Four-week analytics pilot, production maintenance result.

Client type

A mining operation running conveyor systems, with a maintenance team responsible for keeping them available and years of process historian data that it was not yet using to plan maintenance. The operation is not named because the work was delivered under NDA.

Scope

• Review of the existing historian data.
• Failure-mode analysis for conveyor motors.
• Feature engineering on the signals that mattered.
• A dashboard prototype and a maintenance-priority register.
• Support for the production rollout after the pilot.

Constraints

The maintenance team had years of historian data but little usable warning before repeat motor failures. The obvious answer was not a large AI project. The first need was a practical pilot: prove whether the existing signals could identify failure modes early enough for maintenance to act, and keep the scope narrow enough to finish in four weeks.

Standards

There is no drawing code for a pilot like this; the standard the work had to meet was the maintenance team’s own planning process. That meant explainable features rather than a black-box prediction score, a dashboard and priority register that planners could act on, and defined escalation rules for maintenance planners and reliability engineers.

Deliverables

• Cleaned historian tags, with motor load, current, vibration, temperature and stoppage signals aligned.
• A comparison of failure windows against normal operating periods and seasonal load changes.
• A small set of features the maintenance team could understand and trust.
• A dashboard and priority register instead of a black-box score.
• Escalation rules for maintenance planners and reliability engineers.

Timeline

The pilot took four weeks. Its findings justified a production rollout, and the result was measured six months later, once the targeted interventions had been running long enough to show in the downtime and maintenance figures.

Outcome

The project notes focus on the asset-data inputs and the failure-mode questions. Any decision to extend the analysis needs an agreed operational baseline and validation criteria.

Why the pilot worked

The project stayed narrow. We did not try to model every asset in the plant. We selected one failure pattern, proved the signal quality, and made sure the output fit the maintenance team’s planning rhythm. That kept the result operational instead of academic.