We spent the early part of the project trying to build one model that would catch every kind of charger failure. It didn't work, and why it didn't work says something useful about the shape of the problem.
Two failure modes, two signatures
Gradual degradation and sudden hardware failure aren't two ends of one spectrum. They're different events. They show up differently in the data, play out over different timescales, and need different inputs.
Gradual degradation is visible in OCPP. Power delivery efficiency trends down over days. Error codes get more frequent. Sessions complete less often. The signal accumulates in the stream, and by the time the charger actually quits, the evidence has usually been sitting there for weeks.
Sudden hardware failure mostly isn't visible in OCPP. A connector relay burns out, a cooling fan seizes, a power electronics component goes with no precursor at all. The charger reports normally right up until it doesn't. Catching those means having data the protocol doesn't carry: housing temperature, vibration, component-level electrical telemetry. Different problem, different sensors.
Why one model didn't work
Trained on both classes together, the model learned degradation well and quietly gave up on sudden failures. That's a rational thing for it to do, since a sudden failure with no OCPP precursor offers nothing to learn from. The trouble was that the aggregate metrics looked fine. Strong performance on the common class covered for the fact that the model was useless on the hard one.
So we split it:
-
Gradual degradation — an LSTM sequence model plus a gradient-boosted classifier, running on normalized OCPP streams. Deployed in the pilot, and it earns its keep.
-
Sudden failure — a research track. We're working with hardware partners on what telemetry is actually available, what it costs, and whether it's enough to build a classifier worth deploying. Nothing in production.
What operators get
Real predictions for one class of failure, and a straight answer about the other.
If we flag CHG-042 for inspection, that came from a model trained on actual degradation signatures, and you can act on it. If we don't flag a charger, that is not a promise it won't fail tomorrow. It means we see no degradation trend in its OCPP stream. Those are different statements, and we'd rather operators know which one they're getting.
Where the research track stands
Sudden-failure prediction needs hardware integration, so we're talking with telemetry providers about access to the sensor data. The fusion architecture is designed: OCPP stream features combined with time-series sensor data through a multimodal embedding that tolerates the two sources updating at different rates.
That's a design, not a result. The degradation model works today. The other half is genuinely unfinished, and we'll keep saying so until it isn't.