Proc.

24th Chinese Control and Decision Conference

pp. 3528-3533, Taiyuan, China, (2012)
Feedforward Control for Disturbance Rejection:
Model Matching and Other Methods
Hua Zhong1 , Lucy Pao2 , Raymond de Callafon3
1. University of Colorado Boulder, Boulder, CO 80302, USA
E-mail: zhongh@colorado.edu
2. University of Colorado Boulder, Boulder, CO 80302, USA
E-mail: pao@colorado.edu
3. University of California at San Diego, La Jolla, CA 92093, USA
E-mail: callafon@ucsd.edu

Abstract: Feedforward control can improve disturbance rejection performance of a system when the measurement of
the disturbance is available. This paper discusses the design of feedforward control in the discrete-time domain using
the model matching methods that compute optimal and stable feedforward controllers. It is shown that the existence of
a non-zero solution to the model matching problem depends on the difference between the relative degrees of the plant
dynamics and the disturbance dynamics. A number of approximate dynamic inversion techniques commonly used for
feedforward control design are reviewed and compared with the model matching methods. These feedforward control
design methods are then applied to the application of an example tape head track-following servo system where the
feedforward controller aims at reducing the position error caused by the lateral tape motion. Simulation results are
presented to demonstrate the effectiveness of the feedforward control.
Key Words: Feedforward Control, Model Matching, Tape Storage

1 INTRODUCTION One typical design of the feedforward control is to ap-

proximate the inverse of G and compute a stable and causal
Feedforward control can improve disturbance rejection
approximation of −G−1 F . Dynamic inversion and approx-
performance of a system when disturbance measurements
imate dynamic inversion techniques have been widely stud-
are available. As show in Figure 1, the transfer function
ied to design the feedforward controller Cff . Commonly
Ted from the disturbance d to the output error e is given by
used methods include the zero phase error tracking control
(ZPETC) [11], the zero magnitude error tracking control
Ted = F + GCff . (1)
(ZMETC) [10] (and references therein), and the non-causal
where G and F represent the dynamics of the plant and the Taylor series approximation method [8].
disturbance, respectively. The feedforward controller Cff As an alternative to dynamic inversion, H2 and H∞
aims to minimize the impact from d on e, possibly over model matching methods can be used to find an optimal
a desired frequency range. Without loss of generality, the feedforward controller Cff that minimizes the norm of the
plant G is assumed to be stable as a feedback controller transfer function Ted [3] [4]. A generic set of tools to solve
could be added for stabilization otherwise. H2 or H∞ preview control problems has been summarized
in [5]. In this paper, we investigate the H2 model match-
ing method for disturbance rejection in the discrete-time
domain. The existence of the solution for such a prob-
lem depends on the difference between the relative degrees
of the plant dynamics G and the disturbance dynamics F .
We derive the requirements on plant dynamics and distur-
bance dynamics such that the model matching problem has
a non-zero solution. Computing optimal controller by min-
imizing a norm on the transfer function Ted may lead to
large control signal uff if no penalty is imposed on the
Figure 1: The feedforward controller Cff aims to attenuate norm of Cff . This can be solved by augmenting the model
the effects from the disturbance d on the output error e. matching techniques with a simultaneous minimization of
a (weighted form) of Cff to provide a trade-off between
This work was partly supported by the Information Storage Industry minimizing e and the control signal uff .
Consortium (INSIC) and Oracle Corporation. We are also grateful for
support from Professor Sinan Müftü of Northeastern University in Boston, In this paper, a few approximate dynamic inversion tech-
MA for providing us with tape simulation data. niques are reviewed and compared with the augmented

978-1-4577-2074-1/12/$26.00 c 2012 IEEE 3528

model matching method. These controller design methods 2.3 Non-zero Solutions
are then applied to the application of an example tape head When G(z) has zeros outside the unit circle, the in-
track-following servo system where the feedforward con- verse G−1 (z) is unstable and possibly non-causal. Ap-
troller is designed to reduce the position error caused by proximate dynamic inversion techniques can be used to
lateral tape motion. The remainder of the paper is orga- compute the approximate inversion G̃−1 (z) of G(z) and
nized as follows. Section 2 discusses the requirements on Cff (z) = −G̃−1 (z)F (z). However, whether the resulting
the plant dynamics G and the disturbance dynamics F for feedforward controller Cff (z) would reduce the norm of
the H2 model matching method to yield non-zero solutions. Ted (z) is not clear.
We then review the approximate dynamic inversion tech- Improving the disturbance rejection performance of a
niques in Section 3. To demonstrate the effectiveness of system by adding feedforward control is directly addressed
the feedforward control, we present the simulation results in the model matching problem formulation in (2), pro-
of implementing the feedforward control design on the ex- vided a non-zero solution Cff (z) exists. To characterize
ample tape head track-following servo system in Section 4. when such a non-zero solution can be computed, we write
Finally, Section 5 provides conclusions and discussion. the transfer function Ted (z) in (1) in a lower Linear Frac-
2 MODEL MATCHING METHODS tional Transformation (LFT) formulation
2.1 Problem Formulation Ted =Fl (P ,−Cff ) = P11−P12 Cff (I+P22 Cff)−1P21 , (4)
To minimize the variance of the output error e due to the
disturbance d (Figure 1) in the discrete-time domain, the where
model matching methods compute a feedforward controller !
P11 P12
" !
F G
"
Cff (z) such that either the H2 or H∞ norm on the transfer P = = .
P21 P22 −I 0
function Ted (z) is minimized [3], [4], [5]. In this paper, we
restrict ourselves to minimize the H2 norm of the transfer With the state-space model of the LFT plant P given by
function Ted (z), as described in the following equation  
A B1 B2
Cff (z)= arg min "[F (z)+G(z)Cff (z)]Dd (z)"2 . (2) P=ˆ  C1 D11 D12  (5)
Cff (z)∈RH∞
C2 D21 0
Here, the discrete-time filter Dd (z) is stable and stably in-
vertible. Moreover, |Dd (ejω )|2 = Φd (ω), where Φd (ω) is we have a properly specified discrete-time H2 model
the spectral factorization of the spectrum of the disturbance matching problem with the optimality property summa-
signal d. Without loss of generality, we set Dd (z) = I rized in the following lemma.
to simplify the formula and discussion. It should be noted Lemma 1. Let G be stable and exactly proper with a non-
that Cff (z) is required to be stable (∈ RH∞ ) and we allow T
zero feedthrough matrix DG (DG DG > 0). Let F be stable
Cff (z) to be proper. The discrete-time variable z is omitted
and (strictly) proper. Then
in some of the equations in this paper to save space.
2.2 Direct Solution for Minimum-Phase (MP) Systems min "F (z) + G(z)Cff (z)"2 < "F (z)"2
Cff (z)∈RH∞
Analytically, if G(z) only has minimum-phase (MP) ze-
ros and is exactly proper, G(z) is invertible and the feed- Proof. For the minimization of "Fl (P , −Cff )"2 over Cff
forward controller can be computed directly via using the LFT representation (4) of P given in (5), we have
the following four standard conditions to compute an opti-
Cff (z) = −G−1 (z)F (z). (3) mal solution [4]:
When G(z) only has MP zeros and is strictly proper, the
1. (A, B2 ) is stabilizable and (C2 , A) is detectable;
inverse G−1 (z) is stable but non-causal. Denote the rela- ! "
tive degrees of G(z) and F (z) as rG and rF , respectively, A − jωI B2
2. has full column rank for all ω;
we can write G(z) = G(z)z rG z −rG = Ḡ(z)z −rG , where C1 D12
Ḡ(z) = G(z)z rG is invertible. Following (3), the feedfor- ! "
ward controller is A − jωI B1
3. has full row rank for all ω;
C2 D21
Cff (z) = −Ḡ−1 (z)F (z)z rG
T T
and Cff (z) is proper iff F (z)z rG is proper, that is, the rel- 4. D12 D12 > 0 and D21 D21 > 0.
ative degree of F (z) is greater or equal to that of G(z).
The first condition is for the stabilization via output feed-
When rF ≥ rG , the disturbance can be completely re-
back. Since both G and F are stable, this condition is sat-
jected for a MP system G(z) as Ted (z) = 0. In the special
isfied. The second and third conditions are to guarantee
case of rG = 0, the direct solution of the Cff (z) is given
that a solution can be computed via two Riccati equations
in (3). When rF < rG , only the zero solution Cff (z) = 0
[4]. Finally, the last condition guarantees that the proposed
should be used, as a non-zero Cff (z) will worsen the vari-
H2 control problem is nonsingular. With the definition of
ance of the error signal e. This effect will be explained in T
more detail when we examine the conditions for non-zero P in (5), we see that D21 = −I and D21 D21 > 0. Fur-
T
solutions. thermore, D12 D12 > 0 is satisfied as G is exactly proper

2012 24th Chinese Control and Decision Conference (CCDC) 3529

 and has a non-zero feedthrough matrix DG . Hence, we a simultaneous minimization of the (weighted) norm on
have a properly specified discrete-time H2 model matching Cff (z), as described in (6),
problem that computes a non-zero solution of Cff . With ' '
Cff (z) (= 0, we must have "F + GCff "2 < "F "2 as ' F (z) + G(z)Cff (z) '
Cff (z) = arg min ' ' , (6)
"F + GCff "2 = "F "2 only if Cff (z) = 0. It should Cff (z)∈RH∞ ' Cff (z)W (z) '
2
be noted that D11 does not play a role in computing the
optimal solution of the model matching problem. where W (z) is a stable discrete-time filter that acts as a
weighting function for the feedforward control signal uff .
The result in Lemma 1 states that a well-posed model The augmentation provides a trade-off between minimizing
matching problem can yield an optimal feedforward con- the variance of the output error e and the variance of the
troller that reduces the H2 norm of Ted (z). The control signal uff . It is easily verified that the additional
well-posedness of the problem is directly related to the minimization of the variance of uff in (6) does not affect
feedthrough matrix DG of the exactly proper plant G. The the results summarized in Theorem 1.
result can be easily extended to a system G(z) with a rela-
3 APPROXIMATE DYNAMIC INVERSION
tive degree rG ≥ 0.
Dynamic inversion is a commonly-used technique to de-
Theorem 1. Let G and F be stable. Let G have a relative sign feedforward control for disturbance rejection. The
degree rG ≥ 0. If rF ≥ rG , then presence of NMP zeros in the dynamics complicates the
inversion as the NMP zeros become unstable poles in the
min "F (z) + G(z)Cff (z)"2 < "F (z)"2 .
Cff (z)∈RH∞ inverse system. Denote the numerator and denominator of
the plant transfer function G(z) as BG (z) and AG (z), re-
Proof. Define Ḡ(z) = G(z)z rG , Ḡ(z) is then ex- spectively. Denote the numerator as
actly proper with a non-zero feedthrough matrix DG and
T BG (z) = BsG (z)BuG (z),
DG DG > 0. Similarly, define F̄ (z) = F (z)z rF and
F̃ (z) = F̄ (z)z −rF +rG , the H2 norm of Ted (z) can be writ-
where BsG (z) is the polynomial that contains all the zeros
ten as
of G(z) within the unit circle and BuG (z) is the polynomial
"F + GCff "2 = "F̄ z −rF + Ḡz −rG Cff "2 of the NMP zeros, the plant dynamics then is
= "[F̄ z −rF +rG + ḠCff ]z −rG "2 BsG (z)BuG (z)
G(z) = .
= "F̃ + ḠCff "2 . AG (z)

According to Lemma 1, we can conclude Approximate inversion techniques aim to seek an approx-
imate inverse G̃−1 (z) to solve the feedforward controller
min "F̃ (z) + Ḡ(z)Cff (z)"2 < "F̃ (z)"2 and
−1
Cff (z)∈RH∞ AG (z)B̃uG (z)
G̃−1 (z) = ,
BsG (z)
provided the following two conditions hold. First, F̃ (z)
−1
must be (strictly) proper, e.g. rF ≥ rG . Second, Ḡ must where B̃uG (z) is the approximate inverse of BuG (z). The
T goal is to find the approximate inverse of the NMP zero
have a feedthrough matrix DG with DG DG > 0. −1
such that G(z)G̃−1 (z) = BuG (z)B̃uG (z) ≈ 1. We con-
Theorem 1 states that the feedforward controller solved sider the simplest case where G(z) contains only one NMP
by the model matching method can improve the variance zero at z = z0 (|z0 | > 1) and BuG (z) = z − z0 .
(H2 norm) of the error signal e, provided 1) the relative de-
3.1 Zero Phase Error Tracking Control
gree of F (z) is larger than or equal to that of G(z) and 2)
G(z)z rG is well-posed. Furthermore, the larger the differ- The Zero Phase Error Tracking Control (ZPETC) inver-
ence in the relative degrees of F (z) and G(z), the more we sion method is introduced in [11]. This method approxi-
can reduce "F + GCff "2 . This is due to the fact that when mates the inverse by reflecting the NMP zero in the origi-
the effective order of F̄ (z)z −(rF −rG ) increases, a larger nal plant about the unit circle so as to cancel the phase shift
order Cff (z) with more controller parameters to minimize due to the NMP zero, that is,
"F + GC"2 is allowed. Adding extra delays in F (z) can 1 − z0 z
−1
increase the difference between rF and rG . B̃uG (z) =
(1 − z0 )2
2.4 Augmented Model Matching
and
Computing optimal feedforward controller by minimiz- AG (z)(1 − z0 z)
ing a norm on the transfer function Ted (z) may lead to G̃−1 (z) = .
BsG (z)(1 − z0 )2
a large control signal uff , as no penalty is imposed on 1
the norm of Cff (z). In fact, the H2 model matching in Here, the gain (1−z0 )2 is added to maintain unity DC gain
−1
(2) becomes a minimum variance control problem without of G(z)G̃ (z). The resulting dynamic inverse is non-
penalty on the control signal. This can be easily addressed causal and an extra k = rG + 2 steps of delay needs to
by augmenting the model matching techniques in (2) with be added so as to form a causal feedforward controller.

3530 2012 24th Chinese Control and Decision Conference (CCDC)

In general, a total number of rG + 2ku steps of delay is the model matching problem has a non-zero solution), the
needed, where ku is the number of NMP zeros in the plant disturbance affects the feedforward path before the sys-
dynamics. The phase response of G(z)G̃−1 (z) is zero at tem output and the non-causal controller design is feasi-
all frequency while the gain deviates from unity at higher ble. In practice, a feasible feedforward controller needs to
frequency. be strictly proper and this leads to the requirements of a
higher relative degree in F (z).
3.2 Zero Magnitude Error Tracking Control
In contrast to the ZPETC where the NMP zero is con- 4 SIMULATIONS
verted to a stable zero in the inverse system, the Zero Mag- We employ the model matching method and the ap-
nitude Error Tracking Control (ZMETC) reflects the NMP proximate dynamic inversion techniques on the head track-
zero about the unit circle and coverts it to a stable pole in following servo system in a reel-to-reel tape system (Fig-
the inverse system [10]. Thus, the approximate inverse of ure 2). The schematic is modified from the graphical user
BuG (z) interface of a LTM simulation tool (LTMSim) [12].
−1 1
B̃uG (z) =
(z0 z − 1)
and
AG (z)
G̃−1 (z) = .
BsG (z)(z0 z − 1)
To maintain the causality of the feedforward controller,
an additional rG steps of delay is needed. The product
G(z)G̃−1 (z) is
z − z0
G(z)G̃−1 (z) = Figure 2: Schematic of tape head track-following system.
z0 z − 1
and the magnitude response is always unity but there will
be phase lag. Data is recorded in tracks parallel to the edges of the tape
and each data track spans the entire length of the tape. The
3.3 Non-causal Taylor Series Approximation
contemporary tape is half inches wide and there are typi-
The non-causal Taylor series uses a non-causal, stable, cally over a thousand tracks across the width of the tape.
Taylor series expansion to approximate the inverted unsta- The tape winds longitudinally between the two reels and
ble pole. Denote p = zz0 (|p| < 1), passes over the head, where data is read from or written to
the tape. A voice coil actuator moves the head assembly
−1 1 1 laterally across the width of the tape to position the head
BuG (z) = = − · (1 − p)−1 .
z − z0 z0 on the desired track.
When tape transports longitudinally, it can exhibit lateral
Using a nT th -order Taylor series to approximate (1−p)−1 ,
motion that misaligns the head and the desired track and in-
we have
creases position error. The typical peak-to-peak amplitude
1 ( of measured lateral tape motion displacement (LTMD) data
−1
· 1 + p + p 2 + · · · + pn T
)
B̃uG (z) ≈ −
z0 in an operating tape drive is about 10 µm [6] [7] [9]. Both
1 1 − p nT the head and the tape can have out-of-plane motions that
= − · are beyond the scope of this paper. One primary purpose
z0 1−p
1 − ( zz0 )nT of the head positioning servo system is to follow the de-
= sired data track as accurately as possible during read/write
z − z0 operations in spite of disturbances such as the lateral tape
and motion. Feedforward control can be applied to reduce the
z n
AG (z) 1 − ( z0 ) T effects of the LTM on the position error and improve the
G̃−1 (z) = . track-following performance of the head servo mechanism.
BsG (z) z − z0
The tape head track-following system can be modeled
The resulting approximate inverse of G(z) is a non-causal as a spring-mass-damping system with a couple of reso-
transfer function. A total number of rG + nT extra steps nances. A typical transfer function model from the voice
of delay need to be added to the feedforward controller to coil motor current to the head position is
make it causal. The magnitude of G(z)G̃−1 (z) is closer to
unity as nT increases and 1 − ( zz0 )nT gets closer to 1. Kω1 2 ω2 2
The feedforward controller designed using approxi- G(s) = ,
(s2 + 2ζ1 ω1 s + ω1 2 )(s2 + 2ζ2 ω2 s + ω2 2 )
mate dynamic inversion techniques are usually non-causal.
Hence, the knowledge of the disturbance ahead of time where ωi (i = 1, 2) are the two resonances and K is a gain.
is required to achieve good disturbance rejection perfor- In this study, ω1 is chosen to be 100 Hz and ω2 is set to
mance. When the relative degree of the disturbance dy- 1 KHz. The damping ratios are ζ1 = 0.0796 and ζ2 = 0.05
namics is higher than that of the plant dynamics (and so and the gain K = 2000. Using a zero-order hold (ZOH)

2012 24th Chinese Control and Decision Conference (CCDC) 3531

 on all inputs as in [2], with a sampling rate of 10 KHz, the
discrete-time transfer function is

0.12633(z + 9.52)(z + 0.9853)(z + 0.102)

G(z)=
(z 2 − 1.986z + 0.99)(z 2 − 1.569z + 0.9391)

and the Bode plot of the discrete-time model is depicted

in Figure 3. The system is strictly proper with a relative
degree rG = 1. The NMP zero at −9.52 is due to the fast
sampling rate [1].

Figure 4: The feedforward controller designed by the H2

norm model matching method reduces the position error
when the difference between the relative degrees of F (z)
and G(z) is 2.

Figure 3: Bode plot of the tape head track-following model

at a sampling frequency of 10 KHz.

In the simulation, the disturbance d is the lateral tape mo- Figure 5: When the difference between the relative degrees
tion displacements near the head that are measured from an of F (z) and G(z) is 2, the position error is reduced to
industry tape drive. Spectrum analysis shows the ampli- 10−1 µm.
tude of d is on the order of 10 µm and the frequencies of
the components are within 500 Hz. In the presented simula-
tion results, the disturbance dynamics F (z) is chosen to be weighting function W (z) to add more penalty to the con-
a pure delay to better show the improvement in the distur- trol input can further constrain the size of the input at a cost
bance rejection performance as the difference between the of degrading the disturbance rejection performance. If the
relative degrees of F (z) and G(z) increases. When F (z) feedforward control input is allowed to be large, the model
includes more dynamics, the same trend still holds but the matching methods can achieve even better disturbance re-
overall disturbance rejection performance is not as good as jection performance.
when F (z) is a pure delay.
4.2 Approximate Dynamic Inversion Techniques
4.1 Model Matching Methods The authors in [13] developed a number of feedforward
The Matlab robust control toolbox is used to solve for controllers based on approximate dynamic inversion tech-
the optimal Cff (z) that simultaneously minimizes the H2 niques to take into account the lateral tape motion (LTM)
norm of Ted (z) and the control input, as described in (6). disturbances in the tape head positioning system. The ap-
A high-pass weighting filter W (z) is used to limit the size proximate dynamic inversion techniques design Cff (z) in-
of feedforward control input. dependently on the disturbance dynamics F (z). These con-
When F (z) is a unity delay (rF = rG = 1), the solution trollers are non-causal and require as the input the knowl-
of Cff (z) is simply zero. When rF − rG = 2, Figure 4 edge of the disturbance in the future. In the simulations, de-
and 5 show the output error caused by the disturbance is lays are added in F (z) as an alternative to obtaining knowl-
reduced by 1 order of magnitude to the order of 10−1 µm. edge of the disturbance ahead of time.
Increasing the steps of delay in F (z) yields higher oder The disturbance rejection performance of the ZMETC
Cff (z) and better disturbance rejection performance. Fig- and the ZPETC feedforward controller are similar for this
ure 6 shows the output error is further reduced by 1 more application. We present the results of the ZMETC feedfor-
order of magnitude to 10−2 µm when rF − rG = 20. ward controller as it requires less additional delay than the
The maximum feedforward control input current is ZPETC feedforward controller. When F (z) = z −1 , the
around 7 A and the RMS is about 0.9 A. Adjusting the disturbance is reduced by 1 order of magnitude as shown

3532 2012 24th Chinese Control and Decision Conference (CCDC)

Figure 6: The error is further reduced to 10−2 µm when Figure 8: The position error is at 10−3 µm when the feed-
F (z) is a 20-step delay. forward controller is designed using the Taylor series ap-
proximation method.

in Figure 7. Figure 8 shows the simulated output error is

reduced by 3 orders of magnitude when Cff is designed is acceptable. In the special case of minimum-phase sys-
using a third-order Taylor series approximation. tems, the model matching methods and the dynamic in-
version techniques yield the same feedforward control de-
sign. These indicate some equivalence existing in these
two different feedforward control design methods. We ap-
ply these methods to an example tape head track-following
servo system in which the feedforward control is designed
to reject the position error caused by the lateral tape motion
disturbances. Simulation results demonstrate that the feed-
forward control effectively improves disturbance rejection
performance of the servo system.

REFERENCES
[1] K. Åström, P. Hagander, and J. Sternby. Zeros of Sampled-
data Systems. Automatica, vol. 20, no. 1, pp. 31-38, 1984.
[2] G. Franklin, J. Powell, and M. Workman. Digital Control of
Figure 7: ZEMTC feedforward controller reduces the posi-
Dynamic Systems, Prentice Hall, 3rd edition, 1998.
tion error to 10−1 µm. [3] J. Doyle, B. Francis, and A. Tannenbaum. Feedback Control
Theory, Macmillan, 1992.
[4] M. Green and D. Limebeer. Linear Robust Control, Prentice
In both simulations, the maximum value of the feedfor- Hall, 1994.
ward control input current is around 15 A and the RMS is [5] A. Hazell. Discrete-time Optimal Preview Control. Ph.D.
about 3.5 A. The disturbance rejection performance will be Thesis, Univ. of London, 2008.
degraded if the feedforward input is saturated. [6] Information Storage Industry Consortium (INSIC). Interna-
tional Magnetic Tape Storage Roadmap 2008. INSIC, San
5 CONCLUSIONS Diego, September 2008.
[7] V. Kartik. In-plane and Transverse Vibration of Axially-
This paper discusses model matching methods and ap- moving Media with Advanced Guiding and Actuation Ele-
proximate dynamic inversion techniques to design feed- ments, Ph.D. Thesis, Carnegie Mellon University, 2006.
forward controllers for disturbance rejection. The model [8] A. Oppenheim, A. Willsky, and S. Nawab. Signals and Sys-
matching methods convert the disturbance rejection prob- tems, Prentice Hall, 2nd edition, 1996.
lem to an optimization problem in which the feedforward [9] B. Raeymaekers and F. Talke. Measurement and Sources
controller minimizes the norm of the transfer function from of Lateral Tape Motion: A Review. ASME J. Tribology-
the disturbance to the output. Such an optimal controller Transactions, vol. 131, January 2009, pp. 65-68.
can improve disturbance rejection only when the relative [10] B. Rigney. Adaptive Settle-Optimal Control of Servomech-
degree of the disturbance dynamics is higher than that of anisms. Ph.D. Thesis, Univ. of Colorado Boulder, 2008.
[11] M. Tomizuka. Zero Phase Error Tracking Algorithm for
the plant dynamics. The approximate dynamic inversion
Digital Control. ASME J. Dynamics, System, Measurement,
techniques seek an approximate inverse of the plant to con- and Control, vol. 109, March 1987, pp. 1-6.
struct the feedforward controller. The resulting controllers [12] J. Wickert and M. Brake. Tape Path Guiding Simulation Dis-
are usually non-causal. If the disturbance dynamics have tribution. INSIC Report, October 2007.
higher relative degree than the plant dynamics, the dis- [13] H. Zhong and L. Pao. Combined Feedforward/Feedback
turbance affects the feedforward path before it appears in Control for Tape Head Track-Following Servo System. Pro-
the system output and a non-causal feedforward controller ceedings of IFAC World Congress 2011, pp. 4040-4045.

2012 24th Chinese Control and Decision Conference (CCDC) 3533