Picture a process unit where an optimization model has been recommending setpoints for months. Yields creep up, energy use drops, and everyone is pleased. Then someone checks the level transmitter and finds it has been reading a few percent high since spring. The model's math was fine. Its picture of the tank wasn't. That's the quiet truth behind most industrial AI: the data comes from instruments, four of them mostly, measuring pressure, temperature, flow, and level.
Software can only work with what it's given
An algorithm has no direct line to your process. It never sees the tank, the pipe, or the pump. It sees numbers, and those numbers came from a transmitter or probe that somebody selected, installed, and configured. If a reading is off, the model usually can't tell. It treats the bad number as truth.
Most AI in plants today is practical rather than sci-fi: predictive maintenance, anomaly detection, soft sensors that estimate a hard-to-measure quality from easier measurements, and advanced control or optimization layers that sit above the regular PID loops, often recommending moves for operators to approve. Every one of them is trained on, and checked against, real instrument signals. A soft sensor is only as good as the measurements feeding it and the lab samples it was calibrated with. A sloppy measurement used to cause a sloppy loop. Now it can quietly teach a model the wrong behavior.
The big four, and how each one goes wrong
Pressure. Impulse lines that plug, freeze, or trap air produce readings that look plausible and aren't. Range matters too. A device sized for far higher pressure than your process uses has poor resolution where you actually operate, and a model can't recover detail the sensor never captured.
Temperature. A thermowell that's too short, or an element that doesn't reach the flowing fluid, reads closer to the pipe wall than the process. Response time is another quiet culprit. A heavy thermowell can lag real changes, which hurts any controller that depends on seeing trends early.
Flow. Many flow meters need straight pipe upstream and downstream to see a stable velocity profile. Install one right after an elbow or a partly open valve and it can be consistently wrong. A meter that's perfect on clean water may struggle with slurry or a product whose density changes.
Level. Foam, vapor, buildup on a probe, agitation, and changing density can all fool a level device. What suits a clean water tank may be a poor choice for a sticky or coating product.
None of these are exotic. Analytics can sometimes flag a suspicious sensor by comparing it with related measurements, which is useful, but it can't tell you the meter sits too close to an elbow.
Drift is the sneaky one
Sensors rarely fail dramatically. They drift. A pressure transmitter slowly walks off zero, an element ages, a probe picks up a coating. Because the change is gradual, it can look like the process itself is shifting, and a model retrained on that data may learn the drift as normal. Regular calibration keeps the data tied to reality.
Where the human expert comes in
This part doesn't get enough credit. Choosing the right instrument is a judgment call, and it comes from experience. Someone has to ask the questions. What's the fluid? What are the real operating and upset conditions? What does the piping look like? Is the area hazardous, the service corrosive, the process hygienic?
A knowledgeable local instrumentation representative asks those questions before a purchase order is written. They've seen what holds up in a paper mill, a water plant, or a food facility, and they'll tell you when the instrument you asked for isn't the one you need. They also stick around after the sale for startup, calibration, and troubleshooting.
A data sheet tells you what an instrument can do. A good rep tells you what it will do in your plant.
Start with the measurement, then the model
If you're planning an AI or advanced analytics project, walk the process first. Which measurements does the model depend on most? When were those instruments last verified? Are they the right technology for the service, installed where they can actually see the process? Fixing a few weak measurements usually costs less than a software license.
For manufacturers in New England, one resource worth knowing is PSI Controls, the process controls and instrumentation division of Piping Specialties, with offices in Portland, Maine and Danvers, Massachusetts. They handle system design, sales, and service for pressure, temperature, flow, and level measurement along with automated valves. If your plant is getting ready to lean harder on data, that's a good conversation to have first.
