Use cases · Anomaly detection · UAV telemetry
The flight that tells you
it’s failing.
A drone streams hundreds of telemetry values a second. Senua AI learns what normal flight looks like from the stream itself, then flags the moment the aircraft leaves it — on board, offline, with a reason. Built on the one capability every LLM provably lacks: tracking real state in real time.
The scenario
A survey fleet flies beyond the pilot’s line of sight.
A drone-services operator runs a fleet of survey and inspection aircraft. A single in-flight engine failure costs the airframe, the payload, the job — and under beyond-visual-line-of-sight rules, potentially the operating licence. Today’s telemetry alerting is threshold-based: it either floods operators with false alarms or stays silent until the aircraft is already unrecoverable.
Cloud AI monitoring is a non-starter: the fleet flies where there is no reliable link, and per-event cloud inference across a whole fleet is cost-prohibitive. The monitoring has to live on the aircraft — on a companion board with no GPU.
How Senua AI handles it, end to end
The same Console workflow every Senua deployment uses: a telemetry source, a learned gate, an alert — replayed live against real recorded flights.
1 · Learn normal flight
Senua AI reads the aircraft’s own MAVLink telemetry and induces the states of normal flight from the stream itself — no hand-written rules, no thresholds to tune, no labelled training set required.
2 · Watch the stream, on board
In flight, every record is scored against the learned states as it arrives. The engine is native machine code on ordinary CPU — a footprint a Pi-class companion board carries, fully offline.
3 · Flag the departure, with a reason
When the aircraft leaves its learned states, the alert names the state it left, the channels that deviated, and the confidence — a reasoning path an operator can act on and a regulator can audit, not a score from a black box.
Why an LLM cannot do this job
Following state in real time IS the task.
A provable ceiling, not an opinion.
Tracking the evolving state of a machine through a long event stream is a problem class transformers are mathematically limited on — a published complexity-theoretic separation, not a benchmark quibble. Senua AI’s causal-state engine tracks it exactly, with constant memory.
Telemetry-rate economics.
Hundreds of records a second, per aircraft, per fleet. Per-event cloud inference is priced out before it starts; Senua AI runs on the hardware that is already flying.
No link required.
The aircraft flies where the cloud isn’t. Detection happens on board; nothing depends on a round trip.
Alerts you can defend.
Confidence-graded departures instead of threshold floods — and every alert carries its explanation, which is what a BVLOS regulator asks for.
The public proof — run it yourself
Validated against real recorded failures, in the open.
We validate this use case on the ALFA dataset (Carnegie Mellon AirLab): 47 real autonomous fixed-wing flights over MAVLink, including 23 sudden engine failures and 24 control-surface faults — each with the ground-truth fault type and exact onset time. The protocol is pre-registered: learn normal states from the safe flights, replay each fault flight blind at recorded wall-clock rate, measure detection latency against the true onset and false alarms across 66 minutes of normal flight. Don’t take our word for it — the dataset is public and the replay is reproducible.
Your fleet already records the evidence.
Start with a historical audit: we replay your recorded flight logs blind and show you what the telemetry already knew. Then put it on the aircraft.