Honeywell Forge APC-4516

Process Controller Implementation

Dynamic Model Review

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted
1 September 26, 2022
Finite Impulse Response (FIR) Models Honeywell

• Similar in concept to finite step response models

2.5

2 1.96 1.98 1.99 2.00

1.90 1.94
1.84
1.73

1.5 1.55

1.26 U(t)
Y(t)
1
0.79

0.5

0 0.00 0.00 0.00

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 2 September 26, 2022
Step Response Models Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 3 September 26, 2022
FIR Models1 Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 4 September 26, 2022
FIR Models2 Honeywell

2 0.800

1.8
0.700
1.6 This is a step response model of
“a process” 0.600
1.4
0.500
1.2 This is a FIR (response of the process
FSR

to an impulse) I.e. a large change over

FIR
FSR
1 0.400
a short time. This is a theoretical concept FIR

0.8 but one that is useful

0.300
0.6
0.200
0.4
0.100
0.2

0 0.000
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Scan

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 5 September 26, 2022
FIR Models3 Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 6 September 26, 2022
FIR Models4 Honeywell
• Predictor (Positional form)

• Predictor (Velocity form)

y (t ) = b0u (t ) + b1u (t − 1) +  + bnb u (t − nb ) + bias

∆y (t ) = b0 ∆u (t ) + b1∆u (t − 1) +  + bnb ∆u (t − nb )

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 7 September 26, 2022
FIR Models5 Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 8 September 26, 2022
FIR Models6 Honeywell

∆yn = 0.79∆xn −4 + 0.47∆xn −5 + 0.29∆xn −6 + 0.18∆xn −7 + 0.11∆xn −8 + 0.06∆xn −9

+ 0.04∆xn −10 + 0.02∆xn −11 + 0.02∆xn −12 + 0.01∆xn −13 + 0.01∆xn −14

This is the FIR equation in the velocity form

for our earlier response. Note that early
deadtime coefficients are zero and so drop out.

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 9 September 26, 2022
Laplace Transform S-Domain Honeywell

• Converts differential in the time domain to algebraic equations in

the complex s-domain
• Forms the basis for frequency response methods and is the
backbone of classical control theory and root locus theory
• One-to-one correspondence to linear continuous time models
– Transform tables for most common differential equations to be
transformed
• Basis for transfer function modeling of process, electrical and
mechanical systems

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 10 September 26, 2022
Laplace Transforms1 Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 11 September 26, 2022
Laplace Transforms2 Honeywell

dL
A = F1 − F 2
dt
Changing to deviation variables and initial conditions
~

d  L + L0  ~ ~
A = F1+ F10 − F 2 − F 2 0
dt
Where
F1o = Initial flow in
F 2 0 = Inital flow out
Lo = Initial level

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 12 September 26, 2022
Laplace Transforms3 Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 13 September 26, 2022
Laplace Transforms4 Honeywell

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 14 September 26, 2022
Fixed Laplace Form Honeywell

• Within the controller, structure of the Laplace form is limited to:

• Guaranteed stable, no underdamped behavior

k (τ n s + 1)e − ds
• lead dynamics y(s) = u (s)
– positive - Overshoot
s (τ1s + 1)(τ 2 s + 1)
– negative - Inverse response
• Initial s in denominator is present for integrators only

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 15 September 26, 2022
Z-Domain1 Honeywell

• Discrete Transformation
– Converts equations in the Laplace domain to Z domain representation
– Forms the basis for discrete modeling and is the backbone of predictive
control techniques
• One-to-one correspondence between S and Z domain
– Transform tables
• Basis for transfer function modeling

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 16 September 26, 2022
Z-Domain2 Honeywell

Where:
k-1 is previous scan
z is the backward shift operator
Note that the prediction relates to the previous value and the new value in a
similar way to the way the exponential filter works

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 17 September 26, 2022
 Identifier Overview

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted
18 September 26, 2022
Identification Honeywell
• A fundamental problem in system identification is the choice of the
nature of the model which should be used for the system. In general the
model is one of the following:
– Linear time-invariant (lumped parameter) - ODE
– Linear time-varying (lumped parameter) - ODE
– Linear with distributed - PDE
– Non-linear
• System identification still remains both an art and a science. The
science is concerned with parameter estimation, while the art is
concerned with overcoming the following problems:
– Determining structure and /or order
– Excitation requirements
– Accuracy

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 19 September 26, 2022
Parameter Estimation Honeywell

• System identification requires two steps. First a family of

candidate models is decided upon. Then the particular member of
this family that satisfactorily describes the observed data is
selected.
• The first step is the 'art' and is often problem specific.
– The second step is the 'science' and is in fact a parameter estimation
problem.
• Parameter estimation is basically the determination of the 'best'
set of candidate model coefficients such that the model
represents the observed data in a desirable fashion.
• A family of automated estimation tools is contained in Honeywell's
Profit Suite Engineering Studio

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 20 September 26, 2022
RMPCT Features Honeywell
• Gives the engineer a powerful tool to inspect data and generate MIMO
system models
• Considers all dependent and independent variables simultaneously
• Allows operator interaction during plant testing
• Process does not need to have steady initial or terminal conditions
• Multiple model forms and structures are accommodated
• Data segmentation is permissible
• Performance is given in terms of:
• - step response
• - time series prediction per CV
• - residual error
• System models can be automatically chosen based on open-loop
prediction performance
• Both continuous- and discrete- time models are generated
• Cross validation analysis is easily accommodated
• Multiple system models can be merged easily

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 21 September 26, 2022
Identification Overview (1) Honeywell
• Model categories
– Linear or non-linear
– Continuous or discrete
– Parametric or non-parametric
– Time or frequency domain
• Stages of System identification
– Experimental design & execution
– Data processing
– Model order/structure selection
– Parameter estimation
– Model validation

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 22 September 26, 2022
Identification Overview (2) Honeywell

Start

Experimental Design
and execution
Steps, pulses, PRBS, etc.

Identification
- Data processing Correlation, transformation
- Model order/structure
- Parameter estimation
plant models

Model validation Simulation, cross validation

Is model ‘Good’ Use model

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 23 September 26, 2022
Error Sources In Estimation Honeywell
• Error = Bias + Variance
• Bias - Systematic errors caused by
– Input signal characteristics (power, correlation, etc)
– Choice of model structure
• Caused by limiting ourselves to a deadtime and lag model. When you
have inverse response, you may introduce error as you try to fit a
Laplace transform to it unless you allow for a high order model
• Mode of operation (open vs. closed-loop)
– Try to minimize
• Variance - Random errors caused by
– presence of noise
– number of model parameters
– duration of the test
– if you do enough step testing it will be minimized

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 24 September 26, 2022
FIR Properties For Identification Honeywell

Strengths Weaknesses

• Structure free • High raw variance

• Unbiased consistent • Large model orders
estimates in open- • Estimate of settling time
loop experiments • Data for integrators
• Extremely fast must be differenced
• MIMO • Wasted data during
• Discontinuous and non-zero initial conditions
bad data handled • Biased model in closed-
directly loop experiment
• Wide frequency band

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 25 September 26, 2022
FIR Identification Summary Honeywell

• Approach does not introduce systematic error 

• Practical approach for MIMO systems 
• Approach is subject to high raw model variance/random error 

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 26 September 26, 2022
Parametric Model ID1 Honeywell

2.5

1.5

2 e−2 S Parametric
1
y( s ) = u( s ) Noise
CV

2S + 1 FSR

0.5

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Residual “noise” is
-0.5 Laplace-FIR

Scan

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 27 September 26, 2022
Parametric Model ID2 Honeywell
• Benefits of the parametric model fit
– Reduced model variance
• Parametric fit acts as a filter. Rather than filter the raw data we average
out the variance using the parametric fit
– Models are parsimonious (fewer model parameters).
• Computationally more efficient
• Used by on-line controller for prediction
– Easier model manipulation
• If you want to edit the Laplace transform it is very easy to do this and we
can add a filter etc. using the Laplace editor box which we will use
extensively later on in this course to show model manipulation

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 28 September 26, 2022
Parametric Model ID3 Honeywell
• Parametric model fit options
– Laplace (Fixed form)
• The Laplace fit chooses the most appropriate model order for you (1st or
2nd, with or without lead term as required)
– Discrete (ARX or OE)
• The discrete form fits one less zero than the no. of poles I.e. it always
fits a 2nd order model by default but we can fit up to 10th order.
• Experience is that mostly 2nd order is enough with occasional
requirements for 3rd order
– Best of both

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 29 September 26, 2022
Discrete Model Regression Honeywell

ARX / Equation Error

y (t ) =
( )u(t ) + 1 e(t )
B z −1
A (z −1 ) A (z −1 )

Output Error

y (t ) =
( )
B z −1
u (t ) + e(t )
F(z )
−1

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 30 September 26, 2022
‘Best’ Model Selection Honeywell
• Typically multiple sets of FIR generated
– e.g. 3 cases 60, 90 and 120 min. settling time
– can set many more than 3 trials and then pick the best trial or let us do it for
you and then check on its choice and override it at your will!
– One parametric model fit for each CV, MV (or DV) pair per case
• Best combination of models selected to minimize error
– Between predicted CV and plant test results

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 31 September 26, 2022
Summary of RMPCT approach Honeywell
• Hybrid approach
– FIR or GMS
– Parametric
• First pass FIR calculations no structure required (settling time)
• Second pass parametric to FIR
– Continuous (fixed form)
– Discrete (ARX or OE)
– Select best
• Third pass use raw data to automatically select the best model

Honeywell Confidential
Copyright © 2012 Honeywell International Inc.
Honeywell Restricted 32 September 26, 2022