See discussions, stats, and author profiles for this publication at: [Link]

net/publication/260676656

Estimating Loss-of-Control: a Data-Based Predictive Control Approach

Conference Paper · August 2011

DOI: 10.2514/6.2011-6408

CITATIONS READS
24 106

3 authors, including:

Vahram Stepanyan
University of California, Santa Cruz
87 PUBLICATIONS   1,152 CITATIONS   

SEE PROFILE

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

SAFE50 View project

Technologies for Airplane State Awareness (TASA) element of NASA ARMD Airspace Technology Demonstration (ATD) Project View project

All content following this page was uploaded by Vahram Stepanyan on 11 March 2014.

The user has requested enhancement of the downloaded file.

 AIAA Guidance, Navigation, and Control Conference AIAA 2011-6408
08 - 11 August 2011, Portland, Oregon

Estimating Loss-of-Control: a Data-Based Predictive

Control Approach

Jonathan S. Barlow1
Stinger-Gaffarian Technologies, NASA Ames Research Center, Moffett Field, CA, 94035

Vahram Stepanyan 2
Mission Critical Technologies, Inc., NASA Ames Research Center, Moffett Field, CA, 94035

and
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

Kalmanje Krishnakumar3
Intelligent Systems Division, NASA Ames Research Center, Moffett Field, CA, 94035

Loss-of-control is a major contributor to aircraft fatalities. Recent work has been done
to develop quantitative criteria for determining loss-of-control from accident time history
data. This work proposes an approach to estimating boundaries on control actions to
provide information to pilots and/or control systems to assist in avoiding loss-of-control
scenarios. Data-based predictive control theory is used to develop an algorithm that finds
the minimum control input that will result in the aircraft exceeding a safe operating
envelope at various minimum time estimates. The calculated minimum control inputs
become a boundary of a set of safe control inputs. With this information, a pilot could
change flying strategy or an autonomous system could schedule controller gains to prevent
the vehicle from exceeding the envelope.

Nomenclature
A = state transition matrix
A1, A2 = convenience variables
ALat = example lateral state transition matrix
ALon = example longitudinal state transition matrix
B = control matrix
Bd = disturbance input matrix
BLat = example lateral control matrix
BLon = example longitudinal control matrix
C = output matrix
D = direct transmission matrix
G,H,K,V = gains of the dynamic output feedback predictive control laws
J = receeding horizon cost function
Klat, Klon = pole-placement state feedback gains
LS, M1, M2 = convenience variables
P1, P2, W = parameters of multi-step ahead input-output predictor model
Q = output error weighting matrix
R = control input weighting matrix

1
Senior Aerospace Research Engineer, SGT inc., NASA Ames Research Center, Mail Stop 269-1 P.O. Box 1
Moffett Field, CA 94035-1000, AIAA Member.
2
Research Scientist, Mission Critical Technologies, Inc., NASA Ames Research Center, Moffett Field, CA, 94035,
AIAA Senior Member.
3
Group Lead, Intelligent Systems Division, NASA Ames Research Center, Moffett Field, CA, 94035, AIAA
Associate Fellow.
1
American Institute of Aeronautics and Astronautics

Copyright © 2011 by the American Institute of Aeronautics and Astronautics, Inc. The U.S. Government has a royalty-free license to exercise all rights under the copyright claimed herein for Go
 RS = control smoothing weighting matrix
f = number of distinct disturbance frequencies
g = generic variable representing control input or system output
n = system order
ni = number of system inputs
no = number of system outputs
s = length of prediction window
t = time
u = control input
ud = disturbance input
uk-1 = convenience variable
w = generic variable representing the length of a super-vector
x = system state
y = system output
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

z = desired future output

θ = pitch angle
φ = bank angle
β = sideslip angle
ρ = number of past data points, also order of predictive controller
ω = start of prediction window
Γ = convenience variable

I. Introduction

L oss-of-control has been the number one contributing factor to fatal airline accidents, and has resulted in more
fatalities than any other factor during the past ten years[1]. Historically, loss-of-control has been determined to
be a factor in an accident by qualitative judgement based upon accident investigation experience. Recent work[2]
has been done to create quantitative loss-of-control criteria to define loss-of-control events. The quantitative criteria
are in the form of five envelopes relating to airplane flight dynamics, aerodynamics, structural integrity, and flight
control use. Typical maneuvers exceed at most one of the five envelopes, maneuvers bordering on loss-of-control
exceed two envelopes, and maneuvers exceeding three envelopes are considered loss-of-control. These criteria
constitute a numerical methodology that can reliably identify key upset characteristics in a flight test or
accident/incident time history[2].
Reference 2 provides several loss-of-control examples, most of which involve system malfunction and/or
inappropriate control inputs. In many of the loss-of-control examples[2], five or more seconds elapse between
exceeding the first envelope and entering a loss-of-control situation. The envelopes are used to quantitatively
determine whether an accident involved loss-of-control, which, in turn, may help future pilots learn to avoid those
situations.
While it is useful to know whether an accident involved loss-of-control, knowing before hand if loss-of-control
is about to occur may help prevent an accident and save lives. This approach was taken for developing a command
validation algorithm[3,4] (CVA) which monitors commands generated by a fly-by-wire (FBW) system for a
rotorcraft. The CVA uses a low-order model to predict future bank and pitch attitude as well as altitude. If the
attitude prediction exceeds altitude dependent limits, the command is deemed to be invalid, and the FBW system is
automatically disengaged. The CVA was designed for safely testing control algorithms in research aircraft where a
safety pilot is present and ready to take over at any moment, but does not provide any information to the pilot other
than disengaging the FBW system.
The goal of the present work is to provide information during flight that will assist a pilot or autonomous system
to avoid loss-of-control. In particular, the goal of the data-based predictive control approach is to provide a
capability of predicting the control actions that result in exceeding the envelope, and how much time remains before
the envelope is exceeded. The envelopes proposed in [2] are used for this paper, but it is envisioned that any
envelope that describes a boundary in one of the aircraft states could reasonably be used.

2
American Institute of Aeronautics and Astronautics
 II. Data-Based Predictive Control Approach

A. Predictive Control Concept

The safest course in aircraft flight is to operate
as far from the edge of the envelope as possible.
However, maneuvering near the edge of the
envelope may be required in certain circumstances
(e.g. recovering after upset, take-off and landing).
Furthermore, uncertainties introduced into the
system, such as through damage or failure, can
change the boundaries of the envelope, and can
make certain control inputs more likely to take the
aircraft out of the envelope. Control limits can be
estimated using data-based predictive control and
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

various minimum time estimates. The control limits

can be represented as an n-dimensional hypercube,
where n is the number of control inputs. The region
within the box can be treated as a safe region.
A concept 2-dimensional display for a system
with lateral and longitudinal control inputs is shown
in Figure 1. Two notional safe regions are shown
for the minimum control input needed to exceed the
envelope within 1 second and 5 seconds. The
current control input is also shown as being within
the green box. The control limits can also be
updated periodically, even in real time[5], as the Figure 1. Conceptual 2-dimensional control limit display.
limits change. This information may provide lead
time for a pilot to change flying strategy or for an autonomous system to schedule controller gains to prevent the
vehicle from exceeding the envelope.

B. State-space and Input-output Representations

Consider an ni-input, no-output system with the system state x(k) and output y(k) given by

x ( k + 1) = Ax ( k ) + Bu( k ) + Bd u d ( k )
(1)
y ( k ) = Cx€( k ) + Du( k ) €

We assume that neither the system model, defined by A , B , Bd , C , and D , nor the initial state of the system,
x(0) , are known, but a set of sufficiently rich and long excitation input u(k) and possibly disturbance–corrupted
€
output data y(k) is available. The disturbance input, u d (k) , if present, is assumed to be a sum of a finite number of
unknown harmonics. Only an upper bound of the€number € €of harmonics
€ € is known.
The representation of the data history can be simplified by the introduction of “super-vector” notation, defined
€ €
by
€ €
⎡ g( k ) ⎤
⎢ ⎥
⎢ g( k + 1) ⎥
gw (k) ≡ (2)
⎢ M ⎥
⎢ ⎥
⎣ g( k + w − 1)⎦

where g will generally represent an output or control input (column) vector, and w is the length of the vector.
For the system in Eq. (1), the output y(k) is dependent on the initial state x(0) and the disturbance inputs
€
u d (k) . Since the disturbance input and the initial state are assumed unknown, it is beneficial to describe the system
using a relationship between the excitation input and disturbance-corrupted output that does not explicitly include
the terms involving the initial state and the disturbances. In [6], the interaction matrix formulation captures this
€ €
€ 3
American Institute of Aeronautics and Astronautics
 input-output relationship, which does not depend on initial states and disturbances. It was shown that the following
relationship holds for excitation input and possibly disturbance-corrupted output,

y s ( k + q) = P1u ρ ( k − p) − P2 y ρ ( k − p) + Wu s+ω ( k ) (3)

when p is selected such that n o ρ ≥ n + 2 f +1 and 0 ≤ ω ≤ ρ , where n is the system order, f is the number of
distinct disturbance frequencies, and the 1 accounts for a constant disturbance if present. In this context, p is the
€
number of past data points, q is the start of the prediction window, and s is the length of the prediction window. A
conservative value for p can be chosen using an upper bound on the order of the system and the number of distinct
disturbance frequencies.€ €

C. Model-Predictive Controller Design

Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

A predictive controller can be designed to minimize the receding-horizon cost function

⎛ T ⎞
⎜[ y s ( k + q) − z s ( k + q)] Q[ y s ( k + q) − z s ( k + q)] ⎟
⎜ T
J ( k ) = +u s+q ( k ) Ru s+q ( k ) ⎟ (4)
⎜ ⎟
⎜ T ⎟
⎝+[ u s+q ( k ) − u s+q ( k − 1)] RS [ u s+q ( k ) − u s+q ( k − 1)] ⎠

where z s (k + q) is the desired output trajectory to be tracked. The output error cost is evaluated over the interval
from time k + q to k + s + q − 1, with a weight matrix of Q . The control input cost is evaluated over the interval
€
from time k to k + s + q − 1, with a weight matrix of R . The control smoothing cost, which is the weighted norm of
the difference between the control input at time k and the previous control input at time k −1, is evaluated over the
€
interval from time k to k + s + q − 1, with weighting matrix RS .
€ € €
€ € €
The future control input history u s+q (k)
€ that minimizes the resultant cost function
€ can be found by combining Eq.
(3) with Eq.
€ (4),
€ and taking the derivative with respect
€ to u s+q (k) , where u s+q (k) becomes[5]

€ u s+q ( k ) = A1u p ( k − p) + A2 y p ( k − p) + M 1z s ( k + q) + M 2 u k−1 (5)

€ €
T
where A1 = −Γ −1W T QP1 , A2 = Γ −1W T QP2 , M 1 = Γ −1W T Q , M 2 = Γ−1( I − LS ) RS ,
€ ⎡ 0 0 0 0⎤
⎢ ⎥ ⎡ u(k − 1) ⎤
(
Γ = R + ( I − LS )
€
T

€
)
RS ( I − LS ) + W T QW , L
S
I
≡ ⎢ r×r
0
⎢ 0 O 0€
0 0⎥ , and u
0⎥
k−1 = ⎢
⎣ 0(s+q−1)r×1
⎥ .
⎦
⎢ ⎥ €
⎣ 0 0 I r×r 0⎦

€ The optimal control law is extracted from the first ni rows of

€ Eq. (5), resulting in the dynamic output feedback form
shown in Eq. (6). €

u( k ) = Gu p ( k − p) + Hy p ( k − p) + Kz s ( k + q) + Vu k−1 (6)

The gains G, H, K, and V are the first r rows of A1 , A2 , M 1 , and M 2 respectively.

€
D. Cost Function Modifications for Control Limits
To determine the control boundary, the goal is to find the minimum control input that will result in the vehicle
€ €
being at the boundary at a specified time €
t. Rather than€specify an entire desired output trajectory z s (k + q) from the
current time until time t, the start q of the prediction window can be chosen as q=t, and the length s of the prediction
window can be chosen as s=1. Choosing the weighting matrix Q to be a diagonal matrix with non-zero elements
only for the k+t time step can produce the same effect, but results in much more computation in practice. Including
€
4
€
American Institute of Aeronautics and Astronautics
 only the output error cost and control input cost results in a first control input u( k ) that is very conservative, and not
representative of the boundary of the control. By including the control smoothing cost, we can cause the control
input to be nearly constant from the current time until the end of the prediction window. Also, by setting the first ni
diagonal elements of RS to zero, the controller is allowed to select any control input, and afterwards remain nearly
€
constant. It can also be shown that setting the first n i diagonal elements of RS to zero results in Vu k −1 being
identically zero, thus simplifying the computation of the control input. The resulting control is the amplitude of a
step input that,
€ if applied, would result in the vehicle being at the boundary of the envelope at time t. It is important
to note that an under-damped second order system will have overshoot, and for a time t that is chosen beyond the
peak overshoot, the system will leave the envelope before t, and return € to the boundary at time € t. Therefore, the
boundary generated is an estimate, not a guarantee of the minimum control that will result in the vehicle leaving the
envelope. However, multiple algorithms each having a unique time estimate t can be used to create multiple limits,
and the innermost box could be considered the safe region of operation.
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

E. Numerical Example
The operation of the algorithm is demonstrated with a simple model of a generic transport aircraft[7] with
decoupled lateral-directional and longitudinal dynamics. The longitudinal dynamics use elevator as the input and
pitch angle and pitch rate as the states, with state-space matricies:

⎡ 0 1 ⎤ ⎡ 0 ⎤
ALon = ⎢ ⎥ , BLon = ⎢ ⎥ . (7)
⎣−2.6923 −0.7322⎦ ⎣−3.5352⎦

The lateral-directional dynamics use aileron and rudder as inputs and roll rate, yaw rate, bank angle and sideslip
angle as states, with state-space matricies:
€
⎡−1.752 0.509 0 −10.1076⎤ ⎡ 6.33 1.735 ⎤
⎢ ⎥ ⎢ ⎥
0.024 −0.3842 0 2.7698 ⎥ 0.0351 −2.2464 ⎥
ALat = ⎢ , BLat = ⎢ . (8)
⎢ 1 0.0506 0 0 ⎥ ⎢ 0 0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0.0475 −0.9907 0.0404 −0.1578 ⎦ ⎣ 0.0042 0.0476 ⎦

A second lateral-directional model is also used as an off-nominal model. For this example, the off-nominal
model is a damaged aircraft with 25% tip loss of the left wing. The same longitudinal matrices are used, and the
€
modified lateral dynamics are

⎡−1.134 0.4811 0.0001 −9.5195⎤ ⎡ 3.3953 1.7538 ⎤

⎢ ⎥ ⎢ ⎥
0.0636 −0.4107 0 2.9274 ⎥ −0.0069 −2.3722⎥
Aʹ′Lat = ⎢ , Bʹ′Lat = ⎢ . (9)
⎢ 1 0.0892 0 0 ⎥ ⎢ 0 0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0.086 −0.9903 0.0404 −0.1451⎦ ⎣ 0.0054 0.0495 ⎦

State-feedback control is applied to the longitudinal dynamics with gain K Lon = [−1.0064 −0.7928] , to give
the closed-loop
€ longitudinal transfer function a frequency of 2.5 rad/s and damping ratio 0.707. State feedback is
also applied to the lateral-directional channel with gain

⎡−0.0209 0.5414 0.2241€ −1.6017⎤

K Lat = ⎢ ⎥ , (9)
⎣ 0.0482 −1.7509 0.0131 0.117 ⎦

to give the closed-loop lateral-directional transfer function spiral and roll modes time constants of 1 and 0.2857
seconds, respectively, and to give the dutch roll mode a frequency of 1.2 rad/s and damping ratio 0.707. For
simplicity, this example € assumes the sideslip angle (β) command is always zero, and only bank angle is available to
the predictive control algorithm. The envelope boundaries are set at θ=-10° and 25° and φ=±45°, and the control
position limits are set at –20° and 30° for elevator deflection and ±20° for rudder and aileron deflection. The system
is discretized with sampling frequency 100hz.
5
American Institute of Aeronautics and Astronautics
 For this example, separate model-predictive control algorithms are designed for the longitudinal and lateral-
directional axes. For the longitudinal axis, the order of the system is 2, the number of outputs m is 1, and there are
no disturbances, so the minimum p that can be used for the longitudinal channel is p=3. For the lateral axis, the
order of the system is 4, the number of outputs is 1, and the minimum p that can be used is p=5. If only an estimate
of the order of the system and an upper bound on the number of disturbance frequencies is known, then p can be
chosen conservatively. For this example, p is selected to be 10 for both axes as such a conservative choice. As
discussed previously, the length s of the prediction window is selected as s=1, and the values for q used correspond
to 1 and 3 seconds, or q=100 and 300. The output error weighting matrix Q is a single value and chosen to be
Q =1, and the control input weighting matrix R is a diagonal matrix with diagonal elements 0.000001. The control
smoothing weighting matrix RS is a diagonal matrix with first diagonal value 0, and subsequent values 10. A rich
set of control inputs and system outputs was collected for the system, € and the controller gains G, H, and K were
€ computed for each value of q. €
Both the nominal and off-nominal systems were then simulated with the same pre-recorded sequence of pilot
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

€ system. Figures 2 through 4 show a series of snapshots of the control inputs and control
inputs applied to each
limits during six seconds of simulation. A two-second trace of the control history is shown as a continuous black
line, with the earliest point marked as a black dot and the latest point the time at which the snapshot was taken. The
boundaries shown are also the boundaries for the time at which the snapshot was taken. In Figure 2, the aircraft is
flying wings-level, and the reduced control boundary for the off-nominal condition is clearly visible. The 1 second
control boundary for the off-nominal aircraft is less than the 3 second control boundary. This is consistent with the
behavior of an underdamped system. Although the expectation is that a longer time horizon requires less control
input to reach the boundary, in an underdamped system the aircraft will overshoot before settling back at a level
which is within the boundaries of loss-of-control. This illustrates the importance of having multiple boundaries, as
the vehicle would actually leave the safe boundary in one second, and return to the boundary after three seconds. In
Figure 3, the aircraft has rolled slightly to the right, and the 1 second boundary has now saturated at full left lateral
control input, thus indicating that the plane will not leave the boundary within one second, even with full left stick
deflection. Figure 4 shows the condition after the aircraft has rolled to the left, showing the reduced control
boundary and indicating that the aircraft is nearing the boundary.

III. Conclusion
This work allows estimation of the control boundaries that permit safe operation of an aircraft. This work is
envisioned to assist pilots or autonomous systems in predicting and preventing loss-of-control events. Because the
algorithm is based on input-output data, the boundaries can be updated for the current aircraft and flight condition.
Alternatively, the gains of the algorithm can be precomputed and scheduled with the flight control gains. The
control boundaries may help to reduce inappropriate pilot actions, and may help alert the pilot to reduced control
authority.

References
1
“Statistical Summary of Commercial Jet Airplane Accidents – Worldwide Operations 1959-2009”, p. 23, Aviation Safety,
Boeing Commercial Airplanes, Seattle, WA, USA, 2010.
2
Wilborn, J. E. and Foster, J. V., “Defining Commercial Transport Loss-of-Control: A Quantitative Approach,” Proceedings
of the AIAA Flight Mechanics Conference and Exhibit, Providence, RI, August 2004.
3
Gubbels, A. W., Carignan, S. J.R.P. and Ellis, D. K., “The NRC Bell 412 Advanced Systems Research Aircraft - Facility
Description and Results of Safety System Flight Tests,” Proceedings of the AIAA Modeling and Simulation Technologies
Conference, Denver, CO, August 2000.
4
Gubbels, A. W., Carignan, S. J.R.P. and Ellis, D. K., “The NRC Bell 412 Advanced Systems Research Aircraft - A New
Facility for Airborne Simulation,” Proceedings of the 1st AIAA Aircraft, Technology, Integration, and Operations Forum, Los
Angeles, CA, October ,2001
5
Barlow, J. S. and Phan, M. Q., “Direct Data-Based Model Predictive Controller Design,” Proceedings of the AIAA
Guidance, Navigation, and Control Conference and Exhibit, Keystone, Colorado, August, 2006.
6
Darling, R. S. and Phan, M. Q., “Dynamic Output Feedback Predictive Controllers for Vibration Suppression and Periodic
Disturbance Rejection,” Proceedings of the 14th AAS/AIAA Space Flight Mechanics Meeting, Maui, HI, February 2004.
7
T. Jordan and W. Langford, “Development of a Dynamically Scaled Generic Transport Model Testbed for Flight Research
Experiments,” In Proceedings AUVSI Unmanned Systems North America, Arlington, VA, 2004.

6
American Institute of Aeronautics and Astronautics
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

Figure 2. Control input history and control limits, t=0 to 2 seconds.

The aircraft is flying wings-level, and the reduced control boundary for the off-nominal
condition is clearly visible.

7
American Institute of Aeronautics and Astronautics
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

Figure 3a. Control input history and control limits, t=2 to 4 seconds.
The aircraft has rolled slightly to the right, and the 1 second boundary has now saturated at full
left lateral control input, thus indicating that the plane will not leave the boundary within one
second, even with full left stick deflection.

Figure 3b. Aircraft Response, t=2 to 4 seconds.

The aircraft roll angle phi versus time, showing the aircraft has rolled slightly to the right

8
American Institute of Aeronautics and Astronautics
Downloaded by NASA AMES RESEARCH CENTER on March 11, 2014 | [Link] | DOI: 10.2514/6.2011-6408

Figure 4a. Control input history and control limits, t=4 to 6 seconds.
After the aircraft has rolled to the left, showing the reduced control boundary and indicating
that the aircraft is nearing the boundary.

Figure 4b. Aircraft Response, t=4 to 6 seconds.

The aircraft roll angle phi versus time, showing the aircraft has rolled to the left
9
American Institute of Aeronautics and Astronautics

View publication stats