4CM00: Control Engineering

FeedForward control
Dr.ir. Gert Witvoet

September 22, 2015

Where innovation starts
Influence of the setpoint 2/33

Consider a simple motion system

F
m

x
Objective: Move the mass move from position 0 to 1
Question 1: How does the error depend on the setpoint trajectory xs ?
I.e. how do we get from 0 to 1?
Question 2: Can we obtain (nearly) perfect tracking using just a simple
PD-controller?

/w
Closed loop response: 0th order setpoint 3/33

1
1.6

0.8
1.4

0.6
1.2

1 0.4
setpoint

error
0.8 0.2

0.6 0

0.4 −0.2

0.2 −0.4

0 −0.6
0 2 4 6 8 10 12 0 2 4 6 8 10 12
time time

/w
Closed loop response: 0th order setpoint 3/33

1
1.6

0.8
1.4

0.6
1.2

1 0.4
setpoint

error
0.8 0.2

0.6 0

0.4 −0.2

0.2 −0.4

0 −0.6
0 2 4 6 8 10 12 0 2 4 6 8 10 12
time time
Maximal error / overshoot: 0.6

/w
Closed loop response: 1st order setpoint 4/33

0.01

1 0.008

0.006
0.8
0.004

0.002
setpoint

0.6

error
0

0.4 −0.002

−0.004
0.2
−0.006

0 −0.008

−0.01
0 2 4 6 8 10 12 0 2 4 6 8 10 12
time time

/w
Closed loop response: 1st order setpoint 4/33

0.01

1 0.008

0.006
0.8
0.004

0.002
setpoint

0.6

error
0

0.4 −0.002

−0.004
0.2
−0.006

0 −0.008

−0.01
0 2 4 6 8 10 12 0 2 4 6 8 10 12
time time
Maximal error / overshoot: 0.009

/w
Closed loop response: 2nd order setpoint 5/33

−4
x 10

1 4

3
0.8
2

1
setpoint

0.6

error
0

0.4
−1

−2
0.2
−3

0 −4

0 2 4 6 8 10 12 0 2 4 6 8 10 12
time time

/w
Closed loop response: 2nd order setpoint 5/33

−4
x 10

1 4

3
0.8
2

1
setpoint

0.6

error
0

0.4
−1

−2
0.2
−3

0 −4

0 2 4 6 8 10 12 0 2 4 6 8 10 12
time time
Maximal error / overshoot: 0.0004
/w
Trajectory planning 6/33

Very good tracking results can be obtained

I using a simple (low bandwidth) controller
I using a ‘smart’ trajectory for xs
I basic idea: shape trajectory → restrict ẋs , ẍs , etc

/w
Trajectory planning 6/33

Very good tracking results can be obtained

I using a simple (low bandwidth) controller
I using a ‘smart’ trajectory for xs
I basic idea: shape trajectory → restrict ẋs , ẍs , etc

Advantages of trajectory planning:

I less oscillations
I possibly faster settling times
I smaller control signals (less saturation issues)

/w
Trajectory planning 6/33

Very good tracking results can be obtained

I using a simple (low bandwidth) controller
I using a ‘smart’ trajectory for xs
I basic idea: shape trajectory → restrict ẋs , ẍs , etc

Advantages of trajectory planning:

I less oscillations
I possibly faster settling times
I smaller control signals (less saturation issues)

One can wonder...

I Can we quantify the tracking error?
I Is there a way to further reduce this tracking error?

/w
2nd order trajectory 7/33

For now, assume only xs , ẋs and ẍs are restricted

1
xs
0.5

0
0 2 4 6 8 10 12

0.2

0.15
ẋs

0.1

0.05

0
0 2 4 6 8 10 12

0.05
ẍs

−0.05

0 2 4 6 8 10 12
t

/w
Tracking example 1: error dynamics 8/33

Equation of motion: F
m

m ẍ = F x
PD-controller:

F = Kp e + Kv ė
= Kp (xs − x) + Kv (ẋs − ẋ)

Hence

m ẍ + Kv (ẋ − ẋs ) + Kp (x − xs ) = 0
m (ẍs − ẍ) + Kv (ẋs − ẋ) + Kp (xs − x) = m ẍs

Using e = xs − x, the tracking error dynamics becomes

m ë + Kv ė + Kp e = m ẍs
/w
Tracking example 1: error dynamics 9/33

m ë + Kv ė + Kp e = m ẍ
| {z } |{z}s
mass - spring - damper force on mass

Does e go to zero?

Kv mẍs
m
Kp
e

I Iff ẍs = 0 then e → 0

I Tracking error whenever ẍs 6= 0
Whenever the mass needs to be accelerated, there is a tracking error!

/w
Tracking example 1: error dynamics 10/33

−20
Kv mẍs
m −25
Kp
e −30

1/Kp

Magnitude [dB]
−35

m ë + Kv ė + Kp e = u −40
q
Kp
−45 m

1
I DC-gain: Kp
−50

q
Kp −55
I Eigenfrequency: m
−60
I u = m ẍs 10
−1
10
0

Frequency [Hz]
1
10

/w
Tracking example 1: error dynamics 10/33

Kv mẍs mẍs /Kp

m 0.1
error
Kp
e 0.05

m ë + Kv ė + Kp e = u

error e
0

1
I DC-gain: Kp
−0.05

q
Kp
I Eigenfrequency: m
−0.1

I u = m ẍs 0 2 4 6
time t
8 10 12

/w
Tracking example 2: error dynamics 11/33

Suppose the mass encounters friction:

b F
m

e
Equation of motion:

m ẍ + b ẋ = Kp (xs − x) + Kv (ẋs − ẋ)

| {z }
PD-controller
Hence, the error dynamics becomes

m (ẍs − ẍ) + (Kv + b ) (ẋs − ẋ) + Kp (xs − x) = m ẍs + b ẋs

m ë + (Kv + b ) ė + Kp e = m ẍs + b ẋs
/w
Tracking example 2: error dynamics 12/33

m ë + (Kv + b ) ė + Kp e = m ẍ + b ẋ
| {z } | s {z }s
mass - spring - damper force on mass

Does e go to zero?

Kv +b
mẍs +bẋs
m
Kp
e
I Iff ẍs = 0 and ẋs = 0, then e → 0
I Tracking error whenever ẍs 6= 0 and/or ẋs 6= 0

/w
Tracking example 2: error dynamics 13/33

−25
Kv +b
mẍs +bẋs
m −30
Kp
e −35
1/Kp

Magnitude [dB]
m ë + (Kv +b ) ė + Kp e = u
−40
q
Kp
m
−45

1
I DC-gain: Kp
−50

q
Kp −55
I Eigenfrequency: m
−60
I u = m ẍs + b ẋs 10
−1
10
0

Frequency [Hz]
1
10

/w
Tracking example 2: error dynamics 13/33

Kv +b 0.12 (mẍs + bẋs ) /Kp

mẍs +bẋs error
m
0.1
Kp
e 0.08

0.06

m ë + (Kv +b ) ė + Kp e = u 0.04

error e
0.02

1
I DC-gain: Kp
−0.02

q −0.04
Kp
I Eigenfrequency: m −0.06

I u = m ẍs + b ẋs 0 2 4 6
time t
8 10 12

/w
Error dynamics 14/33

How can we decrease this error?

/w
Error dynamics 14/33

How can we decrease this error?

I Using feedback
• increase Kp

/w
Error dynamics 14/33

How can we decrease this error?

I Using feedback
• increase Kp
0.12 error with Kp = 40
error with Kp = 200
0.1

0.08

0.06

0.04
error e

0.02

−0.02

−0.04

−0.06

0 2 4 6 8 10 12
time t

There is still a (steady-state) error. . .

/w
Error dynamics 14/33

How can we decrease this error?

I Using feedback I Using feedforward

• increase Kp
m ë + Kv ė + Kp e = m ẍ
|{z}s
0.12 error with Kp = 40

0.1
error with Kp = 200 is known!
0.08

0.06
I Given xs , can we compensate
0.04 the error beforehand?
error e

0.02

−0.02

−0.04

−0.06

0 2 4 6 8 10 12
time t

There is still a (steady-state) error. . .

/w
Feedback vs feedforward 15/33

Feedback:
I Controller responds to an error
I Error thus already exists
I Feedback = looking back
• you’re always lagging behind

/w
Feedback vs feedforward 15/33

Feedback:
I Controller responds to an error
I Error thus already exists
I Feedback = looking back
• you’re always lagging behind

/w
Feedback vs feedforward 15/33

Feedback:
I Controller responds to an error
I Error thus already exists
I Feedback = looking back
• you’re always lagging behind

Feedforward:
I Use knowledge about the input
I Correct before the error can occur
I Feedforward = looking ahead
• anticipate what happens or what is needed

/w
The inverse problem (1) 16/33

How does feedforward work?

F
m

I Goal: zero error for t → ∞ for all xs :

m ë + Kv ė + Kp e = 0

/w
The inverse problem (1) 16/33

How does feedforward work?

F
m

I Goal: zero error for t → ∞ for all xs :

m ë + Kv ė + Kp e = 0

I Since m ẍ = F , we can obtain this when

F= m ẍ −Kv ė − Kp e
|{z}s | {z }
feedforward PD-controller
/w
Feedforward in the loop (1) 17/33

Note that xs is known and

L {m ẍs } = ms 2 L {xs }

/w
Feedforward in the loop (1) 17/33

Note that xs is known and

L {m ẍs } = ms 2 L {xs }

ms 2

xs 1 x
K p  Kv s
ms 2

/w
Feedforward in the loop (1) 17/33

Note that xs is known and

L {m ẍs } = ms 2 L {xs }

ms 2

xs 1 x
K p  Kv s
ms 2

Forward feed / prescribe in advance . . .

I . . . the force necessary to accelerate the rigid mass.

/w
The inverse problem (2) 18/33

b F
m

e
I Goal: zero error for t → ∞ for all xs :

m ë + Kv ė + Kp e = 0

/w
The inverse problem (2) 18/33

b F
m

e
I Goal: zero error for t → ∞ for all xs :

m ë + Kv ė + Kp e = 0

I Since m ẍ + b ẋ = F , we can obtain this when

F = m ẍs + b ẋs −Kv ė − Kp e

| {z } | {z }
feedforward PD-controller

/w
Feedforward in the loop (2) 19/33

Note that xs is known and

L m ẍs + b ẋs = ms 2 + bs L {xs }


/w
Feedforward in the loop (2) 19/33

Note that xs is known and

L m ẍs + b ẋs = ms 2 + bs L {xs }


ms2 +bs

1 x
xs
Kp +Kv s ms2+bs

/w
Feedforward in the loop (2) 19/33

Note that xs is known and

L m ẍs + b ẋs = ms 2 + bs L {xs }


ms2 +bs

1 x
xs
Kp +Kv s ms2+bs

Forward feed / prescribe in advance . . .

I . . . the force necessary to accelerate the rigid mass,
I . . . the force required to overcome the friction.
/w
Pure feedforward 20/33

In theory, a feedback controller is not necessary:

xs x
H -1 (s) H(s)

Ideal feedforward:
I Cff (s) = H −1 (s)
I x = H (s)H −1 (s)xs = xs
I use the plant inverse to compute the plant input in advance
x

/w
Combination feedback and feedforward 21/33

In practice, feedback is needed to compensate disturbances around xs :

H -1 (s)

xs x
C(s) H(s)

/w
Combination feedback and feedforward 21/33

In practice, feedback is needed to compensate disturbances around xs :

H -1 (s)

xs x
C(s) H(s)

However,
I what if H −1 (s) is unstable?
I what if H (s) is unknown?
Then we try to approximate H −1 (s) in time domain.
/w
General feedforward (1) 22/33

Forward feeding the force required to. . .

I . . . accelerate the rigid body mode of the plant: Fa = Kfa ẍs
I . . . overcome the viscous friction with the fixed world: Fv = Kfv ẋs

/w
General feedforward (1) 22/33

Forward feeding the force required to. . .

I . . . accelerate the rigid body mode of the plant: Fa = Kfa ẍs
I . . . overcome the viscous friction with the fixed world: Fv = Kfv ẋs

xs
K fa

x s
K fv

xs x
C(s) H(s)

/w
Coulomb friction 23/33

What if Coulomb friction (dry friction) is present?

I constant force, opposing direction of motion
I direction dependent

Fc

ẋ

−Fc

Hence, to overcome this force in advance, we need a feedforward

Fc = Kfc · sign (ẋs )

/w
General feedforward (2) 24/33

sign( x s )
K fc

xs
K fa

x s
K fv

xs x
C(s) H(s)

/w
General feedforward (2) 24/33

sign( x s )
K fc

xs
K fa

x s
K fv

xs x
C(s) H(s)

Fff = Kfa ẍs + Kfv ẋs + Kfc sign (ẋs )

/w
Error profile 25/33

Feedforward tuning is based on error profile in time domain

error

time

forward motion backward motion

There is an error due to:

/w
Error profile 25/33

Feedforward tuning is based on error profile in time domain

error

time

forward motion backward motion

There is an error due to:

I lack of force to accelerate (Kfa )

/w
Error profile 25/33

Feedforward tuning is based on error profile in time domain

error

time

forward motion backward motion

There is an error due to:

I lack of force to accelerate (Kfa )
I lack of force to compensate viscous friction (Kfv )

/w
Error profile 25/33

Feedforward tuning is based on error profile in time domain

error

time

forward motion backward motion

There is an error due to:

I lack of force to accelerate (Kfa )
I lack of force to compensate viscous friction (Kfv )
I lack of force to compensate Coulomb friction (Kfc )
/w
Feedforward tuning procedure 26/33

I Define a suitable (2nd or higher order) setpoint trajectory

• both forward and backward movement (with a small pause)
• rather ‘challenging’ (it must generate some error profile)
• approx 1/3 acceleration, 1/3 constant velocity, 1/3 deceleration

/w
Feedforward tuning procedure 26/33

I Define a suitable (2nd or higher order) setpoint trajectory

• both forward and backward movement (with a small pause)
• rather ‘challenging’ (it must generate some error profile)
• approx 1/3 acceleration, 1/3 constant velocity, 1/3 deceleration
I Monitor the error
• the error should be large enough
• each feedforward component should be recognizable
• maybe feedback bandwidth should be lowered?

/w
Feedforward tuning procedure 26/33

I Define a suitable (2nd or higher order) setpoint trajectory

• both forward and backward movement (with a small pause)
• rather ‘challenging’ (it must generate some error profile)
• approx 1/3 acceleration, 1/3 constant velocity, 1/3 deceleration
I Monitor the error
• the error should be large enough
• each feedforward component should be recognizable
• maybe feedback bandwidth should be lowered?
I Tune Kfc
• deceleration forward movement and acceleration backward
movement should have same error

/w
Feedforward tuning procedure 26/33

I Define a suitable (2nd or higher order) setpoint trajectory

• both forward and backward movement (with a small pause)
• rather ‘challenging’ (it must generate some error profile)
• approx 1/3 acceleration, 1/3 constant velocity, 1/3 deceleration
I Monitor the error
• the error should be large enough
• each feedforward component should be recognizable
• maybe feedback bandwidth should be lowered?
I Tune Kfc
• deceleration forward movement and acceleration backward
movement should have same error
I Tune Kfv
• remove slopes during acceleration and deceleration
• no error during constant velocity phase when Kfc and Kfv are correct

/w
Feedforward tuning procedure 26/33

I Define a suitable (2nd or higher order) setpoint trajectory

• both forward and backward movement (with a small pause)
• rather ‘challenging’ (it must generate some error profile)
• approx 1/3 acceleration, 1/3 constant velocity, 1/3 deceleration
I Monitor the error
• the error should be large enough
• each feedforward component should be recognizable
• maybe feedback bandwidth should be lowered?
I Tune Kfc
• deceleration forward movement and acceleration backward
movement should have same error
I Tune Kfv
• remove slopes during acceleration and deceleration
• no error during constant velocity phase when Kfc and Kfv are correct
I Tune Kfa
• minimize error during acceleration and deceleration

/w
Feedforward tuning procedure 26/33

I Define a suitable (2nd or higher order) setpoint trajectory

• both forward and backward movement (with a small pause)
• rather ‘challenging’ (it must generate some error profile)
• approx 1/3 acceleration, 1/3 constant velocity, 1/3 deceleration
I Monitor the error
• the error should be large enough
• each feedforward component should be recognizable
• maybe feedback bandwidth should be lowered?
I Tune Kfc
• deceleration forward movement and acceleration backward
movement should have same error
I Tune Kfv
• remove slopes during acceleration and deceleration
• no error during constant velocity phase when Kfc and Kfv are correct
I Tune Kfa
• minimize error during acceleration and deceleration
I If desired, tune up the feedback controller
/w
Feedforward tuning 27/33

0.025

0.02 I No FF
0.015

0.01

0.005
error

−0.005

−0.01

−0.015

−0.02

−0.025
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time

/w
Feedforward tuning 27/33

0.025

0.02 I No FF
0.015 I Kfc
0.01
tuned
0.005
error

−0.005

−0.01

−0.015

−0.02

−0.025
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time

/w
Feedforward tuning 27/33

0.025

0.02 I No FF
0.015 I Kfc
0.01
tuned
0.005
I Kfv
tuned
error

−0.005

−0.01

−0.015

−0.02

−0.025
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time

/w
Feedforward tuning 27/33

0.025

0.02 I No FF
0.015 I Kfc
0.01
tuned
0.005
I Kfv
tuned
error

0
I Kfa
−0.005 tuned
−0.01

−0.015

−0.02

−0.025
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time

/w
Feedforward tuning 27/33

0.025

0.02 I No FF
0.015 I Kfc
0.01
tuned
0.005
I Kfv
tuned
error

0
I Kfa
−0.005 tuned
Maximal error reduced from 2.5 · 10−2 to
−0.01
3.5 · 10−4
−0.015

−0.02

−0.025
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time

/w
Integral controller action 28/33

To tune feedforward, don’t use a high performance feedback controller!

With PD-controller
With PID-controller
0.1

0.05
error e

−0.05

−0.1

0 5 10 15 20 25
time t

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

I Stability

• Can a stable closed loop be destabilized by feedforward?

• Can an unstable closed loop be stabilized by feedforward?

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

I Stability

•Can a stable closed loop be destabilized by feedforward? No.

•Can an unstable closed loop be stabilized by feedforward? No.
⇒ Feedforward is open loop, in parallel with closed loop

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

I Stability

•Can a stable closed loop be destabilized by feedforward? No.

•Can an unstable closed loop be stabilized by feedforward? No.
⇒ Feedforward is open loop, in parallel with closed loop
I Over-tuning
• What happens if we over-tune feedforward parameters?

error

time

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

I Stability

•Can a stable closed loop be destabilized by feedforward? No.

•Can an unstable closed loop be stabilized by feedforward? No.
⇒ Feedforward is open loop, in parallel with closed loop
I Over-tuning
• What happens if we over-tune feedforward parameters?

⇒ Over-compensation of required forces reverses the error

error

time

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

I Stability

•Can a stable closed loop be destabilized by feedforward? No.

•Can an unstable closed loop be stabilized by feedforward? No.
⇒ Feedforward is open loop, in parallel with closed loop
I Over-tuning
• What happens if we over-tune feedforward parameters?

⇒ Over-compensation of required forces reverses the error

error

time

I Trajectory dependence
• Do the feedforward terms change when the trajectory changes?

/w
Feedforward issues to be aware of 29/33

Feedforward tuning is time domain trial and error.

I Stability

•Can a stable closed loop be destabilized by feedforward? No.

•Can an unstable closed loop be stabilized by feedforward? No.
⇒ Feedforward is open loop, in parallel with closed loop
I Over-tuning
• What happens if we over-tune feedforward parameters?

⇒ Over-compensation of required forces reverses the error

error

time

I Trajectory dependence
• Do the feedforward terms change when the trajectory changes? No.

⇒ Terms represent system parameters, which do not change

/w
Smoothing feedforward 30/33

Second order trajectory may not be sufficient

I jump in ẍs ⇒ jump in feedforward force
I can excite parasitic (higher order) dynamics
I can introduce oscillations
I can cause actuator saturation

/w
Smoothing feedforward 30/33

Second order trajectory may not be sufficient

I jump in ẍs ⇒ jump in feedforward force
I can excite parasitic (higher order) dynamics
I can introduce oscillations
I can cause actuator saturation

How to solve this?

I Higher order setpoints

• gradual increase of feedforward forces (smoothing)

• by limiting higher derivatives (e.g. jerk)
I Here: using 3rd order setpoints

/w
3rd order setpoint 31/33

xs
0.5

0
0 0.2 0.4 0.6 0.8 1

1.5

1
ẋs

0.5

−0.5
0 0.2 0.4 0.6 0.8 1

5
ẍs

−5
0 0.2 0.4 0.6 0.8 1
t

/w
Higher order setpoints 32/33

Momentum = mass × velocity

Force = mass × acceleration

Yank = mass × jerk

Tug = mass × snap

Snatch = mass × crackle

Shake = mass × pop

/w
More advanced methods 33/33

Other advanced techniques:

I Model-based input shaping
I Iterative Learning Control (ILC)
• adapt the feedforward signal after each trial
I Learning input shaping
• adapt the trajectory after each trial

See other courses (4CM60 Advanced Motion Control) and current

research.

/w