Adaptive Control System

ASSESMENT 1

Topics for Adaptive Control Systems

Submitted by: Muhammad Nabeel Hashmi

Submitted to: Dr. Tariq Rehman

Date: 23Rd April ‘23

 TOPICS

1. AN INDUSTRIAL ADAPTIVE PID CONTROLLER

 Introduction
 An Industrial Adaptive Controller
 Auto-Tuning and Gain Scheduling
 Adaptive Feedback Control
 Adaptive Feed forward
 Industrial Experiences
 Conclusion

2. Stability of Adaptive System

 Introduction
 Adaptive Stabilization of Nonlinear Systems in Normal Form
 Model-Reference Adaptive Control of Scalar Linear Systems
 The Standard Adaptive Control Problem
 Uniform Asymptotic Stability of Adaptive Systems
AN INDUSTRIAL ADAPTIVE PID CONTROLLER

1. Introduction

Adaptive control hat been one of the major research areas in automatic control since the early seventies. Two
different ways to use adaptive techniques have emerged. The first is auto-tuning, which means that the
controller is tuned on operator demand. This is done by estimating the process dynamics and calculating the
controller parameters. The second is the true adaptive controller, i.e. a controller which adapt itself to
changing process dynamics without operator demand. Auto tuners, particularly for tuning of PID controllers,
have been accepted and well established among manufacturers of controllers and instrument systems in
recent years. Although many successful applications of adaptive control have been presented, see e.g. Astrom
(1987) and Seborg et al (1986), only few multi-purpose adaptive controllers have appeared on the market.
There are several reasons for this: Their proper use requires a certain expertise. They still are too complicated
to use for non-experts. Since adaptations continuous they require a significant safety network to guarantee
proper operation in all cases. Adaptation speed is often limited for robustness reasons. An adaptive controller
needs some initial parameter settings which are very crucial for the properties of the control, such as sampling
period, model order, and process dead time. If these parameters are given unsuitable values, the adaptive
controller may behave badly. Many adaptive controllers are based on a sampled controller, with sampling
period which is of the same magnitude as the process time constant. These controllers introduce extra dead-
time in the control loop because of the sampling. Such a controller is often not suitable for control loops
subject to load disturbances .See McMillan (1986). An adaptive controller needs some time to find new
parameters when the dynamics change. If a very rapid rate of adaptation is required, a gain-scheduling type of
controller is needed. Gain-scheduling means a kind of feed forward. If we know that the process dynamics are
different at different operating points (due to nonlinearities) or at different times (due to changes in
production rate), the controller can be preprogrammed with different parameter settings for these different
situations. This paper presents a new type of adaptive controller. It can be viewed as a natural extension of
the relay auto-tuner in Astrom and Hagglund (1984). This auto-tuner was based on estimation of a point on
the NY Quist curve where the process has approximately 180· phase shift. In the adaptive controller this point
is tracked by a parameter estimator. The regulator is a PID controller with fast sampling which gives close to
continuous time behavior. The approach has several advantages. First Of all it is very easy to use. No
parameters have to be given by the operator since it is initialized by an auto tuner procedure. It can thus be
used by normal process operators. Since the process model only has two parameters, identifiability is easy to
achieve even under normal operating conditions. It is also possible to get a relatively fast adaptation rate. The
paper is organized as follows. The use of adaptive technique in industrial process control is discussed in
Section 2. The relay auto-tuner used in the initialization phase of the adaptive controller is briefly described in
Section 3, which also contains a discussion of the use of gain scheduling. The adaptive feedback control based
on tracking a point on the Nyquist curve is presented in Section 4. The section contains the principles of the
identification as well as a description of the supervisory procedures. The adaptive feedforward compensation
is presented in Section 5. Section 6 contains a short description of the industrial process controller ECA400,
followed by examples of its use in the process industry.
2. An Industrial Adaptive Controller
There are many different possibilities to use adaptive techniques for industrial process control. There is a wide
range of choices of model structures, identification methods, and control design techniques. Operational
issues like supervisory techniques, safety networks and user interfaces add to the complexity. In this section
we will discuss some of these choices in order to motivate the choices that led to the controller discussed in
this paper. The first adaptive controllers for process control were announced in 1983. There is now about five
years of experience of using such devices. Experience has also shown that PID regulators can handle many of
the industrial problems. The main exceptions are systems whose dynamics is dead-time dominated, systems
with oscillatory dynamics and systems with significant stochastic disturbances. In such cases regulator
structures otherthan the PID may give significant benefits. The PID controller also has the added benefit that
users are familiar with it. The industrial benefits of using feed forward control have also been demonstrated
very clearly particularly in applications with the Novatune regulator. Feedforward is very useful but it require.
Reasonably accurate process models. It has also been demonstrated that automatic tuning which makes it
possible to keep regulators well-tuned is a very desirable feature. Often dynamics does not change very much.
If it does the changes can often be correlated to measurable signals and thus compensated by gain scheduling.
It thus appears that a device capable of realizing pm feedback control and feed forward control could be a
useful component for industrial process control. If the regulator is provided with facilities for automatic
tuning, gain scheduling and adaptation it will also be very easy to use. The diagram in Figure 1 gives another
view on how the different features can be used. First of all it can be stated, that if the requirements on the
control are low, one can often solve the problem using a controller with constant parameters tuned according
to the "worst case". With larger demands on the control, the way in which the adaptive technique should be
used is determined by the way the process dynamics vary. If the process dynamics are constant, the controller
dynamics should also be constant. An autotuner procedure is then useful to set the controller parameters
once and for all. If the process dynamics are varying, the controller should compensate for these variations by
varying the controller parameters. We distinguish between two types of variations, predictable variations and
no predictable variations. The predictable variations are typically caused by nonlinearities in the control loop.
These variations are best handled by using a gain schedule. An autotuner procedure is useful to build the
schedule by finding the different sets of controller parameters. The second type of process dynamics
variations are those which are not predictable. These may be caused by no measurable variations in raw
material, wearing, etc. These variations cannot be handled by gain scheduling. The true adaptive controller is
the only way to make the controller follow the process variations. As will be seen below, an autotuner
procedure is useful even in this case to initialize the adaptive controller. The variations of the process
dynamics may of course consist of both predictable and no predictable parts. A combination of gain
scheduling and adaptive control is then suitable. A diagram analogous to Figure 1 could also be drawn for the
feed forward compensation. It is often difficult to tune the parameters in a feed forward compensator
manually, since the operator often not has access to manipulate the disturbance from which the feedforward
is made. In the feedback case, the control performance can be determined by changing the reference signal. In
the feedforward case, the operator often has to wait for suitable transients in the disturbance signal before he
can decide if the compensator parameters are suitable. Therefore, an adaptive algorithm is particular useful in
the feedforward case, since the adaptive algorithm continuously waits for transients in the disturbance signal,
and adjusts the compensator parameters based on the transient response. Adaptation is therefore useful even
if the dynamics between the disturbance and the measurement signal are constant. Initialization or Pre-tuning
Initialization is an important issue for an industrial controller. Some of the early adaptive regulators were very
demanding on the user. It was necessary to know parameters like sampling rate, dead-time, model order,
desired response time etc. This made use of the controllers very exclusive. It was necessary to have special
skills to commission and use the controllers. Many of the early adaptive regulators were therefore provided
with a pre-tune feature that was intended to help the operators to derive the required knowledge. The
pretune feature W&I often based on an open loop step response measurement or some other transient
response experiment. This will however also require some prior knowledge like the size of the step and how
long we have to wait for steady state. In some cases the pre-tune required a closed loop experiment. To make
this it is however necessary to know values of the regulator parameters that will give a stable response. In an
industrial adaptive controller, the properties of the initialization phase is of great importance. If the controller
is supposed to be operated by personnel not familiar with adaptive control, and perhaps with a limited
knowledge about process dynamics, it is not possible to force the operator to make decisions which will
determine the performance of the adaptive controller. In the system discussed in this paper this is made using
relay feedback which has proven very reliable.

3. Auto-Tuning and Gain Scheduling

The key idea behind the automatic tuning is to use relay feedback. Processes with the dynamics typically
encountered in process control will then exhibit limit cycle oscillations. The autotuner identifies one point on
the Nyquist curve of the process from a simple relay experiment. The autotuner principle is shown in Figure 2.
When the operator decides to tune the controller, he simply presses a button. This switches out the Pill
algorithm and replaces it WIth a nonlinear function which can be described as a relay with hysteresis. The relay
causes the process to oscillate with a small and controlled amplitude. The process perturbation is very close to
the optimum for most of the usual pm design methods. The frequency of the limit cycle is approximately the
ultimate frequency where the process has a phase lag of 180·. The ratio of the amplitude of the limit cycle and
the the relay amplitude is approximately the process gain at that frequency. A relay feedback experiment thus
determines a point on the Nyquist curve associated with the open loop dynamics that is close to the ultimate
point. A reasonable pm regulator can be designed based on this point. This idea, which was presented in Astrom
and Hagglund (1984) and (1988), has been used in commercial auto-tuners since 1984. It has proven to be a
convenient tool for fast tuning of PID controllers. One of its main advantages is that it admits one-button-tuning.
This means that no prior information has to be given from the operator. Tuning is executed simply by pushing
the tuning button.

Another advantage is that the tuning experiment is executed under tight feedback control and that the
experiment generates an input signal that is close to optimal for determining the ultimate point on the
Nyquist curve.
In Figure 3, the use of the autotuner in a flow control loop in a chemical plant is presented. The tuning took
about 15 seconds in this case. The two set-point changes show that the resulting PID controller is well tuned.
(The signals are not smooth because of a sticky valve.) By using the autotuner, we do not only obtain initial
values of the PID parameters which manage to control the process, but as will be seen in the next section, also
valuable process information for the initialization of the identification procedure in the adaptive controller.
The design method
Given the information of one point on the Nyquist curve, many design procedures can be used. Unfortunately,
there is no design method which will suit all types of processes. By deciding to use a PID controller, we have
restricted ourselves to those control problems which are suitable to solve with PID controllers. But even among
these problems, it is desirable to have different design procedures for different control problems.The
identification procedure given in the previous section gives us the information of one point G(iw) on the Nyquist
curve. By introducing the PID controller G PlD( iw) in the control loop, it is possible to give the Nyquist curve of
the compensated system GGPlD a desired location at the frequency w. For most purposes, we have decided to
choose the PID parameters so that G( iw) is moved to the point.

This design method can be viewed as a combination of phase and amplitude-margin specification. Since we have
three adjustable parameters, K, Ti and Td, and the design criterion can be obtained with only two parameters,
we furthermore require that

Some very simple control problems , where the process is approximately a first order system, can be solved
effectively with a PI controller with relatively high gain. For these problems, the D-part of the controller will
not be of any help. Furthermore, since we have a high gain the noise will be much amplified through the D-
part. Therefore, it is desirable to use only a PI controller in these cases. In our controller, we can automatically
detect this kind of processes and thereby switch off the derivative gain in these cases . For this PI controller,
we have chosen the following design:
Where a. = |JG(iw)|. There is also another situation when it is desirable to switch off the derivative part,
namely for processes with long dead-time. If the operator tells the controller that he has a process with long
dead-time, a PI controller with the following design will replace the PID controller.

This controller will give a much faster control than the PID design presented above.

Gain Scheduling
Gain scheduling is an effective method to treat processes with predictable variations in the dynamics. A gain
schedule is a table with several sets of controller parameters, one set for each operating point. (Parameter
schedule would be a more adequate notation than gain schedule!). A reference signal which is related to the
nonlinearity determines when to switch from one set of controller parameters to another. If e.g. the
nonlinearity is caused by a nonlinear valve, the control signal should form the gain scheduling reference, since
it is directly coupled to the valve position. If the nonlinearity is caused by a nonlinear sensor, the
measurement signal should be used as a gain scheduling reference.
Most process control plants contains several nonlinear control loops. In spite of this, gain scheduling is seldom
used in process control. One reason for this is, that is has been regarded as too time consuming to build this
schedule with several different sets of controller parameters. With the use of auto-tuning, this drawback has
disappeared. Using the auto-tuner once at every operating point will automatically provide the schedule.
When the process dynamics are predictable, it is better to use a gain schedule than an adaptive controller. The
gain schedule will instantaneously provide a suitable set of controller parameters as the operating conditions
change, while the adaptive controller needs a fair amount of time before it has adapted itself to the new
conditions. Using adaptive control, the operator provides the controller with the information that the process
dynamics are varying. Using gain scheduling provides the controller with the additional information about how
the dynamics are varying.
4. Adaptive Feedback Control

From the auto-tuner experiment described in the previous section, we obtained the frequency wand the value
of the Nyquist curve at this frequency. In this section, we will describe how the point G(iw) can be tracked
when the dynamics are changing. We will also describe some of the supervisory logic that is included in the
controller.
Tracking a point on the Nyquist curve

The identification principle is illustrated in Figure 4. The control signal u and the measurement signal y are
filtered through narrow band-pass filters at the frequency w. These two signals are then analyzed in a least-
squares estimator which provides an estimate of the point G(iw). The band-pass filters the two band-pass
filters are of the form
This filter will give a relatively high gain at the frequency w, and suppress the signals at other frequencies. We
have given the transfer function in continuous form. In the practical implementation, we use the sampled
version with fast sampling, i.e. the filters are sampled with the same frequency SI the PID controller.
It is well known from practical use of adaptive controller, that a suitable filtering of the signals must always be
used. See e.g. Wittenmark (1986). Low frequencies must be filtered out.

to avoid interactions from load disturbances. High frequencies must be filtered out to avoid high frequency
noise from disturbing the parameter estimates. Since the process model always is simpler than the process
itself, a filtering of the signals must also be made in order to ensure that the interesting part of the dynamics
are identified. By using a narrow band-pass filter, we thus not only obtain the goal to track G(iwJ) at the
frequency w. We also solve the traditional filtering problem in a very effective way.
The least-squares estimator the two narrow band-pass filters produce two signals which we can approximate
with two sine waves with different amplitude and phase. The quotient Q between the amplitudes and the
phase shift V' between the two
Signals give us G(iw).

We have used a least-squares estimator to obtain G(iw) from the signals. The parameters of the second-order
model

are estimated. The sampling period h is determined from the frequency w. We have found that the choice

i.e. eight samples per period, gives good identifiability properties. From the parameters bl and b2, G(iw) can
be determined according to:
The least-squares algorithm is of the constant-trace type. Since the signals u( t - h) and u( t - 2h) entering the
LS-estimator are approximately sine-waves with a constant phase-shift, it is possible to simplify the algorithm.
Instead of identifying the parameters bl and b2 directly, the following scaling is performed:

the expected value of the covariance matrix P becomes diagonal with equal diagonal elements. Hence it is
possible to reduce the P-matrix to a scalar.
Supervision
The adaptive controller cannot run continuously without any supervision. E.g. logic to avoid identification when
no information
Is available must always be present the way this supervision is performed is at least as important as the
underlying basic algorithm. However, there are no general rules or guidelines describing how this supervision is
to be performed. The different manufactures of adaptive controllers have their own tricks. We will now shortly
describe some fundamental procedures at our supervisory level.
First of all, we must ensure that the adaptation mechanism is only active when we have any information in the
signals. Under periods of good control, when both the control signal and the measurement signal are straight
lines, maybe corrupted with high frequency noise, no identification should be made. We have a procedure that
high-pass filters the control signal and the measurement signal. Adaptation is only allowed when both these
Signals have had a transient recently.
Load disturbances are not covered in our description of the process. Implicitly we assume that changes in the
measurement signal are caused by the control actions. To avoid disinformation from high frequency noise and
load disturbances, we put band-pass filters on our signals, as described in section 4. Since we found that this is
not always enough, we have also included a procedure to detect load disturbances, and thereby avoid
adaptation
During the first part of a load disturbance transient.
There is a very simple relation between the parameters estimated in the least squares estimator and the physical
parameters
Q and V'. Therefore, it has been possible to check if the estimates have reasonable values. We have bounded
the parameters in such a way that V' is always inside a sector in the third quadrant and that Q may not vary
more than a specified factor from the initial value given by the autotuner.

3. Adaptive Feed forward

Feed forward, constant or adaptive, is a powerful method to compensate for disturbances before they have
shown up in the measurement signal. We have included an adaptive feed forward compensator in the
controller. The feed forward compensator has the following structure:

Where Uff is the feedforward component of the control signal, Kff is the feed forward gain, and v is the
disturbance signal. The feed forward signal is updated with the same frequency as the control signal, i.e. fast
compared to the time constants of the process. This simple adaptive feed forward compensator has shown to
be very useful. In most cases, it is sufficient to let the feed forward compensation be formed by just a gain
times the disturbance signal. Sometimes, it is desirable to delay the signal v, as will be discussed below.
The gain Kff is determined from the model

where y is the measurement signal. The parameters a and b are determined by an ordinary least-squares
estimator. The signals are both high- and low-pass filtered to get rid of noise and bias terms.
The choice of the time delay d in the model is crucial. If d is not chosen suitably, the model will not capture the
relations between u and y and between v and y. Let the dead-time plus the dominating time constant of the
process be T uy and the dead-time plus the dominating time constant of the transfer function between v and y
be Tvy' The following cases can then be distinguished:
T vy > > T uy
It is desirable to delay the disturbance signal v. Otherwise, the feed forward compensation will influence the
signal V before the disturbance. The disturbance signal should ideally be delayed with the time Tvy – Tuy.

In this case, feedforward is often very efficient. T vy < < Tuy It is not worth wile to use feed forward. We cannot
make any compensation before the disturbance is seen in y. The feedback controller can equally well do the
job.
From these considerations, we can conclude that d should ideally be chosen as Tuy ' i.e. equal to the dead-time
plus the dominating time constant of the process. From the relay experiment in the autotuner, the maximum
time delay between u and 11 is given as half the oscillation period,i.e. Tu /2. (If the process consists of only a
time delay, the oscillation period is two times the time delay!) We have chosen the parameter d = Tu /2. The
sampling interval of the least squares estimator is chosen as in the feedback case, i.e., h = Tu/8. This gives the
following model equation:

We have chosen the gain Kff in the control law as

Where a and b are the estimates of a and b. The adaptive algorithm is surrounded with a security net in the
same way as the feedback algorithm, e.g., high-pass filtering of the signals tells when the information content
in the signals is large enough to allow adaptation.
6. Industrial Experiences

To illustrate the properties of the adaptive controller, some of the experiments from the field tests are
presented below. We will first shortly describe how the different uses of the adaptive technique are
implemented in the controller.
Implementation
The new industrial adaptive controller, named ECA400, is manufactured by Satt Control instruments AB,
Sweden. It is a single station cascade controller. It contains all the adaptive techniques that have been
presented in the previous sections, i.e. the relay autotuner, a gain schedule, adaptive feedback and adaptive
feedforward. The gain schedule, the adaptive feedback and the adaptive feed forward are independent of
each other, and may thus be used separately or together. The autotuner is used to initialize the adaptive
controller and the adaptive feed forward.
In this way, the operator does not have to provide the controller with any information about the process
dynamics. The parameters of the gain schedule are automatically obtained by using the autotuner once at
every operating point. The gain schedule may be combined with the adaptive controller. Adaptation will then
only be performed with the set of controller parameters that are presently used. The result is the same as if
we had several different and independent adaptive controllers.
Temperature control
The first example shows a temperature control loop. Parts of the Experiments are presented in Figure 5.
Water is heated through a heat exchanger, with steam on the primary side. The water temperature is
measured and the controller output determines the steam valve position. The primary disturbances in the
loop are changes in the steam pressure and changes in the water flow rate. The drift in the control signal
during times of constant temperature are due to the changes in steam pressure. The set point is changed
stepwise up and down to activate the adaptation mechanism, and to show the properties of the control.
Before the large change in water flow, we had a PI controller with gain 3.2 and integral time 13 s. When the
water flow decreased, the gain decreased to 2.6 after the first set point change. After three set point changes
the gain had decreased to 2.1. After the fourth set point change, we obtained PID controller with gain 1.5,
integral time S.2 s and derivative time 2.0
These controller parameters were not significantly changed until the next water flow alteration. The
experiment shows that the adaptation to the new process conditions is rather fast . This is especially the case
when the process gain increases, and the controller gain is to be decreased. The reason is that the high loop
gain provides good excitation.

Pulp density control

The second example is taken from the paper industry. The pulp is diluted to a desired density by adding water
to the incoming pulp. Parts of the experiment are shown in Figure 6. The set point of the pulp density was
changed stepwise to activate the adaptation. The process gain is changing according to changes in the pulp
flow. Figure 6 shows a situation where the pulp flow was increased, resulting in a decreased process gain.
Before the flow change, we had a PID controller with a gain of 0.10. As seen in the figure , this low gain gave a
very slow control after the pulp flow change. After five set-point changes, the controller gain had increased to
0.30, resulting in a much faster control

7. Conclusions

This topic concludes that it is possible to combine the ideas of auto-tuning, gain scheduling, feed forward and
adaptation in a simple controller that is easy to use. The key idea is to use relay feedback to obtain the time
scale of the process. When this is known it is possible to simplify many design issues. The structure and
sampling periods of the discrete time model. Used for adaptive feedback and feedforward can then be
determined.
The controller is based on the pm structure. It thus inherits all the advantages and disadvantages of the pm
structure. An advantage is that the operators have a well-known structure that they can relate to. They know
what the parameters mean. They can compare the parameters with values they know from experience. Since
the operators are familiar with the gain, the integral time and the derivative time of a pm controller, they may
very well turn the adaptation off for certain periods and make their own adjustments. The pm structure doe.
However also have some disadvantages particularly when dealing with processes with long dead time.A major
advantage of the new controller is that it is easy to operate. No parameters have to be given in advance. The
auto-tuner is activated simply by pushing the tuning button. The information required for the adaptation is
derived from
the auto-tuning experiment. This is important also from the point of view of robustness. Since all parameters
are calculated automatically, the behavior of the controller is predictable. It cannot deteriorate due to e.g. a
bad choice of the sampling period.

The controller is based on PID control with fast sampling. This has the advantage that the controller will react
very quickly on a disturbance. Adaptive systems based on general linear models often use the same sampling
period for control and estimation. It is common practice to use a fairly long sampling period to make the
estimator robust. There will then be a delay in responding to disturbances. Only two parameters are
estimated. This is less than in most other adaptive controllers. To be able to describe the process dynamics
using only two parameters, these parameters must be chosen properly. This is possible since the initialization
with the auto tuner makes it possible to choose a good structure. The adaptation rate is rather fast since only
two parameters are estimated.
8. References
 ASTROM, K . J. (1987): "Adaptive Feedback Control." Proc. IEEE, 75, 185-486.

 ASTROM, K . J . and T. HXGGLUND (1988): Automatic Thning of PID Controllers, ISA, Research Triangle Park, NC, USA.

 SEBORG, D. E ., T. F . EDGAR , and S. L . SHAH (1986): "Adaptive Control Strategies for Process Control: A Survey," AIChE Journal, 32, 881-913.
Reference

 Margareta, Stefanovic., R., Wang., Michael, G., Safonov. (2004). Stability and convergence in adaptive

systems.

 Rolf, Johansson., Anders, Robertsson. (1998). Stability analysis of adaptive output feedback systems.