Use cases · Anomaly detection · Plant & equipment
The failure your historian
already recorded.
Every plant historian holds years of sensor data — and inside it, every failure the site has ever had, plus the state-changes that preceded each one. Senua AI learns the plant’s normal operating states from its own stream and flags genuine departures — on-prem, on commodity CPU, with an explanation instead of an alarm flood.
The scenario
The bearing failure cost three days of throughput. The alarms fired 4,000 times that week.
A processing plant runs a data historian — a decade of pump vibration, bearing temperatures, pressures, flows and motor currents. When a mill bearing failed, the outage cost millions; afterwards, the review found the precursor signature had been in the data for days. Nobody saw it, because the alarm system cries wolf thousands of times a week and the operators long ago tuned it out.
The cloud predictive-maintenance pilot died in security review: operational data cannot leave the site, and the site’s edge boxes have no GPUs. The analysis has to come to the data.
How Senua AI handles it, end to end
The entry engagement is an audit of data you already own — no deployment risk, no data egress, findings first.
1 · Audit the history, blind
Export two years of historian data for one asset class. Senua AI learns the normal operating states and reports every genuine departure it finds — blind to your incident log. The test: does it rediscover the failures you already paid for, and how early?
2 · Watch the plant live
The same workflow then runs against the live stream, on your hardware: historian source, learned confidence gate, alert sink. One alert when a state genuinely departs — not four thousand threshold crossings.
3 · Explain every alert
Each alert names the learned operating state the asset left, the sensor channels that moved, and the confidence trajectory — an explanation a reliability engineer can defend to management and a regulator.
Why it’s different
Learned states, not tuned thresholds. On-prem, not in someone’s cloud.
No thresholds to tune.
The plant’s normal states are learned from its own history. There is no per-sensor limit to set, drift, and re-tune — the model of normal is the baseline.
Alerts are confidence-graded.
A departure carries a confidence, always available, never a binary trip-wire — which is what shrinks an alarm flood to a handful of signals worth an engineer’s time.
Data never leaves the site.
The engine runs on an existing edge box — commodity CPU, no GPU procurement, no cloud account, no egress. The security review is a one-pager.
The economics are visceral.
Unplanned downtime on a processing train runs into hundreds of thousands of dollars a day. One earlier warning a year pays for the system many times over.
The public proof — run it yourself
Validated on public run-to-failure data, in the open.
We validate this use case on public bearing run-to-failure datasets with known fault onsets — the NASA IMS bearing set and the SKAB industrial testbed benchmark — using a pre-registered protocol: learn normal states from healthy operation, replay the degradation blind, measure how early the state departure is flagged and the false-alarm rate against published baselines. The data is public; the replay is reproducible.
Your historian already knows.
Start with a blind audit of your own historical data — on your hardware, nothing leaves the site. If it doesn’t find what you already know is in there, you’ve lost nothing but our time.