I spent two years helping a gas processing plant build a digital twin of their main compression train. The vendor demos looked incredible. A spinning 3D model with live pressures, temperatures, vibration overlays. Everyone thought we would flip a switch and start predicting failures. We did not. The first month, the twin told us everything was healthy four hours before an actual bearing failure shut down the line.
That failure happened because we modelled the wrong thing. We built the twin from engineering drawings and design specs, not from what the plant actually did. The drawings said this compressor should run at a certain vibration threshold. But after ten years of operations, field technicians had adjusted setpoints, swapped parts, and bypassed a couple of nuisance alarms. The twin didn't know any of that. It was a beautiful, expensive mirror of a compressor that no longer existed.
The lesson for me was that a digital twin is not a CAD model with live data glued on. It has to learn from history. We went back and pulled two years of historian data, including the messy stuff. Periods where sensors dropped out for hours. Times when operators manually logged readings because an instrument was known to drift. We fed that mess into the model and retrained it. Suddenly the twin could tell the difference between normal drift and an actual problem.
In oil and gas, you rarely get clean data. You get corrosive environments, remote sites, and instruments that freeze in winter. Our first twin assumed every tag was perfect. It wasn't. A pressure transmitter on the discharge line had been reading 5 percent high for six months. The twin compensated for it as if it were real. When a real pressure spike came, the twin smoothed it out because it thought the sensor was just misbehaving again. We nearly missed a relief valve chattering.
What made it work eventually was putting operators in the room, not just data scientists. One operator told us the twin was useless because it didn't know the plant had a 'Thursday afternoon quirk' where the fuel gas header pressure dipped when the utility boiler switched modes. That kind of tribal knowledge never shows up in a SCADA tag. We built a rule that accounted for it, and the false alerts dropped way down.
Now I tell clients that a digital twin is 20 percent technology and 80 percent getting the story right. You need to know which sensors you can trust, which setpoints are real, and which alarms operators ignore. If you skip that, you get a very expensive screensaver with a live feed. The compressor failure we had was avoidable. That's the part that stuck with me.