Cross-Machine Transfer of Time-Series Foundation Models for Remaining Useful Life Estimation
Cross-Machine Transfer of Time-Series Foundation Models for Remaining Useful Life Estimation
Description
Details
Context and Problem Statement
RUL models typically require complete degradation histories, yet well-maintained industrial equipment rarely reaches failure under monitored conditions. Time-series foundation models may provide transferable representations, but cross-machine generalization remains poorly understood.
Research Question
To what extent can frozen time-series foundation-model representations transfer across machine families, and how much local data is required to recover operationally useful prognostic performance?
Proposed Approach
Extract frozen foundation-model embeddings, train lightweight regression heads, compare with end-to-end recurrent and convolutional baselines, evaluate cross-machine transfer, and use uncertainty quantification to support maintenance-trigger decisions.
Expected Contribution
A transferability map across equipment families and an uncertainty-aware maintenance decision framework.
Expected Prototype
A prognostics module producing an RUL distribution and intervention recommendation from industrial sensor streams.
Datasets
C-MAPSS, FEMTO-ST / PRONOSTIA, and partner-collected vibration or current signals.
Challenges
Scarce run-to-failure data, heterogeneous sensors and sampling rates, edge inference cost, and long validation cycles.
Research Question
Innovation
Expected Deliverable
Technologies
Required Skills
- Signal Processing and Vibration Analysis
- Deep Learning and Pre-Trained Models
- Statistics and Uncertainty Quantification
- Industrial Instrumentation
Datasets
- C-MAPSS (NASA Prognostics Data Repository)
- FEMTO-ST / PRONOSTIA bearing dataset
- Vibration measurements collected from industrial equipment