Figure 1: (left) normal flooding behavior with increasing band speed
(right) spinning band column with a view of the uninsulated column with
vacuum double jacket and internal
Besides the already mentioned actuators and sensors, the column is
equipped with necessary sensors and actuators to be operated nearly
fully automated. The automation is executed on a PLC by WAGO
Kontakttechnik GmbH & Co. KG, Minden, Germany, and features the
manufacturer independent interface module type package (MTP)
standardized in the VDI/VDE/NAMUR 2658 guidelines. Besides this, the
distillation column offers a service “distill” with a state machine
implemented in the decentralized logic of the modules PLC. Enabled by
the MTP, the distillation column can be integrated to the prototype ABB
process orchestration layer (POL), where the service can be conducted
and the human machine interface (HMI) is automatically visible. Deeper
insights of the MTP concept and its architecture can be taken from , the
architecture combined with the ML implementation is explained in more
detail below.
Regarding the flooding phenomena, the distillation column is behaving
similarly to conventional distillation columns. The pressure drop is
increasing nearly linearly with the gas factor and liquid loading.
Beyond certain gas factors and liquid loadings, called loading point,
the pressure drop is increasing more steeply than before until the
flooding point is reached, which is an undesirable state of the column.,
In the case of the spinning band distillation column, the pressure drop
is also massively influenced by the spinning band speed. This behavior
could be already implemented to a control strategy , but still, the
column is running into the undesired state of flooding. The complex
hydrodynamics of the system with a rotating internal and the two-phase
flow is very hard to model and predict precisely. Hence, the flooding
points need to be examined experimentally. This procedure is very time
consuming and still has some uncertainties, which cannot be described
completely. Hence, it is from high interest to predict the pressure drop
and classify the current operating point with help of a trained ML
algorithm with historical data. Like this, it would be possible to
already get information about a certain parameter set and a
classification even before undesired states like the flooding point is
reached in operation.
For the implementation of ML
scripts, following hard- and software architecture is used. Sensor data
from the SBDC is transferred to and stored on the OPC/UA server of the
PLC. Current and historical values for each PEA can be accessed by the
POL via the MTP interface. In order to implement ML methods to design an
early flooding warning system, the sensor data is directly taken from
the OPC/UA servers via a Python script and fed to the developed
monitoring system to determine the current operating state of the
distillation column by classification. The monitoring concept is
implemented in four steps: i) pressure drop preprocessing and filtering
ii) forecast by supervised learning methods, iii) operating state
classification through clustering, iv) graphical output for the operator
with forecast and classification. Data flow and implementation of the
flooding prevention system are visualized in Figure 2.