AI Predictive Asset Maintenance for Electrical Submersible Pumps
Overview
One of the largest global oil & gas companies implemented BHC3™ Predictive Asset Maintenance to predict failures of electrical submersible pumps (ESPs) for high-value wells in two countries.
Previously the company had no reliable way to predict ESP failures, resulting in lengthy unplanned ESP replacements and several weeks of lost production per incident. To reduce unplanned well downtime through more proactive ESP replacement, the company completed a BHC3 Predictive Asset Maintenance project. The objective was to predict ESP failures at 100 offshore wells across several fields in two countries.
In just 14 weeks from project kickoff, the company was able to deploy a production-ready AI machine learning application. The application enabled the company to identify high-risk wells up to 60 days before an ESP failure with 66% recall, significantly reducing maintenance costs and production downtime. The projected recurring annual economic benefit is estimated at €90 million.
About the Oil and Gas Company
- €150+ billion in annual revenue
- 2+ MBOE/day production
- Over 8 million customers per day
- Over 100,000 employees

"The problems we solve come down to people. While our technology provides significant support, we work very closely with our customers.”
Group VP, Products
Building and deploying a Predictive Asset Maintenance application
In just 14 weeks, BakerHughesC3.ai (BHC3) delivered a production-ready AI application, BHC3 Predictive Asset Maintenance, to address the oil and gas company’s objectives. The team worked with the company to ingest, clean and integrate 10 years of operational data, totaling over 10 billion records, for 100 ESP wells across two countries. The unified data model consists of over 30 models, representing data from 16 sources, including telemetry, ESP operational data and characteristics, outage reports, production data, fluid analyses, operational set points and other sources.
Using this unified dataset, BHC3 generated more than 3,000 machine learning features and configured over 1,500 supervised and unsupervised machine learning model permutations to detect anomalous behaviors and predict unplanned outages up to 60 days in advance of a failure. The team configured a user interface to visualize all data received and identify high-risk ESP wells, with a concise explanation of the top features contributing to the anomalous behavior.
Embedding the BHC3 Predictive Asset Maintenance application in its existing business processes and creating a prediction analysis feedback loop, the oil and gas company can continuously improve these results to further reduce production shortfalls, improve prioritization and planning of costly maintenance operations, and understand how to better design future pumps. With the ability to leverage a centralized, unified data image of its remote wells kept current in real time, the company can take any necessary decisions based on the entire data set from “Operations Smart Rooms” located at the headquarters.
The preliminary estimate of the solution’s recurring annual economic value is €90 million, due to increased production and well pump run-life extension time.
Project Highlights
- 14 weeks from kick-off to pre-production application
- Focused on 100 wells in two countries
- Ingested 10 years of historical data from 16 disparate source systems
- 10 billion normalized records available for advanced analytics
- Created a unified data image using 30 BHC3 Models
- Constructed 3,000+ time-based expressions for machine learning
- Built 1,500+ supervised and unsupervised machine learning models designed for daily predictions