Maintenance agents for manufacturing

The decision this agent takes

For each critical asset, the agent watches vibration, current, temperature and cycle-count signals. When a signal crosses a learned threshold and the failure-mode probability exceeds the policy value your team sets, the agent decides: raise a maintenance work-order in the CMMS, tag it against the specific asset, and propose a slot in the next planned-maintenance window that does not clash with a scheduled production run.

Data the agent needs

• Machine telemetry over OPC UA, Modbus, MQTT or PROFINET (vibration, current, temperature, cycle count, alarm codes).
• Historical downtime and repair logs from the CMMS.
• Production schedule from the MES for planned-maintenance-window alignment.
• Parts availability from the ERP so the agent does not raise a work-order it cannot execute.
• A named failure-mode catalogue per asset class, owned by the reliability engineer.

What it does unsupervised

• Raise a work-order in the CMMS with asset ID, failure-mode hypothesis, and target execution window.
• Notify the assigned technician via email or the MES notification bus.
• Reserve spare-parts inventory in the ERP against the work-order.
• Update its own confidence score using the outcome (was the predicted failure real?).

Where a human stays in the loop

• Every threshold change to the failure-mode policy is signed off by the reliability engineer, not the agent.
• Work-orders above a cost or downtime threshold you set are routed for human approval before execution.
• The agent’s learned thresholds are reviewed monthly by the plant reliability committee.
• Emergency shutdowns are always human-triggered. The agent recommends, it does not stop lines.

Where this agent gets it wrong

Honest failure modes and the data-quality prerequisites we insist on before an agent goes live. If your plant does not meet these, the agent runs in advisory-only mode until it does.

• Sensor drift on a single machine will produce false positives until re-calibrated. Data-quality prereq: quarterly sensor-calibration log per asset.
• A new failure mode never seen in history will not be detected until it is added to the failure-mode catalogue.
• If CMMS work-order feedback (was it a real failure? was the repair effective?) is not entered by technicians, the agent stops learning and starts under-performing within 6-8 weeks.
• If the production schedule is not published to the agent, work-orders will clash with running production and be manually deferred, hurting adoption.

FAQ