Pattern Analysis of Machine Olfactory System

Abstract

Intelligentsensors require a robust recognition paradigm to discriminate and analyse the target species. Artificial olfactory systems (popularly known as E-Nose) typically suffers from poor selectivity of individual sensor elements due to which the final classification often turns out to be poor. The use of chemical sensors in the form of array helps to improve the selectivity. But the problem gets aggressive when real time classification of odors/gases is required. When different classification techniques are implies in real time applications, it is expected that the incoming test/data patterns gets classified in accordance of the distribution of the training data. But in all cases we do not have the initial training samples to train our network and hence we have to use unsupervised classification techniques. These unsupervised techniques use the distance parameter as a key to the classification problem. But the performance limitation of existing unsupervised algorithm like K-means clustering as it is highly sensitive to euclidean distance also limits the performance of E-nose to great extent. Thus it becomes imperative to find a distance parameter that corresponds to the similarity of incoming data patterns and then classify them. In this thesis we have proposed a novel normalized cosine K means clustering technique and Quaternion based approach for classification of data obtained from an E-Nose sensor.