CASE 002 / APPLIED INTELLIGENCE

PREDICTIVE MAINTENANCE · MANUFACTURING

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 OBSERVATION
-12H-8H-4HSTOP

EARLY SIGNAL / TIME TO ACT

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.

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.

MODEL INPUTS
  • Time-series sensor values
  • Sparse fault labels
  • Collection intervals
  • Real-time sensor stream
  1. 01

    Clean

    Separate collection errors and sensor malfunctions from meaningful behavior.

  2. 02

    Represent

    Extract features that reveal structure around rare fault events.

  3. 03

    Select

    Compare classical models and deep architectures through offline experiments.

  4. 04

    Observe

    Run the selected model on real-time data and measure operational behavior.

Advance notice the operation can use.

OBSERVED

80%+

faults predicted

during the live observation period
OBSERVED

6-8h

advance notice

on average before a predicted stop
OBSERVED

8

false alarms

across six months of observation

The model creates a practical planning window: enough time to organize maintenance before a stop, with a false-alarm load the operation can absorb.

What looks fixed
may only be familiar.

Bring us the problem