RULe Life Prediction (Training and Inference) Gateway

An edge AI gateway with a built-in life model training and inference engine. One unit collects sensor data, trains the Health Index (HI) and Remaining Useful Life (RUL) models, and runs online inference, making it ideal for compact, low-point-count deployments.

The RULe.TI series connects 2, 4 or 8 Aurora smart multivariate sensors (such as the 313MM2-R1) and receives the damage-mode features already extracted at the edge, so raw high-frequency vibration never needs to be transmitted and data volume drops by 99.9%. The gateway trains an EHI health index model against the equipment's own normal-operation baseline, then trains an RUL model from the EHI decay trend, answering three questions: how healthy is the equipment now, what health problem may it have, and how long can it keep running? Results can be sent to FDC, SCADA or PHM systems via standard protocols such as Modbus TCP to drive predictive maintenance and interlock protection.

Key features

• Edge gateway module: sensor connection management, and collection and storage of health data and damage-mode feature data.
• Life model training module: trains EHI and RUL models on the gateway, with automatic retraining as the equipment keeps running.
• Life prediction module: HI inference and visualization, damage-mode feature contribution analysis, and RUL inference and visualization.
• Operating-condition filter: ensures training and inference use only data from real running conditions.
• Three-color health light: green 80–100%, yellow 60–79%, red below 60%, paired with an SOP to guide maintenance.

Scenario

We suggest the Life Prediction (Training and Inference) Gateway when you:

• Want to predict the remaining useful life of critical equipment and turn sudden failures into planned maintenance;
• Have a small number of points (2 to 8) and want training and inference on a single unit, without a separate AI training server;
• Need HI and RUL sent straight to the fab FDC / SCADA, with alarms or interlocks when RUL falls below a threshold.

Model training workflow

• Before the software receives data: use the Aurora sensor ID consistently as both the equipment serial and sensor serial; mismatched names keep sensor data out of training. Turn off data collection during maintenance so that service-time values do not make EHI jump.
• Edge data collection setup: select the feature items for training and define operating-condition filters.
• EHI model training: use at least 14 days of running data and validate on a time range excluded from training; a training loss below 0.1 indicates good training. Retrain after the equipment is repaired.
• RUL model training: train from EHI inference results, preferably one month (30+ days) after EHI inference starts. The prediction horizon is half the history length, e.g. 30 days of data predicts the next 15 days.
• Automatic training: if you cannot wait for enough data, build an initial model (at least 1 hour of running data) and enable automatic training so it updates as data accumulates.

Models

All models include model training:

For more than 8 points, or to manage many gateways centrally, see the RULe Life Prediction Inference Gateway with an AI training server.

Case Study

• TSMC EFEM wafer robot life prediction (Level 4): Aurora 313MM2-R1 with RULe.TI002, connected straight to FDC.
• FATC DISCO dicing saw spindle and blade life prediction
• FATC DISCO grinder/polisher spindle life prediction: Aurora 313MM2-R1 with RULe.TI004.

Contact us

If you don't know whether your application is suitable for the Life Prediction (Training and Inference) Gateway, feel free to contact us.

Contact us