Charger components rarely fail instantly. Contacts erode, connections loosen, seals degrade and modules weaken, and each of those produces a measurable change before it produces a fault.
Predictive maintenance is the practice of noticing. It requires recording data nobody currently needs, which is why it is usually implemented after an expensive failure rather than before.
Signals that degrade before failure
None of these is a fault on its own. Each is a trend that becomes one, and the value is entirely in having history to compare against.
- Contactor response time, which lengthens as contacts erode.
- Internal temperature at a given load, which rises as connections loosen.
- Connectivity quality, which falls as antennas or modules degrade.
- Session failure rate for one unit, which climbs before an outright fault.
- Earth continuity or insulation values, where they are measured periodically.
- Time to establish a session, which lengthens as latching wears.
Trend against the unit's own history
Fixed thresholds are set for the worst acceptable case and therefore trigger only once a unit is nearly failed. A unit whose response time has doubled is degrading even if it remains within specification.
Comparing each unit against its own baseline detects that. It also handles the variation between units that makes fleet-wide thresholds either too loose or too noisy.
Environment explains a lot of variation
Temperature rises with ambient and with load, so a rising trend during summer may be seasonal rather than degradation. Comparing against the same period last year, or against nearby units, separates the two.
Without that normalisation, predictive systems generate alerts every summer and are switched off by their second one.
What it is worth
The value is converting an unplanned failure, with its site visit and lost revenue, into planned work done alongside something else. On a dispersed network where travel dominates service cost, that is substantial.
On a dense urban network where an engineer is nearby anyway, it is worth considerably less, and honest assessment of which situation applies should precede investment.
Start by recording, not by predicting
Prediction requires history, and history cannot be created retrospectively. The first step is capturing the signals above, at a sensible interval, and retaining them long enough for trends to be visible.
A year of data makes prediction possible. No amount of modelling substitutes for it, which is why the useful first action is instrumentation rather than analytics.
Act on it or do not build it
A prediction nobody schedules work against changes nothing. Predictive maintenance requires a maintenance function able to act on a signal that is not yet a failure, which is an organisational capability as much as a technical one.
Networks without that will get more information and the same outcomes.