### Phase inversion identification in Electrical Submersible Pumps using mechanical vibrations

#### Resumo

Electric Submersible Pumps (ESPs) are multistage centrifugal pumps used in the artificial lift and transport of multiphase fluid mixtures. The flow regime is a liquid–liquid flow when the fluids correspond to two non-miscible fluids. Liquid–liquid flow is a mixture with a continuous and dispersed phase. As the amount of fluid in the dispersed phase increases, the dispersed phase suddenly becomes continuous and vice-versa. This transition phenomenon is called phase inversion. The flow regimes in oil–water mixtures are oil-in-water (o/w) and water-in-oil (w/o) flow regimes. This work demonstrates a correlation between the flow regime and the flow-induced vibration (FIV) in ESP operating with an oil–water mixture. This research proposes a novelty method to flow regime identification based on the Root Mean Square (RMS) of the vibration acceleration of the Fast Fourier Transform (FFT) signal.

The experimental setup consists of an 8-stage Electrical Submersible Pump (ESP) and a vibration acquisition system with six accelerometers uniformly distributed along the ESP. The experimental procedure consists of changing the water cut (percentage of water) from the oil flow regime to the water flow regime, maintaining stable ESP rotational speed, the total flow rate, and the oil viscosity. For each water cut, mechanical vibration is collected. The operational conditions consider 30, 40, and 50 Hz rotational speeds and viscosities between 70 and 210 cP.

Frequency domain analysis involves studying FFT between 0 and 5000 Hz, considering different water cuts and frequency ranges. Statistical features – mean, variance, geometric mean harmonic mean, and RMS – were extracted from the FFT for each frequency range. Results showed a strong correlation between the RMS of FFT and the phase inversion phenomena considering the rotational speed. A logistic regression model was employed to establish a transition boundary between oil-in-water and water-in-oil using 10% of the data. The model successfully separated at least 95.67% of the remaining data in the least favorable scenario.