You are on page 1of 31

An Adaptive Flight Control system for a Flapping wing

Aircraft

Balaji Kartikeyan Chandrasekaran,1 and James E Steck2


Wichita State University, Wichita, Kansas, 67220, USA

In the present age of increased demand for unmanned aerial vehicles, flapping wing
unmanned aerial vehicle applications have become of interest, primarily because of their
ability to fly silently and at lower speeds. This work explores new territory through the
development of an adaptive flight controller for a bird-like flapping wing aircraft, using
modified strip theory1 to model the aircraft’s aerodynamics and Newtonian equations of
motion for the flight dynamics developed by Rashid2. The aircraft model is validated using
existing data from the Slow Hawk Ornithopter given by zakaria3. The goal of this paper is to
explore various adaptive flight control architectures, such as Model Reference Adaptive
Control and Adaptive Neural Network Inverse Control, leading to an advanced controller to
govern the longitudinal flight characteristics of the flapping wing aircraft. An approximate
math model of the slow hawk ornithopter was developed in MATLAB/Simulink4. A Model
Reference Adaptive Controller with Adaptive Bias Corrector was successfully able to adapt
to uncertainties and improved the tracking performance compared to no adaptation. It was
observed that with a B-Matrix failure the Adaptive controller was not able to reduce the
tracking error to zero. The same observation was also made for system with adaptation and a
PD controller. Another controller architecture in the form of Optimal Control Modification
was utilized to control the system and the performance of different architectures were studied
using the error metrics. OCM was able to adapt to the errors but higher learning rates
exhibited a poor tracking performance and time delay margin. It was observed that OCM
adaptation was able to successfully dampen the oscillations in system response.

I. Nomenclature
ABC = Adaptive Bias Controller
AFCS = Advanced Flight Control System
MST = Modified Strip Theory
𝑑⃗𝑏 𝑖
= 𝐷istance from body cg to point on root of wing aligned with wing cg
𝛤 = Maximum flapping angle amplitude
 = Cycle angle
𝜔 = Flapping frequency
𝛽𝑜 = Magnitude of dynamic Twist’s linear variation
̅
𝜃𝑤 = Mean pitch angle of chord

1
Graduate Research Assistant, Aerospace Engineering, and AIAA Student Member
2
Professor, Aerospace Engineering, and AIAA Senior Member

1
American Institute of Aeronautics and Astronautics
II. Introduction

A. A Brief History
One of the first known attempts of human flight was from Greek mythology in the form of story of Daedalus and
Icarus (figure 1). They were a father-son duo who, in their attempt to escape from prison, built a pair of wings using
twigs and wax. But Icarus flew too close to sun leading to melting of his wing and subsequent crash in sea and death
by drowning. Vymaanika-Shaastra a 4th century BC text written by Maharishi Bhardwaj dealt with operation of ancient
Vimana or flying crafts. One of which was called shakuna-vimana (figure 2) which was supposed to fly like a bird.

Figure 1. Daedalus and Icarus Figure 2. Shakuna Vimana as described in


Vymaanika-Shaastra

In modern history, Leonardo Da Vinci has been known to study the aerodynamics of these ornithopter even though
he did not build any practical machines. The first documented flight of flapping-wing aircraft was achieved by
Alphonse Penaud from France in 1874.

Figure 3. Powered Ornithopter of Lippisch5 Figure 4. Full-Scale Ornithopter “The Great Flapper”
Built by Dr DeLaurier

Due to the idea that humans can only fly by imitating birds led to inventors like Lippisch, Emil Hartman in 1959
etc.,. To develop Human powered models. Lippisch5 designed a human powered ornithopter (figure 3) using the input
given by Dr Brustmann, about the maximum power that can be developed by human arm and legs. He chose a flapping
wing configuration due to its potentially greater efficiency compared to a fixed wing concept. He tested the concept
with a pilot named Hans Werner Krause, whose goal was to reach a predetermined mark at about 250-300 meters from
launch. The importance of this kind of flight was the fact that the acceleration (Thrust) and lift was provided by the
flapping action of the wings alone. Etc. One of the most recent flights of an engine-powered ornithopter (figure 4)
was designed and flown by DeLAurier6 from Institute of Aerospace Studies at University of Toronto . The “Great
Flapper” was a full scale ornithopter powered by a 24hp engine and a flapping wing of 1.05 Hz with a span of 41.2 ft.
The aircraft successfully lifted off the ground for a few seconds.

2
American Institute of Aeronautics and Astronautics
B. The Present
The fixed wing aircraft has come a long way since wright brothers, motivated in part by World War I and World
War II we graduated from simple wood and canvas construction to advanced metal construction and use of composites.
With time even the speed, efficiency, carrying capability and safety also improved leading to present day passenger
aviation. The fixed wing UAV has played an important and decisive role in war, like Iraq war, and in fight against
terrorism. Civilian versions of these UAVs (both fixed wing and rotary wing) are nowadays utilized in agriculture,
off-shore oil platforms, access tall cellphone towers and high placed power lines.
Amid all this success with fixed wing aircraft, the idea of a flapping wing aircraft was forgotten until the advent
of Unmanned Aerial Vehicles (UAV) in recent times. The research in this field led to design and development of
various ornithopters which attempted to bio-mimic birds and insects. Delfly7, is a 3.07 g micro-ornithopter which has
an on-board camera to enable a vision-based navigation. Nano hummingbird8, was a 19g ornithopter designed at
Aerovironment as a part of the Defence Advanced Research Projects Agency’s (DARPA) requirement to develop a
small, biologically inspired UAV that can sustain hover and fly forward at speeds of about 10 m/s. One of the widely
used model9 is the “Slow Hawk Ornithopter “ developed by Sean Kinkade, which is also utilized in this work. A
modified version of this aircraft was developed at University of Maryland Morpheus Lab known as Odyssey which is
presently being used as a test bed for testing passive wing morphing10. Chen11 designed a unique butterfly ornithopter.

C. Bird Flight
The fundamental physics of flight of birds is very similar to that of aircraft, i.e., we need Lift to support weight
and Thrust to overcome drag. The only difference is that instead of separating the lift and thrust producing
mechanisms, a bird produces both by a single flap of wing.
As shown in figure 5, Birds produce their Lift and Thrust by twisting their wing forward in down stroke and tilting
it backward during upstroke. During down stroke it produces positive lift and Thrust but the upstroke produces reduced
lift and possibly negative thrust.
Different flapping species of animals can be distinguished based on wing movement12, which also affects the shape
of their wings. Chatterjee13 studied the aerodynamics of Argentavis, one of the largest flying birds ever thought to
exist based on fossils found in central and northwestern Argentina, and also its flight performance using computer
simulations, which was an accomplished glider cruising at a speed of 67 kmph and a gliding angle close to 3 deg.
Dial14 studied the evolution of avian flight by experimentally studying the wing strokes of ground birds. Hedenstrom15
studied the evolution of flight, the function of tails and also the ecological adaptations that the birds have to suit a
flying lifestyle. Pennycuick16 calculated the power requirements for horizontal flight in pigeon. In a nutshell, a bird
produces lift by moving the wing, thus producing a relative velocity over the airfoil. It produces thrust by active or
passive morphing of the wing, which makes it act as a propeller. Tobalske17 studied the biological aspects of the bird
flight which provided further insight into how the birds fly.

Figure 5. Bird Flapping Cycle

D. Current Work
1. Ornithopter Modelling and Control
Many models of ornithopter have been developed and were eventually controlled successfully. Shigeoka 18
developed an overall ornithopter transfer function model using experimental data acquired by strapping the ornithopter
to a rail thus restricting its movement to forward horizontal direction. Acceleration and position sensor information
were collected at different throttles. They subsequently developed a controller for its velocity and altitude. The velocity
loop was controlled using a PI controller, and so was the altitude loop. Subsequently, a two-dimensional state-space
model was also developed and its control and stability was discussed.

3
American Institute of Aeronautics and Astronautics
Julian19 modelled the flight dynamics of H2 Bird Ornithopter MAV shown in figure 6, a 13g MAV with a wingspan
of 26.5 cm. using system identification techniques. They successfully developed a co-operative controller to help it
traverse through a window with help of on-board sensors and a ground station based camera.
Jackowski20 designed and constructed an autonomous ornithopter
known as Phoenix as part of master’s thesis. The ornithopter was
passively stable in other degrees of freedom but it was found difficult to
control in pitch. A PD controller was found to be sufficient for pitch
control and a proportional control was used for approximate velocity
control.
Krashanitsa21 studied the off-the shelf Cybird P2 ornithopter by
subjecting it to thorough flight testing and developing the flight dynamics
model using these results. The aerodynamic model was developed by
subjecting the model to wind tunnel testing.
2. Model Reference Adaptive Controller
In the work presented here in this paper we use a Model reference
Figure 6. H2 Bird Ornithopter
Adaptive Controller. Model Reference Adaptive Control (MRAC) is an
adaptive flight control architecture which makes a non-linear system
follow desired dynamics even in the presence of modelling error and failures. Early MRAC methods were introduced
in the 1980s. Neural networks were used to adjust for unmodelled system dynamics by Narendra and Annaswamy 22
and Lewis23. The MRAC was implemented for controlling aircraft by Rysdyk24. They combined the adaptive control
techniques with a linear inverse aircraft controller. Since the system adapts to change, it was demonstrated that MRAC
can be utilized to control aircraft that are in loss of control regime as shown by Bosworth25, or if the aircraft is damaged
as shown by Nguyen26.
The General Aviation Flight Lab at Wichita State University has been conducting research on Adaptive control
systems since mid-1990s. An early decoupled control method for general aviation was developed by Duerksen27 as
part of his research funded by NASA’s Advanced General Aviation Transportation Experiment (AGATE) program.
He developed a longitudinal decoupled flight controller with fuzzy logic that commanded flight path angle and
airspeed. Steck28 researched an adaptive control system with decoupled pilot commands and simulated with a general
longitudinal delta wing model and included the pitch attitude and airspeed commands. This technique included a linear
compensator and an Artificial Neural Network that was trained offline.
Since 2001, WSU partnered with Beechcraft Corporation to further develop such Advanced Flight Control System
(AFCS) algorithms for general aviation applications. For purposes of modeling, simulation and flight testing, the
algorithms developed were tested to a Beechcraft CJ-144 Small Aircraft Transportation System (SATS) Fly-By-Wire
test bed. Work by Pesonen29 Advanced the AFCS into MRAC dynamic inverse controller. The more advanced version
of the controller called the Adaptive Biased Controller (ABC) was developed by Steck30. The ABC was an ANN that
included only the bias term of the neural network. The present work also utilizes the same architecture to explore the
ABC’s ability to adapt ornithopter flight dynamics and model errors. The same architecture or a modified form was
utilized in ref [19, 20].

E. Scope of this paper


One of the common characteristics among the above mentioned ornithopter projects and Han’s work 31 was that
they concentrated on developing the flight models for the particular aircraft model they were working with. Here we
develop a general Simulink model for an ornithopter that can be utilized to simulate and develop controllers. Also,
one of the major requirements that was considered was that, considering the advances in the field of morphing wings
and their control21, it should also include the ability of the bird-like ornithopter to morph its wing shape to have an
agile flight and maneuverability.
The aim of this work is to develop a general non-linear model for a bird-like flapping wing aircraft using
appropriate model for aerodynamics, which includes ability to consider the shape of the wings, and flight dynamics,
followed by selection of a suitable specific aircraft model with all the necessary inputs.
After the selection of a suitable model with all the geometric and other inputs required for the aerodynamics and
flight dynamics model, a model follower Adaptive Non-linear Dynamic inversion controller is designed. A simple
ANN is used and then its effect is studied in a scenario of actuator failure and other scenarios.
A model of the ornithopter aerodynamics will be developed in MATLAB/Simulink environment and will be
coupled with a suitable flight dynamics model. This general model will be simulated with an aircraft model to
understand the dynamics of the system and effects of various A-matrix and B-Matrix failures on the system. Then the
Adaptive component will be added to the system to study whether it is able to adapt to this aircraft model with

4
American Institute of Aeronautics and Astronautics
oscillating states. Since the aircraft flight dynamics is similar to that of fixed wing aircraft, with which the adaptation
has been successful in tracking the commands in presence of failures, we will explore whetehr the adaptive controller
will be successful with ornithopter flight dynamics in improving its tracking performance during a commanded pitch
rate doublet.
An OCM based approach for fixed wing aircraft was also adapted to develop a controller for the given system and
its performance will be compared with that of the controller in previous paragraph using the error metrics.

III. Aerodynamic Model


Various researchers have used a wide range of aerodynamic models for flapping wing aircraft. Some used
experimental methods like wind tunnel and motion capture techniques like Krashanitsa 21. Grauer32, 33, and Shigeoka18
used an experimental technique of system identification by visual tracking. Rose34 considered a combined approach
of wind tunnel measurements and free flight measurements using sensors. While Kaviyarasu and Kumar35 utilized an
available simulation platform like X-plane to calculate the aerodynamics. The third approach to development of an
aerodynamic model is a theoretical one based on steady-state aerodynamic models like Rashid2, who uses a quasi-
steady state aerodynamics model and ignores the unsteady wake effects.

A. Wing Aerodynamics Model


The aerodynamics model utilized here in this paper was developed by DeLaurier1. The model presented unsteady
aerodynamics for an ornithopter using modified strip theory (MST). This model has been used by Kim36,Han31, Chalia
and Bharti37 and Beng38. The model breaks down the entire wing into small sections of finite width and then the Lift,
Drag and Moment is calculated for each section, based on the flow properties locally along with the kinematic effects
of the flapping action to get the local flow characteristics. These local aerodynamic properties are then integrated over
the entire wing to obtain the total lift, thrust and power.
1. Program
The programming of the aerodynamic model was done by Beng38 in MATLAB and we have modified the code
by retaining aerodynamic parts but,the Graphical User Interface (GUI) and the optimization module has been removed,
to suit the needs of the present work
The following assumptions was made by Beng during the programming
1. Negative-𝛼 ′ stalling does not occur.
2. Dynamic stall criterion is not incorporated.
3. The crossflow drag coefficient , (𝐶𝑑 )𝑐𝑓 was chosen to be that for a high-AR flat plate, given
by Hoerner39 as 1.98.
4. The texture of wing’s surface is assumed to be such as to produce a full chord turbulent
boundary layer. Thus, the friction drag coefficient, (𝐶𝑑 )𝑓 , obtained from Hoerner40 is given
by
0.89 (1)
(𝐶𝑑 )𝑓 =
[log(𝑅𝑒 )]2.58

B. Tail Aerodynamics Model


The tail module is adopted from the work done by Grauer32. The model was developed using wind tunnel data.
The values of these co-efficients are calculated based on the equations below:
𝐶𝑥 = 𝐶𝑥𝑜 + 𝐶𝑥 2 𝛼 2 (2)
𝛼

𝐶𝑦 = 𝐶𝑦𝛼𝛽 𝛼𝛽 (3)

𝐶𝑧 = 𝐶𝑧𝑜 + 𝐶𝑧𝛼 𝛼 + 𝐶𝑧𝛽 𝛽 (4)

𝐶𝑙 = 𝐶𝑙𝛽 𝛽 + 𝐶𝑙𝛼𝛽 𝛼𝛽 (5)

𝐶𝑚 = 𝐶𝑚𝑜 + 𝐶𝑚 2 𝛽 2 + 𝐶𝑚𝛼 𝛼 + 𝐶𝑚𝛽 𝛽 + 𝐶𝑚𝛼𝑉 𝛼𝑉 (6)


𝛽

5
American Institute of Aeronautics and Astronautics
𝐶𝑛 = 𝐶𝑛𝛽 𝛽 (7)

In order to include the control variables, 𝑡𝑝𝑎 = 𝑇𝑎𝑖𝑙 𝑃𝑖𝑡𝑐ℎ 𝐴𝑛𝑔𝑙𝑒 and 𝑡𝑟𝑎 = 𝑇𝑎𝑖𝑙 𝑅𝑜𝑙𝑙 𝑎𝑛𝑔𝑙𝑒, certain changes
were made to equations (2)-(7). The pitch control variable , 𝑡𝑝𝑎 , was added to α of the tail. Hence, the angle of attack
of the tail is sum of body angle of attack and 𝑡𝑝𝑎 .
The lateral control variable 𝑡𝑟𝑎 , is incorporated in a slightly different way. The lateral control of the ornithopter is
achieved by re-positioning the angle at which the forces act so that some of the lift generated by the tail is utilized for
generating a small side force. Thus, the aerodynamic forces X, Y and Z of that tail are rotated by roll angle of 𝑡𝑟𝑎 .
Thus the new forces considering the roll angle is given by:

F = [1 0 0; 0 cos(t_ra) sin(t_ra); 0 − sin(t_ra) cos(t_ra)] ∗ [X; Y; Z] (8)

Also, there is small change made in the Yawing moment equation, where a moment arm is added to add the effect
of side-force produced by re-directing the lift of tail.
𝑁 = 𝐶𝑛 𝑄𝑆𝑡 𝑏𝑡 + 2 ∗ 𝐹(2) (9)
Where,
2 = longitudinal moment arm in ft
𝐹(2) = Control force in y-direction in body axis

IV. Flight Dynamics

The model presented here can be called the 2-panel method, figure (7). This method assumes that an ornithopter
can be broken down into body, including tail and fin (if any), and 2-wing panels. The forces and moments produced
by wings are coupled to body through the reactionary forces at the hinge points where they are attached to the body
as shown in figure (8)

Figure 7. 2-Panel Method Figure 8. Position of Body and Hinge Joints with
Forces and Moments

The equations of motion are written down separately for body and each of the wings and then they are connected
through the kinematics equations. The final equations are given as a set of 18 Linear equations with 18 unknowns,
namely the states U, V, W, P, Q, R and the reaction forces and moments at the hinge joints between each of the wings
and body. The present work converts them into 6-DOF equations which then becomes easier to model in
MATLAB/Simulink.
The final expressions in vector form is given in equations (10) and (11). These equations are from the author’s
thesis41 and more details can be found there.

6
American Institute of Aeronautics and Astronautics
Force (𝑀𝑏 + 𝑀𝑤1 + 𝑀𝑤2 )𝑉 ⃗⃗̇ + 𝑀 𝜔̇⃗⃗ 𝑋(𝑑⃗ −𝐷 ⃗⃗𝑤1 ) + 𝑀𝑤2 𝜔̇⃗⃗𝑏 𝑋(𝑑⃗𝑏 −𝐷 ⃗⃗𝑤2 ) = (10)
𝑏 𝑤1 𝑏 𝑏𝑤1 𝑤2
vector
Equation 𝐹⃗𝑏 + 𝐹⃗𝑎𝑒𝑟𝑜 + 𝐹⃗𝑎𝑒𝑟𝑜 + 𝐹⃗𝑤1 + 𝐹⃗𝑎𝑒𝑟𝑜 + 𝐹⃗𝑤2 + 𝑀𝑏 (𝜔
𝑏 𝑤1 𝑤2
⃗⃗𝑏 )
⃗⃗𝑏 𝑋𝑉
− 𝑀𝑤1 [(𝜔 ⃗⃗𝑏 𝑋(𝜔⃗⃗𝑏 𝑋𝑑⃗𝑏𝑤1 )) + 𝜔 ⃗⃗𝑏 𝑋𝑉⃗⃗𝑏 + (𝜔⃗⃗𝑏 𝑋𝜔 ⃗⃗𝑏 𝑋(−𝐷 ⃗⃗𝑤1 ))
+ (2𝜔⃗⃗𝑏 𝑋𝜔
⃗⃗𝑐1 𝑋(−𝐷 ⃗⃗𝑤1 )) + 𝜔̇⃗⃗𝑐1 𝑋(−𝐷 ⃗⃗𝑤1 ) + (𝜔 ⃗⃗𝑐1 𝑋𝜔⃗⃗𝑐1 𝑋(−𝐷 ⃗⃗𝑤1 ))
+𝜔⃗⃗𝑤1 𝑋(𝑉⃗⃗𝑏 + 𝜔⃗⃗𝑏 𝑋(−𝐷 ⃗⃗𝑤1 ) + 𝜔 ⃗⃗𝑏 𝑋𝑑⃗𝑏 + 𝜔
𝑤1
⃗⃗𝑐1 𝑋(−𝐷 ⃗⃗𝑤1 ))]
− 𝑀𝑤2 [(𝜔 ⃗⃗𝑏 𝑋(𝜔⃗⃗𝑏 𝑋𝑑⃗𝑏𝑤2 )) + 𝜔 ⃗⃗𝑏 𝑋𝑉⃗⃗𝑏 + (𝜔⃗⃗𝑏 𝑋𝜔 ⃗⃗𝑏 𝑋(−𝐷 ⃗⃗𝑤2 ))
+ (2𝜔⃗⃗𝑏 𝑋𝜔
⃗⃗𝑐2 𝑋(−𝐷 ⃗⃗𝑤2 )) + 𝜔̇⃗⃗𝑐2 𝑋(−𝐷 ⃗⃗𝑤2 ) + (𝜔 ⃗⃗𝑐2 𝑋𝜔⃗⃗𝑐2 𝑋(−𝐷 ⃗⃗𝑤2 ))
+𝜔 ⃗⃗
⃗⃗𝑤2 𝑋(𝑉𝑏 + 𝜔 ⃗⃗
⃗⃗𝑏 𝑋(−𝐷𝑤2 ) + 𝜔 ⃗
⃗⃗𝑏 𝑋𝑑𝑏 + 𝜔 ⃗⃗
⃗⃗𝑐2 𝑋(−𝐷𝑤2 ))]
𝑤2

Moment
𝐼𝑏 𝜔̇⃗⃗𝑏 + 𝐼𝑤2 𝜔̇⃗⃗𝑏 + 𝐼𝑤1 𝜔̇⃗⃗𝑏 + (𝑑⃗𝑏𝑤1 − 𝐷 ⃗⃗̇ + 𝑀 𝜔̇⃗⃗ 𝑋(𝑑⃗
⃗⃗𝑤1 )𝑋 (𝑀𝑤1 𝑉 ⃗⃗ (11)
𝑏 𝑤1 𝑏 𝑏𝑤1 − 𝐷𝑤1 ))
vector
Equation + (𝑑⃗𝑏𝑤2 − 𝐷 ⃗⃗̇ + 𝑀 𝜔̇⃗⃗ 𝑋(𝑑⃗
⃗⃗𝑤2 )𝑋 (𝑀𝑤2 𝑉 ⃗⃗
𝑏 𝑤2 𝑏 𝑏𝑤2 − 𝐷𝑤2 ))

= 𝑀⃗⃗⃗𝑎𝑒𝑟𝑜 + 𝑀 ⃗⃗⃗𝑎𝑒𝑟𝑜 + 𝑀 ⃗⃗⃗𝑎𝑒𝑟𝑜 − 𝐼𝑤1 𝜔̇⃗⃗𝑐1 − 𝐼𝑤2 𝜔̇⃗⃗𝑐2 − 𝜔


⃗⃗𝑏 𝑋𝐼𝑏 𝜔
⃗⃗𝑏
𝑏 𝑤1 𝑤2
−𝜔⃗⃗𝑤1 𝑋𝐼𝑤1 𝜔 ⃗⃗𝑤1 − 𝜔 ⃗⃗𝑤2 𝑋𝐼𝑤2 𝜔 ̇
⃗⃗𝑤2 − 𝐼𝑤1 𝜔 ̇
⃗⃗𝑤1 − 𝐼𝑤2 𝜔⃗⃗𝑤2
+ (𝑑⃗𝑏 −𝐷
𝑤1
⃗⃗𝑤1 )𝑋 (𝑀𝑤1 [(𝜔 ⃗⃗𝑏 𝑋(𝜔⃗⃗𝑏 𝑋𝑑⃗𝑏 )) + 𝜔
𝑤1
⃗⃗𝑏 𝑋𝑉 ⃗⃗𝑏
+ (𝜔
⃗⃗𝑏 𝑋𝜔⃗⃗𝑏 𝑋(−𝐷⃗⃗𝑤1 )) + (2𝜔⃗⃗𝑏 𝑋𝜔
⃗⃗𝑐1 𝑋(−𝐷⃗⃗𝑤1 )) + 𝜔̇⃗⃗𝑐1 𝑋(−𝐷⃗⃗𝑤1 )
+ (𝜔
⃗⃗𝑐1 𝑋𝜔⃗⃗𝑐1 𝑋(−𝐷⃗⃗𝑤1 )) + 𝜔
⃗⃗𝑤1 𝑋(𝑉⃗⃗𝑏 + 𝜔
⃗⃗𝑏 𝑋(−𝐷 ⃗⃗𝑤1 ) + 𝜔⃗⃗𝑏 𝑋𝑑⃗𝑏 𝑤1

+ 𝜔 ⃗⃗𝑤1 ))] + 𝐹⃗𝑎𝑒𝑟𝑜 + 𝐹⃗𝑤1 )


⃗⃗𝑐1 𝑋(−𝐷 𝑤1

+ (𝑑⃗𝑏𝑤2 − 𝐷 ⃗⃗𝑤2 )𝑋(𝑀𝑤2 [(𝜔 ⃗⃗𝑏 𝑋𝑑⃗𝑏𝑤2 )) + 𝜔


⃗⃗𝑏 𝑋(𝜔 ⃗⃗𝑏 𝑋𝑉 ⃗⃗𝑏
+ (𝜔
⃗⃗𝑏 𝑋𝜔⃗⃗𝑏 𝑋(−𝐷 ⃗⃗𝑤2 )) + (2𝜔⃗⃗𝑏 𝑋𝜔
⃗⃗𝑐2 𝑋(−𝐷⃗⃗𝑤2 )) + 𝜔̇⃗⃗𝑐2 𝑋(−𝐷 ⃗⃗𝑤2 )
+ (𝜔
⃗⃗𝑐2 𝑋𝜔⃗⃗𝑐2 𝑋(−𝐷 ⃗⃗𝑤2 )) + 𝜔
⃗⃗𝑤2 𝑋(𝑉⃗⃗𝑏 + 𝜔
⃗⃗𝑏 𝑋(−𝐷 ⃗⃗𝑤2 ) + 𝜔 ⃗⃗𝑏 𝑋𝑑⃗𝑏 𝑤2

+ 𝜔 ⃗⃗𝑤2 ))] + 𝐹⃗𝑎𝑒𝑟𝑜 + 𝐹⃗𝑤2


⃗⃗𝑐2 𝑋(−𝐷 𝑤2

V. Simulink Model and Aircraft Selection

A. Introduction
There are various physical models that the researchers use for ornithopter research that can serve as an example
for our model, but since majority of them are proprietary or the data available are not sufficient to complete the
requirements of our sub-systems, our choice was very limited. One of the options was the “The Great Flapper” that
was built by Dr DeLaurier. Even though the wing section details, like number of sections, chord length of each section,
airfoil of each section etc.,.and mass details were available for building the aerodynamics model, necessary
information like wing CG location, Moment of Inertia etc., were not available to complete the flight dynamics. The
second option was the more common UAV “Slow Hawk Ornithopter” that was built by Sean Kinkade, which is very
common among hobby enthusiasts. The geometric details of this selected aircraft are provided in the following sub-
headings.
In this paper, the aerodynamics model developed by Dr DeLaurier is being used in combination with flight
dynamics model developed by Rashid. The aircraft selected to be simulated is the “slow hawk ornithopter”.

B. Aerodynamics Block
As we can see in the figure (9), the aerodynamics block has the inputs as enumerated in table 1 below. Velocity
and angle of attack are states from the flight dynamics model while the tail angles act as control variables.

7
American Institute of Aeronautics and Astronautics
INPUT OUTPUT
S S

Figure 9. Aerodynamic Block Showing Inputs and Outputs

Table 1. Aerodynamic Block Inputs


Input Variable Units Description
The frequency with which the
Frequency Hz
aircraft wings flap
The velocity vector of aircraft in
𝑉𝑏 ft/s
body axis
Angle of attack of the aircraft body
calculated as the inverse tangent of
alpha rad
ratio of vertical velocity of the body
to forward velocity of the body
The tail setting angle, also referred
𝑡𝑝𝑎 rad
to as tail pitch angle.
The tail roll angle, it is the angle by
which the tail will be rolled about
𝑡𝑟𝑎 rad
the body x-axis for lateral axis
control

The outputs from the aerodynamic model are the forces and moments, and includes other important variables
that are required by flight dynamics block. The details of every output is presented in table (2) below:

8
American Institute of Aeronautics and Astronautics
Table 2 Aerodynamic Block Outputs
Output
Symbol Units Description
Variable
The aerodynamic forces in all three directions developed by body and
Force-Body 𝐹⃗𝑏 lbf
tail combined
The aerodynamic forces in all three directions developed by port
Force-wing 1 𝐹⃗𝑤1 lbf
(left) wing
The aerodynamic forces in all three directions developed by
Force-wing 2 𝐹⃗𝑤2 lbf
starboard (right) wing
Moments-body ⃗⃗⃗𝑎𝑒𝑟𝑜
𝑀 lbf-ft The aerodynamic moments generated by body-tail combination
𝑏

Moments-wing 1 ⃗⃗⃗𝑎𝑒𝑟𝑜
𝑀 lbf-ft The aerodynamic moments generated by port (left) wing
𝑤1

Moments-wing 2 ⃗⃗⃗𝑎𝑒𝑟𝑜
𝑀 lbf-ft The aerodynamic moments generated by starboard (right) wing
𝑤2

Omega_c1 𝜔
⃗⃗𝑐1 rad/sec The angular velocity of the port (left) wing due to flapping
Omega_c2 𝜔
⃗⃗𝑐2 rad/sec The angular velocity of the starboard (right) wing due to flapping
Omega_c1_dot 𝜔̇⃗⃗𝑐1 rad/sec2 The angular acceleration of the port(left) wing

Omega_c2_dot 𝜔̇⃗⃗𝑐2 rad/sec2 The angular acceleration of the starboard (right) wing
Flapping angle of the wing which is a function of cycle angle, ∅ and
Gamma rad
the flapping amplitude in the aerodynamic model

The wing aerodynamic model is introduced in the SIMULINK model as an s-function. The inputs to this s-function
is enumerated in table (3).
Table 3. Inputs to Aerodynamic S-Function
Variable Symbol Units Description
Alpha0In 𝛼𝑜 rad Zero Lift Angle
EtaS 𝜂𝑠 No Dim Leading edge suction efficiency
Cmac 𝐶𝑚𝑎𝑐 No Dim Co-efficient of moment about aerodynamic center of wing
MaxAlphaStallIn (𝛼𝑠𝑡𝑎𝑙𝑙 )𝑚𝑎𝑥 rad Maximum stall angle of the airfoil
Cs cs ft Chord length of each section of wing
ThetaWIn 𝜃𝑤̅ rad Mean pitch angle of chord with respect to pitch axis
𝛽𝑜
Beta0 rad Magnitude of dynamic twist’s linear variation
Ys 𝑦 ft Co-ordinate along semi-span
Dy dy ft Width of each section of a wing

C. Aircraft Parameters-Aerodynamics
1. Wings
As mentioned in the introduction, the aircraft selected for simulation in this thesis is the “Slow Hawk Ornithopter”
built by Kinkade9, 42. Various properties and geometric dimensions required for the aerodynamics model developed in
previous sections are available from the work of Zakaria3.
The values of the variables like flapping angle magnitude, pitch angle of flapping axis and magnitude of dynamics
twist’s linear variation were selected , though different from ones chosen by Zakaria3, in such a way that the lift
generated by the wings matched the weight of the aircraft at the given speed. Their values are given in table 4.

9
American Institute of Aeronautics and Astronautics
Table 4. Values of Initialized Parameters

Variable Description Value

Γ Flapping angle Magnitude 35 deg


𝜃𝑎̅ Pitch angle of flapping axes 5.2 deg
𝛽𝑜 Magnitude of dynamics twist’s linear variation 0.5 deg
Table 5. Input Parameters

̅𝒂 (deg) Lift
𝒇 (Hz) 𝑼(ft/sec) 𝜞(deg) 𝜽 𝜷𝒐 (deg/ft) Thrust(lbs)
(lbs)
3.24 15.65 35 5 35 0.9259415 0.97

The aerodynamic block was populated and run


2.5 and the calculated Lift, Thrust and Moments of the
wing are shown below in figures 10, 11 and 12.
2
Since optimization was not the goal of this work, we
chose a set of values for various design variables to
1.5
generate enough lift whose magnitude is equal and
opposite to the weight of the aircraft. The values of
Lift(in lbs)

these variables are given in table 5.


1

0.5

-0.5
0 1 2 3 4 5 6 7 8 9 10
Time (in secs)

Figure 10. Lift Generated by Wings

1.5 0.9

0.8

1 0.7
Pitching Moment (in lbs-ft)

0.6
0.5
0.5
Thrust (in lbs)

0.4
0

0.3

-0.5 0.2

0.1
-1
0
0 1 2 3 4 5 6 7 8 9 10
Time (in secs)
-1.5
0 1 2 3 4 5 6 7 8 9 10
Time (in secs)

Figure 11. Total Thrust Generated by Wings Figure 12. Total Pitching moment Generated by
Wings

10
American Institute of Aeronautics and Astronautics
1.1 Tail

Figure 13. Tail Aerodynamic Co-Efficients from Grauer


In the model considered in this work, the signs for the values of 𝐶𝑚𝛼 𝑎𝑛𝑑 𝐶𝑚𝛼𝑉 has been changed in order to have
a stable aircraft model after substituting the values for other variables. Various aerodynamic co-efficients are given
below from Grauer32. The forces and moments generated by tail in a trim condition are given in table 6.
Table 6. Tail Forces and Moments Values
Forces Value
Drag -0.0401 lbs
Lift 0.0385 lbs
Side Force 0 lbs
Pitching Moment -0.4892 lbs-ft

D. Flight Dynamics Block


The equations (10) and (11) were coded in SIMULINK. Figure 15 shows the block.

INPUT OUTPUTS
S

Figure 14. Flight Dynamics Block Showing Inputs and Outputs

11
American Institute of Aeronautics and Astronautics
The inputs and outputs are explained in the following tables.
Table 7. Flight Dynamics Block Inputs
Equivalent variable in equations
Block input
10 and 11
Force_body 𝐹⃗𝑏 + 𝐹⃗𝑎𝑒𝑟𝑜 𝑏
Force_wing 1 𝐹⃗𝑤1 + 𝐹⃗𝑎𝑒𝑟𝑜𝑤1
Force_wing 2 𝐹⃗𝑎𝑒𝑟𝑜 + 𝐹⃗𝑤2
𝑤2
Moments_body 𝑀⃗⃗⃗𝑎𝑒𝑟𝑜
𝑏
Moments_wing 1 ⃗⃗⃗𝑎𝑒𝑟𝑜
𝑀 𝑤1
Moments_wing 2 ⃗⃗⃗𝑎𝑒𝑟𝑜
𝑀 𝑤2
Omega_c1 𝜔⃗⃗𝑐1
Omega_c2 𝜔⃗⃗𝑐2
Omega_c1_dot 𝜔̇⃗⃗𝑐1
Omega_c2_dot 𝜔̇⃗⃗𝑐2
dI_w1 𝐼𝑤1̇
dI_w2 𝐼𝑤2 ̇

Table 8. Flight Dynamics Block Outputs


Equivalent variable in equations
Block output
70 and 71
Vb_dot 𝑉⃗⃗̇
𝑏
Wb_dot 𝜔̇⃗⃗𝑏
Euler ∅, 𝜃, 𝛹
Ve 𝑥̇ , 𝑦̇ , 𝑧̇ (From equation 58 and 59)
Xe x,y,z
V_b 𝑉⃗⃗𝑏
W_b 𝜔
⃗⃗𝑏
Alpha 𝛼
A port to tap into the current
moment generated by wing 1 to be
Out 1
forwarded to Dynamic inverse
controller
A port to tap into the current
moment generated by wing 2 to be
Out 2
forwarded to Dynamic inverse
controller

̇ And 𝐼𝑤2
𝐼𝑤1 ̇ are calculated because the flapping wing changes the values due to the flapping motion. A
generalized equation was developed based on the one derived by Rashid2 in appendix E. This equation was then coded
as an S-function in MATLAB.
The model also consists of an S-Function in the end to calculate the linear and angular accelerations. The output
from this S-function is then further integrated to get the linear velocity and angular velocity in body-axis.
These velocities are then used for further processing to solve for Euler angles. These angles are further utilized to
solve for trajectory in the inertial frame of reference. All of these processing utilizes the standard aircraft equations
from Roskam43.

12
American Institute of Aeronautics and Astronautics
E. Aircraft-Inputs Flight Dynamics
There are two more main sets of inputs that are needed for successful execution of the overall SIMULINK model.
The first one is the overall global variables like Mass of various components, density value, etc. Secondly, the inputs
required to initiate the Flight dynamics model.
Table 9. Global Inputs
Variable Value
Mass_body 0.75 lbs
Mass_wing 1 0.0912 lbs
Mass_wing 2 0.0912 lbs
Inertia_wing 1 [0.0217147 0 0;
0 0.0100256 0;
0 0 0.03173936]
Inertia wing 2 [0.0217147 0 0;
0 0.0100256 0;
0 0 0.03173936]
Inertia_body [0.0026713267 0 0;
0 0.09015876 0;
0 0 0.08873731];

Table 10. Miscellaneous Inputs


Variable Value (ft)
⃗⃗𝑤1
𝐷 [0 -0.1 0]’
⃗⃗𝑤2
𝐷 [0 0.1 0]’
𝑑⃗𝑤1 [0 0 0]’
𝑑⃗𝑤2 [0 0 0]’

VI. Controller Design

A. Controller Architectures
Three kinds of adaptive control architectures are utilized here; Adaptive Bias Controller (ABC) with PI Controller,
ABC with PD Controller and OCM-Linear model. Each of these architectures differ either in their linear controller or
the adaptive architecture.

Figure 15. General Overview of Controller architecture

13
American Institute of Aeronautics and Astronautics
1. Model Follower
The first component of the controller is the model follower shown in figure (16). This model depicts the desired
dynamics of the system. It is adapted from the work by Nguyen44

Figure 16. Model Follower

The reference model is a first-order system calculating 𝜔𝑚 for the input vector of commanded aircraft rotational
rates, 𝜔𝑐 and is given by Nguyen44
𝜔̇ 𝑚 + 𝜔𝑛 𝜔𝑚 = 𝜔𝑛 𝜔𝑐 (12)
Where 𝜔𝑛 is the matrix of natural frequencies of each axis along the diagonal. For longitudinal controller, equation
(12) is applied only to the pitch rate command,𝑞𝑚 , which gives us the following equation

𝑞̇ 𝑚 + 𝜔𝑛 𝑞𝑚 = 𝜔𝑛 𝑞𝑐𝑜𝑚 (13)
1
Laplace Transforming, we can see that the above equation is a first-order system with a time constant of . To
𝜔𝑛
design this controller the preferred parameter is the rise time of the system,𝑇𝑟 as
2.2 (14)
𝜔𝑛 =
𝑇𝑟 ,
In order to avoid transmitting a noisy error signal, the derivative portion is taken directly from the model as 𝑞̇ 𝑚 .
This is then added to the output of PI controller and the adaptive signal to form 𝑞̇ 𝑐𝑜𝑚 .
The rise time for the PI controller was chosen to be 0.5 secs considering the type of aircraft and the requirement
to catch up to commanded signal quickly for agility. Hence, the model follower’s rise time was fixed at 0.275 secs so
that it catches up with command sooner than the controller.
2. Inverse Controller
The inverse controller used in this control architecture was derived by inverting equation 11 to solve for t_pi
(input). The input to the controller is the commanded pitch acceleration 𝑞̇𝑐𝑜𝑚 . The inverse controller generate the
required control values of the tail setting angle, t_pi to achieve the desired accelerations. Inversion of equation 11
leads to the equation below, after substituting the values of known variables:

𝑞̇ = 0.0016812 𝑉 2 [−0.3486 − 3.3182 𝛼 + 0.0975 𝛽 − 0.4184 𝛽 2 + 0.3053 𝛼𝑉] + 𝑀𝑤1 + 𝑀𝑤2 (15)
− 𝑝𝑟(𝐼𝑥𝑥𝑏 − 𝐼𝑧𝑧𝑏 ) − 𝑟(𝑝 − 𝜔𝑐1 )(𝐼𝑥𝑥𝑤1 − 𝐼𝑧𝑧𝑤1 ) − 𝑟(𝑝 − 𝜔𝑐2 )(𝐼𝑥𝑥𝑤2 − 𝐼𝑧𝑧𝑤2 )
̇
− 𝑞(𝐼𝑦𝑦 ̇
+ 𝐼𝑦𝑦 )
𝑤1 𝑤2

3. Adaptive Bias Correction


The adaption used here is known as the Adaptive Bias Controller. It is the key component in the Model Reference
Adaptive Controller. The adaptive element adjusts the inputs to the inverse controller by accounting for the modelling
error between aircraft and the inverse controller, which is a very important requirement since the dynamics of an
ornithopter is still uncertain.
The Adaptive Bias Controller was developed at Wichita State University as a simple adaptive element. The ABC
adaptation uses a simple bias neural network element that parametrizes the modelling or tracking-error to form the
corrective signal of 𝑞̇ 𝑎𝑑𝑑 such that

14
American Institute of Aeronautics and Astronautics
𝑞̇ 𝑎𝑑𝑑 𝑛𝑒𝑤 = 𝑞̇ 𝑎𝑑𝑑 𝑜𝑙𝑑 + 𝜂𝑒(𝑡) (16)
Where, 𝜂 is the learning rate and 𝑒(𝑡) is the tracking error.

4. Proportional-Integral (PI) Controller


The fourth component of ABC-PI controller is the PI controller. This was also taken from work by Nguyen44. In a
linear analysis assuming the inverse controller and the aircraft perfectly cancel each other to become an integrator,
using a gain other than unity for 𝐾𝑝 will result in 𝑞 not exactly following, 𝑞𝑚 . The PI controller in the frequency
domain is represented as
𝐾 (17)
𝑞̇ = ( 𝑖 + 𝐾 )𝑞
𝑑𝑒𝑠 𝑝 𝑒
𝑠
Where, 𝑞𝑒 = 𝑞𝑚 − 𝑞.
Nguyen44 recommend relating the gains to the desired dynamics of the system specifying a damping ratio,𝜁, and
natural frequency, 𝜔𝑛 , in the case of the longitudinal controller

𝐾𝑝 = 2𝜁𝜔𝑛 (18)
𝐾𝑖 = 𝜔𝑛 2

The rise time for the system was chosen to be 0.5 secs for the reason as mentioned in model follower, with a
damping ratio of 0.9.
5. Proportional-Derivative (PD) Controller
One of the changes in the architecture that has been studied here is replacing the PI controller with a PD Controller
in fig (17) . The PD controller in time domain is presented as follows
𝑞̇ 𝑑𝑒𝑠 = (𝐾𝑑 𝑠 + 𝐾𝑝 )𝑞𝑒 (19)

Figure 17. PD Controller

6. Optimal Control Modification


Optimal control modification was introduced by Nguyen45 to achieve fast adaptation for MRAC. Reed46 utilized
this concept to develop a OCM architecture and compared its performance with other architectures. The OCM derives
its name from the optimal control techniques used for weight updates.
The variation used here is using the adaptive parameterization given below;

𝑞̇ 𝑎𝑑𝑑 = 𝛩𝑇 ∅ = 𝛩𝑇 [𝑞 𝜃 𝛼]𝑇 (20)

VII. Major model differences

A. Flight Dynamics Model


During early design stages it was found that the tail size and its aerodynamic co-efficients caused it to oscillate
with very large amplitude during trim conditions. The tail oscillates in this model in order to counteract the pitching
moment generated by the model, which is a result of the wing flapping. Hence, in order to reduce these oscillations,
thus reducing the actuator effort, the tail was assumed to be bigger than the original design and placed such that the
effectiveness of the tail is scaled up by a factor of five.
Also, the trim value of the tail setting angle was found to be beyond normal operational limits. Hence, in order to
bring this value within acceptable, practical levels the 𝐶𝑚𝑜 of the tail was scaled up by a factor of three.

15
American Institute of Aeronautics and Astronautics
B. Lateral Control Loop
During initial trials it was observed that the flapping motions of the wings made the aircraft drift slightly sideways,
a small yawing and rolling moment, which was seen to affect the longitudinal variables. Since lateral control was out
of scope of the current work, a Lat-Dir PI controller was chosen and was tuned automatically using
MATLAB/Simulink.

VIII. Results

The results presented are the time response tracking of the final design gains for the controller.
The time response of the controller to a pitch rate doublet will be shown for the pitch rate, angle of attack, airspeed,
aircraft pitch angle along with the control input, and tail setting angle. The results compare the performance of the
system with and without the adaptation.
Stepanyan47 introduced the concept of error metrics which can be used to verify and validate various adaptive
systems. Reed46 utilized these methods to tune the gains for the ABC and OCM controllers. We will be utilizing the
transient performance metric , described in Section II.B(1) by stepanyan 47. This metric will be used to measure the
oscillations of the adaptive system with respect to the model follower and will serve as the method to compare
performance of different architectures in different conditions. The metric is the integration of the second-norm of the
tracking error for a pitch doublet from t = 20 secs to 35 secs
‖𝑞𝑚 (𝑡) − 𝑞(𝑡)‖𝐿2 (21)
𝑀=
‖𝑞𝑚 (𝑡)‖𝐿2

This chapter has four sections. The first section presents the results of the PI control architecture described in the
previous sections for A-Matrix errors in the aircraft model. The second section presents the results of the control
architecture with a PD element in place of the PI controller along with the same modelling errors as introduced for the
PI controller. The third section consists of the response from OCM modification. The fourth section presents the result
of performance of the above mentioned architectures in case of the tail actuator failure, a B-Matrix error, which is
assumed to reduce the effectiveness of the tail by half.

A. PI Controller
The gains of this controller rare given in table 10. With no delay errors the time response of the controller to the pitch
Table 11. PI Controller Gains rate doublet is shown in figure 18 and 19. This controller was able to
track the commanded value of pitch rate. The oscillations observed in
Variable Value the figures are due to the flapping of the wings. The values plotted here
𝐾𝑝 10.38 are the forward velocity u, angle of attack α, pitch angle θ and q.
𝐾𝑖 33.25 The second case presents the performance of the PI controller when
𝜂 0.02 an error is introduced in the model by reducing
the 𝐶𝑚𝛼 to zero in the aircraft model, but Inverse Controller uses the
original value. The controller was unable to exactly follow the commanded pitch rate as we can see in figure 20. The
ANN was introduced again and it can be observed that the ANN adapted to the modelling error and reduced the
tracking error as observed from the 20-25 sec mark in the q plot of fig 21

16
American Institute of Aeronautics and Astronautics
u alpha
15.5
0.05
15.4
forward Velocity (ft/sec)

15.3
0

alpha (rad)
15.2

15.1 -0.05

15

14.9 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Theta q
0.15
actual
0.1 model
0.15

0.05
Theta (rad)

q (rad/sec)

0.1 0

-0.05
0.05
-0.1

0
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 18. PI Controller-Original System

Input

0.8

0.6

0.4
Input (rad)

0.2

-0.2

-0.4
10 15 20 25 30
Time (secs)

Figure 19. Input

17
American Institute of Aeronautics and Astronautics
u alpha

0.05

forward Velocity (ft/sec)


15.4

alpha (rad)
0
15.2
-0.05

15
-0.1
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.15 actual
0.1
model
0.05
Theta (rad)

q (rad/sec)
0.1
0

-0.05
0.05
-0.1

0
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 20. PI Controller +Zero Cm-alpha

u alpha
15.5
0.05
forward Velocity (ft/sec)

15.4

15.3
alpha (rad)

0
15.2

15.1 -0.05
15

14.9 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Theta q
0.15
actual
0.1
0.15 model
0.05
Theta (rad)

q (rad/sec)

0.1
0

-0.05
0.05
-0.1

0
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 21. PI Controller +Zero Cm-alpha+ ANN

18
American Institute of Aeronautics and Astronautics
The third case is to show the performance of the PI controller when an error is introduced in the model by having
a value of 𝐶𝑚𝛼 in the aircraft model that is unstable with a magnitude of 100% of its stable value, but Inverse Controller
used the original value. Consistent with the previous results the system is not able to track the commanded q from 20-
25 sec mark. The ANN was able to adapt to this modelling error and the results are presented in fig 23. Also, consistent
with its previous trend seen above, ANN enabled the system to track the commanded q and the it also improved upon
its error at the 22-25 sec mark.
u alpha
0.05
forward Velocity (ft/sec)

15.4

alpha (rad)
15.2
-0.05

15
-0.1
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.15 actual
0.1
model
0.05
Theta (rad)

q (rad/sec)
0.1
0

0.05 -0.05

-0.1

0
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 22. PI Controller + case 2 unstable value of Cm-alpha

u alpha
forward Velocity (ft/sec)

15.4 0.05
alpha (rad)

0
15.2

-0.05
15
-0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
actual
0.15 0.1
model
Theta (rad)

0.05
q (rad/sec)

0.1 0

-0.05
0.05
-0.1
0
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 23. PI Controller + case 2 unstable value of Cm-alpha+ANN

19
American Institute of Aeronautics and Astronautics
B. PD Controller
The gains of the controller rare given in table 11. With no delay errors the time response of the controller to the pitch
rate doublet is shown in figure 24 and 25. This controller was able
Table 12. PD Controller Gains to track the commanded value of pitch rate. The oscillations
Variable Value observed in the figures are due to the flapping of the wings. The
𝐾𝑝 10 values plotted here are the forward velocity u, angle of attack α,
𝐾𝐷 1.27 pitch angle θ and the pitch rate q.
𝜂 0.02

u alpha
15.5
forward Velocity (ft/sec)

0.05
15.4

alpha (rad)
15.3 0

15.2
-0.05
15.1

15 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.15 actual
0.1
model
Theta (rad)

0.05
q (rad/sec)

0.1
0
0.05
-0.05

0 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 24. PD Controller-Original System

Input

0.8

0.6

0.4
Input (rad)

0.2

-0.2

10 15 20 25 30
Time (secs)

Figure 25. PD Controller Input


The second case is to show the performance of the PD controller when an error is introduced in the model by
reducing the 𝐶𝑚𝛼 to zero in the aircraft model, but the Inverse Controller uses the original value. The controller was
again unable to exactly follow the commanded pitch rate as we can see in figure 26. The ANN was introduced again
and it can be observed that the ANN adapted to the modelling error and reduced the tracking error as observed from
the 20-25 sec mark in the q plot of fig 27. The system responds like a first order system but it can be observed at the

20
American Institute of Aeronautics and Astronautics
22 sec mark and 25 sec mark that the system’s rise time is longer than the first case leading to an error from the pitch
rate model.

u alpha

forward Velocity (ft/sec)


15.5
0.05
15.4

alpha (rad)
15.3 0

15.2 -0.05

15.1
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.1 actual
0.1
model
Theta (rad)

q (rad/sec)
0.05
0.05
0
-0.05
0
-0.1
-0.05
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 26. PD Controller +Zero Cm-alpha

u alpha
0.05
forward Velocity (ft/sec)

15.4

15.3
0
alpha (rad)

15.2

15.1 -0.05
15

14.9 -0.1
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.15 actual
0.1
model
0.05
Theta (rad)

q (rad/sec)

0.1
0

-0.05
0.05
-0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 27. PD Controller +Zero Cm-alpha+ ANN

21
American Institute of Aeronautics and Astronautics
The third case is to show the performance of the PD controller when an error is introduced in the model by having
a value of 𝐶𝑚𝛼 in the aircraft model that is unstable with a magnitude of 100% of its stable value, but the Inverse
Controller uses the original version. Consistent with the previous results the system is not able to track the commanded
q from 20-25 sec mark. The ANN was able to adapt to this modelling error and the results are presented in fig 29.
Also, consistent with its previous trend seen above, ANN enabled the system to track the commanded q but the rise
time is seen to have increased.

forward Velocity (ft/sec) u alpha

15.5
0.05

alpha (rad)
15.4
0
15.3

15.2 -0.05

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.1
actual
0.1
model
0.05
Theta (rad)

q (rad/sec)

0.05
0
0 -0.05

-0.1
-0.05
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 28. PD Controller + case 2 unstable value of Cm-alpha

u alpha
0.05
forward Velocity (ft/sec)

15.4

15.3 0
alpha (rad)

15.2

15.1 -0.05
15

14.9 -0.1
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.15 actual
0.1
model
0.05
Theta (rad)

q (rad/sec)

0.1
0

0.05 -0.05

-0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 29. PD Controller + case 2 unstable value of Cm-alpha+ANN

22
American Institute of Aeronautics and Astronautics
C. OCM Architecture
The OCM architecture that was developed for a fixed wing aircraft was adopted to the present work and its
response to various errors were studied by varying its learning rates. A time delay error metric study was also carried
out to identify trends in robustness of the controller with a particular set of learning rates. Three learning rates are used
here 250, 10000 and 50000. Each of these controllers were coupled with original system and their responses were
recorded. Subsequently, modelling error in the form of unstable 𝐶𝑚𝛼 , similar to third case of the PI and PD controllers
above were introduced and the responses were recorded and presented here. Finally, the reduced tail effectiveness
scenario was also introduced and the response is of each iteration of the controller is presented in the following section.
Figures (30) and (31) present the response of the controller with learning rate 250. It can be observed that the OCM
has no discernible effects on the system response as compared to a PI controller response in figures (18) and (22).

Figure 30. OCM-250 + Original system

Figure 31. OCM-250 + case 2 unstable value of Cm-alpha

23
American Institute of Aeronautics and Astronautics
Figures (32) and (33) present the response of the controller with learning rate 10000. It can be observed that the
OCM has no discernible effects on the system response as compared to a PI controller (figure (18)) response for the
original system with no uncertainty but with modelling error, the OCM performs well as compared to PI controller
(figure(22)). Another feature to notice is that OCM adaptation dampens out the system oscillations, observed in the q
plot of fig 32, in system response while the ABC controller in figure 23 could not.

Figure 32. OCM-10000 + Original system

Figure 33. OCM-10000 + case 2 unstable value of Cm-alpha

24
American Institute of Aeronautics and Astronautics
Figures (34) and (35) present the response of the controller with learning rate 50000. It can be observed that this
iteration of the controller is has a similar response as the original system with no uncertainties but has much better
improved tracking performance for the modelling error case. The improved damping of the system response is also
observed here.

Figure 34. OCM-50000 + Original system

Figure 35. OCM-50000 + case 2 unstable value of Cm-alpha

25
American Institute of Aeronautics and Astronautics
D. Reduced Tail Effectiveness
A B-matrix error that was introduced in the system is where the tail has lost half its effectiveness and the response
of various controller architecture to this scenario is recorded below.
Figure (36) present the responses of the PI controller without adaptation. It can be observed from the q plot that
the oscillations are of higher magnitude in this kind of failure due to the role tail effectiveness plays in counteracting
the moments generated by the wing and the controller is not able to reduce the error arising because of the tail flapping.
This same error, modeling using a PD controller, made the closed system unstable. Hence, no plot is shown.
u alpha

15.3
forward Velocity (ft/sec)

0.05
15.2

alpha (rad)
15.1 0
15
14.9 -0.05
14.8
-0.1
14.7

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.3
actual
0.2 0.2
model
0.1
Theta (rad)

q (rad/sec)
0.15
0
0.1
-0.1

0.05 -0.2

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 36. PI Controller + Reduced Tail effectiveness

Figures (37) and (38) present the responses of the system with ABC adaptation for the PI and PD controller
architectures respectively. In both the cases the ANN was able to successfully adapt to the failure but could not damp
the oscillations in the system.
u alpha
15.6
0.05
forward Velocity (ft/sec)

15.5
alpha (rad)

15.4 0

15.3
-0.05
15.2

15.1 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.3
actual
0.2
0.15 model
0.1
Theta (rad)

q (rad/sec)

0.1 0

-0.1
0.05
-0.2
0
10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 37. PI Controller + ANN

26
American Institute of Aeronautics and Astronautics
u alpha

15.3

forward Velocity (ft/sec)


0.05
15.2
15.1

alpha (rad)
0
15
14.9 -0.05
14.8
14.7 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)
Theta q
0.15
0.2 actual
0.1
model
0.15 0.05
Theta (rad)

q (rad/sec)
0
0.1
-0.05

0.05 -0.1

10 15 20 25 30 10 15 20 25 30
Time (secs) Time (secs)

Figure 38. PD Controller + ANN

The same error was introduced with the OCM adaptations. Only the controller iterations with learning rates 10000
and 50000 are presented here. From figures (39) and (40), we can observe that by increasing the learning rate we can
reduce the oscillations in system response but the tracking error becomes larger.

Figure 39. OCM-10000

27
American Institute of Aeronautics and Astronautics
Figure 40. OCM-50000

E. Time Delay Error Metric Study

In order to identify more suitable learning rates


and gains in OCM architecture, a time delay margin
study using a tracking error metric was done. The
learning rates considered here are same as
considered in above sections. The graphs of the
learning rates 50000 and 10000 have been
terminated at the final value of time delay for which
the system was stable, i.e, the error metric value
becomes infinity after those points.
It can be observed that lower learning rate has a
higher time delay than the higher learning rates.
Learning rate of 250 , though giving a higher time
delay margin could not provide improvement in
Figure 41. error metric study performance of the PI controller while the learning
rate 50000, though has a lower time delay margin
provides a good performance improvement to PI controller in form of damping of system response while also
providing a good tracking performance. The learning rate of 10000 lies in-between with a higher time delay than
50000 and acceptable performance.
It has to be noted that the for this study the PI controller gains were fixed at the values chosen before, hence in
order to increase these time delay margins we can vary those gains to have an acceptable tracking performance
combined with higher time delays.

IX. Conclusion

This paper presented a general model of an ornithopter combining the unsteady aerodynamics with consideration
of wing shape and multi-body flight dynamics model. This work further develops a Model Reference Adaptive
Controller for pitch rate command using the Adaptive Bias configuration as the adaptive element. This controller was
subject to various failures in the form of A matrix error and reduced tail effectiveness. The controller architecture
consisted of a Model follower, PI/PD controller, Dynamic Inverse controller that tracks pitch, velocity accelerations

28
American Institute of Aeronautics and Astronautics
and an adaptive element. These, in combination, generated the necessary tail setting angle commands. These
commands were the inputs to the nonlinear aircraft simulation of a model similar to “Slow Hawk” ornithopter.
Each design involved tuning the PI/PD controller gains along with the learning co-efficient of the ABC
controllers,𝜂.
In addition, this work also adopted the OCM architecture originally developed for fixed wing aircraft and studied
the system response to failures. Subsequently, a time delay error metric study was performed for various versions of
this controller in order to identify optimal learning rates.
Each of the architectures were presented with a commanded pitch rate doublet. Time responses for pitch rate, angle
of attack, pitch angle, forward velocity, tail setting angle were shown for each of the test points. The following are the
conclusions for the controller’s performance in flapping wing aircraft.

A. PI Controller with Adaptive Bias Correction


The PI Controller architecture was able to adapt to the modelling error and tracked the commanded pitch rate
doublet in absence of any kind of failures but in presence of failures, the PI controller tracked the command with a
steady state error. The adaptive element was able to correct this and successfully tracked the commanded doublet in
the presence of modelling error and tail effectiveness error.
One of the observations was that the failure modes did not create any instability in the system and the PI controller
was able to correct the error.

B. PD Controller with Adaptive Bias Correction


The PD Controller architecture was able to track the commanded pitch rate doublet in absence of any kind of
failures but in the presence of failures, the PD controller was not good at tracking the command and in the special case
of reduced tail effectiveness it was not able to stabilize the system. The adaptive element was able to correct this and
successfully tracked the commanded doublet in presence of modelling error and tail effectiveness error, it made the
system’s response to be similar to a first order system but increasing 𝐶𝑚𝛼 error led to increase in the rise time of the
adaptive system.

C. OCM Adaptaion
The OCM controller architecture was also studied as part of this work and it was observed that this controller
requires higher learning rates in order to perform well. The advantage of this controller was seen to be its ability to
dampen out the oscillations in system response. It was observed that with very low learning rates the system mimicked
the original PI system and with higher learning rates it showed an improved response. With higher learning rates it
was observed that the tracking ability of the system was degraded and the time delay that the controller can accept
also degraded. Hence, in order to optimize this controller for this aircraft the proportional and integral gains should
also be varied so that the entire system with PI and OCM adaptation together gives us the required tracking
performance while increasing the time delay.
References
1
DeLaurier, J. D. "An aerodynamic model for flapping-wing flight," The Aeronautical Journal Vol. 97, No. 964,
1993, pp. 125-130.
2
Rashid, T. "The flight dynamics of a full-scale ornithopter." National Library of Canada= Bibliothèque
nationale du Canada, 1995.
3
Zakaria, M., Elshabka, A., Bayoumy, A., and Abd Elhamid, O. "Numerical aerodynamic characteristics of
flapping wings," 13th International Conference on Aerospace Sciences & Aviation Technology, ASAT-13, May. Vol.
26, 2009, p. 28.
4
https://www.mathworks.com, 2017.
5
Lippisch, A. M. "Man Powered Flight in 1929," Journal of Royal Aeronautical Society Vol. 64, No. 595, July
1960, pp. 395-398.
6
DeLaurier, J. D. "The Development and Testing of a Full-Scale Piloted Ornithopter," Canadian Aeronautics
and Space Journal Vol. 45, No. 2, June 1999.
7
De Croon, G., De Clercq, K., Ruijsink, R., Remes, B., and de Wagter, C. "Design, aerodynamics, and vision-
based control of the DelFly," International Journal of Micro Air Vehicles Vol. 1, No. 2, 2009, pp. 71-97.
8
Keennon, M., Klingebiel, K., Won, H., and Andriukov, A. "Development of the nano hummingbird: A tailless
flapping wing micro air vehicle," AIAA aerospace sciences meeting. AIAA Reston, VA, 2012, pp. 1-24.
9
Kinkade, A. "Ornithopter." Google Patents, 2001.

29
American Institute of Aeronautics and Astronautics
10
Billingsley, D., Slipher, G., Grauer, J., and Hubbard, J. "Testing of a passively morphing ornithopter wing,"
AIAA Paper Vol. 1828, 2009.
11
Chen, B.-H., Chen, L.-S., Lu, Y., Wang, Z.-J., and Lin, P.-C. "Design of a Butterfly Ornithopter," Journal of
Minjiang Science and Technology Vol. 19, No. 1, 2016, pp. 7-16.
12
Brown, R. H. J. "THE FLIGHT OF BIRDS," Biological Reviews Vol. 38, No. 4, 1963, pp. 460-489.
doi: 10.1111/j.1469-185X.1963.tb00790.x
13
Chatterjee, S., Templin, R. J., and Campbell, K. E., Jr. "The aerodynamics of Argentavis, the world's largest
flying bird from the Miocene of Argentina," Proc Natl Acad Sci U S A Vol. 104, No. 30, 2007, pp. 12398-403.
14
Dial, K. P., Jackson, B. E., and Segre, P. "A fundamental avian wing-stroke provides a new perspective on the
evolution of flight," Nature Vol. 451, No. 7181, 2008, pp. 985-989.
15
Hedenström, A. "Aerodynamics, evolution and ecology of avian flight," Trends in Ecology & Evolution Vol.
17, No. 9, 2002, pp. 415-422.
16
Pennycuick, C. J. "Power Requirements for Horizontal Flight in the Pigeon <em>Columba
Livia</em&gt," Journal of Experimental Biology Vol. 49, No. 3, 1968, p. 527.
17
Tobalske, B. W. "Biomechanics of bird flight," Journal of Experimental Biology Vol. 210, No. 18, 2007, p.
3135.
18
Shigeoka, K. S. "Velocity and altitude control of an ornithopter micro aerial vehicle." 2007.
19
Julian, R. C., Rose, C. J., Hu, H., and Fearing, R. S. "Cooperative control and modeling for narrow passage
traversal with an ornithopter MAV and lightweight ground station," Proceedings of the 2013 international
conference on Autonomous agents and multi-agent systems. International Foundation for Autonomous Agents and
Multiagent Systems, 2013, pp. 103-110.
20
Jackowski, Z. J. "Design and construction of an autonomous ornithopter." Massachusetts Institute of
Technology, 2009.
21
Krashanitsa, R. Y., Silin, D., Shkarayev, S. V., and Abate, G. "Flight dynamics of a flapping-wing air vehicle,"
International Journal of Micro Air Vehicles Vol. 1, No. 1, 2009, pp. 35-49.
22
Narendra, K., and Annaswamy, A. "A new adaptive law for robust adaptation without persistent excitation,"
IEEE Transactions on Automatic control Vol. 32, No. 2, 1987, pp. 134-145.
23
Lewis, F. L., Yesildirek, A., and Liu, K. "Multilayer neural-net robot controller with guaranteed tracking
performance," IEEE Transactions on Neural Networks Vol. 7, No. 2, 1996, pp. 388-399.
24
Rysdyk, R. T., and Calise, A. J. "Adaptive model inversion flight control for tilt-rotor aircraft," Journal of
Guidance Control and Dynamics Vol. 22, 1999, pp. 402-407.
25
Bosworth, J. T., and Williams-Hayes, P. S. "Flight Test Results from the NF-15B Intelligent Flight Control
System (IFCS) Project with Adaptation to a Simulated Stabilator Failure," 2007.
26
Nguyen, N., Krishnakumar, K., Kaneshige, J., and Nespeca, P. "Dynamics and Adaptive Control for Stability
Recovery of Damaged Aircraft," 2006.
27
Duerksen, N. "Fuzzy logic adaptive decoupled flight controls for General Aviation airplanes," 1996.
28
Steck, J. E., Rokhsaz, K., and Shue, S.-P. "Linear and neural network feedback for flight control decoupling,"
IEEE Control Systems Vol. 16, No. 4, 1996, pp. 22-30.
29
Pesonen, U. J., Steck, J. E., Rokhsaz, K., Bruner, H. S., and Duerksen, N. "Adaptive neural network inverse
controller for general aviation safety," Journal of Guidance, Control, and Dynamics Vol. 27, No. 3, 2004, pp. 434-
443.
30
Steck, J., Rokhsaz, K., Namuduri, K., and Bruner, S. "Exploring Critical Flight Conditions, Controller Modes,
and Parameter Estimation for Adaptive Flight Controls in General Aviation Aircraft," FAA Final Report, 2006.
31
Han, J.-H., Lee, J.-Y., and Kim, D.-K. "Ornithopter modeling for flight simulation," Control, Automation and
Systems, 2008. ICCAS 2008. International Conference on. IEEE, 2008, pp. 1773-1777.
32
Grauer, J. A. Modeling and system identification of an ornithopter flight dynamics model: University of
Maryland, College Park, 2012.
33
Grauer, J., Ulrich, E., Hubbard Jr, J., Pines, D., and Humbert, J. S. "Testing and system identification of an
ornithopter in longitudinal flight," Journal of Aircraft Vol. 48, No. 2, 2011, p. 660.
34
Rose, C., and Fearing, R. S. "Flight simulation of an ornithopter." Master’s thesis, EECS Department,
University of California, Berkeley, May 2013.[Online]. Available: http://www. eecs. berkeley.
edu/Pubs/TechRpts/2013/EECS-2013-60. html, 2013.
35
Kaviyarasu, A., and Kumar, K. S. "Simulation of Flapping-wing Unmanned Aerial Vehicle using X-plane and
Matlab/Simulink," Defence Science Journal Vol. 64, No. 4, 2014, p. 327.

30
American Institute of Aeronautics and Astronautics
36
Kim, D.-K., Lee, J.-S., Lee, J.-Y., and Han, J.-H. "An aeroelastic analysis of a flexible flapping wing using
modified strip theory," Proceedings of SPIE 15th annual symposium smart structures and materials. Vol. 6928,
2008, p. 69281O.
37
Chalia, S., and Bharti, M. K. "A Review on Aerodynamics of Flapping Wings," 2016.
38
BENG, T. W. "Dynamics and control of a flapping wing aircraft." 2004.
39
Malik, M. A., and Ahmad, F. "Effect of different design parameters on lift, thrust, and drag of an ornithopter,"
Proceedings of the World Congress on Engineering. Vol. 2, 2010, pp. 1460-1465.
40
Hoerner, S. F. Fluid-dynamic drag: practical information on aerodynamic drag and hydrodynamic resistance:
Hoerner Fluid Dynamics, 1965.
41
Chandrasekaran, B. K. "Design of an Adaptive flight controller for a bird-like flapping wing aircraft,"
Aerospace Engineering. Vol. Master of Science, Wichita State University, Wichita, 2017.
42
Kinkade, S. "Original Radio Controlled Ornithopter Slow Hawk 2 Instruction Manual." HobbyTechnik.
43
Roskam, J. Airplane flight dynamics and automatic flight controls: DARcorporation, 1995.
44
Nguyen, N., Krishnakumar, K., Kaneshige, J., and Nespeca, P. "Flight dynamics and hybrid adaptive control of
damaged aircraft," Journal of guidance, control, and dynamics Vol. 31, No. 3, 2008, p. 751.
45
Nguyen, N., Krishnakumar, K., and Boskovic, J. "An Optimal Control Modification to Model-Reference
Adaptive Control for Fast Adaptation," AIAA Guidance, Navigation and Control Conference and Exhibit. American
Institute of Aeronautics and Astronautics, 2008.
46
Reed, S. "Demonstration of the optimal control modification for general aviation: design and simulation,"
Aerospace Engineering. Vol. Master of Science, Wichita State University, Wichita, 2010.
47
Stepanyan, V., Krishnakumar, K., Nguyen, N., and Van Eykeren, L. "Stability and Performance Metrics for
Adaptive Flight Control," AIAA Guidance, Navigation, and Control Conference. American Institute of Aeronautics
and Astronautics, 2009.

31
American Institute of Aeronautics and Astronautics

You might also like