Working on a research project (embedded ML + reliability) looking at inferential/soft sensors in process plants (distillation, reactors, etc).

Curious about real practice: when operating conditions drift from what the soft sensor was tuned on, how do you find out? Scheduled recalibration only, redundant hardware check, or something else? And when it does go unnoticed for a while, what does that actually cost (bad control moves, off-spec product, alarms)?

Not selling anything, just trying to understand if this is a real daily pain or something plants have already solved well enough. Any war stories welcome.

 

I'm not sure what a soft sensor is, but this Chemical Safety Board video to fthe BP Texas City refinery fire/explosions (2005?) were in part based false readings from level sensors.

 

The dangerous case is usually not a failed model but a plausible-looking prediction outside its validated operating envelope. A robust setup should monitor input-sensor health, distance from the training domain and prediction residuals whenever laboratory or analyzer measurements become available.

Scheduled recalibration alone is insufficient. Use range/rate checks, multivariate drift statistics, uncertainty limits and periodic comparison with independent samples. This embedded anomaly-detection project also highlights practical problems such as false positives, false negatives and the need to tune a model for its specific equipment.

If confidence falls, the estimate should be flagged or removed from closed-loop control. Otherwise, unnoticed drift can produce biased control moves, excess energy consumption, off-spec product and misleading alarms.

 

So how many process variables are you monitoring and the level of measurement redundancy required, together with the number of control variables being used to achieve the desired outcomes.