Artificial Odour Classification System

Introduction

The visibility of electronic nose (e-nose) in wide range of industrial applications such as food quality or authenticity (Yu et al., 2009; Falasconi et al., 2012; Gliszczyńska-Świglo & Chmielewski, 2016), medical diagnosis (Silvano Dragonieri et al., 2007; Yusuf et al., 2013), and indoor air quality (Abu H. et al., 2012) has been acknowledged by both academicians and practitioners. This electronic sensor has the capability of imitating the human sense of smell using sensor arrays and pattern recognition system, where e-nose assists for quick decision making and overcomes some drawbacks of dependency on olfactory experts or trained human panels (Masnan et al.; 2012). Despite of such advancement, the whole technology on e-nose is still emerging including the mathematical algorithms for classifying diligently smell/odour based on a systematic classification procedure. In details, classification procedure is a form of systematic judgment, often in a mathematical function that deals with a process of training a classification rule to familiarise with the behaviour of groups of instances (Mahat et al., 2015). The classification procedure generally is performed on three main steps; (i) construct an appropriate classification rule using a classified data, (ii) evaluate performance of the constructed rule to reflect its ability in classifying instances to their right groups, and finally (iii) use the constructed rule for allocating future instances with unknown group to one of the predetermined groups.

Human olfactory system includes nose and many parts of the brain. The ability to smell comes from specialised sensory cells, which termed as olfactory sensory neurons found in a small patch of tissue high inside the nose where these cells connect directly to the brain (Haddad et al., 2010). As printed in many biology textbooks, microscopic molecules released by substances will be detected by neurons. Then, these neurons will send messages to brain, which identifies the smell. Such nature procedure was then mimicked by engineers, which later devoted the electronic nose technology commonly known as e-nose. As a first step in producing a valid and reliable instrument, an e-nose needs to be trained with some qualified data sets, in attempt to build a database of reference. Then, the developed instrument may recognise future samples by comparing volatile compounds to those stored in the database. To achieve these, often quantitative method called supervised classification is chosen.

In supervised classification, having sets of raw e-nose data, one needs to construct the best possible classification rule such that the rule would enable one to determine the group of instances correctly. However, in common engineering setting for e-nose, sample data is obtained from an electronic setup consists of p features from g different groups. Unless the conducted experiments deal with adequate number of sample, small number of sample compared to the measured features (n ≤ p) will lead to some mathematical problems. Among of these problems is singular covariance matrix that may hinder one to pursue with more complex analysis as the covariance matrix is a common statistics used in many statistical classification rules. Singularity may occur when most features in a data set are correlated among each other. Even, if the covariance matrix is free from singularity, dealing with many features in an analysis consumes great computational cost, where bias estimators could be produced (Masnan et al., 2015).