Three lessons from rolling out predictive maintenance across 40 sites
Two years, 40 plants, 11 countries. What worked, what did not, and what we would do differently.
Halden Systems Newsroom - - 1 min read
Over the past two years Halden rolled out predictive maintenance across 40 sites for a single pulp and paper customer, including the 14 Nordfjord Paper mills written up on the customer stories page. Here are the three lessons that would have saved the most time had we known them at the start.
1. Sensors are the easy part
Fitting vibration and temperature sensors to rotating equipment was the quickest part of every site. The slow part was agreeing who gets the alert and what they do with it. Where the plant had a named owner for each asset, the system paid for itself within months. Where it did not, alerts went to a shared inbox and were ignored.
We now refuse to switch alerts on for an asset until it has a named owner. It sounds bureaucratic; it is the single biggest predictor of whether the system works.
2. Start with the equipment that has stopped the line before
Every plant has a short list of machines everyone knows will fail again. Starting there builds trust fast, because the first alert the maintenance team gets is one they believe. Starting with a complete asset register, as we did at the first three sites, produced hundreds of alerts on equipment nobody was worried about and taught the team to ignore them.
3. Models are local
A model trained on one site's refiner did not transfer cleanly to the next site's. Same machine, different load profile, different fault signatures. We now treat the first eight weeks at any site as calibration, say so in the contract, and do not report results until the calibration period is over.
The result
Across the 40 sites, unplanned stops on monitored equipment fell by a little under half. The customer's own figure is in their annual report; ours is the one above. The practice area page has the current commercial model.
