Alarm Pattern Analysis Using Computational Intelligence Approach

In today’s high value manufacturing, machines are often equipped with sensors to detect disturbances in
the system, which may trigger the corresponding alarm(s) when limits are exceeded. Apart from displaying on a
terminal for human intervention, these alarm values are also logged to a database. Machine faults or breakdown
may be characterized by a set of alarm patterns. The ability to detect these patterns early can help to alert and
prevent impending machine failure, which is extremely useful for mission critical machines. This project involves
the development of pattern identification concepts and algorithms based on computational intelligence approaches
to predict machine failures. Some of these approaches include Ant Colony System, Genetic Algorithm, and
Statistical methods.
Students with keen interest to learn computational intelligence for pattern recognition and those with knowledge
in manufacture equipment management will have an added advantage for this project

Investigation of the Sequential Accelerator on the Perceptron for Pattern Recognition

In machine learning methods, when the input data becomes extremely large, the current direct methods
require too large learning times and memory. This project investigates one sequential method to overcome this
problem. It is quite a simple method to implement and is tested using the Perceptron as the base classifier. The
perceptron converges very slowly. So it will be interesting to find out if the proposed accelerator can improve
significantly the computational times of this simple classifier. The student will try various investigations on
different very large data sets and to measure their computational complexities. It has been shown that this
sequential method is very fast and only need a small subset of the large data set to complete the learning.

Investigation of the Sequential Accelerator on LDA for Pattern Recognition

In machine learning methods, when the input data becomes extremely large, the current direct methods
require too large learning times and memory. This project investigates one sequential method to overcome this
problem. It is quite a simple method to implement and is tested using the well-known LDA classifier. The student
will try various investigations on different very large data sets and to measure their computational complexities. It
has been shown that this sequential method is very fast and only need a small subset of the large data set to
complete the learning.

Equalization of Fading Channels using RBF Networks

Equalization of Fading Channels using RBF Networks
Summary: Equalization is the process of recovering the true input data from the received data which is corrupted
by noise and passes through a nonlinear communication channel. Earlier work by the supervisor and his group has
successfully used RBF neural networks for complicated nonlinear channels that are stationary. In this project, the
performance of the RBF equalizers will be investigated for time varying ( fading ) channels ( mainly for Raleigh
Fading Channels) and an implementation scheme for the equalizer will be evolved. Performance comparison with
other conventional equalizers will also be made.

Analysis of the quality of fingerprint image features

The most popular fingerprint feature used for matching is the feature points, called minutiae. Several approaches have been proposed which in a way, involves detecting the skeleton image to locate the minutiae. Subsequently, the geometrical and other information are extracted and used to match the minutiae. However, the performance of the minutiae extraction varies. Problem arises when there is noise or the fingerprint image is not clear. This could cause the introduction of false minutiae or the omission of valid minutiae. In this project, the main aim is to devise methods to analyse the quality of the minutiae so that the overall matching is improved. This
will involve analysing the property of the minutiae to provide suitable confidence level and providing consistent information that can be extracted from the minutiae. Good knowledge in C/C++ or Matlab programming is essential for this project.

Neural Network based Color Image Segmentation

Breast cancer is the most common cancer among women, and is the second leading cause of cancer
deaths in women today. Basically, the diagnosis procedure of breast cancer consists of two steps; (1) mammography
based breast abnormality detection; (2) biopsy based diagnosis. Biopsy is the only definitive way to determine
whether cancer is present.
In this project, biopsy color Image segmentation based on neural networks will be studied. The objective of breast
biopsy image segmentation is to segment cells and blood vessels in the image.
As the second part of the project, color image segmentation software will be developed.

Development of a Matlab-based Control System Laboratory

The aim of this project is to develop a Computer Aided Control System Design (CASCD) environment that
are typically utilised in an undergraduate controls laboratory. Due to their popularity and availability, MATLAB,
SIMULINK and the Real-Time Workshop toolbox are chosen as the prototyping environment. The following issues
should be addressed:

1. Standardisation: a consistent hardware interface to various laboratory apparatus and a consistent user
interface, for the following task: modelling, control design, data collection, parameter estimation, and real-time
experiment.

2. Control experiment via Internet: While the local real-time control could be extended for remote control via the
Internet, there will be some issues that are peculiar to Internet lab, e.g. all Internet experiments need to be selfresetting.
In addition, safety, security and user flexibilty are issues that need to be addressed 

Wavelet-based Deconvolution for System Identification

In many practical applications, we are given access to the input and ouput of a system whose
characteristics are unknown. The term ‘deconvolution’ is used to describe the operation of separating the input
function from the characteristics of the system we intend to identify. Conventional techniques for deconvolution
based on direct Fourier or Laplace transfoms suffer from the ill-conditioned nature of the problem in the presence
of noise. The aim of this project is to develop and implement a robust deconvolution technique based on wavelets
and study its performance as compared with other conventional techniques. An appreciation of Matlab and
numerical computation skills would be desirable.

A Highly Efficient DC Lamp Dimmer

The simplest lamp dimmer circuit consists of a rheostat, in series with the lamp, which one
may adjust to obtain the required brightness. Such linear regulators are quite inefficient since
a lot of power is wasted in them. Moreover, in the rheostat the moving contacts are likely to
get damaged in the long run, as its value is frequently adjusted by moving the slider. Such
linear control circuits provide an overall efficiency of no more than 50 per cent. This wastage
of power can be avoided if one uses pulse width modulation (PWM) which can be made to
control an electronic rheostat. The circuit shown here is based on PWM principle. Gate N1
and its associated components constitute an oscillator producing oscillations of approximately
200 Hz with a pulse width of 0.1 ms. This output is fed to transistor T1 for level shifting. At the
output of this transistor is a potentiometer VR2, using which a DC component can be added
to the pulses emerging from transistor T1. By adjusting this potentiometer/trimmer, one can
have a good linear control of the lamp brightness from completely off state to 100 per cent on
state. The signal is inverted by gate N2 and fed to MOSFET 12N10. IC CD40106 provides six
inverting buffers with Schmitt trigger action. The buffers are capable of transforming slowly
changing input signals into sharply defined jitter-free output signals. They are usually used as
wave and pulse shapers. IC CD40106 possesses high immunity and low power consumption
of standard CMOS ICs along with the ability to drive 10 LS-TTL loads. In this circuit loads up
to 24W can be connected between MOSFET drain and 12V supply without using a heatsink.
The loads can even be DC motors, miniature heating elements, etc. If one uses a low RDS
(on) MOSFET, a higher efficiency can be achieved. By using the components as shown in the
circuit, an efficiency of approximately 95 per cent can be achieved. The flexibility of the design
makes it possible to change the MOSFET with a similar one, in case of non-availability of
12N10. The circuit by itself does not draw much current when the load is disconnected.
Ensure proper ESD protection while handling the MOSFET to prevent damage. Lab note: The
circuit was tested using MOSFET IRF640 with RDS (on)=0.18 ohm.