Manufacturing & Distribution · Interactive demo
How much useful life remains, and how uncertain is that estimate?
An engine-held-out remaining-useful-life model that combines multiscale sensor features, calibrated uncertainty, driver proxies, and a transparent maintenance-cost scenario.
Data source and licence
SourcesEngine sensor tracesRun-to-failure cycles · 21 sensors
NigatForecast and decision modelBaseline · uncertainty · monitoring
OutputDecision rangeSchedule on the lower bound, not the estimate
Remaining useful life and uncertainty band
cycle — of —
90% intervalPredicted Benchmark actual
Driver proxies at this cycle
- Test RMSE
- …
- Nominal interval
- 90%
- Measured coverage
- …
- Held-out engines
- …
Business problem
A point prediction makes maintenance look precise when the model is not. One number carries no window, no evidence and no cost, so the same figure is used to justify stripping a healthy machine and to justify running a failing one.
Techniques used
- Multiscale rolling sensor features
- Extra Trees ensemble
- Engine-level train/calibration/test split
- Finite-sample split-conformal interval and cost scenario
What the workflow can improve
- Estimate remaining useful life with uncertainty
- Show risk progression and sensor drivers
- Compare planned and failure exposure
- Create a path to asset-specific maintenance policy
From demonstration to production
Replace the proxy with the customer’s operating evidence.
The production path begins with a bounded workflow scan, source and permission mapping, baseline measurement, user and administrator roles, and an explicit decision owner. In an engagement, your rules, costs and constraints replace the public ones, and the result is measured against your baseline.
More demonstrations
See the same approach in another sector.
Start with the operating problem
Show us where the work slows down.
You do not need a polished technical brief. Describe the decision, handoff, report, document workflow, or AI initiative that is harder than it should be.