ICIUS 2007 ICIUS2007-A009

Oct 24-25, 2007

Bali, Indonesia

Automated Flight Test and System Identification for Rotary Wing

Small Aerial Platform using Frequency Responses Analysis
Widyawardana Adiprawita*, Adang Suwandi Ahmad = and Jaka Semibiring+

*School of Electric Engineering and Informatics

Institut Teknologi Bandung, Bandung, Id.
e-mail: wadiprawita@stei.itb.ac.id
=
School of Electric Engineering and Informatics
Institut Teknologi Bandung, Bandung, Id.
e-mail: asa@isrg.itb.ac.id
+
School of Electric Engineering and Informatics
Institut Teknologi Bandung, Bandung, Id.
e-mail: jaka@ itb.ac.id

Abstract scale rotorcraft. The maneuverability of such vehicles can

offer a tremendous operational advantages if harnessed
This paper proposes an autopilot system that can be used to during autonomous flight. Taking full advantage of the
control the small scale rotorcraft during the flight test for small-scale helicopter's natural abilities, however, has been
linear-frequency-domain system identification. The input a major challenge to its use as a UAV platform.
frequency swept is generated automatically as part of the
autopilot control command. Therefore the bandwidth Simply speaking, helicopters are difficult to fly. They are
coverage and consistency of the frequency swept is unstable, requiring continuous attention from the pilot. Four
guaranteed to produce high quality data for system primary inputs are necessary to control their motions: the
identification. Beside that we can set the safety parameter longitudinal and lateral cyclic for horizontal motions in
during the flight test (maximum roll / pitch value, minimum their respective directions; collective for vertical motion;
altitude, etc) so the safety of the whole flight test is and pedal for yaw (heading) motion.
guaranteed. This autopilot for automated flight test will be 1.1 Vehicle Dynamics Model
tested using hardware in the loop simulator for hover flight A number of available control design methods could, in
condition. theory, be applied to address the rotorcraft control issues
(described later in this chapter). To be successfully applied,
1 Introduction these methods require accurate vehicle dynamics models.
The physical approach to modeling involves deriving the
With the ability to take off and land vertically and hover, equations of motion from the ground up, using the
along with natural agility and controllability, rotorcraft will fundamental laws of mechanics and aerodynamics. This
extend the potential roles for UAVs. Helicopters already approach is referred to as first-principle modeling. For a
have an irreplaceable role among aircraft, indispensable for system involving numerous physical effects, the resulting
a variety of tasks ranging from medical evacuation to equations of motion are typically high-order nonlinear
transportation to construction in confined areas. Such coupled vector differential equations. A first principles
rotorcraft UAVs (RUAV) are already greatly valued by the approach requires considerable knowledge of and
military for applications for a range of battlefield tasks, experience with all the phenomena involved in rotorcraft
such as exploration and even combat operation. There are flight. This makes the first-principles approach difficult to
also numerous examples for civilian applications including do.
filmmaking (allowing both steady and dynamic aerial
views), close up inspection (bridges, buildings, dams) and For control design applications, as well as for flying
digital terrain modeling (where a small vehicle, because of qualities studies, simpler linear models are often sufficient.
its potential for closer proximity to the terrain and Linear model have been sued extensively and successfully
structures, could gather more detailed features). for rotorcraft. At a precise operating point, and even within
a certain region around that point, linear models accurately
Small-size rotorcraft tend to be naturally mode capture the essential effects of the vehicle dynamics. This is
maneuverable and mode responsive than traditional full- advantageous, since numerous analysis and design
ISBN 978-979-16955-0-3 50 © 2007 ICIUS
ICIUS 2007 ICIUS2007-A009
Oct 24-25, 2007
Bali, Indonesia

techniques are available for linear system. However, to

cover the entire flight envelope, multiple linear models and
therefore multiple controllers are necessary. An effective
approach to deriving accurate linear models of a plant is
linear system identification.
1.2 System Identification and CIFER
System identification uses experimental input-output data
collected from a plant to produce a mathematical
representation of the system's dynamics. This method is
attractive because it is direct, it is based on real data, and
integrates validation and model refinement into the
modeling process.
In the 1980s, rotorcraft experts undertook important efforts
in developing effective identification methods and tools.
CIFER (Comprehensive Identification from FrEquency
Responses from the US Army/NASA Rotorcraft Division,
is one of today's standard tools for rotorcraft identification).
Advanced high-bandwidth flight control design requires
models that are accurate at higher frequencies. A very Figure 2: Typical input output pair needed for
effective and accurate way to obtain linear rotorcraft frequency domain system identification
models is trough linear-frequency-domain system
identification. In this methods, frequency responses are
2 Methodology
estimated from the collected input-output data. Then, the
parameters of a linear model capturing the key vehicle The whole experimentation in this research is conducted in
dynamics are identified by tuning them to fit the estimated X-Plane simulation. This approach is taken based on
frequency responses. previous successful development of fixed wing autopilot
based on X-Plane hardware in the loop simulation. So the
big picture of the methodology are :
1. Development of PID based rotary wing autopilot
2. Development of frequency-sweep input algorithm
3. Development of simulation system in X-Plane
4. Processing the simulation data in CIFER
5. Hardware design of automated flight test system
6. Hardware in the loop (HIL) simulation
7. Processing the HIL simulation data in CIFER
8. Real world experimentation to get flight data
Figure 1: Frequency-domain system identification
procedure 9. Processing the real world experimentation data in
CIFER
However, the flight test which is suitable for linear-
frequency-domain system identification is difficult to be 10. Optimal Multiple Input Multiple Output (MIMO)
conducted by human pilot. This especially the case with the control system development based on vehicle's
small scale rotorcraft. The pilot controls the small scale model identified in CIFER
rotorcraft using visual feedback from safe distance. The 11. Real world experimentation to test Optimal
delay of response based from visual feedback make the MIMO control system
pilot difficult to produce the needed quality of frequency
input-output data. Beside that human pilot can not The research has been completed until the 4th step, so step
accurately covers the whole frequency needed for the test. 1 to 4 will be presented in this paper. The remaining steps
hopefully will be completed in near future.
This paper proposes an autopilot system that can be used to
control the small scale rotorcraft during the flight test for
linear-frequency-domain system identification.

ISBN 978-979-16955-0-3 51 © 2007 ICIUS

ICIUS 2007 ICIUS2007-A009
Oct 24-25, 2007
Bali, Indonesia

3 PID Based Rotary Wing Autopilot 2. X-Plane's functionality can be customized using a
plug in. A plug in is executable code that runs
A safe controller for rotary wing platform is needed to inside X-Plane, extending what X-Plane does. Plug
control the platform during frequency swept flight test. It ins are modular, allowing developers to extend the
has been mentioned before that frequency-domain system simulator without having to have the source code
identification is difficult to be conducted by human pilot. to the simulator. Plug ins allow the extension of
For now, only hover condition will be implemented in the the flight simulator's capabilities or gain access to
autopilot. Forward flight condition will be implemented in the simulator's data.
future development.
It has been shown in several previous research conducted
by other institution that PID based controller can be used to
control small rotary wing platform. This controller is
developed based on empirical approach without any
mathematical model. Of course the controller is not optimal
(because no mathematical can be used to justify this
optimality), but this kind of controller proved to be
relatively safe.
The objective of this research is to developed a rotary wing
mathematical model based on flight test, and not first
principle approach. So PID based controller is the logical
choice for the automated flight test. Based on empirical
experience, a PID cascaded controller is implemented.
Figure 4: X-Plane based simulation

Figure 3: PID cascaded controller for rotary wing

This PID controller is tested in X-Plane simulation
environment. There are two main reason why X-Plane is
used :
1. X-Plane is very interesting for non aerodynamicist
developer, because we can make an airframe based
only on its geometric dimension. The physics
model is based on a process called Blade Element
Theory. This set of principles breaks an airframe
down by geometric shape and determines the
number of stress points along its hull and airfoils.
Factors such as drag coefficients are then
calculated at each one of these areas to ensure the Figure 5: PID controller flow chart
entire plane is being affected in some way by
external forces. This system produces figures that After spending several hours of tuning the gain parameters,
are far more accurate than those achieved by the cascaded PID controller can hover the small rotorcraft
taking averages of an entire airfoil, for example. It (RC-02) in X-Plane reliably.
also results in extremely precise physical
properties that can be computed very quickly
during flight, ultimately resulting in a much more
realistic flight model. The X-Plane accuracy of the
flight model is already approved by FAA, for full
motion simulator to train commercial airline pilot.
ISBN 978-979-16955-0-3 52 © 2007 ICIUS
ICIUS 2007 ICIUS2007-A009
Oct 24-25, 2007
Bali, Indonesia

4. All flight parameter need to be logged into CIFER

text input format.

5 Data Processing using CIFER

Frequency responses fully describe the linear dynamics of

a system. When the system has nonlinear dynamics (to
some extent all real physical systems do), system
identification determines the describing functions which are
the best linear fit of the system response based on a first
harmonic approximation of the complete Fourier series. For
the identification, we used a frequency domain method,
developed by the U.S. Army and NASA, known as CIFER
(Comprehensive Identification from Frequency Responses).
While CIFER was developed specifically for rotorcraft
applications, it has been successfully used in a wide range
Figure 6: PID controller hovering RC-02 in X-Plane
of fixed wing and rotary wing, and unconventional aircraft
applications. CIFER provides a set of utilities to support the
4 Automated frequency-sweep input various steps of the identification process. All the tools are
integrated around a database system which conveniently
Reference [] mentioned that computer-generated frequency- organizes the large quantity of data generated throughout
sweep input can be both effective and time efficient in the identification process. In this research we are using
many flight application, but they must be properly CIFER Student Edition.
synthesized in order to obtain the desired frequency
spectrum. Tischler have developed an automated frequency- The steps involved in the identification process are:
sweep testing approach, using an exponential sweep and 1. Collection of flight data. The flight-data is
white noise. Here is the formulation of the automated collected during special flight experiments using
frequency-sweep input : frequency sweeps. This step is completed using
data logged from X-Plane simulation.
2. Frequency response calculation. The frequency
ω = ω min + K (ω max − ω min ) (1) response for each input-output pair is computed
using a Chirp-Z transform. At the same time, the
where coherence function for each frequency response is
calculated. This step is completed using FRESPID
C1t utility from CIFER.
K = C 2 [exp( ) − 1] (2)
Trec 3. Multivariable frequency domain analysis. The
single-input single-output frequency responses are
conditioned by removing the effects from the
The previously developed PID controller need to be secondary inputs. The partial coherence measures
modified to accommodate the this frequency-sweep input. are computed. This step is completed using
Here is the modification needed : MISOSA utility from CIFER.
1. Frequency-sweep input need to conducted in four 4. Window Combination. The accuracy of the low
input channel (roll/lateral cyclic, pitch/longitudinal and high frequency ends of the frequency
cyclic, yaw or pedal and collective axis) responses is improved through optimal
combination of frequency responses generated
2. PID controller need to be active on all control axis using different window lengths. This step is
except the one where the frequency-sweep input is completed using COMPOSITE utility from
conducted CIFER.
3. There will be several safety criteria (maximum 5. State-space identification. The parameters
absolute roll angle, maximum absolute pitch angle, (derivatives) of an a priori-defined state-space
minimum and maximum altitude, and maximum model are identified by solving an optimization
absolute yaw rate) during frequency-sweep input. problem driven by frequency response matching.
If one of the safety criteria is violated, the This step is completed using DERIVID utility
frequency-sweep input is aborted and the PID from CIFER. Due to limitation in the CIFER
controller will be activated on all control axis to Student Edition (5 state, 2 input), we have to
bring the platform back into steady hover separate lateral and longitudinal dynamics.
condition.
ISBN 978-979-16955-0-3 53 © 2007 ICIUS
ICIUS 2007 ICIUS2007-A009
Oct 24-25, 2007
Bali, Indonesia

6. Time Domain Verification. The final verification

of the model accuracy is performed by comparing
the time responses predicted by the model with the
actual helicopter responses collected from flight
experiments using doublet control inputs. This step
is completed using VERIFY utility from CIFER.
5.1 Frequency response calculation
The flight test simulation in X-Plane produced four kind of
frequency-sweep data for each of control axis (lateral,
longitudinal, rudder/pedal and collective). Data is recorded
in 50Hz frequency. After flight test the data is processed
using FRESPID. Here we only present one example from
those 4 control axis (the other are typically same).

Figure 9: Frequency-sweep input auto spectrum

(lateral axis)
The frequency response (magnitude and phase response) for
each input-output pair is computed using a Chirp-Z
transform in this step. This can be viewed as single-input
single-output characteristic of the rotary wing platform.

Figure 7: Frequency-sweep input (lateral axis)

Figure 10: Magnitude frequency response (lateral

axis)

Figure 8: Frequency-sweep output P / roll rate

(lateral axis)
To measure the frequency coverage of the frequency
sweep-input we can see the input auto spectrum plots. This
is the same as probability density function. The following
figure can give us the illustration of how good the coverage
of the automated frequency-sweep input algorithm.

Figure 11: Phase frequency response (lateral axis)

ISBN 978-979-16955-0-3 54 © 2007 ICIUS

ICIUS 2007 ICIUS2007-A009
Oct 24-25, 2007
Bali, Indonesia

A key metric to verify that the flight data is satisfactory for  - .6411E + 02 0.0000E + 00 0.3766E + 02
the purpose of system identification is its coherence.
Typically coherence value of larger than 0.6 is expected. F = 0.0000E + 00 - .6803E + 02 0.0000E + 00
0.6056E + 00 0.0000E + 00 - .5749E + 00 
(7)

0.8700E + 02 0.1000E + 01

G =  0.1000E + 01 0.1713E + 03 (8)
 - .4814E + 00 0.1000E + 01
Then we can verify the resulting state space model using
flight data and VERIFY utility. Here are the result :

Figure 12: Coherence value (lateral axis)

6 State-space identification

After conducting single-input single-output frequency

responses conditioning by removing the effects from the
secondary inputs using MISOSA and optimally combining
frequency responses generated with different window
lengths using COMPOSITE, we can start constructing the
state space model.
Figure 13: Aileron input to verify state space model
The state space form used by CIFER are :
•
M x = Fx + Gu (t − τ ) (3)

And the measurement output

y = Hx + Ju (t − τ ) (4)

In this research only lateral response will be studied and

only basic forces will be used as state variable. M is identity
matrix. J is zero matrix. H is identity matrix. and

P
x =  R  (5) Figure 14: Aileron to P from flight data compared
 Ay  with state space model result

Where τ is zero, so

aileron
u=  (6)
 rudder 
So we only need to find F and G matrix. We use DERIVID
utility to find F and G matrix. Here is the result :

ISBN 978-979-16955-0-3 55 © 2007 ICIUS

ICIUS 2007 ICIUS2007-A009
Oct 24-25, 2007
Bali, Indonesia

Controller Hardware (Internal Report), School of

Electric Engineering and Informatics, Bandung
Institute of Technology, 2006

Figure 15: Aileron to Ay from flight data compared

with state space model result

7 Concluding Remarks

There are still many steps to complete this research. But

there are several concluding remarks that can be drawn until
now, here they are :
1. Simple PID flight controller dan assist the flight
data collection for system identification using
frequency response analysis. The data (input auto
spectrum plots) proved that this system can give
very good frequency input coverage which is quite
difficult to get using manually controlled flight
especially for unmanned system.
2. The data collected give very good coherence for
single-input and single-output frequency response.
3. The verification of state space model gave correct
qualitative result, but still did not gave
satisfactorily result. There are still many
experimentation steps to increase the state space
model accuracy. Among those alternatives are :
o Try different window length in FRESPID
and COMPOSITE.
o Increase the frequency of data recording
o Develop parameterized state space model
template using first principle approach
References

[1] Mark B. Tischler, and Robert K. Remple, Aircraft and

Rotorcraft System Identification : Engineering
Methods with Flight Test Examples, American
Institute of Aeronautics and Astronatics, 2006
[2] Bernard Mettler, Identification Modeling and
Characteristics of Miniature Rotorcraft, Kluwer
Academic Publishers, 2003
[3] Bernard Mettler, Takeo Kanade, and Mark B.
Tischler, System Identification Modeling of a Model-
Scale Helicopter, Carnegie Melon University, 2003
[4] Widyawardana Adiprawita, Development of Simple
Autonomous Fixed Wing Unmanned Aerial Vehicle
ISBN 978-979-16955-0-3 56 © 2007 ICIUS