Predictive Maintenance for Industrial Equipment

Problem

When engineers receive equipment sensor readings, the practical question is not just “plot the data,” but whether the current operating state is close to failure, which failure mode is more likely, and whether maintenance is worth scheduling. predictive-maintenance targets industrial equipment maintenance scenarios. It accepts air_temperature, process_temperature, rotational_speed, torque, tool_wear, and product_type, then returns a failure prediction and optional warnings.

How It Works

The skill connects to the predictive_maintenance service via MCP SSE and exposes four core tools:
- predict_equipment_failure: single-device prediction for questions like “Is this unit at risk?”
- batch_predict_equipment_failure: batch prediction for device lists or CSV-style inputs, returning totals, failure counts, and type distribution.
- get_sensor_normal_ranges: queries normal sensor ranges to validate inputs before prediction.
- get_model_info: shows model version, class definitions, and recognizable failure types.

It covers common mechanical and process failures such as Power Failure, Tool Wear Failure, Overstrain Failure, and Heat Dissipation Failure. The model uses XGBoost with SMOTENC to handle class imbalance. Because missed detections can be costly in maintenance workflows, F2-score places more emphasis on recall.

Boundaries

If inputs fall outside normal ranges, the tool returns warnings but still runs the prediction. The model is trained for a specific industrial manufacturing context, so substantially different equipment types may require retraining. It also relies mainly on sensor-visible signals and does not cover software faults, operator errors, or other non-sensor-caused issues.