Predictive Maintenance Pilot

The problem

A precision manufacturing client was running calendar-based preventative maintenance — every part replaced on a fixed schedule whether or not it needed it. Expensive (replacing components with useful life left) and inadequate (40% of failures were unscheduled and caused production line stops averaging 4 hours each).

What we built

A pilot system that:

• Reads vibration, temperature, and motor-current sensors at 1 Hz from six pieces of equipment
• Computes rolling spectral features and statistical baselines per piece
• Trains a per-equipment XGBoost model on the last 90 days of failure data
• Issues a "high risk in next 14 days" flag, surfaced to the maintenance team via a Slack-integrated dashboard

What we deliberately did not do

• We did not build a deep-learning model. The data did not justify it; gradient-boosted trees performed within 2 percentage points of an LSTM and were 30× faster to train.
• We did not build a full enterprise asset management system. The client already had one. We added one signal to it.
• We did not move inference to the cloud. The factory has reliable power but unreliable internet, so the model runs on a small Linux box on the floor.

Outcome

• Unscheduled downtime dropped from ~14 hours/week to ~4 hours/week over 8 weeks
• Maintenance team adoption was high — the dashboard lived inside Slack, where they already paid attention
• The client extended the system to two more lines after the pilot

Tech stack

Python, scikit-learn, XGBoost, FastAPI, PostgreSQL/TimescaleDB, a small Vue dashboard, all containerised on a single Intel NUC on the factory floor.