Digital Twin Series · Level 4 · Prediction
Predictive Maintenance Digital Twin
Can the twin tell us a machine is going to fail before it does, and when to fix it?
A predictive maintenance twin watches live sensor data and the machine's history, spots abnormal behaviour early, estimates how healthy each part is, forecasts the remaining useful life (RUL) and recommends the right maintenance at the right time. This page has three working tools: an anomaly detection lab, a health index and RUL forecaster that you can check against the hidden truth, and a cost optimiser that compares run-to-failure, fixed-interval and predictive maintenance.
A simulated daily vibration level (RMS, in mm/s) with random noise. The detector learns the normal level and spread from the first 20 healthy days and raises an alarm when the value is more than k standard deviations above normal, for the chosen number of days in a row. A lower k catches faults sooner but gives more false alarms. The example failure limit is a made-up value, not a standard.
The health index (100 is new) is measured with noise. The twin fits a trend to the recent data, extends it to the failure threshold, and reports the remaining useful life with a range. The grey dashed line is the hidden truth that a real twin never sees; here it lets you judge the forecast. An accelerating degradation is under-forecast by a straight line, and a short window is noisier than a long one.
| Strategy | Replace | Chance of failure first | Cost per day |
|---|
Part lives are modelled as a normal spread around the mean life. Fixed-interval (preventive) maintenance has to allow for the whole spread of lives of the fleet. Predictive maintenance knows more about this particular machine, so its uncertainty is smaller and the replacement can be later and safer. The best time is where the cost per day is lowest. A twin that is wrong, or that raises false alarms, earns less: set the twin's uncertainty to see this.
| Component | Condition | Prediction | Recommended action | Schedule |
|---|
A simple rule set. Move the RUL sliders to see how the recommendation changes. A real twin also considers the cost of the stop, the spare part lead time and the production plan.
Figures are indicative and the simulations are simplified. Calibrate any model with your own failure history before you rely on its predictions.