This thesis investigates different signal processing and pattern recognition methods employed in the classification of smells obtained from electronic nose (Enose). While most of them perform well for simple classification tasks and large training sets, the performance degrades significantly for more complex classification tasks or under smaller training sets. For smell classification in clinical experiments, the inconvenience, high-cost and long time duration of the sampling process practically eliminates the possibility of obtaining large training sets, also the difference between the clinical smell samples is subtle, consequently the improvement of classification accuracy under such circumstances will have considerable practical importance in Enose application.; As a compromise solution to this technical challenge, a Multi-Dimension Combination (MDC) method is proposed which incorporates the classification output from each individual classifier (either feature extraction or dimension reduction method). It is shown that the MDC outperforms each individual classifier and increases the classification accuracy. |