Académique Documents
Professionnel Documents
Culture Documents
1
Westin Seattle Hotel, Seattle, Washington, USA
June 11-13, 2008
Abstract— We consider the application of a conditional inte- that does not use gain-scheduling. The application of SMC
grator based sliding mode control design for robust regulation to flight control has been pursued by several others authors,
of minimum-phase nonlinear systems to the control of the see, for example, [5], [6], [19]. Our work differs from earlier
longitudinal flight dynamics of an F-16 aircraft. The design
exploits the modal decomposition of the linearized dynamics ones in that it is based on a recent technique in [17] for
into its short-period and phugoid approximations. The control introducing integral action in SMC. While we design a
design is based on linearization, but is implemented on the nonlinear controller, it is still designed based upon plant
nonlinear multiple-input multiple-output longitudinal model of linearization. In particular, our design exploits the modal
the F-16 aircraft. We consider model following for the angle-of- decomposition of the linearized dynamics into its short-
attack, with the regulation of the aircraft velocity (or the Mach-
hold autopilot) as a secondary objective. It is shown through period and phugoid approximations. Our primary emphasis
extensive simulations that the inherent robustness of the SMC is on the transient and steady-state performance of control
design provides a convenient way to design controllers without of the aircraft’s angle of attack, with the steady-state per-
gain scheduling, with transient performance that is far superior formance and disturbance rejection of the aircraft’s velocity
to that of a conventional gain-scheduled approach with integral as a (minor) secondary objective. The desired transient and
control.
steady-state specifications for the angle of attack are encap-
I. INTRODUCTION sulated in the response of a reference model, and the (SMC)
The dynamic response characteristics of aircraft are highly controller is designed as a model-following controller. As a
nonlinear. Traditionally, flight control systems have been consequence of exploiting the modal decomposition of the
designed using mathematical models of the aircraft linearized aircraft dynamics, the controller has a very simple structure.
at various flight conditions, with the controller parameters It is simply a high-gain PI/PID controller with an “anti-
or gains “scheduled” or varied with the flight operating windup” integrator, followed by saturation. This controller
conditions. Various robust multivariable techniques including structure is a special case of a general design for robust
linear quadratic optimal control (LQR/LQG), H∞ control, output regulation for multiple-input multiple-output (MIMO)
and structured singular value µ-synthesis have been em- nonlinear systems transformable to the normal form, with
ployed in controller design, an excellent and exhaustive com- analytical results for stability and performance described in
pendium of which is available in [10]. In order to guarantee [17]. Through simulations, we show the efficacy of the design
stability and performance of the resulting gain-scheduled and that it outperforms a traditional gain-scheduled controller
controllers, analytical frameworks of gain scheduling have design based on the polynomial approach to model-following
been developed, including the powerful technique of linear- design.
parameter-varying (LPV) control [4], [9], [18], [21]. Nonlin- The rest of this paper is organized as follows. In Section
ear design techniques such as dynamic inversion have been 2, we describe the nonlinear mathematical aircraft model, its
used in [1], [14], [20], while a technique that combines model linearization and the decomposition of the dynamics into the
inversion control with an online adaptive neural network to short-period and phugoid modes. This section is extracted
“robustify” the design is described in [16], and a nonlinear mostly from [22], with the Simulink model for simulation
adaptive design based on backstepping and neural networks purposes based on [15]. Controller design is discussed in
in [7]. A RBFNN based adaptive design with time-scale Section 3, and simulation results showing the efficacy of the
separation between the system and controller dynamics, with design are presented in Section 4. Finally, a summary of
applications to control of both longitudinal (angle-of-attack our work and some suggestions for possible extensions are
command systems) as well as lateral (regulation of the provided in Section 5.
sideslip and roll angles) is described in [23]. A succinct
“industry perspective” on flight control design, including the
techniques of robust control (H∞ , µ-synthesis), LPV control,
dynamic inversion, adaptive control, neural networks, and II. 3-DOF L ONGITUDINAL M ODEL
more, can be found in [3].
Our interest is in the design of robust sliding mode control
(SMC) for the longitudinal flight dynamics of a F-16 aircraft Assuming no thrust-vectoring, the equations for pure lon-
gitudinal motion (pitching and translation) of a high perfor-
*Financially supported in part by the Royal Thai Air Force. mance aircraft can be described by the 5th order nonlinear
1771
to compute the linearization. The flight envelope that we use layer |s| ≤ µ, σ̇ = k1 e1 + k2 e2 + · · · + eρ , which implies
for computing the trim conditions and the linearization is the that ei = 0 at equilibrium, i.e., (5) is the equation of an
cross product set (V̂ , ĥ) ∈ ΩV × Ωh , where ΩV = [300, 900] integrator that provides integral action “conditionally”, inside
ft/s in steps of 100, while Ωh = [5000, 40000] ft in steps the boundary layer. As shown in [17], such a design provides
of 5000. Both our design and the traditional gain-scheduled asymptotic error regulation, while not degrading the transient
design are evaluated on the full fifth order nonlinear state performance, as is common in a conventional design that
model (1). uses the integrator σ̇ = e1 . In the case of relative degree
ρ = 1 and ρ = 2, the controller (3) is simply a specially
III. C ONTROL D ESIGN tuned saturated PI/PID controller with anti-windup (see [17,
Our primary control objective is the design of an angle Section 6]).
of attack command system. For such a system, the entire The control (3) can be extended to the output-feedback
dynamic response is important, and we assume that the case by replacing ei by its estimate êi , obtained using the
desired specifications are encapsulated in a reference model. high-gain observer (HGO)
Whereas a parameter-varying reference model with higher ê˙ i = êi+1 + αi (e1 − ê1 )/ǫi , 1 ≤ i ≤ ρ − 1
bandwidths at higher speeds and lower bandwidths at lower (6)
ê˙ ρ = αρ (e1 − ê1 )/ǫρ
speeds has been used in [4], we use a reference model
with fixed parameters in our simulations. However, it is where ǫ > 0, and the positive constants αi are chosen such
straightforward to include a parameter-varying reference that the roots of λρ + α1 λρ−1 + · · · + αρ−1 λ + αρ = 0
model. Since the original system is MIMO, we also consider, have negative real parts. To complete the design, we need
but as a secondary objective, a Mach-hold autopilot for the to specify how k, µ and ǫ (in the output-feedback case) are
regulation of the velocity V . The reference model that we chosen. The parameter k is chosen “sufficiently large” (to
use for the control of α is similar (but not identical) to the overbound uncanceled terms in ṡ) while µ and ǫ are chosen
one in [4] “sufficiently small”, the former to recover the performance
of ideal (discontinuous) SMC (without an integrator) and
αm (s) 9
Gm (s) = = 2 the latter to recover the performance under state-feedback
αd (s) s + 1.4s + 9 with the continuous SMC. Analytical results for stability
where αd is the angle of attack pilot command. Our approach and performance are given in [17], but are not directly
to control design for α is based (see [17]) on minimum- applicable to this work since they were done for control
phase systems transformable to the normal form η̇ = φ(η, ξ), affine systems. Consequently, we only apply the design to
ξ˙ = Ac ξ+Bc γ(x) [u−α(x)], y = Cc ξ, where x ∈ Rn is the the (control affine) linear approximation (2) and “verify” the
state, u the input, ρ is the system’s relative degree, ξ ∈ Rρ efficacy of the design through simulations. A mention of
the output and its derivatives up to order ρ − 1, η ∈ Rn−ρ stability and boundedness under this design is made in the
the part of the state corresponding to the internal dynamics, concluding paragraph of this section.
and the triple (Ac , Bc , Cc ) a canonical form representation
of a chain of ρ integrators. A SMC design for such a system A. The relative degree ρ = 1 case
was carried out in [17], with the assumption that the internal In order to formally apply the controller design to the
dynamics η̇ = φ(η, ξ) are input-to-state stable (ISS) with ξ short period approximation in (2), with δeδ as input and αδ
as the driving input. For such systems, it is shown in [17] as output, we need to compute this system’s relative degree,
that a continuous sliding mode controller of the form transform it to normal form and check internal stability. The
next assumption states these properties.
k0 σ + k1 e1 + k2 e2 + · · · + eρ
u = −ksign(γ(x)) sat Assumption 2: Consider the short-period approximation
µ
(3)
α̇δ def aαα aαq αδ
b
can be designed to achieve robust regulation, where e1 , ..., = + αδ δeδ
q̇δ aqα aqq qδ bqδ
eρ are the tracking error and its derivatives up to order ρ, the
positive constants ki , i = 1, · · · , ρ − 1 in the sliding surface with output αδ . Then (i) bαδ < 0, so that the relative degree
function aαα aαq
ρ ρ = 1, and (ii) A11 = is Hurwitz, i.e., the
X aqα aqq
s = k0 σ + ki ei + eρ (4) system is minimum-phase.
i=1 While we have only verified Assumption 2 numerically, for
are chosen such that the polynomial λρ−1 + kρ−1 λρ−2 + each trim condition, an analytic discussion based on the
· · · + k1 is Hurwitz, and σi is the output of the “conditional stability derivatives can be found in [22]. Assumption 2
integrator” allows us to design a SMC controller of the form (3) for
the α-dynamics. Because of the integrator, we don’t need
s
σ̇ = −k0 σ + µ sat , σ(0) ∈ [−µ/k0 , µ/k0 ] (5) to add a (gain-scheduled) nominal value to our perturbation
µ input, i.e., we can simply take the output of the controller
where k0 > 0, and µ > 0 is the “width” of the boundary (3) to be u, not uδ . In particular, since ρ = 1, and the high-
layer. From (4) and (5), it is clear that inside the boundary frequency gain bαδ < 0, the control (3) specializes to the
1772
saturated PI controller with the gains kp , kI > 0 chosen to assign the eigenvalues
k0 σ + e
of the resulting 2nd-order system (with states σV and Vδ ) at
δe = k sat(s/µ) = k sat (7) desired pole locations.
µ
As previously mentioned, the analytical results of [17] do
with σ̇ = −k0 σ+µ sat k0 σ+eµ where e = α−αm , k0 , µ > not directly apply to this design, and we do not provide a rig-
0, and k > 0 is simply chosen to be the maximum allowable orous analysis here. However, assuming that the short-period
limit for δe . This completes the design of the controller. approximation holds, a naive argument that the controller
achieves boundedness of all states, and asymptotic error
B. The relative degree ρ = 2 case regulation of the error e is presented below. The SMC (7)
achieves robust regulation of the angle of attack α, provided
While in Assumption 2(i), we said bαδ < 0, it is seen the value of k is “sufficiently large”. The variable q is
(numerically) that bαδ ≈ 0 (this is not simply a “scaling” bounded since the system is minimum-phase. The variable θ
issue, as |bαδ | ≪ |bqδ |). Consequently, the short period evolves according to θ̇ = q, and hence is bounded whenever
approximation (2) with output αδ is “practically” relative q is. The PI controller (10) achieves boundedness of the
degree ρ = 2 3 and not ρ = 1, so that the control (3) now velocity V . Finally, from the equation of ḣ, it follows that
becomes that of a saturated PID controller h is bounded for all finite time whenever V is, so that with
k0 σ + k1 e + ė our SMC and PI controllers for δe and T respectively, all the
u = k sat(s/µ) = k sat (8)
µ states of the closed-loop system are bounded. Our simulation
results, which we present next, appear to validate the above
with k0 , k1 , µ > 0, and as before k > 0 simply chosen to be
conclusions.
the maximum allowable limit for δe . The expression for the
control involves the derivative of α, and it well-known that IV. S IMULATION R ESULTS
the measurement of α is often noisy because of turbulence.
For example, it is common (see, for example, [22]) to assume Numerical values of the SMC parameters that we use in
measurements of α to be corrupted by vertical wind gust all the simulations are k0 = k1 = 1, and that k = 25,
noise wg , with spectral density given by the Dryden model, so that −25 ≤ δe ≤ 25, also used in [15]. The initial
and generated (approximately) by the controllable canonical values in all simulations correspond to trim conditions with
realization of the shaping filter (V̂ , ĥ) = (600f t/s, 20000f t). The control is always tested
on the full 5th order nonlinear model. While we did compare
ż = Aw z + Bw w, wg = Cw z (9) our simulation results against a more classical gain-scheduled
driven by the white noise input w(t) ∼ (0, 1). For our polynomial-based approach to model following (described in
purposes, we simply assume that the derivative ė in (8) is [12, Chapter 3]), they are not presented here, and we only
unavailable as a direct measurement and use the HGO (6) to mention that the SMC design in this paper far outperforms
estimate it (in fact, in simulations with noise, we also replace the gain-scheduled design.
e by its estimate ê obtained from the HGO). A discussion of Our first simulation shows the performance with no in-
the effect of measurement noise on the HGO is discussed in tegral action and assuming ρ = 1, i.e., u = −k sat eµ1 ,
[2], under the assumption that the noise signal is the output with µ = 0.1, when the pilot command αd is a doublet-like
of a linear system driven by a bounded input, which is the signal with value 15 for 0 ≤ t < 5, -5 deg for 5 ≤ t < 10
case here. and 0 for t ≥ 10. The results are shown in Figure 1, and we
Since we are only interested in the Mach-hold autopilot see that the controller achieves good performance, in spite
(for V ) as a secondary objective (of minor importance), and of reaching the saturation limits, and even without integral
this is usually designed simply to meet specifications on control. In particular, the relative error is less than 0.4%
steady-state error and disturbance rejection, we only design when the control is not saturated and roughly 8% for the
a simple PI controller for the thrust T to regulate V . The brief period when the control is saturated 4 .
Vδ -dynamics in (2) are of the form In order to demonstrate the performance of the robustness
def
of the SMC approach to matched disturbances, and the effect
V˙δ = aV α αδ + aV q qδ + aV V Vδ + aV θ θδ + bV T Tδ + bV δ δeδ of integral action on the steady-state performance, we repeat
the first simulation, but with an input additive disturbance
and it can be verified that for each trim condition, aV V < 0,
at the elevator input. Note that this disturbance effectively
i.e., the Vδ -subsystem is stable. We view the term aV α αδ +
replaces δe in every equation in (1) by δe +d. Figure 2 shows
aV q qδ + aV θ θδ + bV δ δeδ as constituting a “matched distur-
the simulation results for d = −5, both without integral
bance”, and simply “augment” the stability of this system by
control, and with the conditional integrator, and for the values
designing Tδ as the PI controller
µ = 0.5 and µ = 0.1. The following inferences can be made
Tδ = −kP Vδ − kI σV , σ̇V = Vδ (10) (i) the transient responses of the controllers are good, even
with the disturbance d, with a maximum absolute error of
3 This assumption corresponds to the observation that the lift derivative
is small with respect to the velocity V and can hence be neglected, and is 4 By comparison, max|e| ≈ 25% with the gain-scheduled polynomial
made in [23] right from the start. approach with integral control even when the control is not saturated.
1773
0.6
α is not corrupted by noise, and use the HGO to estimate
the derivative ė required in the ρ = 2 case. The values of
0.4
the HGO parameters are taken as α1 = 15, α2 = 50, and
e = α − αm
0
magnitude of the high-frequency gain with ρ = 1, which
renders the control less effective than with the assumption
−20
ρ = 2.
−40
0 5 10 15
Time (sec)
0.08 0.01
Fig. 1. Tracking errors with SMC without integral control, µ = 0.1. 0.06 ρ=1 0
m
ρ=2
e=α−α
e=α−α
0.04 −0.01
0.02 −0.02
approximately 0.04 (or a relative error of 0.8%) even without
0 −0.03
integral control and with a value of µ as large as 0.5, (ii)
−0.02 −0.04
the steady-state error is non-zero without integral control, 5 5.5 6 10 10.5 11
and (approximately) zero with integral control. In particular,
in the absence of integral control, |e| = O(µ), and we 20 5
e
However, smaller values of µ can induce chattering when
δ
δ
5
−5
there are switching imperfections such as delays, and has
0
been demonstrated by simulations for the case of pitch-rate
−5 −10
control of the same F-16 model as the one in this paper 5 5.5 6 10 10.5 11
in [13]. On the other hand, the inclusion of integral action Time (sec) Time (sec)
means that we don’t need to make µ very small to achieve
small errors, only small enough to stabilize the equilibrium Fig. 3. Effect of relative degree assumption: tracking errors with controller
designs for ρ = 1 and ρ = 2.
point.
In order to demonstrate the effect of measurement noise on
0.045 the HGO, we repeat the previous simulation with α corrupted
by measurement noise, for ρ = 1 and u = ksat k0 σ+e µ ,
0.04
k0 σ+k1 ê+ėˆ
0.035
and with ρ = 2 and u = ksat µ . We use the noise
5
model (9) with fixed coefficient matrices
(see [22]) Aw =
0.03
0 1 0
, Bw = , Cw = [0.0043 0.0262].
e = α − αm
0.025
SMC without integrator, , µ = 0.1 −0.0823 −0.5737 1
SMC with conditional integrator, µ = 0.1
SMC without integrator, µ = 0.5
The simulation results in Figure 4 clearly show that (i) the
0.02
SMC with conditional integrator, µ = 0.5 presence of noise does not degrade the derivative estimate,
0.015 and in fact, as before (ii) the design with ρ = 2 (that uses the
derivative estimate from the HGO) is marginally better (note
0.01
the error at t = 5 and t = 10, when the doublet changes in
0.005 value) than the (differentiation-free) ρ = 1 based design.
Lastly, before we summarize our results, we present our
0
0 1 2 3 4 5 results for velocity tracking with the PI controller, with gains
Time (sec)
kP = 828.4 and kI = 191.2 chosen to assign the roots of the
closed-loop characteristic polynomial λ2 + bV T kP λ + bV T kI
Fig. 2. Tracking errors with input-additive disturbance, with and without
integral control, µ = = 0.1, 0.5. at -0.3 and -1. The desired velocity reference is the output
1
of the first order filter H(s) = s+1 , to which the input is a
Next, we compare the control designs for the ρ = 1 and 5 This is not really an accurate wind turbulence model, since the matrices
ρ = 2 cases, with µ = 0.1, and for a reference doublet depend on velocity, altitude and atmospheric conditions. Our goal is to
of magnitude 5 deg. We assume that the measurement of simply include a noise model.
1774
0.2
analytical results based on a control-affine approximation of
ρ = 1, no HGO the nonlinear system should be possible. Consequently, we
0.15 ρ = 2, HGO (for error & derivative)
believe that the results presented in this paper are a promising
e = α − αm
0 R EFERENCES
4 4.5 5 5.5 6
[1] R.J. Adams, J.M. Buffington, and S.S Banda. Design of nonlinear
control laws for high-angle-of-attack flight. Jnl. Guidance, Control,
0.02
and Dynamics, 17(4):737–745, 1994.
0 [2] J.H. Ahrens and H.K. Khalil. Output feedback control using high-gain
e = α − αm
600
[12] E. Promtun. Sliding Mode Control of F-16 Longitudinal Dynamics.
550
MS. Thesis, San Diego State University, San Diego, USA, 2007.
500
[13] E. Promtun and S. Seshagiri. Sliding mode control of pitch-rate of
450
15 20 25 30 35 40 45 an f-16 aircraft (submitted). In 17th IFAC World Congress, Seoul, S.
Korea, July 2008.
60 [14] W.C. Reigelsperger and S.S. Banda. Nonlinear simulation of a
40 modified F-16 with full-envelope control laws. Control Engineering
20 Practice, 6:309–320, 1998.
e=V −V
−20
Matlab. Technical report, University of Minnesota, June 2003.
−40
[16] R. Rysdyk and A.J. Calise. Robust nonlinear adaptive flight control for
−60
15 20 25 30 35 40 45 consistent handling qualities. IEEE Trans. Aut. Ctrl., 13(6):896–910,
Time (sec)
2005.
[17] S. Seshagiri and H.K. Khalil. Robust output feedback regulation
of minimum-phase nonlinear systems using conditional integrators.
Fig. 5. Velocity (Mach-hold) autopilot response to velocity command: PI
Automatica, 41(1):43–54, 2005.
controller.
[18] J.S. Shamma and J.S. Cloutier. Gain-scheduled bank-to-turn autopilot
design using linear parameter varying transformations. Jnl. Guidance
Control and Dynamics, 9(5):1056–1063, 1996.
V. C ONCLUSIONS [19] Y. Shtessel, J. Buffington, and S. Banda. Tailless aircraft flight control
We have presented a new SMC design for control of the using multiple time scale reconfigurable sliding modes. IEEE Trans.
Ctrl. Sys. Tech., 10(2):288–62, 2002.
angle of attack of an F-16 aircraft, based on the conditional [20] S.A. Snell, D.F. Enns, and Jr. W.L. Garrard. Nonlinear inversion flight
integrator design of [17]. The design exploits the short- control for a supermaneuverable aircraft. Jnl. Guidance, Control, and
period approximation of the linearized flight dynamics. The Dynamics, 15(4):976–984, 1992.
[21] M.S. Spillman. Robust longitudinal flight control design using linear
robustness of the method to disturbances, modeling uncer- parameter-varying feedback. Jnl. Guidance, Control, and Dynamics,
tainties and measurement noise was demonstrated through 23(1):101–108, 2000.
simulation, with the method outperforming, without any [22] B.L. Stevens and F.L. Lewis. Aircraft Control and Simulation. John
Wiley & Sons, Inc., 2 edition, 2003.
scheduling, the transient and steady-state performance of a [23] A. Young, C. Cao, N. Hovakimyan, and E. Lavretsky. An adaptive
conventional gain-scheduled model-following controller. The approach to nonaffine control design for aircraft applications. In AIAA
method is also applicable to control of the pitch-rate. While Guidance, Navigation, and Control Conference and Exhibit, Keystone,
CO, USA, August 2006.
we did not present any analytical results, we believe, that
1775