Sensor Circuits - The Electronic Nose Artificial Olfaction Technology - page 196

Auto scaling means to scale each sensor to zero-mean and unitary variance.
This operation equalizes the dynamics of the sensor responses avoiding that a
sensor with a larger response range may hide the contribution of other sensors
dynamically limited. Auto scaling is a technique to scale the variables to related
standard deviations and so reduce the effect of signal changes (variance). Further,
auto scaling makes the sensor responses dimensionless; this feature becomes
necessary when sensors whose signals are expressed in different units are joined
in the same array. This is the case of hybrid arrays (different sensor technologies in
the same array) and when electronic noses are fused with other instruments, e.g.,
the fusion of electronic noses and electronic tongues.
There are a number of pattern-recognition methods used to analyze the response
produced by sensor array [ 7 ]. Pattern recognition for odor analysis demands sta-
tistical/mathematical tools and related software as the sensors inherently exhibit
highly

nonlinear characteristics upon introduction to gases with dispersion

occurring at high concentration.
Pattern-recognition techniques are used for data processing of output responses
generated by each sensor of e-nose system. The basic advantages of this method
include the reduction in difficulty of the sensor coating selection, and the capability
to distinguish complex mixtures without the need to identify and quantify individual
components. In this section, we will briefly review the concepts behind this method.
Pattern recognition is a decision vector used to classify class based on a series
of patterns on that class. Usually a matrix is formed from the patterns for a number
of classes and then a decision vector which divides the pattern into an assigned
binary classification is calculated based on standard experiments. This is then used
to categorize unknown patterns. The success of Pattern-recognition techniques can
be improved or simplified by suitable prior handling of the data such that feature
selection and feature extraction are important approaches [ 8 ]. Pattern-recognition
methods are mainly divided into supervised and non-supervised methods, although
a combination of both can be used. The major unsupervised technique is principal
component analysis (PCA) while artificial neural network (ANN) is the best-
known supervised technique [ 6 ].
The signal recorded and the pattern for each sample can make an exploratory
analysis with all the information that is required. The most important multivariable
tool for exploratory analysis is Principal Component Analysis (PCA). Using the
PCA the measured data will be transformed into 2D or 3D coordinates. PCA is a
linear supervised pattern-recognition technique that has often been used to explore
gas sensor array data in conjunction with cluster analysis.
A set of principal components are transformed from a set of correlated variables
such that the first few components define most of the variation in the data set as the
methodology followed in PCA. It is a specific kind of orthogonal projection and its
coordinate system is usually called 'feature space'. A method which is defined as
an unsupervised data reduction method is PCA.
The principle function of the PCA method is to illustrate variations of a mul-
tivariate data set in terms of a set of uncorrelated variables, each of which is a
particular linear combination of the original variables. In short, PCA is a linear