Sunteți pe pagina 1din 5

VALVE DIAGNOSTICS AND AUTOMATIC TUNING

Anders Wallen
Department of Automatic Control, Lund Institute of Technology
Box 118, S-221 00 Lund, Sweden, Phone: + 46 46 2228793
andersw@control.lth.se

Abstract work as an autonomous unit, i.e. it should be able to


install, operate, supervise and revise itself to as high
An integrated system for valve diagnostics and auto-
extent as possible. In a real industrial control system
matic PID tuning is presented. The purpose is to de-
there must of course still be a supervisory level for con-
tect non-linearities such as friction and hysteresis since
trol system (re)configuration, plant start-up and shut-
these may drastically decrease control performance.
down, planning and global fault handling and diagnosis.
Grafcet is used for representing the controller logic.

2. Control Strategy
1. Introduction
The setup used here consists of a standard PID con-
Lately, much attention has been paid to the problem of
troller having a valve as actuator. The controlled vari-
fault detection and isolation (FDI) in process industry.
able could for example be a flow, level or concentration.
For the FDI method to be successful, a lot of effort
It is assumed throughout the paper that the controlled
has to be put into modeling the behavior of the plant,
process is open loop stable with high relative damping
both in normal operating conditions and when a fault
and that there is no integration from the control signal
is present. Different methods require different types
to the output variable. The first assumption is funda-
of modeling, e.g. neural network training, (linearized)
mental for the methods described here, and fortunately
differential equations or di-graphs.
it is fulfilled for most processes found in process indus-
Since a fault occurring in a plant may affect many try. The second assumption could be relaxed, but the
of the measured signals, a global, plant-wide view methods have to be modified accordingly.
is often needed to successfully detect and isolate the
The strategy used for controller installation is now
fault. There are however certain faults which often
described. The user is asked to manually go to the
can be detected locally in the control loop. A very
desired operating point and await stationarity. It is
frequent fault in process industry is sticking valves.
also possible at this stage to give some information to
Since valves are often used as actuators in feedback
the controller, e.g. allowed ranges for input and output
loops having integral action, this non-linearity may
during experiments, crude guesses of the static gain,
induce oscillations in the flow through the valve, see
apparent dead time and apparent time constant of the
for example [5].
process, at least by orders of magnitude. When this has
The problem of non-linearities in valves is especially been done, this automated procedure follows:
crucial when combined with automatic controller tun-
ing. The standard tuning procedures often conceal valve 1. Estimate the measurement noise, i.e. the varia-
problems, but the on-line operation may suffer dramat- tion of the output during a number of samples.
ically from them. Automatic tuning features are ac- 2. Do a series of small open-loop step responses to
cepted and widely spread in todays industrial control get crude estimates of the static gain and time
systems. This paper suggests that these should be ex- scales, and to detect possible hysteresis.
tended with automated tools for detecting severe prob-
lems in the process, both before tuning and in on-line 3. If the hysteresis is not significant, ramp the
operation. control signal slowly to detect possible stiction in
the valve.
The system described in this paper includes sys-
tem guided start-up, tuning, monitoring and trouble- 4. If no stiction is present, tune the controller using
shooting. The ultimate goal is to make the controller the following procedure:
This work has been supported by the Swedish National Board for i. Estimate the Nyquist point using relay feed-
Industrial and Technical Development (NUTEK). back.
ii. Close the loop using a temporary PID tuning. Plant Level Supervisory
Control System
iii. Apply a set point step to estimate the static
gain. Local Control Loop
iv. Calculate new parameters according to the Supervisory Level
method [2].
Real-Time
5. Put the controller in automatic mode. Controller Level
Controlled Process
6. Start monitoring the output in order to detect
degraded control performance, in this case only Figure 1 Hierarchies in the control system.
oscillations using the method presented in [5].

3. Controller Software Structure


For further explanation of how the experiments are
carried out, refer to the example in Section 4. The hierarchical organization of the control system
is shown in Figure 1. The top-level hierarchy deals
The installation procedure controller is now complete, with the plant level planning and supervision and is
and the controller will be operating on-line until the not further discussed here. The second layer contains
performance monitor detects a persistent oscillation. the brain of the local controller, i.e. where the on-
line operation is planned. For example, the strategy
The system assumes the oscillations have one of three in Section 2 is described on this level. The actual
different causes: control algorithms run in the real-time layer which
communicates with the controlled process. The different
External disturbances, for example due to a stuck layers may execute in different computers, but need not
valve in another control loop in the plant, do so. The user typically interacts with the two upper
layers of the system.
Limit cycles due to non-linearities such as stiction The planning and execution of the on-line operation typ-
or hysteresis in the feedback loop, or ically is sequential in its nature. Grafcet, or Sequential
Function Charts, is a Petri Net-based standard indus-
Badly tuned controller.
trial tool for doing sequential control [3]. Here, a mi-
nor modification of the Grafcet standard is used which
In order to rule out the first cause, the controller is substantially increases the descriptive power of the se-
put in manual mode for a time corresponding to a quential chart [1]. Among other things, the extended
few oscillation periods. If the oscillation persists, it is Grafcet allows actions and expressions to be described
assumed to have external causes, and the controller is in a high-level computer language. It also offers a nice
again put into automatic mode. The oscillation must way to parameterize what sequence to use in different
then disappear before there can be another interrupt situations by using procedure steps and Grafcet proce-
from the oscillation detector. dures. [6] shows how extended Grafcet can be used to
structure control algorithms.
If the oscillation disappears in manual mode, the sys-
tem concludes it is either caused by friction or bad tun-
ing. To test for friction, step 3 above is repeated. If no 4. A Session With the System
friction can be detected, the controller is re-tuned ac-
This section shows a simulation of a typical installation,
cording to step 4.
operation and supervision of a controller. The process
The strategy presented in this section will of course transfer function in the example is
not cover all possible situations that may occur in
esL
practice. Some of the experiments will for example fail G (s) = K p (1)
or give faulty results if a major load disturbance occurs. (1 + sT )3
However, the suggested strategy resembles what many
with K p = 0.9, L = 10 and T = 10. White noise with
experienced operators and process engineers do by
unit variance fed through a first-order low-pass filter
hand. By automating the procedure the operators task
has been added to the measurements.
is significantly simplified. Still, the experiments should
in most cases be supervised and acknowledged by an Upon startup, the user is asked to bring the process to
operator in order to abort and redo experiments with the desired operating point and tell the system how far
suspicious or undesired results. from this point the experiments may go. The user is able
Experiment Setup experiments. The first step downwards is performed in
Process variable User supplied order to find suitable step magnitudes. The step size is
estimates initially based on the information supplied by the user,
40.0
Time Delay: 15
but is changed if the output moves too little or too far.
30.0
Time Constant: 40
The second step ensures that the possible backlash in
20.0
Static Gain: 1
the valve is closed when the third step is performed.
Thus, the third step will give a fairly reliable estimate
0:36 0:38 0:40 0:42 0:44 Allowed deviations
during experiments
of the static gain of the process. When the fourth
Control signal step is applied, the output should hopefully go back to
Process variable 10.0
40.0
the same level as after the second step. However, the
30.0 Control signal 10.0
actuator must first close the backlash, and therefore
20.0
the output may exhibit hysteresis. In Figure 3 the
Manual control signal
10.0
estimated static gain is 0.88 and the estimated backlash
39
0:36 0:38 0:40 0:42 0:44
0.0
is 2.15, as compared to 2.0 in the simulated valve. This
is significant compared to the range of operation, and
OK Print . . .
the operator is advised to remove the faulty valve.
When the experiments have been repeated for a valve
Figure 2 Window for the experiment setup. without hysteresis, the ramp response both upwards
and downwards is checked. It is useful to plot the
process variable versus the control signal, see Figure 4.
to supply approximate apparent dead time, apparent Since the output follows the input smoothly, it is
time constant and static gain of the process as well. concluded that no friction is present. The difference in
After the user has confirmed that stationarity has been the output between the ramps up and down is due to
reached, the amplitude of the measurement noise is the fact that the process is not in stationarity while
estimated. This is used for setting the relay hysteresis ramping. The slope of the process input is designed
and the tolerance when checking stationarity in step to give a deviation of approximately 10% of the sweep
response experiments. range. A smaller value would require smaller slope and
thus longer duration for the experiment.
Figure 3 shows a typical sequence of open-loop step
responses. Estimates of the static gain and the amount When this preliminary loop assessment has been per-
of backlash in the valve can easily be extracted from the formed, the controller is tuned using the method.
The method is based on the knowledge of the two points
on the Nyquist curve with 0 and 180 degrees phase
shift, respectively. The Nyquist point is detected using
Process variable

55.0
Process variable vs. control signal
50.0
58.0
45.0
56.0
0:48 0:51 0:54 0:57 1:00
Decreasing u
Control signal 54.0
50.0

52.0 Increasing u
40.0
50.0

30.0 48.0
0:48 0:51 0:54 0:57 1:00 38.0 40.0 42.0 44.0 46.0 48.0

Figure 3 Step responses for an actuator with backlash. Figure 4 Data from ramp responses with no friction.
111
000
Process variable

00001
ysp y
111
u
Process
52.5

PID
50.0

1:14 1:16 1:18 1:20 1:22 1:24


Figure 5 Block diagram in tuning phase.
Control signal

a relay induced oscillation, see block diagram in Fig-


ure 5. Normally you have to use a relay with hysteresis 45.0
to avoid chattering due to noise in the output signal.
Such a relay will not detect the true Nyquist point, 40.0
but rather a point with considerably less than 180 de-
grees phase shift, see Figure 6. Since the method
35.0
is designed for the case when the true Nyquist point 1:14 1:16 1:18 1:20 1:22 1:24
is known, another type of relay has been used here. It
uses the zero crossing level for switching, but requires
the output to go outside the hysteresis band before it
may switch again. Figure 7 Controller tuning procedure.

A sample tuning session is shown in Figure 7. The first


two minutes illustrate the relay feedback experiment. half-periods are detected in sufficiently short time, the
The relay amplitude is modified after a few samples in monitor signals that there is an oscillation.
order to get a reasonable size of the oscillation com-
pared to both the noise magnitude and the allowed de- When input stiction is added to the simulation, the
viations. After the relay experiment, the obtained ul- integrating controller will cause the system to oscillate.
timate gain and period is used to calculate a tempo- A typical valve induced oscillation is shown in Figure 8.
rary PID setting based on Ziegler-Nichols frequency re-
sponse method. This is used during a closed loop set
point step change in order to obtain an estimate of the
static gain of the process. The last part of Figure 7
shows this. With this knowledge, a final PID tuning Process variable
is calculated, and the system is completely installed.
54.0
In order to monitor the control performance on-line,
an oscillation detector is also included in the system.
53.0
The detector calculates the integrated absolute error,
IAE, between two consecutive zero crossings. If the
52.0
IAE is above a certain limit, the monitor consider it
as one half-period of an oscillation. If sufficiently many 2:10 2:12 2:14 2:16 2:18 2:20

Control signal
y
45.0

44.0

43.0

42.0

41.0

2:10 2:12 2:14 2:16 2:18 2:20

Figure 6 The behavior of a relay with hysteresis . Figure 8 Valve induced oscillation.
Process variable vs. control signal
A pply first step
Outside allowed
operating region
Decreasing u A nalyze step
54

Reset

52 Change Step too


direction small

Increasing u
50
A pply second step

48
38.0 40.0 42.0 44.0 46.0
Figure 10 Part of the Grafcet sequence for the step tests.

Figure 9 Data from ramp responses with sticky valve. the control algorithms themselves in order to develop,
debug and maintain the system in an efficient way.

The process is taken off-line, which will cause the


oscillation to disappear, and thus it is concluded that 6. Conclusions
either there is some non-linearity in the process, or the The paper shows the importance of adding diagnostic
controller is not well tuned, for example due to changed features to control systems with automatic tuning. The
process characteristics. problem of friction and backlash in valves is addressed
In this case, a repeated ramp test will find the cause to by adding features for loop assessment before tuning
be a sticky valve, which can be easily concluded from and performance assessment during on-line operation.
the plot in Figure 9. Extended Grafcet is used to structure the execution of
the different experiments.

5. Implementation Issues
7. References
The system has been implemented in G2 [4] using the
Grafchart toolbox [1]. The toolbox has been used to im- [1] K.-E. RZEN. Grafcet for intelligent supervisory
plement the sequential logic described in Section 2. Fig- control applications. Automatica, 30:10, pp. 1513
ure 10 shows a small portion of the logic governing the 1526, 1994.
execution of the step tests along with textual descrip- [2] K. J. STRM AND T. HGGLUND. PID Controllers:
tions of what is done in the steps. The two main building Theory, Design, and Tuning. Instrument Society
blocks are steps and transitions. The steps, drawn as of America, Research Triangle Park, NC, second
squares in Figure 10, represent the states of the sys- edition, 1995.
tem and contain actions to be performed while the step
[3] R. DAVID AND H. ALLA. Petri Nets and Grafcet: Tools
is active. Each transition, drawn as small horizontal
bars in Figure 10, contains a condition for activating for modelling discrete events systems. Prentice-
the step after the transition and deactivating the step Hall, 1992.
before. The steps with filled corners are special pro- [4] GENSYM. G2 Reference Manual, Version 4.0. Gen-
cedure steps, enabling hierarchical and parameterized sym Corporation, 125 CambridgePark Drive, Cam-
sequences. The small vertical bar is an exception tran- bridge, MA 02140, USA, 1995.
sition used for fault handling. It may deactivate the [5] T. HGGLUND. A control-loop performance monitor.
procedure step anytime, whereas ordinary transitions Control Engineering Practice, 3, pp. 15431551,
wait for the procedure to complete its execution. 1995.
In order to take care of all possible behavior that may [6] A. WALLEN. Using Grafcet to structure control
occur during the experiments, the logical safety net algorithms. In Proceedings of The Third European
has to be extensive. It is the authors firm conviction Control Conference, Rome, Italy, 1995.
that the logic must be represented seperately from

S-ar putea să vă placă și