Escolar Documentos
Profissional Documentos
Cultura Documentos
2. Expert Control
current and past information about the process under control. td&(t)
Many current industrial controllers include some heuristics for
safety net procedures [5].These heuristics may only handle extreme
“alarm” type situations, and a prescribed solution may be a plant y(t)
shutdqwn until human intervention can solve the problem. An expert 4output Process
controller should have the ability to udapt to changing situations and
An expert controller should have the capability of using several Figure 1: PID control loop.
different control algorithms as well as the ability to lune the param-
The key parameters in PID control are the proportional constant
eters of each algorithm to the process (loop) under control. Possible
control algorithms might include Proportional, Integral and Deriva- ICp, the integral constant t , and the derivative constant td. These
three “constants” can be manipulated to produce various response
tive (PID), pole-placement, linear observers or algorithms designed
curves from a given process. The whole concept for adaptive PID
for optimal control. Ultimately, the expert controller should have
control revolves around interactively changing these three values.
a library of relevant algorithms for process control and identifica-
The discrete time equivalent expression for PID control used in this
tion. The expert system would orchestrate the application of these
study is shown in equation (2). In equation (2), t is the sample period
algorithms depending on current conditions [5].
for the controller.
‘Now with Techno-Sciences, Inc., Suite 620, 7833 Walker Drive,
Greenbelt, MD 20770.
‘This research was supported in part by a grant of equipment and funds from
the Industrial Systems Division of Texas histruments, in part by the Engineering
Research Center of the College of Enginecring, University of hfaryland, and in
part by the Systems Research Center under NSF Grant CDR-85-00108.
3Specifically, the forward chaining production system YAPS and the object- The approach used in designiiig the Adaptive PID Controller
oriented Flavors system running on a \‘AX ll/T80. (APIDC) is based upon the methodology a human might use, given
Setpjoint
u(t) : determination of a satisfactory response characteristic. The mem-
bership functions have a range of zero to one, where values near
T I 565 I Process one indicate “satisfactory” values. As the response characteristic
* Y(t) i move towards one “bandwidth” unit from the optimal value, “nor-
mal” membership functions return values closer to zero. This may
0
565
the controller as a destab ng element from the feedback loop5 used in building a set of such rules for adaptive tuning.
The monitor procedure then terminates with the stability flag set Through simple, although extensive, trial and error analysis such
t o false (zero). The APIDC enters a waiting phase to give the process trends have been observed. This trial and error procedure consisted
time t o settle naturally before any other action is taken. When the of running tests with a significant cross-section of processes. The
process has settled, the rules are tested sequentially. The first rule testing procedure was as follows: Starting values for the parame-
is the alarm rule which fires. The pointer EOF (for End Of File) ters IC,, t, and t d were chosen arbitrarily and the process response
causes a transition to the end of the rule base, and no further rules to a setpoint change (or unit step disturbance) was monitored and
are tested. There is no need in trying any other rules if an alarm .noted. Then each of the twenty-seven possible choices for tuning
condition exists. The direct consequent of the alarm rule is t o halve the parameters was tried. The amount to adjust the parameters was
IC,, double t d , and leave t , alone. This has been found by experience varied between ten percent and twenty percent in different runs.
to be a good reaction to unstable responses; and it has the general By observing trends in the reactions of the six criteria of settle-
result of bringing the response back under control. time, overshoot, risetime etc. a rule could be derived. As the test-
One of the most important features of the APIDC is a history ing became more complete, the two typical process response types,
file which maintains a body of information for each tuning pass. At named type A and type B, became somewhat clearly defined, and
present the history file is used exclusively for information pertaining which direction t o tune the parameters for specific results became
t o adaptive PID control, but the concept and its usefulness holds more evident. In addition, an indication as to the amount to adjust
for expert control in general. The information saved on each pass each parameter relative to the others could also be ascertained. The
contains the following: the acronym for each rule that fired, the k,, information taken from all the testing for a Type A process response
ti and t d values used as result of the rules that fired, the results from is embodied in the structure Type A in Table 1.
the process response for each of the six criteria, the membership
ratings and direction off for each, and the score for this pass. (For
examples of history file see below.)
This information plays a significant role in the conditions tested
for by about half of the general rules in the system. (Sixteen rules
are currently implemented.) In addition to its internal role in the
system, the history Me, this feature provides a means for the user to
review the events that have taken place up to a given point in time.
The history file is stored internally in memory in a special matrix‘
data structure. It is also periodically written to disk for safe keeping
should the system crash. Table 1: Type A Tuning Rules
Several rules pertain to the situation where it is time to stop This table is used in the MBTA algorithm along with current
actively tuning the process response. One of three conditions may membership ratings to generate a tuning decision consisting of the
apply. These are: (i) when the score has been steady for five passes, percentage and direction each PID parameter should be adjusted.
(ii) when the score has been decreasing for five passes, and (iii) when The application of the information in Table 1 can be conceptually
the current score is higher than 5.5 ( a very high score) and has re- viewed as a body of twelve “if-then” type rules. However, the in-
mained in this approximate region for a t least two passes. The pur- ternal representation of the Algorithm is numerical in form with the
pose of each of these rules is t o have the controller recognize when it “if-then’’ portions buried in the code. Such a rule is (in psuedo-
is no longer making useful progress (perhaps the MBTA algorithm code):
has converged), or when it is actually headed in the wrong direction
after already finding a satisfactory PID control. The action taken as IF (settletime needs to be decreased) THEN <
a result of each of these rules firing is the same. A data structure Xkp-change = .7*(1 - membership(current-settletime))
called “bestpass” which holds an integer (or index) pointer to the %ti-change = . 6 * ( 1 - membership(current-settletime))
pass with the highest score on record is maintained by the APIDC. Xtd-change = .7*(1 - membership(current-settletime))
The APIDC accesses the values in the history file for IC,, t , and t d >
associated with this best score and assigns them as the values to be This rule uses two sources of information to compute the per-
used for subsequent passes. A best score is also saved and a dedi- centage changes for the PID parameters due to a need to decrease
cated monitoring stage is entered by the assignment of a flag called the settletime for a Type A curve. The values 0.7, 0.6 and 0.7 are
“monitor score.”
taken from Table 1, and are referenced by choosing the row for settle-
Rule number two for monitoring the score is then fired on every time and the subcolumn labelled “dec” (for decrease) under each of
subsequent pass of the process, until some significant change occurs
the respective columns for the three PID parameters. These values
in the process output. Should the score for a given pass drop more
are the experience factors learned from the trial and error testing.
than 0.5 below that with the best score, it is assumed the process The other key piece of information is the membership rating for the
has changed. Dedicated monitoring (via rule number two) is disabled
current value of the settletime. The rule for increasing settletime
and the entire tuning process, beginning with Zeigler-Nichols auto-
(usually undesirable as a goal) has the same format except that 0.7,
tuning, is started again.6 0.G and 0.7 are replaced respectively by -0.7, -0.3 and -0.7. Two
3.4.2 Membership Based Tuning Algorithm
companion rules like these can be defined for each of the six criteria,
General observations may be made about the behavior of a given
resulting in twelve rules in all. Twelve similar rules are implemented
process to adjustments in the PID parameters IC, t , and t d . Given
for Type B processes.
enough experience in tuning PID controlled processes, a plant control
For a typical tuning pass, six of the twelve possible rules will
engineer will often know which way t o turn each of the “knobs”
“fire” resulting in six indicated percentage adjustments for each of
to influence the process output. For instance, it may be possible
the three PID parameters. These values are added together to com-
to say that if one wishes to decrease the percentage of overshoot,
pute a total adjustment for each respective PID parameter. It is
then k, should be increased, ti should be decreased and t d should be
important to note that the experience factors in the table differ in
increased. As it turns out, there are some general trends that can be
sign. Thus some force a percentage increase in the parameter and
51n the present implementation we assume that all processes to be controlled some force a decrease. This corresponds to the coupled nature of dif-
have stable natural responses Shutdown or transition to a known stabilizing gain
ferent criteria. That is, action taken with the intention of decreasing
could also be used.
60f course, if auto-tuning has been disallowed by the nser, during controller
settletime is likely increase the percentage overshoot. Hence, the six
initialization, then the nser defined starting values are used Cor the three PID individual adjustments tend to counteract each other with the final
parameters. result representing a compromise.
566
The membership rating of each criteria for a given pass directly the most important criteria. The APIDC continued employing B
influences the final adjustments in the PID parameters. By includ- type tuning until iteration 19 before the steady score rule fired and
ing the arithmetic compliments for membership ratings in the com- PID parameters very close to those found on iteration 11 were settled
putation, the resulting decision or adjustment is reacting directly to upon.
the current process response. Those criteria with the lowest ratings It is important to note that although it took a significant number
dominate the percentage changes, yet they are tempered by the ad- of passes to settle upon the final PID parameters, the process was
justments indicated from all other criteria. The final adjustment for constantly under “good” PID control. That is, control was main-
each parameter is applied as a percentage change over the old value. tained while the controller tried small adjustments in an effort t o
The next pass implements PID control with these new parameters. find the optimal response curve. All coarse adjustments occurred in
This algorithm converges in a large number of cases. As member- the first four or five passes.
ships get higher, the resulting adjustments clearly get smaller and Although the system described here treats it in a very simple
more refined. way, the key to successful “expert control” is model identification.
Not all processes fit so neatly into type A or type B. In general, it . Many techniques exist for process identification but most involve
is likely that a major adjustment in the parameters might be needed repeated “offline” testing of the process. However, “off line” testing
to intialize the tuning procedure. The adjustments in the MBTA al- is not informative when the process is prone t o change. The answer
gorithm, while they reflect the membership ratings, are often rather is to build an expert controller that can learn about the process
small. Making the adjustments too coarse would effect the conver- “on line” while maintaining some level of control. From what it
gence of the algorithm. Also situations such as instability may arise learns, an internal model or configuration of the current process can
and must be dealt with directly. For these purposes a group of more be built. Based on this model and limited only by its completeness
general rules are included in the APIDC. It is also necessary t o have and accuracy, the best possible control can employed. The more
a means to choose which type of MBTA tuning is to be be attempted. the expert controller knows the better the control will be. An expert
The specific type of MBTA tuning is implemented as the result of controller requires the ability to learn about a process; and so, process
one of these general rules firing. One rule exists for each of the two identification is the key factor.
types.
0 5 10 15 20
time (seconds)
‘“Reasonable” is as determined by the user, but scpres of 3.5 or higher are
generally accepted as reasonable.
Natural Response
567
2
1.5
I .5
U j
0
0
0 5 10 15 20 0 5 10 15 20
time (seconds) time (seconds)
Response After Auto-Tuning Tuning Pass No. 6
1.5
.5
0 5 10 15 20
time (seconds)
Tuning Pass No. 8
Criteria Value Membership Direction Off Criteria Value Membership Direction Off
Settling time 8.88 0.0 High Settling time 8.599 0.002 High
Overshoot 0.117 1.0 Overshoot 0.0 1.0
I
569