Predictive maintenance: start with the data, not the algorithm

Predictive maintenance promises to anticipate failures. Models and dashboards are now widely available in off-the-shelf platforms. Yet many projects stall at the pilot stage. The reason is rarely the algorithm: it is almost always the data.

The most common pitfalls

• Data without context: a temperature without knowing which machine, which product, which operating state.
• The wrong sampling rate: one reading per minute misses a vibration degrading within seconds.
• No failure history: without documented failure events, a model has nothing to learn from.
• Collection gaps: unstable gateways, inconsistent timestamps, data lost during stoppages.

The right sequence

1. Connect existing equipment without modifying PLCs (see OPC UA or MQTT).
2. Contextualise: attach each reading to a line, a machine, a state.
3. Monitor with simple thresholds and trends, validated by maintenance.
4. Log events: failures, interventions, part replacements.
5. Only then train prediction models on a reliable history.

What we do, and what we do not sell on its own

We build analytics into our industrial data integration service, but we do not sell it on its own: its value depends entirely on the quality of collection. A one-line pilot checks that quickly.