Predictive Maintenance for Industrial Equipment

The Challenge

A precision manufacturing facility operating 850+ CNC machines was experiencing an average of 14 unplanned equipment failures per week. Each failure caused 2–6 hours of production downtime while maintenance crews diagnosed and repaired the issue. The total annual cost of unplanned downtime — including lost production, expedited parts, and overtime labour — exceeded $6.3M.

The existing maintenance programme was purely calendar-based: bearings were replaced every 2,000 hours, coolant systems were serviced quarterly, and spindle assemblies were inspected annually — regardless of actual condition.

Our Approach

We deployed a predictive maintenance system that monitors machine health in real-time using vibration, thermal, and power consumption sensors, with inference running on edge devices at each machine.

Sensor Network: Each CNC machine was fitted with a triaxial accelerometer (vibration), infrared thermal sensor, and a current transformer on the spindle motor. Sensors sample at 10 kHz (vibration) and 1 Hz (thermal, current) and transmit data to a local edge device via MQTT.

Edge Inference: Each edge device (Raspberry Pi 4 with Coral TPU) runs a TensorFlow Lite model that classifies the machine's health state into four categories: Normal, Watch, Warning, and Critical. The model was trained on 18 months of historical sensor data labelled with maintenance records. Inference runs continuously with 100ms latency, and only anomaly events are transmitted to the cloud — reducing bandwidth by 95%.

Cloud Analytics: Anomaly events are ingested into TimescaleDB and visualised on Grafana dashboards. A secondary model predicts remaining useful life (RUL) for key components (bearings, spindle, coolant pump), allowing maintenance to be scheduled during planned production gaps.

Results