Feedback Control Strategies

Cascade Control systems

• Main purpose: Reject disturbances that affect an

intermediate controlled variable before it hit the
main controlled variable.
 F, Ti

TC
Pc
V,T
Fh, Thi Tj
Fh, Tj

Po
F, T

Feedback Control System

Disturbance in Process side is addressed directly
 F, Ti
L

TC
Pc
V,T
Fh, Thi Tj
Fh, Tj

Po
F, T

Disturbance in Utility side is goes through additional Capacity and thus

control is slower
 F, Ti

T L

TC1 TC2
Pc
V,T
Fh, Thi
Tj
Fh, Tj

Po
F, T

With cascade control corrective action is taken along with the disturbance on
utility side
Cascade Control
Cascade systems contain two feedback loops:

• Primary Loop
• regulates part of the process having slower dynamics
• calculates setpoint for the secondary loop
• e.g. outlet temperature controller for the jacketed reactor

• Secondary Loop
• regulates part of process having faster dynamics
• maintain secondary variable at the desired target given by primary
controller
• e.g. steam flow control for the jacketed reactor example
Cascade control scheme

The simple PID loop that can be implemented, has the limitations if a disturbance arises in the
feed of the cooling fluid. The cascade provides rejection advantages for the disturbance
associated to the manipulated variable
Cascade diagram

L2 L1

GL2 GL1

R1 E1 R2 E2
Ysp P + + C1
+ + C2
Km Gc1 Gc2 GvGp2
+
Gp +
- -

GM2

GM1
Cascade diagram

L2 L1

GL2 GL1

R1 E1 R2 E2
Ysp P + + C1
+ + C2
Km Gc1 Gc2 GvGP2
+
Gp +
- -

GM2

GM1
Cascade diagram
L2

L L1
GL2 2
1+GvGP2GGL2c2GM2 GL1

R1 E1 R2 E2
Ysp Gc2PGvGP2 + + C1
+ + C
Km Gc1 Gc2 G
1+GvGP2Gc2GvM2 +
Gp +
- -

GM2

GM1
Cascade diagram
L2

L2 L1
Gd
GL2 GL1

R1 E1 R2 E2
Ysp P + + C1
C2
Km
+
-
Gc1
+
-
Gc2 Gs G v
+
Gp +

GM2

GM1
Comparison for load change
Cascade Control
Closed-loop transfer function

1. Inner loop
C2 G p2Gv2Gc2
  G cl 2
R2 1  G p2 G v 2 G c2 G m 2

2. Outer loop C1 G p 1 G cl 2 G c 1

R1 1  G p 1 G cl 2 G c 1 G m 1

3. Characteristic equation
1  G p 1 G cl 2 G c 1 G m 1  0
G p2 Gv 2 Gc2
1  G p1 G c1G m 1  0
1  G p2 Gv 2 Gc2 Gm2
1  G p 2 G v 2 G c 2 G m 2  G p 1G p 2 G v 2 G c 2 G c1G m 1  0
Cascade Control

Stability of closed-loop process is governed by

1  G p 2 G v 2 G c 2 G m 2  G p 1G p 2 G v 2 G c 2 G c1G m 1  0
Example

K p1
G p1  , G  K c1 , G v 1  G m 1  1
1 s  1 c 1
K p2
G p2  , G  K c2 , Gm2  1
2 s  1 c 2

K p2 K p2 K p1
1  K c2  K c1 0
2 s  1  2 s  1 1 s  1
( 1 s  1)(  2 s  1 )  K c 2 K p 2 ( 1 s  1)  K c 1 K p 2 K p 1  0

1 2 s 2  ( 1   2  K c 2 K p 2 1 ) s  1  K c 2 K p 2 
K c1 K p 2 K p1  0
 Cascade Control
Design a cascade controller for the following
system:
e 0 .1 s
G p1 ( s )  , G m 1  1,
1. Primary: ( 0 .5 s  1)( s  1 )
 1 
G c 1  K c1  1  
 I s

1
2. Secondary: G p2 
0 .1 s  1
, Gv2  Gm2 1

G c2  K c2
Cascade Control

1. PI controller only
 1 1 e 0 .1 s
G O L 1  K c 1 1  
 s  0 .1 s  1 ( 0 .5 s  1)( s  1)

1 1 1 1
A R  K c1 1  2
 0 .0 1 2  1 0 .2 5 2  1  2  1

1
   tan 1    tan 1 ( 0 .1 )

 tan 1 ( 0 .5 )  tan 1 (  )  0 .1

Critical frequency
c  2 .99 , A R  0 .1 7 8

Maximum gain
K c1  5.61
Cascade Control

Bode Plots

AR

j
j

ln(w)
Cascade Control
2. Cascade Control

• Secondary loop
1
G Ol2  K c2
0 .1 s  1

• no critical frequency gain can be large

• Let Kc2=10.
10
0 .1 s
• Primary loop  1  0 .1 s  1 e
G O L 1  K c 1 1  
 s 1 ( 0 .5 s  1)( s  1)
1 10
0 .1 s  1
10 e 0 .1 s
 K c1
1 1 s ( 0 .5 s  1 )( 0 .1 s  1)
11
Cascade Control
Closed-loop stability:

AR 10 1 1 1

K c1 1 1  2
0 .1 1  0 .25 2
1     2
 11
 0 .1
    0 .1  tan 1    tan 1 ( 0 .5 )
2  11 

c  4 .1 3 , A R  0 .0 9 58

• Bode

• Maximum gain Kc1=10.44

• Secondary loop stabilizes the primary loop.
Cascade Control
Use cascade when:
• conventional feedback loop is too slow at rejecting disturbances
• secondary measured variable is available which
• responds to disturbances
• has dynamics that are much faster than those of the primary variable
• can be affected by the manipulated variable

Implementation

• tune secondary loop first

• operation of two interacting controllers requires more careful
implementation
• switching on and off
Cascade design considerations
Cascade Control systems

• Main purpose: Reject disturbances that affect an

intermediate controlled variable before it hit the
main controlled variable.

• Results: Improve performance in rejecting some

process disturbances.
Install Cascade control on this system
Install Cascade control on this system
FEEDFORWARD CONTROL
SYSTEM
FEEDBACK CONTROL SYSYEM
Feedback control systems have the general form:
D(s)

UR(s)
GD
R(s) + + + +
Gc Gv Gp
+
U(s) Y(s)

Gs
Ym(s)

where UR(s) is an input bias term.

• Feedback controllers
• output of process must change before any action is taken
• disturbances only compensated after they affect the process
FEEDFORWARD CONTROL SYSTEM
 F, Ti
L

TC
Pc
V,T
Fh, Thi Tj
Fh, Tj

Po
F, T

T
 F, Ti
T L

TC
Pc
V,T
Fh, Thi Tj
Fh, Tj

Po
F, T

Disturbance and corrective actions enter together into the Process

FEEDFORWARD CONTROL SYSTEM

Feedforward Controller Design Based on
Dynamic Models

• Objective:
Measure
important load
variables and take
corrective action
before they upset
the process.
Feed-forward
• Main purpose: Reject disturbances that would
affect directly the main controlled variable before
it hit the process.

• Results: Improved performance in rejecting

disturbances compared to simple feedback.
Feedforward- Feedback
Controller Design Based on
Dynamic Models
• Assume that D(s)
• can be measured before it affects the process
• effect of disturbance on process can be described with a model GD(s)
Feedforward Control is possible. Feedforward
Controller
D(s)
Gf
GD
R(s) + + + +
Gc Gv Gp
+
U(s) Y(s)

Gs
Ym(s)

Feedback/Feedforward Controller
Structure
Transfer Function
• Relates the process variable to the disturbance
(load)
• Permits the design of the feedforward controller.
• Stability of the closed loop system.
Feedforward- Feedback Controller Design
Based on Dynamic Models
C= X1+X2

C= GLL+GpM

C= GLL+Gp GvP
C= GLL+Gp Gv[GcE+GfGtL]
C= GLL+Gp Gv[Gc(R’-Gm)+GfGtL]
C= GLL+Gp Gv[Gc(KmR-Gm)+GfGtL]

C= GLL+Gp GvGc(KmR-GmC)+ Gp Gv GfGtL]

C(1+ Gp GvGcGm) = Gp GvGcKm R+ (GL+Gp Gv GfGt)L

C(s)= Gp GvGcKm R(s) + (GL+Gp Gv GfGt) L(s)

(1+ Gp GvGcGm) (1+ Gp GvGcGm)
Feedforward Control
The feedforward controller:
D(s)
Gf
GD
UR(s) + + +
Gv Gp
+ Y(s)
U(s)
Feedforward Control
The feedforward controller:
D(s)
Gf
GD
UR(s) + + +
Gv Gp
+ Y(s)
U(s)

Transfer Function
Y ( s )  G D ( s ) D ( s )  G P ( s ) G v ( s )U ( s )
Y ( s )  G D ( s ) D ( s )  G P ( s ) G v ( s )( U R ( s )  G f ( s ) D ( s ))
Y ( s )  ( G D ( s )  G p ( s ) G v ( s ) G f ( s )) D ( s )  G p ( s ) G v ( s ) U R ( s )
Y ( s )  ( G D ( s )  G p ( s ) G v ( s ) G f ( s )) D ( s )  Y R ( s )

Tracking of YR requires that G D ( s )  G p ( s ) G v ( s ) G f ( s )  0

G D ( s)
 G f ( s)  
G p ( s)G v ( s)
Feedforward Control
Ideal feedforward controller:
G D ( s)
G f ( s)  
G p ( s) G v ( s)

• Exact cancellation requires perfect plant and perfect disturbance models.

G D ( s)  G p ( s)G v ( s) G f ( s)  0

• Feedforward controllers:
• very sensitive to modeling errors
• cannot handle unmeasured disturbances
• cannot implement setpoint changes

• Need feedback control to make control system more robust

Ideal Feedforward Transfer Function
• If the set point is constant, C(s) should be equal to
0 despite the change in L(s).
• The ideal transfer function could not be physically
realizable.
GL  Gt G f Gv G p  0
GL
 Gf  
Gt Gv G p

Here it is important to emphasize that this equation is the ideal case, which should be
modified depending on the models of each of the components. The two main two rules that
must be enforced are (1) the degree of the numerator must be equal or less than the
denominator’s degree; and (2) the overall dead time must be negative (dead time in
numerator > dead time in denominator
 F,Tin
TT
Heated Stirred Tank
TT
TC1

Ps
Steam

Condensate
F,T

• Is this control configuration feedback or feedforward?

• How can we use the inlet stream thermocouple to regulate the inlet folow disturbances
• Will this become a feedforward or feedback controller?
Feedforward Control
A suggestion: TC2
TT
F,Tin +
TT +

TC1
Ps

Steam

Condensate
F,T
How do we design TC2?
Feedforward Control
Feedback/Feedforward Control

D(s)
Gf
GD
R(s) + + + +
Gc Gv Gp
+
U(s) Y(s)

Gs
Ym(s)

What is the impact of Gf on the closed-loop

performance of the feedback control system?
Feedforward Control
Regulatory transfer function of feedforward/feedback loop

C ( s) G D ( s)  G f ( s)G v ( s)G p ( s)

D ( s) 1  G c ( s)G v ( s)G p ( s)G m ( s)

Perfect control requires that (as above)

G D ( s)
G f ( s)  
G v ( s)G p ( s)

Note:
• Feedforward controllers do not affect closed-loop stability
• Feedforward controllers based on plant models can be unrealizable (dead-time or
RHP zeroes)
• Can be approximated by a lead-lag unit or pure gain (rare)
( 1 s  1) KD
G f ( s)  K f G f ( s)  
(  2 s  1) Kv K p
Feedforward Control
Tuning: In absence of disturbance model lead-lag approximation may be good

( 1 s  1)
G f ( s)  K f
(  2 s  1)
• Kf obtained from open-loop data

KD
Kf 
Kv K p
- t1 and t2
• from open-loop data

1   p ,  2   D
• from heuristics

1 1
 0 .5  2 .0
• Trial-and-error 2 2

1   2  c
Feedforward Control
Example:
10
G p ( s) 
(1 0 s  1)( 5 s  1)( s  1)
Plant: 1
G D ( s) 
( 2 .5 s  1)( s  1)

1 0 e 6 s e s
G pm ( s )  G D m ( s) 
Plant Model: 10 s 1 2 .5 s  1

K c  0 .2 6 ,  I  1 3 ,  D  2 .3 1

Feedback Design from plant model: IMC PID tunings

Feedforward Control
Possible Feedforward controllers:

e 5 s ( 10 s  1)
1. From plant models: G f ( s)  
1 0 ( 2 .5 s  1)

• Not realizable

1  1 0 ,  2  2 .5
2. Lead-lag unit 1
Kf 
10

1
3. Feedforward gain controller: Kf 
10
Feedforward Control
For Controller 2 and 3 Disturbance Controller with Feedforward
1.2

0.8
.. - Gain Controller
0.6 -- - Lead-Lag Controller

- - No FF Controller
0.4

0.2

-0.2

-0.4

-0.6
0 20 40 60 80 100 120 140 160 180 200

• Some attenuation observed at first peak

• Difficult problem because disturbance dynamic are much faster
Feedforward Control
• Useful in manufacturing environments if good models are available
• outdoor temperature dependencies can be handle by gain feedforward controllers
• scheduling issues/ supply requirements can be handled

• Benefits are directly related to model accuracy

• rely mainly on feedback control

• Disturbances with different dynamics always difficult to attenuate with PID

• may need advanced feedback control approach (MPC, DMC, QDMC, H4-controllers,
etc…)

• Use process knowledge (and intuition)

Case Study: Mixing in a series of well mixed tanks
 Concentration transmitter
range=0.3-0.7
DP across valve is constant
and maximum flow provided
by valve is 2000LPM
Valve dynamic lag is 0.1min
Densities of all stream are
the same as water
Major disturbance
Steady state value of
composition is 0.472 when
f2=1000
Maximum deviation
stream Flow Mass stream Flow Mass permitted =1.5% of setpoint
(LPM) fraction (LPM) fraction Range(0.465-0.479)
1 1900 0 5 500 0.8
2 1000 0.99 6 3900 0.472
3 2400 0.167 7 500 0.9
4 3400 0.049
Draw the conventional Feedback control block diagram
 LPM

LPM

LPM
Draw the Feedforward control system on the
Mixing process
LPM
Let

Substituting
 IF the flow transmitter is on stream 2 is calibrated from 0-2000 LPM, its transfer
function with fast dynamics is 100%/2000PM

FFC= - 0.032/(0.05 x 1.095) = 0.58

Implementation of feedforward/feedback controller
Ratio Control of blending of two liquid
streams
Override control scheme
Selective (or Auctioneering) control for a plug flow reactor
 Hot oil system

Implement Selective Control strategy

A Large Dead Time Impacts Controller Performance
Dead Time Compensation
• Time delay occurs because of the presence of
distance velocity lags, recycle loops, and dead time
from composition analysis.
• Limits performance of a conventional feedback
control system by adding phase lag.
• Controller gain must be reduced below the value
that could be used if dead time were not present,
hence sluggish response would be obtained
compared to that of no dead time

Tie delay compensation is used to avoid the problems caused by large dead time specially
instability. The main idea is to simulate and control the process without the dead time and then
implement the values after the dead time has passed. This allows the strategy to use larger
gains than the basic PID loop would calculate with the dead time. The result is a more
aggressive control without risking instability.
Dead-time Compensation
Consider feedback loop: D

R C
Gc Gp e-qs

• Dead-time has a de-stabilizing effect on closed-loop

system
• Presence of dead-time requires detuning of controller
• Need a way to compensate for dead-time explicitly
Dead-time Compensation
e  s
G ( s)  , 0 .1    0 .7 5
2
s  3s  2
1
G c ( s )  4 1  
 s

0.75 0.5 0.25 0.1

Dead-time Compensation
D
Use plant model to predict deviation from setpoint
R C
Gc Gp e-qs

Gpm

Result:
• Removes the de-stabilizing effect of dead-time
Problem:
• Cannot compensate for disturbances with just feedback (possible offset)
• Need a very good plant model
Dead-time Compensation
Closed-loop transfer function

C ( s) C ( s) G c G p e s
 1, 
D ( s) R ( s ) 1  G c G pm

Characteristic Equation becomes

1  G c G pm  0

• Effect of dead-time on closed-loop stability is removed

• Controller is tuned to stabilize undelayed process model
• No disturbance rejection
Dead-time Compensation
D

R 1 1 C
4 1   2 e -0.5s
 s s  3s  2

1
s2  3s  2

Servo Response
1.5

0.5

0
0 1 2 3 4 5 6 7 8 9 10

Regulatory Response
1

0.8

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10
 Dead-time Compensation
Include the effect of disturbances using model
predictions D ( s )  Y ( s )  Y( s )

D ( s )  G p e  sU ( s )  G pm e sU ( s )

D
Adding this to previous loop gives
R + C
Gc Gp e-q s

+ - +

+
+
Gpm Gpm e-qs
-
+

This plot shows how the control signal is used in the real process and in the model of the process
without the dead time. The simulated value without the dead time effect is sent back to the
control to compute the next control move. At the same time, the value is sent through the dead
time to simulate the original process. Both values are compared to detect any mismatch or
disturbances. The difference is compared with the set point to compute the real error, which
updates the fictitious error.
Dead-time Compensation
Closed-loop transfer function

C ( s) 1  ( e s  e  s ) G c G pm

D ( s) 1  G G  s  s
c pm  G c ( G p e  G pm e )

C ( s) G c G p e  s

R ( s) 1  G G  s  s
c pm  G c ( G p e  G pm e )

Characteristic Equation

1  G c G pm  G c ( G p e  s  G pm e  s )  0

Fast Slow
Dynamics Dynamics

• Effect of dead-time on stability is removed

• Disturbance rejection is achieved
• Controller tuned for undelayed dynamics
Dead-time Compensation
D

R 1 1 + C
4 1   e
-0.5s

+ -
 s s2  3s  2 +

+ 1 1 +
e -0.5s

+ s2  3s  2 s2  3s  2 -

D ( s )
Servo Response
1.5

0.5

0
0 1 2 3 4 5 6 7 8 9 10

Regulatory Response
1

0.5

-0.5
0 1 2 3 4 5 6 7 8 9 10
Dead-time Compensation
Alternative form D

R + C
Gc Gp e-q s

+ - +

+
Gpm(1-e-qs)
+

Reduces to classical feedback control system with

𝐺𝑐
𝐺∗𝐶 ( 𝑠 )=
1+𝐺 𝑐 𝐺 𝑝𝑚 (1− 𝑒− 𝜃 𝑚 𝑆
)

called a Smith-Predictor
Dead-time compensation
Smith-Predictor Design

1. Determine delayed process model

Y( s )  G pm ( s ) e m s

2. Tune controller Gc for the undelayed transfer function model Gpm

3. Implement Smith-Predictor as
∗ 𝐺𝑐
𝐺 ( 𝑠 )=
𝐶 − 𝜃𝑚 𝑆
1+𝐺 𝑐 𝐺 𝑝𝑚 (1− 𝑒 )

4. Perform simulation studies to tune controller and estimate closed-loop

performance over a range of modeling errors (Gpm and qm)
Dead-time Compensation
Effect of dead-time estimation errors:
D

R 1 1 + C
4 1   e -0.5s

+ -  s s2  3s  2 +

+ 1 1 +
e
-ts

+ s2  3s  2 s2  3s  2 -

D ( s )

t
Smith Predictor - PI Comparison
Internal Model Control (IMC)

Placement of a process model, G*P(s), in parallel with the actual process

Process model G*P (s) receives the actual controller output signal, U(s), and uses it
*

to compute Y*(s), a prediction of the measured process variable, Y(s).

Internal Model Control (IMC)

Y(s) =U(s)GP (s) + D(s)GD (s)

Y*(s) =U(s)GP* (s)

U(s) = E(s)G*C (s)

= [Ysp (s) −Y(s) + Y*(s)]G*C (s)

U(s) = [Ysp (s) −U(s)GP (s) − D(s)GD (s) +U(s)GP* (s)]G*C (s)
= Ysp (s)G*C(s) −U(s)GP (s)G*C (s) − D(s)GD (s)GC* (s) +U(s)GP* (s)G*C (s)

We solve for U(s)

Rearrange into the closed loop transfer function
Three basic steps to derive IMC tuning

Step 1:
Remember that the poles of a transfer function (the roots of the characteristic equation
in the denominator of the transfer function) indicate system stability

The approach we take is to invert the process model to create the controller.

One problem with such an approach is that any roots in the numerator (analogous to
poles, roots in the numerator of a transfer function are called zeros) of the process model
that lie in the right hand of the complex plane, when inverted, become unstable poles. i.e.
unstable system.
 Three basic steps to derive IMC tuning

Step 1:

To avoid creating an unstable controller, factor the process model, GP* (s) , into an
invertible and noninvertible part.

G*P (s) = G*P + (s)G*P − (s)

The invertible part

The noninvertible part, -contains left hand plane zeros

contains all right-hand plane zeros (roots in the

numerator of a transfer function that have
positive real parts) and the dead time term
 low-pass filter
Step 2: with gain
equal to 1
Specify the controller transfer function

t , indicates the speed of

C

the response of a process

to set point changes.

Popular heuristic for achieving a 10% to 15% overshoot

tc is the larger of 0.1t
P or 0.8θP

A heuristic for a more conservative “no overshoot

tc is the larger of 0.5t
P or 4θP
Step 3 Relate the IMC transfer function models to those from classical feedback control

Recall that the closed loop transfer function for a classical feedback control

Equate the two

( )
GP (s)G*C (s)[1+GP (s)GC (s)] = GP (s)GC (s) [1+ GP (s) −G*P (s) G*C (s)]

G*C (s) +G*C (s)GP (s)GC (s)

= GC (s) +G*C (s)GP (s)GC (s) -G*C (s)G*P (s)GC (s)

Rearranging, we obtain Gc(s)!!!

𝐺 ∗𝑐 (𝑠)
𝐺𝑐 ( 𝑠 ) =
1 −G ∗C   (s )G ∗ P   (s )
Assume a process model

Substitute the first-order expansion for e θ s

− P

Factor G*P (s) into invertible and noninvertible parts

G*P (s) = G*P + (s)G*P − (s)

So following the discussion above

G*P + (s) = (1−θ s)

P
Now express the IMC controller model, G*C (s)

In terms of the invertible process model term and a first-order filter term, F(s)

the IMC filter has the form

We can relate this IMC controller model, GC* (s) , to a classical feedback controller model
substituting

Compare it to the classical feedback model for a PI controller

Controller Design by
Direct Synthesis
Controller Design by Direct
Synthesis
• Assume the process + measurement device + final
control element can be represented by G

• Assume also that the closed loop behavior follows

an ideal
Ysp(s)

(s)
Controller Design by Direct
Synthesis
• where
Ysp(s)= q(s) Ysp(s)

• The choice of the reference trajectory q(s) depends

on the desired closed loop response and the open
loop response of the process.
• Next table presents some typical responses
Controller Design by Direct Synthesis
Controller Design by Direct
Synthesis
• What is the form of Gc(s) that produces the desired
response in C(s) i.e. say q(s) , given the G(s) of the
process.

• This can be achieving manipulating the expression

of q(s) to obtain
Gc
Controller Design by Direct
Synthesis
• Assume a first order model for G(s) and an
underdamped second order for the reference
trajectory.
• Therefore, Gc(s) is

 1 
s  1   r2 s2  2 r r s  1 
gc (s)   
k 1 1 
 2 2 
  r s  2 r r s  1 
 Controller Design by Direct
Synthesis
• After some algebra

s  1  1 
gc (s)   2 2 
k  r s  2 r r s 


Model Predictive Control

A dynamic model of the process programmed into the control

architecture.

The function of the model is to predict the future behavior of

the process based upon past controller moves and the current
state of the process

At each sample time, the next controller move is computed from

a comparison of this predicted future behavior with the desired
set point trajectory
MPC strategy begins with a performance objective function

This objective function typically combines controller error and

controller effort into a single formulation.

If the measured process variable is maintained at its set point

over the predicted future (controller error is zero),

If the control effort (the size of each control move) is small, then
the mechanical components of the final control element won't wear
excessively and the process will not experience unsettling sudden
changes
By finding the mathematical minimum of the objective function,
control actions of modest size will be computed to drive the
future predicted controller error to zero

Future set point tracking performance in the objective function is

computed using the dynamic process model

The model predicts the future of the measured process variable

using past and future (yet to be computed) controller output moves.
Only the first of these controller output moves is implemented
before repeating the entire procedure at the next sample time.
Sensors
Sensors are a crucial part of any control system
design, since they provide the necessary information
upon which the controller action is based. They are
the eyes of the controller. Hence, any error, or
significant defect, in the measurement system will
have a significant impact on performance.
Noise
The effect of measurement noise in the nominal loop
is given by

Also, we recall that T0(s) is typically near 1 over the

bandwidth of the system. Thus, given the fact that
noise is typically dominated by high frequencies,
measurement noise usually sets an upper limit on the
bandwidth of the loop.
Filtering the Process Measurement

y f (t )  f ys (t )  (1  f ) y f (t  t )

• Filtering reduces the effect of sensor noise by

approximating a running average.
• Filtering adds lag when the filtered measurement is
used for control.
• Normally, use the minimum amount of filtering
necessary.
• f- filter factor (0-1)
 Feedback Loop with Sensor Filtering
D(s)

G d (s)

E(s) Y(s)
++
Y sp (s) C(s) U(s)
+- G c (s) G a (s) G p (s)

Y f (s) Y s (s)
G f (s) G s (s)
Effect of Filtering on Closed Loop
Dynamics
Characteri stic equation for P  only controller
on first order process with sensor filtering :
 Kp  1 
Kc    1 0
 p s  1  f s  1
  p f
p 
Kc K p  1
p  f
 
2  p  f ( K c K p  1)
Analysis of Example
• tf is equal to Dt (1/f-1) as f becomes small, tf
becomes large.
• As tf is increased, tp’ will increase.
• Critical issue is relative magnitude of tf compare to
tp.
Effect of the Amount of Filtering on
the Open Loop Response
Filtered Temperature

f=0.3

f=0.1
f=0.2

0 20 40 60 80 100
Time (seconds)
Effect of a Noisy Sensor on Controlled
Variable without Filtering

Product Temperature

Manipulated Variable

Time
Effect of a Noisy Sensor on Controlled
Variable with Filtering

Product Temperature

Manipulated Variable

Time
An Example of Too Much and Too
Little Filtering
Temperature (ºC)

104

102 f=0.01
f=0.2
f=0.5
100
0 50 100 150 200
Time (seconds)
Relationship between Filter Factor (f), the Resulting
Repeatability Reduction Ratio (R) and the Filter
Time Constant (tf)

2 2 f
f  2 or R 
R 1 f
1 
 f  t f   1
f 
Key Issues for Sensor Filtering

• To reduce the effect of noise (i.e., R is increased), f

must be reduced, which increases the value of tf.
Filtering slows the closed-loop response
significantly as tf becomes larger than 10% of tp.
• The effect of filtering on the closed-loop response
can be reduced by increasing the frequency with
which the filter is applied, i.e., reducing Dtf.
PID Controller Design Issues
• Over 90% of control loops use PI controller.
• P-only: used for fast responding processes that do
not require offset free operation (e.g., certain level
and pressure controllers)
• PI: used for fast responding processes that require
offset free operation (e.g., certain flow, level,
pressure, temperature, and composition
controllers)