Technology · Data features & models
From raw signal to a model you own.
Hub and nodes feed the same logging, the same dashboard and the same training.
- SenseVision, thermal, depth, sound, vibration
- LogEvery stream time-stamped and hashed
- AnalyseWaveforms, spectra, features
- LabelPer part, asset and fault
- TrainAnomaly to severity
- DeployModel on the edge device
- MonitorResults on the line, retraining
The model ladder
We climb only as far as your data supports.
- Anomaly detectionLearns normal from healthy recordings. Works without a single fault example.
- Fault detectionHealthy versus a named fault, for example a cold weld.
- Multi-fault classificationWhich fault: inner ring, outer ring, rolling element, imbalance.
- Severity gradingHow bad is it? So you can plan the stop.
How we check a model for shortcuts
- Split by recording or asset, never by window
- Balanced accuracy and per-class recall
- Shortcut probes: does speed or load alone explain the result?
- An operating point from the real cost of false alarms and missed faults
Listen: what bearing faults sound like in the vibration signal
Next step
How far do your recordings go?
Send us a sample. We tell you which rung of the ladder is realistic.