You are on page 1of 6

Minimal Modeling of Rocket Roll Response and

Aerodynamic Disturbance in Wind Tunnel


C. E. Hann, A. A. Rao, M. Snowdon, N. Wongvanich, O. Winn and X. Q. Chen
AbstractThe roll dynamics of a 5 kg, 1.3 m high sounding
rocket are analyzed in a vertical wind tunnel. Signicant
turbulence in the tunnel makes the system identication of the
effective inertia, damping and asymmetry with respect to roll
challenging. A novel method is developed which decouples the
disturbance from the rocket frames intrinsic roll dynamics
and allows accurate prediction of roll rate and angle. The
parameter identication method is integral-based, and treats
wind disturbances as equivalent to a movement in the actuator
ns. The method is robust, requires minimal computation,
and gave realistic disturbance distributions reecting the
randomness of the turbulent wind ow. Two models, one with
constant damping and one with n angle dependent damping
were considered. The mean absolute roll rate of the rocket
frame observed in experiments was 16.4 degree/s and both
models predicted the roll rate with a mean absolute error
less than 0.10 degrees/s with a standard deviation less than
0.08 degrees/s. The roll angle (measured by an encoder), was
tracked by the model with a mean absolute error less than 0.20
degrees and a standard deviation less than 0.15 degrees. These
results prove the concept of this minimal modeling approach
which will be extended to varying wind speed, and pitch and
yaw dynamics in the future.
I. INTRODUCTION
Sounding rockets are a fast emerging technology giving
low-cost access to space. Sounding rockets also provide
important data on aerodynamics at speeds up to hypersonic
[1] which are typically very difcult to achieve in wind
tunnels. The idea is to use a minimal mathematical mod-
elling approach (e.g. [2]) to describe the overall effect of
complex dynamics on the inertia, damping and non-linear
spring dynamics of the pitch, yaw and roll axes of the
rocket.
This paper looks at the development of a simpler mod-
elling and system identication method for understanding
the roll dynamics and the effect of complex disturbances on
a small sounding rocket with n actuation. Thus, the paper
concentrates on the computational and algorithm aspects of
identifying parameters from the observed data. The control
development side is left to future work.
To simplify attitude control, this research concentrates on
controlling the roll of the rocket. This approach effectively
decouples the roll dynamics from the pitch and yaw avoid-
ing complex controllers and advanced analysis (e.g.[3]).
C. E. Hann, A. A. Rao, M. Snowdon and N. Wongvanich are with the
Dept. of Electrical and Computer Engineering, University of Canterbury,
Christchurch, New Zealand (phone: +64 3 364 2987; fax: +64 3 364 2761;
e-mail: chris.hann@canterbury.ac.nz).
X. Chen and O. Winn are with the Dept. of Mechanical Engineering,
University of Canterbury, Christchurch, New Zealand (phone: +64 3 364
2987; fax: +64 3 364 2761; e-mail: xiaoqi.chen@canterbury.ac.nz).
To provide an intermediate step from simulation to a
rocket launch, a vertical wind tunnel was built with a fan to
suck air past a rocket airframe with aluminium actuator ns.
There were signicant disturbances in the wind tunnel due
to ow separation and velocity uctuations, which made
the task of analyzing the rocket roll dynamics challenging.
System identication methods for identifying the param-
eters of the rocket roll response commonly involve non-
linear optimization and assume that noise and disturbance
is Gaussian with zero mean [4]. However, it is known that
for excited turbulence which can be achieved with an active
grid in a wind tunnel, the distribution of velocities is sig-
nicantly non-Gaussian [5]. This paper develops a method
that can be applied to any type of random disturbance,
independent of the distribution. The disturbance model is a
time series of effective n angle changes that capture the
observed effects on the rocket roll velocity due to turbulent
effects in the wind tunnel.
II. METHODOLOGY
A. Experimental Setup and Data Acquisition
The rocket was attached to a xed support which allowed
the rocket to roll freely about its primary axis. Control of
the rockets ns was provided by servos, which could be
commanded to move to a set deection (). The exact wind
speed prole inside the wind tunnel was not measured, but
the fan speed was held constant throughout the experiment.
Figure 1 shows the setup.
Fig. 1. Wind Tunnel Assembly
Throughout the test, the roll response was recorded for
several different n positions using an optical encoder
(Hewlett-Packard HEDS-5702-F00). A signicant amount
of time between the transitions was provided to ensure the
system was at steady state. Figures 2(a) and 2(b) show the
input control and output roll response.
2011 9th IEEE International Conference on
Control and Automation (ICCA)
Santiago, Chile, December 19-21, 2011
MonB2.5
978-1-4577-1476-4/11/$26.00 2011 IEEE 213
Fig. 2. (a) Fin angle and (b) Roll Response
Once this raw data had been attained, it was then
analyzed by the following key steps:
1) Fit a piecewise linear model to the roll step position
2) Find a relationship between the steady state angular
velocity and n deection
3) Create a bi-linear model for velocity step response
that ignores external disturbances
4) Apply integral method to identify time constant and
torque constant
5) Fit disturbance model to identify the disturbance
present
6) Re-simulate model with external disturbance
Steps 1-4 involve identifying the rockets intrinsic pa-
rameters. Since these parameters can only be determined
from actuation inputs, the data considered for these steps
was after the rst input deection. Steps 5-6 identify the
external torques on the rocket when the intrinsic parameters
are assumed known, hence for these steps, all data is
considered.
B. Modeling Structure
The roll dynamics of a rocket can be approximated by:
I
xx

P = qSd
_
C

+
d
2v
M
(C
P
P + C
R
R)
_
(1)
where q is dynamic pressure, S is the maximum cross-
sectional area of the rocket and d is the maximum diameter
of the rocket [6]. In general, the yaw-roll coupling in Equa-
tion (1) is weak hence for this paper the yaw contribution
to roll moment is ignored.
After simplifying some notation, the resulting model is
dened:
I v = cv + b +
d
(t) (2)
where:
I = I
xx
rotational inertia, v = P = roll velocity
c =
qSd
2
|C
P
|
2v
M
, b = qSdC

, = n angle (3)

d
(t) external disturbance torque affecting the roll
The parameters c and b in Equations (2) and (3) are
considered as the overall effective roll damping and torque
constant that converts a n angle into a torque about the
roll axis.
In the case of the wind tunnel experiment of Figure
1, the velocity v
m
and the density can be considered
to be constant over time. The effect of turbulence in
the wind tunnel is to give a disturbance on the intrinsic
response of the rocket to a n angle change. In other words,
the turbulence is represented by a time-varying external
disturbance torque
d
(t) in Equation (3).
C. Piecewise Linear Approximation to Roll Step Response
The roll position response of the rocket airframe is close
to piecewise linear as shown in Figure 2(b). Thus, as an
initial approximation, a piecewise linear model with eight
steps was formulated for :

model
(t) = a
1
+ b
1
t, T
1
t < T
2
.
.
. (4)
= a
8
+ b
8
t, T
8
t < T
end
with equality constraints:
a
i
+ b
i
T
i+1
= a
i+1
+ b
i+1
T
i+1
, i = 1, ..., 7 (5)
Let N
i
be the number of time points in each i
th
segment
t [T
i
, T
i+1
) of Equation (4) at the sampling frequency of
10Hz, and dene:
T
vector
= [t
1
, t
2
, ..., t
N
),

N = N
1
+ ... + N
8
(6)
Setting
model
(t) =
data
(t) for t T
vector
of Equation (6)
gives a set of equations in 16 unknowns a
1
, b
1
, ..., a
8
, b
8
which can be solved by linear least squares with the
equality constraints of Equation (5). The result determines
the best t of the model of Equation (4) to the measured roll
data of Figure 2(b). MATLAB was used for all calculations.
The result is depicted in Figure 3.
Fig. 3. Piecewise linear model of the roll response to the eight step-
changes in Figure 2(a)
D. Correlating the steady state angular velocity and n
deection
The stepwise approach of Figure 2(a) allows the steady
state dynamics of the n deection to roll rate to be simply
modelled. Eight step inputs of various magnitudes were
carried out in the experiment, as shown in Figure 2(a). The
slopes b
1
, ..., b
8
in Figure 3 determine eight steady state
values of the roll rate v
ss
corresponding to each steady state
n deection (f
ss
) of Figure 2(a). The values of f
ss
and v
ss
,
are plotted in Figure 4 with a linear best t:
v
ss,new
= f
ss
+ (7)
MonB2.5
214
= 0.6590, = 2.8856 (8)
The correlation coefcient of R
2
= 0.9914 shows that there
is a very strong linear relationship.
Fig. 4. Correlation between steady state deection and roll rate with the
best t line of Equations (7) and (8)
E. Bi-linear Modeling of Velocity Step Response
The roll dynamics of rockets are well known to be over
damped and thus they can be modelled as a rst order
differential equation. The solution to a step response of a
rst order system can be approximated as a bi-linear model.
The general form of a bi-linear model is dened:
v
model
(t) =
_
_
_
vssv0
t0
t + v
0
, 0 < t < t
0
v
ss
, t > t
0
(9)
where it is assumed that v
ss
and v
0
are known and t
0
is
unknown. The velocity step response, which the model of
Equation (9) approximates, is computed by differentiating
the roll angle data in Figure 2(b).
The values of v
ss
at each velocity step are determined
from the corresponding input deection f
ss
using the linear
model of Equation (7). This approach ensures that the
model predicted v
ss
is anti-symmetric about the point of
v
ss
= 0, and essentially acts as a model-based lter on the
external disturbances to obtain a consistent model of the
intrinsic rocket roll steady state response. Values of v
0
are
taken to be the previous value of v
ss
.
The break point t
0
in Equation (9) is assumed to lie less
than a given amount T
0
after the initial time point. For
this experiment T
0
was empirically chosen to be 4s to
minimize the search space, but it is essentially arbitrary.
The approach used to identify t
0
, is to increment from 0
to 4s in steps of 0.01s, and compute the least squares error
between the model of Equation (9) and the interpolated
measured data. Figure 5(a) shows the least squares error
plotted versus the position of the breakpoint for the 8th
velocity step response. The value of t
0
chosen in this case
was 1.61s which corresponds to the lowest error and the
best t of the bi-linear model to the data.
Figure 5(b) plots the approximated velocity (blue dotted
line) using Equation (9) versus the true velocity (green solid
line). This result shows that the model of Equation (9) is
sufcient to capture the overall steady state response even
though there is a lot of external disturbance present in the
raw velocity readings. This disturbance is not sensor error,
but is a real effect coming from turbulence in the wind
tunnel [1].
Fig. 5. (a) Error between model and data while varying the breakpoint
and (b) Bi-linear approximation of velocity step response
F. Modelling and Integral-Based Parameter Identication
Intrinsic Parameters
The bi-linear approximation described by Equation (9)
effectively removes the disturbance from the data. In other
words, this reconstructed velocity can be modelled with
the assumption that roll dynamics are only affected by
the inputs to the servo motors and there is no external
disturbance. Therefore, the rockets roll dynamics can be
described by:
I v = cv + f
ext
() (10)
where I is the rotational inertia, c is damping, and f
ext
() is
the external torque resulting from a change in the n angle
. Consider the following formulation of Equation (10):
I v = c(v ( + )) (11)
At each steady state =
ss
with v = 0 ,v =
ss
+
= v
ss
in Equation (11). Hence, the model of Equation
(11) contains implicitly the linearized steady state model of
Equation (7). Equating Equation (11) with Equation (10),
shows that f
ext
() = c( + ). Both parameters, I and
c, cannot be individually identied from the roll velocity
alone. Thus the model of Equation (11) is rewritten in the
form:
v = a(v ( + )) (12)
where:
a =
c
I
(13)
Equations (12) and (13) lump the two parameters I and c
into a generic damping parameter a.
G. Constant Damping Parameter
It is rst assumed that the unknown parameter a in
Equation (12) is constant for all time. To identify a,
an integral-based method similar to [7]-[8] is formulated.
Equation (12) is rst integrated over each input step period,
MonB2.5
215
which yields:
v(t) = v(T
1
) a
_
t
T1
(v ( + ))dt, T
1
t < T
2
.
.
. (14)
= v(T
8
) a
_
t
T8
(v ( + ))dt, T
8
t < T
end
Denote the measured velocity as v
data
(t). Equating v(t) of
Equation (14) with v
data
(t) for t T
vector
of Equation (6)
yields a set of N equations in the one unknown parameter
a. The integrals in Equation (14) are numerically evaluated
using the trapezium rule and the parameter a is found by
solving the set of simultaneous equations thus obtained.
Equation (12) can then be written in the form:
v = av + b( +

) (15)
where:
b = a ,

= / (16)
The quantity b

in Equation (15) represents the torque offset


that occurs due to asymmetries across the rockets primary
axis. In other words, with no n angle ( = 0) the rocket
will still spin. The parameter

can be interpreted as the
equivalent n angle that would reproduce this spin if the
rocket was perfectly symmetric. The parameter b is the
torque constant that relates a movement in the n angle
to an applied torque on the rocket.
H. Time-Varying Damping Parameter
The unknown parameter a in Equation (15) is next
assumed to vary across each step change of Figure 2(a).
Specically, a(t) is dened by:
a(t) = a
1
, T
1
t < T
2
.
.
. (17)
= a
8
, T
8
t < T
end
This assumption can reveal if the intrinsic parameters of
inertia and damping of the rocket are signicantly altered
by n actuation, and whether there is any clear trend. It is
an example of the overall concept of this research which
is to describe complexity by interactions of simpler models
(e.g. [2]) rather than one highly detailed model.
The method for identifying a time-varying parameter a
follows a similar approach as used for the constant damping
parameter. An integral formulation of Equation (12) is
dened by:
v(t) = v(T
1
) a
1
_
t
T1
(v ( + )) dt, T
1
t < T
2
.
.
. (18)
= v(T
8
) a
8
_
t
Tn
(v ( + )) dt, T
8
t < T
n
where the linear least squares solution denes the best t
of the model of Equation (18) to the measured data v
data
.
I. Modelling System Disturbances
Mathematically the model is dened by the reformulation
of Equation (15).
v = av + b( +

+ u
d
) (19)
where u
d
is the unknown time varying disturbance in the
wind tunnel due to turbulence, and a, b, are dened as in
Equation (16).
To include the fact that parameter a is changing for dif-
ferent n angles in Equation (16), without loss of generality
Equation (19) is reformulated as follows:
v(t) = a(t)(v(t) ((t) +

+ u
d
(t))) (20)
where and

are constants. Equation (20) follows from
substitution of b = a(t) from Equation (16) into Equation
(19) and factorizing for a(t).
J. Identifying Disturbance u
d
(t)
The parameters a(t), v(t), (t), and

in Equation (20)
are assumed known, and the unknown disturbance prole
u
d
(t) is taken to be a piecewise constant function over
intervals of length t which is chosen empirically. For the
experiment the measurements were logged every 0.1s thus
t is chosen to be 0.1s.
Let N
total
be the number of time intervals t that t in the
whole time period on the experiment, and denote v
data
(t) as
the measured data interpolated at the frequency of 100Hz,
or equivalently with the sampling period of t = 0.01s, let:
t
i
= it, i = 1, ..., N
total
(21)
Integrating Equation (20) from t
i1
to t yields:
v(t
i
) v(t
i1
) = a(t
i1
)
_
t
ti1
(v(t) ((t) +

) dt
+ u
d,i
(t
i
t
i1
), t
i1
t < t
i
(22)
u
d,i
= a(t
i1
)u
d,i
(23)
where u
d,i
is an unknown constant across the time interval
[t
i1
, t
i
) of length t and a(t) = a(t
i1
) is constant over
this interval since it is dened by Equation (17). Once u
d,i
is computed, the disturbance u
d,i
can be determined from
Equation (23). The result is a piecewise constant function
given by:
u
d
(t) =
Ntotal

i=1
_
H(t t
i1
) H(t t
i
)
_
u
d,i
(24)
H(t) =
_
1 t > 0
0 t < 0
(25)
where H(t) of Equation (25) denes the standard Heaviside
function and u
d
(t) describes the turbulence prole, or a set
of n angle changes that provides the best model t of
Equation (20) to the measured data v
data
(t).
MonB2.5
216
III. RESULTS AND DISCUSSION
A. Identication of Intrinsic Rocket Roll Parameters
The rocket roll intrinsic parameters are given by a, b,
and

as dened in Equations (15) and (16), without any
disturbance. The parameter a, which in turns determines b
is either constant or can vary across each step change of
the n angle in Figure 2(a). For a constant parameter a in
Equations (15) and (16), applying steps 1-4 as described in
Section A yields:
a = 1.314s
1
, b = 0.8665s
2
,

= 4.38

(26)
These parameters give the best least squares t of the
model of Equation (15) to the bi-linear approximated veloc-
ity step response in Figure 5(b). Figures 6(a) and (b) depict
a close-up of two sections of the overlay of the modeled
response versus this bi-linear approximated response. As is
expected the model precisely captures the steady state and
also captures quite closely the transient responses.
Fig. 6. (a) Close-up of the fth step input in Figure 5 and (b) Close up
of the sixth step input in Figure 5
For a time varying parameter a = a(t) in Equations (15)
and (16), steps 1 - 4 in Section A give a sequence of eight
values of a and b. Figure 7 plots the damping versus the
absolute value of this adjusted n angle
effective
to account
for the asymmetry of the rocket. Thus the quantity |
effective
|
assumes symmetry about =

.
Fig. 7. Plot of the absolute effective n angle, |
effective
|, versus damping
co-efcient denoted by circles. The best t line, ignoring the outlier, is
also shown
With the omission of the clear outlier, there is a strong
negative correlation with R
2
= 0.8280 in Figure 7. The
reason for the outlier is likely due to an increased amount
if disturbance present just at the point where the n angle
changes as shown in Figure 9. Note also that the 5th
and 7th values of a are slightly different even though the
n angles are both 40

. This approach ensures complete


freedom is allowed in the solution to reveal trends without
any presumptions on dynamics. To simplify notation dene:

effective
= +

(27)
which is the sum given in the model of Equation (15).
Equation (27) implies that to obtain zero torque from the
input n angle, an angle of =

= 4.38

on the rocket
is required.
Fig. 8. (a) Close-up of rst period of n angle change with relatively
small disturbance. (b) Close-up of third period of n angle change with
large disturbance.
The result of Figure 7 suggests that the damping coef-
cient a decreases as |
effective
| increases. This behaviour
can be explained by an increase in inertia as the ns
go outwards, since the inertia is on the denominator of
Equation (13).
Physically, as the ns go outwards there is an added mass
due to a greater mass of air being moved by the rockets
air frame. Thus a rocket rotating at a constant speed with
a large n angle will have a greater effective rotational
inertia than a rocket rotating with a small n angle. The
linear relationship in Figure 7 is described by:
a() = | +

| +

, = 0.0289,

= 2.4075 (28)
Replacing a in Equations (15) and (16) with the expres-
sion for a() in Equation (28), and factorizing, gives us an
extended non-linear model of the roll dynamics as follows:
v = a ()
_
v ( +

)
_
(29)
This model of Equation (29) is an example of the
philosophy of this research where more complex models
can be created from simplied model structures as required
to capture measured data and observed trends. Of course
more experiments, and ultimately rocket launches, will be
required to fully validate the accuracy of the model of
Equation (29), but it illustrates the concept. The parameter
a(t) in Equation (20) will also change as a function of the
rocket speed or Mach number, but these cases are left for
future work.
B. Identication of of External Disturbance u
d
(t)
The rocket roll response with external disturbances
present is modelled by Equation (20) where the parameters
a(t), and

are assumed to be known, and v(t) and (t)
are directly measured. The value of the parameter a is either
constant as given in Equation (19) or a function of n angle
MonB2.5
217
as given in Equation (29), which gives two possible u
d
(t)
proles. For the purposes of analysis and discussion dene:
u
d,constant
(t) =identied disturbance based on constant
damping coefcient of Equation (26) (30)
u
d,
(t) =identied disturbance based on dependent
damping coefcient of Equations (28) and (29)
(31)
Figure 9 shows close-ups of the result of applying step 5
from Section A, for two regions of data, R
1
and R
2
. Note
that a 3-point moving average is applied several times to
the velocity in Figure 4(b) to smooth out noise.
Fig. 9. (a) Region 1 (b) Region 2
Although the two disturbances are quite close and follow
the same trend, the disturbance u
d,
(dashed line) has in
general less pronounced peaks and undershoots the distur-
bance u
d,constant
. Specically, the 90% condence interval
of u
d,
(t) and u
d,constant
(t) are [-13.38, 14.28] and [-14.07,
15.15] respectively.
This result shows that for the u
d,constant
case, the algorithm
lumps the error in the constant damping coefcient a into
the disturbance prole, so that sharper changes occur. A
smoother prole of u
d,
suggests that the model of Equa-
tion (29) is potentially better than the model of Equation
(15) which has constant parameters. However, although the
modelling technique may give a more physically accurate
description of rocket roll dynamic, further wind tunnel tests
and rocket launches are required to fully validate whether
there is any signicant practical gain in using the more
complex model of Equation (29).
C. Simulating Model with External Disturbance
Using the identied disturbance u
d,
(t) and u
d,constant
(t)
from Equations (30) and (31), and the parameters in Equa-
tions (26) and (28), the models of Equations (19) and (29)
describing the full roll dynamics are numerically solved
for the velocity. The angular position of the rocket is then
found by integrating the modelled velocity and is compared
to the measured angular position. Both results show that the
two models closely capture the actual data obtained from
the experiment with each model output effectively overlaid.
Specically, the constant damping model tracked the roll
angle with a mean absolute error of 0.13 degrees and a stan-
dard deviation of 0.11 degrees. In comparison, the model of
Equation (29) with the theta dependant damping parameter
tracked the roll angle with a median absolute error of 0.18
degrees and a standard deviation of 0.14 degrees. Both
models give similarly accurate results as expected, since
any potential error in the intrinsic modelled parameters of
the rocket would be lumped into the disturbance prole.
IV. CONCLUSION
A minimal modelling approach and integral based pa-
rameter identication method were used to analyze roll
dynamics of a sounding rocket airframe inside a vertical
wind tunnel. By allowing a sufcient time to elapse after
each step input, the disturbance was decoupled from the
intrinsic dynamics of the rocket. This approach enabled an
accurate calculation of the inertia and damping. Two cases
of constant damping and time varying damping were con-
sidered. For time varying damping, the damping coefcient
identied across each n angle change was correlated with
the absolute value of the effective n angle which was a
correction for asymmetries in the rocket. A very signicant
negative correlation was found which suggested the inertia
of the rocket was increasing as the n angle increased.
The disturbance prole was modelled by a sequence of
effective n angles that reproduce the velocity response
as though the rocket was in a theoretically laminar ow.
The overall outputs for both models closely matched the
measured values for the roll rate and angle, demonstrating
that the algorithms can be generalized to more detailed
and physically accurate models. The identied disturbance
proles gave realistic magnitudes and close to normal
distributions, providing further validation to the overall
approach.
Future work includes investigating how other key vari-
ables such as rocket speed affects roll dynamics as well as
extending the approach to pitch and yaw.
REFERENCES
[1] A. Stamminger, et al., Sounding rockets as a real ight platform for
aerothermodynamic CFD validation of hypersonic ight experiments,
in Proceedings of the Fifth European Symposium on Aerothermody-
namics for Space Vehicles , vol. 563, pp. 431437, 2004.
[2] C.E. Hann, H.R. Sirisena and N. Wongvanich , Simplied Modeling
Approach to System Identication of Non-Linear Boat Dynamics, in
Proceedings of the American Control Conference. Baltimore, MD,
USA, pp. 5218 - 5223, 2010.
[3] H. I. Lee, et al.., Control design of spinning rockets based on co-
evolutionary optimization, in Control Engineering Practice, vol. 9,
pp. 149-157, Feb 2001.
[4] R. Jategaonkar and F. Thielecke, ESTIMA - an integrated software
tool for nonlinear parameter estimation, in Aerospace Science and
Technology, vol. 6, pp. 565578, Dec 2002.
[5] H. Makita, Realization of a large-scale turbulence eld in a small
wind tunnel, in Fluid Dynamics Research, vol. 8, pp. 5364, 1991.
[6] G. M. Siouris, Missile guidance and control systems, Springer Verlag,
2004.
[7] C. Hann et al., Integral-based parameter identication for long-term
dynamic verication of a glucose-insulin system model, in Computer
Methods and Programs in Biomedicine, vol. 77, pp. 259270, Mar
2005.
[8] C. Hann et al., The impact of parameter identication methods on
drug therapy control in an intensive care unit, in Open Med Inform
J, vol. 2, pp. 92104, 2008.
MonB2.5
218

You might also like