Electric submersible pump (ESP) failures are difficult to anticipate ahead of time, therefore having an automated failure detection and diagnostic system that can discover anomalies before they evolve into more troublesome issues is vital to the financial optimization of ESP operations. The objective of this research was to evaluate the application of computer vision as a viable technique to create machine learning models for ESP failure detection and diagnostics.
The intuition for this work comes from observing the diagnostics process followed by trained ESP surveillance engineers; they are able to diagnose problems by visually finding clues in signal behavior and trends. This characteristic visual pattern recognition of ESP diagnostics was the basis to consider image classification as a better alternative to other machine learning approaches for developing machine learning engines for this application. Furthermore, traditional machine learning techniques face major limitations when it comes to analyzing time-series data, because of the constraints on effective feature extraction, or the limited receptive fields.
This paper summarizes the methodology used to train an image classification model to recognize trends in time-series signals for the detection of abnormal ESP conditions using only pictures of the telemetry signals. Transfer learning was used as part of the training strategy to draw from the performance of pre-trained models. This work concludes by presenting the results on the effectiveness of computer vision as a tool to automate ESP failure detection and diagnostics.