetitive compensator thus inverse phase compensator
band-limited satisfies both conditions since t h e f i l t e r
function F(z) can not exist separately from rMand a
high accuracy characteristic of a repetitive control can
be expected in the pass band. A s an example t o demon-
s t r a t e the design procedure and the effectiveness of the
repetitive control, the result of an early experimental
study w i l l be presented below[3].

A band-limited repetitive control was applied t o a

simple dc servomechanism. The basic servo was equipped
with a tacho and a potentiometer feedback. The step
50 100 w 300

response in the control e r r o r voltage was measured a s Fig.9. Gain characteristics of

shown in Fig.7. From this response, two parameters in the f i l t e r f o r band limitation.
2nd-order model f o r t h e servo,
w ,=50 [rad/s] <=1.1.
were read off a s a rough approximation. A slowly damped
higher-frequency oscillation was present in the
response. T h i s was attributable t o its mechanical load,
which consisted of a inertia fixed on a elastic shaft. The
oscillation was ignored in t h e modeling.

The repetitive controller was implemented on a

16-bit micro-processor system. The sampling interval T A A

was 10 milliseconds. The frequency response of t h e

stabilizing compensator is shown in Fig.8. The number of
the summation terms necessary f o r this result was 11. A s
the desired gain, G,(u)=IGI was specified again. Thus, the
variation in the gain characteristics f o r C(z)was small,
and t h e phase shift was leading, cancelling out the lag
in the 2nd-order system. An 11-th order rippleless linear
phase FIR f i l t e r [5] was used f o r band limitation. The
gain Characteristics f o r t h e f i l t e r is plotted in Fig. 9.
To find the best bandwidth, t e s t s had t o be repeated.
Because the band limitation is required when uncertainty
exists, it is difficult t o find analytically the safe and
most efficient bandwidth.
The repetitive system was subjected t o a reference
voltage input as shown i n Fig. lO(a). The period length of

Fig.7. Step response of control Fig.10. Experimental reference input (a)

e r r o r in basic servo. and control e r r o r (b)-(d) waveform.

1676

Authorized licensed use limited to: Columbia University Libraries. Downloaded on November 01,2020 at 03:01:54 UTC from IEEE Xplore. Restrictions apply.
 t h e signal was 10.24 seconds. The control e r r o r waveform Because non-harmonic response can be improved by
in the first (b), the second (c) and the 3rd cycle (d) were t h e smoothing, C(z)G(z) is allowed t o be identical t o
rapidly decaying. For each period, the rms value f o r the unity. Then the response of the repetitive control
control e r r o r voltage was evaluated. The value rested a t system t o a harmonic signal settles (at a constant vector
0.8mV rms. A s the filter F was removed, a high-frequency value) a f t e r J periods from the s t a r t of the operation.
oscillation was excited i n the servomechanism.
If one considers the case t h a t the disturbance is a
5. SMOOTHING OVER PERIODS ripple consisting of a single frequency component within
If the frequency of non-harmonic component i n dis- the interval of u,[O,x), and assumes t h a t the ripple
turbance is unknown, is varying, o r i f such components frequency U, and power pn a r e probabilistic variables,
appear a t wide range of frequencies, the effect of a whose joint distribution function is given by f(un,pn),
repetitive control must be moderated by slowing down its then the expected value f o r the power of t h e control
response. Most simply, it is done by reducing the e r r o r due t o the ripple is given by
amplitude of CC(exp(ju)) a t a l l frequencies. A smoothing
element may be put into t h e repetitive compensator,
instead. Then, the smoothed response can be optimized i n
some sense. A smoothing function can be designed i n the
following way. where C,(exp(ju)) is given by substituting z by ezp(ju) in
Consider a series of a discrete-time variable, xl,xz. the transfer function relating control e r r o r t o the
...,
xk, and organize the samples into M-tuples,
-e, disturbance,
( /{ ~ + z - ~ Bz )( (C(z ) *G(z) -1) }.
G, (Z)=( ~ - z - ~ Bz)) (21)
NO)= [XI, xz, , xw1=,
z(1) [XM-rlv XM+Z? 5ZM1T, It may be f u r t h e r assumed t h a t U, and p , a r e mutually
independent and t h a t un is uniformly distributed on t h e
interval [0, x). Then (20) reduces t o
E(P=) I.E(Pn), (22)
where
Then, the vector z(i) is constant a s i varies if the time
sequence is harmonic. If the vector varies with i, there
a r e non-harmonic components i n the variable. The
variation can be reduced if one uses a weighted sum of t h e 0
vectors
- J-1 If CG(exp(ju))=l a t a l l U, I is simplified t o
Z(i) = bk,,.Z(i-k), (16) J
k=O I = 1 + Cbk2.
k=l
where
It is easily seen here t h a t an even weighting (bk 1/J)
minimizes I (and,thus.E(p,)) under t h e restriction (17).

in place of z ( i ) . The consideration can be applied t o u(t), A high-gain PID feedback around a standard
the output of the repetitive compensator. voltage-output triangular c a r r i e r PWM inverter may
bring about an oscillation a t a frequency much higher
If we define than the c a r r i e r frequency. Such oscillation is trouble-
some because it increases switching loss i n t h e inverter.
If we consider a transistor-output inverter whose
c a r r i e r lies, say, a t 3kHz. it is difficult t o make t h e
we have, as the the frequency response f o r the smoothed feedback bandwidth exceed 1kHz. Because of t h e integral
repetitive compensator i n Fig. 11, term, the loop gain increases at low frequencies. But a
loop gain higher than 20dB, f o r example, can be expected
exp(- jMu) .B(ezp( ju))/{ 1-exp(-jMu) .B(exp(ju))} (19) only a t frequencies lower than 100Hz.
If (17) holds, the amplitude of the frequency response is To remove such a limit, I have applied t h e repetitive
infinitely large a t uk given by (12). and the stability control t o a PWM inverter. A smoothed version w a s used
condition (3) applies t o the control system. Thus, the because there could be non-harmonic disturbances such
smoothed repetitive compensator is a direct generali- a s interference with the ripples in the dc power source
zation of the compensator i n Fig.6(a), not of the t o t h e inverter. The c a r r i e r frequency was 4kHz and the
compensator (b). sampling clock i n t h e controller w a s a t 1kHz. To avoid
aliasing of the c a r r i e r frequency component in t h e
output pulse into the control, feedback detector was
equipped with a demodulator and a LPF. Because the
voltage-output inverter had little smoothing transfer
characteristics, the response t o t h e discrete-time
control could be discontinuous. To prevent this, the
controller output U(t) was linearly interpolated.

The s t e p response of the repetitive loop including

t h e demodulator and t h e interpolator was measured a s
Fig. 11. Repetitive compensator with shown i n Fig.12. It was approximated by a 3 units of delay
smoothing function over periods. so t h a t t h e inverse phase system was simplified. The

1677

Authorized licensed use limited to: Columbia University Libraries. Downloaded on November 01,2020 at 03:01:54 UTC from IEEE Xplore. Restrictions apply.
 experiment was made with J=4, bk=0.25(const.). The power
spectrum f o r the output pulse voltage was measured a s
the response t o a sinusoidal reference input a t 31.244Hz
The spectrum shown in Fig.l3(a) was improved t o (b) by
the repetitive control. Next. a sinusoidal signal of
effective value 5.0V was added t o the output voltage.
The amplitude of the control e r r o r was measured w h i l e c,
the frequency was varied from the 2nd t o the 3rd a
c,
harmonic frequency. The result is shown i n Fig.14. With
0
the original (not smoothed) version of the repetitive
control, the harmonic e r r o r vanished but, a t the the
middle o f the harmonic frequencies, the e r r o r is doubled. Fig.12. Step response f o r
With a smoothed version, the e r r o r voltage was leveled loop transfer characteristics.
a t non-harmonic frequencies
I t has been pointed out i n the above discussion t h a t
a harmonic variation i n a time series makes the corre- U
sponding vector constant-valued. A harmonic variation i n
t h e sampling interval T also keeps t h e vectors of -2 0
sampled variables constant, and keeps the description of
the repetitive control system time invariant. Further, -4 0
t h e interval of time a t which t h e control e r r o r is
sampled and t h a t of time a t which control action takes -6 0
place may vary indebendently of each other. T h i s freedom
makes it possible t o provide a repetitive control with [dBl
unique function s w h a s forcing a servo system, the 0
parameters of which a r e not exactly known, a sequence of
t r u l y finite-time c3ettli;ig point-to-point motions [7]. -2 0

REFERENCES -4 0

[ l ] T.Inoue et al., tfEigh accuracy Control of a Proton -6 0

Synchrotron Magnet Power Supply," i n Proceedings
of the 8th IFAC World Congress, 1981, pp. 3137-3142. 1 3 5 7 9 11 13 1 5
[2] T.Omata et al..tfApplication of Repetitive Control harmonic order
Method t o Mu1 tivariable Systems," Trans. SICE, 20,
pp.795-800, Sept. 1984 (in Japanese). Fig.13. Output power spectrum before (a)
and a f t e r (b) repetitive control.
[3] T.Inoue et al.,"Kigh Accuracy Control of Play-Back
Servo Systems," Trans. IEEJ, C101, pp. 89-96, Apr.
1981(in Japanese).
smoothed
[4] S.Hara et al.,"Repetitive Control System: A New
Type Servo System for Periodic Exogenous Signals,"
IEEE Trans. Autom. Contr., AC-33, pp. 659-667, July
1988.
[5] R.W.Hamming: Digital Filters. Englewood Cliffs NJ:
1977. ch. 7. uncontrolled

:t;] T.Inoue et al.," Improving PWM Inverter Output 0.0

Waveform by Use of Repetitive Control with Regard 6 0.0 70.0 80.0 90.0
t o Asynchronous Interferences," Trans. LEEJ, D107, frequency [ H z j
pp. 994-1001, Aug. 1987 (in Japanese).
Fig.14. Amplitude of harmonic and
[7] T.Inoue et al. ,"Improving Transient Response non-harmonic control error.
through Repeating a Sequence of Positioning
Motions," Trans. SICE, 22, pp.30-35, Jan. 1986 (in
Japanese).

1678

Authorized licensed use limited to: Columbia University Libraries. Downloaded on November 01,2020 at 03:01:54 UTC from IEEE Xplore. Restrictions apply.