Predicting failures before they stop the line

The client manufactures heavy industrial equipment across nine plants on three continents. Machining centers, presses, and paint-line assets ran on calendar-based preventive maintenance, yet unplanned downtime still consumed 11% of scheduled production time — with a single constrained machining line in Monterrey costing an estimated $140K per lost shift.

Plant data existed but was stranded: 20 years of historian tags, maintenance logs in three CMMS systems and two languages, and vibration routes collected monthly on clipboards. Two prior 'Industry 4.0' pilots had produced dashboards but no changed maintenance decisions.

The failure modes that hurt most — spindle bearing degradation, hydraulic press seal failures, paint-line conveyor faults — develop over weeks but were being caught in hours. Detecting them required continuous high-frequency vibration and process data the monthly clipboard routes could not provide, plus models tuned per asset class rather than generic anomaly scores that maintenance planners had learned to ignore.

Operationally, the constraint was trust and workflow: predictions had to arrive as prioritized, evidence-backed work orders inside the CMMS that planners already used — with enough diagnostic context that a technician knew what to inspect — or they would join the previous pilots as ignored dashboards.