In the presented research, an assessment was conducted of how machine learning techniques, in combination with physics-based damage modelling, could be capitalized on to help move towards the goal of developing an effective and automated real-time ESP failure prediction system. A prediction model was developed that has the capability of predicting two groups of failure mechanisms. One group consisted of failures related to pump wear and scale plugging and the other group consisted of failures related to increased drag (rotational resistance) in either the motor or the pump. The model was designed to both predict the remaining useful life (RUL) of the ESP and provide an alarm when the ESP is expected to fail within the next 90 days (PF_90). The base form used for the model was a Long Short-Term Memory (LSTM) neural network. Interesting aspects of the model include: specifically designed training targets allowing for both an RUL estimate to be provided over the life of an ESP and the simultaneous prediction of both RUL and PF_90; specifically designed inputs calculated from the raw data as based on the physics-based damage modelling; and a unique architecture for the LSTM network, allowing for multiple failure mechanisms to be better represented. The performance of the prediction model was promising: an average score of 73% (on a test dataset) was achieved using the chosen performance metric for the research, the F1_Score.

You can access this article if you purchase or spend a download.