Problem
Oil producers in Russia run electric submersible pumps (ESPs, in Russian UETsN) on more than 115,000 wells. A pump fails about once a year, sometimes more or less often, and replacing it takes several days during which the well produces less. Failure is a random process, and operators rely on the planned time to failure, which hurts prediction accuracy. A pump has more than 40 sensors, so there is plenty of telemetry but no reliable early signal.
Example from the article: on one well, liquid flow dropped sharply shortly before the recorded failure, so the failure looked predictable in hindsight.
Solution in brief
Predicting time to failure or a failed/working label with supervised models gave useless metrics. The team switched to anomaly detection: an LSTM autoencoder learns to reconstruct normal telemetry, and a high reconstruction error flags a pump that is about to fail. The alert threshold is left to the user.
Questions this case answers
- How to predict rare equipment failures from multivariate telemetry when labeled failures are rare and noisy and supervised models give useless metrics
- How to choose an alert threshold for an anomaly detector given the business cost of false alarms
- How to predict failures of electric submersible pumps from multivariate telemetry
- How to detect equipment anomalies when supervised failure prediction does not work
Tools and methods
- LSTM autoencoder — final model, trained on normal data, flags anomalies by reconstruction error
- SVD (singular value decomposition) — compresses features to 3 components before the model
- CatBoost, XGBClassifier, Isolation Forest, CUSUM — baselines that did not work well enough
- TSFresh — extra time-series features for the baselines
- Streamlit — demo app with a threshold slider
Google or a one-time email code. No password.