Sensor Technologies - Artificial Human Sensors: Science and Applications - page 98

handling of the signal processing.
4.5.1 Multivariate Data Analysis
Nature is multivariate, i.e., a single phenomenon normally depends on several
other occurrences, and the measurement object of interest in a sensor application
is expected to behave in a similar manner. Univariate methods, where a serial
approach deals with one variable at a time, are rarely found in nature. This phe-
nomenon is often of limited interest in complex data analysis. However, in all
cases of measuring procedures it is a necessity to control, or at least be aware of, all
these factors, in order to understand the multivariate behaviours of the dynamical
sensor systems.
Multivariate data analysis in a powerful tool to increase the knowledge about
complex behaviours of the operational principle and to find properties in different
measurement conditions. Further, it may also include the possibility to learn more
about the behaviours and how a pattern, e.g., the measured response picture, are
expected to occur in future operations. Software techniques, that use multivariate
analysis, have been developed and applied in numerous applications, with more
or less success. The main objective is however, as always to find, explain and learn
as much as possible about the sensor specifications, limitations and behaviours
including the properties of the complete multi-sensor system.
When analysing multivariate techniques for classification of data, a structural
approach can be used by the following road map. The analysing parameters begin
with the procedure closest to the sensor unit, and may be an indicative procedure
to approach when organising the system structure as given below.
— Pre-processing , describes the computational techniques to structure the data
in a shape that is convenient to use in the following analysis. For example,
normalisation of sensor data, compensation for a reference level and a com-
mon reduction of biases can result in a better performance in later analysing
stages.
— Exploratory analysis , is considered to be an initial estimation of the measured
data and provides a direction of advice of significant variables respectively
possible outliers, i.e., the discrimination between measuring variables and er-
roneous data, the variables and noise. Suitable techniques, like for example
feature selection, feature extraction can be found in the literature, Petersson
(2008).
— Classification , is a technique to analyse processes in order to determine the
existing relationship between the multi-sensor data, i.e., a set of indepen-
dent, selective single sensor variables, and their relationship to each other, in
other words a set of dependent classes, Esbensen (2000). A number of classi-
fication models exist. For example, statistical techniques by Multiple Regres-
sion Technique, MLR, Partial Least Sqares, PLS, or Cluster Analysis, CA, is
further described in, e.g., Batagelj (2006), Loutfi (2006). However, for non-linear