Catching Process Drift
A catalyst slowly degrades. The control loop compensates — temperature stays flat, the mean barely shifts. But pressure oscillations grow from ±1 kPa to ±50 kPa. A winkComposer flow with three detectors catches the instability while it is still developing.
This is a real fault from the Tennessee Eastman Process — the most widely used benchmark in process fault detection. A reactor converts gaseous feed into liquid product under high pressure (~2,800 kPa); unreacted gas recycles through a compressor.
The pipeline streams raw sensor data at 1-minute resolution and monitors two signals: reactor pressure, where variance tells the story, and compressor work, where the correlation with pressure shifts.
Drag the slider and watch the 2σ band widen as the process destabilises.
What You’re Seeing
The chart shows reactor pressure (cyan) tracked by an exponentially smoothed mean (lavender dashed). The filled 2σ band around the mean is the smoothed mean ± twice the smoothed standard deviation — tight during normal operation, widening as variance grows post-fault.
Three detection events unfold in sequence:
-
|z| > 3 exceedance fires first (amber blocks). The adaptive z-score divides each pressure deviation by the smoothed standard deviation. Because the smoothed estimate lags behind the growing variance, early pressure swings punch through ±3σ — catching both positive and negative excursions while the 2σ band is still narrow.
-
Correlation shift fires next (orange vertical). The pressure–compressor correlation starts drifting as the physics decouple. A change-point detector accumulates evidence until the shift is confirmed.
-
Compressor shift fires last (red vertical). An independent change-point test on the smoothed compressor mean detects the energy level drop — a slower, confirmatory signal from a different physical variable.
By the time the 2σ band visibly explodes, all three detectors have already fired.
How It Works
Three pipelines, three detectors — each answers a different question using the same composable building blocks:
Pressure Exceedance catches individual swings. The exponentially smoothed standard deviation updates with a half-life of 90 samples (~1.5 hours), so it lags behind growing variance. Early pressure spikes produce large z-scores even while the mean barely moves — and the detector fires whenever |z| crosses ±3. The amber regions on the chart mark every excursion.
Correlation Drift catches structural change. A Page-Hinkley change-point test accumulates evidence that the pressure–compressor correlation is shifting. A trend detector watches the rate of that accumulation — when evidence climbs consistently, it fires as an early warning before the change-point test confirms the structural shift.
Compressor Shift provides independent confirmation from a different physical variable. The compressor’s smoothed mean energy drops as slowing reactions reduce the gas volume to recycle. This is a subtle shift (~1–2%), so the change-point test fires late — serving as confirmation rather than early detection.
References
- Downs, J.J. & Vogel, E.F. (1993). A plant-wide industrial process control problem. Computers & Chemical Engineering, 17(3), 245–255. doi:10.1016/0098-1354(93)80018-I
- Page, E.S. (1954). Continuous inspection schemes. Biometrika, 41(1/2), 100–115. doi:10.2307/2333009
- Dataset: mv-per/tennessee-eastman-dataset — 3,001 samples at 1-minute resolution (50 hours), fault injected at hour 8. Reactor pressure is XMEAS 7 (kPa), compressor work is XMEAS 20 (kW). Variable definitions from the original FORTRAN simulation .
Next Steps
- Detecting Bearing Failure — predict bearing failure from vibration data
- Composition Patterns — the patterns behind this pipeline