Most writing about industrial soft sensors ends at the point where the model validates. A quality variable that used to arrive from a laboratory hours later now arrives every minute from a model, the held-out error is acceptable, and the project is declared finished. The plants that actually run soft sensors describe something different. For them the validation is the start of a maintenance problem, and the interesting engineering is everything that happens afterwards.
An industrial soft sensor is a model that infers a variable you cannot measure online from variables you can. It is the lowest layer of process AI, the one that measures and does not act. It exists because the quantity that matters commercially, a concentration, an impurity level, a tar loading, is measured offline in a laboratory, infrequently and late. The soft sensor replaces that delay with an estimate. What it does not replace is the reason the laboratory measurement existed, and that distinction decides whether the model survives its first year.
How widely this is already in use
This is not an emerging technique. Engineers at Dow report more than 300 soft sensors deployed globally across the company, inferring concentration and impurity levels from existing temperature, pressure and flow measurements. The motivation they give is the same one that applies to a gasifier or a fermenter: the quality variable is measured offline, the sample count is limited and the delay is significant, which makes it difficult to adjust the process in time.
The same account is candid about how those models are maintained. Updates at Dow are predominantly carried out by engineers retraining the model by hand. The only automated adjustment is a bias term that tracks recent laboratory samples, and once the bias between model and plant grows too large, a new model is required. Rebuilding one means data collection, data processing and retraining, which is to say the maintenance cost is paid in engineering time rather than in hardware.
Why the model decays and the plant does not
A soft sensor is a statistical fit between inputs and an output, and that fit holds only while the relationship it captured holds. Processes rarely sit at a true steady state. Equipment changes, feedstock changes, sensors drift, operating strategy changes, and these move on a slower timescale than the window the model was trained on. In an industrial survey of soft sensor implementations by Kano and Fujiwara, accuracy deterioration caused by changes in process characteristics accounted for 29 percent of reported problems. It is the single largest category, and it is not a modelling error. It is the model being correct about a plant that no longer exists.
The decay has been measured over a realistic horizon. A case study on an industrial process built a partial least squares soft sensor on 735 samples collected over thirteen months, then tested it against 750 samples from the following fifteen months, spanning 29 months of operation in total and several changes in operating condition. Left alone, the original model returned a coefficient of determination of 0.61, a root mean squared error of 0.75 and a mean absolute percentage error of 15.8 percent. With an online update mechanism that retrained it ten times across the test period, the same model reached 0.71, 0.64 and 12.9 percent. The updated model is not dramatically better. The useful number is the other one: a model that was acceptable at commissioning was explaining 61 percent of the variance two years later.
Decay can also be abrupt rather than gradual. A long short-term memory soft sensor trained on an industrial-scale penicillin fermentation dataset, sixty clean batches from a 100,000 litre bioreactor, was then run on ten batches containing process deviations. On three of them the coefficient of determination fell below 0.33 and the average prediction error reached between 2.4 and 2.7 grams per litre against peak concentrations near 13. The model had not degraded slowly. It met a condition outside its training distribution and stopped being informative, in a way that nothing inside the model announced.
Telling a bad model from a bad day
That last point is the practical difficulty. A soft sensor whose prediction has drifted away from the laboratory value is reporting one of three different things, and they call for opposite responses.
The first is a short process upset, which will correct itself and needs no model change. The second is a prolonged process fault, a catalyst change or a turnaround, where the plant has genuinely moved and the model has to follow it. The third is a model fault, where the original fit was poor in a way that only became visible later. Retraining on the first wastes engineering time and can make the model worse by fitting it to a transient. Ignoring the second leaves a controller acting on a number that is no longer true.
The monitoring literature has converged on watching aggregate statistics rather than individual errors. One recent framework runs four indicators over segmented windows, Hotelling's statistic, the squared prediction error, the recovery of the coefficient of determination, and a detection ratio, then uses a second layer to judge whether the behaviour is consistent enough to be a process fault or a model fault. On a pilot-scale distillation case study, short-term faults showed low detection rates near 7 and 8 percent with a quick recovery in the coefficient of determination, marking them as self-correcting. Prolonged process faults showed detection rates near 19 percent with no recovery. A model fault showed a persistent bias in the squared prediction error of about 67 percent. The specific thresholds belong to that process. The structure, a diagnosis before a decision to retrain, transfers.
The best model is usually not the most sophisticated
There is a standing assumption that a soft sensor gets better as the model class gets richer. The measured comparisons do not support it reliably.
A 2025 study developed soft sensors for sulfur content in the product of a full-scale refinery hydrodesulfurization unit, comparing linear system identification models, nonlinear ones, and machine learning models including long short-term memory and gated recurrent unit networks. A static artificial neural network gave the best prediction accuracy of the set. The recurrent networks, which capture temporal dependency and cost considerably more to build and maintain, gave only marginal improvements, and the authors attribute that to the relatively low dynamic complexity of the process. The simpler output error and Hammerstein-Wiener models were described as robust with interpretable structures.
The lesson is not that recurrent networks are useless. It is that model capacity should be matched to the dynamics actually present, and that interpretability has an operational value which does not appear in a held-out error score. A model an engineer can reason about is a model they can diagnose at three in the morning.
What a working soft sensor is worth
The economic case, when it is reported rather than projected, is modest per unit and large in aggregate. A long-running industrial application combined just-in-time soft sensors with commercial model predictive control software at Showa Denko in Japan, running for more than a year on a cracked gasoline fractionator and on a purification section of an acetyl plant. On the fractionator the reported result was a cut of about 0.6 percent in operating cost, alongside a reduction in environmental burden.
Six tenths of one percent sounds small until it is applied to a continuous unit for a year. It is also the right order of magnitude to keep in mind when scoping a soft sensor project, because it sets the budget for everything above: if the model earns fractions of a percent, the maintenance regime has to be cheap enough to leave the benefit intact. That, rather than estimation accuracy, is the constraint that decides whether an industrial soft sensor is worth building.
Where our own stack sits
Our control work treats the estimator and the thing that consumes it as one deliverable rather than two. A soft sensor that feeds an operator display can be wrong occasionally and be corrected by the person reading it. A soft sensor that feeds a controller is a different object, because its error becomes an action, which is why a learned estimate in our stack sits behind a safety layer that answers in under 2 milliseconds and can veto an action before it reaches an actuator, and why the end-to-end path from sensor to command is held at 285 milliseconds, reduced from 1.2 seconds.
The validation record matters for the same reason. On the 500 litre pilot basin the learned model reaches an R-squared above 0.95, and sim-to-real validation ran across 400 simulated years with zero recorded safety violations. Read the simulated years as coverage of the state space rather than elapsed time. Against the decay described above, coverage is the point: a model is only as durable as the range of conditions it was shown, and the honest version of that claim names the range rather than the headline accuracy.
Building an industrial soft sensor that survives its first year
Four things separate a soft sensor still in service after two years from one quietly switched off.
Budget the maintenance before the model. If nobody owns retraining, the model has an expiry date that nobody has written down. Instrument the decay, not just the accuracy, with an indicator that distinguishes a transient from a genuine shift in the plant, so that retraining is triggered by diagnosis rather than by someone noticing. Match the model class to the dynamics, and prefer the structure an engineer can interpret when the accuracy difference is marginal. And decide at the design stage whether the estimate will ever drive an actuator, because that single question determines whether a validation record and a hard safety layer are optional or mandatory.
Deriving the equations of a process first, fitting only what the equations cannot give you, and handing over the record that lets someone else re-run the whole thing, is the method we bring to contract research and development for industrial control. If an industrial soft sensor is the piece you are missing, talk to the team.