Philip Picton

Biological data mining with neural networks: implementation and application of a flexible decision tree extraction algorithm to genomic problem domains

In the past, neural networks have been viewed as classification and regression systems whose internal representations were extremely difficult to interpret. It is now becoming apparent that algorithms can be designed which extract understandable representations from trained neural networks, enabling them to be used for data mining, i.e. the discovery and explanation of previously unknown relationships present in data. This paper reviews existing algorithms for extracting comprehensible representations from neural networks and describes research to generalize and extend the capabilities of one of these algorithms. The algorithm has been generalized for application to bioinformatics datasets, including the prediction of splice site junctions in Human DNA sequences. Results generated on this datasets are compared with those generated by a conventional data mining technique (C5) and conclusions drawn.

Automated control of an actively compensated Langmuir probe system using simulated annealing

A simulated annealing (SA) method has been developed to deduce 14 Fourier terms in a radio frequency waveform for active compensation of a Langmuir probe system. The active compensation system uses seven harmonics to generate a required waveform. Therefore, 14 heavily interacting continuous parameters need to be tuned before measurements can be taken. Because of the magnitude of the resulting search space, it is virtually impossible to test all possible solutions within an acceptable time. An automated control system employing SA has been developed for online tuning of the waveform. This control system has been shown to find better solutions in less time than skilled human operators. The results are also more reproducible and hence more reliable.

Multistage Neural Network Ensembles

Neural network ensembles (some times referred to as committees or classifier ensembles) are effective techniques to improve the generalization of a neural network system. Combining a set of neural network classifiers whose error distributions are diverse can lead to generating more accurate results than any single network. Combination strategies commonly used in ensembles include simple averaging, weighted averaging, majority voting and ranking. However, each method has its limitations, dependent either on the application areas it is suited to, or due to its effectiveness. This paper proposes a new ensembles combination scheme called multistage neural network ensembles. Experimental investigations based on multistage neural network ensembles are presented, and the benefit of using this approach as an additional combination method in ensembles is demonstrated.

Optimization of plasma etch processes using evolutionary search methods with in situ diagnostics

This paper presents several approaches that have been used to control, optimize and characterize a low pressure (10–300 mTorr) plasma processing system. Methods such as contour following and differential evolution have been used to find contours of DC bias, total ion flux, ion energy flux, quadrupole mass spectrum (QMS) intensity ratios and line intensity ratios of the optical emission spectrum (OES) in argon and nitrogen plasmas. A mapping for a 4 × 4 multi-dimensional parameter space is also presented, in which the relationship between four control parameters (power, pressure, mass flow rates of two supplied gases) and four measurement outputs (DC bias, ion flux, QMS ratios and OES line intensity ratios) is determined in a plasma etching process. The use of these methods significantly reduces the time needed to re-configure the processing system and will benefit transfer of processes between different systems. A similar approach has also been used to find quickly an optimum condition for directional etching of a silicon wafer.

Extraction of short-latency evoked potentials using a combination of wavelets and evolutionary algorithms.

Somatosensory evoked potentials, recorded at the spine or scalp of a patient, are contaminated by noise. It is common practice to use ensemble averaging to remove the noise, which usually requires a large number of responses to produce one averaged signal. In this paper a post-processing technique is shown which uses a combination of wavelets and evolutionary algorithms to produce a representative waveform with fewer responses. The most suitable wavelets and a set of weights are selected by an evolutionary algorithm to form a filter bank, which enhances the extraction of evoked potentials from noisy recordings.

Review of Artificial Neural Networks (ANN) applied to corrosion monitoring

The assessment of corrosion within an engineering system often forms an important aspect of condition monitoring but it is a parameter that is inherently difficult to measure and predict. The electrochemical nature of the corrosion process allows precise measurements to be made. Advances in instruments, techniques and software have resulted in devices that can gather data and perform various analysis routines that provide parameters to identify corrosion type and corrosion rate. Although corrosion rates are important they are only useful where general or uniform corrosion dominates. However, pitting, inter-granular corrosion and environmentally assisted cracking (stress corrosion) are examples of corrosion mechanisms that can be dangerous and virtually invisible to the naked eye. Electrochemical noise (EN) monitoring is a very useful technique for detecting these types of corrosion and it is the only non-invasive electrochemical corrosion monitoring technique commonly available. Modern instrumentation is extremely sensitive to changes in the system and new experimental configurations for gathering EN data have been proven. In this paper the identification of localised corrosion by different data analysis routines has been reviewed. In particular the application of Artificial Neural Network (ANN) analysis to corrosion data is of key interest. In most instances data needs to be used with conventional theory to obtain meaningful information and relies on expert interpretation. Recently work has been carried out using artificial neural networks to investigate various types of corrosion data in attempts to predict corrosion behaviour with some success. This work aims to extend this earlier work to identify reliable electrochemical indicators of localised corrosion onset and propagation stages.

Improved simulated annealing with step width adaptation for Langmuir probe tuning

In a previous investigation, a simulated annealing (SA) method was developed to optimize 14 Fourier terms in a radio-frequency waveform for active compensation of a Langmuir probe system. This approach was shown to find better solutions in less time than skilled human operators. However, variations in fitness indicated that the SA algorithm did not always find the precise global solution, although it came consistently close to it. This variability was caused by the limited number of fitness evaluations available due to time constraints. In this research, the chosen maximum step width has been shown to have a significant effect on the overall performance of the algorithm. A scaling function has been developed to adapt the maximum step width of the SA algorithm, on-line, as a function of the number of elapsed iterations. The modified algorithm has been shown to find fitter solutions with reduced variability in fitness.

Development of an image analysis technique for measurement of Poisson’s ratio for viscoelastic materials: application to leather

The application of a computerised image analysis system to simultaneously measure longitudinal and axial strain in naturally available anisotropic-viscoelastic material like leather is described. The results are used to calculate Poisson’s ratio and indicate that this parameter varies with position and orientation on the hide and depends on the degree of pre-existing fibre orientation. The origins of this dependency can be explained by a simple microstructural model. The practical implication of the findings for detecting lines of tightness (maximum fibre’s orientation) is discussed.

The in-service inspection of coated steel welds using Eddy-Current Techniques

Traditionally surface crack detection in coated Ferritic Steel Welds with Eddy-Current Techniques has been difficult due to the change in material properties in the Heat Affected Zone. These typically produce signals larger than crack signals. Sophisticated probe design and construction, combined with modern electronic equipment, have largely overcome the traditional problems and now enable the advantages of Eddy-Current Techniques to be applied to In-Service Inspection of Coated Ferritic Steel Structures in the as-we!ded conditions. Specifically, the advantage of the technique is that under quantifiable conditions an inspection may now be carried out through corrosion protection systems. It is the intention of this paper to review the current information available, establish the limiting parameters of the technique and detail the practical experiments conducted to determine the extent of the limiting parameters. The results of these experiments are detailed. Having determined the limiting factors, outline testing procedures have been established together with relative sensitivity settings.

The coupled nonlinear dynamics of a lift system

Coupled lateral and longitudinal vibrations of suspension and compensating ropes in a high-rise lift system are often induced by the building motions due to wind or seismic excitations. When the frequencies of the building become near the natural frequencies of the ropes, large resonance motions of the system may result. This leads to adverse coupled dynamic phenomena involving nonplanar motions of the ropes, impact loads between the ropes and the shaft walls, as well as vertical vibrations of the car, counterweight and compensating sheave. Such an adverse dynamic behaviour of the system endangers the safety of the installation. This paper presents two mathematical models describing the nonlinear responses of a suspension/ compensating rope system coupled with the elevator car / compensating sheave motions. The models accommodate the nonlinear couplings between the lateral and longitudinal modes, with and without longitudinal inertia of the ropes. The partial differential nonlinear equations of motion are derived using Hamilton Principle. Then, the Galerkin method is used to discretise the equations of motion and to develop a nonlinear ordinary differential equation model. Approximate numerical solutions are determined and the behaviour of the system is analysed.