Problem
Maintenance was reactive — equipment was repaired after it failed. Repeated unplanned stops increased downtime and reduced productivity. A way to detect problems before failure was needed.
Data collected
Operating data from fastening and inspection equipment was collected, and a data assessment established which signals corresponded to which conditions.
Modules applied
E² provides real-time monitoring of equipment data and A² builds the anomaly detection models. An auto-encoder model was applied to fastening equipment and a deep neural network (CNN) to inspection equipment.
Result
Real-time anomaly detection and pre-emptive servicing shifted the workflow from reactive repair to predictive maintenance. Faults and replacement cycles can be anticipated, so stoppages can be planned.
A different model for each machine
Even on one line, machines of different character need different approaches.
Fastening equipment — auto-encoder Learns the normal operating state and measures how far current behaviour departs from it. Suited to equipment where failure cases are hard to accumulate.
Inspection equipment — deep neural network (CNN) Classifies the signal pattern itself, applied where anomaly types need to be distinguished.
What happens after the alert matters
An “anomaly detected” alert changes nothing if nobody has decided who does what. This project designed the workflow alongside the system: real-time detection, review, then pre-emptive servicing. Settling that sequence usually matters more than the system itself.