Figure 3: Sliding window method for the transformation of time series
data.
This transformation results
in many additional data columns, which must be considered for the choice
of an appropriate ML model. For the de-noising of the pressure drop
signal, an exponentially weighted moving average (EWMA) is used that
weighs the most recent measurements stronger as they are more important
to detect changes in trend and level of the pressure drop. For some ML
methods that are based on distance (SVM, Clustering), an additional
scaling step is necessary that normalizes the input data. Linear
regression and decision tree regressors do not require this scaling
step.
Feature extraction via machine learning
ML
The spinning band distillation column contains the following sensors and
control variables summarized in Table 1.
Table 2: Sensors and control
inputs of spinning band distillation column relevant for ML.