On Using CFD and Experimental Data to Train an Artificial Neural Network to Reconstruct ECVT Images: Application for Fluidized Bed Reactors

• Published in: Processes Vol. 12; no. 2; p. 386
• Main Authors: Montilla, Carlos, Ansart, Renaud, Majji, Anass, Nadir, Ranem, Cid, Emmanuel, Simoncini, David, Negny, Stephane
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.02.2024; MDPI
• Subjects: Algorithms; Approximation; Artificial neural networks; Capacitance; Chemical and Process Engineering; Configurations; Electrodes; Engineering Sciences; Fluid Dynamics; Fluidized bed reactors; Fluidized beds; Image reconstruction; Learning strategies; Machine learning; Measurement techniques; Neural networks; Neurons; Physics; Simulation; Supervised learning; Tomography; deep learning; 3D ECT; ECVT 3D ECT fluidization deep learning multi-phase flow; multi-phase flow; ECVT; fluidization
• ISSN: 2227-9717; 2227-9717
• DOI: 10.3390/pr12020386

Electrical capacitance volume tomography (ECVT) is an experimental technique capable of reconstructing 3D solid volume fraction distribution inside a sensing region. This technique has been used in fluidized beds as it allows for accessing data that are very difficult to obtain using other experimental devices. Recently, artificial neural networks have been proposed as a new type of reconstruction algorithm for ECVT devices. One of the main drawbacks of neural networks is that they need a database containing previously reconstructed images to learn from. Previous works have used databases with very simple or limited configurations that might not be well adapted to the complex dynamics of fluidized bed configurations. In this work, we study two different approaches: a supervised learning approach that uses simulated data as a training database and a reinforcement learning approach that relies only on experimental data. Our results show that both techniques can perform as well as the classical algorithms. However, once the neural networks are trained, the reconstruction process is much faster than the classical algorithms.