ONE PROBLEM, DECODED.
Six to eight hours
before failure.
A fault-prediction system that turns noisy, sparse sensor history into enough advance notice for maintenance teams to act.
TWO YEARS OF SENSOR DATA · SIX MONTHS LIVE OBSERVATIONEARLY SIGNAL / TIME TO ACT
01 / THE PROBLEM
Rare failures.
Noisy signals.
Production stops matter precisely because they are uncommon. That also makes them difficult to learn: fault examples are sparse, while the sensor stream around them is large and uneven.
The existing in-house model produced too many false positives to use in the field. Missing intervals, sensor malfunctions, and collection changes obscured the weak signals that appeared before a stop.
02 / THE METHOD
Learn the normal. Detect the deviation.
The system had to recognize an emerging fault early enough to change the maintenance decision - without overwhelming the operation with alarms.
We cleaned the time-series history while preserving the structure around known faults, then engineered features that could carry information despite sparse labels.
Models of increasing complexity were tested offline. A carefully tuned deep autoencoder was selected and deployed against the live sensor stream for operational observation.
- Time-series sensor values
- Sparse fault labels
- Collection intervals
- Real-time sensor stream
- 01
Clean
Separate collection errors and sensor malfunctions from meaningful behavior.
- 02
Represent
Extract features that reveal structure around rare fault events.
- 03
Select
Compare classical models and deep architectures through offline experiments.
- 04
Observe
Run the selected model on real-time data and measure operational behavior.
03 / THE RESULT
Advance notice the operation can use.
80%+
faults predicted
during the live observation period6-8h
advance notice
on average before a predicted stop8
false alarms
across six months of observationThe model creates a practical planning window: enough time to organize maintenance before a stop, with a false-alarm load the operation can absorb.
04 / YOUR PROBLEM