Case studies / Manufacturing
Case studyEarly warning of machine failure that planners act on
A plastics moulding operator combined sensor readings with maintenance history to warn planners before machines failed, and tuned its alerts until they were trusted.
Overview
A family-owned injection-moulding business in Poland making food-grade containers and closures for European brands, running two plants on a three-shift, six-day pattern.
Scale. About 340 machines, including moulding presses, robots, chillers and dryers, with a maintenance team of 28 across both sites.
The challenge
On a busy Tuesday, a hydraulic pump failed on one of the largest presses. The tool had to be pulled, an urgent customer order slipped by four days, and the replacement part came by courier from Germany. It was the third such failure that year, and each one cost far more in lost output and penalties than the part itself.
Maintenance was almost entirely reactive. Planned servicing followed fixed intervals that bore little relation to how hard each machine actually worked. The presses had sensors, but their readings were stored locally and rarely looked at, and the maintenance system held years of work orders that nobody had connected to them.
The operations director had tried an off-the-shelf monitoring tool. It produced dozens of alerts a day, most of them irrelevant, and within a month planners had stopped reading them.
The solution
Readings from presses, chillers and dryers now flow every few minutes into a central data platform, alongside the maintenance history and the production plan. Models for each machine type watch for the patterns that preceded past failures and score how likely a problem is in the coming days.
When a score crosses its threshold, the planner receives a single alert on their phone and in the maintenance system, showing which reading has changed, how it compares with the machine's normal behaviour, and similar past cases. They can raise a work order with one action, schedule it into a planned changeover and order parts in advance.
Thresholds are not fixed. Each alert carries a feedback step, and planners mark whether it was useful. The data team reviews this monthly and adjusts sensitivity per machine, so a press that runs hot by nature does not flood the queue, and a quiet chiller that starts to drift is noticed.
The engagement
- Seven months
- Five specialists, alongside the client's maintenance engineers
- Data collection from machines and records, failure pattern analysis, alert design, maintenance work-order integration, threshold tuning and extension to further machines.
Integrations
- Machine sensors and controllers
- Maintenance management system
- Production planning
- Mobile alerts
- Central data platform
How we delivered it
- 01
Combine history and choose critical machines
Six weeks
We joined two years of sensor readings with the maintenance work-order history and the production schedule. With the maintenance manager we ranked machines by the cost of an unplanned stop and picked the forty that mattered most for the first phase.
- 02
Build and test against past failures
Seven weeks
For each machine type we looked for patterns in the readings before past failures: rising oil temperature, pressure drift, heater current changes, longer cycle times. Models were tested by replaying history and checking whether they would have warned in time without crying wolf.
- 03
Send alerts to planners who decide
Eight weeks
Alerts went to maintenance planners with the evidence attached and a suggested inspection. Planners decided whether to act, when, and how to fit it around production. Every decision and its outcome was recorded.
- 04
Tune thresholds and extend
Eight weeks
Each fortnight we reviewed every alert with the planners: useful, too early, too late or noise. Thresholds were adjusted machine by machine, and once alert quality held, coverage grew to 120 machines across both plants.
Results
- Fewer unplanned stoppages on critical presses
- Maintenance planned around production rather than against it
- Parts ordered ahead, with fewer emergency deliveries
- Alerts trusted because they were tuned with planners
Unplanned stoppages on monitored machines fell noticeably over the first six months of full operation. Several failures that would previously have stopped a press mid-order were caught days ahead and fixed during scheduled tool changes.
Maintenance planning has become calmer and more deliberate. Parts are ordered before they are needed, overtime call-outs have dropped, and fixed-interval servicing is being replaced with condition-based checks on the machines where the data supports it.
Most importantly, people read the alerts. Because they were tuned with the planners rather than for them, most now lead to a planned action. The maintenance manager describes the alert queue as a short list worth reading, not a noise to be filtered.
The first tool we tried was clever but we stopped looking at it. This one tells us a few things a week, and they are usually right.
What made the difference
Alert tuning mattered more than the model: too many alerts get ignored. The fortnightly review with planners did more for results than any improvement in prediction accuracy.
