A fruit fly's novelty detector on wind turbine SCADA
The fly decides whether a smell is new in two steps: about 150 projection neurons (PNs) fan out to about 2,000 Kenyon cells (KCs), only the 5 % most excited KCs stay active (a sparse "tag"), and the synapses from the active KCs to one output neuron weaken each time the smell comes back. A tag full of unweakened synapses means "never smelled this". Here the smell is the 10-minute state of a wind turbine, the wiring is the real right mushroom body from the FlyWire connectome, and the question is whether that finds faults early, at what cost, and whether the real wiring beats a random one.
The expansion: 149 PNs to 2,364 Kenyon cells
Share of KCs of each block (columns, 20 KCs each, grouped by KC class) that receive input from each PN (rows, grouped by glomerulus). Real FlyWire wiring, the same after degree-preserving shuffling, and FlyHash (every KC samples 4 PNs at random).
Alarms on real faults
When each method is in alarm over the prediction section (red: share of the hour's producing rows flagged; every method with its own threshold from the calibration rows). Dashed: the labelled window leading up to the fault (anomaly datasets).
Live replay
Methods on CARE to Compare (95 datasets, 45 faults)
CARE score (Gück et al. 2024; 0.5 is what a random detector reaches, and a method that detects no fault at all scores 0). Fly variants: mean of three seeds, the dots are the seeds. Same thresholds and smoothing for every method. Skill: share of the faults detected minus share of the normal datasets flagged.
Synthetic faults in the training data (label-free)
Share of injected faults detected (12 h of net alarms), 95 datasets x 2 component temperatures, and the false alarms on the clean data.
Operation
Cost per method on one dataset (wind farm B, 34 signals). The fly detector's learned state is one weight per KC.
Share of trials (95 datasets) with an alarm (12 h net) in the six weeks after nothing changes (control), after a change of normal (+5 K step in one temperature, e.g. after service) and during a slow fault (drift to +10 K), for four learning policies.
Wiring: FlyWire Consortium, connectome v783, doi:10.5281/zenodo.10676866, and cell annotations of Schlegel et al. 2024, CC BY 4.0. Faults and SCADA: CARE to Compare, C. Gück and C. Roelofs (Fraunhofer IEE), doi:10.5281/zenodo.15846963, CC BY-SA 4.0; the scores and results shown on this page are derived from it and are CC BY-SA 4.0 as well. Method: Dasgupta, Stevens, Navlakha 2017 (Science) and Dasgupta, Sheehan, Stevens, Navlakha 2018 (PNAS). Code (MIT): fly-novelty-scada.