Fleet-Wide Predictive Maintenance

Remaining-useful-life models trained across the whole fleet predict component degradation earlier and more sharply than any single unit's history allows.

Remaining useful life, learned across units

Predictive maintenance estimates each component's remaining useful life (RUL) from its condition signals. A single early unit has little failure history, so its RUL estimates are wide. L7 trains RUL models across every unit in the fleet, so a degradation signature seen on one breeder informs the RUL estimate for the same component on every other breeder, and likewise across the burner fleet.

From signal to RUL

# RUL as time until a health index crosses a limit
# health index h(t) from twin (strain, resistance, thermal margin ...)
def rul(component, now):
    traj = degradation_model(component).forecast(h_history(component))
    for t in future_times(now):
        if traj.mean(t) - k*traj.std(t) <= h_limit(component):
            return t - now                 # conservative RUL (lower band)
    return horizon                          # beyond horizon: no action
# model is shared across fleet units of the same component type

The estimate is deliberately conservative: RUL is taken where the lower confidence band, not the mean, crosses the limit, so maintenance is triggered before the expected failure, with margin for model uncertainty. As the fleet accumulates hours the band tightens, allowing later, better-timed intervention without raising risk.

Component types map to specific physics. Breeder RUL focuses on REBCO magnet strain and quench-precursor trends toward the 16.84 T peak field, first-wall and divertor erosion under 14 MeV neutron and heat load, and blanket integrity. Burner RUL focuses on the 26.49 T plug and 17 T throat coils, direct-energy-conversion electrode wear, and neutron-activated structures.

RUL outputs feed the maintenance loop and the firm-power forecast: a shortening RUL both schedules an intervention and de-rates the availability a unit can firmly commit. This keeps grid commitments consistent with the true condition of the machines.

Fleet predictive maintenance runs today on twin-simulated degradation and analogous historical data. The models are structured for transfer so that, as real units come online from ~2030, early-unit experience immediately benefits later NOAK and BOAK units.