Multi-Modal Fusion: Heat, Gas, and Smoke
Each sensor modality has different blind spots. Multi-modal fusion is not a buzzword — it is the only architecture that survives the marine failure modes.
A single-modality detector has a single failure mode that can take the layer offline. A thermal grid alone is degraded by occlusion. A gas mesh alone is degraded by ventilation. A camera alone is degraded by smoke obscuration. Fusion is not a marketing claim — it is the answer to the question "what happens when one layer is wrong?"
What fusion looks like
Each modality runs independently with its own anomaly detector. A fusion stage aggregates the signals with weights that reflect each sensor's confidence at that moment, and produces a single trust state — green, amber, red. The fusion logic is rule-based, not learned, because the decision has to be auditable.
Why rules, not models, for the trust state
- Auditable behaviour under class-society scrutiny.
- Predictable response when one sensor is in known failure (e.g. occluded camera).
- Inspectable test methodology against fixed event catalogue.
The weighting scheme
- Per-vehicle thermal grid: highest weight when ambient is stable, lower under solar gain.
- Gas sensing: highest weight in sealed compartments and low-ventilation states.
- Video analytics: highest weight on access lanes and ramps.
- Fiber-optic LHD: highest weight on weather decks and open spaces.
How RoRoSAFE helps
Fusion is how RoRoSAFE decides when to alarm. Each vehicle's thermal and vent-gas channels are combined under deterministic rules, so a blind spot in one modality is covered by the other and a nuisance on one channel does not trip the bridge. It works alongside the ship's smoke detection, which remains the class-mandated layer.
Pilot: one deck · installed alongside the berth · no drydock · 6 months of dashboard access
Questions, answered
Why is multi-modal sensor fusion necessary rather than a single detector?+
Because a single-modality detector has a single failure mode that can take the whole layer offline: a thermal grid is degraded by occlusion, a gas mesh by ventilation, a camera by smoke obscuration. Fusion is the answer to what happens when one layer is wrong — each modality covers what the others miss.
How does the fusion stage produce a single trust state?+
Each modality runs its own independent anomaly detector, then a fusion stage aggregates the signals with weights reflecting each sensor's confidence at that moment and outputs a single green/amber/red trust state. The fusion logic is rule-based, not learned, so the decision stays auditable.
Why use rules rather than a model for the fused trust state?+
For auditable behaviour under class-society scrutiny, predictable response when one sensor is in known failure (such as an occluded camera), and an inspectable test methodology against a fixed event catalogue. A monolithic AI detection box is the wrong answer; a small set of inspectable rules is the right one.
Continue the thread
The Four Stages of Li-Ion Thermal Runaway
Li-ion runaway moves from SEI breakdown near 80–120 °C through venting to fire. Useful detection works in the venting window, before smoke and flame.
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.
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.