Implementation of a Control Strategy for Hydrodynamics of a Stirred Liquid–Liquid Extraction Column Based on Convolutional Neural Networks

Abstract

Online process supervision is enabled by smart sensor development, hence attracting increasing attention in pharmaceutical and chemical process engineering. Additional sensor data enable more precise process control as additional process parameters can be monitored. They are easy to integrate into modular plants, and their provided additional process parameters enable a more flexible operation of the apparatus due to the quick and more sensitive reaction to changing circumstances. An artificial intelligence-based optical sensor for the investigation of different operating states and droplet sizes within a liquid−liquid stirred DN32 extraction column in counter current flow is developed and presented in this work. Two operating states, the flooding state and the column's regular operating state, are differentiated as observable states. Additionally, the diameter of the rising liquid droplets of the disperse phase is categorized into different diameter classes. A control strategy for the extraction column is derived based on the results of the convolutional neural network-based image analysis. Thus, a robust soft sensor controlling the hydrodynamics of an extraction column was developed. The developed control strategy automatically leads the extraction column into a favorable hydrodynamically stable operation state.

Detecting Crystals in Suspensions: Convolutional Neural Networks vs. Gravity‐Based Approach for Size Distribution Detection

“…A method to implement machine learning (ML) inspired by biological neural networks are artificial neural networks (ANN). Convolutional neural networks (CNNs) are a subgroup of ANNs, especially designed for processing data with a grid-like topology, such as images [21]. ANNs consist of so-called artificial neurons, functions that calculate an out-put from several weighted inputs.…”

“…ANNs consist of so-called artificial neurons, functions that calculate an out-put from several weighted inputs. Deep learning (DL) deals with neural networks that consist of several layers of neurons and are therefore referred to as deep [21].…”

“…To use it the terms true positives t p , where positive cases are identified as positive, true negatives t n , where negative cases are identified as negative, false positives f p , where negative cases are identified as positive, and false negatives f n , where positive cases are identified as negative need to be introduced. Then the accuracy is defined as 21, 23: …”

P50 - Künstliche Intelligenz (KI)-basierte optische Sensorik für flüssig-flüssig Systeme

“…Über 99 % Genauigkeit bei der Klassifizierung ermöglichen die online Überwachung des Betriebs der Extraktionskolonne. Für weitere Details sei hierzu ebenfalls auf Referenz [7] In Zukunft ist geplant die Tropfen erkennende KI ebenfalls anzuwenden um Tropfen in Schläuchen analysieren. Da Schläuche, beispielsweise als Coiled Flow Inverter (CFI), ein vielversprechendes Reaktordesign bieten, um Batch-Prozessen auf kontinuierliche Prozesse zu übertragen [9] kann auch hier eine automatisierte optische Prozessüberwachung einen Mehrwert bieten.…”

“…8: Überlagerung der von dem KI-Algorithmus Mask-RCNN[8] erkannten Tropfen im Stoffsystem Butylacetat/deionisiertes Wasser in der gerührten Extraktionskolonne Die daraus resultierende durchschnittliche Tropfengröße des Systems wird berechnet, um zu sehen, ob die aktuellen Einstellungen für eine optimale Hydrodynamik der Extraktionskolonne sorgen. In der Kolonne zeigte sich durch frühere Arbeiten eine durchschnittliche Tropfengröße von 1 mm als der hydrodynamisch bevorzugte Betriebszustand [7]. Ergebnisse dieser Untersuchung werden in Abb.…”