Biomedical Engineering Reference
In-Depth Information
In each application it is very important to apply process understanding at the
data pre-processing stage and then to decide which method of scaling is appro-
priate for a particular application and will not result in the introduction of artificial
features into the data set.
2.3 Data Reconciliation
Data scaling and pre-processing methods described in
Sects. 2.1
and
2.2
represent
only a small proportion of the extensive variety of data reconciliation methods
applied in a range of industrial sectors. A plethora of methods for eliminating
random and gross errors, such as calibration errors, malfunction in devices or post-
calibration drifts, from process data employ model-based approaches to data
reconciliation and treat this problem as a model identification and parameter
estimation problem [
3
]. Various model representations have been used for this
purpose, ranging from straightforward material and energy balances [
58
], extended
Kalman filters [
20
], through to some of the MVDA approaches, such as support
vector regression [
35
], described in more detail in
Sect. 3.3
.
The selection of an appropriate model structure for data reconciliation is not a
straightforward issue, in particular in bioprocessing, where a range of models from
black box to metabolic to synthetic mechanistic models can be used (see Chap. 6
and
Sect. 3.3
). The choice of the right level of model complexity is crucial, but
might in reality also be influenced by the availability of measurements, although
recent advances in real-time process measurement (Chaps. 1 and 9) have reduced
this challenge significantly.
3 Feature Extraction Methods
Exploratory data analysis, feature extraction, pattern recognition, classification and
clustering are very important in bioprocess data analysis for a number of reasons.
The complexity and the amount of data obtained from the various measurements
taken during bioprocess operation or an experiment preclude a comprehensive and
robust analysis of trends 'by eye' even if carried out by the best trained and most
experienced personnel. Traditional chemometric methods, such as principal
component analysis (PCA), allow significant data reduction to eliminate correla-
tion and noise in the data and thus provide a clearer depiction of the relationships
captured by the data. However, the basic assumption of a linear correlation within
the data structures is often highlighted as a major limitation of such techniques
within bioprocess applications, as the clear non-linear and complex nature of the
studied systems is assumed (and proven in a number of applications) to be difficult
to capture by linear methods. Hence a range of non-linear feature extraction
techniques have also been applied to bioprocess data analysis [
39
,
42
].
Search WWH ::
Custom Search