Bacteria classification with an electronic nose employing artificial neural networks
This PhD thesis describes research for a medical application of electronic nose technology. There is a need at present for early detection of bacterial infection in order to improve treatment. At present, the clinical methods used to detect and classify bacteria types (usually using samples of infected matter taken from patients) can take up to two or three days. Many experienced medical staff, who treat bacterial infections, are able to recognise some types of bacteria from their odours. Identification of pathogens (i.e. bacteria responsible for disease) from their odours using an electronic nose could provide a rapid measurement and therefore early treatment. This research project used existing sensor technology in the form of an electronic nose in conjunction with data pre-processing and classification methods to classify up to four bacteria types from their odours. Research was performed mostly in the area of signal conditioning, data pre-processing and classification. A major area of interest was the use of artificial neural networks classifiers. There were three main objectives. First, to classify successfully a small range of bacteria types. Second, to identify issues relating to bacteria odour that affect the ability of an artificially intelligent system to classify bacteria from odour alone. And third, to establish optimal signal conditioning, data pre-processing and classification methods. The Electronic Nose consisted of a gas sensor array with temperature and humidity sensors, signal conditioning circuits, and gas flow apparatus. The bacteria odour was analysed using an automated sampling system, which used computer software to direct gas flow through one of several vessels (which were used to contain the odour samples), into the Electronic Nose. The electrical resistance of the odour sensors were monitored and output as electronic signals to a computer. The purpose of the automated sampling system was to improve repeatability and reduce human error. Further improvement of the Electronic Nose were implemented as a temperature control system which controlled the ambient gas temperature, and a new gas sensor chamber which incorporated improved gas flow. The odour data were collected and stored as numerical values within data files in the computer system. Once the data were stored in a non-volatile manner various classification experiments were performed. Comparisons were made and conclusions were drawn from the performance of various data pre-processing and classification methods. Classification methods employed included artificial neural networks, discriminant function analysis and multi-variate linear regression. For classifying one from four types, the best accuracy achieved was 92.78%. This was achieved using a growth phase compensated multiple layer perceptron. For identifying a single bacteria type from a mixture of two different types, the best accuracy was 96.30%. This was achieved using a standard multiple layer perceptron. Classification of bacteria odours is a typical `real world' application of the kind that electronic noses will have to be applied to if this technology is to be successful. The methods and principles researched here are one step towards the goal of introducing artificially intelligent sensor systems into everyday use. The results are promising and showed that it is feasible to used Electronic Nose technology in this application and that with further development useful products could be developed. The conclusion from this thesis is that an electronic nose can detect and classify different types of bacteria.