International Society for Industrial Process Tomography

Long Short-Term Memory Neural Networks for Flow Regime Identification using ECT

Rafael Johansen¹, Torbjørn Grande Østby¹, Antoine Dupré², Saba Mylvaganam^1*

¹Department of Electrical Engineering, IT and Cybernetics, Faculty of Technology, Natural Sciences and Maritime Sciences, University of South-Eastern Norway, 3901 Porsgrunn, Norway
²Private Practice, Jouquetti, 05400 Furmeyer, France

*Email: saba.mylvaganam@usn.no

ABSTRACT

In multiphase flow related experiments and simulations, flow regime identification (FRI) and the transitions from one flow regime to the other are of interest to process engineers as well as researchers dealing with computational fluid dynamics (CFD). This paper presents some results using deep learning analytics in identifying flow regimes in water-air flow measurements, using an Electrical Capacitance Tomographic (ECT) system. The ECT system consists of single plane 12-electrodes, operating at 500 fps delivering 66 independent capacitance values in 2 ms. By carefully selecting the corresponding mass flow rates of each phase, an experimental matrix is formed showing the different flow regimes in the flow rig. All measurement series from experiment #1 to #84 were divided into two data groups, training data with 7500 frames and validation data with 7499 frames. All 84 experiments were repeated forming a set of data for testing, whereas the former set was used for training and validation. Pre-processing the data from the experiments referred to in Table 1 was necessary to achieve the right balance of flow regimes before using them in the machine learning algorithms. As the algorithms used in ANN use gradients, which in a typical recurrent neural network (RNN) with many hidden layers can vanish or escalate, the usage of Long Short-Term Memory (LSTM) neural networks circumvent to a certain extent, problems associated with such gradient excursions, at the same retaining memory over short durations.After introducing the neural network architecture consisting of 2 LSTM in hidden layers and the 66 capacitance values as inputs delivering 5 outputs for the annular, plug, slug, stratified and wavy flow regimes, the success scenario of the algorithm in FRI of the five regimes are presented.The presented results show that the LSTM network is successful in FRI. Some of the wrong FRI results due to anomalies and ambiguities observed in the FRI algorithms can be filtered out by fusing other FRI methods such as the one based on eigenvalues.

Keywords ECT, flow regime identification (FRI), long short-term memory (LSTM), neural networks (NN), machine learning, deep learning.

Industrial Application Oil & Gas, Multiphase flow studies, CFD.