Professional Documents
Culture Documents
Model-free adaptive (MFA) control, as its name suggests, is an SISO MFA controller, and a feedback loop. The control
adaptive control method that does not require process models. An objective is for the controller to produce an output u(t) to
MFA control system is defined to have the following properties: force the process variable y(t) to track the given trajectory
of its setpoint r(t) under variations of setpoint, disturbance,
1. No precise quantitative knowledge of the process is and process dynamics. In other words, the task of the MFA
available. controller is to minimize the error e(t) in an online fashion,
2. No process identification mechanism or identifier is where e(t) is the difference between the setpoint r(t) and the
included in the system. process variable y(t). The minimization of error e(t) is
3. No controller design for a specific process is needed. achieved by (i) the regulatory control capability of the MFA
4. No manual tuning of controller parameters is required. controller, and (ii) the adjustment of the MFA controller
5. Closed-loop system stability analysis and criteria are weighting factors that allow the controller to deal with the
1–3
available to guarantee the system stability. dynamic changes, disturbances, and other uncertainties of the
control system.
Derivations of the core MFA control technology address
specific control problems as described here:
4–14 MFA Controller Architecture
224
E0 = 1 wij
e(t)
N(.)
E1 p1 q1
Σ ϕ(.)
hi
Z–1
E2 p2 q2
Σ ϕ(.) q0 = 1
Z–1 o(t)
E3 p3 q3 v(t)
Σ ϕ(.) Σ ϕ(.) ψ(.) Kc
+
+
.....
.....
Z–1 qN
EN pN
Σ ϕ(.)
FIG. 2.15b
Architecture of a SISO MFA controller.
almost no memory, and MFA possesses the memory that is The weighting factors can be updated online at every
essential to a “smart” controller. sample interval using the following formulas:
FIG. 2.15c
Comparison of MFA and PID.
condition. If both controllers start from a sluggish situation, 2. Controller gain—use of a default value is recom-
MFA will control the process faster and better. mended.
3. Time constant—a rough estimate of the process time
MFA Control System Requirements constant in seconds.
4. Acting type—direct or reverse action of the process.
As a feedback control system, MFA requires the process to If the process input increases and then its output
have the following behavior:
increases, it is direct acting, and vice versa. However,
M FA controllers embedded in various platforms
1. The process is controllable. always use the vendor’s definition. Sometimes, con-
2. The process is open-loop stable. troller acting type is used, which is different than the
3. The process is either direct or reverse acting (process process acting type.
does not change its sign).
According to the principles in the information theory, it
If the process is not controllable, improvement of the process is required that sample interval be less than or equal to one
structure or its variable pairing is required. If the process is not third of the time constant. That is,
open-loop stable, it is always a good practice to stabilize it first.
However, for certain simple open-loop unstable processes such
1
as a non-self-regulating level loop, no special treatment is Ts ≤ T, 2.15(7)
required when using MFA. If a process changes its sign within 3 c
1
its operating range, special MFA controllers are required.
where Ts is the sample interval, and Tc is the time constant.
SISO MFA Configuration Once the configuration is done, MFA can be launched at
any time and will control the process immediately. MFA does
A SISO MFA controller has only a few parameters to configure: not require process identification and is not a dynamic modeling-
based controller; there is no need to first collect data to train
1. Sample interval—the interval between two samples the model. MFA controllers can be switched between auto-
or calculations in seconds. A high-speed MFA con- matic and manual at any time. No specific bumpless transfer
troller can run at a 1 millisecond rate. procedure is required.
d(t)
uf (t) Feed-
forward
MFA
Process
Gp2
+ + y2(t)
r(t) e(t) MFA uc(t) u(t) Process y1(t) y(t)
+ controller + Gp1 +
–
FIG. 2.15f
Feedback and feedforward MFA control system.
FIG. 2.15e
Configuration menu of MFA pH controller. general-purpose feedforward controller. It does not attempt a
perfect cancellation of the disturbances, which is very difficult
to implement in industrial applications due to changing process
for the large nonlinear gain changes. In addition, it can control dynamics and operating conditions. A feedback/feedforward
the full pH range with high precision and enables automatic MFA control system diagram is illustrated in Figure 2.15f,
control of acid or alkaline concentration, both of which are where Gp1 is the main process and Gp2 is the process with dis-
critical quality variables for the chemical process industry. turbance input and the process variable as output.
FEEDFORWARD MFA CONTROLLER The feedforward MFA time constant can be an estimate of
the time constant of Gp2. This is related to how fast the
Feedforward is a control scheme to take advantage of distur- disturbance will affect the process variable (PV).
bance signals. If a process has a significant measurable dis-
turbance, a feedforward controller can be used to reduce its
effect before the feedback loop takes corrective action. A good ANTIDELAY MFA CONTROLLER
feedforward controller can improve the control system perfor-
mance economically. Many processes have large time delays due to the delay in
Feedforward compensation can be as simple as a ratio the transformation of heat, materials, and signals. No matter
between two signals. It could also involve complicated energy what control action is taken, its effect is not measurable
or material balance calculations. The feedforward MFA is a during a period of time delay. This is equivalent to disabling
d(t) to a certain point, cools off, forms its shape, etc. Antidelay
r(t) e(t) u(t) Process x(t) + y(t) MFA makes it possible for process industries to achieve six
MFA
+ controller
with large
+ sigma or zero defects quality control objectives.
– time delays In a semicontinuous production environment, the process
line speed may change as many as 100 times or more, which
yc(t)
will cause the delay time to change on a similar scale. Since
Delay
predictor the line speed is measurable, the delay time can be easily
calculated and provided to the antidelay MFA controller in
real time. In this way, the control performance will not sac-
FIG. 2.15g rifice much even during large line speed changes.
Antidelay MFA control system.
On the other hand, if the delay time of a process changes
on a scale of more than five times, and the delay time infor-
mation cannot be provided to the controller, the time-varying
the feedback for a period of time, where feedback informa- MFA controller will be more suitable for this application.
tion is essential to automatic control.
If a PID is used to control a process with significant time
delays, the controller output will keep growing during the ROBUST MFA CONTROLLER
delay time and cause a large overshoot in system responses
or even make the system unstable. Typically, a PID has to be In complex control applications, the following challenges
detuned significantly in order to stay in automatic but will may occur:
sacrifice control performance. Generally speaking, a PID con-
troller usually works for the process if its τ-T ratio (delay 1. A large change in the system dynamics occurs, so that
time/time constant) is smaller than one, unless it is detuned. a prompt control action is required to meet the control
When a controller is detuned, it loses the sharpness of its performance criteria.
control capability, so the process cannot be tightly controlled. 2. The dominant disturbance to the system cannot be
The Smith predictor is a useful control scheme to deal with economically measured, and therefore feedforward
processes with large time delays. However, a precise process compensation cannot be easily implemented.
model is usually required to construct a Smith predictor. Oth- 3. A controller purposely detuned to minimize the vari-
erwise, its performance may not be satisfactory. ations in its manipulated variable may lose control
Figure 2.15g shows a block diagram for a SISO antidelay when a large disturbance or significant dynamic
MFA control system with an antidelay MFA controller and behavior change occurs.
a process with large time delays. 4. The system dynamic behavior or load change does not
A special delay predictor is designed to produce a dynamic provide triggering information to allow the control
signal yc(t) to replace the process variable y(t) as the feedback system to switch operating modes.
signal. The idea here is to produce an e(t) signal for the con-
troller and let it “feel” its control action without much delay For instance, controlling the reaction temperature for a batch
so that it will keep producing proper control signals. In other reactor is always a challenge due to the complex nature of the
words, the artificial dynamic signal yc(t) is able to keep the process, large potential disturbances, interactions between key
feedback loop working even when there is a large time delay. variables, and multiple operating conditions. A large percent-
Since the MFA controller in the system has adaptive capability, age of batch reactors running today cannot keep the reactor
the delay predictor can be designed in a simple form without temperature in automatic control throughout the entire operat-
1,2
knowing the quantitative information of the process. ing period, thus resulting in lower efficiency, wasted manpower
Compared to the traditional Smith predictor, the antidelay and materials, and inconsistent product quality.
MFA controller does not need a precise process model. It An exothermal batch reactor process typically has four
only needs an estimated delay time as the basic information operating stages:
for its delay predictor. If the delay time used in the MFA delay
predictor has a mismatch with the actual process delay time, 1. Startup stage: ramps up the reactor temperature by use
the controller is robust enough to deal with the difference. of steam to a predefined reaction temperature.
Typically, it can deal with the situation where the delay time 2. Reaction and holding stage: holds the temperature by
is two to five times larger or smaller than the actual delay time use of cooling water while chemical reaction is taking
with satisfactory control performance. In addition, there is no place and heat is being generated.
real limitation on how large the τ-T ratio is, as long as an 3. No-reaction and holding stage: holds the temperature
estimated delay time is provided. by use of steam after the main chemical reaction is
The antidelay MFA controller is especially useful in con- complete and heat is not being generated.
trolling process quality variables since a quality variable is 4. Ending stage: ramps down the reactor temperature for
typically measured after the product or process material travels discharging the products.
During the transition period from Stage 2 to Stage 3, the reactor output (OP), where a hard limit or constraint can be
can change its nature rapidly from a heat-generation process set. PV is a process variable that can only be controlled
to a heat-consumption process. This change happens without by manipulating the OP. Therefore, the upper and
any triggering signal because the chemical reaction can end at lower bounds for PV are very different from the OP
any time depending on the types of chemicals, their concen- constraints.
tration, the catalyst, and the reaction temperature. Within a 2. Gain ratio—The coefficient to increase or decrease
very short period of time, the reactor temperature can drop the MFA control action. Typically, you want to enter
significantly. The control system must react quickly to cut off gain ratio = 3, which implies that the MFA gain work-
the cooling water and send in a proper amount of steam to ing in abnormal situations is three times higher than
drive the reactor temperature back to normal. A regular feed- the regular MFA gain setting. It is important to under-
back controller is not able to automatically control a batch stand that this is not a gain scheduling approach,
reactor during this transition. In practice, batch reactors are although it appears to be this way. Gain scheduling
usually switched to manual control and rely on well-trained will not be able to resolve the complex problems
operators during critical transitions. It is a tedious and described.
nerve-wracking job that can result in low product quality
and yield.
The robust MFA controller is able to control the problematic TIME-VARYING MFA CONTROLLER
processes described. Without the need to redesign a controller,
using feedforward compensation, or retune the controller param- The time-varying M FA controller is used to control a pro-
eters, the robust MFA controller is able to keep the system in cess with large time constant and /or delay variations. For
automatic control through normal and extreme operating condi- instance, a temperature control loop usually has a shorter
tions when there are significant disturbances or system dynamic time constant when it heats up and a much longer time
changes. constant when it cools down because adding heat to the
process is much faster than taking it away. Also, a line speed
or flow rate change will cause the process delay time to
Robust MFA Controller Configuration vary significantly.
1. Upper and lower bound—the bounds for the process As shown in Figure 2.15i, the time-varying MFA controller
variable (PV) being controlled. They provide “intelli- can be easily configured with an estimated minimum and
gent” upper and lower boundaries for the PV. These maximum process time constant plus delay time. The con-
bounds are typically the marginal values that the PV troller is able to deal with the large time constant and/or
should not go beyond. PV is unlike the controller delay time changes without having to retune any parameters.
Without losing generality, we will show how a multivariable 1. The regulatory control capability of the MFA control-
model-free adaptive control system works with a two-input two- lers
output (2 × 2) system as illustrated in Figure 2.15k, which is the 2. The decoupling capability of the MFA compensators
2 × 2 arrangement of Figure 2.15j. In the 2 × 2 MFA control 3. The adjustment of the MFA weighting factors, which
system, the MFA controller set consists of two controllers—C11 allow the controllers to deal with the dynamic changes,
and C22 —and two compensators—C21 and C12. The process has large disturbances, and other uncertainties
four subprocesses—G11, G21, G12, and G22.
3. Time constant—a rough estimate of the process time In natural science, the combination of physics, mathe-
constant in seconds. matics, and philosophy plays an integral part in developing
4. Acting type—direct or reverse acting of the process. a theory that is practically useful. Physics is the foundation
5. Compensator gain—to deal with the interaction from for the study of the physical process or environment, math-
the other loop. ematics provides the tools to precisely describe the physical
process or phenomenon, and equally important is the philos-
ophy that provides directions.
MIMO MFA CONTROLLER APPLICATION GUIDE The development of model-free adaptive control technol-
ogy started from a simple desire to develop a new controller
A MIMO system can be much more complex than a SISO that could easily and effectively solve various industrial con-
system; precautious have to be taken when applying a MIMO trol problems. The actual development process has evolved
MFA controller. When designing a multivariable control sys- from a prolonged interest in the study of combined intelli-
tem, the first step is to decide which process variable is paired gence methodology. Since model-free adaptive control does
with a manipulated variable. A MIMO MFA control system not follow the traditional path of model-based adaptive con-
should follow these pairing rules: trol, the philosophy behind the combined intelligence has led
the way up this long and rocky road.
1. Each process of the main loops has to be controllable,
open-loop stable, and either reverse or direct acting.
2. A process with a large static gain should be included SUMMARY
in the main loop as the main process (G11, G22), and
a process with a small static gain should be treated as To see how the MFA control method is developed based on
a subprocess (G21, G12). the combined intelligence methodology, we will relate MFA
3. A faster process should be paired as the main process to each of four key points.
and a slower process, and processes with time delays
should be treated as subprocesses. Simple Solution
4. If pairing rules 2 and 3 should result in a conflict, a
PID control is simple since it is a general-purpose controller
tradeoff is the only option.
and its algorithm is easy to understand. However, PID is
almost too simple to control complex systems. In this regard,
In addition, an MFA control system should be designed based PID cannot be considered an effective solution to the more
on the degree of interactions between the loops. Table 2.15l difficult control problems. On the other hand, model-based
lists the control system design strategy based on the degree of advanced control methods have proven themselves too com-
interaction of a MIMO process. plex to launch and maintain since they depend on either a
first principle or an identification-based process model. A
dream controller has to be powerful enough to control various
MFA CONTROL METHODOLOGY complex processes yet simple enough to use, launch, and
maintain. MFA is a solution that fits these requirements.
“All roads lead to Rome.” A problem usually has multiple
possible solutions, and a process can usually be controlled Use All Information Available
using different controllers based on different control methods.
Almost every control method has its merits and weakness. Model-free adaptive control, as its name suggests, is a con-
What is important is to use the right controller to fit the trol method that does not depend on either first principle or
application at a minimum cost. identification-based process models. However, we do try to
use all the process information available. For this reason, it
can be considered an information-based controller.
For instance, process time constant defines how fast a
TABLE 2.15l dynamic system responds to its input. A slow process might
MIMO System Design Strategy have a 10-hour time constant and a fast process might have
a 10-millisecond time constant. It would be unwise not to
Interaction Measure Control Strategy
use this information for the controller. In addition, it is rela-
Small to no interaction Tighten both loops with SISO MFA tively easy to estimate the time constant by reading a trend
Moderate interaction Tighten important loops with SISO MFA chart. Other important yet easily obtained information about
and detune less important loops or use a process includes its acting type (either direct or reverse),
MIMO MFA for better overall control static gain, and delay time if any. An MFA controller is
Severe interaction Use MIMO MFA to control the process; designed to use the process parameters that can be easily
may need to de-tune less important loops estimated.
A process can be classified as a white, gray, or black box. If its 1. Cheng, G. S., MFA in Control with CyboCon, Rancho Cordova, CA:
input–output relationship is clear, the process is a white box. CyboSoft, General Cybernation Group, Inc., March 2002.
We can easily use existing well-established control methods 2. VanDoren, V., Techniques of Adaptive Control, Burlington, MA:
and tools to design a controller for this process. Butterworth-Heinemann, 2003.
3. VanDoren, V., “Model Free Adaptive Control—This New Technique
When we are not sure if the process input–output relation- for Adaptive Control Addresses a Variety of Technical Challenges,”
ship is accurate, or if the process has potential disturbances, Control Engineering Europe, March 2001.
dynamic changes, and uncertainties, the process is a gray box. 4. Major, M., “Model-Free Adaptive Control on an Evaporator,” Con-
In this case, MFA’s adaptive capability is able to handle such trol, September 1998.
changes and uncertainties. PID or model-based control methods 5. Cheng, G. S., He, M., and Li, D. L., “Model-Free Coking Furnace
Adaptive Control,” Hydrocarbon Processing, December 1999.
will have a much tougher time or higher cost addressing these 6. Seiver, D., and Marin, O., “Air Separation Advances with MFA
uncertainties. Control,” Control, May 2001.
7. Cheng, G. S., and Huo, L. Q., “Control System Optimizes EOR Steam
Generator Output,” Oil & Gas Journal, September 2003.
8. Cheng, G. S., and Huo, L. Q., “Leading Edge Plant Control System—
Technique That Fits the Application Model-Free Adaptive Control Puts Ling-Yuan Iron and Steel in
League of Its Own,” InTech, October 2003.
MFA is neither model based nor rule based. We might say 9. Cheng, G. S., CyboCon CE User Manual, Rancho Cordova, CA:
that it is an information-based control method. If the argu- CyboSoft, General Cybernation Group, Inc., 2001.
ment is made that the process information used is equivalent 10. Cheng, G. S., “Tomato Process Improvements Using Model-Free
Adaptive Controllers,” Tomato News, April 2004.
to a process model, that is perfectly acceptable. The key to 11. Harris, S., “Model-Free Adaptive Control Improves Productivity and
this approach is that we focus on delivering a simple, adap- Efficiency,” Control, June 2004.
tive, and effective solution. 12. Cheng, G. S., “Model-Free Adaptive (MFA) Control,” IEE Comput-
To extend this idea, a series of MFA controllers, many of ing and Control Engineering, June/July 2004.
which are described here, has been developed to address dif- 13. Cheng, G. S., and Zhang, Z. W., “Model-Free Adaptive Control in
Water Treatment,” Control Engineering Europe, September 2004.
ferent difficult control problems. Users can simply select the 14. Cheng, G. S., and Zhang, W. D., “Model-Free Adaptive Technology
appropriate MFA, configure its parameters, launch the control- Improves Distillation Column Chain Control,” Hydrocarbon Process-
ler, and reap the benefits. ing, October 2004.