Recently
a new approach has been proposed by several groups for reducing the number of
required sensors [204]-[214].
This approach is based on the exploitation of the time specific information
of sensor responses. If various analytes show different kinetics for the sorption
into the sensitive layer, the resulting sensor response recorded versus time
features a different shape for these analytes. This additional temporal information
can render the parallel information of different sensors in an array redundant
allowing reducing the number of the sensors. For that purpose the shape of the
sensor responses has to be digitized by recording the sensor responses over
time (further referred to as time points) and by performing a multivariate analysis
of these time points. These time-resolved measurements were performed in various
areas of sensor research. For example, Yan et al. [204]
quantified binary mixtures of solvents in water by a single reflectometric interference
spectroscopic sensor whereby the time shift of the highest signal after analyte
exposure depended on the composition of the mixture similar to gas or liquid
chromatography. The components of binary and ternary mixtures of organic analytes
in water could also be determined by the use of a single amperometric sensor
[207],[214]. Thereby, the consumption
of oxygen by the metabolism of microorganisms with different time constants
for the analytes was detected. In the gaseous phase, time-resolved measurements
were used in combination with sensor arrays to obtain additional variables.
Using time-resolved measurements with an array of quartz microbalances coated
with three different polymer films the classification of six solvent vapors
was improved compared with the classification using only the saturation mass
[215].
Johnson et al. [210] classified 20 different analytes
with only 4 fiber optic sensors and also classified these analytes semi-quantitatively
into low, medium and high analyte concentrations using ten fiber optic sensors
with 90 % of the test data being assigned to the correct concentration
class. Podgorsek et al. [216]
used a glassy polyimide for the detection of methanol and ethanol and showed
the difference of the response times, which could be used for the discrimination
of both analytes. However, a detection of both analytes in mixtures was not
performed.
Yet,
all these publications used the time-resolved measurements as phenomenological
tool and no systematic research was performed concerning the optimization of
the different mechanisms and components like the time delaying effects, the
time-resolution of the recorded sensor response, the feature extraction, the
data preparation and the multivariate data analysis, which was rather basic
in most of the publications cited. For example in most of the approaches the
sensor responses were recorded using a rather coarse time resolution [205],[206],[212]
or even a coarse resolution combined with other more or less arbitrary features
[210],[211],[217].
Although a coarse time-resolution grants an easily manageable quantity of information,
the risk of losing important information can be high, especially if several
analytes show similar or very fast sensor responses. Thus, a fine time grid
should be generally preferred, which nevertheless needs a more sophisticated
data analysis. Additionally, all approaches cited above were isolated applications
and no transfer to other systems was performed.
In
this work, an extensive and systematic research on time-resolved measurements
is performed from the investigation of the effects causing the time delays to
an optimization of the data analysis. The principle of time-resolved
measurements is applied to several analytical problems and to several different
sensor devices. In this work, the time-delaying effects are based on a
microporous polymer. For the data analysis several methods are developed, which
allow a highly efficient evaluation of a fine time grid of the sensor
responses. These methods are embedded in frameworks, which allow a simultaneous
variable selection and calibration of nonlinear data and which can be applied
to any linear and nonlinear multivariate relationship.
Ph. D. Thesis