TruckVitals: detect the fault, not the ramp

Condition monitoring on a haul truck has a problem that fixed machinery does not: the truck changes what it is doing constantly. Payload, road grade and speed all move, and every monitored channel moves with them. Run a change detector straight on that telemetry and the biggest changes it finds are the truck loading, climbing and tipping.

TruckVitals puts the segmentation first: raw telemetry, then regimes, then the within-regime residual, then the change point, then the onset. The regimes are learned from context channels (payload, grade, speed) that are disjoint from the monitored ones, and that disjointness is the whole point, because a segmentation trained on the monitored channel would absorb the very fault it exists to expose.

What it does not claim is on the page too. Regime conditioning does not improve onset localisation, and that null is published (paired over five seeds: -0.08 ± 0.74 in chance-corrected skill, ahead in 2 of 5). No false-alarm-reduction claim is made anywhere, because on the synthetic lane the raw arm wins that metric outright. And the missing lane is named rather than papered over: there is no public, redistributable dataset of continuous haul-truck telemetry with labelled faults, so onset error is measured on a synthetic fleet and the app says so.

The headline reads the way it does because I commissioned a review to refute it. It confirmed the mechanism through eight attacks and broke three of the four numbers I had, and every one of those three had flattered the result. That is a separate story, and a better one than the numbers it replaced. Engine is the separately published regimecpd. Live · source.