Where AI Anomaly Detection Beats Rules
We use both. The interesting question is which decisions belong to which approach. The split is not where most marketing decks would put it.
The "AI-powered" claim is doing a lot of work in our space. The honest answer is that we use machine-learned models for some decisions and explicit rules for others, and we are quite particular about which is which.
Where rules win
- Trip conditions for a red-state alert. These have to be inspectable, auditable, and class-society-presentable. A neural network is not.
- Network health (cable break detection, node liveness). Deterministic, not probabilistic.
- Calibration drift compensation. Linear, well-understood physics.
Where models help
- False-positive suppression. The space of nuisance signatures is too varied for explicit rules.
- Per-deck baselining. Different ships, different cargo mixes, different baselines.
- Anomaly clustering for post-voyage analysis.
How RoRoSAFE helps
RoRoSAFE uses that split in practice. Models help build per-vehicle baselines and spot patterns across the fleet, while the decision to alarm stays with deterministic, reviewable rules that class and crews can audit. For an operator, that means earlier warnings of a developing battery event without an alarm path that nobody can explain afterwards.
Pilot: one deck · installed alongside the berth · no drydock · 6 months of dashboard access
Questions, answered
Does RoRoSAFE use AI or rule-based detection?+
Both — and the split matters. Machine-learned models handle some decisions and explicit rules handle others, chosen deliberately rather than by marketing. The honest answer is that AI-powered is doing a lot of unearned work in this space.
Which decisions stay rule-based and why?+
The red-state alert trip conditions, which must be inspectable, auditable, and class-society-presentable (a neural network is not); network health such as cable-break detection and node liveness, which is deterministic; and calibration-drift compensation, which is linear, well-understood physics.
Where do machine-learned models actually help?+
False-positive suppression, where the space of nuisance signatures is too varied for explicit rules; per-deck baselining across different ships and cargo mixes; and anomaly clustering for post-voyage analysis. The reverse split — models for alarm trips, rules for suppression — produces systems that are both noisy and opaque.
Continue the thread
Building Per-Vehicle Thermal Baselines
Anomaly detection is only as good as its baseline. How a per-vehicle rolling reference is built, warmed up, and held steady across a voyage.
Fire Detection vs Video Monitoring on Ro-Ro Ships
Fire detection raises the alarm; video monitoring shows what raised it. SOLAS 2026 asks new ro-pax ships for both, and neither can stand in for the other.
How Are Thermal Hotspots Detected?
Measuring temperature on a cargo deck is easy; deciding which rise is real is not. Relative thresholds and neighbour checks separate faults from noise.