Snapshot
Who: Global components plant network (anonymized) — mixed legacy PLCs and a newer MES layer.
Problem: PdM pilots had stalled: too many false positives, tribal thresholds per site, and no shared playbook when engineers rotated.
What we shipped: Multivariate anomaly models per asset class, explainability overlays for shop-floor tablets, and a closed loop from work-order outcomes back into feature stores.
Challenge
Each plant had a different “yellow band” on charts. Corporate reliability wanted a defensible standard without ripping out PLCs.
Approach
We clustered assets by kinematic family, trained on 14 months of historian data, and shipped shadow scoring for 90 days before any wrench-time KPIs counted.
Solution components
Rollout timeline
Phase 0 · Data audit
Tagged 860 assets; retired 12% of sensors that never calibrated cleanly.
Phase 1 · Two plants
Shadow mode; weekly calibration with reliability SMEs.
Phase 2 · Network scale
Blue/green model deploys; multilingual UI strings for EMEA sites.
Phase 3 · Finance tie-in
Linked downtime minutes to loaded labor rates for board reporting.
“We stopped chasing ghosts. The top of the queue is boring now—and that’s exactly what we needed.”
— Director of Plant Reliability, anonymized
Results & controls
- Unplanned downtime −62% YoY on instrumented lines (cohort vs. control lines).
- Emergency WO volume −28%; PM compliance actually rose because crews had bandwidth.
- Safety near-misses tied to mechanical failures down 19% in the same period.
- Model cards stored with each release: data slices, metrics, and sign-offs from quality.
Representative scenario for marketing purposes. Metrics are illustrative composites unless covered by a separate, signed case release.