Sensor Circuits - The Electronic Nose Artificial Olfaction Technology - page 197

feature extraction technique which shrinks the dimensionality of data with the least
loss of information. This technique is achieved by projecting the data onto fewer
dimensions which are chosen to utilize the relationships between the variables, so
that the maximum amount of information is retained in the smallest number of
dimensions. This technique allows the similarities and differences between sam-
ples to be better reviewed.
Exploratory (Multivariate) Techniques
The most widespread multivariate tool for exploratory analysis is still Principal
Component Analysis (PCA). There are sets of data given related to a number of
measurements, multivariate techniques that aim at studying the intrinsic charac-
teristics of the data to discover data internal properties. Multivariate analysis
supports the attitude of an electronic nose to be utilized for a specified application,
leaving to the supervised classification the task of building a model to be used to
predict the class membership of unknown samples. Two main groups of multi-
variate analysis may be identified: dimensionality reduction techniques and clus-
tering techniques.
Considering and understanding multivariate data is the key to success. The
available list of multivariate methods is long (as we listed same in Chap. 5 ) , but
only a limited number of methods are normally used in gas sensing. A group of
algorithms aimed at providing a representation of the data in a space of dimensions
lower than the original sensor space is defined as dimensionality reduction
technique.
All techniques are based on specific propositions about the nature of the data
and the sensor space. Each technique is responsible to maintain some particular
and defined characteristic of the data. The simplest, calculus, and interpretation of
results are based on the strongest assumption about the statistical distribution of
the data. A neural network is necessary for data representation where assumption
of data distribution is removed.
Clustering techniques are generally based on the theory of connection expressed
through the definition of a metric (distances calculus rule) in the sensor space. The
most insignificant and common choice is to express the similarity as a Euclidean
distance [ 9 ].
Principal Component Analysis
From neuroscience to computer graphics a simple, non-parametric method of
all types of data analysis is one and only Principal Component Analysis (PCA).
PCA is required for information extracting from confusing data sets. PCA is a
method to recognize patterns in data, and state the data in a way to highlight their
similarities and dissimilarities [ 10 ]. In simple ways, PCA is a numerical method
for analyzing the basis of variation present in a multidimensional data set [ 11 ].
From applied linear algebra most valuable results have been described by PCA.
For reducing complex data sets to a lower dimension to reveal many hidden,
simplified structure that often are most useful, PCA is implemented with minimal