Predictive maintenance in FTTH: see the curve before the outage

Almost no fiber trouble ticket comes out of nowhere. A connector that gets dirty, a fiber poorly seated on a pole, a splice that absorbs moisture: the power the ONU receives drops little by little, over weeks, until one day it crosses the threshold and the subscriber calls. That curve was in the data the whole time. Predictive maintenance is looking at it sooner.

What's worth forecasting

Not everything you measure deserves a forecast. The useful series are the ones that degrade slowly and have a clear limit:

• Each ONU's receive power, against its threshold (for example, −27 dBm).
• The laser bias current of each PON port, as a percentage of the module's maximum: a laser that needs more and more current to emit the same output is aging.
• The utilization of each uplink and PON port, better as a daily 95th percentile than as an average, against a reasonable ceiling (for example, 80% of its speed).
• The maximum temperature of each OLT, which flags problems with fans or equipment rooms.

The crossing date, and how much confidence it has

A useful forecast isn't a line on a chart: it's a concrete answer. "This ONU would cross its threshold in 23 days." With that, you schedule the visit during normal hours, before the complaint, instead of rushing out on a Sunday.

But not every series is equally predictable. One that declines steadily gives a reliable date; one that jumps all over the place doesn't. That's why every forecast has to come with its confidence level, and when confidence is low, the honest thing is to show it as uncertain instead of making up a date.

Sudden changes matter too

Some degradations don't have time to form a curve: an ONU that lost 3 dB overnight is still above the threshold, but something happened to it. Comparing each ONU against its own history over the last seven days catches those jumps even before any alarm fires.

How much data you need

A forecast needs history. With less than a week of data, any trend is noise. A newly connected network will show "no data" for the first few weeks, and that's correct: it's better not to forecast than to forecast badly.

Forecasting isn't alerting

A forecast is recalculated every few hours and is meant for planning: building the week's route, ordering spare parts, prioritizing areas. It's not how you find out something went down. That's what real-time alerts are for. The two complement each other: the alert handles what's urgent and the forecast reduces how many emergencies there are.

How eCloud OLT Controller does it

The dashboard projects those four series 90 days out, with the crossing date and its confidence, and notifies you when an ONU drops 3 dB below its own seven-day history. It needs at least seven days of data and is recalculated every six hours.

You can see how it works in the forecasts section of the platform page.