Fault identification in an electrical submersible pump

The research article 'Fault identification using a chain of decision trees in an electrical submersible pump operating in a liquid-gas flow' has been published in Elsevier's Journal of Petroleum Science and Engineering (Volume 184, January 2020, 106490).


The monitoring of centrifugal pumps is essential for the suitable operation of several industrial applications. The reliability of petroleum artificial lifting systems that use Electrical Submersible Pumps (ESP) depends substantially on the performance of these pumps. ESP can operate subjected to severe operating conditions such as viscous and two-phase flow. In recent years, real-time technologies based on machine learning algorithms have gained importance due to the capability to take advantage of historical data for future predictions.

The present work proposes a particular assembly of Classification and Regression Trees (CART) for the detection and classification of incipient faults in a pumping system. Experiments were carried out on a ten-stage ESP to simulate, monitor and label the faults. The pump worked at 1800, 2400, 3000 and 3500 rpm, with a two-phase liquid-gas mixture. The gas-phase was nitrogen, and the liquid-phase was a heavy oil with a viscosity between 200 and 1000 cP. The proposed methodology, named here as Chain of Decision Trees, observe the system behavior based on the monitoring of the pressure, flow, torque, and temperature only. The algorithm has two steps. The first determines whether the system is in a fault state; if it is, the second determines the type of fault. The failures considered were the unexpected closure of the choke valve, the input pressure decreasing, the fluid viscosity increasing and the gas flow rate increasing.

The proposed approach intends to improve the balance in classification and the interpretation of the cause of failure. The Chain of Decision Trees and the Decision Tree were compared regarding the overall accuracy and the individual fault misclassification getting a reduction in individual misclassification and better comprehensibility for the Chain of Decision Tree.

Access the complete article on ScienceDirect.