How it works
How a prediction is made, checked and recorded.
For lab, quality and data teams who need to see the working. Ten stages, from logger file to audit trail. One of them is a rule, not a model.
- 01
Logger data
Readings come in from the data logger the lab already uses.
In the buildApproved logger export by upload. Direct connection only once the lab confirms the interface, security and network route.
- 02
Session identity
Each test is tied to one unit, one model and one intended band.
In the buildUnit, model, rated values used by the formula, intended band, revision, lab and test station.
- 03
Quality gate
Bad data stops here, before any prediction.
In the buildTimestamps, channel completeness, sensor error values, voltage, ambient conditions, missing data.
- 04
Test state
It works out which phase the test is in.
In the buildTest phases, cycles or load points completed, and whether the product has stabilised.
- 05
Features
It turns the trace into measurable behaviour.
In the buildFor example temperature, power, cumulative energy, duty cycle, cycle timing, flows, voltage, ambient and humidity, chosen per category.
- 06
Prediction with a range
A model for that product family predicts the final energy result.
In the buildOne validated model per product family, with a calibrated prediction range.
- 07
Coverage check
It checks whether it should predict at all.
In the buildValidated design? Enough evidence so far? Range narrow enough? If not, no read.
- 08
Rating rules
The band comes from the regulation's formula, not from a model.
In the buildYour standard's published formula, the approved rated values and the rule version in force.
- 09
Early read and audit
The engineer sees one of four reads, with the reasons.
In the buildTarget match, Outside target, Continue testing or Engineer review. Every input and rule version is logged.
- 10
Learning loop
Each certified result is compared with what was predicted.
In the buildMonitoring of prediction against final results, and governed retraining with sign-off.
When is it ready to predict?
No fixed clock time is assumed.
Readiness differs by product and test. Several points in the run are tested, and the earliest one that holds up on held-out units is kept.
Time checkpoints, for example in a refrigeration test
Event checkpoints, set per category
The target is the earliest point where the predicted range agrees with the final certified result often enough to be useful. That point is found in your data, not promised in advance.
How it is measured
Reported on units the model has never seen.
- Prediction error on the final energy result, measured on whole held-out units.
- Range coverage: how often the final result actually falls inside the predicted range, and how wide the range is.
- False target matches: cases where it predicted on target and the full test disagreed. This is the number that matters most.
- Early-read coverage: the share of tests that get a confident early read at all.
- Time to a reliable read, as a median and 90th percentile.
- Correct holding back on unvalidated models, poor data or wide ranges.
Data it needs
Mostly what the lab already has.
- Exports from your existing data logger, in the format your lab already approves.
- Past completed tests with their final certified results.
- For each model: the rated values your standard's formula uses, the intended band, revision and key components.
- Unit identity, lab and test station for every session.
Data it will not use
Excluded on purpose, and logged as excluded.
- Corrupt or unreadable sessions, and duplicate copies of the same test.
- Sessions with an unknown model or unit, or a missing final result.
- Sensor error values treated as real readings.
- Sessions with unresolved timestamp resets or major gaps.
- Old and new engineering revisions pooled together without versioning.
- Readings from one unit split across training and testing. Units are held out whole.
Lab assessment
Find out if your test data can predict your ratings.
We look at your logger exports and completed tests, and tell you plainly whether Energy Mark can work in your lab.