NIU Marine · Artificial intelligence · 2025
Predicting maintenance beforethe vessel tells you it is late
18d
average early warning before a failure
31%
fewer unplanned dry-dock days
100%
of the scheduling view works offline
6
vessel classes covered by one model
01 — The challenge
The telemetry was already there. Nobody could act on it.
Sensors had been streaming for three years into a store nobody queried. Maintenance still ran on a fixed calendar, which meant servicing vessels that were fine and missing ones that were not.
Any tool had to work at the dock, where connectivity is unreliable and nobody is going to wait for a page to load.
Must work fully offline at the dock
Crews would not adopt a second login
Predictions had to be explainable to engineers
02 — What we did
Earn the crew's trust before automating anything.
Backtest against three years of history
Before anyone saw a prediction, we showed the model would have caught the failures that actually happened — and admitted the ones it would have missed.
Offline-first, not offline-tolerant
The scheduling view is built to work with no signal and reconcile later. That is the difference between a tool crews open and one they do not.
Explain every flag
Each prediction shows the readings behind it, so an engineer can overrule it with a reason the model learns from.
Ship to one class, then widen
One vessel class ran the whole season before we extended to six. Nothing about the rollout required a leap of faith.
Got a bottleneck you can't quite name yet?
That's usually the best time to talk. Thirty minutes, no pitch deck.