Maintenance

Before predictive maintenance,
build the equipment data first

Predictive maintenance most often stalls on data, not algorithms. When failure history is scattered across documents and maintenance records vary by person, no model gets past that wall.

Request a site surveyAll guides

Preparation sequence

  1. Bring failure and maintenance records into one format

    History scattered across spreadsheets, paper logs and individual notes gets unified into the same fields — equipment, timestamp, symptom, cause, action taken, downtime.

  2. Check what signals each machine can actually provide

    Survey machine by machine whether signals related to failure — vibration, temperature, current, run time — can be collected today.

  3. Establish a baseline for normal operation

    Recording only failure moments leaves nothing to compare against.

  4. Write down the rules per maintenance type

    Document preventive maintenance intervals, the criteria for reactive maintenance, and which equipment moves to predictive maintenance.

Why projects stall here

Predictive maintenance conversations usually start with "we want AI to warn us before a failure." In practice, many plants cannot even produce a table of how many failures occurred last year or which part caused each one.

A predictive model needs data that can tell a failure state apart from a normal one. Bringing in AI before that data exists means the project spends most of its time on data collection rather than modelling.

Structuring maintenance (CMMS) data comes before predictive maintenance, not after. Once that structure is in place, predictive maintenance sits naturally on top of it.

Procedure

Following the order matters more than perfecting each step.

  1. STEP 01

    Bring failure and maintenance records into one format

    History scattered across spreadsheets, paper logs and individual notes gets unified into the same fields — equipment, timestamp, symptom, cause, action taken, downtime. Records that do not share a format cannot be merged later.

  2. STEP 02

    Check what signals each machine can actually provide

    Survey machine by machine whether signals related to failure — vibration, temperature, current, run time — can be collected today. Where there is no communication, consider attaching an external sensor.

  3. STEP 03

    Establish a baseline for normal operation

    Recording only failure moments leaves nothing to compare against. Signal ranges during normal operation need to be collected over a period first, to serve as the baseline that makes an anomaly recognisable as one.

  4. STEP 04

    Write down the rules per maintenance type

    Document preventive maintenance intervals, the criteria for reactive maintenance, and which equipment moves to predictive maintenance. Do not move every machine to predictive maintenance at once.

Common mistakes

Patterns that recur in implementation discussions.

Common mistake

Installing sensors before there is any failure history

Signals accumulate, but without failure records to compare against there is no basis for judging what counts as an anomaly. Organising history always comes first.

Common mistake

Treating every machine as a predictive-maintenance target at once

Starting with the one or two machines that fail most often and cost the most downtime builds data faster and shows results sooner.

Common mistake

Keeping maintenance records as one person's personal notes

The reasoning disappears along with the person when they leave or hand over the role. Maintenance history has to live in a system, not an individual, to be usable as data.

Common mistake

Setting alert thresholds once and never revisiting them

Equipment condition, season and load keep changing. Initial thresholds need periodic adjustment based on actual false-positive and false-negative results in operation.

Checklist

Confirm these are in place before starting.

  • Failure and maintenance history per machine is recorded in one consistent format
  • Failure causes and actions taken are captured as codes or standard fields
  • Collectible signals and collection methods have been surveyed per machine
  • A baseline from normal operating periods has been collected over time
  • Priority equipment for predictive maintenance is selected by failure frequency and downtime cost
  • A review cycle and owner are assigned for revisiting alert thresholds

Tools for this stage

The procedure is the same without any product. These shorten the steps.

Tool

U-CMMS

Builds equipment history and maintenance records into a standard format — the base data predictive maintenance is built on.

Tool

SRP-100W

Collects vibration, current and other equipment signals to build the normal-operation baseline.

Quick Answers

Preparing data for predictive maintenance — frequently asked

Questions that come up during evaluation.

We have very little failure history — is predictive maintenance still possible?

With few recorded failures, threshold-based anomaly detection is a more realistic starting point than statistical prediction. Refine it as history accumulates.

Can we adopt predictive maintenance without CMMS first?

It is possible, but failure and maintenance history still has to be structured somewhere, which ends up rebuilding what CMMS already does. Designing both together is more efficient.

How much data do we need before a model is usable?

It depends on the equipment and failure type. The usual approach is to first build a solid baseline from normal-operation data, then improve the model incrementally as failure cases accumulate.

Is CMMS worth using on its own, without predictive maintenance?

Yes. Standardising failure history and maintenance records alone helps immediately with finding root causes of recurring failures and tuning preventive maintenance intervals. Predictive maintenance is the step after that.

How does this play out in your plant?

Equipment mix and current workflow determine how the procedure is adapted.

Request a site survey
KR KO