Predictive Maintenance Dashboard

Business Challenge

Ironclad Manufacturing ran maintenance on a fixed calendar schedule — every 90 days regardless of actual equipment condition — leading to both wasted maintenance on healthy machines and unplanned failures on ones that needed attention sooner.

Objectives

• Predict equipment failures before they cause unplanned downtime
• Reduce unnecessary scheduled maintenance on healthy machines
• Give floor supervisors a clear, non-technical dashboard, not a data science tool

Approach

Ironclad already had vibration and temperature sensors on their newer machines but wasn't using the data beyond basic threshold alerts. We started by proving the concept on their 20 newest machines before expanding sensor coverage further.

Deliverables

A predictive maintenance model ingesting live sensor data, a Grafana dashboard for floor supervisors, and an automated alert system flagging machines trending toward failure.

Solution

Planning

We prioritized the machines with the highest historical downtime cost first, since that's where prediction accuracy would matter most to the business case.

Design

The dashboard shows a simple health score per machine (green/yellow/red) rather than raw sensor readings — supervisors needed a decision tool, not a data science interface.

Development

Built a Python pipeline ingesting sensor data over MQTT into PostgreSQL, with anomaly-detection models trained on 18 months of historical failure data, and Grafana as the supervisor-facing dashboard layer.

Testing

Validated predictions against six months of historical failures the model hadn't seen during training, tuning the alert threshold until false positives dropped to a rate supervisors considered trustworthy.

Deployment

Rolled out to the initial 20 machines first, then expanded to the remaining 44 once floor supervisors confirmed the health scores matched their own hands-on assessment of machine condition.

Optimization

Post-launch, we added a secondary model for a class of intermittent failures the first version missed — these required a different detection window than the steady-degradation failures the original model handled well.