Ride health — Predictive maintenance dashboard
Centralised dashboard that ingests telemetry from ride-sensor-monitoring nodes, runs anomaly detection (statistical baselines + simple ML), and surfaces failure predictions weeks before they happen. Replaces the 'wait until it breaks' model of mechanical maintenance. Built on the open-source observability stack: Prometheus for metrics, Grafana for visualisation, a Python service that computes rolling baselines per ride and flags drift.
Problem
Even with sensor loggers on every ride, raw vibration data isn't actionable. Maintenance teams get a stream of numbers but no signal of when something is starting to fail. The data lives in dashboards nobody checks daily.
Solution
A Python service that ingests ride telemetry, computes rolling baselines per ride (median + std of vibration at each speed, current draw at each phase, etc.), and flags drift weeks before a hard failure. Outputs predictions to a maintenance dashboard with prioritised work orders.
Key takeaways
- Ingests ride-sensor-monitoring data via MQTT
- Rolling baselines per ride, anomaly detection on drift
- Failure predictions weeks in advance, prioritised work orders
- Open observability stack: Prometheus, Grafana, scikit-learn
Concept
This is the back end that turns the stream of ride telemetry into decisions. A Python service subscribes to the MQTT broker where the ride-sensor-monitoring nodes publish, stores the readings in InfluxDB (designed for time-series data), and computes per-ride baselines: median vibration at each speed, current draw at each phase of the ride cycle, temperature at each point in the day. Each new reading is compared to its baseline; deviations above a configurable threshold become alerts.
The ML piece is intentionally simple: rolling statistics, threshold rules, and a scikit-learn model trained on the ride’s own history to flag drift that humans would miss. The output is a maintenance dashboard built on Grafana, with prioritised work orders: this ride’s vibration signature has drifted 12% over the last 30 days, inspect the bearings next week. The system is containerised, runs on the same infrastructure as the rest of the park’s observability stack, and the data is open: park engineers can query it, export it, build on it.
Why this matters in themed entertainment
Predictive maintenance is well established in aerospace and heavy industry, but parks have been slow to adopt it because the failure modes are different (rides run hundreds of cycles a day, not a handful) and the cost of being wrong is measured in guest complaints, not safety incidents. A small, open-source tool that learns per-ride baselines and surfaces drift early is a way to bring the discipline into the park without the multi-million-dollar systems the big players have. The first park to do it cleanly will have a measurable downtime win.