Tuning Coherence Windows to Kill False Positives
Cross-cell coherence suppresses false alarms; the window length is the lever. Too short and solar gain trips the deck; too long and you lose lead time.
A per-cell rolling baseline gives you a candidate trip. Cross-cell coherence is what turns the candidate into an alarm or a discard. The mechanic is simple — if the same delta is shared by neighbours, suppress; if it is local, escalate — but the window length over which "shared" is evaluated is where the engineering happens.
The trade space
- Short window (sub-second): coherence is too tight; transient cross-cell environmental events look like vehicular ones.
- Long window (10+ seconds): real cell-local events get suppressed because the environmental delta catches up.
- Variable window: per-deck tuning to the ventilation and solar profile of the specific vessel.
Where we landed
How we tune in the field
Three voyages of calibration data per deck, replayed against the trip engine, with the window swept across the ±1 s range. The optimal window minimises false positives without losing any of the staged true-positive replays. Process is documented and reproducible — a class-society auditor can run it themselves.
Questions, answered
What is a coherence window in the detection system?+
It is the time interval over which the system evaluates whether a temperature delta is shared across neighbouring cells. If the same delta is shared by neighbours it is suppressed as environmental; if it stays local it escalates. A per-cell rolling baseline produces the candidate trip; the coherence window turns it into an alarm or a discard.
Why is the window length so sensitive?+
Too short (sub-second) and transient cross-cell environmental events look vehicular; too long (10+ seconds) and real cell-local events get suppressed as the environmental delta catches up. The default lands at about 3 seconds with a ±1-second per-vessel tuning range, giving roughly 0.04% residual false positives against the bench-rig catalogue.
How is the window tuned for a specific vessel?+
Three voyages of calibration data per deck are replayed against the trip engine with the window swept across the ±1-second range. The optimum minimises false positives without losing any staged true-positive replays. The process is documented and reproducible, so a class auditor can run it themselves, and it should be re-tuned after major deck changes.
Continue the thread
How Are Thermal Hotspots Detected?
The detection problem on a cargo deck is not measurement — IR sensors are commodity. It is deciding which delta in which cell at which time is real.
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.
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.