Proposed Validation Protocol

From 574 available SCADA channels, TorqueScope identifies the 15–20 signals that map to its standard sensor roles: power output, wind speed, nacelle ambient temperature, main bearing temperatures (front and rear), gearbox oil temperature, generator winding temperatures, converter temperatures, and hydraulic system temperature.

This mapping requires no bespoke code. TorqueScope's signal mapper ingests the channel description file and applies keyword matching against its internal sensor taxonomy — the same process used for Kelmarsh, Penmanshiel, and Hill of Towie. The 7 MW Samsung drivetrain uses a different bearing and gearbox configuration from the 2 MW Senvion fleet, so this step also validates that the sensor taxonomy generalises across turbine classes.

Using a 12-month rolling window from a period with no logged alarm events, TorqueScope trains its Normal Behaviour Model (NBM): a 200-bin operational grid (20 power bins × 10 ambient temperature bins) storing median and standard deviation of each temperature sensor under normal conditions.

The met mast data adds a significant advantage over the onshore validation datasets: dedicated multi-height wind speed measurements (rather than relying on the nacelle anemometer) produce a cleaner power-wind curve and more accurate operational binning. The standard deviation metric within each bin is computed from the 10-min stdev aggregate field directly — no re-processing is needed.

TorqueScope computes Lomb–Scargle periodograms on each mapped temperature sensor over the 12-month training window. Dominant periodic components are identified (expected: ~24 h diurnal thermal cycle, ~annual seasonal envelope) and the coefficient of variation (CV) is measured across 4 temporal sub-windows. Sensors with CV < 5 % qualify as having a stable periodic baseline that feeds the heuristic scoring stage.

The offshore environment introduces periodic signals absent from onshore datasets: tidal loading on the foundation (approximately 12.4 h M2 tidal period) may be detectable in structural vibration sensors, and sea-state seasonality will modulate thermal gradients differently from land. This makes Levenmouth the first dataset where the Ab Astris cross-domain methodology — validated independently on oceanographic tidal data and wind turbine SCADA — may find its two domains simultaneously present in a single asset.

The trained v5 hybrid pipeline runs forward across the full dataset from Jan 2017 to present: heuristic scoring (LS periodic analysis) + NBM residual scoring + criticality counter (alarm threshold: 72). Every timestamp where the criticality counter reaches threshold is logged as a detection event, recording the timestamp, the dominant sensor signal, and the hybrid score trajectory leading up to alarm.

The 1 Hz data is used selectively for validation rather than routine processing — specifically to examine signal morphology in the 48–72 hour window before each alarm, checking whether faster-timescale precursors are visible that the 10-min aggregates might smooth over. This is an exploratory analysis rather than a core pipeline step.

TorqueScope cross-references its detection events against the ORE Catapult alarm log. For each logged alarm event, the following metrics are computed:

Lead time — hours between TorqueScope's first criticality threshold crossing and the alarm log entry.
Detection rate — fraction of logged alarm events that had a TorqueScope detection within the preceding 72 hours.
False positive rate — detection events with no corresponding alarm log entry within ±24 hours.
Fault type breakdown — detection performance segmented by alarm category (electrical, mechanical, thermal, control system).

This produces the first ground-truth earliness measurement on real offshore SCADA data outside a competition benchmark. The result is directly comparable to TorqueScope's CARE benchmark performance (current CARE Earliness score: 0.636) but on a named, independently attributable asset.

One acknowledged limitation: as a single turbine, the alarm event count will be small, giving the detection rate and false positive rate lower statistical power than a fleet-scale evaluation. The per-fault-category breakdown mitigates this by grounding each result in a specific, reproducible event.