Examination Committee

Prof Ricky S W LEE, MAE/HKUST (Chairperson)
Prof Amine BERMAK, ECE/HKUST (Thesis Supervisor)
Prof Kea Tiong TANG, Department of Electrical Engineering, National Tsing Hua University (External Examiner)
Prof Chi Ying TSUI, ECE/HKUST
Prof Levent YOBAS, ECE/HKUST
Prof Chih-Chen CHANG, CIVL/HKUST

Abstract

Artificial olfactory systems, referred to as electronic nose systems, have been developed to mimic the functionality of the mammalian olfactory system, in which odorant receptors play a key role to transform odor molecules into electrical spikes, through olfactory transduction. Odor information is encoded in these spike patterns, which are processed by the brain to identify and quantify tens of thousands of odors. Research efforts to obtain a similar performance between electronic olfaction and its biological counterpart, have been focused on two fronts. The first deals with the fabrication of miniaturized sensor arrays to replicate the functionality of odorant receptors, while the second targets the development of algorithms with a potentially equivalent level of odor identification performance to that of the brain.

Miniaturized sensor arrays are now feasible due to the great advancement in fabrication and characterization techniques of sensing materials in recent decade. But, state-of-the-art pattern recognition algorithms, on the other hand, are mostly investigated for odor classification. Although, these algorithms perform well, a low power and portable electronic nose system remains a challenge due to the algorithmic’s complexity and its computationally intensive nature. Moreover, field deployment of electronic nose systems is another issue given the fact that these algorithms require manually tuning of many parameters as well as the requirement for sensors' calibration.

We follow two different approaches to dealing with the challenges of developing odor identification algorithms with high accuracy, namely (i) a closed-form solution for E-nose systems and (ii) quantitative feedback and hardware friendly implementation. One utilizes recent experimental findings about odor identification in the mammalian olfactory system, and the other focuses on some state-of-art pattern recognition algorithms to meet these challenges. In all these classifiers, we transform the multi-gas identification problem into pair-wise classification problems and a quantitative feedback is integrated in each binary classifier to avoid the misclassifications at the cost of rejection of uncertain predictions.

Regarding our first approach, we propose three classifiers. Firstly, we propose probabilistic rank scoring classifier that ranks the sensors features to form rank codes and then a simple probabilistic approach is used to identify the test vectors. In the second classifier, we search pairs of those sensors' features whose difference results in opposite signs for the two gases in each binary classification problem, and these are later used for classification of the test feature vectors. The performance of these two classifiers may be limited when discriminatory information is not found in the ranks or no signed pair is found due to limited number of sensors and overlapping features in the electronic nose systems. To handle this challenge, we propose multivariate Bayesian classifier by overcoming its inherent limitation of poor estimation of covariance matrix through applications of random matrix theory. By using statistical principles, we introduce a computationally efficient classifier by using the mean values of the sensors' features with their weights based on their capability to discriminate the gases. The performance of this classifier is limited when mean values of the sensors’ features do not contain discriminatory information then we propose a cluster k-nearest neighbors classifier by reducing its inherent computational complexity through clustering the data into subclasses and using representatives of the subclasses instead of the whole data set for the classification of a new test vector. The performance of all these classifiers is evaluated by developing three sensor arrays, containing commercial gas sensors, and acquiring data of multiple gases in the laboratory under controlled operating conditions. Our classifiers perform comparable to computations-intensive pattern recognition algorithms despite their hardware friendly implementation and that is further increased to around 100 % at the cost of rejection of uncertain predictions.