Вы находитесь на странице: 1из 5

Instrumentation

Account for Uncertainty


with Robust Control Design
Part 2
Rakesh Joshi
Konstantinos Tsakalis
Arizona State Univ.
J. Ward MacArthur
Honeywell International Co.

Robust control can be applied to design


controllers that can handle process uncertainty.
Learn how to use identification experiments to
obtain estimates of this uncertainty.

Sachi Dash
Honeywell Process Solutions

obust control has emerged as a means to design


control systems that can handle the complexities
that arise in the chemical process industries (CPI). It
does so by accounting for uncertainties in the models used to
design the controller. This allows for the design of a control
system that maintains stability and achieves the specified
performance under all operating conditions.
This two-part article discusses robust proportionalintegral-derivative (PID) control design and its usefulness
for the CPI. Part 1, which appeared in the November issue,
explained the basics of robust control and outlined the
steps involved in robust control design. It also illustrated
the method with an example the classical pH-control
problem involving a first-principles model. The step
responses of the model used in that example produced
severe nonlinear behavior at two operating points of interest. This nonlinearity was also reflected in the computation
of the differences between the linearized models at these
two operating points. Treating these differences as uncertainty and applying robust control design tools revealed
that no single controller designed at a specific operating
point (e.g., pH = 6 or pH = 8) would be able to perform
satisfactorily in terms of the intended design specifications.

46

www.aiche.org/cep December 2014 CEP

pH Probe

Overflow Beaker

Microcontroller/
Computer

u
Buffer

Pump

Pump

Strong
Base
Strong
Acid

p Figure 10. A pH controller is designed for a 500-mL magnetically stirred


laboratory reactor with three input streams (acid, base, and buffer).

Copyright 2014 American Institute of Chemical Engineers (AIChE)

Example 2:
Controller design based on experimental analysis
Consider a 500-mL magnetically stirred reactor with
three input streams (acid, base, and buffer) as shown in
Figure 10. An overflow tube keeps the volume inside the
reactor constant. The flowrate of the base stream is used to
control the pH inside the reactor and is varied by adjusting
the pumps motor voltage; the other two flowrates (acid
and buffer) are fixed at 2.45 mL/min.
The example in Part 1 designed and evaluated pH
controllers based on modeling at two specific pH levels
pH = 6 and pH = 8. For this example, laboratory experiments were performed over a pH range of 59. The nominal models were designed at pH values of 6 and 8, but they
were evaluated over a wider range of pH. This wider set
of conditions is required to illustrate the highly nonlinear
characteristics of the pH response and the restrictions this
nonlinearity can impose on the selection of a single controller to operate at various operating conditions.
To identify the system model, a pseudo-random binary
sequence (PRBS) is used for step testing, with the flowrate
of the base as the input variable, a switching time of 100 sec,
and the magnitude of input variation chosen to perturb the
system around pH values of 59. (The interest in the two
extreme values, which are outside the operating range of the
application, is to enable the assessment of the model beyond
the nominal operating range.) The output (pH) sequences

pH x 103

1.5

pH 5

1.0

pH 9

pH 6
0
0

pH 7
1

2
Time, min

p Figure 11. The step responses of the models for the lab-scale reactor
(Figure 10) for pH values of 59 are shown over 4 min.

Copyright 2014 American Institute of Chemical Engineers (AIChE)

pH

p Figure 12. The identification experiment can provide the integration rate for
the lab-scale reactor shown in Figure 10 as a function of pH.

were measured and then fitted to linear dynamic models.


The reactor size is 500 mL and the total flowrate into the
reactor is about 7.35 mL/min, so the approximate settling
time for the step change in flowrate is about 3.4 hr, which
is three times the systems time constant of 1.13 hr. This
response time is significantly slower than that of the simulation model discussed in Part 1. Based on the equipment
limitations and the desire to vary the pH from 5 through 9,
an integrating model over a time scale of a few minutes was
selected to represent the data.
Figure 11 shows the step responses of the models at
each operating point over 4 min, and Figure 12 shows the
integration rate as a function of pH. The integration rate
(the slope of the lines in Figure 11) is the steady-state rate
of change of the pH with respect to time. The lowest integration rate is at pH = 7, the highest at pH = 5.
The quality of the models can be determined by plotting
the predicted pH vs. the actual pH at the various pH levels
(Figure 13). Clearly, in all cases, the predictions are very
close to the actual data, which illustrates the high quality
of the models. The transfer functions of the models at pH
values of 59 have the form of:
P(s) =

pH 8

0.5

Integration Rate, 1 x 104

Here, in Part 2 of the article, the focus shifts to the problem of computing realistic estimates of the model uncertainty from data obtained via identification experiments.
The data are then used to determine feasible performance
objectives (e.g., bandwidth) such that the corresponding
controller will produce the desired response with a high
degree of confidence.

K(1s+1) Td s
e
s( 2 s+1)

(19)

where K is the process gain, s is the Laplace variable, is a


time constant, and Td is the transport delay.
Table 2 lists the parameters for Eq. 19, where the subscripts denote the pH.
Next, the uncertainty analysis examines the ratio of the
spectral power of the residuals (|FFT (y ym)|) to the model
output (|FFT (ym)|), which provides an estimate of the signalto-noise ratio (SNR) in the frequency domain. (A variety of
CEP December 2014 www.aiche.org/cep

47

Instrumentation

spectral methods can be used for this computation.) Figure 14


shows the inverse of the multiplicative uncertainty estimate,
which serves as an upper bound on the loop complementary
sensitivity, which, in turn, provides an upper bound on the
controller gain. For the robust stability condition to be satisfied, the loop should have a bandwidth less than 1 rad/min
(the point at which the inverse multiplicative uncertainty,
1/Dm, equals one, as shown in the Figure 14). At higher
frequencies, the inverse multiplicative uncertainty becomes
less than one, which indicates that the controller should be
attenuating the loop signals.
Based on the uncertainty estimate, a loop bandwidth of
0.6 rad/min is chosen for the controller design for both process models (pH = 6 and pH = 8), as it satisfies robustness
with some margin and sampling time constraints, and is consistent with actuator saturation limits and quantization noise
levels (the last one is verified by a few loop simulations).
Two controllers were designed, one for pH = 6 and the
other for pH = 8. The closed-loop systems are tested at
nominal (i.e., design) and off-nominal conditions. The transfer functions of the two PI controllers are:

1
Cm6 (s) = 12,450 1+

(20)
2.94 s

9.22
pH Setpoint = 9

pH

8.89
8.56
8.23

8.21
pH Setpoint = 8
pH

8.07
7.92
7.78

7.01

pH

pH Setpoint = 7
6.87
6.74
6.60

pH

6.30

1
Cm8 (s) = 3,596 1+

2.36 s

5.88
5.46

Both controllers satisfy the small-gain theorem (Eq. 4)


for their respective model uncertainty and provide reasonable responses during a simulation-based validation. Figure 15 shows the experimentally measured step responses
of the closed-loop system with the controllers Cm6 and Cm8

pH Setpoint = 6
5.04
5.06

pH

4.43

101

3.80
pH Setpoint = 5

1/m = 1

3.16
25

50

75

100

Time, min

p Figure 13. This plot of the pH values predicted by the open-loop model
(ym, blue) and the actual measured pH values (y, green) as a function of
time reveals the high quality of the models at predicting process behavior.
Table 2. Parameters for the transfer function (Eq. 19)
of the models at pH values of 59.
K

Td

t1

t2

P5

3.77 x 104

0.267

1.6 x 105

P6

6.69 x

105

0.367

2.65

2.8

P7

4.96 x 105

0.217

0.0625

P8

2.02 x

104

0.183

0.124

P9

2.48 x 104

0.367

0.125

0.0334

www.aiche.org/cep December 2014 CEP

100
1/m

48

(21)

101

102
102

101

100
Frequency, rad/min

101

102

p Figure 14. This plot of the inverse multiplicative uncertainty estimate for
the model at pH = 8 shows that the magnitude of the uncertainty crosses
the value of one at a frequency of approximately 1 rad/min. Similar results
are obtained at pH = 6. This figure provides a quantitative characterization
of the model quality. Based on the robust stability condition (Eq. 4), it also
indicates that the controller can be designed with a closed-loop bandwidth
of 0.6 rad/min.

Copyright 2014 American Institute of Chemical Engineers (AIChE)

Steps toward improvement


The pH control problem exhibits several interesting
characteristics that are associated with process nonlinearity
and with the uncertainty (or complexity) in the description
of the practical components, such as sensors and actuators.
Experimental results show that the pH control performance degrades when the process operates away from its
nominal design conditions. The process itself undergoes
large gain variations, which means that simple controllers,
such as a standard PI/PID system, would require either
significant performance compromises or restriction to a tight
operating range. The experimental results also show that
when the process is operating in the vicinity of an operating
point, the process model and the limitations on achievable
closed-loop bandwidth can be identified with confidence,

0.2
0.1
pH = 6
0
0.1

10

15

Time, min

p Figure 16. The closed-loop step responses to a setpoint step change


of 0.12 (black) are plotted as a function of time for the Cm6 controller when
the process is at pH = 6 (red) and pH = 8 (blue). The considerable increase
in the speed of the closed-loop response with the Cm6 controller at pH = 8
is not consistent with the design objectives and indicates lower robustness
margins.
0.2

pH

pH = 8

0.1

0.1
pH = 6
0

pH = 6

0.1
0.1
0

10

Time, min

10
15
Time, min

25

20

p Figure 17. The closed-loop step responses to a setpoint step change of


0.12 (black) are plotted as a function of time for the Cm8 controller when the
process is at pH = 6 (red) and pH = 8 (blue). The response of the closedloop system with the Cm8 controller at pH = 6 is slower than the design
objective, indicating inadequate disturbance attenuation.

pH = 6

3,000

15

4,000

pH = 8

2,000

6
1,000

10

15

Time, min
*Base Flowrate (mL/min) = (7 x DAC Setpoint)/4,095

p Figure 15. These graphs show the performance of the closed-loop


systems at their respective design conditions, which closely matches the
design objectives. Top: The closed-loop step responses to a setpoint step
change of 0.12 (black) are plotted as a function of time for the Cm6
controller when the process is at pH = 6 (red) and the Cm8 controller when
the process is at a pH of 8 (blue). Bottom: This graph shows the flowrate of
the base is proportional to the voltage applied to the pump by the digital-toanalog converter (DAC).

Copyright 2014 American Institute of Chemical Engineers (AIChE)

pH

DAC Setpoint*, V

pH = 8

pH = 8

0.2

pH

and a systematic procedure can be used to design a controller to achieve this bandwidth. Performance degradation can
be expected due to the nonlinear behavior of the process
or changes in the process parameters, and becomes more
pronounced as the process moves farther from its nominal
operating conditions.
This raises an important question: Can control performance be improved over a wider range of operating conditions by online controller scheduling or adaptation? Based
on preliminary work (beyond the scope of this article),
we expect that uniform performance can be achieved by

pH

at their respective nominal pH conditions. These agree


well with the simulated closed-loop responses.
Figures 16 and 17 compare the closed-loop responses
of the controllers at their alternative operating points (offnominal conditions, i.e., Cm6 at pH = 8 and Cm8 at pH = 6).
The performance degrades becoming either too slow or
too fast, which in some cases can also result in oscillatory
behavior or instability. Figure 18 shows one such extreme
case where the closed-loop system is destabilized when the
Cm6 controller operates near pH = 4.5, where the process
gain is high.

Step Response

5
4
3

10

15
20
Time, min

25

30

35

p Figure 18. The closed-loop step response (red) to a pH setpoint change


from 6 to 4.5 (black) is plotted as a function of time for the Cm6 controller. The
process has a very high gain at a pH of 4.5, which, because the controller is
designed for a low-gain model, results in oscillatory behavior that is constrained
only by the high and low limits of the base flow actuator.

CEP December 2014 www.aiche.org/cep

49

Instrumentation

using more sophisticated, nonlinear controller schemes. For


example, the nonlinear model predictive controller paradigm
can be followed, whereby the control input is chosen based
on the identification of full nonlinear process models such
that it minimizes a suitable error (e.g., setpoint tracking) for
a given prediction horizon. Details of this method can be
found in Ref. 9.
A different approach would be to invoke the general
principle of gain scheduling. Different controllers would be
designed for different operating points and scheduled for
use based on independent measurements that determine the
operating point.
Alternatively, one can apply adaptive control principles
to estimate the controller gains based on input-output measurements. An adaptive PID controller is of particular interest here, because the controller complexity is low and the
effective variability can be attributed to a single parameter,
which can be estimated reliably with very modest excitation
requirements. Details of this method are found in Ref. 10
and an application to the pH problem will be the subject of
CEP
future work.

Literature Cited
1. Ljung, L., System Identification: Theory for the User, 2nd ed.,
PTR Prentice Hall, Upper Saddle River, NJ (1999).
2. Zhan, C. Q., and K. Tsakalis, System Identification for Robust
Control, American Control Conference, 2007, New York, NY,
pp. 846851 (July 913, 2007).
3. Alvarez, H., et al., pH Neutralization Process as a Benchmark
for Testing Nonlinear Controllers, Industrial & Engineering
Chemistry Research, 40 (11), pp. 24672473 (2001).
4. strm, K. J., and K. J. Hagglund, PID Controllers: Theory,
Design, and Tuning, Instrument Society of America, Research
Triangle Park, NC (1995).
5. Doyle, J. C., et al., Feedback Control Theory, MacMillan
Publishing, New York, NY (1992).
6. Grassi, E., et al., Integrated System Identification and PID Controller Tuning by Frequency Loop-Shaping, IEEE Transactions
on Control Systems Technology, 9 (2), pp. 285294 (Mar. 2001).
7. Rivera, D. E., et al., Internal Model Control: PID Controller
Design, Industrial & Engineering Chemistry Process Design
and Development, 25 (1), pp. 252265 (1986).
8. Henson, M. A., and D. E. Seborg, Adaptive Nonlinear Control
of a pH Neutralization Process, IEEE Transactions on Control
Systems Technology, 2 (3), pp. 169182 (Sept. 1994).
9. MacArthur, J. W., A New Approach for Nonlinear Process
Identification Using Orthonormal Bases and Ordinal Splines,
Journal of Process Control, 22, pp. 375389 (2012).
10. Tsakalis, K., and D. Sachi, Approximate H Loop Shaping in
PID Parameter Adaptation, International Journal of Adaptive
Control Signal Process, 27 (12), pp. 136152 (2013).

50

www.aiche.org/cep December 2014 CEP

Acknowledgments
The authors acknowledge the financial support from SERDP Environmental
Restoration Projects #2237 and #2239. Special thanks to Amy Childress
(Univ. of Southern California), Cesar Torres and Sudeep Popat (Arizona State
Univ.), and Eric Marchand (Univ. of Nevada, Reno) for their valuable input in
identifying the specific issues with pH control in the wastewater system.

Rakesh Joshi is a PhD student in electrical engineering at Arizona State


Univ., Tempe (Email: rsjoshi1@asu.edu). His research interests include
system identification, estimation theory, process control, robust control, and adaptive control.
Konstantinos Tsakalis is a professor in the School of Electrical,
Computer, and Energy Systems Engineering at Arizona State Univ.
(Email: tsakalis@asu.edu). His research is focused on the theory and
application of control systems, adaptive control, and system identification and optimization. In collaboration with Semy Engineering, he
developed an integrated identification and controller design procedure
for the temperature control of diffusion furnaces used in semiconductor
manufacturing.
J. Ward MacArthur is a Senior Engineering Fellow at Honeywell
International (Email: ward.macarthur@gmail.com). He has worked at
Honeywell for over 36 years, during which he has held various positions
at both the corporate research center in Minneapolis and the Industrial
Automation and Controls (IAC) Div. in Phoenix.
Sachi Dash is a Fellow at Honeywell Process Solutions (Phone: (480)
297-5387; Email: sachi.dash@honeywell.com). He currently holds the
positions of Global Leader, Process Control Center of Excellence and
Global Leader, Engineering Excellence. His research includes theory and
application of process control for the refining, petrochemical, pulp and
paper, power generation, and oil and gas industries. He is also involved
in designing controls for new or challenging processes.
The authors complete bios appear at the end of Part 1 (Nov. 2014, pp. 3138).

Nomenclature
C
Cm
FFT
j
K
M
P

= transfer function of the controller


= transfer function of the controller model
= fast Fourier transform
= imaginary number, square root of 1
= process gain (equilibrium constant for Eq. 68)
= nominal closed-loop transfer function
= transfer function describing the relationship
between an input (u) and an output (y)
Pm (s) = transfer function of the model (min)
= transfer function at a nominal operating point
Po
r
= controller setpoint
s
= time (min)
T
= complementary sensitivity
V
= tank volume (mL)
= reaction invariant a for each stream (i)
Wai
= reaction invariant b for each stream (i)
Wbi
y
= process output
= output of the process model
ym
Greek Letters
D
= frequency response of the uncertainty
= frequency response of the multiplicative
Dm
uncertainty
t
= process time constant
w
= frequency

Copyright 2014 American Institute of Chemical Engineers (AIChE)

Вам также может понравиться