562
R. W. Kessler and W. Kessler
processor nodes in the system and provide the operator screens. Level 3 represents the
production control level, which, however, does not directly control the process, but is
concerned with monitoring production and setting targets. Level 4 is the production
planning level. The future sensor levels may well be more decentralized, but the
decision hierarchies will be retained (see also Fig. 23.4) [24].
For the future intelligent process analyzers, the following important goals apply
in addition to miniaturization and cost reduction described in [22, 23, 26] and
summarized in [1]:
• Plug-and-play capability.
• Self-configuration, self-calibration, and self-optimization.
• Higher sensitivity for inline trace analysis.
• Imaging ability.
• Multimodality with “all-in-one” functionality.
New and miniaturized instruments can drastically reduce costs and thus open
up further possibilities for use. The user should also be given the possibility of
“plug-and-play.” Since miniaturization allows the integration of different measurement technologies, multimodal information can be accommodated in a single sensor
[1, 26].
Data fusion, i.e., the combination of different data sources for classification, can
be carried out at different levels. At the low level, the concatenation of data matrices
at the variable level is possible. At the middle level, data matrices are concatenated
using a feature selection, for example, at score level after applying a PCA. And at
the high level, model predictions are combined, e.g., predictions from PLS models
[27–29].
In addition to the already established multivariate data analysis and other chemometric tools in the spectroscopically based control systems, the concepts of machine
learning (ML) and deep learning (DL) are also used in “Industry 4.0” applications which is described in more detail in [1, 30]. ML and DL algorithms are
used to describe the system behavior instead of the previous physical–technical or
chemometric-based models. This is usually referred to as predictive analytics. It must
be emphasized that feature engineering is included in the ML workflow, whereas it
is not available in DL. DL data acquired from the sensors (“raw measurements”)
can be entered directly into the deep learning algorithms. Thus, ML is more oriented
towards traditional engineering methodology, while DL algorithms are largely based
on artificial neural networks.
R. W. Kessler and W. Kessler
processor nodes in the system and provide the operator screens. Level 3 represents the
production control level, which, however, does not directly control the process, but is
concerned with monitoring production and setting targets. Level 4 is the production
planning level. The future sensor levels may well be more decentralized, but the
decision hierarchies will be retained (see also Fig. 23.4) [24].
For the future intelligent process analyzers, the following important goals apply
in addition to miniaturization and cost reduction described in [22, 23, 26] and
summarized in [1]:
• Plug-and-play capability.
• Self-configuration, self-calibration, and self-optimization.
• Higher sensitivity for inline trace analysis.
• Imaging ability.
• Multimodality with “all-in-one” functionality.
New and miniaturized instruments can drastically reduce costs and thus open
up further possibilities for use. The user should also be given the possibility of
“plug-and-play.” Since miniaturization allows the integration of different measurement technologies, multimodal information can be accommodated in a single sensor
[1, 26].
Data fusion, i.e., the combination of different data sources for classification, can
be carried out at different levels. At the low level, the concatenation of data matrices
at the variable level is possible. At the middle level, data matrices are concatenated
using a feature selection, for example, at score level after applying a PCA. And at
the high level, model predictions are combined, e.g., predictions from PLS models
[27–29].
In addition to the already established multivariate data analysis and other chemometric tools in the spectroscopically based control systems, the concepts of machine
learning (ML) and deep learning (DL) are also used in “Industry 4.0” applications which is described in more detail in [1, 30]. ML and DL algorithms are
used to describe the system behavior instead of the previous physical–technical or
chemometric-based models. This is usually referred to as predictive analytics. It must
be emphasized that feature engineering is included in the ML workflow, whereas it
is not available in DL. DL data acquired from the sensors (“raw measurements”)
can be entered directly into the deep learning algorithms. Thus, ML is more oriented
towards traditional engineering methodology, while DL algorithms are largely based
on artificial neural networks.