Home/News & Publications/Guide

Predictive Maintenance Without Historical Data: Why You Don't Need to Collect for 12 Months

The most common project killer in predictive maintenance is one sentence: 'First we need 6 to 12 months of data.' It isn't true as stated. What works from week one without any history — and what honestly comes later.

Predictive MaintenanceData CollectionDrift DetectionGetting Started

The Sentence That Ends Projects Before They Start

Many predictive maintenance initiatives don't die from budget or technology. They die from one sentence in the first vendor meeting: "Before we can detect anything, we need to collect six to twelve months of data."

That means: a year of cost without value, a year of project momentum evaporating, a year in which the managing director asks at every status meeting what all this is actually delivering. No wonder many plants walk away at that point.

The good news: that sentence describes one particular class of approaches — not the technology itself.

What Works Without Any History

There is a category of monitoring that needs no historical data and no trained model: statistical monitoring of every signal against its own normal range.

Our system learns each signal's baseline from the first readings in live operation — and monitors continuously from then on. Three independent statistical methods run side by side: one for sudden spikes, one for slow sustained drift, one for abrupt level shifts. The drift method is the decisive one: it accumulates small deviations instead of waiting for one big one — and thereby sees exactly the creeping developments that fixed alarm limits are structurally blind to (why those are the most expensive failures).

Concretely: value from week one, with no data collection project. No waiting period in which nothing happens.

What Honestly Comes Later

To avoid any false impression — the staged approach has two stages, and the second one does need data:

Stage 1 — immediately: the universal watchdog described above. It doesn't know your machine, but it recognizes when it behaves differently from its own past. That covers a surprisingly large share of real-world cases.

Stage 2 — as soon as your data supports it: machine-specific trained models that have learned the particularities of your line. If you already have data in PLC logs, MES, or a historian — and most plants do, often without knowing it — that training can start immediately. If not, stage 1 collects the data as a side effect while already monitoring.

The difference to the "collect first" approach: collection is not a pre-project here. It is a byproduct of running value.

How to Spot the Dubious Promises

Two warning signs in a vendor meeting:

  1. "We first need X months of data" without the counter-question of what data your controllers already produce. Anyone who doesn't ask about PLC, MES, and historian first is planning your project past your existing assets.
  2. An accuracy promise before ever seeing your data. Accuracy depends on your machines, sensors, and data. The only honest order is the reverse: measure first, then quantify — our Predictive Maintenance module (fixed price after scoping, 7 weeks) ends with an accuracy report for your own equipment. If it comes out poorly, that is what it says.

Whether your existing data is enough for stage 2 is exactly what the AI Readiness Assessment answers — €9,500 flat, 1–1.5 weeks, open-ended. Or just ask us directly.

← Back to News & Publications