Machine learning · Discrete manufacturing
Situation
SomaTech Components makes precision sub-assemblies for automotive and medical OEMs. Unplanned spindle and bearing failure caused line halts, scrap, and last-minute contract manufacturer overflow. The maintenance team had vibration snapshots and CMMS work orders, but not a single ranked view of “next 48–72h risk” for critical spindles.
What we did
NeuralFaaruuq integrated high-frequency vibration and thermal readouts with MES context (part mix, run speed, lot changes). We trained gradient-boosted models with back-tested labels from historical failures and tattletale alerts. Explainability packs translate model drivers into what the floor can act on. Playbooks: yellow vs. red work orders, part kits staged before weekend runs, and capital budget triggers when recurring alerts cluster on a single asset class.
- Data pipeline from shop-floor historians with drift checks
- Per-plant validation so models respect different machine vintages
- Integration with the CMMS to open templated work orders in one click
Outcomes (six-month controlled pilot, three plants)
- 27% reduction in unplanned stoppage hours across pilot lines (vs. same season prior year, weather- and load-normalized)
- Fewer emergency part air-freights; maintenance crews report higher confidence in weekend planning
Tech stack (high level)
Time-series feature store, Python/SQL training jobs, on-prem edge scoring where latency is tight, and dashboards the plant manager reviews daily.
