Neural networks find process applications

Scanzorosciate, Italy-Neural networks have already found important applications in a number of technological areas, including machine vision, signal processing, and medical instrumentation, among others. Now, an Italian company, TRI srl, of Scanzorosciate, has developed a number of applications in the field of process engineering.

Neural networks are generally used to learn, from experimental data, the relationship between a given parameter and a number of measurable variables. They are particularly superior to conventional curvefitting methods when the variables are many, the function is nonlinear, and the number of mixed partial derivatives in the equation is relatively large.

One of the uses of neural networks is in constructing a "soft sensor" or "virtual analyzer" by training the network with data describing the real instruments response to a series of known inputs. TRI's applications are based on software products developed by Pavilion Technology, Austin, Tex. One replaces, with a software sensor, analyzers used to monitor the emissions of CO and NOx from large industrial chimneys and stacks. Other soft-sensor applications predict the melt index of extruded polymer mixtures and the composition of effluentform distillation columns.

The validity of a parameter estimated by neural networks, made on the basis of many measured values, can be called into question since the probability is high that at least one of the measurements is in error. To solve this problem, another type of neural network, called "sensor validation network," can be used. In this case, a network is trained to reproduce each of the measured variables as a function of the complete set of the same variables. The value of a variable that deviates from the prediction is then recognized as inconsistent, an alarm is given, and the measured value is replaced by the predicted one.

This method requires that some functional redundancy is present in the set of variables (i.e., the number of variables contains at least two correlated variables more than the number strictly necessary to describe the status of the system). In the particular case of the prediction of stack emissions, the combination of a sensor validator and of a soft sensor typically allows an availability greater than 99% to be achieved, compared with the 75%-85% values given by many commercial, hardware-only analyzers. Sensor validation by neural networks is also an important application in its own right since it provides a way to detect sensor failures that cannot be recognized by the standard tests performed by most types of controller software.

One of the most sophisticated uses of neural networks in process control engineering is set-point optimization. Neural network models are used in combination with numerical optimization techniques to advise on the best control settings to, for example, increase product yield, reduce energy consumption, or maintain the concentration of a certain impurity within limits.

Improvements of a few percent, or even less in very large process plants, can produce millions of dollars in savings in one year, and a number of applications already made achieved payback in less than four months. In general, most of the applications have been open-loop, but some closed-loop installations are now operating in a group of refineries in the U.S.

- Cris Whetton

Copyright Instrument Society of America Jan 1997
Provided by ProQuest Information and Learning Company. All rights Reserved