Finite sample properties of system identification for control.
The asymptotic convergences properties of system identification methods
are well known, but comparatively little is known about their behaviour
for the practical case when only a finite number of data points are available
for system identification. One open problem is the following: Assume the
true system is a linear system of known order but with unknown parameters.
The system parameters are estimated using least squares identification,
and a controller, e.g. a pole placement controller, is designed based on
the identified model. What is the probability that the resulting closed
loop is stable? How many data points do we need for identification in order
to guarantee with high probability, say 0.9995, that the closed loop system
is stable? Given that we know that the model has been obtained by a system
identification experiment, can we design the controller in such a way that
the closed loop system is stabilised with the highest possible probability?
System identification for pole placement control
System identification and control design are an iterative process,
and several schemes for iterative identification and control have been
developed. The basic steps in such schemes are: First a model is
identified using open loop data, and a controller is designed. The
controller is implemented, and the system is operated in closed
loop. During closed loop operation more data is collected and the
system is re-identified using closed loop data. The controller is then
redesigned based on the new model. The controller is implemented, and
the whole process is repeated until the controller performance is
satisfactory. In this project we will focus on pole placement control,
and question we will address is: How should the identification
criterion be chosen? How should the input signal (in open loop) be
chosen? Should the data be filtered in a particular way before we use
them for system identification? What is the best way of combining data
collected in open and closed loop? Does the parameter estimate
converge as the number of identification/control iterations tends to
infinty?
Modelling and control
of open water channels.
A number of projects are available in this area. Some of them are listed
below.
Control of large irrigation networks.
Irrigation networks can be large with hundreds of kilometers of
channels and a large number of gates for regulation of water levels
and flows. A typical approach to control design is to start with a
small portion of the channel and design a controller for this portion,
and then extend the design to incorporate more and more gates and
water levels. However, there is no guarantee that a control strategy
which works well for a small part of the channel will also work well
for the total channel networks. In this project we will investigate
which control strategies scales well and if particular conditions must
be satisfied for the controllers to scale well.
Controller configurations for decentralised control of irrigation
channels.
A feature of a decentralised control strategy is that one gate
controls one water level. Advantages of such a control strategy is
that the design of the controllers is relatively easy and the
communication requirements, i.e. the amount of data that has to be
sent over a radio network, are small. A drawback is that the
performance will not be as good as for a centralised multivariable
controller. With a decentralised controller there is still a choice of
configuration, e.g. which water level should a gate control. Two
common strategies are upstream control, where the gate controls the
immediate upstream level, or distant downstream control where it
controls the water levels immediately upstream of the next downstream
gate. Another strategy is very distant downstream control where a gate
controls the water level upstream of a gate far downstream, and the
intermediate gates are in upstream control mode. In this project we
will study different controller configuration for decentralised
control, and assess there performance and compare with multivariable
controllers. Another research problem to be adressed is the
development of guidelines for decentralised controller configurations
taking into account the geometry of the channel and the control
objectives
Multivariable control of irrigation channels
In this project we will study design of multivariable controllers for
irrigation channels. Multivariable controllers differs from
decentralised controllers where one gate is responsible for the
control one water level, in the sense that all gates are contributing
to the regulation of all water levels. Generally we expect to get
better performance out of a multivariable controller, but they are
also more difficult to design. As data from irrigation channels has
to be sent over a radio network, there are communication constraints,
and design of multivariable control systems where the signals are
sampled at different rates becomes important. Another issue which
must be addressed in this project is controller reduction, that is
based upon the designed multivariable controller, find a controller
which is simpler and easier to tune and implement, but with nearly the
same performance as the full multivariable controller.
Initial control design using physical data only.
Relevant operational data are quite often missing when an automatic
control system is installed for the first time on an irrigation
channel. The initial choice of controller parameters must therefore be
based on geometric data, such as length, width, slope and roughness of
the irrigation channel. In this project we will investigate how to
tune controllers based on such physical information, and how to
improve the controllers when operational data such as water levels and
gate positions, become available.
Performance monitoring and fault detection in irrigation channels.
Due to the sheer size of irrigation networks and the number of gates
and sensors involved, faults are bound to occur. The most typical
faults are sensor failure (water levels and gate positions) and
actuator failure (motors driving the gates). Some failures are easy to
detect, but other such as a sensor slowly drifting off can be
difficult to detect. In other cases it can be easy to detect that
something is wrong, but not what has gone wrong, e.g. has the gate got
stuck or is it the gate position sensor that is faulty? In this
project we will develop a fault detection and performance monitoring
system. Such a system will monitor the data from the irrigation
channel, extract relevant information and make comparison with the
expected behaviour of the irrigation channel. If the difference
between what is observed and what is measure is large an alarm is
raised.
More projects are listed on CSSIP's webpage.
Interested in environmental problems, modelling, control and signal
processing?
More projects are available, both of a theoretical and
practical nature, contact e.weyer@ee.unimelb.edu.au
Back to home page