A Simple Method for On-Line Identification

and Controller Tuning

Cheng-Liang Cben
Chemical Engineering Department
National Taiwan University
Taipei, Taiwan 10764, R.O.C.

The three-mode (PID) controller is still widely used in chemi- Simulation examples are supplied to demonstrate the robustness
cal industries because it is robust and is easy to operate. Many and applicability of this identification method.
tuning guidelines have been recommended for the PID control-
lers. These include the Ziegler-Nichols closed-loop cycling The Method
method (Zeigler and Nichols, 1942), the Cohen-Coon open-loop Consider a single-input/single-output(SISO) feedback con-
reaction curve method (Cohen and Coon, 1952), and many trol system as shown in Figure 1. With P mode controller, it is
other model-based minimum error integral methods (Lopez et assumed that the closed-loop system is underdamped if a feed-
al., 1967). The Ziegler-Nichols continuous cycling procedure is back gain large enough is chosen. The typical response to step
usually criticized for requiring the controlled process to be set-point change is shown in Figure 2. The response approxi-
forced to the level of marginal stability. On the other hand, the mates that of an underdamped second-order system with some
advantage of the Ziegler-Nichols tuning formulae is that it need element of dead time, as in Eq. 1,
not characterize the process by parametric models, the results of
which are known to depend on testing conditions. Many algo- C(s ) K e-dr
rithms have been presented to obtain the critical data (ultimate G,(s) = -=
gain and period) under acceptable conditions (e.g., Krishna-
R(s) 72s2 + 2<7s + 1
swamy et al., 1987). Here,
Recently, Yuwana and Seborg (1982) proposed a simple on-
line algorithm which used the closed-loop response and the Pade
approximation of the dead-time element to evaluate the parame-
ters of a first-order process model, then the critical data of the
model are used for subsequent controller tuning by Ziegler-
-In (H)
Nichols rules. Lee (1989) modified the identification algorithms
by matching the dominate poles of the closed-loop model to the
<= ./r2+ In2 (H)
(3)
poles of observed process transfer function. This modification
enables the method to be used for processes with large dead
times such as were not comprehended in the original paper.
However, the applicability of Lee’s modification to under-
damped processes has not been demonstrated.
Instead of using a low-order parametric model in process d = 2 f p , - t,,
characterization, this article proposes determining the process
critical data directly from the closed-loop response during step
set-point change under P control mode. The modified method is
expected to have two advantages over the algorithms of Yuwa-
na-Seborg and Lee:
1) It can provide more robust process critical data under
many different testing conditions.
2) It is applicable to underdamped processes and processes In these equations, A is the magnitude of input disturbance in
with dominant dead times. set point, and H i s the overshoot. Note that the new steady state

AIChE Journal December 1989 Vol. 35, No. 12 2037

 t
Figure 1. Conventional feedback control system.

(c,) after step change in set point can be inferred from the tran- Q)
u)
sient response data (Yuwana and Seborg, 1982). The testing c
0
period can thus be reduced dramatically. c1
u)
Instead of estimating an open-loop parametric model from Q)
L
the closed-loop transfer function, the ultimate gain and fre-
quency of the open-loop system, which are critical for subse-
quent controller settings, are determined directly. The phase
cross-over frequency (w,) of the closed-loop transfer function
can be easily obtained by solving the following nonlinear equa-
tion,
O tq tml Time
- do, - tan-' (;-)=-r (8) Figure 2. Typical underdamped closed-loop response to
a step-setpoint change.
The corresponding magnitude at this frequency is,
Here, the process gain, K,,,, can be determined from the above-
mentioned simple closed-loop test (Yuwana and Seborg, 1982),

Note that w, is also the ultimate frequency of the loop transfer

function C,(s)C,(s). Thus, the ultimate gain of the open-loop
system is, Simulated Examples
Three simulated examples are supplied to illustrate the pro-
posed identification and controller tuning algorithm. For com-
parison, the examples are the same as those used by Yuwana
and Seborg (YS, 1982) and Lee (1989). The first example uses
Equation (10) is the reciprocal of gain margin (GM) of the sys- first-order with dead-time processes which give no inherent
tem with proportional controller gain K,. Thus, the ultimate structural error according to methods of YS and Lee. The sec-
feedback controller gain can be obtained as, ond example illustrates the applicability of the method when
used in a process with high-order dynamics and large dead time
for which the YS method is inadequate. The third example dem-
onstrates the adequacy of the method when applied to under-
The resulting ultimate frequency and feedback controller damped process.
gain can then be used directly to determine controller settings
according to Ziegler-Nichols rules (Ziegler and Nichols, 1942). Example 1
For those tuning relations based on some form of parametric
process model, such as Cohen and Coon rules and Minimum This example considers a first-order process (Lee, 1989)
Error Integral tuning rules, the ultimate data can also be used to
e-dps
determine an equivalent approximation. For a first-order model Gp(S)= -
this is. s+ 1

Table 1. Estimates of Model Parameters for Example 1

The parameters of the first-order model which give the resulting

ultimate data can be obtained from the following relations,
1 1 0.5 1.5 1 1.02 0.507 1 1.01 0.516 1 1.08 0.480
2.0 1 1.03 0.554 1 0.98 0.520 1 1.05 0.490
1 1.0 1 1.15 0.988 1 0.99 1.01 1 1.10 0.971
1.5 1 1.27 1.14 1 1.03 1.05 1 1.04 0.993
2 0.5 1 1.46 1.67 1 0.98 2.07 1 1.13 1.922
1.0 1 1.59 2.09 1 1.04 2.03 1 0.99 1.991

2038 December 1989 Vol. 35, No. 12 AIChE Journal

 Table 2. Estimates of Model Parameters for Example 2 Table 3. Estimates of Model Parameters for Example 3

Model Ultimate Data Model Ultimate Data

Method K, K, 7, dm Pm K- Method Ke K, 7, dm Pu Km
Yuwana-Seborg 1.0 1 3.92 4.69 14.1 2.01 Yuwana-Seborg 2 1 4.69 4.20 13.2 2.45
0.25 1 3.57 3.19 10.1 2.44 I I 3.89 4.90 14.6 1.95
Lee 1.0 1 2.78 4.74 13.4 1.64 Lee 2 1 3.65 3.62 11.2 2.24
0.25 1 2.60 4.87 13.5 1.57 1 1 2.67 4.82 13.5 1.57
Proposed 1.0 1 3.01 4.44 12.9 1.77 Proposed 2 1 3.96 3.35 10.7 2.51
0.25 1 2.95 4.48 12.9 1.74 1 1 4.21 3.27 10.5 2.66
Ziealer-Nichols 1 2.89, 4.49* 12.9 1.73 Ziealer-Nichols 1 3.98* 3.36* 10.7 2.54
*Calculated from ultimate data *Calculated from ultimate data

Table 1 gives estimates for tests of two feedback gains and Notation .
three different dead times. The results show that the proposed A = magnitude of set-point change
method can provide consistent estimates subject even to various c, C(s) = controlled variable and its Laplace transform
degrees of process lag and testing condition. rm,.cpl. cp2 = first minimum, first and second peaks of c
c, = steady state of c
d , d,, d, = dead times of closed-loop response, model and process
G,(s), Gd(s) = controller and closed-loop transfer functions
Example 2 C,(s), GJs) = model and process transfer functions
The considered process dynamics is (Yuwana and Seborg, GM = gain margin
1982; Lee, 1989), H = overshoot
K , Kc, K,,, K, = closed-loop gain, controller gain, ultimate controller
gain and model gain
&s
P. = ultimate period
Gp(s) = (s + 1)2(2s+ 1) R (s ) = Laplace transform of set point
t,,, rp, = time of first minimum and first peak of closed-loop
response
Table 2 reveals testing conditions and the estimates obtained, { = damping factor
T , T,,,= process and model time constants
including results based on methods of YS and Lee. These results w, = ultimate frequency
also show that the present method can provide critical data and
PID settings which are more consistent with the Ziegler-Nichols Literature Cited
closed-loop cycling method.
Cohen, G. H., and G. A. Coon, “Theoretical Investigation of Retarded
Control,” Trans. ASME. 75,827 (1953).
Krishnaswamy, P. R., B. E. Mary Chan, and G. P. Rangaiah, “Closed-
Example 3 Loop Tuning of Process Control Systems,” Chem. Eng. Sci., 42,2173
The underdamped process of Yuwana and Seborg (1 982) is (1987).
Lopez, A. M. P., P. W. Murrill, and C. L. Smith, “Controller Tuning
considered, Relationships Based on Integral Performance Criteria,” inst. Tech.,
14,57 (1 967).
e-’ Lee, J., “On-Line PID Controller Tuning from a Single, Closed-Loop
Test,” AIChE J., 35,329 (1989).
Gp(s) = 9 s2 + 2.4 s +1 Yuwana, M., and D. E. Seborg, “A New Method for On-Line Con-
troller Tuning,” AIChE J., 28,434 (1982).
The comparison of results of YS, Lee and present methods, Ziegler, J. G., and N. B. Nichols, “Optimum Settings for Automatic
Controllers,” Trans. ASME, 64,759 (1942).
Table 3, demonstrates the robustness of proposed process char-
acterization method. Manuscript received May 23, 1989.

AlChE Journal December 1989 Vol. 35, No. 12 2039