Fault diagnosis of electric submersible pumps using vibration signals

The present work assesses the feasibility of using vibration signals to establish correlations between different types of faults (a state) and related failure modes (an event) in electric submersible pumps (ESPs). Most available diagnosis software strives to distinguish between normal and abnormal conditions in centrifugal pumps, but are not capable of directly correlating a detected type of failure to an existing mechanical fault. Here, several types of controlled mechanical faults are applied to seven different ESPs as a means to try to correlate via spectrum analysis typical signatures to given fault-failure pairs. Pressure and flow rate data together with vibration signals were collected to detect the wear state and operational conditions of known pumps. The acceleration data were analyzed using Spectral Analysis and Power Spectral Density techniques. This paper introduces an approach based on Random Forests, an algorithm that uses decision trees for classification and regression. The work shows that the proposed procedure is accurate and general enough to allow fault-failure identification and classification.