Emission calibration

Emissions results come out of a cycle, but the explanation is spread across the recording: how long the catalyst took to light off, how much enrichment the start needed, where the engine actually spent its time, and which transients pushed lambda away from the target. Answering that quickly is a measurement-tool question.

A cold start is the clearest example. Battery voltage collapses while the starter turns, enrichment drives lambda rich, ignition advance is pulled back hard to dump heat into the exhaust, and the catalyst walks up while the coolant is still cold. Every one of those is a channel you can put a cursor on, and the difference between two soak temperatures is visible strip by strip.

Where the effort goes is also a cycle-structure question. Because consumption and emissions are reported per kilometre, a phase that takes a long time to cover a short distance carries more reporting weight per kilometre than a fast one — and on a modern downsized engine that slow phase is also the furthest from its best efficiency. The levers that remain are usually outside the combustion envelope altogether: fuel burned at standstill, fuel burned while the driver is asking for no torque, enrichment delivered during transients, start fuelling, and the mechanical work taken by auxiliaries while the engine is fuelled.

The gate that catches the most errors is the one applied before a number is quoted at all. Two runs are only comparable if they were driven in the same thermal state, measured through the same quantification chain, corrected for any difference in electrical charge balance, and executed closely enough against the prescribed trace that the driver is not the variable under test.

Each of those is a real failure mode, not a formality. Comparing a bag result against a controller-integrated fuel figure mixes two chains that can sit several percent apart. A run that ended with a fuller battery than it started with has moved mechanical work around rather than saved it. And a comparison read at the wrong point in a manoeuvre reports an alignment artefact as an effect.

Sweeping a full grid on the dyno spends hours measuring regions the engine will never run. A designed experiment covers the space with far fewer points, a response surface is fitted to what was measured, and the optimiser is asked for the best point that still respects the constraints — knock limit, exhaust temperature, torque demand.

The value of the method is in where the next measurement goes: each point is chosen because the model says it is worth measuring. The result exports as A2L, CDF or HEX and carries into the calibration image with its provenance intact.

An emissions dataset is reviewed long after the person who changed it has moved on. HELIX treats a change as a package: source A2L and base dataset, the approved changes, validation status and a frozen release state, so the question "why is this value what it is?" has an answer that survives the programme.

For teams that cannot let prototype data leave the building, the desktop instruments run fully offline and the agent-based tools deploy on-premise.