Context
Degradation in rotating machinery shows up gradually in operating data. The hard part is turning that drift into a number of cycles remaining.
Getting the number wrong costs in both directions. Call it early and the unit goes to the shop with life still in it. Call it late and the shop visit becomes an unplanned one, with the schedule rebuilt around it.
The methods available for this each involve a trade. Statistical approaches are light and easy to interpret, but capturing nonlinear degradation spread across several sensors takes real design and tuning effort. Deep learning delivers the accuracy, at the cost of the compute needed to train it and the environment needed to run it on site.
What we tested
The input is the sensor data already recorded during operation. No additional instrumentation is required.
The model estimates the number of cycles remaining before failure. It is built on a class of lightweight machine-learning techniques (reservoir computing) suited to time-series signals.
Validation used a public benchmark covering several engines run to failure. The data is generated from an engine simulation and is a standard reference in the prognostics field. Engines held out from training were used for evaluation.
What it shows
Remaining cycles could be estimated on engines the model had not seen during training. As degradation progresses, the estimate converges towards the actual remaining life.
This may apply to your equipment
If you operate a fleet of identical machines and hold records of units that reached failure or a shop visit, the same approach may transfer.
It can support levelling out shop induction dates, ordering spares ahead of need, and deferring work on units that still have life in them.
A short conversation is usually enough to tell whether your operating data is in a usable form.