Example system · Manufacturing
Predictive maintenance
Hear a machine failing weeks before it stops.
Example system Manufacturing
Hear a machine failing weeks before it stops.
Sense. Vibration, sound, temperature and motor current from every critical machine, sampled thousands of times a second and condensed on the factory floor.
Illustrative system. Simulated data, not client results.
Inside the engineering
The problem
Machines rarely fail without warning, but the warning is too faint for people to hear. By the time a bearing is loud, it is days from seizing, and an unplanned stop idles the line, the shift and the order behind it.
Why it is hard
- 01Failures are rare, so there is little labelled history to learn from.
- 02Every machine sounds different, even two of the same model.
- 03Factory networks drop out, and raw vibration is far too much data to send anywhere.
- 04False alarms cost trust quickly. A week of crying wolf and the alerts get ignored.
What we build
From signal to decision.
- 01
Sense
Accelerometers, contact microphones and current clamps on each critical machine. An edge gateway on site computes spectra and features, buffers through network outages, and sends only what matters.
- 02
Learn
A transformer pre-trained on each machine’s normal running, then refined as real faults are confirmed. Bearing fault frequencies and other physics go in alongside, so the model starts from what engineers already know.
- 03
Predict
An anomaly score per machine, a likely fault, and a remaining-life estimate with a confidence band. Alert thresholds are set against your tolerance for false alarms, not ours.
- 04
Act
Alerts on WhatsApp or a simple app, a health board for the plant, and work orders pushed to your maintenance system or a shared calendar. Each alert asks the technician what they found, which trains the next version.
Before we start
What needs to be in place.
- Access to fit sensors on the machines that matter most, usually five to twenty.
- Two to four weeks of normal running before alerts are switched on.
- Someone on site to confirm or reject the first alerts.
- Patience on the time estimate: it detects drift reliably from the start, and estimates time left better as confirmed faults accumulate.
How the project runs
From survey to live.
Fixed price once the specification is agreed. Every repository, model and account is yours from the first day.
- Survey1–2 weeksChoose the machines whose failure costs most. Map power, network and mounting points.
- Instrument2–3 weeksFit sensors and the gateway. Start recording normal operation.
- Model4–6 weeksTrain per-machine baselines, calibrate thresholds, run silently beside your team.
- Go live1–2 weeksSwitch on alerts, connect work orders, hand over the runbooks.