One of the early pioneers of soft sensors was Pavilion Technologies, Austin, Texas. Over the last two to three years the company has put a lot of effort into developing a hybrid approach to inferential sensors, says senior product manager Ric Snyder. "There was no perfect first principles model, and no perfect empirical model. Both have their pluses and minuses, but we have a lot of runtime now on the hybrid model and use it built into our existing MPC [model predictive control] solution. All of the chemical plants where we have our control software tend to have at least one or two inferential sensors running. It may be just for a purity measurement on a distillation column, estimating impurities online, or something like melt index in a polymer reactor or extruder — typically for a quality measure that you're actually selling against."
The hybrid model also has opened up other markets for Pavilion, most notably in the pharmaceuticals sector. For instance, by providing online measurements from fermentation processes, notes Snyder, Pavilion's model helps operators optimize their feed into the fermenter (Figure 4). "We've been able to reduce batch times by anywhere from 10% to 20%," he says.
Figure 4. Use of both empirical and first-principles models has speeded up fermentation batches. Source: Pavilion Technologies
While Pavilion's solutions are still essentially neural-network-based, Matrikon's ProcessMonitor product uses a partial least squares (PLS) algorithm to process its data. "The beauty of doing it this way," contends Brown, "is that the PLS structure we use is very good at providing a tremendous amount of insight into how the model you're developing ties back into the fundamental variables of the process that you understand."
Back at Dow, Kordon says two types of inferential sensors are in use around the company. Earlier ones are based on neural networks, such as those from Pavilion that have been used successfully on NOx emission control duties (see What’s in the air for continuous emissions monitoring? for more on predictive emissions monitoring systems). "It's not closed-loop control," he says, "but is generally accepted for the limits of the [emission] tests involved."
However, for other solutions for estimating properties and for closed loop control, Kordon says Dow "is gradually migrating to a genetic programming technology, especially its capability to derive a symbolic regression model — which looks not only at the accuracy of the data but its complexity as well, so we can find solutions that are relatively robust compared to neural nets."
This technology for the moment remains internal to Dow and the company filed patents for it earlier this year. However, Kordon believes more than technology might be needed to improve the acceptance of inferential sensors generally across industry.
"The key challenge is the psychology, the credibility, the risk-taking culture that's required," he says. "It's not really a technical issue — it's relatively mature — but it won't ‘turn your data into gold' as was claimed at the beginning. But it is a viable solution, and we're using it."