Using an electrical submersible pump mechanical vibration and a Fourier convolution neural network to estimate the water cut in two-phase liquid-liquid flows
Abstract
In oil exploitation, water in the production is common and increases with the field life. For oil
production to be economically feasible, artificial lifting techniques may be required. In this context,
electrical submersible pumps have been widely used to provide energy to the fluid for decades.
Despite the recent development on ESP vibration signal usage, most studies use the vibration signal
for fault diagnosis algorithm development. On the other hand, the water cut is essential to obtain
several volumetric variables, such as the flow rate of each phase, the effective viscosity,
and the pressure drop. This study aims to obtain an artificial neural network that correlates the ESP
mechanical vibrations with the water-liquid ratio of an oil-water two-phase flow. The artificial neural
network uses a Fourier convolution neural network architecture, where the convolutions are
performed in the frequency domain rather than the time domain. The experimental data obtained
resulted in a non-uniform dataset on different spatial projection, which significantly affected the
data-driven technique performance. Then, by filtering out spaced experimental points and using only
the samples inside the sample convex hull, it was possible to successfully obtain a regression
between the ESP vibration spectrum and the water cut. The results showed that data filtering and
selection were crucial for the artificial neural network performance.