This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Rotorcraft flight control system (FCS) is currently designed in a point-by-point man-
ner using linear principles. Gain scheduling is used to phase-in and phase-out control
effects during flight. The flight control computer is also responsible for controlling struc-
tural loads on the aircraft. This is achieved, in the V-22 for example, by limiting control
surface commands (displacements and rates) produced by pilot stick inputs. Apart from
being a complex task, control rate and position limiting have serious effects on aircraft-
handling qualities. Pilots evaluating FCS with rate limiting usually mark them down. Infact, control rate limiting is a known trigger mechanism for aircraft-pilot coupling (APC)
or pilot-induced oscillation (PIO), which can cause fatalities.
Model Predictive Control (MPC) is a multiinput multioutput digital-control design
methodology [6, 7] that uses a highly intuitive principle. At any time step, the control
inputs to be applied depend on the current state, control objectives to be achieved, and
the current operational constraints. So, a rational procedure is to formulate a constrained
optimization problem at each time step and obtain control inputs by solving it. This has the
advantage of explicitly accounting for time domain control constraints such as rate limits
and state-dependent constraints such as limits on normal acceleration. Frequency domainspecifications can also be brought in through MPC design parameters discussed later. Since
the control objectives and operational constraints are always updated and accounted for,
MPC can provide a holistic approach to rotorcraft flight control. Our aim is to demonstrate
this using realistic control problems.
This chapter is organized as follows. The next section describes the MPC problem
and implementation in detail. For nonlinear systems, the online optimization problem for-
mulated by MPC is generally nonconvex and cannot be solved in real time, especially at
the rates required for flight control. We say that a problem is solvable if there is an anytime
algorithm [13] that terminates in polynomial time for a given accuracy. A procedure fordesigning real-time implementable MPC using linear-parameter varying (LPV) approxi-
mations of rotorcraft dynamics is presented in the next section. Section 3 gives a brief
description of active control of aeromechanical instability using MPC. Our results show
the effectiveness of MPC in suppressing vibrations in the presence of actuator saturation
limits. Section 4 presents the design of an MPC-based flight control system (FCS) for
the XV-15 tilt rotor. FCS design specifications include frequency-domain MIL-F-83300
and MIL-F-8785C requirements and time-domain safety and control limits. MPC-based
FCS was implemented at 50 Hz on a real-time 6-DOF XV-15 simulator at Bell Helicopter.
Results of piloted simulations and evaluations by test pilots are presented. The chapterconcludes with some recommendations for future work. Due to space limitations, some
details have been omitted; a complete version of this chapter can be obtained from the
authors.
2 MODEL PREDICTIVE CONTROL OF NONLINEAR SYSTEMS
This section begins with a description of the control problem and the MPC approach. We
then indicate the computational difficulties associated with MPC and present an algorithm
that can be readily implemented. The algorithm uses a linear parameter varying (LPV) ap-
proximation of the nonlinear system to reduce the computational requirements significantly
from those for nonconvex nonlinear programming to that of convex quadratic program-
ming (QP). Thus, our approach trades off optimality for polynomial-time computability.
The MPC implementation is described in detail. Although the presentation is geared to-
ward rotorcraft applications, the algorithm can be applied to problems in other areas as
well.
2.1 Control Problem
Consider the nonlinear system:
xk+1 = f (xk, uk) (1a)
yk = g(xk) (1b)
zk = h(xk) (1c)
ulbd ≤ uk ≤ uubd, for all k (1d)
zlbd ≤ zk ≤ zubd, for all k (1e)
where xk ∈ IRnx is the state vector, uk ∈ IRnu is the control input, yk ∈ IRny is thesignal that must track a specified command, and zk ∈ IRnz is the vector of signals that
are constrained (the control inputs are not included in this vector). It is possible to make
the constraint bounds ulbd, zlbd, uubd and zubd depend on time, but for simplicity they are
taken as constants. As part of the control problem, we are also given a desired trajectory or
commanded trajectory ykcmd.
Control Problem: Generate a feedback control law so that yk tracks the desired
trajectory ykcmd subject to the system dynamics and constraints given in (1).
MPC solution to this problem is the following: At time k , define the performance
index J as follows:
J (xk, δuk, δuk+1, · · ·, δuk+N −1) = J i + J p + J u (2a)
J i =N −1l=0
l
j=0
yk+j+1cmd − yk+j+1
T Qi
yk+j+1cmd − yk+j+1
(2b)
J p =N −1i=0
yk+i+1cmd − yk+i+1
T Q
yk+i+1cmd − yk+i+1
(2c)
J u = δuT k Rδuk +
N −2
i=0
(δuk+i+1 − δuk+i)T
R (δuk+i+1 − δuk+i) (2d)
and minimize it over control moves subject to the system dynamics and constraints in (1).
Apply the control input uk so obtained, increment k , and repeat the procedure.
The terms J i and J p in (2b) and (2c) can be thought of as discrete versions of integral
error and tracking error penalties over the control horizon. These terms are expected to
yield a proportional-integral (PI) effect in the closed-loop system. Like in classical con-
trol, the weights Qi and Q appearing in (2b) and (2c) can be tuned to obtain satisfactory
steady state and transient tracking performances. The third term J u in (2d) is a penalty for
excessive control rate which smoothes out the control inputs. The control inputs are also
constrained to satisfy the saturation limits (1d).
Note that the performance index requires knowledge of the desired or commanded
trajectory ykcmd over the prediction horizon. When the application is rotorcraft stability
and control augmentation, the commands are pilot stick inputs passed through a command
shaping filter. So, their future values yk+1cmd, · · ·, yk+N cmd needed in the performance index will
not be known at time k. Therefore, we have to make an assumption, and the simplest (in
terms of computations) assumption is that
ykcmd = yk+1cmd = · · · = yk+N cmd .
This may also be reasonable if the horizon length is chosen to be relatively small. For
example, when aircraft dynamics is discretized at 0.02 seconds (50 Hz), a horizon length
N = 10 corresponds to 0.2 seconds (5 Hz), which is comparable to the pilot’s bandwidth
so that changes in the command inputs from its current value of ykcmd during the horizon
of interest can be neglected. We may be able to extrapolate past pilot commands by fitting
polynomials to arrive at better estimates, but this will not be used here. In other applications
such as autopilot design, the trajectory is defined by the controller itself and, hence, will
be known over the prediction horizon. Both problems fit into the MPC framework and the
only difference is in the data used by MPC.
2.2 Linear Parameter-Varying Approximation
When the tracked and constrained variables are written in terms of the control inputs (which
are the unknowns), the optimization problem to be solved online becomes nonconvex and
corresponds to a general nonlinear programming problem. This difficulty is usually cir-
cumvented by linearizing about the current state. The linearized system leads to a convex
quadratic programming (QP) that can be readily solved. However, this procedure is flawedfor a number of reasons: (a) linearization is performed about a nonequilibrium state, but the
linearized system is treated as time-invariant, (b) the quality of approximation provided by
linearization is not known because it depends on the states and their derivatives, (c) plant
model and state estimates have uncertainties, and (d) linearization requires the solution of
nonlinear algebraic equations, which can be difficult.
Here, we use a linear parameter varying (LPV) approximation of the nonlinear sys-
tem:
δxk+1 = A(θk)δxk + B(θk)δuk (3a)δyk = C y(θk)δxk (3b)
δzk = C z(θk)δxk (3c)
xk = xtrim(θk) + δxk, uk = utrim(θk) + δuk (3d)
yk = ytrim(θk) + δyk, zk = ztrim(θk) + δzk (3e)
ulbd ≤ utrim(θk) + δuk ≤ uubd, k = 0, 1, · · ·, N − 1 (3f)
zlbd ≤ ztrim(θk) + δzk ≤ zubd, k = 1, 2, · · ·, N , (3g)
where θk ∈ IRnθ is the scheduling variable (time-varying parameter). This parameter
defines the flight condition, and the set of all its possible values defines the flight envelope.
There are several ways to obtain an LPV approximation, some of which depend on
the application. See [9, 10, 11, 12] for details on LPV systems. A procedure that confirms
to the current practice in aerospace is the following:
The conversion-decel maneuver began at approximately 145 knots, 2500 feet altitude
and 50% power. After bringing the aircraft closer to the landing area, the pilot began
converting at which time the MPC-controller was turned on. Sample results are shown in
Fig. 5. The pilot noted that the pitch axis control was tight and reasonably good, although
the pitch attitude excursions during the maneuver were at times about 4 degrees more than
what he would have liked. Over several runs, he rated MPC performance from good to
excellent.
5 CONCLUSIONS
We have described a model predictive control paradigm that uses a linear parameter-varying
approximation of nonlinear systems instead of on-line linearization. This method has sev-eral advantages, the most notable of which are its quality of approximation and polynomial-
time computability. These features are essential for successful flight control design. As a
demonstration, MPC was implemented on a real-time XV-15 tilt rotor flight simulator.
Even though the simulator math model is highly nonlinear, the implementation performed
very well in piloted simulations. A coordinated turn maneuver was used to evaluate the
multiaxis capability of MPC-based SCAS; while a highly nonlinear conversion maneuver
was used to evaluate the autopilot capabilities of MPC. Test pilots rated MPC performance
from good to excellent in both cases.
MPC-based flight control system design used several MIL-F-83300 andMIL-F-8785C requirements as well as time-domain saturation type constraints. However,
a number of issues such as verification of these requirements were not considered in this
study. We have identified the following areas for future research and development:
1. Further Piloted Simulations and Hardware-in-the-Loop (HIL) Testing of MPC-
based FCS — During the piloted simulations, the test pilots suggested that further simula-
tions are needed for a complete evaluation of MPC-based FCS. This is also the natural step
before HIL testing. Future work should include (a) piloted simulations, (b) implementation
of MPC-based FCS on a flight computer, and (c) piloted simulations with a flight computer
in the loop.
2. Applications of Virtual Tilt Rotor to Predictive Control1 — This is a nonlinear
dynamical model of a tilt rotor that resides in the flight computer and is executed at rates
faster than real time. A high-fidelity virtual tilt rotor model has been developed by Bell
Helicopter. This concept has many applications including predictive control. Future work
should consider (a) MPC implementations with the virtual tilt rotor as the internal model,
and (b) computational requirements for real-time implementation of MPC-based FCS.
3. Multivariable Control of Quad Tilt Rotor (QTR) — QTR is the next generation
of tilt rotors being developed by the rotorcraft industry. The QTR has more control in-puts than the number of equations/states and represents significant control challenges and
opportunities for reconfiguration under control failures.
1Proposed by Dr. Richard Bennett of Bell Helicopter.
This work was funded by NASA Ames Research Center under contract NAS2-98022. We
are grateful to Mr. Stephen Jacklin of NASA Ames for support and encouragement dur-
ing the project. The real-time simulations were flown at Bell Helicopter Textron Inc. by
V-22/XV-15 test pilots Thomas Warren and Roy Hopkins. Their comments and insights
were of great help to us in evaluating the controller as well as in identifying future researchissues. Dr. Richard Bennett, David Neckels, and Dr. Mark Wasikowski of Bell Helicopter
also contributed to this project. We thank them for their efforts.
REFERENCES
1. Military specification flying qualities of piloted airplanes, MIL-F-8785C, November
1980.
2. Military specification flying qualities of piloted V/STOL airplanes, MIL-F-83300, De-cember 1970.
3. W.Y. Chan and I. Chopra, “Aeromechanical stability of hingeless helicopter rotors in
forward flight,” AHS specialists meeting, March 1991.
4. F. Gandhi and W.H. Weller, “Active aeromechanical stability augmentation using fuse-
lage state feedback,” AHS specialists meeting, 1997.
5. D. Hodges, “An aeromechanical stability analysis for bearingless rotor helicopters,”
AHS specialists meeting, 1978.
6. M. Kothare, V. Balakrishnan and M. Morari, “Robust constrained model predictive
control using linear matrix inequalities,” Automatica, Vol. 32, No. 10, 1996.7. D.Q. Mayne, “Constrained model predictive control: stability and optimality,” Auto-
matica, Vol. 36, No. 6, 2000.
8. Yu. Nesterov and A. Nemirovskii, “Interior-point polynomial methods in convex pro-
gramming,” SIAM , Philadelphia, 1994.
9. A. Packard, “Gain scheduling via linear fractional transformations,” Systems and Con-
trol Letter , Vol. 22, 1994.
10. W. Rugh and J. Shamma, “Research on gain scheduling,” Automatica, Vol. 36, No. 9,
2000.
11. J. Shamma and M. Athans, “Guaranteed properties of gain scheduled control for linearparameter varying plants,” Automatica, Vol. 27, 1991.
12. J. Shamma and J.R. Cloutier, “A linear parameter varying approach to gain scheduled
missile autopilot design,” Proceedings of ACC , 1992.
13. S. Zilberstein, “Using anytime algorithms in intelligent systems,” AI Magazine, Fall