Remaining Useful Life Estimation

Estimating how long a component will keep working, with honest uncertainty, so maintenance lands before failure not after.

RUL as a distribution

Remaining useful life (RUL) is the time a component can keep operating before it crosses a failure threshold. The right output is not a single number but a distribution, because acting on a point estimate that is optimistically wrong causes exactly the unplanned failure predictive maintenance exists to avoid. Decisions use a conservative lower quantile of the RUL distribution.

Two model families

• Physics-of-failure: a degradation model (fatigue, creep, irradiation damage) integrated forward from measured stressors
• Data-driven: a learned model mapping health features to time-to-threshold from run-to-failure examples
• Hybrid: a physics model whose parameters are updated online from telemetry - preferred pre-FOAK where run-to-failure data is scarce

import numpy as np
def rul_quantile(health, rate_samples, threshold, q=0.1):
    # Monte-Carlo forward: many plausible degradation rates -> RUL samples
    remaining = (threshold - health) / np.maximum(rate_samples, 1e-9)
    return float(np.quantile(remaining, q))   # conservative lower bound

# schedule against the 10th-percentile RUL, not the mean

Honest uncertainty

For the breeder magnets and first wall, and especially for any burner component in the 166-830x extrapolated plug regime, the degradation models are poorly constrained by data. We propagate that ignorance into wide RUL distributions and act on the conservative tail rather than pretending to precision. A wide, honest interval is more useful than a narrow, wrong one.

RUL estimates drive the maintenance schedule and feed spares planning. Every estimate is reconciled against teardown findings so the models improve across the FOAK-NOAK-BOAK sequence and propagate through the fleet.