Setting Thermal Anomaly Detection Thresholds: ΔT and Dwell
A trip condition has three axes — magnitude, persistence, and spatial coherence. Get the first two wrong and the third can't save you.
A thermal anomaly trip condition is not a single number. It is three independent axes evaluated together: magnitude (how far above baseline), persistence (how long the rise holds), and spatial coherence (whether neighbours share it). Coherence gets the attention, but if the magnitude and dwell thresholds are wrong, the coherence layer is suppressing or escalating the wrong candidates in the first place.
Magnitude: delta, not absolute
The magnitude axis is a delta above a per-vehicle rolling baseline, never an absolute temperature. Absolute thresholds fail on a cargo deck because engine bays, ambient conditions, and solar gain all shift the floor. A relative trip of roughly 6 °C above a vehicle's own EWMA baseline is the working figure — large enough to clear sensor noise and normal post-loading cooling, small enough to catch an early cell fault well before flame.
Dwell: rejecting the transient
The persistence axis requires the delta to hold for a minimum dwell time before it counts. A real cell fault heats and stays hot; a reflection, a passing exhaust plume, or a sensor glitch spikes and decays. Setting dwell too short re-admits exactly the transients the magnitude threshold was meant to exclude; too long and you spend hard-won lead time waiting for confirmation. The dwell is a per-deck parameter, swept against replayed calibration voyages like the coherence window.
How the three axes combine
A candidate is raised only when the magnitude and dwell conditions are both met. The spatial coherence check then decides escalate-or-discard: a rise shared across neighbouring cells is environmental and suppressed; a rise confined to one vehicle escalates to a red-state alert. Magnitude and dwell control how many candidates reach the coherence stage; coherence controls how many candidates survive it. Tuning one axis in isolation moves the false-positive and lead-time numbers in ways the others have to absorb.
Tuning in the field
Thresholds are not shipped as universal constants. Per deck, calibration-voyage data is replayed against the trip engine with magnitude and dwell swept across their ranges; the optimum minimises false positives without losing any staged true-positive replay. The procedure is documented and reproducible so a class auditor can re-run it, and it should be re-tuned after major structural or ventilation changes to the deck.
What it means for class and underwriters
For a classification-society reviewer, the value of a three-axis trip condition is that each axis is separately inspectable and the tuning method is reproducible — no black box on the alarm path. For underwriters, documented threshold tuning is what makes a low false-alarm claim defensible rather than marketing, and it ties directly to the fault-to-response interval that drives total-loss outcomes like Felicity Ace (2022) and Fremantle Highway (2023).
Sources
- 1. RoRoSafe bench testing — per-vehicle baseline magnitude/dwell vs deck-zone smoke detection (internal, [VERIFY: 6 °C trip, 60–120 s dwell, 18–25 min lead])
- 2. Felicity Ace (2022) and Fremantle Highway (2023) casualty context — commercial.allianz.com, gcaptain.com, maritime-executive.com
- 3. Companion method detail — see RoRoSafe 'Tuning Coherence Windows to Kill False Positives'
Questions, answered
What are the thresholds in thermal anomaly detection?+
Three axes evaluated together: magnitude (a delta above a per-vehicle baseline, around 6 °C), persistence (a minimum dwell time the rise must hold), and spatial coherence (whether neighbouring vehicles share the rise). A candidate is raised when magnitude and dwell are both met; coherence then decides whether it escalates to an alarm or is discarded as environmental.
Why use a delta instead of an absolute temperature limit?+
Because absolute limits fail on a cargo deck. Engine bays cool at different rates, ambient conditions vary by voyage, and solar gain through deck openings shifts the floor. A delta above each vehicle's own rolling baseline stays valid across those conditions, so the same threshold works in a cold North Atlantic crossing and a hot tropical sailing.
What is dwell time and why does it matter?+
Dwell is the minimum duration a temperature delta must persist before it counts as a candidate. A real cell fault heats and stays hot; reflections, exhaust plumes, and sensor glitches spike and decay. Too short a dwell re-admits those transients; too long wastes lead time. It is tuned per deck against replayed calibration voyages, typically in the 60–120 second range.
Are these thresholds AI or rule-based?+
Rule-based, deliberately. The trip condition for a red-state alert must be inspectable, auditable, and presentable to a classification society — a neural network is not. Machine-learned models assist with false-positive suppression and per-deck baselining, but the magnitude and dwell thresholds on the alarm path stay explicit and documented.
Continue the thread
What Is Thermal Anomaly Detection on a Car Deck?
Thermal anomaly detection flags a vehicle heating abnormally against its own baseline — minutes before smoke, the trigger for most RoRo fire alarms.
Tuning Coherence Windows to Kill False Positives
Cross-cell coherence is the suppression layer. The window length is the lever. Two seconds too short and solar gain trips the deck; two seconds too long and you spend lead time on noise.
Per-Vehicle Thermal Baselines and 6 °C Anomaly Detection
A deck-wide threshold misses the early-stage signature. A per-vehicle rolling baseline catches it 18–25 minutes earlier in our bench tests.