72% less unplanned downtime across a diagnostics fleet

The diagnostics division of a global pharmaceutical leader operates a large fleet of clinical chemistry and immunochemistry analyzers across hospital and reference laboratories. The instruments are workhorses: hundreds of assays a day, tight QC windows, and laboratories that plan their staffing around them.

Maintenance was reactive. An analyzer failed, the lab logged a service call, and a field engineer was dispatched with whatever parts the error code suggested. Each unplanned outage meant halted assays, samples couriered to other sites, recalibration and QC reruns after the repair, and an emergency dispatch that cost a multiple of a scheduled visit.

The service organization could see the pattern in its own data. It could not see the failures coming.

We started with the data the instruments were already producing: sensor telemetry, error and event logs, reagent and calibration records, and a decade of service history sitting in the field service system.

The core is an anomaly detection layer that models normal behavior per instrument and per assay type. Photometer lamp output decays on a curve. Pipettor pressure profiles drift before dispensing errors appear. Temperature excursions in reagent storage precede a specific family of QC failures. The models flag the deviation days before it becomes an error code, and a remaining-useful-life estimate turns the flag into a date.

On top of that sits intervention scheduling. Predictions are ranked by failure probability, clinical impact, and parts availability, then pushed into the dispatch queue the service organization already uses. An engineer arrives during a planned window with the right part, instead of after a failure with a diagnostic guess.

Production was the hard part, as it usually is. The telemetry pipeline had to be qualified as a data source under the client's quality system, which meant documented lineage from instrument to model input. Alert thresholds were tuned with the dispatchers, deliberately: a model that cries wolf gets muted within a month, so we launched conservative and widened coverage as precision held.

We ran a pilot on roughly 10% of the fleet and let the service organization audit every prediction against the actual outcome for a full quarter. Model changes go through the client's change control process, and retraining runs on an approved schedule with documented performance review.

The system is owned by the client's service organization today. Our team transferred operations after twelve months of joint running.