pilot model for wake vortex encounter fotofoto...

S. Amelsberg – 2 June 2010WakeNet3 - Europe Specific Workshop: Models and Methods for WVE Simulations

Berlin Institute ofTechnology

FotoFotoPILOT MODEL FOR WAKE VORTEX ENCOUNTER SIMULATIONS FOR TAKEOFF AND DEPARTURE

Name Surname 3. Juni 2010

Introduction

MotivationRisk analysis in with Monte Carlo

Simulation (MCS)

Application: Wake Vortex Encounter (WVE)

Models needed: - Aircraft - Vortex (Evolution, Interaction)-- Pilot Pilot behaviourbehaviour

Pilot modeldeterministic

andprobabilistic

Aircraft simulation model

AircraftreactionPilot inputs

AileronElevatorRudderThrust

Desiredflight path

-Δ

Wake vortex simulation

model-40-30-20-10010203040

-20

-15

-10

-5

0

5

10

15

20

Lamb-OseenBurnham & HallockProctorWinckelmans

Vortex induced accelerations

Elements of MCS

Objectiveo Modelling of human pilot behaviour for tracking and WVE recovery

o Validated pilot model


Content

• Development process

• Simulator data for pilot modelling

• Structure of pilot model

• Neural network part of pilot model

o Topology

o Parameter optimisation

• Validation

• Conclusion


Development process and application

Pilot-in-the-loop Simulator tests

PM

Test evaluation(Time histories &Questionnaires)

Model optimization

Pilot task

Pilot model

Wake vortex

Normal / special procedure

WVE Risk analysis / Hazard rating

PMControl inputs

Monte Carlo Simulation

Wake Vortex

0 10 20 30 40 50 60 70 80-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

time [s]

norm

. SS

RO

[-]

test file: 777-0.tub330.8simulation

Normal / special procedure

FCOM

Flight Crew Operational Manual


Flight phases to model

The variety of tasks requires a pilot model consistent out of many submodules

TRACKING TASK WVE

0 10 20 30 40 50 60 700

2

4

6

8

Ang

le o

f atta

ck [°

]

0 10 20 30 40 50 60 7020

40

60

80

100

True

airs

peed

[m/s

]

0 10 20 30 40 50 60 70-1

-0.5

0

0.5

1

Sid

estic

k ro

ll []

0 10 20 30 40 50 60 70-20

-10

0

10

20

Ban

k an

gle

[deg

]

Disconnect AP in case of high deviation from the nominal flight path

Stabilize the a/c after encountering a wake vortex

Retract flaps, landing gear and reduce thrust setting depending on altitude and speeds

Match the stationary aircraft climb gradient to the actual thrust level

Correct lateral deviations

Climb along the desired flight path

Rotation

Take-off run on the runway centre line

CONTROL TASK


Pilot model

Flap/LG/ Thrust lever

A/Ccy ey c y

AP

Take-off run

Rotation

SID tracking

WV recoverAP disc.model

APc

d

Mc

Flap/LG/ Thrust lever

A/CA/Ccy ey c y

AP

Take-off run

Rotation

SID tracking

WV recoverAP disc.model

APc

d

Mc

AP disconnect button

Sidestick pitchSidestick roll

Throttle lever (TL)Flap lever (FL)Landing gear lever (LGL)

Sidestick pitchSidestick roll

Sidestick pitch

Pedal position

OUTPUT (CONTROL ELEMENT)

AP disconnect modelDisconnect AP in case of high deviation from the flight path

Wake vortex recoveryStabilize the a/c after encountering a wake vortex

Flap, landing gear, thrust lever control

Retract flaps, landing gear and reduce thrust setting depending on altitude and speeds

Match the stationary aircraft climb gradient to the actual thrust

Correct lateral deviations

SID trackingClimb along the desired flight path

RotationRotation

Take-off runTake-off run on the runway centre line

SUB-MODULECONTROL TASK


Neural network

• Training- backpropagation with Matlab (cost function=MSE)

- optimization with MOPS (user-defined cost function), post-processing

Problems: time-consuming calculation for evolutional algorithm because of the high number of parameters (tuners) in neural networks

-stops training by the use of cross validation

division of simulator data: 60% training, 20% validation, 20% test data

• Model approaches- static feed-forward network

- recurrent network

• Design parameters (Topology)a) number/ choice of inputs

b) number of layers

c) type of activation function

f

f

f

f

f

f

f

Inputs Hidden Outputs


Evaluate net performance

84 86 88 90 92 94 96 98 100 102-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1Neural net vs. Pilot

Time

Sid

estic

k ro

ll

Neural netPilot

95 100 105 110 115 120-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1Neural net vs. Pilot

Time

Sid

estic

k pi

tch



Average Angle Measure AAM

Mean square error MSE

( )2*

1( ) ( )

Nk

ny n y n

MSEN

=−∑

=

y* recorded human pilot outputsyk neural network modelled outputsN number of data samples

1

1

** 2 2

( ) ( )41 ;

( )

( ) ( )( ) arccos ; ( ) ( ) ( )

2 ( )

N

nN

n

kk

D n nAAM

D n

y n y nn D n y n y n

D n

α

π

α

=

=

⎡ ⎤⋅∑⎢ ⎥

⎢ ⎥= −⎢ ⎥∑⎢ ⎥⎣ ⎦⎡ ⎤+⎢ ⎥= = +⎢ ⎥⎢ ⎥⎣ ⎦

Correlation coefficient



Dynamic Time Warping (DTW)

• Using dynamic time warping to find patterns in time series• Algorithm for measuring similarity between two sequences which may vary in time and

speed Sequences are warped non-linearity in the time dimension • Measure of their similarities independent of certain non-linear variations in the time

dimension• Target: Find the optimal alignment between two time series, which minimizes the sum of

distances between aligned elements

• While sequences have an overall similar shape, they are not aligned in time axis



Dynamic Time Warping

0 5 10 15 20-2

-1.5

-1

-0.5

0

0.5

1

1.5

2

Zeit [sec]

δ col [g

rad]

CC 0.133AAM 0.092581DTW 0.70023

PilotPM

30 32 34 36 38 40 42 44 46-0.5

0

0.5

δ row [d

eg]

PilotPM

CC -0.15AAM 0.03DTW 0.63

Synthetic case: demonstrate differences between evaluation criteria

• two sinus signals, differ in amplitude and time• No similarity given by CC= 0.13 and AAM=0.09

• The DTW= 0.7 implies a reasonable similarity

CC 0.13AAM 0.09DTW 0.70

Real case:• E.g. phenomenon of variable time delayed reaction

• No similarity given by CC= -0.15 and AAM=0.03

• The DTW= 0.63 implies a reasonable similarity


Chosen net topologies

WVE Mode: Roll axis WVE Mode: Pitch axis

rollSS

rollFD

p

p&

jf

jf

jf

jf

jf

jf

kf

pitchFD

q&

pitchSS

jf

jf

jf

jf

kf

)(

,)(1

jjj

l

ijiijj

zfy

wxbz

=

∑ ⋅+==

Functionality of a neuron

,)(1

∑ ⋅+===

m

jkjjkkk wybzy

Neural net output signal

jy

ky

rollSS

rollFD

p

jf

jf

kf rollSS

rollFD

p

jf jf

jf jf

kfkf

Tracking Mode: Roll axis


Validation

Pilot-in-the-loop Simulator tests

Time histories:A/C statePilot task

Online Validation• A330 simulator (Berlin)• A320 simulator (Hamburg)

Wake vortex (repeatable “fixed encounter”)

PM-in-the-loop Simulator tests

© Technische Universität Berlin. All rights reserved. Confidential and proprietary document. Printed with: TUBsimplot_v1.m18/08/09 11:28

TU Berlin(FMRA)Page No.:2/3

AircraftA/C massXcg

: A330-300: 65970 kg: 25 MAC

Pilot Test EngineerSession

: Pilot A-N: S. Kauertz: 1-14

Vortex data AirportRunw ay

: 212-2511: EDDF: RWY25R

CREDOS Longitudinal 09-07-2009 10:30

0 10 20 30 40 50 60 700

5

10

15

20

Time [s]

Pitc

h an

gle

[deg

]

Pitch angle

0 10 20 30 40 50 60 700

2

4

6

8

10

Time [s]

Ang

le o

f atta

ck [°

]

Angle of attack

0 10 20 30 40 50 60 70-2

-1

0

1

2

3

Time [s]

Pitc

h ra

te [d

eg/s

]

Pitch rate

0 10 20 30 40 50 60 70-6

-4

-2

0

2

4

6

Time [s]Pitc

h ac

cele

ratio

n [d

eg/s

] Pitch acceleration

0 10 20 30 40 50 60 70-1

-0.5

0

0.5

1

Time [s]

Sid

estic

k pi

tch

[]

Sidestick pitch


Validation PM-in-the-Loop

0 10 20 30 40 50 60 70-20

-10

0

10

20

Time [s]

Bank

ang

le [d

eg]

Bank angle

PilotsPM

0 10 20 30 40 50 60 70-6

-4

-2

0

2

4

Time [s]

Sid

eslip

ang

le [°

]

Sideslip angle

PilotsPM

0 10 20 30 40 50 60 70-30

-20

-10

0

10

20

30

Time [s]

Rol

l rat

e [d

eg/s

]

Roll rate

PilotsPM

0 10 20 30 40 50 60 70-60

-40

-20

0

20

40

60

Time [s]

Rol

l acc

eler

atio

n [d

eg/s

]

Roll acceleration

PilotsPM

0 10 20 30 40 50 60 70-1.5

-1

-0.5

0

0.5

1

1.5

Time [s]

Side

stic

k ro

ll []

Sidestick roll

PilotsPM

0 10 20 30 40 50 60 700

5

10

15

20

25

Time [s]

Pitc

h an

gle

[deg

]

Pitch angle

PilotsPM

0 10 20 30 40 50 60 700

2

4

6

8

10

12

Time [s]

Ang

le o

f atta

ck [°

]

Angle of attack

PilotsPM

0 10 20 30 40 50 60 70-6

-4

-2

0

2

4

6

8

Time [s]

Pitc

h ra

te [d

eg/s

]

Pitch rate

PilotsPM

0 10 20 30 40 50 60 70-15

-10

-5

0

5

10

Time [s]Pitc

h ac

cele

ratio

n [d

eg/s

]

Pitch acceleration

PilotsPM

0 10 20 30 40 50 60 70-1

-0.5

0

0.5

1

Time [s]

Sid

estic

k pi

tch

[]

Sidestick pitch

PilotsPM


Validation PM-in-the-Loop

0 10 20 30 40 50 60 70-20

-10

0

10

20

30

40

50

Time [s]

Bank

ang

le [d

eg]

Bank angle

PilotsPM

0 10 20 30 40 50 60 70-10

-5

0

5

10

Time [s]

Sid

eslip

ang

le [°

]

Sideslip angle

PilotsPM

0 10 20 30 40 50 60 70-40

-20

0

20

40

60

Time [s]

Rol

l rat

e [d

eg/s

]

Roll rate

PilotsPM

0 10 20 30 40 50 60 70-100

-50

0

50

100

Time [s]

Rol

l acc

eler

atio

n [d

eg/s

]

Roll acceleration

PilotsPM

0 10 20 30 40 50 60 70-1.5

-1

-0.5

0

0.5

1

1.5

Time [s]

Sid

estic

k ro

ll []

Sidestick roll

PilotsPM

0 10 20 30 40 50 60 70-5

0

5

10

15

20

25

Time [s]

Pitc

h an

gle

[deg

]

Pitch angle

PilotsPM

0 10 20 30 40 50 60 70-2

0

2

4

6

8

10

12

Time [s]

Ang

le o

f atta

ck [°

]

Angle of attack

PilotsPM

0 10 20 30 40 50 60 70-6

-4

-2

0

2

4

6

Time [s]

Pitc

h ra

te [d

eg/s

]

Pitch rate

PilotsPM

0 10 20 30 40 50 60 70-15

-10

-5

0

5

10

15

Time [s]Pitc

h ac

cele

ratio

n [d

eg/s

]

Pitch acceleration

PilotsPM

0 10 20 30 40 50 60 70-1

-0.5

0

0.5

1

Time [s]

Side

stic

k pi

tch

[]

Sidestick pitch

PilotsPM


Conclusion

• Successful implementation of the pilot model into the A320 simulation

• A very good agreement between human pilot and PM (on the basis of a feed-forward neural network) behaviour was achieved in the roll axis

• The deterministic pilot model lies, in most of the cases, inside the standard deviation of human pilots for sidestick roll, roll rate, roll acceleration and bank angle

• The PM based on recurrent networks delivered comparable good results

• The general structure for a PM probabilistic add-on is considered in the A320 simulation


End

Present activities


Adaptive Pilot Model

Objectiveo Only one model for tracking tasks AND compensatory tasks

o no discrete switches, avoid possible instable aircraft reactions

Motivation/ Ideao Approximate human pilot adaptation abilities between tracking- and compensatory

tasks

Input Adaptiver Manuel Control

(Pilot)

Input Adaptation

& Prediction

Controlled Element

Adaptation

TaskAdaptation

Aircraft

Disturbance

∆ OutputInput Adaptiver Manuel Control

(Pilot)

Input Adaptation

& Prediction

Controlled Element

Adaptation

Controlled Element

Adaptation

TaskAdaptation

Aircraft

Disturbance

∆ Output



Approacho Applied on a controlled aircraft (VFW-614) simulation

o Adaptive part is described by a neural networ

o Human pilots diversity

- needed for statistical investigations

- can be covered by different learning rates, required precision of adaptation, variations in time delay for input and output data of the neural net

o An extension on a predictive module (Luenberger observer, Kalman filter, or 2nd neural net) is intended

o The model should be applicable at least for real-time simulations



Problemso Differences between open-loop (static) and closed-loop (dynamic

simulation) neural net training

o Closed-loop adaptation is highly learning rate (gradient) sensitive (aircraft instability)

o Training epoch > 1

o Dead time compensation is needed (error of cost function and inputs contains the dead time of a/c, actuator, sensor, etc)

o Predictive part still missing

Controller AircraftAdaptive PM

WV

Reference Aircraft

yc ye

Δy

yyr

cM c

L/A

Controller AircraftAdaptive PM

WV

Reference Aircraft

yc ye

Δy

yyr

cM c

L/A

pilot model for wake vortex encounter fotofoto...

Documents