Context
A Singapore-headquartered precision-engineering exporter ran two production lines (CNC + post-process inspection) producing high-mix, low-volume parts for the medical-device and aerospace supply chain. Roughly 18% of planned production hours were lost to unplanned downtime — split between bearing failures on the spindle heads and cosmetic-defect false-rejects from the human inspection station.
The cost: ~3 days of production lost per month per line, plus inventory carrying cost on customer reservations they couldn’t ship on time. They’d evaluated three “predictive maintenance” SaaS products. None of them tied into their Yonyou NC Cloud ERP, none adapted to the specific failure modes their spindles exhibited, and none could close the work-order loop without a human re-keying the alert into their existing maintenance system.
What we did
Spindle vibration monitoring
Instrumented the spindle-head bearings with three-axis accelerometers, streaming via OPC UA → MQTT into Azure IoT Hub. Trained a per-asset anomaly model that tied vibration signatures to specific failure modes (not generic “outlier” scoring) — bearing wear vs. lubrication loss vs. misalignment each surfaces a distinct frequency-domain fingerprint, and the maintenance team’s response is different for each.
Vision-AI inspection station
Deployed a Vision-AI inspection rig next to the line — small convolutional model running on an edge box, fail-safe to manual inspection if confidence dropped below threshold. 60 units/minute throughput, sub-2-second decision. The team kept the human inspector station running in parallel for the first 8 weeks; we measured agreement rate, tuned, then transitioned the human to exception review only.
Yonyou NC Cloud work-order integration
Alerts flowed into the existing Yonyou maintenance work-order module — pre-filled with asset ID, failure-mode prediction, recommended parts, and a suggested 48-hour scheduling window that minimized production impact. The maintenance team confirmed and dispatched from the same screen they’d been using for years.
Results
- -41% unplanned downtime across the two pilot lines
- < 2 seconds defect detection at 60 units/min sustained throughput
- Zero false-rejects flagged as costly customer returns over the first 6 months in production
- 92% alert-to-scheduled-work-order rate within 48 hours of prediction (vs. 38% before)
- Customer NPS lift attributable to on-time delivery improvement; customer reservations stopped slipping
What we learned
The biggest unlock wasn’t the model — it was getting maintenance to trust the alerts enough to schedule the work. We spent more time on the human-process change than on the ML itself: every false-positive investigation reviewed publicly, every saved bearing failure celebrated, the model’s confidence threshold tuned with the maintenance lead in the room.
The Yonyou integration mattered more than we expected. The previous SaaS tools had failed at adoption because alerts arrived as email — which got filtered, ignored, and eventually distrusted. Arriving as a pre-filled work order in the system the team already used closed the loop in days, not months.