Predictive maintenance with foundation models
Forecast sensor readings — vibration, temperature, pressure, current draw — and use prediction intervals to estimate when equipment will cross failure thresholds. No per-machine model training required.
{
"model": "amazon/chronos-bolt-base",
"inputs": [
{
"item_id": "pump-07-vibration-mm-s",
"target": [2.1, 2.3, 2.2, 2.5, 2.8, 3.1, 3.4, 3.6, 3.9, 4.2],
"start": "2026-04-01T00:00:00Z"
},
{
"item_id": "pump-07-bearing-temp-C",
"target": [62, 63, 62, 64, 66, 68, 70, 72, 74, 76],
"start": "2026-04-01T00:00:00Z"
}
],
"parameters": {
"prediction_length": 72,
"freq": "h",
"quantiles": [0.1, 0.5, 0.9]
}
}72-hour forecast for vibration and bearing temperature with prediction intervals for threshold breach estimation.
Zero-shot
Forecast any sensor series without per-equipment model training
Multi-sensor
Batch multiple sensor channels in a single API request
Threshold planning
Prediction intervals estimate when readings will breach limits
Why foundation models improve predictive maintenance
Traditional predictive maintenance requires failure history, labeled degradation data, and per-equipment model fitting. Foundation models remove these prerequisites.
No failure history required
Classical PdM models need labeled examples of failures to train on. Foundation models forecast sensor trends zero-shot — you just need the recent sensor readings and a known failure threshold.
Works across equipment types
The same API call forecasts vibration data from a pump, temperature from a bearing, current draw from a motor, and pressure from a hydraulic system. No domain-specific model configuration.
Prediction intervals for maintenance windows
Quantile forecasts tell you not just when the median prediction crosses a threshold, but the range of likely crossing times. Use the upper quantile for conservative scheduling and the lower quantile for early warning.
View quantile docsGetting started with predictive maintenance
Most maintenance teams start with one critical asset, validate against known degradation events, then expand to a fleet.
- 1
Select a critical asset with sensor data
Choose a pump, motor, compressor, or other asset that has regular sensor readings (vibration, temperature, pressure, current) and known failure thresholds from the manufacturer or your maintenance team.
- 2
Forecast sensor trends and check thresholds
Send recent sensor history to the API. Compare the forecast trajectory and prediction intervals against your failure thresholds. If the upper quantile crosses the threshold within your planning horizon, schedule maintenance.
- 3
Scale to your asset fleet
Use the batch endpoint to forecast all sensors across all assets in a single request. Integrate with your CMMS or EAM system to automatically generate work orders when forecasts breach thresholds.
Sensor types that work well
Any regularly sampled sensor measurement can be forecasted. These are the most common in predictive maintenance programs.
Vibration analysis
Bearing vibration (velocity in mm/s or acceleration in g) is the most common PdM signal. Foundation models capture the gradual increase pattern that precedes bearing failure.
Temperature monitoring
Bearing temperature, winding temperature, and coolant temperature all follow predictable degradation curves. Forecast these to estimate when thermal limits will be reached.
Electrical signatures
Current draw, power factor, and impedance measurements from motors and drives. Increasing current draw at constant load often indicates mechanical degradation that the motor compensates for.
Frequently Asked Questions
Forecast your first sensor
Send a vibration or temperature series to the API and see when the forecast crosses your failure threshold. No model training required.