At a glance
Oil and gas producers risk costly production stoppages when poppet valves fail without warning. SoftServe built a machine learning model to predict valve failures using time-series data from more than 100 sensors. The model generates daily predictions for a 10 to 14 day window ahead.
During a four-week PoC, the model reached a test AUC (area under the curve, a measure of prediction accuracy for unbalanced classes) between 0.62 and 0.69. The prediction capability lets producers schedule maintenance in advance and increase equipment uptime.
Client background
The client delivers an enterprise cloud platform built on a private network, with roots in nearly 20 years of work across security, network architecture, collaboration, artificial intelligence, and open-source software. Customers in more than 150 countries use the client's tools to modernize their computing environments.
The client wanted to give its oil and gas customers a way to predict valve failures in advance, so maintenance teams could act before a costly shutdown.
Business challenges
The client's oil and gas customers had no reliable way to anticipate valve failures before they happened, so unplanned stoppages cost time and revenue on the factory floor. Poppet valve failures during pressurization forced factories to stop or pause production, and no existing model translated input from over 100 sensors into a failure prediction. These customers needed a way to turn diverse sensor inputs into an early warning signal for maintenance teams.
Activities & Solutions
SoftServe built a machine learning model that predicts poppet valve failures 10 to 14 days before they occur, using time-series data from over 100 sensors across the factory. The model makes a fresh prediction each day, drawing on historical failure patterns to flag rising risk in individual valves.
Real-world data introduced key modeling challenges: sensor noise, too few individual failures to reveal clear patterns, and little visible difference between failure and maintenance periods. These challenges produced errors in the labeled training data that the team resolved to keep predictions reliable.
Training data included:
- Cooler, plunger, and hyper-section temperatures
- Plunger positions, discharge pressures, and gas flow rates
- Statistics on about 100 historical failures across eight components
Technologies used:
- XGBoost, a gradient-boosting algorithm, with group-based cross-validation
- Python for data processing and modeling
- Google Cloud Platform (GCP) for data storage
Value delivered
SoftServe delivered a proof of concept that gives oil and gas producers a data-driven way to anticipate valve failures and schedule maintenance before a shutdown occurs. The model turns dense sensor data into a daily, actionable forecast within a matter of weeks.
| Metric |
Before | After |
| Valve failure prediction window | None | 10 to 14 days in advance |
| Daily prediction accuracy (test AUC) | Not measured | 0.62 to 0.69 |
| Time-series features generated daily |
0 | Nearly 5,000 |
| Proof of concept timeline |
N/A | Delivered in four weeks |
Ready to predict failures before they happen? SoftServe builds machine learning models that turn sensor data into early warnings for critical equipment. Contact us to explore what predictive maintenance could do for your operations.
