Professional Documents
Culture Documents
net/publication/322309590
CITATIONS READS
8 1,739
4 authors, including:
Mark B. Tischler
US Army Aviation Development Directorate - AFDD
157 PUBLICATIONS 2,686 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
High Speed Rotorcraft Modeling, Control, and Handling Qualities View project
All content following this page was uploaded by Marit Elisabeth Knapp on 01 March 2019.
Tom Berger †
‡
Mark B. Tischler
U.S. Army Aviation Development Directorate, Moffett Field, CA
§
M. Christopher Cotting
Aaron Marcus ¶
USAF Test Pilot School, Edwards AFB, CA
I. Introduction
he demands of highly maneuverable and responsive modern aircraft have led to configurations that are
T 1
aerodynamically unstable. Performing system identification to obtain accurate bare-airframe models
can be challenging when flight testing must be performed with a control system engaged, thus yielding
highly correlated control surface inputs. This paper explores system identification of the Calspan NF-16D
Variable-stability In-flight Simulator Test Aircraft (VISTA), see Figure 1. The VISTA operates with a
Variable Stability System (VSS) that provides a control architecture allowing the aircraft stability to be
varied with adjustable feedback gains that can produce controlled aircraft responses by suppressing aerody-
namic characteristics (e.g., unstable phugoid and spiral modes) that cause instability. Identification of the
aerodynamic contributions of the individual control surfaces and the overall aerodynamic characteristics of
the VISTA were of primary interest for building an accurate simulation model that behaves identically to
the aircraft. System identification of the VISTA was conducted in collaboration with the USAF Test Pilot
School. The project was to develop a full flight envelope “stitched model”2, 3 to accurately represent the
VISTA bare-airframe flight dynamics and trim characteristics across its flight and loading envelopes.
∗ Research Engineer
† Aerospace Engineer, Senior Member AIAA
‡ Flight Control Technology Group Leader, Senior Scientist, U.S. Army AMRDEC, Associate Fellow AIAA
§ Master Instructor of Flying Qualities, Associate Fellow AIAA
¶ Flight Sciences Technical Expert
1 of 27
In many applications, it is desired to identify the open-loop characteristics of the aircraft or plant when the
system is operating with a controller engaged (closed-loop). The inherent problem with closed-loop data is
that it typically has less information about the open-loop system. When a system is operating closed-loop,
as the VISTA does, the measured responses may not excite the dynamics of the system adequately since
the feedback is meant to suppress undesired modes/characteristics to generate controlled responses (e.g.,
unstable systems). Additionally, output noise and process noise inside the feedback loop become correlated
with the bare-airframe input.3 The data correlation among dynamic variables of the control system and sup-
pression of important dynamics can result in inaccurate parameter identification with erroneous derivatives
that could be highly correlated due to the feedback action.3, 4
Specifically for fixed-wing applications, the aileron-to-rudder interconnect (ARI) present in the control sys-
tem poses a problem for bare-airframe identification because it directly correlates the aileron and rudder.3
When data is generated for system identification purposes of a dynamic multi-input system operating in
closed loop, the feedback causes correlations between the input variables,3, 4 making it difficult or impossible
to determine the input-output relationship associated with each of the individual inputs. This is especially
problematic for system identification from measured input-output data, as it is necessary that the measured
data contain accurate information about the modes of the system which requires determining the portion of
the response attributable to each input. As a result, it may not be possible to use a standard identification
approach5 alone to identify the open loop system parameters independently when a system is operating
closed-loop.
Even for plants that are asymptotically stable, opening the loop for identification may not be feasible due
to operational constraints. Identification of an open loop unstable plant, in which case opening the loop
for identification is prohibited, poses an even greater challenge. Alternate methods of system identification
for unstable systems operating closed-loop have been researched to accurately derive the open-loop system
dynamics.3, 6 In standard open-loop identification, the model is captured for the aerosurface command to
response behavior (y/δA ), as depicted in Figure 2 and demonstrated in Ref. 3. Generally, there are three ap-
2 of 27
When correlation of inputs is high, alternate methods of identification can be used. Closed-loop identi-
fication using the Indirect approach incorporates the feedback (K) loop equations in the identification. To
extract the bare-airframe model from a closed-loop identification model with the Indirect method requires
accurate knowledge of the control system. Alternatively, the Joint Input/Output (JIO) method7, 8, 9 requires
no knowledge of the control system. Instead, the inputs and outputs of the bare-airframe are treated as
outputs of a system driven by uncorrelated piloted inputs. Herein, the JIO method was used to generate
open-loop frequency responses of the MIMO closed-loop system. This additional conditioning of the data
enables identification of bare-airframe derivatives from closed-loop flight data.
This paper covers the development of a full flight envelope stitched model for the VISTA. Section II re-
views the standard frequency-response system identification methodology, shown in Figure 3, along with
additional processing using the JIO method. Next, the point model identification process is presented in
Section III with example results of a point model for one of the flight conditions. A brief background on
the stitched model is given in Section IV followed by the integration of the identified point models to the
stitched model architecture. Each identified model and associated trim data was scaled to a common loading
configuration for the stitched model integration, presented in Section V. Lastly, the models were “stitched”
together to produce a continuous full flight envelope simulation model. Results and analysis of the full flight
envelope stitched model are presented in Section VI.
Conditioned
Transfer-Function Frequency Responses
Modeling &
Partial Coherences
Frequency-Response +
Identification
Algorithm Identification
Criterion –
Sensitivity Analysis
&
Dissimilar flight Model Structure
data not used in Verification Determination
identification
The flight test plan of maneuvers and flight conditions for the aircraft is outlined in Section II.A. The primary
flight data used for identification were frequency sweeps and longitudinal static stability data. Before any
flight testing occurred, data consistency was checked, which is covered in Section II.B. System identification
requires an analyze of the input data for input cross-correlation. Correlation requirements and limitations
are discussed in Section II.C along with evaluation of the VISTA cross-correlation. Based on the level of
3 of 27
1. Aircraft Description
The aircraft used in this study is the Calspan NF-16D VISTA, a modified F-16 Fighting Falcon, single-engine
supersonic multirole fighter aircraft that carries an evaluation pilot and safety pilot, shown in Figure 1. The
lightweight fighter has relaxed static stability, with a cruise speed of 300 KCAS, maximum speed at sea
level of Mach 1.2 and a service ceiling of 50,000 ft. This unique aircraft has a custom digital flight control
computer. With the VSS engaged, the evaluation pilot flies the aircraft from the front cockpit, while the
safety pilot located in the back cockpit manipulates the VSS settings (e.g. freezing certain control surfaces).
Other modifications of the VISTA configuration include a heavy weight landing gear and a larger capacity
hydraulic pump and lines to accommodate the increased surface motions needed for in-flight simulation.
The VISTA is flown with the VSS, which operates symmetric and asymmetric horizontal tails and flap-
erons, rudder, and throttle control. The only surfaces not controlled by the VSS are the leading edge flaps,
which are scheduled with angle of attack. The VSS flight envelope is more constrained than the F-16 en-
velope, limited by the numerous safety trips that turn off the VSS and return nominal F-16 control law
augmentation in the event of exceedance of aircraft or aircrew safety limits. This study focused on the
operational flight envelope with the VSS engaged, which is limited to Mach 0.9 or 440 KCAS. The sensors
available for data collection on the VISTA measure pilots stick and throttle controls (δlat , δlon , δthrt ), ac-
tuator commands and corresponding surface deflections (δa , δe , δr , δf , δh ), air data (α, β, VT AS ), aircraft
states (φ, θ, ψ, p, q, r), and accelerometers (ax , ay , az ), all data signals were sampled at 64 Hz.
4 of 27
5
0
10
0
−10
0
−50
−100
0 20 40 60 80 100 120 140 160
Time [sec]
Figure 5 shows the test flight conditions and maneuvers flown. At each test point used in model identification,
frequency sweep, doublet, longitudinal static stability data and steady heading sideslip data were collected.
Doublet maneuvers were performed in the pitch, roll, and yaw axes for model validation. Additional trim
data was collected during the Longitudinal Static Stability (LSS) maneuver, which produces 10 second trim
records at various airspeeds. The LSS maneuver flight data was used to augment the low-frequency iden-
tification of the longitudinal responses. Steady Heading Sideslip (SHSS) maneuver flight data was used to
verify the low-frequency identification of the lateral/directional responses.
Test point 3*, indicated with the pink diamond in Figure 5, was the anchor point for parameter identification
and was repeated in a heavy and light fuel weight configuration to enable calculations of the moments of
inertia to scale the stitched model.5, 3 Trim shot data were collected in 20 KIAS increments at 20 kft as
presented in Figure 5. Longitudinal and lateral/directional state-space models were identified at all of the
test points 1 − 5, indicated as the Model Identification points, covering the flight envelope shown in Figure 5.
5 of 27
Altitude [ft]
40000
1 2 3$ 4 5
20000
0
0 0:2 0:4 0:6 0:8 1
Mach
Test points and maneuvers to cover full flight envelope (flight envelope depicted for the
Figure 5.
VISTA with VSS engaged).
q = sθ (1)
The angular consistency check was performed by forming the measured attitude-to-angular rate frequency
response extracted from flight data. The consistency of the pitch rate and angle responses was evaluated
with a gain and time delay transfer function (TF) fit:
1q
Ke−τ s = (s) (2)
sθ
over the frequency range of 0.5 − 7.2 rad/sec, where the flight data is indicated in blue and the transfer func-
tion fit is in red in Figure 6. Any scale factor or time shift difference would indicate kinematic inconsistencies
between the responses. The resulting gain and time delay are 1.04 and 15.5 ms respectively, indicating good
kinematic consistency between the two signals. The results of the angular kinematic data consistency check
showed good agreement between the attitude and angular rate in both the time domain and frequency do-
main. Generally, the angular-rate response had higher coherence for the entire frequency range of interest.
The pitch attitude had higher coherence at low frequency so it was included for low frequency range in the
longitudinal identification.
For this application, checking the consistency of the feedback sensors was crucial so that time delays could
be incorporated into the identification. Figure 7 shows the angle of attack consistency check relation:
ẇ = sUo α (3)
6 of 27
20 20
Magnitude [dB]
Magnitude [dB]
10 10
0 0
-10 -10
q=3s -20 w=sU
_ o,
-20
100 TF -t Ke!=s = 1:04e!0:0155s (J = 2:82) 100 TF -t Ke!=s = 1:05e!0:0453s (J = 6:63)
Phase [deg]
Phase [deg]
50 50
0 0
-50 -50
-100 -100
1 1
Coherence [-]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
Figure 6. Pitch rate/attitude consistency check. Figure 7. Angle of attack/vertical velocity consis-
tency check.
The effective time delay on angle of attack response, τα cannot be neglected in the state-space model because
it is one of the outputs. This was taken into account by adding a sensor delay, represented with an equivalent
first-order Padé filter in the CIFER
R
identification set-up. The delayed angle of attack αdelay that is fed
back to the controller is used in the identification with Equation 4.
αdelay −s + 2/τα
(s) = (4)
α s + 2/τα
The delayed angle of sideslip was used for identification with this same method.
7 of 27
The level of control input correlation was determined by coherence and autospectrum evaluation of the in-
puts. The coherence function is a measure of the portion of the output that is linearly attributable to the
input of a frequency response.3 The level of correlation between two signals (x,y) can be evaluated with the
2
coherence function, where low average coherence (i.e., (γx,y )ave ≤ 0.53 ) indicates low correlation. Figure 9
shows the cross-control coherence between asymmetric flaperons and rudder during a closed-loop sweep of
the rudder. The low average coherence over the entire frequency range in Figure 9 during the rudder sweep
1
./2a ;/r : Asym Flaperons /a
0:9
0:8
0:7
0:6
Coherence
0:5
0:4
0:3
0:2
0:1
0
10!1 100 101 102
Frequency [rad/sec]
means the asymmetric flaperons were not heavily correlated with the rudder, indicating Direct MIMO iden-
tification method could be used.
In this study, the ARI on the VISTA flight controller was turned off, but the feedback correlates the rudder
and asymmetric flaperon surface motions. Figure 10 shows that during roll sweeps performed with the asym-
metric flaperons, the two control inputs are highly correlated, indicated by the high cross-control coherence.
Figure 11 shows that the rudder was heavily excited during the asymmetric flaperon sweeps, because its
power spectral density was within 10 − 20 dB of the swept aerosurface.3 During the asymmetric flaperon
roll sweeps, the rudder cannot be conditioned out with conventional multi-input processing because of the
high input correlation.3 As such, the control input correlation would compromise the bare-airframe identifi-
cation in the lateral/directional axis if the Direct method was used. The same analysis was repeated for the
longitudinal axis, where correlation was present between the symmetric flaperons and horizontal tail during
the symmetric flaperon sweep (horizontal tail moves due to α and α̇ feedback).
8 of 27
Magnitude [dB]
0:6
Coherence
!20
0:5
0:4 !30
0:3
!40
0:2
!50
0:1 ./2r ;/a : Rudder /r
0 !60 !1
10!1 100 101 102 10 100 101 102
Frequency [rad/sec] Frequency [rad/sec]
Figure 10. Rudder coherence during asymmetric Figure 11. Comparison of rudder and flaperon
frequency sweep of the flaperons. aerosurface autospectrum during asymmetric fre-
quency sweep of the flaperons.
9 of 27
Magnitude [dB]
Magnitude [dB]
!30
!20
!40
!40
!50
!60 !60
-90 -90
Phase [deg]
Phase [deg]
-180 -180
-270 -270
-360 -360
-450
1 -450
1
Coherence [-]
Coherence [-]
0:5 0:5
q=/e Direct Method r=/r Direct Method
q=/e Joint I/O Method r=/r Joint I/O Method
0 0
10!1 100 101 102 10!1 100 101 102
Figure 12. Pitch rate response from Direct method Figure 13. Yaw rate response from Direct method
compared to the response generated with the Joint compared to the response generated with the Joint
Input/Output method. Input/Output method.
ẋ = Ax + Bu(t − τ )
(6)
y = Cx + Du(t − τ )
The longitudinal model is given by Equation 7 and the lateral/directional model by Equation 8, where the
aerosurface time delays τ are not shown but were included in identification.
u̇ Xu Xw Xq − Wo −g cos(Θo ) u Xδe Xδf XδT
δe
ẇ Zu Zw Zq + Uo −g sin(Θo ) w Zδe Zδf ZδT
= + δf
q̇ Mu Mw Mq 0 q Mδe Mδf MδT
δT
θ̇ 0 0 1 0 θ 0 0 0
q 0 0 1 0 0 0 0 (7)
α 0 1/Uo 0 0 u 0 0 0
δe
ax Xu Xw X q 0 w X X XδT
+ δe δf
= δf
az Z Zw Zq 0 q
Zδe Zδf ZδT
u
δT
u̇ Xu Xw Xq − Wo −g cos(Θo ) θ Xδe Xδf XδT
ẇ Zu Zw Zq + Uo −g sin(Θo ) Zδe Zδf ZδT
10 of 27
The derivatives of the body velocities (e.g., u̇) were reconstructed from inertial measurements for use in
the model identification from inertial data.3 The thrust input δT [lbs] was calculated from the measured
pilot throttle input δt [deg] and modeled with a second order system given in Equation 9.2
ωT2
δT = δt (9)
s2 + 2ζT ωT s + ωT2
Longitudinal static stability data were gathered using the stabilized point technique10 and used for the iden-
tification of the u stability derivatives. The aircraft was trimmed around the identification point (300 KIAS,
20 kft), throttle position was held constant, and airspeed was varied by climbing and descending. Airspeed
was stabilized at three points above and three points below the trim condition (approximately 10 kts apart).
Figure 14 shows the full maneuver time history. Highlighted in red on the time history data are sections
where the airspeed was stabilized. Trim pitch attitude θ, angle of attack α, horizontal tail surface deflection
δe , and throttle δt were averaged over each of the highlight sections. These trim points are plotted against
airspeed to determine the trim gradients with respect to airspeed as shown in Figure 15. The trim gradients
were used to determine stability derivatives (Xu , Zu , and Mu ) in the state-space model identification. This
is done by recognizing that in trim the states are constant, ẋ = 0. Then, the state-space equations can be
used to solve for the u derivatives as a function of the other derivatives and the trim gradients, as is shown
here for the Xu derivative:3
∆u̇ = Xu ∆u + Xw ∆w + (Xq − Wo )∆q − g cos Θo ∆θ + Xδe ∆δe + Xδf ∆δf + XδT ∆δT (10)
11 of 27
800 6
U [ft/sec]
3 [deg]
600
4
400
2
10
3 [deg]
60
0
w [ft/ses]
"w="u = -0.074059 ft/s / ft/s
!10 40
10 20
, [deg]
5 2:5
/e [deg]
0
2
4 "/e ="u = 0.0012908 deg / ft/s
/e [deg]
2 1:5
0 2
/f [deg]
"/f ="u = 0.00056088 deg / ft/s
4 1.8
/f [deg]
2
1.6
0
44
50
/t [deg]
"/t ="u = 0.001123 deg / ft/s
/t [deg]
43
40
30 42
0 50 100 150 200 250 300 350 400 500 550 600 650 700 750
Figure 14. Longitudinal static stability time history. Figure 15. Longitudinal static stability selected
trim data.
Table 1 shows that many of the outputs have duplicate measurements, like az and az2 , where two accelerator
measurements from different locations in the aircraft were used. This redundancy gave the software addi-
tional information to identify the state-space model. The identified parameter values as well as insensitivity
and Cramér-Rao bounds are given in Table 2. All identified parameters have good accuracy and are not
correlated with other parameters as indicated by their low insensitivities (I ≤ 10%) and Cramér-Rao bounds
(CR ≤ 20%). The zero value for Mw shows that the VISTA has no static stability (no static margin), as
expected for the relaxed static stability F-16. Eigenvalues presented in Table 3 show that at 300 KIAS with
a full fuel tank, the longitudinal axis is stable for the short term modes, it is slightly unstable for the long
term, phugoid mode. This process of identifying a model from frequency responses from the flight data was
repeated for the lateral/directional axis and at each flight condition. The eigenvalues from the 300 KIAS
lateral/directional axis flight identification are presented in Table 3. All of the lateral/directional modes are
stable for the 300 KIAS with a full fuel tank loading configuration.
12 of 27
Magnitude [dB]
Magnitude [dB]
Frequency response Frequency response
!20 Model identi-cation !20 Model identi-cation
!30 !40
!40 !60
0 0
Phase [deg]
Phase [deg]
-180 -180
-360 -360
-540 -540
1 1
Coherence [-]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
a) b)
ax =/e az =/e
30 60
Magnitude [dB]
Magnitude [dB]
20
Frequency response Frequency response
Model identi-cation 40 Model identi-cation
10
20
0
!10 0
0 0
Phase [deg]
Phase [deg]
-180 -180
-360 -360
-540 -540
1 1
Coherence [-]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
c) d)
u=/
_ e w=/
_ e
30 50
Magnitude [dB]
Magnitude [dB]
20
Frequency response Frequency response
Model identi-cation 40 Model identi-cation
10
30
0
!10 20
0 0
Phase [deg]
Phase [deg]
-180 -180
-360 -360
-540 -540
1 1
Coherence [-]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
e) f)
Figure 16. a) q/δe b) α/δe c) ax /δe d) az /δe e) u̇/δe f ) ẇ/δe Longitudinal frequency response and
identification results for 300 KIAS, heavy configuration at 20 kft.
13 of 27
14 of 27
1
40
ax [ft/sec2 ]
/lon [inch]
0
!1
!2 -40
10
0
az [ft/sec2 ]
/e [deg]
-40
0
!5 -80
10 15
Flight Data Model
5 10
, [deg]
3 [deg]
0 5
-5 0
20
0
q [deg/sec]
10
w [ft/sec]
0
!50
-10
!100
-20
0 1 2 3 4 5 0 1 2 3 4 5
Figure 17. Time domain verification results of pitch doublet (300 KIAS, 20 kft, heavy loading config-
uration).
of JRMS ≈ 1.0 to 2.0 reflect an acceptable level of accuracy.3 Time validation with doublets was performed
for all of the aerosurfaces of every point model. The RMS fit error costs were all below JRMS = 2, which
gives confidence in the identified models for both the longitudinal and lateral/directional axes.
The stitched model architecture enables scaling of the forces and moments for changes in mass, moments
of inertia, and center of gravity (cg) location through the equations of motion, thus allowing for simulation
of changes in those parameters without explicitly including them as scheduling parameters. The resulting
model is accurate for typical aircraft flight dynamics for the entire maneuvering flight regime, except for
15 of 27
U Uf
Nonlinear
Control input • [identified
specific Variable Description
gravity forces
channels Trim values control derivatives]
(look-up table) (look-up table)
U Total longitudinal body axis velocity
m or msim Uf Filtered velocity
+ Δu BΔu
Pilot commands m or msim or
sim ∆u Control perturbations
m Total
+ + Forces
Nonlinear
∆x State perturbations
U Specific Aero + + (lb)
equations
Trim forces
+ Moments of motion M Mass and inertia matrix
(ft - lb)
Aircraft state Δx AΔx A Dimensional stability derivatives
+
Uf Low-pass U B Dimensional control derivatives
Aircraft state • [identified filter
Trim values stability derivatives]
m Aircraft mass
(look-up table) (look-up table)
I Aircraft inertia matrix
U Uf
Aircraft state
total instantaneous x-body axis velocity state U or the true airspeed for typical angle of attack ranges. This
highlights the need for a fine grid of trim data around each point model to accurately represent the speed
stability derivatives (Xu , Zu , and Mu ). Additionally, the stitched model is able to simulate take-off and
landing through the inclusion of a simple landing gear and spoilers models, and different levels of wind and
turbulence.
Model stitching is accomplished by implementing lookup tables of the aircraft state trim values, control
input trim values, and stability and control derivatives based on the identified point models and trim data.
Perturbation aerodynamic and control forces and moments are determined from the perturbation states ∆x
and controls ∆u. The aerodynamic trim forces and moments are then summed to the perturbation values
to yield the total aerodynamic forces and moments acting on the aircraft in body axes.
16 of 27
Figure 19. LSS data and repurposed LSS data to constant altitude for 300 KIAS at 20 kft.
B. Baselining Models
After individual point models at each flight condition were identified, the models needed to be scaled to a
common loading configuration for implementation in the stitched model. The scaling technique demonstrated
in this section is explained in detail by Ref. 5, and performed in the same manner for the VISTA. The model
scaling technique was tested using the point models from the flight condition, 300 KIAS at 20 kft, at two
loading configuration: full and empty fuel tank.
p=/a
0
Magnitude [dB]
-10
-20
-30
-40
0
Light/Fwd .ight data
Phase [deg]
0:5
0
10!1 100 101 102
Frequency [rad/sec]
Figure 20. Roll rate response to asymmetric flaperon for heavy and light loading configurations.
17 of 27
To test the ability of the stitched model to scale from one loading configuration to another the point model
identification of the light-weight/fwd-cg configuration was scaled to form the stitched model in the heavy-
weight/aft-cg loading configuration of the 300 KIAS flight condition. Figure 21 provided validation of the
scaling technique from the good agreement seen in the dynamic responses between the flight data and scaled
stitched model. All of the point models were scaled to the common loading configuration of heavy-weight/aft-
cg loading configuration for the final stitched model. When the point models were scaled to the common
loading configuration, the trim data at each flight condition was also updated to the common loading con-
figuration.
18 of 27
Magnitude [dB]
-10 20
10
-20
0
-30
-10
-40 -20
-50 -30
0 0
Phase [deg]
Phase [deg]
-180 Light/Fwd Flight Data -180 Light/Fwd Flight Data
Light/Fwd ID Model Light/Fwd ID Model
Heavy/Aft Flight Data Heavy/Aft Flight Data
-360 -360
Heavy/Aft Stitched Model Heavy/Aft Stitched Model
-540 -540
1 1
Coherence [-]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
a) b)
p=/h r=/a
0 -20
Magnitude [dB]
Magnitude [dB]
-10 -30
-20 -40
-30 -50
-40 -60
-50 -70
0 180
Phase [deg]
Phase [deg]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
c) d)
p=/a r=/r
0 -20
Magnitude [dB]
Magnitude [dB]
-10 -30
-20 -40
-30 -50
-40 -60
0 0
Phase [deg]
Phase [deg]
Coherence [-]
0:5 0:5
0 0
10!1 100 101 102 10!1 100 101 102
e) f)
Figure 21. a) q/δe b) w/δf c) p/δa d) r/δr e) p/δh f ) r/δa comparison between scaled stitched model,
flight identified model and flight data at 300 KIAS at 20 kft.
19 of 27
u w q /e /f /T
0 0:08 1 1 0:05 0:17
0 !0:4 2 0 0 0:018
!0:05 !0:6 !1
0 !2 0:016
Z
!0:1 !0:8 !2
!0:15 !1 !2 !3 !4 0:014
#10!4 #10!3
5 10 0 0 0:02 1
Heavy/Aft ID models
0 5 0 Baselined stitched model
M
!1 !0:2 0
!5 0 !0:02
Figure 22. Longitudinal axis stitched model derivatives baselined to heavy-weight/aft-cg loading con-
figuration at 20 kft.
Whereas, the lateral/directional axis flight data was collected during the second half of the flights when the
aircraft was lighter. For this reason, the stitched model and identified parameters in Figure 23 show good
agreement in overall trends, but not exact values because the stitched model was baselined to a heavier
loading configuration than the light loading configurations of the lateral/directional identified models.
20 of 27
4
0:05 10 0 2 !0:5
Y
0 5 !1 0 0 !1
!2 2 !0:5
!0:06 !0:5 0:2
L
!3 1:5 !1
!0:08 !4 1 !1:5 !1 0
#10!3 #10!3
10 0:05 !0:2 0 0 0
!0:4
5 0 !2 !0:05 !0:1
N
Figure 23. Lateral/directional axis stitched model derivatives baselined to heavy-weight/aft-cg loading
configuration at 20 kft.
21 of 27
Figure 24. Repurposed LSS trim data sets compared to trim shot data collected at a constant altitude.
22 of 27
ax [ft/sec2 ]
/lon [inch]
0
0
!1
!2 -40
10
0
az [ft/sec2 ]
/e [deg]
5
-40
0
-80
!5
15 15
10 10
, [deg]
3 [deg]
5 5
0 0
-5 !5
20
0
q [deg/sec]
10
0 w [ft/sec] !50
-10
!100
-20
0 2 4 6 8 0 2 4 6 8
Figure 25. Stitched model pitch doublet validation (300 KIAS, 20 kft, empty loading configuration).
Variation in the roll axis with airspeed is shown in Figure 27. As expected, the roll mode/Dutch roll
mode dipole increases in frequency with increased airspeed.3 Figure 30 shows that the lightly damped Dutch
roll mode becomes less damped with increased speed. From the shift in the yaw responses to the right
(Figure 28), one can deduce that the frequency of the Dutch roll mode frequency increases with speed, which
is also seen in Figure 30 by the increasing natural frequency of the Dutch roll mode.
23 of 27
!10
!10
Magnitude [dB]
Magnitude [dB]
!20
!20
!30
!30
!40
!50 !40
Increasing speed Increasing speed
!60 !50
Phase [deg]
300 KIAS 300 KIAS
360 KIAS 360 KIAS
420 KIAS 420 KIAS
-270 -270
-450 !1 -450 !1
10 100 101 102 10 100 101 102
Figure 26. Pitch rate response identification results Figure 27. Roll rate response identification results
at varying speed. at varying speed.
r=/r
!10
!20
Magnitude [dB]
!30
!40
!50
!60
Increasing speed
!70
200 KIAS
-90 250 KIAS
Phase [deg]
300 KIAS
360 KIAS
420 KIAS
-270
-450 !1
10 100 101 102
Frequency [rad/sec]
4 4
Figure 29. Longitudinal axis eigenvalues. Figure 30. Lateral/directional axis eigenvalues.
24 of 27
Figure 31 suggests that at low Mach number the nondimensional derivatives change very little, but dramatic
changes occur above Mach 0.6. Therefore, the point model identification speeds could be arranged further
apart over the low Mach number range and more finely at higher speeds, corroborating conclusions made in
References 2 and 3.
Next, the ability of the stitched model to extrapolate for different weight and cg was demonstrated. Here,
the stitched model was used to predict the pitch rate frequency response for the full range of loading con-
figurations (from full to empty fuel tank). Figures 32 and 33 show a mixture of heavy and light identified
models and stitched model extrapolated for various loading configurations.
q=/e w=/f
0 60
28400 lbs Stitched model 28400 lbs Stitched model
-10 27387 lbs Heavy ID model 40 27387 lbs Heavy ID model
Magnitude [dB]
Magnitude [dB]
-50 !40
0 0
-90 -90
Phase [deg]
Phase [deg]
-180 -180
-270 -270
-360 -360
-450 !1 0 1 2
-450 !1
10 10 10 10 10 100 101 102
Figure 32. Stitched model pitch rate responses for Figure 33. Stitched model vertical velocity re-
different fuel levels for 300 KIAS at 20 kft. sponses for different fuel levels for 300 KIAS at
20 kft.
25 of 27
VII. Conclusion
This paper covered the development of a full flight envelope simulation model of the NF-16D VISTA from
flight data. The stitched model was developed from discrete point models identified with the control system
engaged. Additional conditioning of the data was performed to deal with highly correlated aerosurface inputs.
A scaling method was implemented to baseline the models to a common loading configuration, making it
possible to interpolate and extrapolate the final stitched model over the full flight envelope for various flight
conditions and loading configurations.
1. Due to the highly correlated aerosurface, decorrelation of aerosurface inputs using Joint Input/Output
method was a necessary and effective method to extract frequency responses for system identification.
These results show it is possible to identify accurate bare-airframe models from closed-loop data with
high levels of correlation.
2. Implementation of the frequency response system identification method produced accurate point models
at each flight condition, which reached up to approximately Mach 0.9. The identify reliable stability
and control derivatives are very reliable as evident in the low Cramér-Rao bounds and low time domain
RMS fit cost of the models.
3. Longitudinal static stability data provided trim gradients of the states and controls for the determi-
nation of speed derivatives, Xu , Zu and Mu . Repurposed LSS data also provided a richer set of trim
data as needed for the stitched model.
4. The final stitched model was formed with stability and control derivatives and trim data baselined to
a common loading configuration. Validation results in the time and frequency domain show excellent
agreement between the stitched model and flight data. These results show the ability of stitched model
to simulate the aircraft over its full flight envelope, capturing the aircraft’s dynamics from low speed
up to high subsonic speeds with five identified point models and trim data.
Acknowledgments
The authors would like to thank Jason Kirkpatrick and Jay Kemper (Calspan Corporation) for their hard
work on the simulator and flight tests, as well the test pilots Roger Roscoe Tanner, Evan Ivan Thomas and
Evil Bill Gray for flying excellent frequency sweeps.
References
1 Hehs, E., “F-16 Designer Harry Hillaker,” Code One Online Articles, Lockheed Martin Aeronautics Company, April 1991.
2 Tobias, E. L. and Tischler, M. B., “A Model Stitching Architecture for Continuous Full Flight-Envelope Simulation of
Fixed-Wing Aircraft and Rotorcraft from Discrete-Point Linear Models,” U.S. Army AMRDEC Special Report RDMR-AF-16-
01, April 2016.
3 Tischler, M. B. and Remple, R. K., Aircraft and Rotorcraft System Identification: Engineering Methods and Flight Test
Business Jet Simulation Model Using a Stitching Architecture,” presented at AIAA Modeling and Simulation Technologies
Conference, January 2017.
6 Jategaonkar, R. V., Unstable Aircraft Identification, Flight Vehicle System Identification: A Time-Domain Methodology,
AIAA, 2015.
7 Hersey, S., Celi, R., Juhasz, O., Tischler, M. B., Rand, O., and Khromov, V., “State-Space Inflow Model Identification
and Flight Dynamics Coupling for an Advanced Coaxial Rotorcraft Configuration,” presented at the American Helicopter
Society 73rd Annual Forum, Fort Worth, TX, May 2017.
8 Forssell, U. and Ljung, L., “Closed-Loop Identification Revisited,” Automatica, Vol. 35, Issue 7, July 1999.
26 of 27
27 of 27