Sensor Circuits - The Electronic Nose Artificial Olfaction Technology - page 201

probability distribution. This means that when the same odor is accessible in the
extended time, all codebooks are moved toward the cluster that represents the
current input, and even to others associated to different odors. Thus, other code-
books (related to different odor classes) become closer to the current one,
destroying the historic knowledge base and making the system unstable for the
task at hand. Other disadvantages of this technique is, the SOM that uses fixed
network architecture in terms of number and arrangement of neural processing
elements, has to be defined prior to training.
This problem has been solved by using an architecture based on multiple
SOMs, each associated to a single odor to be recognized [ 13 ]. This network adjusts
itself to new changes of the input probability distribution by means of repetitive
self-training processes. Once each map self-organizes its codebook vectors, it
refines them by using a learning vector quantization algorithm in order to reduce
the high uncertainty accumulated at the borders of two or more different clusters.
The self-training processes are carried out in an autonomous fashion during the
testing phase to track odor patterns with changing statistical distributions [ 14 ].
For the case of mostly unknown input data characteristics it is available as
insignificant to determine the network architecture that allows for satisfying
results. During the unsupervised training process there is worth to consider a
neural network models that determine the number and arrangement of units.
Cluster Analysis
The basic theory with these methods is that measurements made for related
samples be likely to be similar. For similar samples the distance between samples
is smaller compared to unrelated samples. It is a data reduction tool that creates
subgroups that are more manageable than individual datum.
In cluster analysis there is no previous information about which elements
belong to which clusters. The grouping or clusters are defined through an analysis
of the data. For organizing observed data cluster analysis is an exploratory data
analysis tool. Figure 7.24 illustrates the steps involved in the cluster analysis.
There are unsupervised cluster methods:
• Univariate clustering:
It calculates individual variables and groups samples into homogeneous classes.
Univariate clustering is useful to provide a proportional basis for exploring the
data. Performing univariate study is a common way of exploring the data at
hand, before turning to more complex analysis, such as multivariate analysis.
The univariate analysis is applied on observable attributes.
• Hierarchical cluster analysis:
In this analysis, reduction of multiple variables of a sample is done to a single
'distance' value. Rank and link samples are based on relative distances. For
finding relatively homogeneous clusters of cases based on measured charac-
teristics main statistical method is hierarchical cluster analysis. In this process, it
starts with the each case as an element cluster, there are many clusters and these
clusters are combines sequentially and it reduces the number of clusters at each