You are on page 1of 21

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225020564

Open-loop reference models for nonlinear control with applications to


unmanned helicopter flight

Conference Paper · August 2010


DOI: 10.2514/6.2010-7860 · Source: DLR

CITATIONS READS

5 94

1 author:

Sven Lorenz
German Aerospace Center (DLR)
32 PUBLICATIONS   139 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Autonomous Research Testbed for Intelligent Systems (ARTIS) View project

SHEFEX View project

All content following this page was uploaded by Sven Lorenz on 30 May 2014.

The user has requested enhancement of the downloaded file.


Open-loop reference models for nonlinear control with
applications to unmanned helicopter flight

Sven Lorenz∗
German Aerospace Center (DLR), Institute of Flight Systems, Braunschweig, Germany

A novel way to design a nonlinear feedforward controller is introduced. Contrary to


input-output linearization and pseudo control hedging techniques, the new formulation
proposed is based on the states of the reference system instead on the states of the plant. By
reshaping the structure commonly used in literature the similarities between the resulting
scheme and what is made in model following control are utilized to get a pure nonlinear
feedforward system which is completely decoupled from system outputs. This permits the
application of the nonlinear control law when knowledge of the system states is not available.
Due to the strict separation of plant and reference dynamics in the proposed approach
a stability analysis of the reference system can be simplified. This new control scheme
has been used for the control of an unmanned helicopter demonstrating its applicability.
Results of these flight-test are presented in this paper.

Nomenclature

∗b index “b” stands for body fixed frame


∗c index “c” stands for a command
∗g index “g” stands for geodetic fixed frame
Δ(x, u) error term effected by the state and input vector
Δ̂A (xA , uc ) estimate of the derivation of the system input due to the actuator dynamics
F̂(. . .) represents (in figures) the systems dynamics
ˆ (symbol: hat) represents estimated values
A, B state space matrices of a linear model
AR reference system dynamics matrix
bold bold written symbols represent matrices
BR reference system input matrix
f (x(t)), g(x(t)) nonlinear functions of the state vector
K error feedback gain matrix
N(x) matrix of functions (contains Lie derivatives)
φ, θ, ψ Euler angles
uc input signal to the actuator
∗ a underline represents a vector
ω gb
b represents a rotation vector, expressed in body frame, describes rotation of body against
geodetic
e error vector
m(x) vector of functions (contains Lie derivatives)
u(t) input vector
v pseudo control vector
x(t) state vector
xA actuator state vector
y(t) output vector
y R (t) reference (command) output vector
∗ Research engineer at the Institute of Flight Systems, Department of Unmanned Vehicles, sven.lorenz@dlr.de.

1 of 20

American Institute of Aeronautics and Astronautics


yc desired output vector (input to the reference system)
ai positive constants (roots of the linear dynamics of the reference system)
bi constants of the reference model input matrix
ci positive constants (error feedback gain values)
eCR control error between reference model input and reference model output
eC overall control error (between reference model input and system output)
GA (x, uc ) represents the actuator dynamics
GR (. . .) represents (in figures) the reference system
Lf h Lie derivative of h with respect to f
m dimension of the output vector
n dimension of the state vector
r relative degree of a system
(r )
ym m the r-th derivation of a m-dimensional vector
vh pseudo hedging control signal
vDC error feedback control signal (from dynamic compensator)

I. Introduction
The availability of powerful and low-cost computers continues to increase the interest in nonlinear control
techniques. A great deal of theoretical investigation has been applied to the method of feedback linearization
used here to control an unmanned helicopter.
UAV missions in urban environments and their associated constraints (e.g. obstacle avoidance) gener-
ally require very good manoeuvrability, hover capabilities, and fast changes of the flight directions. These
requirements often lead to prefer helicopters to fixed-wing aircrafts. Furthermore the aggressive maneu-
vers resulting from high-speed flight in such environments force to look for controllers reaching very high
performances.
The equations of motion of an aircraft are nonlinear. Linear approximations of these dynamics, may
limit the flight regime in particular for aggressive maneuvers. An algebraic transformation called feedback
linearization can be used to transform a nonlinear system into a (fully or partly) linear one, as illustrated for
example by Slotine1 and Khalil.2 Therefore, after using feedback linearization to handle system’s nonlinear-
ities, linear control techniques can be used to tune the dynamics of the system without the aforementioned
restriction. In addition, the decoupling of the dynamics resulting from the transformation leads to simple
state or output feedback structures. Unfortunately, this method is very sensitive to unconsidered actuator
dynamics, delays, and parameter uncertainties.
A way to reduce the sensitivity of feedback linearization has been introduced by Johnson.3, 4 The combina-
tion of feedback linearization with Pseudo Control Hedging (PCH) deals with the challenge of compensating
unconsidered input dynamics when applying input-output linearization to a real system. In combination
with adaptive elements, PCH is used by Hovakimyan to provide stable adaptation during periods of control
saturation.5
As stated by Johnson the hedging signal does not influence the error dynamics. Moreover, the resulting
error signals are ultimately bounded, as proven by Kim.6 This approach makes use of the difference between
a commanded and an achieved control signal. The fact it results in a hidden feedback loop will be illustrated
in this paper. The loop itself will be analyzed and be reshaped in order to make appear more clearly its
internal functioning. As a consequence, this novel representation eases the stability analysis and provides
grasp concerning the influence of the hedging loop to the reference dynamics.
For classical helicopters the control input to the main rotor is affected by the rotor tilt angle and introduces
dynamics that are not present in rigid body approximations. Without the knowledge of the rotor states and,
therefore, by using only states of a rigid body system the achievable performance will be reduced. The
method of PCH was used by Johnson et al. to demonstrate aggressive flight maneuvers with unmanned
helicopters by utilizing a linear hover domain model. However, as the rotor flapping motion is not taken into
account in such a model the performance that can be achieved using this approach is lower than system’s
real physical limitations.
This paper presents the way the PCH can be clarified by reshaping and then transformed into a purely
nonlinear feedforward system. With the proposed approach the inversion of the rotor dynamics is made
without measuring or observing rotor states. The novel interpretation of the control system results in a

2 of 20

American Institute of Aeronautics and Astronautics


feedforward scheme which is based on the reference system’s states, the feedback linearization control law,
and the PCH in the modified form as announced before. The dynamical behavior of the nonlinear feedforward
law (based on FBL and PCH) can be analyzed independently from the choice of a feedback controller. This
feedback controller is required to reject disturbances, improve robustness, or stabilize the plant if necessary.
The strict separation between feedforward and feedback, makes the analysis of the overall system easier
because the imbrication between the PCH loop and the feedback loop has been removed.
Beside this theoretical discussion the concept has already been demonstrated in flight using an unmanned
helicopter. Implementation issues and flight test results will be discussed.
This paper starts with a compact review of the PCH control scheme including feedback linearization,
reference trajectory generation, error feedback compensation and pseudo control hedging in section II. The
reshaping presented in section III will clarify the influence of the pseudo control hedging system to the
reference dynamics. Existing similarities between the resulting system at hand and common feedforward –
feedback control systems are explained in section IV. Furthermore, a transformation of the PCH system
into a feedforward system can be achieved. Section V presents the implementation and flight test results for
an unmanned helicopter testbed followed by the summary and conclusion remarks in section VI.

II. Common Control Scheme based on Feedback Linearization


As many different terms were used for the method of feedback linearization a short historical summary
follows to point out the roots and progress of this method. According to Holzpafel7 a method called Computed
Torque was introduced in robotics around 1960. Based on the physical view of robot manipulators this
technique leads to a control law similar to the later developed fundamentals of feedback linearization.1, 2
Since then this method has become the most common technique of state transformation used in robotics
control. To the best of author’s knowledge, Singh and Schy have been the firsts to apply the method
of feedback linearization to a flying vehicle,8 under the name Dynamic Inversion. After Singh and Schy,
the technique has received increasing interest with a wealth of publications generated. Since 1993 the
synthesis of adaptive helicopter control by approximate feedback linearization has been investigated.5, 9–12
The NASA Johnson Space Center (JSC) published guidelines to address the flight control design for vehicles
with highly nonlinear physical descriptions of motions in 2002.13 Addressing the need for techniques to
transform a nonlinear system into an equivalent linear system, this document provides details of the research
on robustness and performance analyses for the X-38 reentry vehicle. The Intelligent Flight Control System
(IFCS) program from the NASA Dryden Flight Research Center also used a highly modified F-15B aircraft as
the testbed for adaptive control with the aim of compensating unexpected failures in flight.14 Outstanding
demonstrations applying this method for unmanned vehicles were performed for example by the Georgia
Institute of Technology.4–6, 11, 15–17
The basic idea behind feedback linearization is to transform a nonlinear system into a linear one by
applying an adequate feedback. The approach differs drastically from conventional linearization: feedback
linearization is achieved by an exact transformation using the system states, whereas linearization consists
in computing a local linear approximation around an operating point. In particular noticeable differences
would appear between these models in the two following cases: when the influence of nonlinear terms in the
dynamic of the system is high and when the deviations to the operating point are high enough to invalidate
the local approximation made.
Consider a MIMO nonlinear system in first order form:

ẋ(t) = f (x(t)) + g(x(t)) · u(t) (1)


y(t) = h(x(t)) (2)

where x ∈ n is the state vector of dimension n, f(x(t)) and g(x(t)) representing smooth vector fields, and
u(t) ∈ m is a vector of control inputs. The output vector y(t) ∈ m of this general system is a function
of the states as well. Note that the above system is linear or affine1 in the control input u. Without
further discussion on this hypothesis, we will assume hereafter that the considered systems are input-output
linearizable. For simplicity of notation the time dependence of each vector will be omitted.
We now consider the problem of designing a controller so that the output vector y follows a reference
trajectory y R . Usually the output equations are not straightforward invertible towards the system input due
to differential dependencies. The method of input-output linearization is based on the direct and algebraic

3 of 20

American Institute of Aeronautics and Astronautics


relation of time derivatives of the the outputs y, y1 , y2 , . . . with respect to the input vector u. In the next
section, we remind the reader, how this relation can be obtained and used to linearize the system by feedback.

A. Input-Output Linearization
To find an explicit relationship between the outputs and the inputs, one has to derivate successively the
output expressions with respect to time. When considering yi (the i-th output of the system that has to
be controlled), these successive derivations are stopped as soon as an algebraic relation between one of the
inputs and yi has been found. The number ri is then defined as the number of times is has been needed
to derive the expression of yi to obtain this algebraic relation. It is called the relative degree of the system
with respect to output yi . The scalar r = r1 + . . . + rm is called the total relative degree. Let n be the order
of the system. Notice that in the general case r ≤ n. In the case where r = n, there is no internal dynamics
and we obtain a input-state linearization of the original nonlinear system.
The differentiation leads to a system of equations represented by
⎡ ⎤
(r )
y1 1
⎢ . ⎥
⎢ . ⎥ = m(x) + N(x)u (3)
⎣ . ⎦
(rm )
ym

giving a relation between the input and the output. The derivation of the output Eq. (2) for each yi leads
to a vector of functions m(x) of the dimension m:
⎡ ⎤
Lrf1 h1
⎢ ⎥
⎢ Lrf2 h2 ⎥
m(x) = ⎢
⎢ ..

⎥ (4)
⎣ . ⎦
Lrfm hm

and a matrix N(x) of the dimension m × m:

⎡ ⎤
Lg1 Lrf1 −1 h1 (x) ··· Lgm Lrf1 −1 h1 (x)
⎢ ⎥
⎢ Lg1 Lrf2 −1 h2 (x) ··· Lgm Lrf2 −1 h2 (x) ⎥
N(x) = ⎢


⎥ (5)
⎣ ... ..
. ... ⎦
Lg1 Lfrm −1 hm (x) ··· Lgm Lfrm −1 hm (x)

where Lf h represents the Lie-Derivative of h with respect to f and Lg h the Lie-Derivative of h with respect
to g.2 Assuming the matrix N to have full rank for all possible x, the matrix is invertible and the inversion
of Eq. (3) becomes

u = N(x)−1 [v − m(x)] . (6)

The ri -th output derivative is with the control law from Eq. (6) linearly related to the control input vi :

yi (ri ) = vi . (7)

Plant states, which are not observable using the output measurements, are called unobservable. To fulfill
the control task (i. e. the system outputs follow the reference commands) it is required that these states are
not associated to any unstable internal dynamics. For the stability of the internal dynamics of a nonlinear
system the zero dynamics must be stable. The zero-dynamics are defined to be the internal dynamics of the
system when the system output is kept at zero by the inputs. A nonlinear system whose zero-dynamics are
asymptotically stable will be called as asymptotically minimum phase system.1, 2
For the system input vector v with pure integral influence on the output vector y a reference trajectory
has to be designed, which will be discussed in the following section.

4 of 20

American Institute of Aeronautics and Astronautics


B. Reference Model
The reference trajectory for the new control input is usually generated by a reference model. For simplicity
(r )
it is common to use linear reference models. Since the input vi only affects the output yi i the reference
 T
(r −1)
signal vector y R,i = yR,i , . . . , yR,ii can be calculated without regarding other outputs. Consequently,
the index “i” will be dropped hereafter. Therefore, we consider a single input multiple output system in the
following. For a system with a relative degree r a reference model written in state space form will contain r
first order equations to calculate the required control signal vR . By integration, all required feedback loop
(r−1)
reference trajectories yR to yR are calculated as used for the error feedback (following section).
The dynamics of the linear reference system in state space form is represented by

ẋR = AR · xR + BR · yc (8)
yR = xR (9)

with
⎡ ⎤
0 1 ... 0 ⎡ ⎤
⎢ .. .. .. ⎥ 0
⎢ . ⎥ ⎢ . ⎥
AR = ⎢

. . ⎥
⎥ (10) BR = ⎢ ⎥
⎣ .. ⎦ (11)
⎣ 0 0 ... 1 ⎦
b0
−a0 −a1 . . . −ar−1
where −a0 , −a1 , . . . , −ar−1 are positive constants (the roots of the linear dynamics) and b0 is a constant of
the input matrix. The input to the reference model will be the desired output of the system yc .

C. Feedback Compensator and Error Dynamics


A linear feedback loop should lead to an exponentially stable tracking error justified by external disturbances
as well as different initial conditions for the plant and the reference model. The tracking error vector is defined
as:
⎡ ⎤ ⎡ ⎤
e1 y1 − yR,1
⎢ . ⎥ ⎢ .. ⎥
e=⎢ ⎥ ⎢
⎣ .. ⎦ = ⎣ .
⎥.
⎦ (12)
em ym − yR,m

The ri -th derivative of the error ei is:


(ri ) (ri ) (r )
∀i ∈ {1, . . . , m}, ei = yi − yR,ii . (13)

Assuming a fully or exactly linearized system the new control signal vi must include the reference signal
(r )
yR,ii , as well as the error feedback:
(r )
∀i ∈ {1, . . . , m}, vi = yR,ii − c0,i · (yi − yR,i ) − c1,i · (y˙i − ẏR,i ) −


(r −1) (r −1)
. . . − cri −1 · yi i − yR,ii (14)
i −1
r
(r )
= yR,ii − cj,i · ei (j) (15)
j=0
= vR,i + vDC,i . (16)
(ri ) (r )
Finally, using ei = yi (ri ) − yR,ii leads to exponentially stable error dynamics:

∀i ∈ {1, . . . , m}, ei (ri ) + cri −1,i ei (ri −1) + . . . + c1,i e˙i + c0,i ei = 0 (17)

if the constants cj,i are positive. In the case of incomplete or approximately inversion an error term, Δ(x, u),
appears on the right hand side of Eq. (17). Furthermore, this additional error source can be faster compen-
sated by using a universal approximator as introduced by Johnson3, 4 or Kim6 for example.

5 of 20

American Institute of Aeronautics and Astronautics


D. Pseudo Control Hedging
Thus far it has been assumed that the control inputs operate directly on the plant inputs. If actuators
are present, their states should be considered during the inversion process. Including these dynamics in
the input-output linearization of the whole system would increase the relative degree r and measurements
or estimates of actuator states are required. Moreover, some effects, for instance hard-limitations, are not
invertible due their nondifferentiability. On the other hand, if the reference model and the inversion of
the plant dynamics were designed without considering the plant input dynamics or actuator saturations,
significant errors between reference states and plant states may occur.
The technique of Pseudo Control Hedging (PCH) was developed to account for these dynamics when
generating the reference and pseudo control signals. Here the vivid way Johnson in16 the PCH method
describes:
The reference model is moved in the opposite direction (hedged) by an estimate of the amount the
plant did not move due to system characteristics the control designer does not want the adaptive
control element to see.a
Figure 3 shows the scheme of the real plant in combination with the plant model as presented by
Holzapfel.7 With the actuator states xA assumed to be immeasurable the outputs u are estimated us-
ing an actuator model ĜA . The estimated plant inputs û given to the plant model F̂ (the same as used
for the inversion) give an estimate of the pseudo control vector v̂. The difference is then fed back into the
reference model GR to manipulate – or so called “hedge” – the reference states. This signal is fed into the
reference model as an additional component onto the r-th derivative; as shown in Fig. 2 for a linear reference
model. For simplification the PCH will be exemplified for a SISO system. The highest derivation of the
output reference calculated by the reference model becomes:
(r)
yR = vR − vh . (18)
According to Holzapfel7 the relation between reference signal y (r) , pseudo control reference signal vR , and
the hedging vh , see Fig. 3, can by expressed as:
y (r) = v − vh . (19)
The pseudo control signal v contains the reference signal vR and the signals from error feedback vDC :

r−1
v = vR + vDC = vR − c0 (y − yR ) − c1 (ẏ − ẏR ) . . . = vR − cj e(j) . (20)
j=0

By replacing v in Eq. (19) by its expression of Eq. (20), the r-th derivative of the output reads:

r−1
y (r) = vR − cj e(j) − vh + Δ (21)
j=0

(r)
and finally replacing vR by its expression from Eq. (18): vR = yR + vh

(r)

r−1
y (r) = yR + vh − cj e(j) − vh + Δ (22)
j=0

(r)

r−1
= yR − cj e(j) + Δ (23)
j=0

(r)
the influence of the hedging signal to y (r) disappears completely. The error e(r) = y (r) − yR with


r−1
e(r) + cj e(j) = Δ (24)
j=0

a Even the linear feedback does not see the characteristics and therefore the statement fits into this discussion without dealing

with adaptive elements in this paper.

6 of 20

American Institute of Aeronautics and Astronautics


will not be affected by the hedging at all. But, as there is a feedback of plant states into the reference model
the proof of Bounded-Input Bounded-Output (BIBO) stability between the commanded output yc and the
plant output yR will be required:

eCR = yR − yc . (25)

The resulting error dynamics are the superposition of eCR and e as defined above.

eC = eCR + e. (26)

The proof of ultimate boundedness of an error in the presence for approximately inverted dynamics is
given by Kim.6 To determine the influence of plant states as well as the effect of the actuator model to
the reference dynamics the following conversion will help to clarify the relationships between the reference
model and the hedging extension concerning the stability of the reference dynamics.

III. Reshaping the PCH Scheme


From the scheme shown in the previous section it does not seem easy to distinguish the influence of the
actuator and plant model to the reference model output. Moreover, the difference between a commanded and
an achieved control signal is used to affect the reference through a hidden feedback loop, which has major
influence on the reference system response - as intended. In this section, we introduce a rearrangement that
will lead to a equivalent representation permitting to emphasize the significance of the actuator and plant
models and their influence on the reference system response.
A dynamical model of a system (the symbol x̂ denotes an estimate) is represented by

x̂˙ = m̂(x) + N̂(x)u. (27)

Given that ĝ(x)−1 exists the above equation can be transposed into

u = N̂(x)−1 x̂˙ − m̂(x) . (28)

The correlation between the commanded input uc,i and real control input ui is represented by

ui = GA,i (xA,i , uc,i ). (29)

Returning back to Fig. 3, the hedge signal v h is calculated by

vh = v − v̂ (30)
= v R + v DC − v̂. (31)
(r )
Going on to Fig. 2 the ri -th derivation of the reference output vector yR,ii is given by
(r )
yR,ii = vR,i − vh,i . (32)

Up to here everything is identical compared to the referred literature. Remember from Fig. 3, the estimate
of the pseudo-control control input v̂ is defined as

v̂ = F̂(x, û) (33)


= m̂(x) + N̂(x) · û. (34)

Combining Eq. (31), (32), and (34) the reference output becomes

⎡ (r1 )

yR,1
⎢ . ⎥
⎢ . ⎥ = v R − v R − v DC + v̂ (35)
⎣ . ⎦
(rm )
yR,m
= v̂ − v DC (36)
= m̂(x) + N̂(x) · û − v DC , (37)

7 of 20

American Institute of Aeronautics and Astronautics


and therefore, the influence of v R disappears completely from all reference signals! In fact, the reference
model output is purely influenced by the actuator and plant model responses for the nominal case (i. e.
without disturbances and uncertainties).
By recognizing that, the next logical step seems to rearrange the constituents to end up with the new
scheme as shown in Fig. 4. The actuator and the plant models are becoming parts of a new reference
system which is no longer only a linear reference model. Taking into account the model of the plant, this
augmented reference system provides a signal uc which permits to follow the reference yc in accordance with
the desired behavior of the system (specified by means of the AR and BR matrices of the reference model)
and with the physical limitations of the system. For the inversion of the system dynamics the real plant
states are used as shown in section II. To compensate control errors due to disturbances and uncertainties
the error feedback loop provides an additional signal vDC to the pseudo-control input. As the plant as well
as the actuator model are directly included in the augmented reference system the analysis of the resulting
reference dynamics enables statements about the BIBO stability of the reference system (i. e. between input
uR = yc and the outputs vR and y R ).
We now consider the relation between the commanded input u and the estimate of the real input of the
plant û (i. e. the outputs of the actuators). This estimate û can be obtained by simulating an actuator model
ĜA as introduced in Eq. (29). We define Δ̂A as:

û = ĜA (xA , uC ) := uc − Δ̂A (x̂A , uc ), (38)

together with the control value uc as shown in Eq. (28):

uc = N̂−1 (x) [v − m̂(x)] , (39)

following Eq. (37), the ri -th time derivative of reference systems’s output is:
(ri )
yR = m̂(x) + N̂(x) · uc − . . .
−N̂(x) · Δ̂A (xA , uc ) − v DC (40)
= −m̂(x) + N̂(x) · N̂−1 (x) [v − m̂(x)] − . . .
N̂(x) · Δ̂A (xA , uc ) − v DC (41)
= v − N̂(x) · Δ̂A (xA , uc ) − v DC . (42)

The response of the reference system can be chosen by the constants ai and bi as shown in section B.
(r) (r)
Using yR = xR = vR and v = vR + vDC Eq. (42) becomes (simplified for a scalar output of a SISO system)

(r) (r)

r−1
yR = xR = −ai · xR,i + b0 · uR +vDC − N̂ (x) · Δ̂A (x̂A , uc ) − vDC
i=0
 
(r)
xR =vR


r−1
= −ai xR,i + b0 uR − N̂ (x)Δ̂A (x̂A , uc ) (43)
i=0

In this SISO case, N̂ (x) and Δ̂A (x̂A , uc ) are both scalars and therefore are neither written in bold nor
underlined in Eq. (43). The potential major influences of this term on the reference system response must
be considered when placing the reference system poles. In many cases these extra dynamics are fast and
stable with respect to the desired reference response and therefore not really a problem.
Assuming that the positive constants ai as well as the plant model are known Eq. (43) provides the
potential to prove the BIBO stability between yc and yR explicitly in the presence of the PCH system. By
considering the uncertainties of N̂ a stability analysis may become possible, but a stability proof is difficult
due to the dependency of N̂ on the state vector x.
Similarities between the control law for the input-output linearization used before and a model following
control system presented afterwards are used to introduce a conversion into a pure feedforward control
system. This will make the reference system independent from system states and therefore ease the stability
analysis. The following section will give a compact and simplified review of the method to design a explicit
model following control system.

8 of 20

American Institute of Aeronautics and Astronautics


IV. Separation into Feedforward and Feedback
A. Control Law of an Explicit Model Following Control System
This section presents the development of a control law by explicit model following control. For simplification
only linear case without uncertainty is considered in this subsection. The analogy is later used to separate
the PCH scheme into feedforward and feedback components. Beside, the illustration of the process allows
to explain the resulting error dynamics in details as well as to discuss the properties of the resulting control
architecture.
For a well modeled process, the application of a suitable feedforward signal enables excellent command
following and trajectory tracking performance. The response to setpoint changes for a control loop can be
manipulated by adding a feedforward control signal to an existing error feedback as illustrated in Fig. 1.
Having the exact inversion of the plant model as feedforward (with no uncertainties), and assuming to have
no disturbances z as well as identical initial conditions x0 , the plant states x will follow exactly a commanded
signal xR without the need of a feedback loop.
In real world scenarios, however, model uncertainties, disturbances, and unknown initial conditions will
exist. Therefore, a feedback of the state or output vector has to be applied. Usually, the feedback dynamics
are adjusted by linear state or output feedback. Various methods have been developed to account for
robustness against sensor noise, parameter uncertainties, and unmodeled dynamics while determining the
feedback gains. But, as suggested in the title, this paper will focus on the feedforward control component.
In explicit model following design, the aim is to force the real system to react as the reference model that
has been chosen. As a consequence, the reference is added to the feedforward part. This principle is also used
under the name of in-flight simulations when doing flight tests with a flight control computer running such a
model following scheme based on reference model that is in reality the model of a different aircraft. In-flight
simulation can be used for training purposes (using aircrafts that are less expensive to operate than the real
ones during pilots’ formations) or to even to fly aircrafts that does not exist (e.g. which developments is not
yet finished).
This technique was demonstrated for planes (e. g. by Bauschat,18 Saager19 or Duda20, 21 ) as well as for
helicopters (e. g. by Bouwer et al.22, 23 ). As pointed out by Moennich24 an exact command following of the
plant outputs can only be achieved if the number of independent outputs correspond with the number of
independent inputs to the system. Moreover, stability of all existing zero dynamics is required.
To emphasize the principles of explicit model following control system design, first of all, the derivation
of the control law for a linear system will follow. The linear state-space representation of the system will be
represented by:

ẋ(t) = Ax(t) + Bu(t), x(0) = x0 , (44)


y(t) = x(t). (45)

The desired reference response is represented by a reference model:

ẋR (t) = AR xR (t) + BR y c (t), xR (0) = xR0 , (46)


y R (t) = xR (t). (47)

Aim of the explicit model following control is to adjust the system states to the reference states:

xR (t) = x(t), (48)


ẋR (t) = ẋ(t). (49)

Insertion of the reference model states into Eq. (44) by using Eq. (48) and Eq. (49) gives:

ẋR (t) = AxR (t) + Bu(t) (50)

and leads to the desired relationship for explicit model following:

u(t) = B−1 [ẋR (t) − AxR (t)] . (51)

As explained by Moennich:24 ẋR is the desired acceleration and AxR is the acceleration that would be
obtained without command input. The command input must correct the difference ẋR − AxR between
desired and natural acceleration of the system.

9 of 20

American Institute of Aeronautics and Astronautics


Inverting the input matrix B, the adequate input vector u can be obtained as shown in Eq. (51). This
obviously use the assumption that B is invertible which is made here for the sake of simplicity but not
always required. Indeed a necessary conditions for exact explicit model following is that the number of
independent control inputs is equal or greater than the number of independent outputs to be controlled (also
called Lagrange Condition and mentioned above). If there are less control inputs than outputs to control,
only an approximate output following can be achieved.
To analyze the error dynamics, using the control law Eq. (51) in Eq. (44):

ẋ(t) = Ax(t) + BB−1 (ẋR (t) − AxR (t)) (52)


= ẋR (t) + A(x(t) − xR (t)) (53)

the error between reference model and plant states can be defined as:

ė(t) = ẋR (t) − ẋ(t) (54)


= A(xR (t) − x(t)) (55)
= Ae. (56)

Under the aforementioned assumptions, it becomes obvious that the error between reference model and plant
states depends only on the plant dynamics itself and not on the commanded signals at all. By feeding back
the error vector e to the input u the error dynamics can be manipulated.
Considering a full error state feedback, the input vector becomes:

u(t) = B−1 [ẋR (t) − AxR (t)] + K(xR (t) − x(t))


= B−1 [ẋR (t) − AxR (t)] + Ke(t). (57)

Using Eq. (57) in Eq. (44):

ẋ(t) = Ax(t) + BB−1 [ẋR (t) − AxR (t)] + BKe(t)


= Ax(t) + ẋR (t) − AxR (t) + BKe (58)

and substitute e(t) = xR (t) − x(t) and ė(t) = ẋR (t) − ẋ(t) it is evident, that the error dynamics:

ė(t) = ẋR (t) − ẋ(t) (59)


= AxR (t) − Ax(t) − BKe(t) (60)
= (A − BK)e (61)

can be manipulated by the feedback gain K independently from the reference dynamics. As long as the
reference model is stable, the stability of the overall system and the error dynamics only depends on the
chosen feedback gains. By the strict separation of feedforward and feedback the plant states do not have
any influence on the subsystem that generates the reference trajectory. This seems to be a major advantage
and makes this control technique preferable for industrial applications but this conclusion is strongly linked
to the linear assumption utilized in this section.
A disadvantage of the separation of (plant-)state independent and reference state dependent feedforward
control is, that effects on the real states (disturbances, or measurement errors) which are not considered in
the reference model may lead to larger control errors and will be compensated only via the feedback loop.
A sudden (fast) change in the airspeed caused by gust for example, even if detectable, will not cause any
feedforward signal to compensate for the changed condition. Moreover, the design of the feedback loop for
a nonlinear plant with guaranteed robustness and stability can be challenging.
The basic idea of using the knowledge of the plant (i. e. as model) as well as its actuators in a feedforward
way leads to the transformation introduced hereafter.

B. Transformation of the modified PHC System into a Feedforward Control System


As illustrated in section A, a feedforward control part can increase the controller performance without
affecting the error dynamics. Occasionally, it may be useful to adopt the error feedback gains first, for example
within flight tests and without any mathematical model of the plant. To improve the flight performance as

10 of 20

American Institute of Aeronautics and Astronautics


needed the upgrade with feedforward elements may happen afterwards, subsequent to a system identification
process which will provide a mathematical representation of the system.
For the case where a part of the state vector is not available, the decoupling of the feedforward control law
from plant states enables the consideration of the additional dynamics by using the reference states as follows.
This is for instance the case in the application considered in this paper (see section V) in which helicopter’s
rotor states are not available. Similarities between the model following control law Eq. (51) and the one of
the feedback linearization Eq. (6) have to be noticed. The idea behind the transformation of the PCH that
is proposed in this paper is a direct consequence of this observation. The aim of expanding the feedback
linearization method with pseudo control hedging was to create reference signals according to the ability of
the real plant. Assuming the error feedback system is able to reduce the control error to zero, reference-
and plant will tend to be equal. Without considering the error feedback, the PCH reference system yields in
a decoupled reference system by replacing the real measurements of the feedback linearization control law
with the reference system states:

x = xR . (62)

This results in the “new” control law

uc = N̂−1 (xR ) [v − m̂(xR )] , (63)

where only states of the reference system are used. The control law of the input-output linearization can be
used to generate a nonlinear model following control system. Furthermore, this will eliminate the need of
the plant state vector, illustrated as dashed line in Fig. 4.
The final result is a fully separated reference system as illustrated in Fig. 5, which, in addition to the
desired reference states, provides an feedforward signal based on feedback linearization. Since no plant
states are involved, the reference system is a closed system itself. Reference dynamics and the inversion of
an uncertainly known plant model are separated from the error feedback.
In this new formulation the BIBO stability of the reference system is completely independent from
uncertainties. Compared to the “classical” PCH system the stability analysis is here much simpler, because
it is not required to consider the entire system. Moreover, the PCH system requires a robust stability analysis
to take into account the differences between the model that permits to compute N̂(x) and the system that
would result in N(x). Such an analysis will be much more complex to conduct than the stability proof one
can conduct in the new formulation.
This formulation offers the new possibility to a control designer where the FBL and PCH are used for
the generation of the reference trajectory only. Improvements of robustness and disturbance rejection are
let for the feedback controller. The desired response of the reference system can be designed using any
kind of criterion (e.g. handling qualities) and the performances will be kept as long as the physics of the
system allows it. Then the PCH included in the feedforward will prevent the instabilities that could result
from excessive control inputs. The plant output feedback loop gains must be determined by the amount
of uncertainties, state estimation errors and all other robustness criteria. As the feedback loop must be
designed for a nonlinear system its tuning may be more complex then with the original approach. The
symbolic illustration of the feedback loop in Fig. 5 represents a feasible method to design disturbance
rejection, which is furthermore important to compensate for uncertainly known initial conditions as well
as system parameters. Application of this approach to a linear system will result in the error dynamics
presented in Eq. (61).
By removing the feedback of control errors into the reference system, limitations of the real plant are not
regarded in the determination of the reference states. Therefore, slightly conservative limits in the reference
system might be used to prevent the system from exhibiting excessive performance deteriorations or even
loss of stability, as well as an antiwindup mechanism for the feedback controller if it can be subject to windup
(e. g. with integral part). The system as defined so far was tested in flight with constant feedback gains on
an unmanned rotorcraft testbed. Obtained results are presented in the following section.

V. Implementation for an Unmanned Helicopter and Flight Test Results


The Autonomous Rotorcraft Testbed for Intelligent Systems (ARTIS), under development at the DLR
Institute of Flight Systems, provides a low-cost research platform to investigate methods and technologies

11 of 20

American Institute of Aeronautics and Astronautics


for future VTOL (Vertical Takeoff and Landing) UAVs. Various publications cover different aspects of the
challenge of developing an autonomous rotorcraft at the DLR.25–29 Concerning the approach of using open-
loop reference systems for the flight control for this UAV in the following the velocity reference system will
be considered in more details.
In general, a classical helicopter has four control inputs. Three of them are used to change the attitude
and one is used to manipulate the magnitude of the thrust vector. To manage six degrees of freedom with
less than six control inputs, the differences in terms of frequency response will be utilized to separate the
rotational motion from the translational one. Usually, this is done by providing accelerations from an outer
loop to determine the required change in attitude for an inner loop. The inner loop dynamics will be seen
as an actuator from the outer loop. Without further discussion the required pitch and roll angles can be
determined by:

   
ẍgg ,c −ÿgg ,c cos θ
θ = arctan φ = arctan . (64)
z̈gg ,c z̈gg ,c

The term ẍgg represents an acceleration expressed in an around ψ rotated geodetic frame with respect to
the geodetic frame.
As pointed out in the introduction the knowledge of the system model is fairly limited. Usually, these
kind of technical demonstrators are based on radio controlled helicopter models. A complete identification
of the systems dynamics based either on numerical or on experimental data appears uneconomically when
comparing the magnitude of the system costs to the identification expenses. In addition the complexity to
achieve a mathematical model for the fast forward flight regime increases even more when the system must
be exited from a person standing fix on the flight field. Due to these constraints the control design for such
low-cost helicopter models is often based on a linear model around hover condition.

A. Mathematical description of the translational motion at hover flight domain


The translational motion is driven by the forces at the main rotor in combination with the Earth’s gravity
force. The approximation of the dynamics of the utilized demonstrator is based on estimates of force and
moment derivatives obtained from flight test data recorded during frequency sweep excitations. It can be
represented by the following differential equations:
⎡ ⎤ ⎡ ⎤
u̇ u
⎢ ⎥ ⎢ ⎥
⎣ v̇ ⎦ = m̂trans (x) + N̂trans uc ; y=⎣ v ⎦ (65)
ẇ w

that link the accelerations (u̇,v̇,ẇ) to the state x and the command vector uc . Note that due to the linear
approximation made N̂trans does not depend on the state vector x. As this set of equations represents the
system model in the form of Eq. (2) the required change in the vector uc can be directly deducted from the
required change in the acceleration.
The influence of the Earth’s gravity to the motion of the helicopter is nonlinear due to the frame trans-
formation which depends on the attitude of the vehicle. The acceleration becomes the superposition of the
linear approximation and the nonlinear gravity effect. Due to the hover domain a extension due the rotation
of the body frame is not required, but can be added to enhance the model.
Written in the body fixed frame the derivation of the velocity output vector against the body fixed frame
becomes
⎡ ⎤b ⎡ ⎤
 b u̇ 0
dv ⎢ ⎥ ⎢ ⎥
= ⎣ v̇ ⎦ = Âtrans xtrans + N̂col ⎣ 0 ⎦ + Tbg g −ω gb ×v . (66)
dt b b b
ẇ b δcol f or hover ≈0

Equation (66) gives the relation of the control inputs for the collective control as well as the command to a
underlying attitude reference system. The collective control input has major influence in the acceleration in
body fixed vertical direction. To establish a commanded acceleration u̇b , v̇b a change of the attitude will be

12 of 20

American Institute of Aeronautics and Astronautics


required. The change in attitude can be determined by the knowledge of the actual state of ẇbb . The state
vector xtrans is composed of the system states as well as the rest of the system inputs:
T
xtrans = [uf , vf , wf , p, q, r, ṗ, q̇, δlat , δlon , δped ] . (67)

The matrix Tbg represents the transformation from geodetic into body frame. Due to the rotation of the
body frame compared to the geodetic frame a term depending on the rotation and translational speed is
added.
The matrix Âtrans , representing the linear part of the approximation of mtrans introduced in Eq. (66),
contains the derivatives assigned to the corresponding state or input signal. It reads:

⎡ ⎤
Xu Xv Xw Xp Xq Xr Xṗ Xq̇ Xlat Xlon Xped
Âtrans = ⎣ Yu Yv Yw Yp Yq Yr Yṗ Yq̇ Ylat Ylon Yped ⎦ . (68)
Zu Zv Zw Zp Zq Zr Zṗ Zq̇ Zlat Zlon Zped

The resulting acceleration depending on the collective control input is approximated by a factor N̂col
which is multiplied by the collective control input value in the linear part of Eq. (66). The remaining
elements of the given expression of N̂col are equal to zero.

B. Reference System for the translational motion


T
The aim is to control the output vector which contains the body fixed velocities: y = [u, v, w]b . The
corresponding accelerations required to derive the change in attitude are determined by a reference system.
As the accelerations and speeds are linked by first-order equations, the reference system can be based on
three first-order systems in parallel. Written in state space representation the change in the acceleration
depending on the velocity command and the (reference) velocity state becomes:
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤⎡ ⎤
u̇R −au 0 0 uR −bu 0 0 ucmd
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
v R = ⎣ v̇R ⎦ = ⎣ 0 −av 0 ⎦ ⎣ vR ⎦ + ⎣ 0 −bv 0 ⎦ ⎣ vcmd ⎦ . (69)
ẇR 0 0 −aw wR 0 0 −bw wcmd
   
AR BR

The parameters ai and bi must be tuned to get the desired transfer behavior, e. g. to a step in the
velocity command. To adjust the resulting acceleration to large changes in the velocity command the
reference acceleration v R will be limited by appropriate values. This is required because the linear dynamics
approximation does not contain such limitations and in consequence the pseudo control hedging will not
prevent accelerations outside the desired flight envelope.
A inner loop represented by the attitude reference system, which is not discussed in detail, acts as the
actuator to the outer loop reference system. Therefore, the dynamics of the first order translational reference
system is “hedged” by the inner model, as depicted by Johnson.16
The overall reference system is illustrated in Fig. 6. The open-loop FBL/PCH scheme introduced
in section B will be used with the reference model defined in Eq. (69). From the velocity command
generated by an outer guidance system the change in acceleration will be derived. First of all, the velocity
command is converted into a body fixed frame and limited to the desired range. As the relative degree is
one the acceleration reference signals are calculated accordingly. After the limitation of the acceleration
the vertical part is used to derive the necessary thrust respectively the required collective control input.
Horizontal accelerations will be achieved by changing the attitude of the helicopter. Therefore by using a
algebraic relation between acceleration and attitude the change in attitude can be computed for the desired
accelerations. The resulting acceleration includes the part depending on the current attitude as well as the
one from the gravity force in body fixed frame. By the integration of the sum the velocity reference signals
are computed.

C. Flight test result


A snapshot of a typical trajectory is extracted from flight test data and shown in the Fig. 7 to 9. For each
degree-of-freedom three variables are presented: Command, Reference Signal, and State. The Command

13 of 20

American Institute of Aeronautics and Astronautics


represents the input to the reference system uR = y c . For the given example, the velocity commands are a
combination of the path direction and a compensation for path errors, represented as solid (dark/blue) line.
Together with some acceleration limits, the reference system produces the Reference Signals, displayed
as a solid (gray/green) line. The velocity reference commands are calculated based on a inner loop reference
system, which is considered for the outer loop as “actuator”. This inner loop as well as the path following
guidance system will not be discussed in this paper at all.
A sensor fusion algorithm based on an Extended Kalman Filter recovers from the noisy raw measurements
the current State vector, represented by a dashed (red) line. The Velocity error, is fed back by a proportional
and integral feedback loop to produce an additional control signal accounting for model uncertainties and for
the rejection of disturbances as well as for static errors. By combining the feedforward and feedback control
signals, the helicopter performs the given trajectory with acceptable control errors as can be seen from Fig.
8, especially when considering that the control system is based only on a linear hover domain model.
Figure 9 presents the feedforward and feedback signals from the velocity reference system. As solid
(dark/blue) line the attitude feedforward commands ûc as Euler angles. For the collective input the solid
(dark/blue) line presents the feedforward of the collective control signal. The solid (gray/green) lines are
the signals resulting from the sum of feedforward and feedback signals uc . The dashed (red) lines are the
differences between the entire and the feedforward commands uc − ûc . Therefore, they represent the feedback
signal due to control errors between reference system and plant states. Compared to the feedforward signals
the feedback have small values which shows that the feedforward signal was suitable defined with respect
to systems limitations. Larger feedback signals could result from disturbances like wind and gust or from
modeling errors.
From a path guidance system perspective, the reference system acts as a parameter dependent nonlinear
low-pass filter. Therefore, velocity reference states as well as plant velocity states are lagging compared to
the commands provided by the guidance system. The flight path resulting from the plant states will differ
from one which are based on the guidance system outputs. As a consequence the flight path control loop
may be adapted to take into account the closed loop response of the (reference) system. The adaptation
of the guidance system depending on the reference dynamics is one of the topics of the current research at
DLR.

VI. Summary and Concluding Remarks


In this paper a new method based on feedback linearization and pseudo control hedging has been intro-
duced to use them for the design of a nonlinear feedforward controller.
The paradigm shift is feasible after reshaping the structure presented in the literature to put into relief
the interaction between reference system and model based control hedging.
Motivated by the difficulty of a stability analysis of the common PCH system, caused by having a
feedback of real states into the reference system, a method to encapsulate the reference dynamics has been
presented. The benefit of the approach proposed is a strict separation of plant and reference dynamics,
which simplifies in conclusion a stability analysis. As a feedforward can neither compensate model errors nor
reject disturbances, this must be made by a feedback controller. Discussion of the error feedback as well as
the resulting error dynamics are presented briefly. Its tuning has been let outside of the scope of this paper
whose focus is the design of the feedforward part.
By applying the new control architecture for an unmanned helicopter testbed some flight test results
are presented and show the applicability of this method. The reference tracking errors were acceptable,
in particular when considering the fact only a hover domain model was used to define the feedforward.
Further research ware made to combine adaptive elements with this approach with the aim of improving the
capabilities of this control method. This should also permit to extend the flight region limitations caused by
poor plant models without having a need to use the adaptive components afterwards outside the research
field.

VII. Acknowledgments
The author gratefully acknowledges the contribution of the ARTIS team members as well as the fellows
of the unmanned systems department. In particular Jörg Dittrich, Florian M. Adolf, Franz Andert, Lukas
Goormann, Johann Dauer, Dr.-Ing. Gordon Strickert, Christian Greiner-Perth, Fabian Klüßendorf, and our
safety pilot Jörg Rösner. Special thanks to Dr. Nicolas Fezans for the valuable feedback and the constructive
discussions.
14 of 20

American Institute of Aeronautics and Astronautics


References
1 Slotine, J.-J. E. and Li, W., Applied Nonlinear Control, Prentice Hall, Englewood Cliffs, NJ, 1991.
2 Kahlil, H., Nonlinear Systems, Prentice Hall, Upper Saddle River, NJ, 2nd ed., 1996.
3 Johnson, E. and Calise, A., “Neural Network Adaptive Control of Systems with Input Saturation,” Tech. rep., Georgia

Institute of Technology, Atlanta, Georgia, Juni 2001.


4 Johnson, E., Calise, A., El-Shirbiny, H., and Rysdyk, R., “Feedback Linearization with Neural Network Augmentation

applied to X-33 Attitude Control,” AIAA Guidance, Navigation and Control Conference and Exhibit, No. AIAA-2000-4157,
American Institute of Aeronautics and Astronautics, Inc., August 2000.
5 Hovakimyan, N., Kim, N., Calise, A. J., Prasad, J. V. R., and Corban, E., “Adaptive Output Feedback for High-

Bandwidth Control of an Unmanned Helicopter,” AIAA Guidance, Navigation, and Control Conference, No. AIAA 2001-4181,
American Institute of Aeronautics and Astronautics, Inc., Montreal, Canada, August 2001.
6 Kim, N., Improved Methods in Neural Network Based Adaptive Output Feedback Control, with Applications to Flight

Control, Ph.D. thesis, School of Aerospace Enginieering, Georgia Institute of Technology, Juli 2003.
7 Holzapfel, F., Nichtlineare adaptive Regelung eines unbemannten Fluggeräts, Ph.D. thesis, Technische Universität

München, 2004.
8 Singh, S. N. and Schy, A. A., “Output feedback nonlinear decoupled control synthesis and observer design for manoeu-

vering aircraft,” International Journal of Control, Vol. 31, 1980, pp. 781–806.
9 Lipp, A. M. and Prasad, J. V. R., “Synthesis of a Helicopter Nonlinear Flight Controller Using Approximate Model

Inversion,” Mathematical and Computer Modelling, Vol. 18, August 1993, pp. 89–100.
10 Schrage, D. P., Yillikci, Y. K., Liu, S., Prasad, J. V. R., and Hanagud, S. V., “Instrumentation of the Yamaha R-

50/RMAX Helicopter testbeds for Airloads Identification and follow-on research,” 25th European Rotorcraft Forum, 1999.
11 Corban, J. E., Calise, A. J., Prasad, J. V. R., Heynen, G., Koenig, B., and Hur, J., “Flight Evaluation of an Adaptive

Velocity Command System for Unmanned Helicopters,” AIAA Guidance, Navigation, and Control Conference and Exhibit,
No. AIAA-2003-5594, American Institute of Aeronautics and Astronautics, Inc., Austin, TX, August 2003.
12 Kim, N., Calise, A. J., Corban, J. E., and Prasad, J., “Adaptive Output Feedback for Altitude Control of an Unmanned

Helicopter Using Rotor RPM,” AIAA Guidance, Navigation and Control Conference, 2004.
13 Ito, D., Georgie, J., Valasek, J., and Ward, D. T., “Reentry Vehicle Flight Controls Design Guidelines: Dynamic Inver-

sion,” Tech. Rep. NASA/TP-2002-210771, NASA Flight Simulation Laboratory, Houston, Texas, March 2002.
14 NASA, “Fact Sheets: F-15B Research Testbed,” http://www.nasa.gov/centers/dryden/news/FactSheets/FS-055-DFRC.

html, 2007.
15 Johnson, E., “Pseudo-Control Hedging: A new Method for Adaptive Control,” Tech. rep., Georgia Institute of Technology,

Atlanta, Georgia, November 2000.


16 Johnson, E., Limited Authority Adaptive Flight Control, Ph.D. thesis, Georgia Institute of Technology, Atlanta, GA,

November 2000.
17 Johnson, E. N. and Turbe, M. A., “Modeling, Control, and Flight Testing of a Small Ducted-Fan Aircraft,” Journal of

Guidance, Control, and Dynamics, Vol. 29, 2006, pp. 769–779.


18 Bauschat, J.-M. and Lange, H.-H., “ATTAS and its Contributions to System Design and In-Flight Simulation.” Interna-

tional Conference on Aircraft Flight Safety, Zhukovsky, Russia, 31 August - 5 September 1993 , 1993.
19 Saager, P. and Heutger, H., “Ground-Based System Simulation for an Inflight Simulator Aircraft.” Proceedings of the

15th ADIUS Annual Conference, Ann Arbor, MICH, USA, 12 - 15 June 1994 , 1994, pp. 203 – 219.
20 Duda, H., Bouwer, G., Bauschat, J.-M., and Hahn, K.-U., “Autopilot Design Based on the Model Following Control

Approach.” Robust Flight Control - A Design Challenge. Lecture Notes in Control and Information Sciences, Springer, London,
1997, pp. 360 – 378.
21 Duda, H., Bouwer, G., Bauschat, J.-M., and Hahn, K.-U., “A Review of the Model Following Control Approach.” Robust

Flight Control - A Design Challenge. Lecture Notes in Control and Information Sciences, Springer-Verlag, London, 1997, 1997,
pp. 116 – 124.
22 Bouwer, G., “The Application of Model Following Control to Helicopter Tracking Tasks.” Tech. Rep. 111-91/23, DLR

Braunschweig, Institut für Flugsystemtechnik, 1991.


23 Bouwer, G., von Grünhagen, W., and Pausder, H.-J., “Model Following Control for Tailoring Handling Qualities. - ACT

Experience with ATTHeS -,” AGARD-CP-560 , AGARD, 1995.


24 Mönnich, W., “Vorsteuerungsansätze für Modellfolgesystem,” Internal Report IB 111-1999/20, German Aerospace Center,

Braunschweig, 1999.
25 Dittrich, J. S., Bernatz, A., and Thielecke, F., “Intelligent Systems Research Using a Small Autonomous Rotorcraft

Testbed,” 2nd AIAA ”Unmanned Unlimited” Systems, Technologies, and Operations, No. AIAA 2003-6561, American Institute
of Aeronautics and Astronautics, Inc., San Diego, California, September 2003.
26 Chowdhary, G. and Lorenz, S., “Non-linear Model Identification for a Miniature Rotorcraft, Preliminary Results,” AHS

Forum 61 , AHS International, Grapevine, Texas, June 2005.


27 Koch, A., Wittich, H., and Thielecke, F., “A Vision-based Navigation Algorithm for a VTOL-UAV,” AIAA Guidance,

Navigation and Control Conference, edited by AIAA, 08 2006.


28 Adolf, F.-M. and Thielecke, F., “A Sequence Control System for Onboard Mission Management of an Unmanned Heli-

copter,” AIAA Infotech@Aerospace 2007 Conference and Exhibit, No. AIAA 2007-2769, American Institute of Aeronautics and
Astronautics, Inc., Rohnert Park, California, Mai 2007.
29 Andert, F. and Goormann, L., “Stereo-Based Obstacle Mapping from a Helicopter Platform,” CEAS European Air and

Space Conference, 09 2007, pp. 3273 – 3278.

15 of 20

American Institute of Aeronautics and Astronautics


H
H
- HH

 
 z
Feedforward
H ?
HH ... ?
r - .................-
xR e + u x
H-.............. - r -
+ -
  +
6  
Feedback Plant

Figure 1. Simplified scheme of a feedforward and feedback loop

vR -
(n)
yR -
(n−1)
yR -
(n−2)
yR -

...
vh
ẏR -
a    
uR- aa - +  +?
t-
- yR
b0 !a
! +  t- t- t. . . t- t -
!! 6


 
 

` `−[a ...a ]
...

```r−1 0
`` ` 


Figure 2. Diagram of a linear reference system with a hedging signal

vh  - v̂
 +
6
- ĜA (x̂A , uc ) û- F̂ (x, û)

Aktuator Model Plant Model

- vR -+ v -
r F̂ −1 (x, v) uc -
r GA (xA , uc ) u- F (x, u)
y (r) 
- ...

-
r
ẏ 
-
y
ẋR = AR xR + BR uR +
HH v 6 Inversion
R  - -
yR = xR Actuator Plant Integrators
-
uR ξ- e
HDC
+ - P D
6 
...

Error Feedback
Reference Model

ξ = y ẏ ÿ . . . y (r−1)

Figure 3. Scheme of pseudo control hedging

16 of 20

American Institute of Aeronautics and Astronautics


Integrators

y (r) ẏ y
uc
- GA (xA , uc ) u
- m(x) + N (x)u
- 
...

s- 
s-

Actuator Plant

...
x 

?
m̂(x)

s

νDC
s  ID e ?

-

HPHH +
H
Error Feedback

HH (r)
+ νR +?ν--?
-?
+ ν̂ -?
- b0 H    
yc yR
H-  + - - N̂ −1 (x) s- -
û -  - -
s s. . . s-
+  +  +  +
ĜA (x̂A , uc ) N̂ (x)
 6


Actuator Model



```−[ar−1 ...a0 ]  s

...
``` 
`` 
yR

Figure 4. Feedback linearisation of a n-th order system in due consideration of actuator dynamic

...

-
+ -
- - m(x) + N (x)u -  
s- 
-
+
 GA (xA , uc ) ...
6uC u y (r) ẏ y
Actuator Plant Integrators

HH (r)
v = v + +
H + R -
- b0 H    
yc û û xR
 - +  - s - - - - -
s s. . . s-
C
N̂ −1 (xR ) ĜA (x̂A , ûc ) N̂ (xR ) 

- +
 6 6 6
Actuator Model
s m(xR )


xR



```−[ar−1 ...a0 ]  s
...

``` 
`` 
?
Feedforward and Reference System
xR → yR


e ?


+
HP ID  f
HH -
H

Figure 5. Feedback linearisation of a n-th order system used as feedforward control

17 of 20

American Institute of Aeronautics and Astronautics


⎡ ⎤ δcol -
u   ⎡ ⎤
⎣ ⎦ q u
v   ⎣ ⎦
v
w g,c
?
w
HH
H −1  f,R
- Tgf (q) -  +
- +
- - Ncol (xR ) s - Ncol (xR ) -+  -
 - - ω0,R H  s
 - +
 - 
6  6 6

 Thrust Inversion Actuator- and
Transformation Limiter Limiter
Engine Dynamics

 
- Δv̇ → Δq - q̇ = f q (...) - q
ω̇ = f (...)  
ω

Attitude Reference System


s

18 of 20
- m(xR ) +
 - 
q +
  6
?
g - Tgf (q)

Transformation

American Institute of Aeronautics and Astronautics


s

Reference System

Figure 6. Scheme of the velocity reference system for ARTIS


Flight Test ARTIS, 14.12.2007, Flight 2
10
ub in m/s Command
5 Reference Signal
State
0
285 290 295 300 305 310 315 320 325 330 335 340
5
vb in m/s

−5
285 290 295 300 305 310 315 320 325 330 335 340
5
wb in m/s

−5
285 290 295 300 305 310 315 320 325 330 335 340
Time in s

Figure 7. Flight test result using the ARTIS technology demonstrator

Flight Test ARTIS, 14.12.2007, Flight 2


eu in m/s

2 Velocity error
0
−2
290 300 310 320 330 340
ev in m/s

2
0
−2
290 300 310 320 330 340
ew in m/s

2
0
−2
290 300 310 320 330 340
Time in s

Figure 8. Velocity error for a sample flight test snapshot

19 of 20

American Institute of Aeronautics and Astronautics


Flight Test ARTIS, 14.12.2007, Flug 2
collective input roll angle in ° pitch angle in °

40
20
0
−20 uc,hat
−40 uc
290 300 310 320 330
uc−uc,hat 340
20

10

−10
290 300 310 320 330 340

0.2
0
−0.2
−0.4
−0.6
290 300 310 320 330 340
Time in s

Figure 9. Input signals obtained from the velocity reference system

20 of 20

American Institute of Aeronautics and Astronautics

View publication stats

You might also like